1 /*
2 * kernel/sched/core.c
3 *
4 * Kernel scheduler and related syscalls
5 *
6 * Copyright (C) 1991-2002 Linus Torvalds
7 *
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
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
27 */
29 #include <linux/mm.h>
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
78 #include <asm/tlb.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
83 #endif
85 #include "sched.h"
86 #include "../workqueue_sched.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
93 {
94 unsigned long delta;
95 ktime_t soft, hard, now;
97 for (;;) {
98 if (hrtimer_active(period_timer))
99 break;
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
109 }
110 }
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
118 {
119 s64 delta;
121 if (rq->skip_clock_update > 0)
122 return;
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
125 rq->clock += delta;
126 update_rq_clock_task(rq, delta);
127 }
129 /*
130 * Debugging: various feature bits
131 */
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
138 0;
140 #undef SCHED_FEAT
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
144 #name ,
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
148 };
150 #undef SCHED_FEAT
152 static int sched_feat_show(struct seq_file *m, void *v)
153 {
154 int i;
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
158 seq_puts(m, "NO_");
159 seq_printf(m, "%s ", sched_feat_names[i]);
160 }
161 seq_puts(m, "\n");
163 return 0;
164 }
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
176 };
178 #undef SCHED_FEAT
180 static void sched_feat_disable(int i)
181 {
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
184 }
186 static void sched_feat_enable(int i)
187 {
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
190 }
191 #else
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp)
197 {
198 int i;
199 int neg = 0;
201 if (strncmp(cmp, "NO_", 3) == 0) {
202 neg = 1;
203 cmp += 3;
204 }
206 for (i = 0; i < __SCHED_FEAT_NR; i++) {
207 if (strcmp(cmp, sched_feat_names[i]) == 0) {
208 if (neg) {
209 sysctl_sched_features &= ~(1UL << i);
210 sched_feat_disable(i);
211 } else {
212 sysctl_sched_features |= (1UL << i);
213 sched_feat_enable(i);
214 }
215 break;
216 }
217 }
219 return i;
220 }
222 static ssize_t
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224 size_t cnt, loff_t *ppos)
225 {
226 char buf[64];
227 char *cmp;
228 int i;
230 if (cnt > 63)
231 cnt = 63;
233 if (copy_from_user(&buf, ubuf, cnt))
234 return -EFAULT;
236 buf[cnt] = 0;
237 cmp = strstrip(buf);
239 i = sched_feat_set(cmp);
240 if (i == __SCHED_FEAT_NR)
241 return -EINVAL;
243 *ppos += cnt;
245 return cnt;
246 }
248 static int sched_feat_open(struct inode *inode, struct file *filp)
249 {
250 return single_open(filp, sched_feat_show, NULL);
251 }
253 static const struct file_operations sched_feat_fops = {
254 .open = sched_feat_open,
255 .write = sched_feat_write,
256 .read = seq_read,
257 .llseek = seq_lseek,
258 .release = single_release,
259 };
261 static __init int sched_init_debug(void)
262 {
263 debugfs_create_file("sched_features", 0644, NULL, NULL,
264 &sched_feat_fops);
266 return 0;
267 }
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
271 /*
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
274 */
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
277 /*
278 * period over which we average the RT time consumption, measured
279 * in ms.
280 *
281 * default: 1s
282 */
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
285 /*
286 * period over which we measure -rt task cpu usage in us.
287 * default: 1s
288 */
289 unsigned int sysctl_sched_rt_period = 1000000;
291 __read_mostly int scheduler_running;
293 /*
294 * part of the period that we allow rt tasks to run in us.
295 * default: 0.95s
296 */
297 int sysctl_sched_rt_runtime = 950000;
301 /*
302 * __task_rq_lock - lock the rq @p resides on.
303 */
304 static inline struct rq *__task_rq_lock(struct task_struct *p)
305 __acquires(rq->lock)
306 {
307 struct rq *rq;
309 lockdep_assert_held(&p->pi_lock);
311 for (;;) {
312 rq = task_rq(p);
313 raw_spin_lock(&rq->lock);
314 if (likely(rq == task_rq(p)))
315 return rq;
316 raw_spin_unlock(&rq->lock);
317 }
318 }
320 /*
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
322 */
323 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
324 __acquires(p->pi_lock)
325 __acquires(rq->lock)
326 {
327 struct rq *rq;
329 for (;;) {
330 raw_spin_lock_irqsave(&p->pi_lock, *flags);
331 rq = task_rq(p);
332 raw_spin_lock(&rq->lock);
333 if (likely(rq == task_rq(p)))
334 return rq;
335 raw_spin_unlock(&rq->lock);
336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
337 }
338 }
340 static void __task_rq_unlock(struct rq *rq)
341 __releases(rq->lock)
342 {
343 raw_spin_unlock(&rq->lock);
344 }
346 static inline void
347 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
348 __releases(rq->lock)
349 __releases(p->pi_lock)
350 {
351 raw_spin_unlock(&rq->lock);
352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
353 }
355 /*
356 * this_rq_lock - lock this runqueue and disable interrupts.
357 */
358 static struct rq *this_rq_lock(void)
359 __acquires(rq->lock)
360 {
361 struct rq *rq;
363 local_irq_disable();
364 rq = this_rq();
365 raw_spin_lock(&rq->lock);
367 return rq;
368 }
370 #ifdef CONFIG_SCHED_HRTICK
371 /*
372 * Use HR-timers to deliver accurate preemption points.
373 *
374 * Its all a bit involved since we cannot program an hrt while holding the
375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
376 * reschedule event.
377 *
378 * When we get rescheduled we reprogram the hrtick_timer outside of the
379 * rq->lock.
380 */
382 static void hrtick_clear(struct rq *rq)
383 {
384 if (hrtimer_active(&rq->hrtick_timer))
385 hrtimer_cancel(&rq->hrtick_timer);
386 }
388 /*
389 * High-resolution timer tick.
390 * Runs from hardirq context with interrupts disabled.
391 */
392 static enum hrtimer_restart hrtick(struct hrtimer *timer)
393 {
394 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
396 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
398 raw_spin_lock(&rq->lock);
399 update_rq_clock(rq);
400 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
401 raw_spin_unlock(&rq->lock);
403 return HRTIMER_NORESTART;
404 }
406 #ifdef CONFIG_SMP
407 /*
408 * called from hardirq (IPI) context
409 */
410 static void __hrtick_start(void *arg)
411 {
412 struct rq *rq = arg;
414 raw_spin_lock(&rq->lock);
415 hrtimer_restart(&rq->hrtick_timer);
416 rq->hrtick_csd_pending = 0;
417 raw_spin_unlock(&rq->lock);
418 }
420 /*
421 * Called to set the hrtick timer state.
422 *
423 * called with rq->lock held and irqs disabled
424 */
425 void hrtick_start(struct rq *rq, u64 delay)
426 {
427 struct hrtimer *timer = &rq->hrtick_timer;
428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 hrtimer_set_expires(timer, time);
432 if (rq == this_rq()) {
433 hrtimer_restart(timer);
434 } else if (!rq->hrtick_csd_pending) {
435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
436 rq->hrtick_csd_pending = 1;
437 }
438 }
440 static int
441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
442 {
443 int cpu = (int)(long)hcpu;
445 switch (action) {
446 case CPU_UP_CANCELED:
447 case CPU_UP_CANCELED_FROZEN:
448 case CPU_DOWN_PREPARE:
449 case CPU_DOWN_PREPARE_FROZEN:
450 case CPU_DEAD:
451 case CPU_DEAD_FROZEN:
452 hrtick_clear(cpu_rq(cpu));
453 return NOTIFY_OK;
454 }
456 return NOTIFY_DONE;
457 }
459 static __init void init_hrtick(void)
460 {
461 hotcpu_notifier(hotplug_hrtick, 0);
462 }
463 #else
464 /*
465 * Called to set the hrtick timer state.
466 *
467 * called with rq->lock held and irqs disabled
468 */
469 void hrtick_start(struct rq *rq, u64 delay)
470 {
471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
472 HRTIMER_MODE_REL_PINNED, 0);
473 }
475 static inline void init_hrtick(void)
476 {
477 }
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq *rq)
481 {
482 #ifdef CONFIG_SMP
483 rq->hrtick_csd_pending = 0;
485 rq->hrtick_csd.flags = 0;
486 rq->hrtick_csd.func = __hrtick_start;
487 rq->hrtick_csd.info = rq;
488 #endif
490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
491 rq->hrtick_timer.function = hrtick;
492 }
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq *rq)
495 {
496 }
498 static inline void init_rq_hrtick(struct rq *rq)
499 {
500 }
502 static inline void init_hrtick(void)
503 {
504 }
505 #endif /* CONFIG_SCHED_HRTICK */
507 /*
508 * resched_task - mark a task 'to be rescheduled now'.
509 *
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
512 * the target CPU.
513 */
514 #ifdef CONFIG_SMP
516 #ifndef tsk_is_polling
517 #define tsk_is_polling(t) 0
518 #endif
520 void resched_task(struct task_struct *p)
521 {
522 int cpu;
524 assert_raw_spin_locked(&task_rq(p)->lock);
526 if (test_tsk_need_resched(p))
527 return;
529 set_tsk_need_resched(p);
531 cpu = task_cpu(p);
532 if (cpu == smp_processor_id())
533 return;
535 /* NEED_RESCHED must be visible before we test polling */
536 smp_mb();
537 if (!tsk_is_polling(p))
538 smp_send_reschedule(cpu);
539 }
541 void resched_cpu(int cpu)
542 {
543 struct rq *rq = cpu_rq(cpu);
544 unsigned long flags;
546 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
547 return;
548 resched_task(cpu_curr(cpu));
549 raw_spin_unlock_irqrestore(&rq->lock, flags);
550 }
552 #ifdef CONFIG_NO_HZ
553 /*
554 * In the semi idle case, use the nearest busy cpu for migrating timers
555 * from an idle cpu. This is good for power-savings.
556 *
557 * We don't do similar optimization for completely idle system, as
558 * selecting an idle cpu will add more delays to the timers than intended
559 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
560 */
561 int get_nohz_timer_target(void)
562 {
563 int cpu = smp_processor_id();
564 int i;
565 struct sched_domain *sd;
567 rcu_read_lock();
568 for_each_domain(cpu, sd) {
569 for_each_cpu(i, sched_domain_span(sd)) {
570 if (!idle_cpu(i)) {
571 cpu = i;
572 goto unlock;
573 }
574 }
575 }
576 unlock:
577 rcu_read_unlock();
578 return cpu;
579 }
580 /*
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
589 */
590 void wake_up_idle_cpu(int cpu)
591 {
592 struct rq *rq = cpu_rq(cpu);
594 if (cpu == smp_processor_id())
595 return;
597 /*
598 * This is safe, as this function is called with the timer
599 * wheel base lock of (cpu) held. When the CPU is on the way
600 * to idle and has not yet set rq->curr to idle then it will
601 * be serialized on the timer wheel base lock and take the new
602 * timer into account automatically.
603 */
604 if (rq->curr != rq->idle)
605 return;
607 /*
608 * We can set TIF_RESCHED on the idle task of the other CPU
609 * lockless. The worst case is that the other CPU runs the
610 * idle task through an additional NOOP schedule()
611 */
612 set_tsk_need_resched(rq->idle);
614 /* NEED_RESCHED must be visible before we test polling */
615 smp_mb();
616 if (!tsk_is_polling(rq->idle))
617 smp_send_reschedule(cpu);
618 }
620 static inline bool got_nohz_idle_kick(void)
621 {
622 int cpu = smp_processor_id();
623 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
624 }
626 #else /* CONFIG_NO_HZ */
628 static inline bool got_nohz_idle_kick(void)
629 {
630 return false;
631 }
633 #endif /* CONFIG_NO_HZ */
635 void sched_avg_update(struct rq *rq)
636 {
637 s64 period = sched_avg_period();
639 while ((s64)(rq->clock - rq->age_stamp) > period) {
640 /*
641 * Inline assembly required to prevent the compiler
642 * optimising this loop into a divmod call.
643 * See __iter_div_u64_rem() for another example of this.
644 */
645 asm("" : "+rm" (rq->age_stamp));
646 rq->age_stamp += period;
647 rq->rt_avg /= 2;
648 }
649 }
651 #else /* !CONFIG_SMP */
652 void resched_task(struct task_struct *p)
653 {
654 assert_raw_spin_locked(&task_rq(p)->lock);
655 set_tsk_need_resched(p);
656 }
657 #endif /* CONFIG_SMP */
659 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
660 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
661 /*
662 * Iterate task_group tree rooted at *from, calling @down when first entering a
663 * node and @up when leaving it for the final time.
664 *
665 * Caller must hold rcu_lock or sufficient equivalent.
666 */
667 int walk_tg_tree_from(struct task_group *from,
668 tg_visitor down, tg_visitor up, void *data)
669 {
670 struct task_group *parent, *child;
671 int ret;
673 parent = from;
675 down:
676 ret = (*down)(parent, data);
677 if (ret)
678 goto out;
679 list_for_each_entry_rcu(child, &parent->children, siblings) {
680 parent = child;
681 goto down;
683 up:
684 continue;
685 }
686 ret = (*up)(parent, data);
687 if (ret || parent == from)
688 goto out;
690 child = parent;
691 parent = parent->parent;
692 if (parent)
693 goto up;
694 out:
695 return ret;
696 }
698 int tg_nop(struct task_group *tg, void *data)
699 {
700 return 0;
701 }
702 #endif
704 static void set_load_weight(struct task_struct *p)
705 {
706 int prio = p->static_prio - MAX_RT_PRIO;
707 struct load_weight *load = &p->se.load;
709 /*
710 * SCHED_IDLE tasks get minimal weight:
711 */
712 if (p->policy == SCHED_IDLE) {
713 load->weight = scale_load(WEIGHT_IDLEPRIO);
714 load->inv_weight = WMULT_IDLEPRIO;
715 return;
716 }
718 load->weight = scale_load(prio_to_weight[prio]);
719 load->inv_weight = prio_to_wmult[prio];
720 }
722 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
723 {
724 update_rq_clock(rq);
725 sched_info_queued(p);
726 p->sched_class->enqueue_task(rq, p, flags);
727 }
729 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
730 {
731 update_rq_clock(rq);
732 sched_info_dequeued(p);
733 p->sched_class->dequeue_task(rq, p, flags);
734 }
736 void activate_task(struct rq *rq, struct task_struct *p, int flags)
737 {
738 if (task_contributes_to_load(p))
739 rq->nr_uninterruptible--;
741 enqueue_task(rq, p, flags);
742 }
744 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
745 {
746 if (task_contributes_to_load(p))
747 rq->nr_uninterruptible++;
749 dequeue_task(rq, p, flags);
750 }
752 static void update_rq_clock_task(struct rq *rq, s64 delta)
753 {
754 /*
755 * In theory, the compile should just see 0 here, and optimize out the call
756 * to sched_rt_avg_update. But I don't trust it...
757 */
758 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
759 s64 steal = 0, irq_delta = 0;
760 #endif
761 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
762 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
764 /*
765 * Since irq_time is only updated on {soft,}irq_exit, we might run into
766 * this case when a previous update_rq_clock() happened inside a
767 * {soft,}irq region.
768 *
769 * When this happens, we stop ->clock_task and only update the
770 * prev_irq_time stamp to account for the part that fit, so that a next
771 * update will consume the rest. This ensures ->clock_task is
772 * monotonic.
773 *
774 * It does however cause some slight miss-attribution of {soft,}irq
775 * time, a more accurate solution would be to update the irq_time using
776 * the current rq->clock timestamp, except that would require using
777 * atomic ops.
778 */
779 if (irq_delta > delta)
780 irq_delta = delta;
782 rq->prev_irq_time += irq_delta;
783 delta -= irq_delta;
784 #endif
785 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
786 if (static_key_false((¶virt_steal_rq_enabled))) {
787 u64 st;
789 steal = paravirt_steal_clock(cpu_of(rq));
790 steal -= rq->prev_steal_time_rq;
792 if (unlikely(steal > delta))
793 steal = delta;
795 st = steal_ticks(steal);
796 steal = st * TICK_NSEC;
798 rq->prev_steal_time_rq += steal;
800 delta -= steal;
801 }
802 #endif
804 rq->clock_task += delta;
806 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
807 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
808 sched_rt_avg_update(rq, irq_delta + steal);
809 #endif
810 }
812 void sched_set_stop_task(int cpu, struct task_struct *stop)
813 {
814 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
815 struct task_struct *old_stop = cpu_rq(cpu)->stop;
817 if (stop) {
818 /*
819 * Make it appear like a SCHED_FIFO task, its something
820 * userspace knows about and won't get confused about.
821 *
822 * Also, it will make PI more or less work without too
823 * much confusion -- but then, stop work should not
824 * rely on PI working anyway.
825 */
826 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
828 stop->sched_class = &stop_sched_class;
829 }
831 cpu_rq(cpu)->stop = stop;
833 if (old_stop) {
834 /*
835 * Reset it back to a normal scheduling class so that
836 * it can die in pieces.
837 */
838 old_stop->sched_class = &rt_sched_class;
839 }
840 }
842 /*
843 * __normal_prio - return the priority that is based on the static prio
844 */
845 static inline int __normal_prio(struct task_struct *p)
846 {
847 return p->static_prio;
848 }
850 /*
851 * Calculate the expected normal priority: i.e. priority
852 * without taking RT-inheritance into account. Might be
853 * boosted by interactivity modifiers. Changes upon fork,
854 * setprio syscalls, and whenever the interactivity
855 * estimator recalculates.
856 */
857 static inline int normal_prio(struct task_struct *p)
858 {
859 int prio;
861 if (task_has_rt_policy(p))
862 prio = MAX_RT_PRIO-1 - p->rt_priority;
863 else
864 prio = __normal_prio(p);
865 return prio;
866 }
868 /*
869 * Calculate the current priority, i.e. the priority
870 * taken into account by the scheduler. This value might
871 * be boosted by RT tasks, or might be boosted by
872 * interactivity modifiers. Will be RT if the task got
873 * RT-boosted. If not then it returns p->normal_prio.
874 */
875 static int effective_prio(struct task_struct *p)
876 {
877 p->normal_prio = normal_prio(p);
878 /*
879 * If we are RT tasks or we were boosted to RT priority,
880 * keep the priority unchanged. Otherwise, update priority
881 * to the normal priority:
882 */
883 if (!rt_prio(p->prio))
884 return p->normal_prio;
885 return p->prio;
886 }
888 /**
889 * task_curr - is this task currently executing on a CPU?
890 * @p: the task in question.
891 */
892 inline int task_curr(const struct task_struct *p)
893 {
894 return cpu_curr(task_cpu(p)) == p;
895 }
897 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
898 const struct sched_class *prev_class,
899 int oldprio)
900 {
901 if (prev_class != p->sched_class) {
902 if (prev_class->switched_from)
903 prev_class->switched_from(rq, p);
904 p->sched_class->switched_to(rq, p);
905 } else if (oldprio != p->prio)
906 p->sched_class->prio_changed(rq, p, oldprio);
907 }
909 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
910 {
911 const struct sched_class *class;
913 if (p->sched_class == rq->curr->sched_class) {
914 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
915 } else {
916 for_each_class(class) {
917 if (class == rq->curr->sched_class)
918 break;
919 if (class == p->sched_class) {
920 resched_task(rq->curr);
921 break;
922 }
923 }
924 }
926 /*
927 * A queue event has occurred, and we're going to schedule. In
928 * this case, we can save a useless back to back clock update.
929 */
930 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
931 rq->skip_clock_update = 1;
932 }
934 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
936 void register_task_migration_notifier(struct notifier_block *n)
937 {
938 atomic_notifier_chain_register(&task_migration_notifier, n);
939 }
941 #ifdef CONFIG_SMP
942 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
943 {
944 #ifdef CONFIG_SCHED_DEBUG
945 /*
946 * We should never call set_task_cpu() on a blocked task,
947 * ttwu() will sort out the placement.
948 */
949 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
950 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
952 #ifdef CONFIG_LOCKDEP
953 /*
954 * The caller should hold either p->pi_lock or rq->lock, when changing
955 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
956 *
957 * sched_move_task() holds both and thus holding either pins the cgroup,
958 * see task_group().
959 *
960 * Furthermore, all task_rq users should acquire both locks, see
961 * task_rq_lock().
962 */
963 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
964 lockdep_is_held(&task_rq(p)->lock)));
965 #endif
966 #endif
968 trace_sched_migrate_task(p, new_cpu);
970 if (task_cpu(p) != new_cpu) {
971 struct task_migration_notifier tmn;
973 if (p->sched_class->migrate_task_rq)
974 p->sched_class->migrate_task_rq(p, new_cpu);
975 p->se.nr_migrations++;
976 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
978 tmn.task = p;
979 tmn.from_cpu = task_cpu(p);
980 tmn.to_cpu = new_cpu;
982 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
983 }
985 __set_task_cpu(p, new_cpu);
986 }
988 struct migration_arg {
989 struct task_struct *task;
990 int dest_cpu;
991 };
993 static int migration_cpu_stop(void *data);
995 /*
996 * wait_task_inactive - wait for a thread to unschedule.
997 *
998 * If @match_state is nonzero, it's the @p->state value just checked and
999 * not expected to change. If it changes, i.e. @p might have woken up,
1000 * then return zero. When we succeed in waiting for @p to be off its CPU,
1001 * we return a positive number (its total switch count). If a second call
1002 * a short while later returns the same number, the caller can be sure that
1003 * @p has remained unscheduled the whole time.
1004 *
1005 * The caller must ensure that the task *will* unschedule sometime soon,
1006 * else this function might spin for a *long* time. This function can't
1007 * be called with interrupts off, or it may introduce deadlock with
1008 * smp_call_function() if an IPI is sent by the same process we are
1009 * waiting to become inactive.
1010 */
1011 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1012 {
1013 unsigned long flags;
1014 int running, on_rq;
1015 unsigned long ncsw;
1016 struct rq *rq;
1018 for (;;) {
1019 /*
1020 * We do the initial early heuristics without holding
1021 * any task-queue locks at all. We'll only try to get
1022 * the runqueue lock when things look like they will
1023 * work out!
1024 */
1025 rq = task_rq(p);
1027 /*
1028 * If the task is actively running on another CPU
1029 * still, just relax and busy-wait without holding
1030 * any locks.
1031 *
1032 * NOTE! Since we don't hold any locks, it's not
1033 * even sure that "rq" stays as the right runqueue!
1034 * But we don't care, since "task_running()" will
1035 * return false if the runqueue has changed and p
1036 * is actually now running somewhere else!
1037 */
1038 while (task_running(rq, p)) {
1039 if (match_state && unlikely(p->state != match_state))
1040 return 0;
1041 cpu_relax();
1042 }
1044 /*
1045 * Ok, time to look more closely! We need the rq
1046 * lock now, to be *sure*. If we're wrong, we'll
1047 * just go back and repeat.
1048 */
1049 rq = task_rq_lock(p, &flags);
1050 trace_sched_wait_task(p);
1051 running = task_running(rq, p);
1052 on_rq = p->on_rq;
1053 ncsw = 0;
1054 if (!match_state || p->state == match_state)
1055 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1056 task_rq_unlock(rq, p, &flags);
1058 /*
1059 * If it changed from the expected state, bail out now.
1060 */
1061 if (unlikely(!ncsw))
1062 break;
1064 /*
1065 * Was it really running after all now that we
1066 * checked with the proper locks actually held?
1067 *
1068 * Oops. Go back and try again..
1069 */
1070 if (unlikely(running)) {
1071 cpu_relax();
1072 continue;
1073 }
1075 /*
1076 * It's not enough that it's not actively running,
1077 * it must be off the runqueue _entirely_, and not
1078 * preempted!
1079 *
1080 * So if it was still runnable (but just not actively
1081 * running right now), it's preempted, and we should
1082 * yield - it could be a while.
1083 */
1084 if (unlikely(on_rq)) {
1085 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1087 set_current_state(TASK_UNINTERRUPTIBLE);
1088 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1089 continue;
1090 }
1092 /*
1093 * Ahh, all good. It wasn't running, and it wasn't
1094 * runnable, which means that it will never become
1095 * running in the future either. We're all done!
1096 */
1097 break;
1098 }
1100 return ncsw;
1101 }
1103 /***
1104 * kick_process - kick a running thread to enter/exit the kernel
1105 * @p: the to-be-kicked thread
1106 *
1107 * Cause a process which is running on another CPU to enter
1108 * kernel-mode, without any delay. (to get signals handled.)
1109 *
1110 * NOTE: this function doesn't have to take the runqueue lock,
1111 * because all it wants to ensure is that the remote task enters
1112 * the kernel. If the IPI races and the task has been migrated
1113 * to another CPU then no harm is done and the purpose has been
1114 * achieved as well.
1115 */
1116 void kick_process(struct task_struct *p)
1117 {
1118 int cpu;
1120 preempt_disable();
1121 cpu = task_cpu(p);
1122 if ((cpu != smp_processor_id()) && task_curr(p))
1123 smp_send_reschedule(cpu);
1124 preempt_enable();
1125 }
1126 EXPORT_SYMBOL_GPL(kick_process);
1127 #endif /* CONFIG_SMP */
1129 #ifdef CONFIG_SMP
1130 /*
1131 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1132 */
1133 static int select_fallback_rq(int cpu, struct task_struct *p)
1134 {
1135 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1136 enum { cpuset, possible, fail } state = cpuset;
1137 int dest_cpu;
1139 /* Look for allowed, online CPU in same node. */
1140 for_each_cpu(dest_cpu, nodemask) {
1141 if (!cpu_online(dest_cpu))
1142 continue;
1143 if (!cpu_active(dest_cpu))
1144 continue;
1145 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1146 return dest_cpu;
1147 }
1149 for (;;) {
1150 /* Any allowed, online CPU? */
1151 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1152 if (!cpu_online(dest_cpu))
1153 continue;
1154 if (!cpu_active(dest_cpu))
1155 continue;
1156 goto out;
1157 }
1159 switch (state) {
1160 case cpuset:
1161 /* No more Mr. Nice Guy. */
1162 cpuset_cpus_allowed_fallback(p);
1163 state = possible;
1164 break;
1166 case possible:
1167 do_set_cpus_allowed(p, cpu_possible_mask);
1168 state = fail;
1169 break;
1171 case fail:
1172 BUG();
1173 break;
1174 }
1175 }
1177 out:
1178 if (state != cpuset) {
1179 /*
1180 * Don't tell them about moving exiting tasks or
1181 * kernel threads (both mm NULL), since they never
1182 * leave kernel.
1183 */
1184 if (p->mm && printk_ratelimit()) {
1185 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1186 task_pid_nr(p), p->comm, cpu);
1187 }
1188 }
1190 return dest_cpu;
1191 }
1193 /*
1194 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1195 */
1196 static inline
1197 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1198 {
1199 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1201 /*
1202 * In order not to call set_task_cpu() on a blocking task we need
1203 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1204 * cpu.
1205 *
1206 * Since this is common to all placement strategies, this lives here.
1207 *
1208 * [ this allows ->select_task() to simply return task_cpu(p) and
1209 * not worry about this generic constraint ]
1210 */
1211 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1212 !cpu_online(cpu)))
1213 cpu = select_fallback_rq(task_cpu(p), p);
1215 return cpu;
1216 }
1218 static void update_avg(u64 *avg, u64 sample)
1219 {
1220 s64 diff = sample - *avg;
1221 *avg += diff >> 3;
1222 }
1223 #endif
1225 static void
1226 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1227 {
1228 #ifdef CONFIG_SCHEDSTATS
1229 struct rq *rq = this_rq();
1231 #ifdef CONFIG_SMP
1232 int this_cpu = smp_processor_id();
1234 if (cpu == this_cpu) {
1235 schedstat_inc(rq, ttwu_local);
1236 schedstat_inc(p, se.statistics.nr_wakeups_local);
1237 } else {
1238 struct sched_domain *sd;
1240 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1241 rcu_read_lock();
1242 for_each_domain(this_cpu, sd) {
1243 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1244 schedstat_inc(sd, ttwu_wake_remote);
1245 break;
1246 }
1247 }
1248 rcu_read_unlock();
1249 }
1251 if (wake_flags & WF_MIGRATED)
1252 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1254 #endif /* CONFIG_SMP */
1256 schedstat_inc(rq, ttwu_count);
1257 schedstat_inc(p, se.statistics.nr_wakeups);
1259 if (wake_flags & WF_SYNC)
1260 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1262 #endif /* CONFIG_SCHEDSTATS */
1263 }
1265 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1266 {
1267 activate_task(rq, p, en_flags);
1268 p->on_rq = 1;
1270 /* if a worker is waking up, notify workqueue */
1271 if (p->flags & PF_WQ_WORKER)
1272 wq_worker_waking_up(p, cpu_of(rq));
1273 }
1275 /*
1276 * Mark the task runnable and perform wakeup-preemption.
1277 */
1278 static void
1279 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1280 {
1281 trace_sched_wakeup(p, true);
1282 check_preempt_curr(rq, p, wake_flags);
1284 p->state = TASK_RUNNING;
1285 #ifdef CONFIG_SMP
1286 if (p->sched_class->task_woken)
1287 p->sched_class->task_woken(rq, p);
1289 if (rq->idle_stamp) {
1290 u64 delta = rq->clock - rq->idle_stamp;
1291 u64 max = 2*sysctl_sched_migration_cost;
1293 if (delta > max)
1294 rq->avg_idle = max;
1295 else
1296 update_avg(&rq->avg_idle, delta);
1297 rq->idle_stamp = 0;
1298 }
1299 #endif
1300 }
1302 static void
1303 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1304 {
1305 #ifdef CONFIG_SMP
1306 if (p->sched_contributes_to_load)
1307 rq->nr_uninterruptible--;
1308 #endif
1310 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1311 ttwu_do_wakeup(rq, p, wake_flags);
1312 }
1314 /*
1315 * Called in case the task @p isn't fully descheduled from its runqueue,
1316 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1317 * since all we need to do is flip p->state to TASK_RUNNING, since
1318 * the task is still ->on_rq.
1319 */
1320 static int ttwu_remote(struct task_struct *p, int wake_flags)
1321 {
1322 struct rq *rq;
1323 int ret = 0;
1325 rq = __task_rq_lock(p);
1326 if (p->on_rq) {
1327 ttwu_do_wakeup(rq, p, wake_flags);
1328 ret = 1;
1329 }
1330 __task_rq_unlock(rq);
1332 return ret;
1333 }
1335 #ifdef CONFIG_SMP
1336 static void sched_ttwu_pending(void)
1337 {
1338 struct rq *rq = this_rq();
1339 struct llist_node *llist = llist_del_all(&rq->wake_list);
1340 struct task_struct *p;
1342 raw_spin_lock(&rq->lock);
1344 while (llist) {
1345 p = llist_entry(llist, struct task_struct, wake_entry);
1346 llist = llist_next(llist);
1347 ttwu_do_activate(rq, p, 0);
1348 }
1350 raw_spin_unlock(&rq->lock);
1351 }
1353 void scheduler_ipi(void)
1354 {
1355 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1356 return;
1358 /*
1359 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1360 * traditionally all their work was done from the interrupt return
1361 * path. Now that we actually do some work, we need to make sure
1362 * we do call them.
1363 *
1364 * Some archs already do call them, luckily irq_enter/exit nest
1365 * properly.
1366 *
1367 * Arguably we should visit all archs and update all handlers,
1368 * however a fair share of IPIs are still resched only so this would
1369 * somewhat pessimize the simple resched case.
1370 */
1371 irq_enter();
1372 sched_ttwu_pending();
1374 /*
1375 * Check if someone kicked us for doing the nohz idle load balance.
1376 */
1377 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1378 this_rq()->idle_balance = 1;
1379 raise_softirq_irqoff(SCHED_SOFTIRQ);
1380 }
1381 irq_exit();
1382 }
1384 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1385 {
1386 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1387 smp_send_reschedule(cpu);
1388 }
1390 bool cpus_share_cache(int this_cpu, int that_cpu)
1391 {
1392 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1393 }
1394 #endif /* CONFIG_SMP */
1396 static void ttwu_queue(struct task_struct *p, int cpu)
1397 {
1398 struct rq *rq = cpu_rq(cpu);
1400 #if defined(CONFIG_SMP)
1401 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1402 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1403 ttwu_queue_remote(p, cpu);
1404 return;
1405 }
1406 #endif
1408 raw_spin_lock(&rq->lock);
1409 ttwu_do_activate(rq, p, 0);
1410 raw_spin_unlock(&rq->lock);
1411 }
1413 /**
1414 * try_to_wake_up - wake up a thread
1415 * @p: the thread to be awakened
1416 * @state: the mask of task states that can be woken
1417 * @wake_flags: wake modifier flags (WF_*)
1418 *
1419 * Put it on the run-queue if it's not already there. The "current"
1420 * thread is always on the run-queue (except when the actual
1421 * re-schedule is in progress), and as such you're allowed to do
1422 * the simpler "current->state = TASK_RUNNING" to mark yourself
1423 * runnable without the overhead of this.
1424 *
1425 * Returns %true if @p was woken up, %false if it was already running
1426 * or @state didn't match @p's state.
1427 */
1428 static int
1429 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1430 {
1431 unsigned long flags;
1432 int cpu, success = 0;
1434 smp_wmb();
1435 raw_spin_lock_irqsave(&p->pi_lock, flags);
1436 if (!(p->state & state))
1437 goto out;
1439 success = 1; /* we're going to change ->state */
1440 cpu = task_cpu(p);
1442 if (p->on_rq && ttwu_remote(p, wake_flags))
1443 goto stat;
1445 #ifdef CONFIG_SMP
1446 /*
1447 * If the owning (remote) cpu is still in the middle of schedule() with
1448 * this task as prev, wait until its done referencing the task.
1449 */
1450 while (p->on_cpu)
1451 cpu_relax();
1452 /*
1453 * Pairs with the smp_wmb() in finish_lock_switch().
1454 */
1455 smp_rmb();
1457 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1458 p->state = TASK_WAKING;
1460 if (p->sched_class->task_waking)
1461 p->sched_class->task_waking(p);
1463 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1464 if (task_cpu(p) != cpu) {
1465 wake_flags |= WF_MIGRATED;
1466 set_task_cpu(p, cpu);
1467 }
1468 #endif /* CONFIG_SMP */
1470 ttwu_queue(p, cpu);
1471 stat:
1472 ttwu_stat(p, cpu, wake_flags);
1473 out:
1474 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1476 return success;
1477 }
1479 /**
1480 * try_to_wake_up_local - try to wake up a local task with rq lock held
1481 * @p: the thread to be awakened
1482 *
1483 * Put @p on the run-queue if it's not already there. The caller must
1484 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1485 * the current task.
1486 */
1487 static void try_to_wake_up_local(struct task_struct *p)
1488 {
1489 struct rq *rq = task_rq(p);
1491 if (WARN_ON_ONCE(rq != this_rq()) ||
1492 WARN_ON_ONCE(p == current))
1493 return;
1495 lockdep_assert_held(&rq->lock);
1497 if (!raw_spin_trylock(&p->pi_lock)) {
1498 raw_spin_unlock(&rq->lock);
1499 raw_spin_lock(&p->pi_lock);
1500 raw_spin_lock(&rq->lock);
1501 }
1503 if (!(p->state & TASK_NORMAL))
1504 goto out;
1506 if (!p->on_rq)
1507 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1509 ttwu_do_wakeup(rq, p, 0);
1510 ttwu_stat(p, smp_processor_id(), 0);
1511 out:
1512 raw_spin_unlock(&p->pi_lock);
1513 }
1515 /**
1516 * wake_up_process - Wake up a specific process
1517 * @p: The process to be woken up.
1518 *
1519 * Attempt to wake up the nominated process and move it to the set of runnable
1520 * processes. Returns 1 if the process was woken up, 0 if it was already
1521 * running.
1522 *
1523 * It may be assumed that this function implies a write memory barrier before
1524 * changing the task state if and only if any tasks are woken up.
1525 */
1526 int wake_up_process(struct task_struct *p)
1527 {
1528 WARN_ON(task_is_stopped_or_traced(p));
1529 return try_to_wake_up(p, TASK_NORMAL, 0);
1530 }
1531 EXPORT_SYMBOL(wake_up_process);
1533 int wake_up_state(struct task_struct *p, unsigned int state)
1534 {
1535 return try_to_wake_up(p, state, 0);
1536 }
1538 /*
1539 * Perform scheduler related setup for a newly forked process p.
1540 * p is forked by current.
1541 *
1542 * __sched_fork() is basic setup used by init_idle() too:
1543 */
1544 static void __sched_fork(struct task_struct *p)
1545 {
1546 p->on_rq = 0;
1548 p->se.on_rq = 0;
1549 p->se.exec_start = 0;
1550 p->se.sum_exec_runtime = 0;
1551 p->se.prev_sum_exec_runtime = 0;
1552 p->se.nr_migrations = 0;
1553 p->se.vruntime = 0;
1554 INIT_LIST_HEAD(&p->se.group_node);
1556 /*
1557 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1558 * removed when useful for applications beyond shares distribution (e.g.
1559 * load-balance).
1560 */
1561 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1562 p->se.avg.runnable_avg_period = 0;
1563 p->se.avg.runnable_avg_sum = 0;
1564 #endif
1565 #ifdef CONFIG_SCHEDSTATS
1566 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1567 #endif
1569 INIT_LIST_HEAD(&p->rt.run_list);
1571 #ifdef CONFIG_PREEMPT_NOTIFIERS
1572 INIT_HLIST_HEAD(&p->preempt_notifiers);
1573 #endif
1575 #ifdef CONFIG_NUMA_BALANCING
1576 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1577 p->mm->numa_next_scan = jiffies;
1578 p->mm->numa_next_reset = jiffies;
1579 p->mm->numa_scan_seq = 0;
1580 }
1582 p->node_stamp = 0ULL;
1583 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1584 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1585 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1586 p->numa_work.next = &p->numa_work;
1587 #endif /* CONFIG_NUMA_BALANCING */
1588 }
1590 #ifdef CONFIG_NUMA_BALANCING
1591 #ifdef CONFIG_SCHED_DEBUG
1592 void set_numabalancing_state(bool enabled)
1593 {
1594 if (enabled)
1595 sched_feat_set("NUMA");
1596 else
1597 sched_feat_set("NO_NUMA");
1598 }
1599 #else
1600 __read_mostly bool numabalancing_enabled;
1602 void set_numabalancing_state(bool enabled)
1603 {
1604 numabalancing_enabled = enabled;
1605 }
1606 #endif /* CONFIG_SCHED_DEBUG */
1607 #endif /* CONFIG_NUMA_BALANCING */
1609 /*
1610 * fork()/clone()-time setup:
1611 */
1612 void sched_fork(struct task_struct *p)
1613 {
1614 unsigned long flags;
1615 int cpu = get_cpu();
1617 __sched_fork(p);
1618 /*
1619 * We mark the process as running here. This guarantees that
1620 * nobody will actually run it, and a signal or other external
1621 * event cannot wake it up and insert it on the runqueue either.
1622 */
1623 p->state = TASK_RUNNING;
1625 /*
1626 * Make sure we do not leak PI boosting priority to the child.
1627 */
1628 p->prio = current->normal_prio;
1630 /*
1631 * Revert to default priority/policy on fork if requested.
1632 */
1633 if (unlikely(p->sched_reset_on_fork)) {
1634 if (task_has_rt_policy(p)) {
1635 p->policy = SCHED_NORMAL;
1636 p->static_prio = NICE_TO_PRIO(0);
1637 p->rt_priority = 0;
1638 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1639 p->static_prio = NICE_TO_PRIO(0);
1641 p->prio = p->normal_prio = __normal_prio(p);
1642 set_load_weight(p);
1644 /*
1645 * We don't need the reset flag anymore after the fork. It has
1646 * fulfilled its duty:
1647 */
1648 p->sched_reset_on_fork = 0;
1649 }
1651 if (!rt_prio(p->prio))
1652 p->sched_class = &fair_sched_class;
1654 if (p->sched_class->task_fork)
1655 p->sched_class->task_fork(p);
1657 /*
1658 * The child is not yet in the pid-hash so no cgroup attach races,
1659 * and the cgroup is pinned to this child due to cgroup_fork()
1660 * is ran before sched_fork().
1661 *
1662 * Silence PROVE_RCU.
1663 */
1664 raw_spin_lock_irqsave(&p->pi_lock, flags);
1665 set_task_cpu(p, cpu);
1666 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1668 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1669 if (likely(sched_info_on()))
1670 memset(&p->sched_info, 0, sizeof(p->sched_info));
1671 #endif
1672 #if defined(CONFIG_SMP)
1673 p->on_cpu = 0;
1674 #endif
1675 #ifdef CONFIG_PREEMPT_COUNT
1676 /* Want to start with kernel preemption disabled. */
1677 task_thread_info(p)->preempt_count = 1;
1678 #endif
1679 #ifdef CONFIG_SMP
1680 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1681 #endif
1683 put_cpu();
1684 }
1686 /*
1687 * wake_up_new_task - wake up a newly created task for the first time.
1688 *
1689 * This function will do some initial scheduler statistics housekeeping
1690 * that must be done for every newly created context, then puts the task
1691 * on the runqueue and wakes it.
1692 */
1693 void wake_up_new_task(struct task_struct *p)
1694 {
1695 unsigned long flags;
1696 struct rq *rq;
1698 raw_spin_lock_irqsave(&p->pi_lock, flags);
1699 #ifdef CONFIG_SMP
1700 /*
1701 * Fork balancing, do it here and not earlier because:
1702 * - cpus_allowed can change in the fork path
1703 * - any previously selected cpu might disappear through hotplug
1704 */
1705 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1706 #endif
1708 rq = __task_rq_lock(p);
1709 activate_task(rq, p, 0);
1710 p->on_rq = 1;
1711 trace_sched_wakeup_new(p, true);
1712 check_preempt_curr(rq, p, WF_FORK);
1713 #ifdef CONFIG_SMP
1714 if (p->sched_class->task_woken)
1715 p->sched_class->task_woken(rq, p);
1716 #endif
1717 task_rq_unlock(rq, p, &flags);
1718 }
1720 #ifdef CONFIG_PREEMPT_NOTIFIERS
1722 /**
1723 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1724 * @notifier: notifier struct to register
1725 */
1726 void preempt_notifier_register(struct preempt_notifier *notifier)
1727 {
1728 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1729 }
1730 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1732 /**
1733 * preempt_notifier_unregister - no longer interested in preemption notifications
1734 * @notifier: notifier struct to unregister
1735 *
1736 * This is safe to call from within a preemption notifier.
1737 */
1738 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1739 {
1740 hlist_del(¬ifier->link);
1741 }
1742 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1744 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1745 {
1746 struct preempt_notifier *notifier;
1747 struct hlist_node *node;
1749 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1750 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1751 }
1753 static void
1754 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1755 struct task_struct *next)
1756 {
1757 struct preempt_notifier *notifier;
1758 struct hlist_node *node;
1760 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1761 notifier->ops->sched_out(notifier, next);
1762 }
1764 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1766 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1767 {
1768 }
1770 static void
1771 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1772 struct task_struct *next)
1773 {
1774 }
1776 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1778 /**
1779 * prepare_task_switch - prepare to switch tasks
1780 * @rq: the runqueue preparing to switch
1781 * @prev: the current task that is being switched out
1782 * @next: the task we are going to switch to.
1783 *
1784 * This is called with the rq lock held and interrupts off. It must
1785 * be paired with a subsequent finish_task_switch after the context
1786 * switch.
1787 *
1788 * prepare_task_switch sets up locking and calls architecture specific
1789 * hooks.
1790 */
1791 static inline void
1792 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1793 struct task_struct *next)
1794 {
1795 trace_sched_switch(prev, next);
1796 sched_info_switch(prev, next);
1797 perf_event_task_sched_out(prev, next);
1798 fire_sched_out_preempt_notifiers(prev, next);
1799 prepare_lock_switch(rq, next);
1800 prepare_arch_switch(next);
1801 }
1803 /**
1804 * finish_task_switch - clean up after a task-switch
1805 * @rq: runqueue associated with task-switch
1806 * @prev: the thread we just switched away from.
1807 *
1808 * finish_task_switch must be called after the context switch, paired
1809 * with a prepare_task_switch call before the context switch.
1810 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1811 * and do any other architecture-specific cleanup actions.
1812 *
1813 * Note that we may have delayed dropping an mm in context_switch(). If
1814 * so, we finish that here outside of the runqueue lock. (Doing it
1815 * with the lock held can cause deadlocks; see schedule() for
1816 * details.)
1817 */
1818 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1819 __releases(rq->lock)
1820 {
1821 struct mm_struct *mm = rq->prev_mm;
1822 long prev_state;
1824 rq->prev_mm = NULL;
1826 /*
1827 * A task struct has one reference for the use as "current".
1828 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1829 * schedule one last time. The schedule call will never return, and
1830 * the scheduled task must drop that reference.
1831 * The test for TASK_DEAD must occur while the runqueue locks are
1832 * still held, otherwise prev could be scheduled on another cpu, die
1833 * there before we look at prev->state, and then the reference would
1834 * be dropped twice.
1835 * Manfred Spraul <manfred@colorfullife.com>
1836 */
1837 prev_state = prev->state;
1838 vtime_task_switch(prev);
1839 finish_arch_switch(prev);
1840 perf_event_task_sched_in(prev, current);
1841 finish_lock_switch(rq, prev);
1842 finish_arch_post_lock_switch();
1844 fire_sched_in_preempt_notifiers(current);
1845 if (mm)
1846 mmdrop(mm);
1847 if (unlikely(prev_state == TASK_DEAD)) {
1848 /*
1849 * Remove function-return probe instances associated with this
1850 * task and put them back on the free list.
1851 */
1852 kprobe_flush_task(prev);
1853 put_task_struct(prev);
1854 }
1855 }
1857 #ifdef CONFIG_SMP
1859 /* assumes rq->lock is held */
1860 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1861 {
1862 if (prev->sched_class->pre_schedule)
1863 prev->sched_class->pre_schedule(rq, prev);
1864 }
1866 /* rq->lock is NOT held, but preemption is disabled */
1867 static inline void post_schedule(struct rq *rq)
1868 {
1869 if (rq->post_schedule) {
1870 unsigned long flags;
1872 raw_spin_lock_irqsave(&rq->lock, flags);
1873 if (rq->curr->sched_class->post_schedule)
1874 rq->curr->sched_class->post_schedule(rq);
1875 raw_spin_unlock_irqrestore(&rq->lock, flags);
1877 rq->post_schedule = 0;
1878 }
1879 }
1881 #else
1883 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1884 {
1885 }
1887 static inline void post_schedule(struct rq *rq)
1888 {
1889 }
1891 #endif
1893 /**
1894 * schedule_tail - first thing a freshly forked thread must call.
1895 * @prev: the thread we just switched away from.
1896 */
1897 asmlinkage void schedule_tail(struct task_struct *prev)
1898 __releases(rq->lock)
1899 {
1900 struct rq *rq = this_rq();
1902 finish_task_switch(rq, prev);
1904 /*
1905 * FIXME: do we need to worry about rq being invalidated by the
1906 * task_switch?
1907 */
1908 post_schedule(rq);
1910 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1911 /* In this case, finish_task_switch does not reenable preemption */
1912 preempt_enable();
1913 #endif
1914 if (current->set_child_tid)
1915 put_user(task_pid_vnr(current), current->set_child_tid);
1916 }
1918 /*
1919 * context_switch - switch to the new MM and the new
1920 * thread's register state.
1921 */
1922 static inline void
1923 context_switch(struct rq *rq, struct task_struct *prev,
1924 struct task_struct *next)
1925 {
1926 struct mm_struct *mm, *oldmm;
1928 prepare_task_switch(rq, prev, next);
1930 mm = next->mm;
1931 oldmm = prev->active_mm;
1932 /*
1933 * For paravirt, this is coupled with an exit in switch_to to
1934 * combine the page table reload and the switch backend into
1935 * one hypercall.
1936 */
1937 arch_start_context_switch(prev);
1939 if (!mm) {
1940 next->active_mm = oldmm;
1941 atomic_inc(&oldmm->mm_count);
1942 enter_lazy_tlb(oldmm, next);
1943 } else
1944 switch_mm(oldmm, mm, next);
1946 if (!prev->mm) {
1947 prev->active_mm = NULL;
1948 rq->prev_mm = oldmm;
1949 }
1950 /*
1951 * Since the runqueue lock will be released by the next
1952 * task (which is an invalid locking op but in the case
1953 * of the scheduler it's an obvious special-case), so we
1954 * do an early lockdep release here:
1955 */
1956 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1957 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1958 #endif
1960 context_tracking_task_switch(prev, next);
1961 /* Here we just switch the register state and the stack. */
1962 switch_to(prev, next, prev);
1964 barrier();
1965 /*
1966 * this_rq must be evaluated again because prev may have moved
1967 * CPUs since it called schedule(), thus the 'rq' on its stack
1968 * frame will be invalid.
1969 */
1970 finish_task_switch(this_rq(), prev);
1971 }
1973 /*
1974 * nr_running, nr_uninterruptible and nr_context_switches:
1975 *
1976 * externally visible scheduler statistics: current number of runnable
1977 * threads, current number of uninterruptible-sleeping threads, total
1978 * number of context switches performed since bootup.
1979 */
1980 unsigned long nr_running(void)
1981 {
1982 unsigned long i, sum = 0;
1984 for_each_online_cpu(i)
1985 sum += cpu_rq(i)->nr_running;
1987 return sum;
1988 }
1990 unsigned long nr_uninterruptible(void)
1991 {
1992 unsigned long i, sum = 0;
1994 for_each_possible_cpu(i)
1995 sum += cpu_rq(i)->nr_uninterruptible;
1997 /*
1998 * Since we read the counters lockless, it might be slightly
1999 * inaccurate. Do not allow it to go below zero though:
2000 */
2001 if (unlikely((long)sum < 0))
2002 sum = 0;
2004 return sum;
2005 }
2007 unsigned long long nr_context_switches(void)
2008 {
2009 int i;
2010 unsigned long long sum = 0;
2012 for_each_possible_cpu(i)
2013 sum += cpu_rq(i)->nr_switches;
2015 return sum;
2016 }
2018 unsigned long nr_iowait(void)
2019 {
2020 unsigned long i, sum = 0;
2022 for_each_possible_cpu(i)
2023 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2025 return sum;
2026 }
2028 unsigned long nr_iowait_cpu(int cpu)
2029 {
2030 struct rq *this = cpu_rq(cpu);
2031 return atomic_read(&this->nr_iowait);
2032 }
2034 unsigned long this_cpu_load(void)
2035 {
2036 struct rq *this = this_rq();
2037 return this->cpu_load[0];
2038 }
2041 /*
2042 * Global load-average calculations
2043 *
2044 * We take a distributed and async approach to calculating the global load-avg
2045 * in order to minimize overhead.
2046 *
2047 * The global load average is an exponentially decaying average of nr_running +
2048 * nr_uninterruptible.
2049 *
2050 * Once every LOAD_FREQ:
2051 *
2052 * nr_active = 0;
2053 * for_each_possible_cpu(cpu)
2054 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2055 *
2056 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2057 *
2058 * Due to a number of reasons the above turns in the mess below:
2059 *
2060 * - for_each_possible_cpu() is prohibitively expensive on machines with
2061 * serious number of cpus, therefore we need to take a distributed approach
2062 * to calculating nr_active.
2063 *
2064 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2065 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2066 *
2067 * So assuming nr_active := 0 when we start out -- true per definition, we
2068 * can simply take per-cpu deltas and fold those into a global accumulate
2069 * to obtain the same result. See calc_load_fold_active().
2070 *
2071 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2072 * across the machine, we assume 10 ticks is sufficient time for every
2073 * cpu to have completed this task.
2074 *
2075 * This places an upper-bound on the IRQ-off latency of the machine. Then
2076 * again, being late doesn't loose the delta, just wrecks the sample.
2077 *
2078 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2079 * this would add another cross-cpu cacheline miss and atomic operation
2080 * to the wakeup path. Instead we increment on whatever cpu the task ran
2081 * when it went into uninterruptible state and decrement on whatever cpu
2082 * did the wakeup. This means that only the sum of nr_uninterruptible over
2083 * all cpus yields the correct result.
2084 *
2085 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2086 */
2088 /* Variables and functions for calc_load */
2089 static atomic_long_t calc_load_tasks;
2090 static unsigned long calc_load_update;
2091 unsigned long avenrun[3];
2092 EXPORT_SYMBOL(avenrun); /* should be removed */
2094 /**
2095 * get_avenrun - get the load average array
2096 * @loads: pointer to dest load array
2097 * @offset: offset to add
2098 * @shift: shift count to shift the result left
2099 *
2100 * These values are estimates at best, so no need for locking.
2101 */
2102 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2103 {
2104 loads[0] = (avenrun[0] + offset) << shift;
2105 loads[1] = (avenrun[1] + offset) << shift;
2106 loads[2] = (avenrun[2] + offset) << shift;
2107 }
2109 static long calc_load_fold_active(struct rq *this_rq)
2110 {
2111 long nr_active, delta = 0;
2113 nr_active = this_rq->nr_running;
2114 nr_active += (long) this_rq->nr_uninterruptible;
2116 if (nr_active != this_rq->calc_load_active) {
2117 delta = nr_active - this_rq->calc_load_active;
2118 this_rq->calc_load_active = nr_active;
2119 }
2121 return delta;
2122 }
2124 /*
2125 * a1 = a0 * e + a * (1 - e)
2126 */
2127 static unsigned long
2128 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2129 {
2130 load *= exp;
2131 load += active * (FIXED_1 - exp);
2132 load += 1UL << (FSHIFT - 1);
2133 return load >> FSHIFT;
2134 }
2136 #ifdef CONFIG_NO_HZ
2137 /*
2138 * Handle NO_HZ for the global load-average.
2139 *
2140 * Since the above described distributed algorithm to compute the global
2141 * load-average relies on per-cpu sampling from the tick, it is affected by
2142 * NO_HZ.
2143 *
2144 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2145 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2146 * when we read the global state.
2147 *
2148 * Obviously reality has to ruin such a delightfully simple scheme:
2149 *
2150 * - When we go NO_HZ idle during the window, we can negate our sample
2151 * contribution, causing under-accounting.
2152 *
2153 * We avoid this by keeping two idle-delta counters and flipping them
2154 * when the window starts, thus separating old and new NO_HZ load.
2155 *
2156 * The only trick is the slight shift in index flip for read vs write.
2157 *
2158 * 0s 5s 10s 15s
2159 * +10 +10 +10 +10
2160 * |-|-----------|-|-----------|-|-----------|-|
2161 * r:0 0 1 1 0 0 1 1 0
2162 * w:0 1 1 0 0 1 1 0 0
2163 *
2164 * This ensures we'll fold the old idle contribution in this window while
2165 * accumlating the new one.
2166 *
2167 * - When we wake up from NO_HZ idle during the window, we push up our
2168 * contribution, since we effectively move our sample point to a known
2169 * busy state.
2170 *
2171 * This is solved by pushing the window forward, and thus skipping the
2172 * sample, for this cpu (effectively using the idle-delta for this cpu which
2173 * was in effect at the time the window opened). This also solves the issue
2174 * of having to deal with a cpu having been in NOHZ idle for multiple
2175 * LOAD_FREQ intervals.
2176 *
2177 * When making the ILB scale, we should try to pull this in as well.
2178 */
2179 static atomic_long_t calc_load_idle[2];
2180 static int calc_load_idx;
2182 static inline int calc_load_write_idx(void)
2183 {
2184 int idx = calc_load_idx;
2186 /*
2187 * See calc_global_nohz(), if we observe the new index, we also
2188 * need to observe the new update time.
2189 */
2190 smp_rmb();
2192 /*
2193 * If the folding window started, make sure we start writing in the
2194 * next idle-delta.
2195 */
2196 if (!time_before(jiffies, calc_load_update))
2197 idx++;
2199 return idx & 1;
2200 }
2202 static inline int calc_load_read_idx(void)
2203 {
2204 return calc_load_idx & 1;
2205 }
2207 void calc_load_enter_idle(void)
2208 {
2209 struct rq *this_rq = this_rq();
2210 long delta;
2212 /*
2213 * We're going into NOHZ mode, if there's any pending delta, fold it
2214 * into the pending idle delta.
2215 */
2216 delta = calc_load_fold_active(this_rq);
2217 if (delta) {
2218 int idx = calc_load_write_idx();
2219 atomic_long_add(delta, &calc_load_idle[idx]);
2220 }
2221 }
2223 void calc_load_exit_idle(void)
2224 {
2225 struct rq *this_rq = this_rq();
2227 /*
2228 * If we're still before the sample window, we're done.
2229 */
2230 if (time_before(jiffies, this_rq->calc_load_update))
2231 return;
2233 /*
2234 * We woke inside or after the sample window, this means we're already
2235 * accounted through the nohz accounting, so skip the entire deal and
2236 * sync up for the next window.
2237 */
2238 this_rq->calc_load_update = calc_load_update;
2239 if (time_before(jiffies, this_rq->calc_load_update + 10))
2240 this_rq->calc_load_update += LOAD_FREQ;
2241 }
2243 static long calc_load_fold_idle(void)
2244 {
2245 int idx = calc_load_read_idx();
2246 long delta = 0;
2248 if (atomic_long_read(&calc_load_idle[idx]))
2249 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2251 return delta;
2252 }
2254 /**
2255 * fixed_power_int - compute: x^n, in O(log n) time
2256 *
2257 * @x: base of the power
2258 * @frac_bits: fractional bits of @x
2259 * @n: power to raise @x to.
2260 *
2261 * By exploiting the relation between the definition of the natural power
2262 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2263 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2264 * (where: n_i \elem {0, 1}, the binary vector representing n),
2265 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2266 * of course trivially computable in O(log_2 n), the length of our binary
2267 * vector.
2268 */
2269 static unsigned long
2270 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2271 {
2272 unsigned long result = 1UL << frac_bits;
2274 if (n) for (;;) {
2275 if (n & 1) {
2276 result *= x;
2277 result += 1UL << (frac_bits - 1);
2278 result >>= frac_bits;
2279 }
2280 n >>= 1;
2281 if (!n)
2282 break;
2283 x *= x;
2284 x += 1UL << (frac_bits - 1);
2285 x >>= frac_bits;
2286 }
2288 return result;
2289 }
2291 /*
2292 * a1 = a0 * e + a * (1 - e)
2293 *
2294 * a2 = a1 * e + a * (1 - e)
2295 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2296 * = a0 * e^2 + a * (1 - e) * (1 + e)
2297 *
2298 * a3 = a2 * e + a * (1 - e)
2299 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2300 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2301 *
2302 * ...
2303 *
2304 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2305 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2306 * = a0 * e^n + a * (1 - e^n)
2307 *
2308 * [1] application of the geometric series:
2309 *
2310 * n 1 - x^(n+1)
2311 * S_n := \Sum x^i = -------------
2312 * i=0 1 - x
2313 */
2314 static unsigned long
2315 calc_load_n(unsigned long load, unsigned long exp,
2316 unsigned long active, unsigned int n)
2317 {
2319 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2320 }
2322 /*
2323 * NO_HZ can leave us missing all per-cpu ticks calling
2324 * calc_load_account_active(), but since an idle CPU folds its delta into
2325 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2326 * in the pending idle delta if our idle period crossed a load cycle boundary.
2327 *
2328 * Once we've updated the global active value, we need to apply the exponential
2329 * weights adjusted to the number of cycles missed.
2330 */
2331 static void calc_global_nohz(void)
2332 {
2333 long delta, active, n;
2335 if (!time_before(jiffies, calc_load_update + 10)) {
2336 /*
2337 * Catch-up, fold however many we are behind still
2338 */
2339 delta = jiffies - calc_load_update - 10;
2340 n = 1 + (delta / LOAD_FREQ);
2342 active = atomic_long_read(&calc_load_tasks);
2343 active = active > 0 ? active * FIXED_1 : 0;
2345 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2346 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2347 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2349 calc_load_update += n * LOAD_FREQ;
2350 }
2352 /*
2353 * Flip the idle index...
2354 *
2355 * Make sure we first write the new time then flip the index, so that
2356 * calc_load_write_idx() will see the new time when it reads the new
2357 * index, this avoids a double flip messing things up.
2358 */
2359 smp_wmb();
2360 calc_load_idx++;
2361 }
2362 #else /* !CONFIG_NO_HZ */
2364 static inline long calc_load_fold_idle(void) { return 0; }
2365 static inline void calc_global_nohz(void) { }
2367 #endif /* CONFIG_NO_HZ */
2369 /*
2370 * calc_load - update the avenrun load estimates 10 ticks after the
2371 * CPUs have updated calc_load_tasks.
2372 */
2373 void calc_global_load(unsigned long ticks)
2374 {
2375 long active, delta;
2377 if (time_before(jiffies, calc_load_update + 10))
2378 return;
2380 /*
2381 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2382 */
2383 delta = calc_load_fold_idle();
2384 if (delta)
2385 atomic_long_add(delta, &calc_load_tasks);
2387 active = atomic_long_read(&calc_load_tasks);
2388 active = active > 0 ? active * FIXED_1 : 0;
2390 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2391 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2392 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2394 calc_load_update += LOAD_FREQ;
2396 /*
2397 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2398 */
2399 calc_global_nohz();
2400 }
2402 /*
2403 * Called from update_cpu_load() to periodically update this CPU's
2404 * active count.
2405 */
2406 static void calc_load_account_active(struct rq *this_rq)
2407 {
2408 long delta;
2410 if (time_before(jiffies, this_rq->calc_load_update))
2411 return;
2413 delta = calc_load_fold_active(this_rq);
2414 if (delta)
2415 atomic_long_add(delta, &calc_load_tasks);
2417 this_rq->calc_load_update += LOAD_FREQ;
2418 }
2420 /*
2421 * End of global load-average stuff
2422 */
2424 /*
2425 * The exact cpuload at various idx values, calculated at every tick would be
2426 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2427 *
2428 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2429 * on nth tick when cpu may be busy, then we have:
2430 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2431 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2432 *
2433 * decay_load_missed() below does efficient calculation of
2434 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2435 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2436 *
2437 * The calculation is approximated on a 128 point scale.
2438 * degrade_zero_ticks is the number of ticks after which load at any
2439 * particular idx is approximated to be zero.
2440 * degrade_factor is a precomputed table, a row for each load idx.
2441 * Each column corresponds to degradation factor for a power of two ticks,
2442 * based on 128 point scale.
2443 * Example:
2444 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2445 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2446 *
2447 * With this power of 2 load factors, we can degrade the load n times
2448 * by looking at 1 bits in n and doing as many mult/shift instead of
2449 * n mult/shifts needed by the exact degradation.
2450 */
2451 #define DEGRADE_SHIFT 7
2452 static const unsigned char
2453 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2454 static const unsigned char
2455 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2456 {0, 0, 0, 0, 0, 0, 0, 0},
2457 {64, 32, 8, 0, 0, 0, 0, 0},
2458 {96, 72, 40, 12, 1, 0, 0},
2459 {112, 98, 75, 43, 15, 1, 0},
2460 {120, 112, 98, 76, 45, 16, 2} };
2462 /*
2463 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2464 * would be when CPU is idle and so we just decay the old load without
2465 * adding any new load.
2466 */
2467 static unsigned long
2468 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2469 {
2470 int j = 0;
2472 if (!missed_updates)
2473 return load;
2475 if (missed_updates >= degrade_zero_ticks[idx])
2476 return 0;
2478 if (idx == 1)
2479 return load >> missed_updates;
2481 while (missed_updates) {
2482 if (missed_updates % 2)
2483 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2485 missed_updates >>= 1;
2486 j++;
2487 }
2488 return load;
2489 }
2491 /*
2492 * Update rq->cpu_load[] statistics. This function is usually called every
2493 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2494 * every tick. We fix it up based on jiffies.
2495 */
2496 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2497 unsigned long pending_updates)
2498 {
2499 int i, scale;
2501 this_rq->nr_load_updates++;
2503 /* Update our load: */
2504 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2505 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2506 unsigned long old_load, new_load;
2508 /* scale is effectively 1 << i now, and >> i divides by scale */
2510 old_load = this_rq->cpu_load[i];
2511 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2512 new_load = this_load;
2513 /*
2514 * Round up the averaging division if load is increasing. This
2515 * prevents us from getting stuck on 9 if the load is 10, for
2516 * example.
2517 */
2518 if (new_load > old_load)
2519 new_load += scale - 1;
2521 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2522 }
2524 sched_avg_update(this_rq);
2525 }
2527 #ifdef CONFIG_NO_HZ
2528 /*
2529 * There is no sane way to deal with nohz on smp when using jiffies because the
2530 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2531 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2532 *
2533 * Therefore we cannot use the delta approach from the regular tick since that
2534 * would seriously skew the load calculation. However we'll make do for those
2535 * updates happening while idle (nohz_idle_balance) or coming out of idle
2536 * (tick_nohz_idle_exit).
2537 *
2538 * This means we might still be one tick off for nohz periods.
2539 */
2541 /*
2542 * Called from nohz_idle_balance() to update the load ratings before doing the
2543 * idle balance.
2544 */
2545 void update_idle_cpu_load(struct rq *this_rq)
2546 {
2547 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2548 unsigned long load = this_rq->load.weight;
2549 unsigned long pending_updates;
2551 /*
2552 * bail if there's load or we're actually up-to-date.
2553 */
2554 if (load || curr_jiffies == this_rq->last_load_update_tick)
2555 return;
2557 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2558 this_rq->last_load_update_tick = curr_jiffies;
2560 __update_cpu_load(this_rq, load, pending_updates);
2561 }
2563 /*
2564 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2565 */
2566 void update_cpu_load_nohz(void)
2567 {
2568 struct rq *this_rq = this_rq();
2569 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2570 unsigned long pending_updates;
2572 if (curr_jiffies == this_rq->last_load_update_tick)
2573 return;
2575 raw_spin_lock(&this_rq->lock);
2576 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2577 if (pending_updates) {
2578 this_rq->last_load_update_tick = curr_jiffies;
2579 /*
2580 * We were idle, this means load 0, the current load might be
2581 * !0 due to remote wakeups and the sort.
2582 */
2583 __update_cpu_load(this_rq, 0, pending_updates);
2584 }
2585 raw_spin_unlock(&this_rq->lock);
2586 }
2587 #endif /* CONFIG_NO_HZ */
2589 /*
2590 * Called from scheduler_tick()
2591 */
2592 static void update_cpu_load_active(struct rq *this_rq)
2593 {
2594 /*
2595 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2596 */
2597 this_rq->last_load_update_tick = jiffies;
2598 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2600 calc_load_account_active(this_rq);
2601 }
2603 #ifdef CONFIG_SMP
2605 /*
2606 * sched_exec - execve() is a valuable balancing opportunity, because at
2607 * this point the task has the smallest effective memory and cache footprint.
2608 */
2609 void sched_exec(void)
2610 {
2611 struct task_struct *p = current;
2612 unsigned long flags;
2613 int dest_cpu;
2615 raw_spin_lock_irqsave(&p->pi_lock, flags);
2616 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2617 if (dest_cpu == smp_processor_id())
2618 goto unlock;
2620 if (likely(cpu_active(dest_cpu))) {
2621 struct migration_arg arg = { p, dest_cpu };
2623 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2624 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2625 return;
2626 }
2627 unlock:
2628 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2629 }
2631 #endif
2633 DEFINE_PER_CPU(struct kernel_stat, kstat);
2634 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2636 EXPORT_PER_CPU_SYMBOL(kstat);
2637 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2639 /*
2640 * Return any ns on the sched_clock that have not yet been accounted in
2641 * @p in case that task is currently running.
2642 *
2643 * Called with task_rq_lock() held on @rq.
2644 */
2645 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2646 {
2647 u64 ns = 0;
2649 if (task_current(rq, p)) {
2650 update_rq_clock(rq);
2651 ns = rq->clock_task - p->se.exec_start;
2652 if ((s64)ns < 0)
2653 ns = 0;
2654 }
2656 return ns;
2657 }
2659 unsigned long long task_delta_exec(struct task_struct *p)
2660 {
2661 unsigned long flags;
2662 struct rq *rq;
2663 u64 ns = 0;
2665 rq = task_rq_lock(p, &flags);
2666 ns = do_task_delta_exec(p, rq);
2667 task_rq_unlock(rq, p, &flags);
2669 return ns;
2670 }
2672 /*
2673 * Return accounted runtime for the task.
2674 * In case the task is currently running, return the runtime plus current's
2675 * pending runtime that have not been accounted yet.
2676 */
2677 unsigned long long task_sched_runtime(struct task_struct *p)
2678 {
2679 unsigned long flags;
2680 struct rq *rq;
2681 u64 ns = 0;
2683 rq = task_rq_lock(p, &flags);
2684 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2685 task_rq_unlock(rq, p, &flags);
2687 return ns;
2688 }
2690 /*
2691 * This function gets called by the timer code, with HZ frequency.
2692 * We call it with interrupts disabled.
2693 */
2694 void scheduler_tick(void)
2695 {
2696 int cpu = smp_processor_id();
2697 struct rq *rq = cpu_rq(cpu);
2698 struct task_struct *curr = rq->curr;
2700 sched_clock_tick();
2702 raw_spin_lock(&rq->lock);
2703 update_rq_clock(rq);
2704 update_cpu_load_active(rq);
2705 curr->sched_class->task_tick(rq, curr, 0);
2706 raw_spin_unlock(&rq->lock);
2708 perf_event_task_tick();
2710 #ifdef CONFIG_SMP
2711 rq->idle_balance = idle_cpu(cpu);
2712 trigger_load_balance(rq, cpu);
2713 #endif
2714 }
2716 notrace unsigned long get_parent_ip(unsigned long addr)
2717 {
2718 if (in_lock_functions(addr)) {
2719 addr = CALLER_ADDR2;
2720 if (in_lock_functions(addr))
2721 addr = CALLER_ADDR3;
2722 }
2723 return addr;
2724 }
2726 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2727 defined(CONFIG_PREEMPT_TRACER))
2729 void __kprobes add_preempt_count(int val)
2730 {
2731 #ifdef CONFIG_DEBUG_PREEMPT
2732 /*
2733 * Underflow?
2734 */
2735 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2736 return;
2737 #endif
2738 preempt_count() += val;
2739 #ifdef CONFIG_DEBUG_PREEMPT
2740 /*
2741 * Spinlock count overflowing soon?
2742 */
2743 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2744 PREEMPT_MASK - 10);
2745 #endif
2746 if (preempt_count() == val)
2747 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2748 }
2749 EXPORT_SYMBOL(add_preempt_count);
2751 void __kprobes sub_preempt_count(int val)
2752 {
2753 #ifdef CONFIG_DEBUG_PREEMPT
2754 /*
2755 * Underflow?
2756 */
2757 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2758 return;
2759 /*
2760 * Is the spinlock portion underflowing?
2761 */
2762 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2763 !(preempt_count() & PREEMPT_MASK)))
2764 return;
2765 #endif
2767 if (preempt_count() == val)
2768 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2769 preempt_count() -= val;
2770 }
2771 EXPORT_SYMBOL(sub_preempt_count);
2773 #endif
2775 /*
2776 * Print scheduling while atomic bug:
2777 */
2778 static noinline void __schedule_bug(struct task_struct *prev)
2779 {
2780 if (oops_in_progress)
2781 return;
2783 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2784 prev->comm, prev->pid, preempt_count());
2786 debug_show_held_locks(prev);
2787 print_modules();
2788 if (irqs_disabled())
2789 print_irqtrace_events(prev);
2790 dump_stack();
2791 add_taint(TAINT_WARN);
2792 }
2794 /*
2795 * Various schedule()-time debugging checks and statistics:
2796 */
2797 static inline void schedule_debug(struct task_struct *prev)
2798 {
2799 /*
2800 * Test if we are atomic. Since do_exit() needs to call into
2801 * schedule() atomically, we ignore that path for now.
2802 * Otherwise, whine if we are scheduling when we should not be.
2803 */
2804 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2805 __schedule_bug(prev);
2806 rcu_sleep_check();
2808 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2810 schedstat_inc(this_rq(), sched_count);
2811 }
2813 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2814 {
2815 if (prev->on_rq || rq->skip_clock_update < 0)
2816 update_rq_clock(rq);
2817 prev->sched_class->put_prev_task(rq, prev);
2818 }
2820 /*
2821 * Pick up the highest-prio task:
2822 */
2823 static inline struct task_struct *
2824 pick_next_task(struct rq *rq)
2825 {
2826 const struct sched_class *class;
2827 struct task_struct *p;
2829 /*
2830 * Optimization: we know that if all tasks are in
2831 * the fair class we can call that function directly:
2832 */
2833 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2834 p = fair_sched_class.pick_next_task(rq);
2835 if (likely(p))
2836 return p;
2837 }
2839 for_each_class(class) {
2840 p = class->pick_next_task(rq);
2841 if (p)
2842 return p;
2843 }
2845 BUG(); /* the idle class will always have a runnable task */
2846 }
2848 /*
2849 * __schedule() is the main scheduler function.
2850 *
2851 * The main means of driving the scheduler and thus entering this function are:
2852 *
2853 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2854 *
2855 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2856 * paths. For example, see arch/x86/entry_64.S.
2857 *
2858 * To drive preemption between tasks, the scheduler sets the flag in timer
2859 * interrupt handler scheduler_tick().
2860 *
2861 * 3. Wakeups don't really cause entry into schedule(). They add a
2862 * task to the run-queue and that's it.
2863 *
2864 * Now, if the new task added to the run-queue preempts the current
2865 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2866 * called on the nearest possible occasion:
2867 *
2868 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2869 *
2870 * - in syscall or exception context, at the next outmost
2871 * preempt_enable(). (this might be as soon as the wake_up()'s
2872 * spin_unlock()!)
2873 *
2874 * - in IRQ context, return from interrupt-handler to
2875 * preemptible context
2876 *
2877 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2878 * then at the next:
2879 *
2880 * - cond_resched() call
2881 * - explicit schedule() call
2882 * - return from syscall or exception to user-space
2883 * - return from interrupt-handler to user-space
2884 */
2885 static void __sched __schedule(void)
2886 {
2887 struct task_struct *prev, *next;
2888 unsigned long *switch_count;
2889 struct rq *rq;
2890 int cpu;
2892 need_resched:
2893 preempt_disable();
2894 cpu = smp_processor_id();
2895 rq = cpu_rq(cpu);
2896 rcu_note_context_switch(cpu);
2897 prev = rq->curr;
2899 schedule_debug(prev);
2901 if (sched_feat(HRTICK))
2902 hrtick_clear(rq);
2904 raw_spin_lock_irq(&rq->lock);
2906 switch_count = &prev->nivcsw;
2907 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2908 if (unlikely(signal_pending_state(prev->state, prev))) {
2909 prev->state = TASK_RUNNING;
2910 } else {
2911 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2912 prev->on_rq = 0;
2914 /*
2915 * If a worker went to sleep, notify and ask workqueue
2916 * whether it wants to wake up a task to maintain
2917 * concurrency.
2918 */
2919 if (prev->flags & PF_WQ_WORKER) {
2920 struct task_struct *to_wakeup;
2922 to_wakeup = wq_worker_sleeping(prev, cpu);
2923 if (to_wakeup)
2924 try_to_wake_up_local(to_wakeup);
2925 }
2926 }
2927 switch_count = &prev->nvcsw;
2928 }
2930 pre_schedule(rq, prev);
2932 if (unlikely(!rq->nr_running))
2933 idle_balance(cpu, rq);
2935 put_prev_task(rq, prev);
2936 next = pick_next_task(rq);
2937 clear_tsk_need_resched(prev);
2938 rq->skip_clock_update = 0;
2940 if (likely(prev != next)) {
2941 rq->nr_switches++;
2942 rq->curr = next;
2943 ++*switch_count;
2945 context_switch(rq, prev, next); /* unlocks the rq */
2946 /*
2947 * The context switch have flipped the stack from under us
2948 * and restored the local variables which were saved when
2949 * this task called schedule() in the past. prev == current
2950 * is still correct, but it can be moved to another cpu/rq.
2951 */
2952 cpu = smp_processor_id();
2953 rq = cpu_rq(cpu);
2954 } else
2955 raw_spin_unlock_irq(&rq->lock);
2957 post_schedule(rq);
2959 sched_preempt_enable_no_resched();
2960 if (need_resched())
2961 goto need_resched;
2962 }
2964 static inline void sched_submit_work(struct task_struct *tsk)
2965 {
2966 if (!tsk->state || tsk_is_pi_blocked(tsk))
2967 return;
2968 /*
2969 * If we are going to sleep and we have plugged IO queued,
2970 * make sure to submit it to avoid deadlocks.
2971 */
2972 if (blk_needs_flush_plug(tsk))
2973 blk_schedule_flush_plug(tsk);
2974 }
2976 asmlinkage void __sched schedule(void)
2977 {
2978 struct task_struct *tsk = current;
2980 sched_submit_work(tsk);
2981 __schedule();
2982 }
2983 EXPORT_SYMBOL(schedule);
2985 #ifdef CONFIG_CONTEXT_TRACKING
2986 asmlinkage void __sched schedule_user(void)
2987 {
2988 /*
2989 * If we come here after a random call to set_need_resched(),
2990 * or we have been woken up remotely but the IPI has not yet arrived,
2991 * we haven't yet exited the RCU idle mode. Do it here manually until
2992 * we find a better solution.
2993 */
2994 user_exit();
2995 schedule();
2996 user_enter();
2997 }
2998 #endif
3000 /**
3001 * schedule_preempt_disabled - called with preemption disabled
3002 *
3003 * Returns with preemption disabled. Note: preempt_count must be 1
3004 */
3005 void __sched schedule_preempt_disabled(void)
3006 {
3007 sched_preempt_enable_no_resched();
3008 schedule();
3009 preempt_disable();
3010 }
3012 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3014 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3015 {
3016 if (lock->owner != owner)
3017 return false;
3019 /*
3020 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3021 * lock->owner still matches owner, if that fails, owner might
3022 * point to free()d memory, if it still matches, the rcu_read_lock()
3023 * ensures the memory stays valid.
3024 */
3025 barrier();
3027 return owner->on_cpu;
3028 }
3030 /*
3031 * Look out! "owner" is an entirely speculative pointer
3032 * access and not reliable.
3033 */
3034 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3035 {
3036 if (!sched_feat(OWNER_SPIN))
3037 return 0;
3039 rcu_read_lock();
3040 while (owner_running(lock, owner)) {
3041 if (need_resched())
3042 break;
3044 arch_mutex_cpu_relax();
3045 }
3046 rcu_read_unlock();
3048 /*
3049 * We break out the loop above on need_resched() and when the
3050 * owner changed, which is a sign for heavy contention. Return
3051 * success only when lock->owner is NULL.
3052 */
3053 return lock->owner == NULL;
3054 }
3055 #endif
3057 #ifdef CONFIG_PREEMPT
3058 /*
3059 * this is the entry point to schedule() from in-kernel preemption
3060 * off of preempt_enable. Kernel preemptions off return from interrupt
3061 * occur there and call schedule directly.
3062 */
3063 asmlinkage void __sched notrace preempt_schedule(void)
3064 {
3065 struct thread_info *ti = current_thread_info();
3067 /*
3068 * If there is a non-zero preempt_count or interrupts are disabled,
3069 * we do not want to preempt the current task. Just return..
3070 */
3071 if (likely(ti->preempt_count || irqs_disabled()))
3072 return;
3074 do {
3075 add_preempt_count_notrace(PREEMPT_ACTIVE);
3076 __schedule();
3077 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3079 /*
3080 * Check again in case we missed a preemption opportunity
3081 * between schedule and now.
3082 */
3083 barrier();
3084 } while (need_resched());
3085 }
3086 EXPORT_SYMBOL(preempt_schedule);
3088 /*
3089 * this is the entry point to schedule() from kernel preemption
3090 * off of irq context.
3091 * Note, that this is called and return with irqs disabled. This will
3092 * protect us against recursive calling from irq.
3093 */
3094 asmlinkage void __sched preempt_schedule_irq(void)
3095 {
3096 struct thread_info *ti = current_thread_info();
3098 /* Catch callers which need to be fixed */
3099 BUG_ON(ti->preempt_count || !irqs_disabled());
3101 user_exit();
3102 do {
3103 add_preempt_count(PREEMPT_ACTIVE);
3104 local_irq_enable();
3105 __schedule();
3106 local_irq_disable();
3107 sub_preempt_count(PREEMPT_ACTIVE);
3109 /*
3110 * Check again in case we missed a preemption opportunity
3111 * between schedule and now.
3112 */
3113 barrier();
3114 } while (need_resched());
3115 }
3117 #endif /* CONFIG_PREEMPT */
3119 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3120 void *key)
3121 {
3122 return try_to_wake_up(curr->private, mode, wake_flags);
3123 }
3124 EXPORT_SYMBOL(default_wake_function);
3126 /*
3127 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3128 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3129 * number) then we wake all the non-exclusive tasks and one exclusive task.
3130 *
3131 * There are circumstances in which we can try to wake a task which has already
3132 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3133 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3134 */
3135 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3136 int nr_exclusive, int wake_flags, void *key)
3137 {
3138 wait_queue_t *curr, *next;
3140 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3141 unsigned flags = curr->flags;
3143 if (curr->func(curr, mode, wake_flags, key) &&
3144 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3145 break;
3146 }
3147 }
3149 /**
3150 * __wake_up - wake up threads blocked on a waitqueue.
3151 * @q: the waitqueue
3152 * @mode: which threads
3153 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3154 * @key: is directly passed to the wakeup function
3155 *
3156 * It may be assumed that this function implies a write memory barrier before
3157 * changing the task state if and only if any tasks are woken up.
3158 */
3159 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3160 int nr_exclusive, void *key)
3161 {
3162 unsigned long flags;
3164 spin_lock_irqsave(&q->lock, flags);
3165 __wake_up_common(q, mode, nr_exclusive, 0, key);
3166 spin_unlock_irqrestore(&q->lock, flags);
3167 }
3168 EXPORT_SYMBOL(__wake_up);
3170 /*
3171 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3172 */
3173 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3174 {
3175 __wake_up_common(q, mode, nr, 0, NULL);
3176 }
3177 EXPORT_SYMBOL_GPL(__wake_up_locked);
3179 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3180 {
3181 __wake_up_common(q, mode, 1, 0, key);
3182 }
3183 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3185 /**
3186 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3187 * @q: the waitqueue
3188 * @mode: which threads
3189 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3190 * @key: opaque value to be passed to wakeup targets
3191 *
3192 * The sync wakeup differs that the waker knows that it will schedule
3193 * away soon, so while the target thread will be woken up, it will not
3194 * be migrated to another CPU - ie. the two threads are 'synchronized'
3195 * with each other. This can prevent needless bouncing between CPUs.
3196 *
3197 * On UP it can prevent extra preemption.
3198 *
3199 * It may be assumed that this function implies a write memory barrier before
3200 * changing the task state if and only if any tasks are woken up.
3201 */
3202 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3203 int nr_exclusive, void *key)
3204 {
3205 unsigned long flags;
3206 int wake_flags = WF_SYNC;
3208 if (unlikely(!q))
3209 return;
3211 if (unlikely(!nr_exclusive))
3212 wake_flags = 0;
3214 spin_lock_irqsave(&q->lock, flags);
3215 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3216 spin_unlock_irqrestore(&q->lock, flags);
3217 }
3218 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3220 /*
3221 * __wake_up_sync - see __wake_up_sync_key()
3222 */
3223 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3224 {
3225 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3226 }
3227 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3229 /**
3230 * complete: - signals a single thread waiting on this completion
3231 * @x: holds the state of this particular completion
3232 *
3233 * This will wake up a single thread waiting on this completion. Threads will be
3234 * awakened in the same order in which they were queued.
3235 *
3236 * See also complete_all(), wait_for_completion() and related routines.
3237 *
3238 * It may be assumed that this function implies a write memory barrier before
3239 * changing the task state if and only if any tasks are woken up.
3240 */
3241 void complete(struct completion *x)
3242 {
3243 unsigned long flags;
3245 spin_lock_irqsave(&x->wait.lock, flags);
3246 x->done++;
3247 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3248 spin_unlock_irqrestore(&x->wait.lock, flags);
3249 }
3250 EXPORT_SYMBOL(complete);
3252 /**
3253 * complete_all: - signals all threads waiting on this completion
3254 * @x: holds the state of this particular completion
3255 *
3256 * This will wake up all threads waiting on this particular completion event.
3257 *
3258 * It may be assumed that this function implies a write memory barrier before
3259 * changing the task state if and only if any tasks are woken up.
3260 */
3261 void complete_all(struct completion *x)
3262 {
3263 unsigned long flags;
3265 spin_lock_irqsave(&x->wait.lock, flags);
3266 x->done += UINT_MAX/2;
3267 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3268 spin_unlock_irqrestore(&x->wait.lock, flags);
3269 }
3270 EXPORT_SYMBOL(complete_all);
3272 static inline long __sched
3273 do_wait_for_common(struct completion *x, long timeout, int state)
3274 {
3275 if (!x->done) {
3276 DECLARE_WAITQUEUE(wait, current);
3278 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3279 do {
3280 if (signal_pending_state(state, current)) {
3281 timeout = -ERESTARTSYS;
3282 break;
3283 }
3284 __set_current_state(state);
3285 spin_unlock_irq(&x->wait.lock);
3286 timeout = schedule_timeout(timeout);
3287 spin_lock_irq(&x->wait.lock);
3288 } while (!x->done && timeout);
3289 __remove_wait_queue(&x->wait, &wait);
3290 if (!x->done)
3291 return timeout;
3292 }
3293 x->done--;
3294 return timeout ?: 1;
3295 }
3297 static long __sched
3298 wait_for_common(struct completion *x, long timeout, int state)
3299 {
3300 might_sleep();
3302 spin_lock_irq(&x->wait.lock);
3303 timeout = do_wait_for_common(x, timeout, state);
3304 spin_unlock_irq(&x->wait.lock);
3305 return timeout;
3306 }
3308 /**
3309 * wait_for_completion: - waits for completion of a task
3310 * @x: holds the state of this particular completion
3311 *
3312 * This waits to be signaled for completion of a specific task. It is NOT
3313 * interruptible and there is no timeout.
3314 *
3315 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3316 * and interrupt capability. Also see complete().
3317 */
3318 void __sched wait_for_completion(struct completion *x)
3319 {
3320 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3321 }
3322 EXPORT_SYMBOL(wait_for_completion);
3324 /**
3325 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3326 * @x: holds the state of this particular completion
3327 * @timeout: timeout value in jiffies
3328 *
3329 * This waits for either a completion of a specific task to be signaled or for a
3330 * specified timeout to expire. The timeout is in jiffies. It is not
3331 * interruptible.
3332 *
3333 * The return value is 0 if timed out, and positive (at least 1, or number of
3334 * jiffies left till timeout) if completed.
3335 */
3336 unsigned long __sched
3337 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3338 {
3339 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3340 }
3341 EXPORT_SYMBOL(wait_for_completion_timeout);
3343 /**
3344 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3345 * @x: holds the state of this particular completion
3346 *
3347 * This waits for completion of a specific task to be signaled. It is
3348 * interruptible.
3349 *
3350 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3351 */
3352 int __sched wait_for_completion_interruptible(struct completion *x)
3353 {
3354 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3355 if (t == -ERESTARTSYS)
3356 return t;
3357 return 0;
3358 }
3359 EXPORT_SYMBOL(wait_for_completion_interruptible);
3361 /**
3362 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3363 * @x: holds the state of this particular completion
3364 * @timeout: timeout value in jiffies
3365 *
3366 * This waits for either a completion of a specific task to be signaled or for a
3367 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3368 *
3369 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3370 * positive (at least 1, or number of jiffies left till timeout) if completed.
3371 */
3372 long __sched
3373 wait_for_completion_interruptible_timeout(struct completion *x,
3374 unsigned long timeout)
3375 {
3376 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3377 }
3378 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3380 /**
3381 * wait_for_completion_killable: - waits for completion of a task (killable)
3382 * @x: holds the state of this particular completion
3383 *
3384 * This waits to be signaled for completion of a specific task. It can be
3385 * interrupted by a kill signal.
3386 *
3387 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3388 */
3389 int __sched wait_for_completion_killable(struct completion *x)
3390 {
3391 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3392 if (t == -ERESTARTSYS)
3393 return t;
3394 return 0;
3395 }
3396 EXPORT_SYMBOL(wait_for_completion_killable);
3398 /**
3399 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3400 * @x: holds the state of this particular completion
3401 * @timeout: timeout value in jiffies
3402 *
3403 * This waits for either a completion of a specific task to be
3404 * signaled or for a specified timeout to expire. It can be
3405 * interrupted by a kill signal. The timeout is in jiffies.
3406 *
3407 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3408 * positive (at least 1, or number of jiffies left till timeout) if completed.
3409 */
3410 long __sched
3411 wait_for_completion_killable_timeout(struct completion *x,
3412 unsigned long timeout)
3413 {
3414 return wait_for_common(x, timeout, TASK_KILLABLE);
3415 }
3416 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3418 /**
3419 * try_wait_for_completion - try to decrement a completion without blocking
3420 * @x: completion structure
3421 *
3422 * Returns: 0 if a decrement cannot be done without blocking
3423 * 1 if a decrement succeeded.
3424 *
3425 * If a completion is being used as a counting completion,
3426 * attempt to decrement the counter without blocking. This
3427 * enables us to avoid waiting if the resource the completion
3428 * is protecting is not available.
3429 */
3430 bool try_wait_for_completion(struct completion *x)
3431 {
3432 unsigned long flags;
3433 int ret = 1;
3435 spin_lock_irqsave(&x->wait.lock, flags);
3436 if (!x->done)
3437 ret = 0;
3438 else
3439 x->done--;
3440 spin_unlock_irqrestore(&x->wait.lock, flags);
3441 return ret;
3442 }
3443 EXPORT_SYMBOL(try_wait_for_completion);
3445 /**
3446 * completion_done - Test to see if a completion has any waiters
3447 * @x: completion structure
3448 *
3449 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3450 * 1 if there are no waiters.
3451 *
3452 */
3453 bool completion_done(struct completion *x)
3454 {
3455 unsigned long flags;
3456 int ret = 1;
3458 spin_lock_irqsave(&x->wait.lock, flags);
3459 if (!x->done)
3460 ret = 0;
3461 spin_unlock_irqrestore(&x->wait.lock, flags);
3462 return ret;
3463 }
3464 EXPORT_SYMBOL(completion_done);
3466 static long __sched
3467 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3468 {
3469 unsigned long flags;
3470 wait_queue_t wait;
3472 init_waitqueue_entry(&wait, current);
3474 __set_current_state(state);
3476 spin_lock_irqsave(&q->lock, flags);
3477 __add_wait_queue(q, &wait);
3478 spin_unlock(&q->lock);
3479 timeout = schedule_timeout(timeout);
3480 spin_lock_irq(&q->lock);
3481 __remove_wait_queue(q, &wait);
3482 spin_unlock_irqrestore(&q->lock, flags);
3484 return timeout;
3485 }
3487 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3488 {
3489 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3490 }
3491 EXPORT_SYMBOL(interruptible_sleep_on);
3493 long __sched
3494 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3495 {
3496 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3497 }
3498 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3500 void __sched sleep_on(wait_queue_head_t *q)
3501 {
3502 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3503 }
3504 EXPORT_SYMBOL(sleep_on);
3506 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3507 {
3508 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3509 }
3510 EXPORT_SYMBOL(sleep_on_timeout);
3512 #ifdef CONFIG_RT_MUTEXES
3514 /*
3515 * rt_mutex_setprio - set the current priority of a task
3516 * @p: task
3517 * @prio: prio value (kernel-internal form)
3518 *
3519 * This function changes the 'effective' priority of a task. It does
3520 * not touch ->normal_prio like __setscheduler().
3521 *
3522 * Used by the rt_mutex code to implement priority inheritance logic.
3523 */
3524 void rt_mutex_setprio(struct task_struct *p, int prio)
3525 {
3526 int oldprio, on_rq, running;
3527 struct rq *rq;
3528 const struct sched_class *prev_class;
3530 BUG_ON(prio < 0 || prio > MAX_PRIO);
3532 rq = __task_rq_lock(p);
3534 /*
3535 * Idle task boosting is a nono in general. There is one
3536 * exception, when PREEMPT_RT and NOHZ is active:
3537 *
3538 * The idle task calls get_next_timer_interrupt() and holds
3539 * the timer wheel base->lock on the CPU and another CPU wants
3540 * to access the timer (probably to cancel it). We can safely
3541 * ignore the boosting request, as the idle CPU runs this code
3542 * with interrupts disabled and will complete the lock
3543 * protected section without being interrupted. So there is no
3544 * real need to boost.
3545 */
3546 if (unlikely(p == rq->idle)) {
3547 WARN_ON(p != rq->curr);
3548 WARN_ON(p->pi_blocked_on);
3549 goto out_unlock;
3550 }
3552 trace_sched_pi_setprio(p, prio);
3553 oldprio = p->prio;
3554 prev_class = p->sched_class;
3555 on_rq = p->on_rq;
3556 running = task_current(rq, p);
3557 if (on_rq)
3558 dequeue_task(rq, p, 0);
3559 if (running)
3560 p->sched_class->put_prev_task(rq, p);
3562 if (rt_prio(prio))
3563 p->sched_class = &rt_sched_class;
3564 else
3565 p->sched_class = &fair_sched_class;
3567 p->prio = prio;
3569 if (running)
3570 p->sched_class->set_curr_task(rq);
3571 if (on_rq)
3572 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3574 check_class_changed(rq, p, prev_class, oldprio);
3575 out_unlock:
3576 __task_rq_unlock(rq);
3577 }
3578 #endif
3579 void set_user_nice(struct task_struct *p, long nice)
3580 {
3581 int old_prio, delta, on_rq;
3582 unsigned long flags;
3583 struct rq *rq;
3585 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3586 return;
3587 /*
3588 * We have to be careful, if called from sys_setpriority(),
3589 * the task might be in the middle of scheduling on another CPU.
3590 */
3591 rq = task_rq_lock(p, &flags);
3592 /*
3593 * The RT priorities are set via sched_setscheduler(), but we still
3594 * allow the 'normal' nice value to be set - but as expected
3595 * it wont have any effect on scheduling until the task is
3596 * SCHED_FIFO/SCHED_RR:
3597 */
3598 if (task_has_rt_policy(p)) {
3599 p->static_prio = NICE_TO_PRIO(nice);
3600 goto out_unlock;
3601 }
3602 on_rq = p->on_rq;
3603 if (on_rq)
3604 dequeue_task(rq, p, 0);
3606 p->static_prio = NICE_TO_PRIO(nice);
3607 set_load_weight(p);
3608 old_prio = p->prio;
3609 p->prio = effective_prio(p);
3610 delta = p->prio - old_prio;
3612 if (on_rq) {
3613 enqueue_task(rq, p, 0);
3614 /*
3615 * If the task increased its priority or is running and
3616 * lowered its priority, then reschedule its CPU:
3617 */
3618 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3619 resched_task(rq->curr);
3620 }
3621 out_unlock:
3622 task_rq_unlock(rq, p, &flags);
3623 }
3624 EXPORT_SYMBOL(set_user_nice);
3626 /*
3627 * can_nice - check if a task can reduce its nice value
3628 * @p: task
3629 * @nice: nice value
3630 */
3631 int can_nice(const struct task_struct *p, const int nice)
3632 {
3633 /* convert nice value [19,-20] to rlimit style value [1,40] */
3634 int nice_rlim = 20 - nice;
3636 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3637 capable(CAP_SYS_NICE));
3638 }
3640 #ifdef __ARCH_WANT_SYS_NICE
3642 /*
3643 * sys_nice - change the priority of the current process.
3644 * @increment: priority increment
3645 *
3646 * sys_setpriority is a more generic, but much slower function that
3647 * does similar things.
3648 */
3649 SYSCALL_DEFINE1(nice, int, increment)
3650 {
3651 long nice, retval;
3653 /*
3654 * Setpriority might change our priority at the same moment.
3655 * We don't have to worry. Conceptually one call occurs first
3656 * and we have a single winner.
3657 */
3658 if (increment < -40)
3659 increment = -40;
3660 if (increment > 40)
3661 increment = 40;
3663 nice = TASK_NICE(current) + increment;
3664 if (nice < -20)
3665 nice = -20;
3666 if (nice > 19)
3667 nice = 19;
3669 if (increment < 0 && !can_nice(current, nice))
3670 return -EPERM;
3672 retval = security_task_setnice(current, nice);
3673 if (retval)
3674 return retval;
3676 set_user_nice(current, nice);
3677 return 0;
3678 }
3680 #endif
3682 /**
3683 * task_prio - return the priority value of a given task.
3684 * @p: the task in question.
3685 *
3686 * This is the priority value as seen by users in /proc.
3687 * RT tasks are offset by -200. Normal tasks are centered
3688 * around 0, value goes from -16 to +15.
3689 */
3690 int task_prio(const struct task_struct *p)
3691 {
3692 return p->prio - MAX_RT_PRIO;
3693 }
3695 /**
3696 * task_nice - return the nice value of a given task.
3697 * @p: the task in question.
3698 */
3699 int task_nice(const struct task_struct *p)
3700 {
3701 return TASK_NICE(p);
3702 }
3703 EXPORT_SYMBOL(task_nice);
3705 /**
3706 * idle_cpu - is a given cpu idle currently?
3707 * @cpu: the processor in question.
3708 */
3709 int idle_cpu(int cpu)
3710 {
3711 struct rq *rq = cpu_rq(cpu);
3713 if (rq->curr != rq->idle)
3714 return 0;
3716 if (rq->nr_running)
3717 return 0;
3719 #ifdef CONFIG_SMP
3720 if (!llist_empty(&rq->wake_list))
3721 return 0;
3722 #endif
3724 return 1;
3725 }
3727 /**
3728 * idle_task - return the idle task for a given cpu.
3729 * @cpu: the processor in question.
3730 */
3731 struct task_struct *idle_task(int cpu)
3732 {
3733 return cpu_rq(cpu)->idle;
3734 }
3736 /**
3737 * find_process_by_pid - find a process with a matching PID value.
3738 * @pid: the pid in question.
3739 */
3740 static struct task_struct *find_process_by_pid(pid_t pid)
3741 {
3742 return pid ? find_task_by_vpid(pid) : current;
3743 }
3745 /* Actually do priority change: must hold rq lock. */
3746 static void
3747 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3748 {
3749 p->policy = policy;
3750 p->rt_priority = prio;
3751 p->normal_prio = normal_prio(p);
3752 /* we are holding p->pi_lock already */
3753 p->prio = rt_mutex_getprio(p);
3754 if (rt_prio(p->prio))
3755 p->sched_class = &rt_sched_class;
3756 else
3757 p->sched_class = &fair_sched_class;
3758 set_load_weight(p);
3759 }
3761 /*
3762 * check the target process has a UID that matches the current process's
3763 */
3764 static bool check_same_owner(struct task_struct *p)
3765 {
3766 const struct cred *cred = current_cred(), *pcred;
3767 bool match;
3769 rcu_read_lock();
3770 pcred = __task_cred(p);
3771 match = (uid_eq(cred->euid, pcred->euid) ||
3772 uid_eq(cred->euid, pcred->uid));
3773 rcu_read_unlock();
3774 return match;
3775 }
3777 static int __sched_setscheduler(struct task_struct *p, int policy,
3778 const struct sched_param *param, bool user)
3779 {
3780 int retval, oldprio, oldpolicy = -1, on_rq, running;
3781 unsigned long flags;
3782 const struct sched_class *prev_class;
3783 struct rq *rq;
3784 int reset_on_fork;
3786 /* may grab non-irq protected spin_locks */
3787 BUG_ON(in_interrupt());
3788 recheck:
3789 /* double check policy once rq lock held */
3790 if (policy < 0) {
3791 reset_on_fork = p->sched_reset_on_fork;
3792 policy = oldpolicy = p->policy;
3793 } else {
3794 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3795 policy &= ~SCHED_RESET_ON_FORK;
3797 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3798 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3799 policy != SCHED_IDLE)
3800 return -EINVAL;
3801 }
3803 /*
3804 * Valid priorities for SCHED_FIFO and SCHED_RR are
3805 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3806 * SCHED_BATCH and SCHED_IDLE is 0.
3807 */
3808 if (param->sched_priority < 0 ||
3809 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3810 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3811 return -EINVAL;
3812 if (rt_policy(policy) != (param->sched_priority != 0))
3813 return -EINVAL;
3815 /*
3816 * Allow unprivileged RT tasks to decrease priority:
3817 */
3818 if (user && !capable(CAP_SYS_NICE)) {
3819 if (rt_policy(policy)) {
3820 unsigned long rlim_rtprio =
3821 task_rlimit(p, RLIMIT_RTPRIO);
3823 /* can't set/change the rt policy */
3824 if (policy != p->policy && !rlim_rtprio)
3825 return -EPERM;
3827 /* can't increase priority */
3828 if (param->sched_priority > p->rt_priority &&
3829 param->sched_priority > rlim_rtprio)
3830 return -EPERM;
3831 }
3833 /*
3834 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3835 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3836 */
3837 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3838 if (!can_nice(p, TASK_NICE(p)))
3839 return -EPERM;
3840 }
3842 /* can't change other user's priorities */
3843 if (!check_same_owner(p))
3844 return -EPERM;
3846 /* Normal users shall not reset the sched_reset_on_fork flag */
3847 if (p->sched_reset_on_fork && !reset_on_fork)
3848 return -EPERM;
3849 }
3851 if (user) {
3852 retval = security_task_setscheduler(p);
3853 if (retval)
3854 return retval;
3855 }
3857 /*
3858 * make sure no PI-waiters arrive (or leave) while we are
3859 * changing the priority of the task:
3860 *
3861 * To be able to change p->policy safely, the appropriate
3862 * runqueue lock must be held.
3863 */
3864 rq = task_rq_lock(p, &flags);
3866 /*
3867 * Changing the policy of the stop threads its a very bad idea
3868 */
3869 if (p == rq->stop) {
3870 task_rq_unlock(rq, p, &flags);
3871 return -EINVAL;
3872 }
3874 /*
3875 * If not changing anything there's no need to proceed further:
3876 */
3877 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3878 param->sched_priority == p->rt_priority))) {
3879 task_rq_unlock(rq, p, &flags);
3880 return 0;
3881 }
3883 #ifdef CONFIG_RT_GROUP_SCHED
3884 if (user) {
3885 /*
3886 * Do not allow realtime tasks into groups that have no runtime
3887 * assigned.
3888 */
3889 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3890 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3891 !task_group_is_autogroup(task_group(p))) {
3892 task_rq_unlock(rq, p, &flags);
3893 return -EPERM;
3894 }
3895 }
3896 #endif
3898 /* recheck policy now with rq lock held */
3899 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3900 policy = oldpolicy = -1;
3901 task_rq_unlock(rq, p, &flags);
3902 goto recheck;
3903 }
3904 on_rq = p->on_rq;
3905 running = task_current(rq, p);
3906 if (on_rq)
3907 dequeue_task(rq, p, 0);
3908 if (running)
3909 p->sched_class->put_prev_task(rq, p);
3911 p->sched_reset_on_fork = reset_on_fork;
3913 oldprio = p->prio;
3914 prev_class = p->sched_class;
3915 __setscheduler(rq, p, policy, param->sched_priority);
3917 if (running)
3918 p->sched_class->set_curr_task(rq);
3919 if (on_rq)
3920 enqueue_task(rq, p, 0);
3922 check_class_changed(rq, p, prev_class, oldprio);
3923 task_rq_unlock(rq, p, &flags);
3925 rt_mutex_adjust_pi(p);
3927 return 0;
3928 }
3930 /**
3931 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3932 * @p: the task in question.
3933 * @policy: new policy.
3934 * @param: structure containing the new RT priority.
3935 *
3936 * NOTE that the task may be already dead.
3937 */
3938 int sched_setscheduler(struct task_struct *p, int policy,
3939 const struct sched_param *param)
3940 {
3941 return __sched_setscheduler(p, policy, param, true);
3942 }
3943 EXPORT_SYMBOL_GPL(sched_setscheduler);
3945 /**
3946 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3947 * @p: the task in question.
3948 * @policy: new policy.
3949 * @param: structure containing the new RT priority.
3950 *
3951 * Just like sched_setscheduler, only don't bother checking if the
3952 * current context has permission. For example, this is needed in
3953 * stop_machine(): we create temporary high priority worker threads,
3954 * but our caller might not have that capability.
3955 */
3956 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3957 const struct sched_param *param)
3958 {
3959 return __sched_setscheduler(p, policy, param, false);
3960 }
3962 static int
3963 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3964 {
3965 struct sched_param lparam;
3966 struct task_struct *p;
3967 int retval;
3969 if (!param || pid < 0)
3970 return -EINVAL;
3971 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3972 return -EFAULT;
3974 rcu_read_lock();
3975 retval = -ESRCH;
3976 p = find_process_by_pid(pid);
3977 if (p != NULL)
3978 retval = sched_setscheduler(p, policy, &lparam);
3979 rcu_read_unlock();
3981 return retval;
3982 }
3984 /**
3985 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3986 * @pid: the pid in question.
3987 * @policy: new policy.
3988 * @param: structure containing the new RT priority.
3989 */
3990 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3991 struct sched_param __user *, param)
3992 {
3993 /* negative values for policy are not valid */
3994 if (policy < 0)
3995 return -EINVAL;
3997 return do_sched_setscheduler(pid, policy, param);
3998 }
4000 /**
4001 * sys_sched_setparam - set/change the RT priority of a thread
4002 * @pid: the pid in question.
4003 * @param: structure containing the new RT priority.
4004 */
4005 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4006 {
4007 return do_sched_setscheduler(pid, -1, param);
4008 }
4010 /**
4011 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4012 * @pid: the pid in question.
4013 */
4014 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4015 {
4016 struct task_struct *p;
4017 int retval;
4019 if (pid < 0)
4020 return -EINVAL;
4022 retval = -ESRCH;
4023 rcu_read_lock();
4024 p = find_process_by_pid(pid);
4025 if (p) {
4026 retval = security_task_getscheduler(p);
4027 if (!retval)
4028 retval = p->policy
4029 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4030 }
4031 rcu_read_unlock();
4032 return retval;
4033 }
4035 /**
4036 * sys_sched_getparam - get the RT priority of a thread
4037 * @pid: the pid in question.
4038 * @param: structure containing the RT priority.
4039 */
4040 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4041 {
4042 struct sched_param lp;
4043 struct task_struct *p;
4044 int retval;
4046 if (!param || pid < 0)
4047 return -EINVAL;
4049 rcu_read_lock();
4050 p = find_process_by_pid(pid);
4051 retval = -ESRCH;
4052 if (!p)
4053 goto out_unlock;
4055 retval = security_task_getscheduler(p);
4056 if (retval)
4057 goto out_unlock;
4059 lp.sched_priority = p->rt_priority;
4060 rcu_read_unlock();
4062 /*
4063 * This one might sleep, we cannot do it with a spinlock held ...
4064 */
4065 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4067 return retval;
4069 out_unlock:
4070 rcu_read_unlock();
4071 return retval;
4072 }
4074 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4075 {
4076 cpumask_var_t cpus_allowed, new_mask;
4077 struct task_struct *p;
4078 int retval;
4080 get_online_cpus();
4081 rcu_read_lock();
4083 p = find_process_by_pid(pid);
4084 if (!p) {
4085 rcu_read_unlock();
4086 put_online_cpus();
4087 return -ESRCH;
4088 }
4090 /* Prevent p going away */
4091 get_task_struct(p);
4092 rcu_read_unlock();
4094 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4095 retval = -ENOMEM;
4096 goto out_put_task;
4097 }
4098 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4099 retval = -ENOMEM;
4100 goto out_free_cpus_allowed;
4101 }
4102 retval = -EPERM;
4103 if (!check_same_owner(p)) {
4104 rcu_read_lock();
4105 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4106 rcu_read_unlock();
4107 goto out_unlock;
4108 }
4109 rcu_read_unlock();
4110 }
4112 retval = security_task_setscheduler(p);
4113 if (retval)
4114 goto out_unlock;
4116 cpuset_cpus_allowed(p, cpus_allowed);
4117 cpumask_and(new_mask, in_mask, cpus_allowed);
4118 again:
4119 retval = set_cpus_allowed_ptr(p, new_mask);
4121 if (!retval) {
4122 cpuset_cpus_allowed(p, cpus_allowed);
4123 if (!cpumask_subset(new_mask, cpus_allowed)) {
4124 /*
4125 * We must have raced with a concurrent cpuset
4126 * update. Just reset the cpus_allowed to the
4127 * cpuset's cpus_allowed
4128 */
4129 cpumask_copy(new_mask, cpus_allowed);
4130 goto again;
4131 }
4132 }
4133 out_unlock:
4134 free_cpumask_var(new_mask);
4135 out_free_cpus_allowed:
4136 free_cpumask_var(cpus_allowed);
4137 out_put_task:
4138 put_task_struct(p);
4139 put_online_cpus();
4140 return retval;
4141 }
4143 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4144 struct cpumask *new_mask)
4145 {
4146 if (len < cpumask_size())
4147 cpumask_clear(new_mask);
4148 else if (len > cpumask_size())
4149 len = cpumask_size();
4151 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4152 }
4154 /**
4155 * sys_sched_setaffinity - set the cpu affinity of a process
4156 * @pid: pid of the process
4157 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4158 * @user_mask_ptr: user-space pointer to the new cpu mask
4159 */
4160 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4161 unsigned long __user *, user_mask_ptr)
4162 {
4163 cpumask_var_t new_mask;
4164 int retval;
4166 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4167 return -ENOMEM;
4169 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4170 if (retval == 0)
4171 retval = sched_setaffinity(pid, new_mask);
4172 free_cpumask_var(new_mask);
4173 return retval;
4174 }
4176 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4177 {
4178 struct task_struct *p;
4179 unsigned long flags;
4180 int retval;
4182 get_online_cpus();
4183 rcu_read_lock();
4185 retval = -ESRCH;
4186 p = find_process_by_pid(pid);
4187 if (!p)
4188 goto out_unlock;
4190 retval = security_task_getscheduler(p);
4191 if (retval)
4192 goto out_unlock;
4194 raw_spin_lock_irqsave(&p->pi_lock, flags);
4195 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4196 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4198 out_unlock:
4199 rcu_read_unlock();
4200 put_online_cpus();
4202 return retval;
4203 }
4205 /**
4206 * sys_sched_getaffinity - get the cpu affinity of a process
4207 * @pid: pid of the process
4208 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4209 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4210 */
4211 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4212 unsigned long __user *, user_mask_ptr)
4213 {
4214 int ret;
4215 cpumask_var_t mask;
4217 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4218 return -EINVAL;
4219 if (len & (sizeof(unsigned long)-1))
4220 return -EINVAL;
4222 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4223 return -ENOMEM;
4225 ret = sched_getaffinity(pid, mask);
4226 if (ret == 0) {
4227 size_t retlen = min_t(size_t, len, cpumask_size());
4229 if (copy_to_user(user_mask_ptr, mask, retlen))
4230 ret = -EFAULT;
4231 else
4232 ret = retlen;
4233 }
4234 free_cpumask_var(mask);
4236 return ret;
4237 }
4239 /**
4240 * sys_sched_yield - yield the current processor to other threads.
4241 *
4242 * This function yields the current CPU to other tasks. If there are no
4243 * other threads running on this CPU then this function will return.
4244 */
4245 SYSCALL_DEFINE0(sched_yield)
4246 {
4247 struct rq *rq = this_rq_lock();
4249 schedstat_inc(rq, yld_count);
4250 current->sched_class->yield_task(rq);
4252 /*
4253 * Since we are going to call schedule() anyway, there's
4254 * no need to preempt or enable interrupts:
4255 */
4256 __release(rq->lock);
4257 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4258 do_raw_spin_unlock(&rq->lock);
4259 sched_preempt_enable_no_resched();
4261 schedule();
4263 return 0;
4264 }
4266 static inline int should_resched(void)
4267 {
4268 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4269 }
4271 static void __cond_resched(void)
4272 {
4273 add_preempt_count(PREEMPT_ACTIVE);
4274 __schedule();
4275 sub_preempt_count(PREEMPT_ACTIVE);
4276 }
4278 int __sched _cond_resched(void)
4279 {
4280 if (should_resched()) {
4281 __cond_resched();
4282 return 1;
4283 }
4284 return 0;
4285 }
4286 EXPORT_SYMBOL(_cond_resched);
4288 /*
4289 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4290 * call schedule, and on return reacquire the lock.
4291 *
4292 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4293 * operations here to prevent schedule() from being called twice (once via
4294 * spin_unlock(), once by hand).
4295 */
4296 int __cond_resched_lock(spinlock_t *lock)
4297 {
4298 int resched = should_resched();
4299 int ret = 0;
4301 lockdep_assert_held(lock);
4303 if (spin_needbreak(lock) || resched) {
4304 spin_unlock(lock);
4305 if (resched)
4306 __cond_resched();
4307 else
4308 cpu_relax();
4309 ret = 1;
4310 spin_lock(lock);
4311 }
4312 return ret;
4313 }
4314 EXPORT_SYMBOL(__cond_resched_lock);
4316 int __sched __cond_resched_softirq(void)
4317 {
4318 BUG_ON(!in_softirq());
4320 if (should_resched()) {
4321 local_bh_enable();
4322 __cond_resched();
4323 local_bh_disable();
4324 return 1;
4325 }
4326 return 0;
4327 }
4328 EXPORT_SYMBOL(__cond_resched_softirq);
4330 /**
4331 * yield - yield the current processor to other threads.
4332 *
4333 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4334 *
4335 * The scheduler is at all times free to pick the calling task as the most
4336 * eligible task to run, if removing the yield() call from your code breaks
4337 * it, its already broken.
4338 *
4339 * Typical broken usage is:
4340 *
4341 * while (!event)
4342 * yield();
4343 *
4344 * where one assumes that yield() will let 'the other' process run that will
4345 * make event true. If the current task is a SCHED_FIFO task that will never
4346 * happen. Never use yield() as a progress guarantee!!
4347 *
4348 * If you want to use yield() to wait for something, use wait_event().
4349 * If you want to use yield() to be 'nice' for others, use cond_resched().
4350 * If you still want to use yield(), do not!
4351 */
4352 void __sched yield(void)
4353 {
4354 set_current_state(TASK_RUNNING);
4355 sys_sched_yield();
4356 }
4357 EXPORT_SYMBOL(yield);
4359 /**
4360 * yield_to - yield the current processor to another thread in
4361 * your thread group, or accelerate that thread toward the
4362 * processor it's on.
4363 * @p: target task
4364 * @preempt: whether task preemption is allowed or not
4365 *
4366 * It's the caller's job to ensure that the target task struct
4367 * can't go away on us before we can do any checks.
4368 *
4369 * Returns true if we indeed boosted the target task.
4370 */
4371 bool __sched yield_to(struct task_struct *p, bool preempt)
4372 {
4373 struct task_struct *curr = current;
4374 struct rq *rq, *p_rq;
4375 unsigned long flags;
4376 bool yielded = 0;
4378 local_irq_save(flags);
4379 rq = this_rq();
4381 again:
4382 p_rq = task_rq(p);
4383 double_rq_lock(rq, p_rq);
4384 while (task_rq(p) != p_rq) {
4385 double_rq_unlock(rq, p_rq);
4386 goto again;
4387 }
4389 if (!curr->sched_class->yield_to_task)
4390 goto out;
4392 if (curr->sched_class != p->sched_class)
4393 goto out;
4395 if (task_running(p_rq, p) || p->state)
4396 goto out;
4398 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4399 if (yielded) {
4400 schedstat_inc(rq, yld_count);
4401 /*
4402 * Make p's CPU reschedule; pick_next_entity takes care of
4403 * fairness.
4404 */
4405 if (preempt && rq != p_rq)
4406 resched_task(p_rq->curr);
4407 }
4409 out:
4410 double_rq_unlock(rq, p_rq);
4411 local_irq_restore(flags);
4413 if (yielded)
4414 schedule();
4416 return yielded;
4417 }
4418 EXPORT_SYMBOL_GPL(yield_to);
4420 /*
4421 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4422 * that process accounting knows that this is a task in IO wait state.
4423 */
4424 void __sched io_schedule(void)
4425 {
4426 struct rq *rq = raw_rq();
4428 delayacct_blkio_start();
4429 atomic_inc(&rq->nr_iowait);
4430 blk_flush_plug(current);
4431 current->in_iowait = 1;
4432 schedule();
4433 current->in_iowait = 0;
4434 atomic_dec(&rq->nr_iowait);
4435 delayacct_blkio_end();
4436 }
4437 EXPORT_SYMBOL(io_schedule);
4439 long __sched io_schedule_timeout(long timeout)
4440 {
4441 struct rq *rq = raw_rq();
4442 long ret;
4444 delayacct_blkio_start();
4445 atomic_inc(&rq->nr_iowait);
4446 blk_flush_plug(current);
4447 current->in_iowait = 1;
4448 ret = schedule_timeout(timeout);
4449 current->in_iowait = 0;
4450 atomic_dec(&rq->nr_iowait);
4451 delayacct_blkio_end();
4452 return ret;
4453 }
4455 /**
4456 * sys_sched_get_priority_max - return maximum RT priority.
4457 * @policy: scheduling class.
4458 *
4459 * this syscall returns the maximum rt_priority that can be used
4460 * by a given scheduling class.
4461 */
4462 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4463 {
4464 int ret = -EINVAL;
4466 switch (policy) {
4467 case SCHED_FIFO:
4468 case SCHED_RR:
4469 ret = MAX_USER_RT_PRIO-1;
4470 break;
4471 case SCHED_NORMAL:
4472 case SCHED_BATCH:
4473 case SCHED_IDLE:
4474 ret = 0;
4475 break;
4476 }
4477 return ret;
4478 }
4480 /**
4481 * sys_sched_get_priority_min - return minimum RT priority.
4482 * @policy: scheduling class.
4483 *
4484 * this syscall returns the minimum rt_priority that can be used
4485 * by a given scheduling class.
4486 */
4487 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4488 {
4489 int ret = -EINVAL;
4491 switch (policy) {
4492 case SCHED_FIFO:
4493 case SCHED_RR:
4494 ret = 1;
4495 break;
4496 case SCHED_NORMAL:
4497 case SCHED_BATCH:
4498 case SCHED_IDLE:
4499 ret = 0;
4500 }
4501 return ret;
4502 }
4504 /**
4505 * sys_sched_rr_get_interval - return the default timeslice of a process.
4506 * @pid: pid of the process.
4507 * @interval: userspace pointer to the timeslice value.
4508 *
4509 * this syscall writes the default timeslice value of a given process
4510 * into the user-space timespec buffer. A value of '0' means infinity.
4511 */
4512 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4513 struct timespec __user *, interval)
4514 {
4515 struct task_struct *p;
4516 unsigned int time_slice;
4517 unsigned long flags;
4518 struct rq *rq;
4519 int retval;
4520 struct timespec t;
4522 if (pid < 0)
4523 return -EINVAL;
4525 retval = -ESRCH;
4526 rcu_read_lock();
4527 p = find_process_by_pid(pid);
4528 if (!p)
4529 goto out_unlock;
4531 retval = security_task_getscheduler(p);
4532 if (retval)
4533 goto out_unlock;
4535 rq = task_rq_lock(p, &flags);
4536 time_slice = p->sched_class->get_rr_interval(rq, p);
4537 task_rq_unlock(rq, p, &flags);
4539 rcu_read_unlock();
4540 jiffies_to_timespec(time_slice, &t);
4541 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4542 return retval;
4544 out_unlock:
4545 rcu_read_unlock();
4546 return retval;
4547 }
4549 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4551 void sched_show_task(struct task_struct *p)
4552 {
4553 unsigned long free = 0;
4554 int ppid;
4555 unsigned state;
4557 state = p->state ? __ffs(p->state) + 1 : 0;
4558 printk(KERN_INFO "%-15.15s %c", p->comm,
4559 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4560 #if BITS_PER_LONG == 32
4561 if (state == TASK_RUNNING)
4562 printk(KERN_CONT " running ");
4563 else
4564 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4565 #else
4566 if (state == TASK_RUNNING)
4567 printk(KERN_CONT " running task ");
4568 else
4569 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4570 #endif
4571 #ifdef CONFIG_DEBUG_STACK_USAGE
4572 free = stack_not_used(p);
4573 #endif
4574 rcu_read_lock();
4575 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4576 rcu_read_unlock();
4577 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4578 task_pid_nr(p), ppid,
4579 (unsigned long)task_thread_info(p)->flags);
4581 show_stack(p, NULL);
4582 }
4584 void show_state_filter(unsigned long state_filter)
4585 {
4586 struct task_struct *g, *p;
4588 #if BITS_PER_LONG == 32
4589 printk(KERN_INFO
4590 " task PC stack pid father\n");
4591 #else
4592 printk(KERN_INFO
4593 " task PC stack pid father\n");
4594 #endif
4595 rcu_read_lock();
4596 do_each_thread(g, p) {
4597 /*
4598 * reset the NMI-timeout, listing all files on a slow
4599 * console might take a lot of time:
4600 */
4601 touch_nmi_watchdog();
4602 if (!state_filter || (p->state & state_filter))
4603 sched_show_task(p);
4604 } while_each_thread(g, p);
4606 touch_all_softlockup_watchdogs();
4608 #ifdef CONFIG_SCHED_DEBUG
4609 sysrq_sched_debug_show();
4610 #endif
4611 rcu_read_unlock();
4612 /*
4613 * Only show locks if all tasks are dumped:
4614 */
4615 if (!state_filter)
4616 debug_show_all_locks();
4617 }
4619 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4620 {
4621 idle->sched_class = &idle_sched_class;
4622 }
4624 /**
4625 * init_idle - set up an idle thread for a given CPU
4626 * @idle: task in question
4627 * @cpu: cpu the idle task belongs to
4628 *
4629 * NOTE: this function does not set the idle thread's NEED_RESCHED
4630 * flag, to make booting more robust.
4631 */
4632 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4633 {
4634 struct rq *rq = cpu_rq(cpu);
4635 unsigned long flags;
4637 raw_spin_lock_irqsave(&rq->lock, flags);
4639 __sched_fork(idle);
4640 idle->state = TASK_RUNNING;
4641 idle->se.exec_start = sched_clock();
4643 do_set_cpus_allowed(idle, cpumask_of(cpu));
4644 /*
4645 * We're having a chicken and egg problem, even though we are
4646 * holding rq->lock, the cpu isn't yet set to this cpu so the
4647 * lockdep check in task_group() will fail.
4648 *
4649 * Similar case to sched_fork(). / Alternatively we could
4650 * use task_rq_lock() here and obtain the other rq->lock.
4651 *
4652 * Silence PROVE_RCU
4653 */
4654 rcu_read_lock();
4655 __set_task_cpu(idle, cpu);
4656 rcu_read_unlock();
4658 rq->curr = rq->idle = idle;
4659 #if defined(CONFIG_SMP)
4660 idle->on_cpu = 1;
4661 #endif
4662 raw_spin_unlock_irqrestore(&rq->lock, flags);
4664 /* Set the preempt count _outside_ the spinlocks! */
4665 task_thread_info(idle)->preempt_count = 0;
4667 /*
4668 * The idle tasks have their own, simple scheduling class:
4669 */
4670 idle->sched_class = &idle_sched_class;
4671 ftrace_graph_init_idle_task(idle, cpu);
4672 #if defined(CONFIG_SMP)
4673 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4674 #endif
4675 }
4677 #ifdef CONFIG_SMP
4678 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4679 {
4680 if (p->sched_class && p->sched_class->set_cpus_allowed)
4681 p->sched_class->set_cpus_allowed(p, new_mask);
4683 cpumask_copy(&p->cpus_allowed, new_mask);
4684 p->nr_cpus_allowed = cpumask_weight(new_mask);
4685 }
4687 /*
4688 * This is how migration works:
4689 *
4690 * 1) we invoke migration_cpu_stop() on the target CPU using
4691 * stop_one_cpu().
4692 * 2) stopper starts to run (implicitly forcing the migrated thread
4693 * off the CPU)
4694 * 3) it checks whether the migrated task is still in the wrong runqueue.
4695 * 4) if it's in the wrong runqueue then the migration thread removes
4696 * it and puts it into the right queue.
4697 * 5) stopper completes and stop_one_cpu() returns and the migration
4698 * is done.
4699 */
4701 /*
4702 * Change a given task's CPU affinity. Migrate the thread to a
4703 * proper CPU and schedule it away if the CPU it's executing on
4704 * is removed from the allowed bitmask.
4705 *
4706 * NOTE: the caller must have a valid reference to the task, the
4707 * task must not exit() & deallocate itself prematurely. The
4708 * call is not atomic; no spinlocks may be held.
4709 */
4710 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4711 {
4712 unsigned long flags;
4713 struct rq *rq;
4714 unsigned int dest_cpu;
4715 int ret = 0;
4717 rq = task_rq_lock(p, &flags);
4719 if (cpumask_equal(&p->cpus_allowed, new_mask))
4720 goto out;
4722 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4723 ret = -EINVAL;
4724 goto out;
4725 }
4727 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4728 ret = -EINVAL;
4729 goto out;
4730 }
4732 do_set_cpus_allowed(p, new_mask);
4734 /* Can the task run on the task's current CPU? If so, we're done */
4735 if (cpumask_test_cpu(task_cpu(p), new_mask))
4736 goto out;
4738 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4739 if (p->on_rq) {
4740 struct migration_arg arg = { p, dest_cpu };
4741 /* Need help from migration thread: drop lock and wait. */
4742 task_rq_unlock(rq, p, &flags);
4743 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4744 tlb_migrate_finish(p->mm);
4745 return 0;
4746 }
4747 out:
4748 task_rq_unlock(rq, p, &flags);
4750 return ret;
4751 }
4752 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4754 /*
4755 * Move (not current) task off this cpu, onto dest cpu. We're doing
4756 * this because either it can't run here any more (set_cpus_allowed()
4757 * away from this CPU, or CPU going down), or because we're
4758 * attempting to rebalance this task on exec (sched_exec).
4759 *
4760 * So we race with normal scheduler movements, but that's OK, as long
4761 * as the task is no longer on this CPU.
4762 *
4763 * Returns non-zero if task was successfully migrated.
4764 */
4765 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4766 {
4767 struct rq *rq_dest, *rq_src;
4768 int ret = 0;
4770 if (unlikely(!cpu_active(dest_cpu)))
4771 return ret;
4773 rq_src = cpu_rq(src_cpu);
4774 rq_dest = cpu_rq(dest_cpu);
4776 raw_spin_lock(&p->pi_lock);
4777 double_rq_lock(rq_src, rq_dest);
4778 /* Already moved. */
4779 if (task_cpu(p) != src_cpu)
4780 goto done;
4781 /* Affinity changed (again). */
4782 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4783 goto fail;
4785 /*
4786 * If we're not on a rq, the next wake-up will ensure we're
4787 * placed properly.
4788 */
4789 if (p->on_rq) {
4790 dequeue_task(rq_src, p, 0);
4791 set_task_cpu(p, dest_cpu);
4792 enqueue_task(rq_dest, p, 0);
4793 check_preempt_curr(rq_dest, p, 0);
4794 }
4795 done:
4796 ret = 1;
4797 fail:
4798 double_rq_unlock(rq_src, rq_dest);
4799 raw_spin_unlock(&p->pi_lock);
4800 return ret;
4801 }
4803 /*
4804 * migration_cpu_stop - this will be executed by a highprio stopper thread
4805 * and performs thread migration by bumping thread off CPU then
4806 * 'pushing' onto another runqueue.
4807 */
4808 static int migration_cpu_stop(void *data)
4809 {
4810 struct migration_arg *arg = data;
4812 /*
4813 * The original target cpu might have gone down and we might
4814 * be on another cpu but it doesn't matter.
4815 */
4816 local_irq_disable();
4817 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4818 local_irq_enable();
4819 return 0;
4820 }
4822 #ifdef CONFIG_HOTPLUG_CPU
4824 /*
4825 * Ensures that the idle task is using init_mm right before its cpu goes
4826 * offline.
4827 */
4828 void idle_task_exit(void)
4829 {
4830 struct mm_struct *mm = current->active_mm;
4832 BUG_ON(cpu_online(smp_processor_id()));
4834 if (mm != &init_mm)
4835 switch_mm(mm, &init_mm, current);
4836 mmdrop(mm);
4837 }
4839 /*
4840 * Since this CPU is going 'away' for a while, fold any nr_active delta
4841 * we might have. Assumes we're called after migrate_tasks() so that the
4842 * nr_active count is stable.
4843 *
4844 * Also see the comment "Global load-average calculations".
4845 */
4846 static void calc_load_migrate(struct rq *rq)
4847 {
4848 long delta = calc_load_fold_active(rq);
4849 if (delta)
4850 atomic_long_add(delta, &calc_load_tasks);
4851 }
4853 /*
4854 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4855 * try_to_wake_up()->select_task_rq().
4856 *
4857 * Called with rq->lock held even though we'er in stop_machine() and
4858 * there's no concurrency possible, we hold the required locks anyway
4859 * because of lock validation efforts.
4860 */
4861 static void migrate_tasks(unsigned int dead_cpu)
4862 {
4863 struct rq *rq = cpu_rq(dead_cpu);
4864 struct task_struct *next, *stop = rq->stop;
4865 int dest_cpu;
4867 /*
4868 * Fudge the rq selection such that the below task selection loop
4869 * doesn't get stuck on the currently eligible stop task.
4870 *
4871 * We're currently inside stop_machine() and the rq is either stuck
4872 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4873 * either way we should never end up calling schedule() until we're
4874 * done here.
4875 */
4876 rq->stop = NULL;
4878 for ( ; ; ) {
4879 /*
4880 * There's this thread running, bail when that's the only
4881 * remaining thread.
4882 */
4883 if (rq->nr_running == 1)
4884 break;
4886 next = pick_next_task(rq);
4887 BUG_ON(!next);
4888 next->sched_class->put_prev_task(rq, next);
4890 /* Find suitable destination for @next, with force if needed. */
4891 dest_cpu = select_fallback_rq(dead_cpu, next);
4892 raw_spin_unlock(&rq->lock);
4894 __migrate_task(next, dead_cpu, dest_cpu);
4896 raw_spin_lock(&rq->lock);
4897 }
4899 rq->stop = stop;
4900 }
4902 #endif /* CONFIG_HOTPLUG_CPU */
4904 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4906 static struct ctl_table sd_ctl_dir[] = {
4907 {
4908 .procname = "sched_domain",
4909 .mode = 0555,
4910 },
4911 {}
4912 };
4914 static struct ctl_table sd_ctl_root[] = {
4915 {
4916 .procname = "kernel",
4917 .mode = 0555,
4918 .child = sd_ctl_dir,
4919 },
4920 {}
4921 };
4923 static struct ctl_table *sd_alloc_ctl_entry(int n)
4924 {
4925 struct ctl_table *entry =
4926 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4928 return entry;
4929 }
4931 static void sd_free_ctl_entry(struct ctl_table **tablep)
4932 {
4933 struct ctl_table *entry;
4935 /*
4936 * In the intermediate directories, both the child directory and
4937 * procname are dynamically allocated and could fail but the mode
4938 * will always be set. In the lowest directory the names are
4939 * static strings and all have proc handlers.
4940 */
4941 for (entry = *tablep; entry->mode; entry++) {
4942 if (entry->child)
4943 sd_free_ctl_entry(&entry->child);
4944 if (entry->proc_handler == NULL)
4945 kfree(entry->procname);
4946 }
4948 kfree(*tablep);
4949 *tablep = NULL;
4950 }
4952 static int min_load_idx = 0;
4953 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4955 static void
4956 set_table_entry(struct ctl_table *entry,
4957 const char *procname, void *data, int maxlen,
4958 umode_t mode, proc_handler *proc_handler,
4959 bool load_idx)
4960 {
4961 entry->procname = procname;
4962 entry->data = data;
4963 entry->maxlen = maxlen;
4964 entry->mode = mode;
4965 entry->proc_handler = proc_handler;
4967 if (load_idx) {
4968 entry->extra1 = &min_load_idx;
4969 entry->extra2 = &max_load_idx;
4970 }
4971 }
4973 static struct ctl_table *
4974 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4975 {
4976 struct ctl_table *table = sd_alloc_ctl_entry(13);
4978 if (table == NULL)
4979 return NULL;
4981 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4982 sizeof(long), 0644, proc_doulongvec_minmax, false);
4983 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4984 sizeof(long), 0644, proc_doulongvec_minmax, false);
4985 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4986 sizeof(int), 0644, proc_dointvec_minmax, true);
4987 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4988 sizeof(int), 0644, proc_dointvec_minmax, true);
4989 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4990 sizeof(int), 0644, proc_dointvec_minmax, true);
4991 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4992 sizeof(int), 0644, proc_dointvec_minmax, true);
4993 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4994 sizeof(int), 0644, proc_dointvec_minmax, true);
4995 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4996 sizeof(int), 0644, proc_dointvec_minmax, false);
4997 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4998 sizeof(int), 0644, proc_dointvec_minmax, false);
4999 set_table_entry(&table[9], "cache_nice_tries",
5000 &sd->cache_nice_tries,
5001 sizeof(int), 0644, proc_dointvec_minmax, false);
5002 set_table_entry(&table[10], "flags", &sd->flags,
5003 sizeof(int), 0644, proc_dointvec_minmax, false);
5004 set_table_entry(&table[11], "name", sd->name,
5005 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5006 /* &table[12] is terminator */
5008 return table;
5009 }
5011 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5012 {
5013 struct ctl_table *entry, *table;
5014 struct sched_domain *sd;
5015 int domain_num = 0, i;
5016 char buf[32];
5018 for_each_domain(cpu, sd)
5019 domain_num++;
5020 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5021 if (table == NULL)
5022 return NULL;
5024 i = 0;
5025 for_each_domain(cpu, sd) {
5026 snprintf(buf, 32, "domain%d", i);
5027 entry->procname = kstrdup(buf, GFP_KERNEL);
5028 entry->mode = 0555;
5029 entry->child = sd_alloc_ctl_domain_table(sd);
5030 entry++;
5031 i++;
5032 }
5033 return table;
5034 }
5036 static struct ctl_table_header *sd_sysctl_header;
5037 static void register_sched_domain_sysctl(void)
5038 {
5039 int i, cpu_num = num_possible_cpus();
5040 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5041 char buf[32];
5043 WARN_ON(sd_ctl_dir[0].child);
5044 sd_ctl_dir[0].child = entry;
5046 if (entry == NULL)
5047 return;
5049 for_each_possible_cpu(i) {
5050 snprintf(buf, 32, "cpu%d", i);
5051 entry->procname = kstrdup(buf, GFP_KERNEL);
5052 entry->mode = 0555;
5053 entry->child = sd_alloc_ctl_cpu_table(i);
5054 entry++;
5055 }
5057 WARN_ON(sd_sysctl_header);
5058 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5059 }
5061 /* may be called multiple times per register */
5062 static void unregister_sched_domain_sysctl(void)
5063 {
5064 if (sd_sysctl_header)
5065 unregister_sysctl_table(sd_sysctl_header);
5066 sd_sysctl_header = NULL;
5067 if (sd_ctl_dir[0].child)
5068 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5069 }
5070 #else
5071 static void register_sched_domain_sysctl(void)
5072 {
5073 }
5074 static void unregister_sched_domain_sysctl(void)
5075 {
5076 }
5077 #endif
5079 static void set_rq_online(struct rq *rq)
5080 {
5081 if (!rq->online) {
5082 const struct sched_class *class;
5084 cpumask_set_cpu(rq->cpu, rq->rd->online);
5085 rq->online = 1;
5087 for_each_class(class) {
5088 if (class->rq_online)
5089 class->rq_online(rq);
5090 }
5091 }
5092 }
5094 static void set_rq_offline(struct rq *rq)
5095 {
5096 if (rq->online) {
5097 const struct sched_class *class;
5099 for_each_class(class) {
5100 if (class->rq_offline)
5101 class->rq_offline(rq);
5102 }
5104 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5105 rq->online = 0;
5106 }
5107 }
5109 /*
5110 * migration_call - callback that gets triggered when a CPU is added.
5111 * Here we can start up the necessary migration thread for the new CPU.
5112 */
5113 static int __cpuinit
5114 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5115 {
5116 int cpu = (long)hcpu;
5117 unsigned long flags;
5118 struct rq *rq = cpu_rq(cpu);
5120 switch (action & ~CPU_TASKS_FROZEN) {
5122 case CPU_UP_PREPARE:
5123 rq->calc_load_update = calc_load_update;
5124 break;
5126 case CPU_ONLINE:
5127 /* Update our root-domain */
5128 raw_spin_lock_irqsave(&rq->lock, flags);
5129 if (rq->rd) {
5130 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5132 set_rq_online(rq);
5133 }
5134 raw_spin_unlock_irqrestore(&rq->lock, flags);
5135 break;
5137 #ifdef CONFIG_HOTPLUG_CPU
5138 case CPU_DYING:
5139 sched_ttwu_pending();
5140 /* Update our root-domain */
5141 raw_spin_lock_irqsave(&rq->lock, flags);
5142 if (rq->rd) {
5143 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5144 set_rq_offline(rq);
5145 }
5146 migrate_tasks(cpu);
5147 BUG_ON(rq->nr_running != 1); /* the migration thread */
5148 raw_spin_unlock_irqrestore(&rq->lock, flags);
5149 break;
5151 case CPU_DEAD:
5152 calc_load_migrate(rq);
5153 break;
5154 #endif
5155 }
5157 update_max_interval();
5159 return NOTIFY_OK;
5160 }
5162 /*
5163 * Register at high priority so that task migration (migrate_all_tasks)
5164 * happens before everything else. This has to be lower priority than
5165 * the notifier in the perf_event subsystem, though.
5166 */
5167 static struct notifier_block __cpuinitdata migration_notifier = {
5168 .notifier_call = migration_call,
5169 .priority = CPU_PRI_MIGRATION,
5170 };
5172 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5173 unsigned long action, void *hcpu)
5174 {
5175 switch (action & ~CPU_TASKS_FROZEN) {
5176 case CPU_STARTING:
5177 case CPU_DOWN_FAILED:
5178 set_cpu_active((long)hcpu, true);
5179 return NOTIFY_OK;
5180 default:
5181 return NOTIFY_DONE;
5182 }
5183 }
5185 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5186 unsigned long action, void *hcpu)
5187 {
5188 switch (action & ~CPU_TASKS_FROZEN) {
5189 case CPU_DOWN_PREPARE:
5190 set_cpu_active((long)hcpu, false);
5191 return NOTIFY_OK;
5192 default:
5193 return NOTIFY_DONE;
5194 }
5195 }
5197 static int __init migration_init(void)
5198 {
5199 void *cpu = (void *)(long)smp_processor_id();
5200 int err;
5202 /* Initialize migration for the boot CPU */
5203 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5204 BUG_ON(err == NOTIFY_BAD);
5205 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5206 register_cpu_notifier(&migration_notifier);
5208 /* Register cpu active notifiers */
5209 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5210 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5212 return 0;
5213 }
5214 early_initcall(migration_init);
5215 #endif
5217 #ifdef CONFIG_SMP
5219 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5221 #ifdef CONFIG_SCHED_DEBUG
5223 static __read_mostly int sched_debug_enabled;
5225 static int __init sched_debug_setup(char *str)
5226 {
5227 sched_debug_enabled = 1;
5229 return 0;
5230 }
5231 early_param("sched_debug", sched_debug_setup);
5233 static inline bool sched_debug(void)
5234 {
5235 return sched_debug_enabled;
5236 }
5238 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5239 struct cpumask *groupmask)
5240 {
5241 struct sched_group *group = sd->groups;
5242 char str[256];
5244 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5245 cpumask_clear(groupmask);
5247 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5249 if (!(sd->flags & SD_LOAD_BALANCE)) {
5250 printk("does not load-balance\n");
5251 if (sd->parent)
5252 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5253 " has parent");
5254 return -1;
5255 }
5257 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5259 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5260 printk(KERN_ERR "ERROR: domain->span does not contain "
5261 "CPU%d\n", cpu);
5262 }
5263 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5264 printk(KERN_ERR "ERROR: domain->groups does not contain"
5265 " CPU%d\n", cpu);
5266 }
5268 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5269 do {
5270 if (!group) {
5271 printk("\n");
5272 printk(KERN_ERR "ERROR: group is NULL\n");
5273 break;
5274 }
5276 /*
5277 * Even though we initialize ->power to something semi-sane,
5278 * we leave power_orig unset. This allows us to detect if
5279 * domain iteration is still funny without causing /0 traps.
5280 */
5281 if (!group->sgp->power_orig) {
5282 printk(KERN_CONT "\n");
5283 printk(KERN_ERR "ERROR: domain->cpu_power not "
5284 "set\n");
5285 break;
5286 }
5288 if (!cpumask_weight(sched_group_cpus(group))) {
5289 printk(KERN_CONT "\n");
5290 printk(KERN_ERR "ERROR: empty group\n");
5291 break;
5292 }
5294 if (!(sd->flags & SD_OVERLAP) &&
5295 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5296 printk(KERN_CONT "\n");
5297 printk(KERN_ERR "ERROR: repeated CPUs\n");
5298 break;
5299 }
5301 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5303 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5305 printk(KERN_CONT " %s", str);
5306 if (group->sgp->power != SCHED_POWER_SCALE) {
5307 printk(KERN_CONT " (cpu_power = %d)",
5308 group->sgp->power);
5309 }
5311 group = group->next;
5312 } while (group != sd->groups);
5313 printk(KERN_CONT "\n");
5315 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5316 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5318 if (sd->parent &&
5319 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5320 printk(KERN_ERR "ERROR: parent span is not a superset "
5321 "of domain->span\n");
5322 return 0;
5323 }
5325 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5326 {
5327 int level = 0;
5329 if (!sched_debug_enabled)
5330 return;
5332 if (!sd) {
5333 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5334 return;
5335 }
5337 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5339 for (;;) {
5340 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5341 break;
5342 level++;
5343 sd = sd->parent;
5344 if (!sd)
5345 break;
5346 }
5347 }
5348 #else /* !CONFIG_SCHED_DEBUG */
5349 # define sched_domain_debug(sd, cpu) do { } while (0)
5350 static inline bool sched_debug(void)
5351 {
5352 return false;
5353 }
5354 #endif /* CONFIG_SCHED_DEBUG */
5356 static int sd_degenerate(struct sched_domain *sd)
5357 {
5358 if (cpumask_weight(sched_domain_span(sd)) == 1)
5359 return 1;
5361 /* Following flags need at least 2 groups */
5362 if (sd->flags & (SD_LOAD_BALANCE |
5363 SD_BALANCE_NEWIDLE |
5364 SD_BALANCE_FORK |
5365 SD_BALANCE_EXEC |
5366 SD_SHARE_CPUPOWER |
5367 SD_SHARE_PKG_RESOURCES)) {
5368 if (sd->groups != sd->groups->next)
5369 return 0;
5370 }
5372 /* Following flags don't use groups */
5373 if (sd->flags & (SD_WAKE_AFFINE))
5374 return 0;
5376 return 1;
5377 }
5379 static int
5380 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5381 {
5382 unsigned long cflags = sd->flags, pflags = parent->flags;
5384 if (sd_degenerate(parent))
5385 return 1;
5387 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5388 return 0;
5390 /* Flags needing groups don't count if only 1 group in parent */
5391 if (parent->groups == parent->groups->next) {
5392 pflags &= ~(SD_LOAD_BALANCE |
5393 SD_BALANCE_NEWIDLE |
5394 SD_BALANCE_FORK |
5395 SD_BALANCE_EXEC |
5396 SD_SHARE_CPUPOWER |
5397 SD_SHARE_PKG_RESOURCES);
5398 if (nr_node_ids == 1)
5399 pflags &= ~SD_SERIALIZE;
5400 }
5401 if (~cflags & pflags)
5402 return 0;
5404 return 1;
5405 }
5407 static void free_rootdomain(struct rcu_head *rcu)
5408 {
5409 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5411 cpupri_cleanup(&rd->cpupri);
5412 free_cpumask_var(rd->rto_mask);
5413 free_cpumask_var(rd->online);
5414 free_cpumask_var(rd->span);
5415 kfree(rd);
5416 }
5418 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5419 {
5420 struct root_domain *old_rd = NULL;
5421 unsigned long flags;
5423 raw_spin_lock_irqsave(&rq->lock, flags);
5425 if (rq->rd) {
5426 old_rd = rq->rd;
5428 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5429 set_rq_offline(rq);
5431 cpumask_clear_cpu(rq->cpu, old_rd->span);
5433 /*
5434 * If we dont want to free the old_rt yet then
5435 * set old_rd to NULL to skip the freeing later
5436 * in this function:
5437 */
5438 if (!atomic_dec_and_test(&old_rd->refcount))
5439 old_rd = NULL;
5440 }
5442 atomic_inc(&rd->refcount);
5443 rq->rd = rd;
5445 cpumask_set_cpu(rq->cpu, rd->span);
5446 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5447 set_rq_online(rq);
5449 raw_spin_unlock_irqrestore(&rq->lock, flags);
5451 if (old_rd)
5452 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5453 }
5455 static int init_rootdomain(struct root_domain *rd)
5456 {
5457 memset(rd, 0, sizeof(*rd));
5459 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5460 goto out;
5461 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5462 goto free_span;
5463 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5464 goto free_online;
5466 if (cpupri_init(&rd->cpupri) != 0)
5467 goto free_rto_mask;
5468 return 0;
5470 free_rto_mask:
5471 free_cpumask_var(rd->rto_mask);
5472 free_online:
5473 free_cpumask_var(rd->online);
5474 free_span:
5475 free_cpumask_var(rd->span);
5476 out:
5477 return -ENOMEM;
5478 }
5480 /*
5481 * By default the system creates a single root-domain with all cpus as
5482 * members (mimicking the global state we have today).
5483 */
5484 struct root_domain def_root_domain;
5486 static void init_defrootdomain(void)
5487 {
5488 init_rootdomain(&def_root_domain);
5490 atomic_set(&def_root_domain.refcount, 1);
5491 }
5493 static struct root_domain *alloc_rootdomain(void)
5494 {
5495 struct root_domain *rd;
5497 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5498 if (!rd)
5499 return NULL;
5501 if (init_rootdomain(rd) != 0) {
5502 kfree(rd);
5503 return NULL;
5504 }
5506 return rd;
5507 }
5509 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5510 {
5511 struct sched_group *tmp, *first;
5513 if (!sg)
5514 return;
5516 first = sg;
5517 do {
5518 tmp = sg->next;
5520 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5521 kfree(sg->sgp);
5523 kfree(sg);
5524 sg = tmp;
5525 } while (sg != first);
5526 }
5528 static void free_sched_domain(struct rcu_head *rcu)
5529 {
5530 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5532 /*
5533 * If its an overlapping domain it has private groups, iterate and
5534 * nuke them all.
5535 */
5536 if (sd->flags & SD_OVERLAP) {
5537 free_sched_groups(sd->groups, 1);
5538 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5539 kfree(sd->groups->sgp);
5540 kfree(sd->groups);
5541 }
5542 kfree(sd);
5543 }
5545 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5546 {
5547 call_rcu(&sd->rcu, free_sched_domain);
5548 }
5550 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5551 {
5552 for (; sd; sd = sd->parent)
5553 destroy_sched_domain(sd, cpu);
5554 }
5556 /*
5557 * Keep a special pointer to the highest sched_domain that has
5558 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5559 * allows us to avoid some pointer chasing select_idle_sibling().
5560 *
5561 * Also keep a unique ID per domain (we use the first cpu number in
5562 * the cpumask of the domain), this allows us to quickly tell if
5563 * two cpus are in the same cache domain, see cpus_share_cache().
5564 */
5565 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5566 DEFINE_PER_CPU(int, sd_llc_id);
5568 static void update_top_cache_domain(int cpu)
5569 {
5570 struct sched_domain *sd;
5571 int id = cpu;
5573 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5574 if (sd)
5575 id = cpumask_first(sched_domain_span(sd));
5577 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5578 per_cpu(sd_llc_id, cpu) = id;
5579 }
5581 /*
5582 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5583 * hold the hotplug lock.
5584 */
5585 static void
5586 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5587 {
5588 struct rq *rq = cpu_rq(cpu);
5589 struct sched_domain *tmp;
5591 /* Remove the sched domains which do not contribute to scheduling. */
5592 for (tmp = sd; tmp; ) {
5593 struct sched_domain *parent = tmp->parent;
5594 if (!parent)
5595 break;
5597 if (sd_parent_degenerate(tmp, parent)) {
5598 tmp->parent = parent->parent;
5599 if (parent->parent)
5600 parent->parent->child = tmp;
5601 destroy_sched_domain(parent, cpu);
5602 } else
5603 tmp = tmp->parent;
5604 }
5606 if (sd && sd_degenerate(sd)) {
5607 tmp = sd;
5608 sd = sd->parent;
5609 destroy_sched_domain(tmp, cpu);
5610 if (sd)
5611 sd->child = NULL;
5612 }
5614 sched_domain_debug(sd, cpu);
5616 rq_attach_root(rq, rd);
5617 tmp = rq->sd;
5618 rcu_assign_pointer(rq->sd, sd);
5619 destroy_sched_domains(tmp, cpu);
5621 update_top_cache_domain(cpu);
5622 }
5624 /* cpus with isolated domains */
5625 static cpumask_var_t cpu_isolated_map;
5627 /* Setup the mask of cpus configured for isolated domains */
5628 static int __init isolated_cpu_setup(char *str)
5629 {
5630 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5631 cpulist_parse(str, cpu_isolated_map);
5632 return 1;
5633 }
5635 __setup("isolcpus=", isolated_cpu_setup);
5637 static const struct cpumask *cpu_cpu_mask(int cpu)
5638 {
5639 return cpumask_of_node(cpu_to_node(cpu));
5640 }
5642 struct sd_data {
5643 struct sched_domain **__percpu sd;
5644 struct sched_group **__percpu sg;
5645 struct sched_group_power **__percpu sgp;
5646 };
5648 struct s_data {
5649 struct sched_domain ** __percpu sd;
5650 struct root_domain *rd;
5651 };
5653 enum s_alloc {
5654 sa_rootdomain,
5655 sa_sd,
5656 sa_sd_storage,
5657 sa_none,
5658 };
5660 struct sched_domain_topology_level;
5662 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5663 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5665 #define SDTL_OVERLAP 0x01
5667 struct sched_domain_topology_level {
5668 sched_domain_init_f init;
5669 sched_domain_mask_f mask;
5670 int flags;
5671 int numa_level;
5672 struct sd_data data;
5673 };
5675 /*
5676 * Build an iteration mask that can exclude certain CPUs from the upwards
5677 * domain traversal.
5678 *
5679 * Asymmetric node setups can result in situations where the domain tree is of
5680 * unequal depth, make sure to skip domains that already cover the entire
5681 * range.
5682 *
5683 * In that case build_sched_domains() will have terminated the iteration early
5684 * and our sibling sd spans will be empty. Domains should always include the
5685 * cpu they're built on, so check that.
5686 *
5687 */
5688 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5689 {
5690 const struct cpumask *span = sched_domain_span(sd);
5691 struct sd_data *sdd = sd->private;
5692 struct sched_domain *sibling;
5693 int i;
5695 for_each_cpu(i, span) {
5696 sibling = *per_cpu_ptr(sdd->sd, i);
5697 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5698 continue;
5700 cpumask_set_cpu(i, sched_group_mask(sg));
5701 }
5702 }
5704 /*
5705 * Return the canonical balance cpu for this group, this is the first cpu
5706 * of this group that's also in the iteration mask.
5707 */
5708 int group_balance_cpu(struct sched_group *sg)
5709 {
5710 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5711 }
5713 static int
5714 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5715 {
5716 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5717 const struct cpumask *span = sched_domain_span(sd);
5718 struct cpumask *covered = sched_domains_tmpmask;
5719 struct sd_data *sdd = sd->private;
5720 struct sched_domain *child;
5721 int i;
5723 cpumask_clear(covered);
5725 for_each_cpu(i, span) {
5726 struct cpumask *sg_span;
5728 if (cpumask_test_cpu(i, covered))
5729 continue;
5731 child = *per_cpu_ptr(sdd->sd, i);
5733 /* See the comment near build_group_mask(). */
5734 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5735 continue;
5737 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5738 GFP_KERNEL, cpu_to_node(cpu));
5740 if (!sg)
5741 goto fail;
5743 sg_span = sched_group_cpus(sg);
5744 if (child->child) {
5745 child = child->child;
5746 cpumask_copy(sg_span, sched_domain_span(child));
5747 } else
5748 cpumask_set_cpu(i, sg_span);
5750 cpumask_or(covered, covered, sg_span);
5752 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5753 if (atomic_inc_return(&sg->sgp->ref) == 1)
5754 build_group_mask(sd, sg);
5756 /*
5757 * Initialize sgp->power such that even if we mess up the
5758 * domains and no possible iteration will get us here, we won't
5759 * die on a /0 trap.
5760 */
5761 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5763 /*
5764 * Make sure the first group of this domain contains the
5765 * canonical balance cpu. Otherwise the sched_domain iteration
5766 * breaks. See update_sg_lb_stats().
5767 */
5768 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5769 group_balance_cpu(sg) == cpu)
5770 groups = sg;
5772 if (!first)
5773 first = sg;
5774 if (last)
5775 last->next = sg;
5776 last = sg;
5777 last->next = first;
5778 }
5779 sd->groups = groups;
5781 return 0;
5783 fail:
5784 free_sched_groups(first, 0);
5786 return -ENOMEM;
5787 }
5789 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5790 {
5791 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5792 struct sched_domain *child = sd->child;
5794 if (child)
5795 cpu = cpumask_first(sched_domain_span(child));
5797 if (sg) {
5798 *sg = *per_cpu_ptr(sdd->sg, cpu);
5799 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5800 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5801 }
5803 return cpu;
5804 }
5806 /*
5807 * build_sched_groups will build a circular linked list of the groups
5808 * covered by the given span, and will set each group's ->cpumask correctly,
5809 * and ->cpu_power to 0.
5810 *
5811 * Assumes the sched_domain tree is fully constructed
5812 */
5813 static int
5814 build_sched_groups(struct sched_domain *sd, int cpu)
5815 {
5816 struct sched_group *first = NULL, *last = NULL;
5817 struct sd_data *sdd = sd->private;
5818 const struct cpumask *span = sched_domain_span(sd);
5819 struct cpumask *covered;
5820 int i;
5822 get_group(cpu, sdd, &sd->groups);
5823 atomic_inc(&sd->groups->ref);
5825 if (cpu != cpumask_first(sched_domain_span(sd)))
5826 return 0;
5828 lockdep_assert_held(&sched_domains_mutex);
5829 covered = sched_domains_tmpmask;
5831 cpumask_clear(covered);
5833 for_each_cpu(i, span) {
5834 struct sched_group *sg;
5835 int group = get_group(i, sdd, &sg);
5836 int j;
5838 if (cpumask_test_cpu(i, covered))
5839 continue;
5841 cpumask_clear(sched_group_cpus(sg));
5842 sg->sgp->power = 0;
5843 cpumask_setall(sched_group_mask(sg));
5845 for_each_cpu(j, span) {
5846 if (get_group(j, sdd, NULL) != group)
5847 continue;
5849 cpumask_set_cpu(j, covered);
5850 cpumask_set_cpu(j, sched_group_cpus(sg));
5851 }
5853 if (!first)
5854 first = sg;
5855 if (last)
5856 last->next = sg;
5857 last = sg;
5858 }
5859 last->next = first;
5861 return 0;
5862 }
5864 /*
5865 * Initialize sched groups cpu_power.
5866 *
5867 * cpu_power indicates the capacity of sched group, which is used while
5868 * distributing the load between different sched groups in a sched domain.
5869 * Typically cpu_power for all the groups in a sched domain will be same unless
5870 * there are asymmetries in the topology. If there are asymmetries, group
5871 * having more cpu_power will pickup more load compared to the group having
5872 * less cpu_power.
5873 */
5874 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5875 {
5876 struct sched_group *sg = sd->groups;
5878 WARN_ON(!sd || !sg);
5880 do {
5881 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5882 sg = sg->next;
5883 } while (sg != sd->groups);
5885 if (cpu != group_balance_cpu(sg))
5886 return;
5888 update_group_power(sd, cpu);
5889 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5890 }
5892 int __weak arch_sd_sibling_asym_packing(void)
5893 {
5894 return 0*SD_ASYM_PACKING;
5895 }
5897 /*
5898 * Initializers for schedule domains
5899 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5900 */
5902 #ifdef CONFIG_SCHED_DEBUG
5903 # define SD_INIT_NAME(sd, type) sd->name = #type
5904 #else
5905 # define SD_INIT_NAME(sd, type) do { } while (0)
5906 #endif
5908 #define SD_INIT_FUNC(type) \
5909 static noinline struct sched_domain * \
5910 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5911 { \
5912 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5913 *sd = SD_##type##_INIT; \
5914 SD_INIT_NAME(sd, type); \
5915 sd->private = &tl->data; \
5916 return sd; \
5917 }
5919 SD_INIT_FUNC(CPU)
5920 #ifdef CONFIG_SCHED_SMT
5921 SD_INIT_FUNC(SIBLING)
5922 #endif
5923 #ifdef CONFIG_SCHED_MC
5924 SD_INIT_FUNC(MC)
5925 #endif
5926 #ifdef CONFIG_SCHED_BOOK
5927 SD_INIT_FUNC(BOOK)
5928 #endif
5930 static int default_relax_domain_level = -1;
5931 int sched_domain_level_max;
5933 static int __init setup_relax_domain_level(char *str)
5934 {
5935 if (kstrtoint(str, 0, &default_relax_domain_level))
5936 pr_warn("Unable to set relax_domain_level\n");
5938 return 1;
5939 }
5940 __setup("relax_domain_level=", setup_relax_domain_level);
5942 static void set_domain_attribute(struct sched_domain *sd,
5943 struct sched_domain_attr *attr)
5944 {
5945 int request;
5947 if (!attr || attr->relax_domain_level < 0) {
5948 if (default_relax_domain_level < 0)
5949 return;
5950 else
5951 request = default_relax_domain_level;
5952 } else
5953 request = attr->relax_domain_level;
5954 if (request < sd->level) {
5955 /* turn off idle balance on this domain */
5956 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5957 } else {
5958 /* turn on idle balance on this domain */
5959 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5960 }
5961 }
5963 static void __sdt_free(const struct cpumask *cpu_map);
5964 static int __sdt_alloc(const struct cpumask *cpu_map);
5966 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5967 const struct cpumask *cpu_map)
5968 {
5969 switch (what) {
5970 case sa_rootdomain:
5971 if (!atomic_read(&d->rd->refcount))
5972 free_rootdomain(&d->rd->rcu); /* fall through */
5973 case sa_sd:
5974 free_percpu(d->sd); /* fall through */
5975 case sa_sd_storage:
5976 __sdt_free(cpu_map); /* fall through */
5977 case sa_none:
5978 break;
5979 }
5980 }
5982 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5983 const struct cpumask *cpu_map)
5984 {
5985 memset(d, 0, sizeof(*d));
5987 if (__sdt_alloc(cpu_map))
5988 return sa_sd_storage;
5989 d->sd = alloc_percpu(struct sched_domain *);
5990 if (!d->sd)
5991 return sa_sd_storage;
5992 d->rd = alloc_rootdomain();
5993 if (!d->rd)
5994 return sa_sd;
5995 return sa_rootdomain;
5996 }
5998 /*
5999 * NULL the sd_data elements we've used to build the sched_domain and
6000 * sched_group structure so that the subsequent __free_domain_allocs()
6001 * will not free the data we're using.
6002 */
6003 static void claim_allocations(int cpu, struct sched_domain *sd)
6004 {
6005 struct sd_data *sdd = sd->private;
6007 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6008 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6010 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6011 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6013 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6014 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6015 }
6017 #ifdef CONFIG_SCHED_SMT
6018 static const struct cpumask *cpu_smt_mask(int cpu)
6019 {
6020 return topology_thread_cpumask(cpu);
6021 }
6022 #endif
6024 /*
6025 * Topology list, bottom-up.
6026 */
6027 static struct sched_domain_topology_level default_topology[] = {
6028 #ifdef CONFIG_SCHED_SMT
6029 { sd_init_SIBLING, cpu_smt_mask, },
6030 #endif
6031 #ifdef CONFIG_SCHED_MC
6032 { sd_init_MC, cpu_coregroup_mask, },
6033 #endif
6034 #ifdef CONFIG_SCHED_BOOK
6035 { sd_init_BOOK, cpu_book_mask, },
6036 #endif
6037 { sd_init_CPU, cpu_cpu_mask, },
6038 { NULL, },
6039 };
6041 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6043 #ifdef CONFIG_NUMA
6045 static int sched_domains_numa_levels;
6046 static int *sched_domains_numa_distance;
6047 static struct cpumask ***sched_domains_numa_masks;
6048 static int sched_domains_curr_level;
6050 static inline int sd_local_flags(int level)
6051 {
6052 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6053 return 0;
6055 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6056 }
6058 static struct sched_domain *
6059 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6060 {
6061 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6062 int level = tl->numa_level;
6063 int sd_weight = cpumask_weight(
6064 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6066 *sd = (struct sched_domain){
6067 .min_interval = sd_weight,
6068 .max_interval = 2*sd_weight,
6069 .busy_factor = 32,
6070 .imbalance_pct = 125,
6071 .cache_nice_tries = 2,
6072 .busy_idx = 3,
6073 .idle_idx = 2,
6074 .newidle_idx = 0,
6075 .wake_idx = 0,
6076 .forkexec_idx = 0,
6078 .flags = 1*SD_LOAD_BALANCE
6079 | 1*SD_BALANCE_NEWIDLE
6080 | 0*SD_BALANCE_EXEC
6081 | 0*SD_BALANCE_FORK
6082 | 0*SD_BALANCE_WAKE
6083 | 0*SD_WAKE_AFFINE
6084 | 0*SD_SHARE_CPUPOWER
6085 | 0*SD_SHARE_PKG_RESOURCES
6086 | 1*SD_SERIALIZE
6087 | 0*SD_PREFER_SIBLING
6088 | sd_local_flags(level)
6089 ,
6090 .last_balance = jiffies,
6091 .balance_interval = sd_weight,
6092 };
6093 SD_INIT_NAME(sd, NUMA);
6094 sd->private = &tl->data;
6096 /*
6097 * Ugly hack to pass state to sd_numa_mask()...
6098 */
6099 sched_domains_curr_level = tl->numa_level;
6101 return sd;
6102 }
6104 static const struct cpumask *sd_numa_mask(int cpu)
6105 {
6106 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6107 }
6109 static void sched_numa_warn(const char *str)
6110 {
6111 static int done = false;
6112 int i,j;
6114 if (done)
6115 return;
6117 done = true;
6119 printk(KERN_WARNING "ERROR: %s\n\n", str);
6121 for (i = 0; i < nr_node_ids; i++) {
6122 printk(KERN_WARNING " ");
6123 for (j = 0; j < nr_node_ids; j++)
6124 printk(KERN_CONT "%02d ", node_distance(i,j));
6125 printk(KERN_CONT "\n");
6126 }
6127 printk(KERN_WARNING "\n");
6128 }
6130 static bool find_numa_distance(int distance)
6131 {
6132 int i;
6134 if (distance == node_distance(0, 0))
6135 return true;
6137 for (i = 0; i < sched_domains_numa_levels; i++) {
6138 if (sched_domains_numa_distance[i] == distance)
6139 return true;
6140 }
6142 return false;
6143 }
6145 static void sched_init_numa(void)
6146 {
6147 int next_distance, curr_distance = node_distance(0, 0);
6148 struct sched_domain_topology_level *tl;
6149 int level = 0;
6150 int i, j, k;
6152 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6153 if (!sched_domains_numa_distance)
6154 return;
6156 /*
6157 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6158 * unique distances in the node_distance() table.
6159 *
6160 * Assumes node_distance(0,j) includes all distances in
6161 * node_distance(i,j) in order to avoid cubic time.
6162 */
6163 next_distance = curr_distance;
6164 for (i = 0; i < nr_node_ids; i++) {
6165 for (j = 0; j < nr_node_ids; j++) {
6166 for (k = 0; k < nr_node_ids; k++) {
6167 int distance = node_distance(i, k);
6169 if (distance > curr_distance &&
6170 (distance < next_distance ||
6171 next_distance == curr_distance))
6172 next_distance = distance;
6174 /*
6175 * While not a strong assumption it would be nice to know
6176 * about cases where if node A is connected to B, B is not
6177 * equally connected to A.
6178 */
6179 if (sched_debug() && node_distance(k, i) != distance)
6180 sched_numa_warn("Node-distance not symmetric");
6182 if (sched_debug() && i && !find_numa_distance(distance))
6183 sched_numa_warn("Node-0 not representative");
6184 }
6185 if (next_distance != curr_distance) {
6186 sched_domains_numa_distance[level++] = next_distance;
6187 sched_domains_numa_levels = level;
6188 curr_distance = next_distance;
6189 } else break;
6190 }
6192 /*
6193 * In case of sched_debug() we verify the above assumption.
6194 */
6195 if (!sched_debug())
6196 break;
6197 }
6198 /*
6199 * 'level' contains the number of unique distances, excluding the
6200 * identity distance node_distance(i,i).
6201 *
6202 * The sched_domains_nume_distance[] array includes the actual distance
6203 * numbers.
6204 */
6206 /*
6207 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6208 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6209 * the array will contain less then 'level' members. This could be
6210 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6211 * in other functions.
6212 *
6213 * We reset it to 'level' at the end of this function.
6214 */
6215 sched_domains_numa_levels = 0;
6217 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6218 if (!sched_domains_numa_masks)
6219 return;
6221 /*
6222 * Now for each level, construct a mask per node which contains all
6223 * cpus of nodes that are that many hops away from us.
6224 */
6225 for (i = 0; i < level; i++) {
6226 sched_domains_numa_masks[i] =
6227 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6228 if (!sched_domains_numa_masks[i])
6229 return;
6231 for (j = 0; j < nr_node_ids; j++) {
6232 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6233 if (!mask)
6234 return;
6236 sched_domains_numa_masks[i][j] = mask;
6238 for (k = 0; k < nr_node_ids; k++) {
6239 if (node_distance(j, k) > sched_domains_numa_distance[i])
6240 continue;
6242 cpumask_or(mask, mask, cpumask_of_node(k));
6243 }
6244 }
6245 }
6247 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6248 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6249 if (!tl)
6250 return;
6252 /*
6253 * Copy the default topology bits..
6254 */
6255 for (i = 0; default_topology[i].init; i++)
6256 tl[i] = default_topology[i];
6258 /*
6259 * .. and append 'j' levels of NUMA goodness.
6260 */
6261 for (j = 0; j < level; i++, j++) {
6262 tl[i] = (struct sched_domain_topology_level){
6263 .init = sd_numa_init,
6264 .mask = sd_numa_mask,
6265 .flags = SDTL_OVERLAP,
6266 .numa_level = j,
6267 };
6268 }
6270 sched_domain_topology = tl;
6272 sched_domains_numa_levels = level;
6273 }
6275 static void sched_domains_numa_masks_set(int cpu)
6276 {
6277 int i, j;
6278 int node = cpu_to_node(cpu);
6280 for (i = 0; i < sched_domains_numa_levels; i++) {
6281 for (j = 0; j < nr_node_ids; j++) {
6282 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6283 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6284 }
6285 }
6286 }
6288 static void sched_domains_numa_masks_clear(int cpu)
6289 {
6290 int i, j;
6291 for (i = 0; i < sched_domains_numa_levels; i++) {
6292 for (j = 0; j < nr_node_ids; j++)
6293 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6294 }
6295 }
6297 /*
6298 * Update sched_domains_numa_masks[level][node] array when new cpus
6299 * are onlined.
6300 */
6301 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6302 unsigned long action,
6303 void *hcpu)
6304 {
6305 int cpu = (long)hcpu;
6307 switch (action & ~CPU_TASKS_FROZEN) {
6308 case CPU_ONLINE:
6309 sched_domains_numa_masks_set(cpu);
6310 break;
6312 case CPU_DEAD:
6313 sched_domains_numa_masks_clear(cpu);
6314 break;
6316 default:
6317 return NOTIFY_DONE;
6318 }
6320 return NOTIFY_OK;
6321 }
6322 #else
6323 static inline void sched_init_numa(void)
6324 {
6325 }
6327 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6328 unsigned long action,
6329 void *hcpu)
6330 {
6331 return 0;
6332 }
6333 #endif /* CONFIG_NUMA */
6335 static int __sdt_alloc(const struct cpumask *cpu_map)
6336 {
6337 struct sched_domain_topology_level *tl;
6338 int j;
6340 for (tl = sched_domain_topology; tl->init; tl++) {
6341 struct sd_data *sdd = &tl->data;
6343 sdd->sd = alloc_percpu(struct sched_domain *);
6344 if (!sdd->sd)
6345 return -ENOMEM;
6347 sdd->sg = alloc_percpu(struct sched_group *);
6348 if (!sdd->sg)
6349 return -ENOMEM;
6351 sdd->sgp = alloc_percpu(struct sched_group_power *);
6352 if (!sdd->sgp)
6353 return -ENOMEM;
6355 for_each_cpu(j, cpu_map) {
6356 struct sched_domain *sd;
6357 struct sched_group *sg;
6358 struct sched_group_power *sgp;
6360 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6361 GFP_KERNEL, cpu_to_node(j));
6362 if (!sd)
6363 return -ENOMEM;
6365 *per_cpu_ptr(sdd->sd, j) = sd;
6367 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6368 GFP_KERNEL, cpu_to_node(j));
6369 if (!sg)
6370 return -ENOMEM;
6372 sg->next = sg;
6374 *per_cpu_ptr(sdd->sg, j) = sg;
6376 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6377 GFP_KERNEL, cpu_to_node(j));
6378 if (!sgp)
6379 return -ENOMEM;
6381 *per_cpu_ptr(sdd->sgp, j) = sgp;
6382 }
6383 }
6385 return 0;
6386 }
6388 static void __sdt_free(const struct cpumask *cpu_map)
6389 {
6390 struct sched_domain_topology_level *tl;
6391 int j;
6393 for (tl = sched_domain_topology; tl->init; tl++) {
6394 struct sd_data *sdd = &tl->data;
6396 for_each_cpu(j, cpu_map) {
6397 struct sched_domain *sd;
6399 if (sdd->sd) {
6400 sd = *per_cpu_ptr(sdd->sd, j);
6401 if (sd && (sd->flags & SD_OVERLAP))
6402 free_sched_groups(sd->groups, 0);
6403 kfree(*per_cpu_ptr(sdd->sd, j));
6404 }
6406 if (sdd->sg)
6407 kfree(*per_cpu_ptr(sdd->sg, j));
6408 if (sdd->sgp)
6409 kfree(*per_cpu_ptr(sdd->sgp, j));
6410 }
6411 free_percpu(sdd->sd);
6412 sdd->sd = NULL;
6413 free_percpu(sdd->sg);
6414 sdd->sg = NULL;
6415 free_percpu(sdd->sgp);
6416 sdd->sgp = NULL;
6417 }
6418 }
6420 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6421 struct s_data *d, const struct cpumask *cpu_map,
6422 struct sched_domain_attr *attr, struct sched_domain *child,
6423 int cpu)
6424 {
6425 struct sched_domain *sd = tl->init(tl, cpu);
6426 if (!sd)
6427 return child;
6429 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6430 if (child) {
6431 sd->level = child->level + 1;
6432 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6433 child->parent = sd;
6434 }
6435 sd->child = child;
6436 set_domain_attribute(sd, attr);
6438 return sd;
6439 }
6441 /*
6442 * Build sched domains for a given set of cpus and attach the sched domains
6443 * to the individual cpus
6444 */
6445 static int build_sched_domains(const struct cpumask *cpu_map,
6446 struct sched_domain_attr *attr)
6447 {
6448 enum s_alloc alloc_state = sa_none;
6449 struct sched_domain *sd;
6450 struct s_data d;
6451 int i, ret = -ENOMEM;
6453 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6454 if (alloc_state != sa_rootdomain)
6455 goto error;
6457 /* Set up domains for cpus specified by the cpu_map. */
6458 for_each_cpu(i, cpu_map) {
6459 struct sched_domain_topology_level *tl;
6461 sd = NULL;
6462 for (tl = sched_domain_topology; tl->init; tl++) {
6463 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6464 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6465 sd->flags |= SD_OVERLAP;
6466 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6467 break;
6468 }
6470 while (sd->child)
6471 sd = sd->child;
6473 *per_cpu_ptr(d.sd, i) = sd;
6474 }
6476 /* Build the groups for the domains */
6477 for_each_cpu(i, cpu_map) {
6478 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6479 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6480 if (sd->flags & SD_OVERLAP) {
6481 if (build_overlap_sched_groups(sd, i))
6482 goto error;
6483 } else {
6484 if (build_sched_groups(sd, i))
6485 goto error;
6486 }
6487 }
6488 }
6490 /* Calculate CPU power for physical packages and nodes */
6491 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6492 if (!cpumask_test_cpu(i, cpu_map))
6493 continue;
6495 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6496 claim_allocations(i, sd);
6497 init_sched_groups_power(i, sd);
6498 }
6499 }
6501 /* Attach the domains */
6502 rcu_read_lock();
6503 for_each_cpu(i, cpu_map) {
6504 sd = *per_cpu_ptr(d.sd, i);
6505 cpu_attach_domain(sd, d.rd, i);
6506 }
6507 rcu_read_unlock();
6509 ret = 0;
6510 error:
6511 __free_domain_allocs(&d, alloc_state, cpu_map);
6512 return ret;
6513 }
6515 static cpumask_var_t *doms_cur; /* current sched domains */
6516 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6517 static struct sched_domain_attr *dattr_cur;
6518 /* attribues of custom domains in 'doms_cur' */
6520 /*
6521 * Special case: If a kmalloc of a doms_cur partition (array of
6522 * cpumask) fails, then fallback to a single sched domain,
6523 * as determined by the single cpumask fallback_doms.
6524 */
6525 static cpumask_var_t fallback_doms;
6527 /*
6528 * arch_update_cpu_topology lets virtualized architectures update the
6529 * cpu core maps. It is supposed to return 1 if the topology changed
6530 * or 0 if it stayed the same.
6531 */
6532 int __attribute__((weak)) arch_update_cpu_topology(void)
6533 {
6534 return 0;
6535 }
6537 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6538 {
6539 int i;
6540 cpumask_var_t *doms;
6542 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6543 if (!doms)
6544 return NULL;
6545 for (i = 0; i < ndoms; i++) {
6546 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6547 free_sched_domains(doms, i);
6548 return NULL;
6549 }
6550 }
6551 return doms;
6552 }
6554 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6555 {
6556 unsigned int i;
6557 for (i = 0; i < ndoms; i++)
6558 free_cpumask_var(doms[i]);
6559 kfree(doms);
6560 }
6562 /*
6563 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6564 * For now this just excludes isolated cpus, but could be used to
6565 * exclude other special cases in the future.
6566 */
6567 static int init_sched_domains(const struct cpumask *cpu_map)
6568 {
6569 int err;
6571 arch_update_cpu_topology();
6572 ndoms_cur = 1;
6573 doms_cur = alloc_sched_domains(ndoms_cur);
6574 if (!doms_cur)
6575 doms_cur = &fallback_doms;
6576 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6577 err = build_sched_domains(doms_cur[0], NULL);
6578 register_sched_domain_sysctl();
6580 return err;
6581 }
6583 /*
6584 * Detach sched domains from a group of cpus specified in cpu_map
6585 * These cpus will now be attached to the NULL domain
6586 */
6587 static void detach_destroy_domains(const struct cpumask *cpu_map)
6588 {
6589 int i;
6591 rcu_read_lock();
6592 for_each_cpu(i, cpu_map)
6593 cpu_attach_domain(NULL, &def_root_domain, i);
6594 rcu_read_unlock();
6595 }
6597 /* handle null as "default" */
6598 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6599 struct sched_domain_attr *new, int idx_new)
6600 {
6601 struct sched_domain_attr tmp;
6603 /* fast path */
6604 if (!new && !cur)
6605 return 1;
6607 tmp = SD_ATTR_INIT;
6608 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6609 new ? (new + idx_new) : &tmp,
6610 sizeof(struct sched_domain_attr));
6611 }
6613 /*
6614 * Partition sched domains as specified by the 'ndoms_new'
6615 * cpumasks in the array doms_new[] of cpumasks. This compares
6616 * doms_new[] to the current sched domain partitioning, doms_cur[].
6617 * It destroys each deleted domain and builds each new domain.
6618 *
6619 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6620 * The masks don't intersect (don't overlap.) We should setup one
6621 * sched domain for each mask. CPUs not in any of the cpumasks will
6622 * not be load balanced. If the same cpumask appears both in the
6623 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6624 * it as it is.
6625 *
6626 * The passed in 'doms_new' should be allocated using
6627 * alloc_sched_domains. This routine takes ownership of it and will
6628 * free_sched_domains it when done with it. If the caller failed the
6629 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6630 * and partition_sched_domains() will fallback to the single partition
6631 * 'fallback_doms', it also forces the domains to be rebuilt.
6632 *
6633 * If doms_new == NULL it will be replaced with cpu_online_mask.
6634 * ndoms_new == 0 is a special case for destroying existing domains,
6635 * and it will not create the default domain.
6636 *
6637 * Call with hotplug lock held
6638 */
6639 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6640 struct sched_domain_attr *dattr_new)
6641 {
6642 int i, j, n;
6643 int new_topology;
6645 mutex_lock(&sched_domains_mutex);
6647 /* always unregister in case we don't destroy any domains */
6648 unregister_sched_domain_sysctl();
6650 /* Let architecture update cpu core mappings. */
6651 new_topology = arch_update_cpu_topology();
6653 n = doms_new ? ndoms_new : 0;
6655 /* Destroy deleted domains */
6656 for (i = 0; i < ndoms_cur; i++) {
6657 for (j = 0; j < n && !new_topology; j++) {
6658 if (cpumask_equal(doms_cur[i], doms_new[j])
6659 && dattrs_equal(dattr_cur, i, dattr_new, j))
6660 goto match1;
6661 }
6662 /* no match - a current sched domain not in new doms_new[] */
6663 detach_destroy_domains(doms_cur[i]);
6664 match1:
6665 ;
6666 }
6668 if (doms_new == NULL) {
6669 ndoms_cur = 0;
6670 doms_new = &fallback_doms;
6671 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6672 WARN_ON_ONCE(dattr_new);
6673 }
6675 /* Build new domains */
6676 for (i = 0; i < ndoms_new; i++) {
6677 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6678 if (cpumask_equal(doms_new[i], doms_cur[j])
6679 && dattrs_equal(dattr_new, i, dattr_cur, j))
6680 goto match2;
6681 }
6682 /* no match - add a new doms_new */
6683 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6684 match2:
6685 ;
6686 }
6688 /* Remember the new sched domains */
6689 if (doms_cur != &fallback_doms)
6690 free_sched_domains(doms_cur, ndoms_cur);
6691 kfree(dattr_cur); /* kfree(NULL) is safe */
6692 doms_cur = doms_new;
6693 dattr_cur = dattr_new;
6694 ndoms_cur = ndoms_new;
6696 register_sched_domain_sysctl();
6698 mutex_unlock(&sched_domains_mutex);
6699 }
6701 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6703 /*
6704 * Update cpusets according to cpu_active mask. If cpusets are
6705 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6706 * around partition_sched_domains().
6707 *
6708 * If we come here as part of a suspend/resume, don't touch cpusets because we
6709 * want to restore it back to its original state upon resume anyway.
6710 */
6711 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6712 void *hcpu)
6713 {
6714 switch (action) {
6715 case CPU_ONLINE_FROZEN:
6716 case CPU_DOWN_FAILED_FROZEN:
6718 /*
6719 * num_cpus_frozen tracks how many CPUs are involved in suspend
6720 * resume sequence. As long as this is not the last online
6721 * operation in the resume sequence, just build a single sched
6722 * domain, ignoring cpusets.
6723 */
6724 num_cpus_frozen--;
6725 if (likely(num_cpus_frozen)) {
6726 partition_sched_domains(1, NULL, NULL);
6727 break;
6728 }
6730 /*
6731 * This is the last CPU online operation. So fall through and
6732 * restore the original sched domains by considering the
6733 * cpuset configurations.
6734 */
6736 case CPU_ONLINE:
6737 case CPU_DOWN_FAILED:
6738 cpuset_update_active_cpus(true);
6739 break;
6740 default:
6741 return NOTIFY_DONE;
6742 }
6743 return NOTIFY_OK;
6744 }
6746 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6747 void *hcpu)
6748 {
6749 switch (action) {
6750 case CPU_DOWN_PREPARE:
6751 cpuset_update_active_cpus(false);
6752 break;
6753 case CPU_DOWN_PREPARE_FROZEN:
6754 num_cpus_frozen++;
6755 partition_sched_domains(1, NULL, NULL);
6756 break;
6757 default:
6758 return NOTIFY_DONE;
6759 }
6760 return NOTIFY_OK;
6761 }
6763 void __init sched_init_smp(void)
6764 {
6765 cpumask_var_t non_isolated_cpus;
6767 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6768 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6770 sched_init_numa();
6772 get_online_cpus();
6773 mutex_lock(&sched_domains_mutex);
6774 init_sched_domains(cpu_active_mask);
6775 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6776 if (cpumask_empty(non_isolated_cpus))
6777 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6778 mutex_unlock(&sched_domains_mutex);
6779 put_online_cpus();
6781 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6782 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6783 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6785 /* RT runtime code needs to handle some hotplug events */
6786 hotcpu_notifier(update_runtime, 0);
6788 init_hrtick();
6790 /* Move init over to a non-isolated CPU */
6791 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6792 BUG();
6793 sched_init_granularity();
6794 free_cpumask_var(non_isolated_cpus);
6796 init_sched_rt_class();
6797 }
6798 #else
6799 void __init sched_init_smp(void)
6800 {
6801 sched_init_granularity();
6802 }
6803 #endif /* CONFIG_SMP */
6805 const_debug unsigned int sysctl_timer_migration = 1;
6807 int in_sched_functions(unsigned long addr)
6808 {
6809 return in_lock_functions(addr) ||
6810 (addr >= (unsigned long)__sched_text_start
6811 && addr < (unsigned long)__sched_text_end);
6812 }
6814 #ifdef CONFIG_CGROUP_SCHED
6815 struct task_group root_task_group;
6816 LIST_HEAD(task_groups);
6817 #endif
6819 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6821 void __init sched_init(void)
6822 {
6823 int i, j;
6824 unsigned long alloc_size = 0, ptr;
6826 #ifdef CONFIG_FAIR_GROUP_SCHED
6827 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6828 #endif
6829 #ifdef CONFIG_RT_GROUP_SCHED
6830 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6831 #endif
6832 #ifdef CONFIG_CPUMASK_OFFSTACK
6833 alloc_size += num_possible_cpus() * cpumask_size();
6834 #endif
6835 if (alloc_size) {
6836 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6838 #ifdef CONFIG_FAIR_GROUP_SCHED
6839 root_task_group.se = (struct sched_entity **)ptr;
6840 ptr += nr_cpu_ids * sizeof(void **);
6842 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6843 ptr += nr_cpu_ids * sizeof(void **);
6845 #endif /* CONFIG_FAIR_GROUP_SCHED */
6846 #ifdef CONFIG_RT_GROUP_SCHED
6847 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6848 ptr += nr_cpu_ids * sizeof(void **);
6850 root_task_group.rt_rq = (struct rt_rq **)ptr;
6851 ptr += nr_cpu_ids * sizeof(void **);
6853 #endif /* CONFIG_RT_GROUP_SCHED */
6854 #ifdef CONFIG_CPUMASK_OFFSTACK
6855 for_each_possible_cpu(i) {
6856 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6857 ptr += cpumask_size();
6858 }
6859 #endif /* CONFIG_CPUMASK_OFFSTACK */
6860 }
6862 #ifdef CONFIG_SMP
6863 init_defrootdomain();
6864 #endif
6866 init_rt_bandwidth(&def_rt_bandwidth,
6867 global_rt_period(), global_rt_runtime());
6869 #ifdef CONFIG_RT_GROUP_SCHED
6870 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6871 global_rt_period(), global_rt_runtime());
6872 #endif /* CONFIG_RT_GROUP_SCHED */
6874 #ifdef CONFIG_CGROUP_SCHED
6875 list_add(&root_task_group.list, &task_groups);
6876 INIT_LIST_HEAD(&root_task_group.children);
6877 INIT_LIST_HEAD(&root_task_group.siblings);
6878 autogroup_init(&init_task);
6880 #endif /* CONFIG_CGROUP_SCHED */
6882 #ifdef CONFIG_CGROUP_CPUACCT
6883 root_cpuacct.cpustat = &kernel_cpustat;
6884 root_cpuacct.cpuusage = alloc_percpu(u64);
6885 /* Too early, not expected to fail */
6886 BUG_ON(!root_cpuacct.cpuusage);
6887 #endif
6888 for_each_possible_cpu(i) {
6889 struct rq *rq;
6891 rq = cpu_rq(i);
6892 raw_spin_lock_init(&rq->lock);
6893 rq->nr_running = 0;
6894 rq->calc_load_active = 0;
6895 rq->calc_load_update = jiffies + LOAD_FREQ;
6896 init_cfs_rq(&rq->cfs);
6897 init_rt_rq(&rq->rt, rq);
6898 #ifdef CONFIG_FAIR_GROUP_SCHED
6899 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6900 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6901 /*
6902 * How much cpu bandwidth does root_task_group get?
6903 *
6904 * In case of task-groups formed thr' the cgroup filesystem, it
6905 * gets 100% of the cpu resources in the system. This overall
6906 * system cpu resource is divided among the tasks of
6907 * root_task_group and its child task-groups in a fair manner,
6908 * based on each entity's (task or task-group's) weight
6909 * (se->load.weight).
6910 *
6911 * In other words, if root_task_group has 10 tasks of weight
6912 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6913 * then A0's share of the cpu resource is:
6914 *
6915 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6916 *
6917 * We achieve this by letting root_task_group's tasks sit
6918 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6919 */
6920 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6921 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6922 #endif /* CONFIG_FAIR_GROUP_SCHED */
6924 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6925 #ifdef CONFIG_RT_GROUP_SCHED
6926 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6927 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6928 #endif
6930 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6931 rq->cpu_load[j] = 0;
6933 rq->last_load_update_tick = jiffies;
6935 #ifdef CONFIG_SMP
6936 rq->sd = NULL;
6937 rq->rd = NULL;
6938 rq->cpu_power = SCHED_POWER_SCALE;
6939 rq->post_schedule = 0;
6940 rq->active_balance = 0;
6941 rq->next_balance = jiffies;
6942 rq->push_cpu = 0;
6943 rq->cpu = i;
6944 rq->online = 0;
6945 rq->idle_stamp = 0;
6946 rq->avg_idle = 2*sysctl_sched_migration_cost;
6948 INIT_LIST_HEAD(&rq->cfs_tasks);
6950 rq_attach_root(rq, &def_root_domain);
6951 #ifdef CONFIG_NO_HZ
6952 rq->nohz_flags = 0;
6953 #endif
6954 #endif
6955 init_rq_hrtick(rq);
6956 atomic_set(&rq->nr_iowait, 0);
6957 }
6959 set_load_weight(&init_task);
6961 #ifdef CONFIG_PREEMPT_NOTIFIERS
6962 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6963 #endif
6965 #ifdef CONFIG_RT_MUTEXES
6966 plist_head_init(&init_task.pi_waiters);
6967 #endif
6969 /*
6970 * The boot idle thread does lazy MMU switching as well:
6971 */
6972 atomic_inc(&init_mm.mm_count);
6973 enter_lazy_tlb(&init_mm, current);
6975 /*
6976 * Make us the idle thread. Technically, schedule() should not be
6977 * called from this thread, however somewhere below it might be,
6978 * but because we are the idle thread, we just pick up running again
6979 * when this runqueue becomes "idle".
6980 */
6981 init_idle(current, smp_processor_id());
6983 calc_load_update = jiffies + LOAD_FREQ;
6985 /*
6986 * During early bootup we pretend to be a normal task:
6987 */
6988 current->sched_class = &fair_sched_class;
6990 #ifdef CONFIG_SMP
6991 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6992 /* May be allocated at isolcpus cmdline parse time */
6993 if (cpu_isolated_map == NULL)
6994 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6995 idle_thread_set_boot_cpu();
6996 #endif
6997 init_sched_fair_class();
6999 scheduler_running = 1;
7000 }
7002 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7003 static inline int preempt_count_equals(int preempt_offset)
7004 {
7005 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7007 return (nested == preempt_offset);
7008 }
7010 void __might_sleep(const char *file, int line, int preempt_offset)
7011 {
7012 static unsigned long prev_jiffy; /* ratelimiting */
7014 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7015 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7016 system_state != SYSTEM_RUNNING || oops_in_progress)
7017 return;
7018 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7019 return;
7020 prev_jiffy = jiffies;
7022 printk(KERN_ERR
7023 "BUG: sleeping function called from invalid context at %s:%d\n",
7024 file, line);
7025 printk(KERN_ERR
7026 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7027 in_atomic(), irqs_disabled(),
7028 current->pid, current->comm);
7030 debug_show_held_locks(current);
7031 if (irqs_disabled())
7032 print_irqtrace_events(current);
7033 dump_stack();
7034 }
7035 EXPORT_SYMBOL(__might_sleep);
7036 #endif
7038 #ifdef CONFIG_MAGIC_SYSRQ
7039 static void normalize_task(struct rq *rq, struct task_struct *p)
7040 {
7041 const struct sched_class *prev_class = p->sched_class;
7042 int old_prio = p->prio;
7043 int on_rq;
7045 on_rq = p->on_rq;
7046 if (on_rq)
7047 dequeue_task(rq, p, 0);
7048 __setscheduler(rq, p, SCHED_NORMAL, 0);
7049 if (on_rq) {
7050 enqueue_task(rq, p, 0);
7051 resched_task(rq->curr);
7052 }
7054 check_class_changed(rq, p, prev_class, old_prio);
7055 }
7057 void normalize_rt_tasks(void)
7058 {
7059 struct task_struct *g, *p;
7060 unsigned long flags;
7061 struct rq *rq;
7063 read_lock_irqsave(&tasklist_lock, flags);
7064 do_each_thread(g, p) {
7065 /*
7066 * Only normalize user tasks:
7067 */
7068 if (!p->mm)
7069 continue;
7071 p->se.exec_start = 0;
7072 #ifdef CONFIG_SCHEDSTATS
7073 p->se.statistics.wait_start = 0;
7074 p->se.statistics.sleep_start = 0;
7075 p->se.statistics.block_start = 0;
7076 #endif
7078 if (!rt_task(p)) {
7079 /*
7080 * Renice negative nice level userspace
7081 * tasks back to 0:
7082 */
7083 if (TASK_NICE(p) < 0 && p->mm)
7084 set_user_nice(p, 0);
7085 continue;
7086 }
7088 raw_spin_lock(&p->pi_lock);
7089 rq = __task_rq_lock(p);
7091 normalize_task(rq, p);
7093 __task_rq_unlock(rq);
7094 raw_spin_unlock(&p->pi_lock);
7095 } while_each_thread(g, p);
7097 read_unlock_irqrestore(&tasklist_lock, flags);
7098 }
7100 #endif /* CONFIG_MAGIC_SYSRQ */
7102 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7103 /*
7104 * These functions are only useful for the IA64 MCA handling, or kdb.
7105 *
7106 * They can only be called when the whole system has been
7107 * stopped - every CPU needs to be quiescent, and no scheduling
7108 * activity can take place. Using them for anything else would
7109 * be a serious bug, and as a result, they aren't even visible
7110 * under any other configuration.
7111 */
7113 /**
7114 * curr_task - return the current task for a given cpu.
7115 * @cpu: the processor in question.
7116 *
7117 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7118 */
7119 struct task_struct *curr_task(int cpu)
7120 {
7121 return cpu_curr(cpu);
7122 }
7124 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7126 #ifdef CONFIG_IA64
7127 /**
7128 * set_curr_task - set the current task for a given cpu.
7129 * @cpu: the processor in question.
7130 * @p: the task pointer to set.
7131 *
7132 * Description: This function must only be used when non-maskable interrupts
7133 * are serviced on a separate stack. It allows the architecture to switch the
7134 * notion of the current task on a cpu in a non-blocking manner. This function
7135 * must be called with all CPU's synchronized, and interrupts disabled, the
7136 * and caller must save the original value of the current task (see
7137 * curr_task() above) and restore that value before reenabling interrupts and
7138 * re-starting the system.
7139 *
7140 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7141 */
7142 void set_curr_task(int cpu, struct task_struct *p)
7143 {
7144 cpu_curr(cpu) = p;
7145 }
7147 #endif
7149 #ifdef CONFIG_CGROUP_SCHED
7150 /* task_group_lock serializes the addition/removal of task groups */
7151 static DEFINE_SPINLOCK(task_group_lock);
7153 static void free_sched_group(struct task_group *tg)
7154 {
7155 free_fair_sched_group(tg);
7156 free_rt_sched_group(tg);
7157 autogroup_free(tg);
7158 kfree(tg);
7159 }
7161 /* allocate runqueue etc for a new task group */
7162 struct task_group *sched_create_group(struct task_group *parent)
7163 {
7164 struct task_group *tg;
7165 unsigned long flags;
7167 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7168 if (!tg)
7169 return ERR_PTR(-ENOMEM);
7171 if (!alloc_fair_sched_group(tg, parent))
7172 goto err;
7174 if (!alloc_rt_sched_group(tg, parent))
7175 goto err;
7177 spin_lock_irqsave(&task_group_lock, flags);
7178 list_add_rcu(&tg->list, &task_groups);
7180 WARN_ON(!parent); /* root should already exist */
7182 tg->parent = parent;
7183 INIT_LIST_HEAD(&tg->children);
7184 list_add_rcu(&tg->siblings, &parent->children);
7185 spin_unlock_irqrestore(&task_group_lock, flags);
7187 return tg;
7189 err:
7190 free_sched_group(tg);
7191 return ERR_PTR(-ENOMEM);
7192 }
7194 /* rcu callback to free various structures associated with a task group */
7195 static void free_sched_group_rcu(struct rcu_head *rhp)
7196 {
7197 /* now it should be safe to free those cfs_rqs */
7198 free_sched_group(container_of(rhp, struct task_group, rcu));
7199 }
7201 /* Destroy runqueue etc associated with a task group */
7202 void sched_destroy_group(struct task_group *tg)
7203 {
7204 unsigned long flags;
7205 int i;
7207 /* end participation in shares distribution */
7208 for_each_possible_cpu(i)
7209 unregister_fair_sched_group(tg, i);
7211 spin_lock_irqsave(&task_group_lock, flags);
7212 list_del_rcu(&tg->list);
7213 list_del_rcu(&tg->siblings);
7214 spin_unlock_irqrestore(&task_group_lock, flags);
7216 /* wait for possible concurrent references to cfs_rqs complete */
7217 call_rcu(&tg->rcu, free_sched_group_rcu);
7218 }
7220 /* change task's runqueue when it moves between groups.
7221 * The caller of this function should have put the task in its new group
7222 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7223 * reflect its new group.
7224 */
7225 void sched_move_task(struct task_struct *tsk)
7226 {
7227 struct task_group *tg;
7228 int on_rq, running;
7229 unsigned long flags;
7230 struct rq *rq;
7232 rq = task_rq_lock(tsk, &flags);
7234 running = task_current(rq, tsk);
7235 on_rq = tsk->on_rq;
7237 if (on_rq)
7238 dequeue_task(rq, tsk, 0);
7239 if (unlikely(running))
7240 tsk->sched_class->put_prev_task(rq, tsk);
7242 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7243 lockdep_is_held(&tsk->sighand->siglock)),
7244 struct task_group, css);
7245 tg = autogroup_task_group(tsk, tg);
7246 tsk->sched_task_group = tg;
7248 #ifdef CONFIG_FAIR_GROUP_SCHED
7249 if (tsk->sched_class->task_move_group)
7250 tsk->sched_class->task_move_group(tsk, on_rq);
7251 else
7252 #endif
7253 set_task_rq(tsk, task_cpu(tsk));
7255 if (unlikely(running))
7256 tsk->sched_class->set_curr_task(rq);
7257 if (on_rq)
7258 enqueue_task(rq, tsk, 0);
7260 task_rq_unlock(rq, tsk, &flags);
7261 }
7262 #endif /* CONFIG_CGROUP_SCHED */
7264 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7265 static unsigned long to_ratio(u64 period, u64 runtime)
7266 {
7267 if (runtime == RUNTIME_INF)
7268 return 1ULL << 20;
7270 return div64_u64(runtime << 20, period);
7271 }
7272 #endif
7274 #ifdef CONFIG_RT_GROUP_SCHED
7275 /*
7276 * Ensure that the real time constraints are schedulable.
7277 */
7278 static DEFINE_MUTEX(rt_constraints_mutex);
7280 /* Must be called with tasklist_lock held */
7281 static inline int tg_has_rt_tasks(struct task_group *tg)
7282 {
7283 struct task_struct *g, *p;
7285 do_each_thread(g, p) {
7286 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7287 return 1;
7288 } while_each_thread(g, p);
7290 return 0;
7291 }
7293 struct rt_schedulable_data {
7294 struct task_group *tg;
7295 u64 rt_period;
7296 u64 rt_runtime;
7297 };
7299 static int tg_rt_schedulable(struct task_group *tg, void *data)
7300 {
7301 struct rt_schedulable_data *d = data;
7302 struct task_group *child;
7303 unsigned long total, sum = 0;
7304 u64 period, runtime;
7306 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7307 runtime = tg->rt_bandwidth.rt_runtime;
7309 if (tg == d->tg) {
7310 period = d->rt_period;
7311 runtime = d->rt_runtime;
7312 }
7314 /*
7315 * Cannot have more runtime than the period.
7316 */
7317 if (runtime > period && runtime != RUNTIME_INF)
7318 return -EINVAL;
7320 /*
7321 * Ensure we don't starve existing RT tasks.
7322 */
7323 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7324 return -EBUSY;
7326 total = to_ratio(period, runtime);
7328 /*
7329 * Nobody can have more than the global setting allows.
7330 */
7331 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7332 return -EINVAL;
7334 /*
7335 * The sum of our children's runtime should not exceed our own.
7336 */
7337 list_for_each_entry_rcu(child, &tg->children, siblings) {
7338 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7339 runtime = child->rt_bandwidth.rt_runtime;
7341 if (child == d->tg) {
7342 period = d->rt_period;
7343 runtime = d->rt_runtime;
7344 }
7346 sum += to_ratio(period, runtime);
7347 }
7349 if (sum > total)
7350 return -EINVAL;
7352 return 0;
7353 }
7355 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7356 {
7357 int ret;
7359 struct rt_schedulable_data data = {
7360 .tg = tg,
7361 .rt_period = period,
7362 .rt_runtime = runtime,
7363 };
7365 rcu_read_lock();
7366 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7367 rcu_read_unlock();
7369 return ret;
7370 }
7372 static int tg_set_rt_bandwidth(struct task_group *tg,
7373 u64 rt_period, u64 rt_runtime)
7374 {
7375 int i, err = 0;
7377 mutex_lock(&rt_constraints_mutex);
7378 read_lock(&tasklist_lock);
7379 err = __rt_schedulable(tg, rt_period, rt_runtime);
7380 if (err)
7381 goto unlock;
7383 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7384 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7385 tg->rt_bandwidth.rt_runtime = rt_runtime;
7387 for_each_possible_cpu(i) {
7388 struct rt_rq *rt_rq = tg->rt_rq[i];
7390 raw_spin_lock(&rt_rq->rt_runtime_lock);
7391 rt_rq->rt_runtime = rt_runtime;
7392 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7393 }
7394 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7395 unlock:
7396 read_unlock(&tasklist_lock);
7397 mutex_unlock(&rt_constraints_mutex);
7399 return err;
7400 }
7402 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7403 {
7404 u64 rt_runtime, rt_period;
7406 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7407 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7408 if (rt_runtime_us < 0)
7409 rt_runtime = RUNTIME_INF;
7411 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7412 }
7414 long sched_group_rt_runtime(struct task_group *tg)
7415 {
7416 u64 rt_runtime_us;
7418 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7419 return -1;
7421 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7422 do_div(rt_runtime_us, NSEC_PER_USEC);
7423 return rt_runtime_us;
7424 }
7426 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7427 {
7428 u64 rt_runtime, rt_period;
7430 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7431 rt_runtime = tg->rt_bandwidth.rt_runtime;
7433 if (rt_period == 0)
7434 return -EINVAL;
7436 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7437 }
7439 long sched_group_rt_period(struct task_group *tg)
7440 {
7441 u64 rt_period_us;
7443 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7444 do_div(rt_period_us, NSEC_PER_USEC);
7445 return rt_period_us;
7446 }
7448 static int sched_rt_global_constraints(void)
7449 {
7450 u64 runtime, period;
7451 int ret = 0;
7453 if (sysctl_sched_rt_period <= 0)
7454 return -EINVAL;
7456 runtime = global_rt_runtime();
7457 period = global_rt_period();
7459 /*
7460 * Sanity check on the sysctl variables.
7461 */
7462 if (runtime > period && runtime != RUNTIME_INF)
7463 return -EINVAL;
7465 mutex_lock(&rt_constraints_mutex);
7466 read_lock(&tasklist_lock);
7467 ret = __rt_schedulable(NULL, 0, 0);
7468 read_unlock(&tasklist_lock);
7469 mutex_unlock(&rt_constraints_mutex);
7471 return ret;
7472 }
7474 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7475 {
7476 /* Don't accept realtime tasks when there is no way for them to run */
7477 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7478 return 0;
7480 return 1;
7481 }
7483 #else /* !CONFIG_RT_GROUP_SCHED */
7484 static int sched_rt_global_constraints(void)
7485 {
7486 unsigned long flags;
7487 int i;
7489 if (sysctl_sched_rt_period <= 0)
7490 return -EINVAL;
7492 /*
7493 * There's always some RT tasks in the root group
7494 * -- migration, kstopmachine etc..
7495 */
7496 if (sysctl_sched_rt_runtime == 0)
7497 return -EBUSY;
7499 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7500 for_each_possible_cpu(i) {
7501 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7503 raw_spin_lock(&rt_rq->rt_runtime_lock);
7504 rt_rq->rt_runtime = global_rt_runtime();
7505 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7506 }
7507 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7509 return 0;
7510 }
7511 #endif /* CONFIG_RT_GROUP_SCHED */
7513 int sched_rt_handler(struct ctl_table *table, int write,
7514 void __user *buffer, size_t *lenp,
7515 loff_t *ppos)
7516 {
7517 int ret;
7518 int old_period, old_runtime;
7519 static DEFINE_MUTEX(mutex);
7521 mutex_lock(&mutex);
7522 old_period = sysctl_sched_rt_period;
7523 old_runtime = sysctl_sched_rt_runtime;
7525 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7527 if (!ret && write) {
7528 ret = sched_rt_global_constraints();
7529 if (ret) {
7530 sysctl_sched_rt_period = old_period;
7531 sysctl_sched_rt_runtime = old_runtime;
7532 } else {
7533 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7534 def_rt_bandwidth.rt_period =
7535 ns_to_ktime(global_rt_period());
7536 }
7537 }
7538 mutex_unlock(&mutex);
7540 return ret;
7541 }
7543 #ifdef CONFIG_CGROUP_SCHED
7545 /* return corresponding task_group object of a cgroup */
7546 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7547 {
7548 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7549 struct task_group, css);
7550 }
7552 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7553 {
7554 struct task_group *tg, *parent;
7556 if (!cgrp->parent) {
7557 /* This is early initialization for the top cgroup */
7558 return &root_task_group.css;
7559 }
7561 parent = cgroup_tg(cgrp->parent);
7562 tg = sched_create_group(parent);
7563 if (IS_ERR(tg))
7564 return ERR_PTR(-ENOMEM);
7566 return &tg->css;
7567 }
7569 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7570 {
7571 struct task_group *tg = cgroup_tg(cgrp);
7573 sched_destroy_group(tg);
7574 }
7576 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7577 struct cgroup_taskset *tset)
7578 {
7579 struct task_struct *task;
7581 cgroup_taskset_for_each(task, cgrp, tset) {
7582 #ifdef CONFIG_RT_GROUP_SCHED
7583 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7584 return -EINVAL;
7585 #else
7586 /* We don't support RT-tasks being in separate groups */
7587 if (task->sched_class != &fair_sched_class)
7588 return -EINVAL;
7589 #endif
7590 }
7591 return 0;
7592 }
7594 static void cpu_cgroup_attach(struct cgroup *cgrp,
7595 struct cgroup_taskset *tset)
7596 {
7597 struct task_struct *task;
7599 cgroup_taskset_for_each(task, cgrp, tset)
7600 sched_move_task(task);
7601 }
7603 static void
7604 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7605 struct task_struct *task)
7606 {
7607 /*
7608 * cgroup_exit() is called in the copy_process() failure path.
7609 * Ignore this case since the task hasn't ran yet, this avoids
7610 * trying to poke a half freed task state from generic code.
7611 */
7612 if (!(task->flags & PF_EXITING))
7613 return;
7615 sched_move_task(task);
7616 }
7618 #ifdef CONFIG_FAIR_GROUP_SCHED
7619 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7620 u64 shareval)
7621 {
7622 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7623 }
7625 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7626 {
7627 struct task_group *tg = cgroup_tg(cgrp);
7629 return (u64) scale_load_down(tg->shares);
7630 }
7632 #ifdef CONFIG_CFS_BANDWIDTH
7633 static DEFINE_MUTEX(cfs_constraints_mutex);
7635 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7636 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7638 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7640 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7641 {
7642 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7643 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7645 if (tg == &root_task_group)
7646 return -EINVAL;
7648 /*
7649 * Ensure we have at some amount of bandwidth every period. This is
7650 * to prevent reaching a state of large arrears when throttled via
7651 * entity_tick() resulting in prolonged exit starvation.
7652 */
7653 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7654 return -EINVAL;
7656 /*
7657 * Likewise, bound things on the otherside by preventing insane quota
7658 * periods. This also allows us to normalize in computing quota
7659 * feasibility.
7660 */
7661 if (period > max_cfs_quota_period)
7662 return -EINVAL;
7664 mutex_lock(&cfs_constraints_mutex);
7665 ret = __cfs_schedulable(tg, period, quota);
7666 if (ret)
7667 goto out_unlock;
7669 runtime_enabled = quota != RUNTIME_INF;
7670 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7671 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7672 raw_spin_lock_irq(&cfs_b->lock);
7673 cfs_b->period = ns_to_ktime(period);
7674 cfs_b->quota = quota;
7676 __refill_cfs_bandwidth_runtime(cfs_b);
7677 /* restart the period timer (if active) to handle new period expiry */
7678 if (runtime_enabled && cfs_b->timer_active) {
7679 /* force a reprogram */
7680 cfs_b->timer_active = 0;
7681 __start_cfs_bandwidth(cfs_b);
7682 }
7683 raw_spin_unlock_irq(&cfs_b->lock);
7685 for_each_possible_cpu(i) {
7686 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7687 struct rq *rq = cfs_rq->rq;
7689 raw_spin_lock_irq(&rq->lock);
7690 cfs_rq->runtime_enabled = runtime_enabled;
7691 cfs_rq->runtime_remaining = 0;
7693 if (cfs_rq->throttled)
7694 unthrottle_cfs_rq(cfs_rq);
7695 raw_spin_unlock_irq(&rq->lock);
7696 }
7697 out_unlock:
7698 mutex_unlock(&cfs_constraints_mutex);
7700 return ret;
7701 }
7703 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7704 {
7705 u64 quota, period;
7707 period = ktime_to_ns(tg->cfs_bandwidth.period);
7708 if (cfs_quota_us < 0)
7709 quota = RUNTIME_INF;
7710 else
7711 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7713 return tg_set_cfs_bandwidth(tg, period, quota);
7714 }
7716 long tg_get_cfs_quota(struct task_group *tg)
7717 {
7718 u64 quota_us;
7720 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7721 return -1;
7723 quota_us = tg->cfs_bandwidth.quota;
7724 do_div(quota_us, NSEC_PER_USEC);
7726 return quota_us;
7727 }
7729 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7730 {
7731 u64 quota, period;
7733 period = (u64)cfs_period_us * NSEC_PER_USEC;
7734 quota = tg->cfs_bandwidth.quota;
7736 return tg_set_cfs_bandwidth(tg, period, quota);
7737 }
7739 long tg_get_cfs_period(struct task_group *tg)
7740 {
7741 u64 cfs_period_us;
7743 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7744 do_div(cfs_period_us, NSEC_PER_USEC);
7746 return cfs_period_us;
7747 }
7749 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7750 {
7751 return tg_get_cfs_quota(cgroup_tg(cgrp));
7752 }
7754 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7755 s64 cfs_quota_us)
7756 {
7757 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7758 }
7760 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7761 {
7762 return tg_get_cfs_period(cgroup_tg(cgrp));
7763 }
7765 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7766 u64 cfs_period_us)
7767 {
7768 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7769 }
7771 struct cfs_schedulable_data {
7772 struct task_group *tg;
7773 u64 period, quota;
7774 };
7776 /*
7777 * normalize group quota/period to be quota/max_period
7778 * note: units are usecs
7779 */
7780 static u64 normalize_cfs_quota(struct task_group *tg,
7781 struct cfs_schedulable_data *d)
7782 {
7783 u64 quota, period;
7785 if (tg == d->tg) {
7786 period = d->period;
7787 quota = d->quota;
7788 } else {
7789 period = tg_get_cfs_period(tg);
7790 quota = tg_get_cfs_quota(tg);
7791 }
7793 /* note: these should typically be equivalent */
7794 if (quota == RUNTIME_INF || quota == -1)
7795 return RUNTIME_INF;
7797 return to_ratio(period, quota);
7798 }
7800 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7801 {
7802 struct cfs_schedulable_data *d = data;
7803 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7804 s64 quota = 0, parent_quota = -1;
7806 if (!tg->parent) {
7807 quota = RUNTIME_INF;
7808 } else {
7809 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7811 quota = normalize_cfs_quota(tg, d);
7812 parent_quota = parent_b->hierarchal_quota;
7814 /*
7815 * ensure max(child_quota) <= parent_quota, inherit when no
7816 * limit is set
7817 */
7818 if (quota == RUNTIME_INF)
7819 quota = parent_quota;
7820 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7821 return -EINVAL;
7822 }
7823 cfs_b->hierarchal_quota = quota;
7825 return 0;
7826 }
7828 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7829 {
7830 int ret;
7831 struct cfs_schedulable_data data = {
7832 .tg = tg,
7833 .period = period,
7834 .quota = quota,
7835 };
7837 if (quota != RUNTIME_INF) {
7838 do_div(data.period, NSEC_PER_USEC);
7839 do_div(data.quota, NSEC_PER_USEC);
7840 }
7842 rcu_read_lock();
7843 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7844 rcu_read_unlock();
7846 return ret;
7847 }
7849 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7850 struct cgroup_map_cb *cb)
7851 {
7852 struct task_group *tg = cgroup_tg(cgrp);
7853 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7855 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7856 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7857 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7859 return 0;
7860 }
7861 #endif /* CONFIG_CFS_BANDWIDTH */
7862 #endif /* CONFIG_FAIR_GROUP_SCHED */
7864 #ifdef CONFIG_RT_GROUP_SCHED
7865 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7866 s64 val)
7867 {
7868 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7869 }
7871 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7872 {
7873 return sched_group_rt_runtime(cgroup_tg(cgrp));
7874 }
7876 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7877 u64 rt_period_us)
7878 {
7879 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7880 }
7882 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7883 {
7884 return sched_group_rt_period(cgroup_tg(cgrp));
7885 }
7886 #endif /* CONFIG_RT_GROUP_SCHED */
7888 static struct cftype cpu_files[] = {
7889 #ifdef CONFIG_FAIR_GROUP_SCHED
7890 {
7891 .name = "shares",
7892 .read_u64 = cpu_shares_read_u64,
7893 .write_u64 = cpu_shares_write_u64,
7894 },
7895 #endif
7896 #ifdef CONFIG_CFS_BANDWIDTH
7897 {
7898 .name = "cfs_quota_us",
7899 .read_s64 = cpu_cfs_quota_read_s64,
7900 .write_s64 = cpu_cfs_quota_write_s64,
7901 },
7902 {
7903 .name = "cfs_period_us",
7904 .read_u64 = cpu_cfs_period_read_u64,
7905 .write_u64 = cpu_cfs_period_write_u64,
7906 },
7907 {
7908 .name = "stat",
7909 .read_map = cpu_stats_show,
7910 },
7911 #endif
7912 #ifdef CONFIG_RT_GROUP_SCHED
7913 {
7914 .name = "rt_runtime_us",
7915 .read_s64 = cpu_rt_runtime_read,
7916 .write_s64 = cpu_rt_runtime_write,
7917 },
7918 {
7919 .name = "rt_period_us",
7920 .read_u64 = cpu_rt_period_read_uint,
7921 .write_u64 = cpu_rt_period_write_uint,
7922 },
7923 #endif
7924 { } /* terminate */
7925 };
7927 struct cgroup_subsys cpu_cgroup_subsys = {
7928 .name = "cpu",
7929 .css_alloc = cpu_cgroup_css_alloc,
7930 .css_free = cpu_cgroup_css_free,
7931 .can_attach = cpu_cgroup_can_attach,
7932 .attach = cpu_cgroup_attach,
7933 .exit = cpu_cgroup_exit,
7934 .subsys_id = cpu_cgroup_subsys_id,
7935 .base_cftypes = cpu_files,
7936 .early_init = 1,
7937 };
7939 #endif /* CONFIG_CGROUP_SCHED */
7941 #ifdef CONFIG_CGROUP_CPUACCT
7943 /*
7944 * CPU accounting code for task groups.
7945 *
7946 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7947 * (balbir@in.ibm.com).
7948 */
7950 struct cpuacct root_cpuacct;
7952 /* create a new cpu accounting group */
7953 static struct cgroup_subsys_state *cpuacct_css_alloc(struct cgroup *cgrp)
7954 {
7955 struct cpuacct *ca;
7957 if (!cgrp->parent)
7958 return &root_cpuacct.css;
7960 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7961 if (!ca)
7962 goto out;
7964 ca->cpuusage = alloc_percpu(u64);
7965 if (!ca->cpuusage)
7966 goto out_free_ca;
7968 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7969 if (!ca->cpustat)
7970 goto out_free_cpuusage;
7972 return &ca->css;
7974 out_free_cpuusage:
7975 free_percpu(ca->cpuusage);
7976 out_free_ca:
7977 kfree(ca);
7978 out:
7979 return ERR_PTR(-ENOMEM);
7980 }
7982 /* destroy an existing cpu accounting group */
7983 static void cpuacct_css_free(struct cgroup *cgrp)
7984 {
7985 struct cpuacct *ca = cgroup_ca(cgrp);
7987 free_percpu(ca->cpustat);
7988 free_percpu(ca->cpuusage);
7989 kfree(ca);
7990 }
7992 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7993 {
7994 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7995 u64 data;
7997 #ifndef CONFIG_64BIT
7998 /*
7999 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8000 */
8001 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8002 data = *cpuusage;
8003 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8004 #else
8005 data = *cpuusage;
8006 #endif
8008 return data;
8009 }
8011 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8012 {
8013 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8015 #ifndef CONFIG_64BIT
8016 /*
8017 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8018 */
8019 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8020 *cpuusage = val;
8021 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8022 #else
8023 *cpuusage = val;
8024 #endif
8025 }
8027 /* return total cpu usage (in nanoseconds) of a group */
8028 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8029 {
8030 struct cpuacct *ca = cgroup_ca(cgrp);
8031 u64 totalcpuusage = 0;
8032 int i;
8034 for_each_present_cpu(i)
8035 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8037 return totalcpuusage;
8038 }
8040 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8041 u64 reset)
8042 {
8043 struct cpuacct *ca = cgroup_ca(cgrp);
8044 int err = 0;
8045 int i;
8047 if (reset) {
8048 err = -EINVAL;
8049 goto out;
8050 }
8052 for_each_present_cpu(i)
8053 cpuacct_cpuusage_write(ca, i, 0);
8055 out:
8056 return err;
8057 }
8059 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8060 struct seq_file *m)
8061 {
8062 struct cpuacct *ca = cgroup_ca(cgroup);
8063 u64 percpu;
8064 int i;
8066 for_each_present_cpu(i) {
8067 percpu = cpuacct_cpuusage_read(ca, i);
8068 seq_printf(m, "%llu ", (unsigned long long) percpu);
8069 }
8070 seq_printf(m, "\n");
8071 return 0;
8072 }
8074 static const char *cpuacct_stat_desc[] = {
8075 [CPUACCT_STAT_USER] = "user",
8076 [CPUACCT_STAT_SYSTEM] = "system",
8077 };
8079 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8080 struct cgroup_map_cb *cb)
8081 {
8082 struct cpuacct *ca = cgroup_ca(cgrp);
8083 int cpu;
8084 s64 val = 0;
8086 for_each_online_cpu(cpu) {
8087 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8088 val += kcpustat->cpustat[CPUTIME_USER];
8089 val += kcpustat->cpustat[CPUTIME_NICE];
8090 }
8091 val = cputime64_to_clock_t(val);
8092 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8094 val = 0;
8095 for_each_online_cpu(cpu) {
8096 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8097 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8098 val += kcpustat->cpustat[CPUTIME_IRQ];
8099 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8100 }
8102 val = cputime64_to_clock_t(val);
8103 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8105 return 0;
8106 }
8108 static struct cftype files[] = {
8109 {
8110 .name = "usage",
8111 .read_u64 = cpuusage_read,
8112 .write_u64 = cpuusage_write,
8113 },
8114 {
8115 .name = "usage_percpu",
8116 .read_seq_string = cpuacct_percpu_seq_read,
8117 },
8118 {
8119 .name = "stat",
8120 .read_map = cpuacct_stats_show,
8121 },
8122 { } /* terminate */
8123 };
8125 /*
8126 * charge this task's execution time to its accounting group.
8127 *
8128 * called with rq->lock held.
8129 */
8130 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8131 {
8132 struct cpuacct *ca;
8133 int cpu;
8135 if (unlikely(!cpuacct_subsys.active))
8136 return;
8138 cpu = task_cpu(tsk);
8140 rcu_read_lock();
8142 ca = task_ca(tsk);
8144 for (; ca; ca = parent_ca(ca)) {
8145 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8146 *cpuusage += cputime;
8147 }
8149 rcu_read_unlock();
8150 }
8152 struct cgroup_subsys cpuacct_subsys = {
8153 .name = "cpuacct",
8154 .css_alloc = cpuacct_css_alloc,
8155 .css_free = cpuacct_css_free,
8156 .subsys_id = cpuacct_subsys_id,
8157 .base_cftypes = files,
8158 };
8159 #endif /* CONFIG_CGROUP_CPUACCT */
8161 void dump_cpu_task(int cpu)
8162 {
8163 pr_info("Task dump for CPU %d:\n", cpu);
8164 sched_show_task(cpu_curr(cpu));
8165 }