Line data Source code
1 : // SPDX-License-Identifier: GPL-2.0-only
2 : /*
3 : * kernel/sched/core.c
4 : *
5 : * Core kernel scheduler code and related syscalls
6 : *
7 : * Copyright (C) 1991-2002 Linus Torvalds
8 : */
9 : #define CREATE_TRACE_POINTS
10 : #include <trace/events/sched.h>
11 : #undef CREATE_TRACE_POINTS
12 :
13 : #include "sched.h"
14 :
15 : #include <linux/nospec.h>
16 :
17 : #include <linux/kcov.h>
18 : #include <linux/scs.h>
19 :
20 : #include <asm/switch_to.h>
21 : #include <asm/tlb.h>
22 :
23 : #include "../workqueue_internal.h"
24 : #include "../../fs/io-wq.h"
25 : #include "../smpboot.h"
26 :
27 : #include "pelt.h"
28 : #include "smp.h"
29 :
30 : /*
31 : * Export tracepoints that act as a bare tracehook (ie: have no trace event
32 : * associated with them) to allow external modules to probe them.
33 : */
34 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
35 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
36 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
37 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
38 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
39 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
40 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
41 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
42 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
43 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
44 :
45 : DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
46 :
47 : #ifdef CONFIG_SCHED_DEBUG
48 : /*
49 : * Debugging: various feature bits
50 : *
51 : * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52 : * sysctl_sched_features, defined in sched.h, to allow constants propagation
53 : * at compile time and compiler optimization based on features default.
54 : */
55 : #define SCHED_FEAT(name, enabled) \
56 : (1UL << __SCHED_FEAT_##name) * enabled |
57 : const_debug unsigned int sysctl_sched_features =
58 : #include "features.h"
59 : 0;
60 : #undef SCHED_FEAT
61 : #endif
62 :
63 : /*
64 : * Number of tasks to iterate in a single balance run.
65 : * Limited because this is done with IRQs disabled.
66 : */
67 : const_debug unsigned int sysctl_sched_nr_migrate = 32;
68 :
69 : /*
70 : * period over which we measure -rt task CPU usage in us.
71 : * default: 1s
72 : */
73 : unsigned int sysctl_sched_rt_period = 1000000;
74 :
75 : __read_mostly int scheduler_running;
76 :
77 : /*
78 : * part of the period that we allow rt tasks to run in us.
79 : * default: 0.95s
80 : */
81 : int sysctl_sched_rt_runtime = 950000;
82 :
83 :
84 : /*
85 : * Serialization rules:
86 : *
87 : * Lock order:
88 : *
89 : * p->pi_lock
90 : * rq->lock
91 : * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
92 : *
93 : * rq1->lock
94 : * rq2->lock where: rq1 < rq2
95 : *
96 : * Regular state:
97 : *
98 : * Normal scheduling state is serialized by rq->lock. __schedule() takes the
99 : * local CPU's rq->lock, it optionally removes the task from the runqueue and
100 : * always looks at the local rq data structures to find the most eligible task
101 : * to run next.
102 : *
103 : * Task enqueue is also under rq->lock, possibly taken from another CPU.
104 : * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
105 : * the local CPU to avoid bouncing the runqueue state around [ see
106 : * ttwu_queue_wakelist() ]
107 : *
108 : * Task wakeup, specifically wakeups that involve migration, are horribly
109 : * complicated to avoid having to take two rq->locks.
110 : *
111 : * Special state:
112 : *
113 : * System-calls and anything external will use task_rq_lock() which acquires
114 : * both p->pi_lock and rq->lock. As a consequence the state they change is
115 : * stable while holding either lock:
116 : *
117 : * - sched_setaffinity()/
118 : * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
119 : * - set_user_nice(): p->se.load, p->*prio
120 : * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
121 : * p->se.load, p->rt_priority,
122 : * p->dl.dl_{runtime, deadline, period, flags, bw, density}
123 : * - sched_setnuma(): p->numa_preferred_nid
124 : * - sched_move_task()/
125 : * cpu_cgroup_fork(): p->sched_task_group
126 : * - uclamp_update_active() p->uclamp*
127 : *
128 : * p->state <- TASK_*:
129 : *
130 : * is changed locklessly using set_current_state(), __set_current_state() or
131 : * set_special_state(), see their respective comments, or by
132 : * try_to_wake_up(). This latter uses p->pi_lock to serialize against
133 : * concurrent self.
134 : *
135 : * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
136 : *
137 : * is set by activate_task() and cleared by deactivate_task(), under
138 : * rq->lock. Non-zero indicates the task is runnable, the special
139 : * ON_RQ_MIGRATING state is used for migration without holding both
140 : * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
141 : *
142 : * p->on_cpu <- { 0, 1 }:
143 : *
144 : * is set by prepare_task() and cleared by finish_task() such that it will be
145 : * set before p is scheduled-in and cleared after p is scheduled-out, both
146 : * under rq->lock. Non-zero indicates the task is running on its CPU.
147 : *
148 : * [ The astute reader will observe that it is possible for two tasks on one
149 : * CPU to have ->on_cpu = 1 at the same time. ]
150 : *
151 : * task_cpu(p): is changed by set_task_cpu(), the rules are:
152 : *
153 : * - Don't call set_task_cpu() on a blocked task:
154 : *
155 : * We don't care what CPU we're not running on, this simplifies hotplug,
156 : * the CPU assignment of blocked tasks isn't required to be valid.
157 : *
158 : * - for try_to_wake_up(), called under p->pi_lock:
159 : *
160 : * This allows try_to_wake_up() to only take one rq->lock, see its comment.
161 : *
162 : * - for migration called under rq->lock:
163 : * [ see task_on_rq_migrating() in task_rq_lock() ]
164 : *
165 : * o move_queued_task()
166 : * o detach_task()
167 : *
168 : * - for migration called under double_rq_lock():
169 : *
170 : * o __migrate_swap_task()
171 : * o push_rt_task() / pull_rt_task()
172 : * o push_dl_task() / pull_dl_task()
173 : * o dl_task_offline_migration()
174 : *
175 : */
176 :
177 : /*
178 : * __task_rq_lock - lock the rq @p resides on.
179 : */
180 1148 : struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
181 : __acquires(rq->lock)
182 : {
183 1148 : struct rq *rq;
184 :
185 2296 : lockdep_assert_held(&p->pi_lock);
186 :
187 1148 : for (;;) {
188 1148 : rq = task_rq(p);
189 1148 : raw_spin_lock(&rq->lock);
190 1148 : if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
191 1148 : rq_pin_lock(rq, rf);
192 1148 : return rq;
193 : }
194 0 : raw_spin_unlock(&rq->lock);
195 :
196 0 : while (unlikely(task_on_rq_migrating(p)))
197 0 : cpu_relax();
198 : }
199 : }
200 :
201 : /*
202 : * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
203 : */
204 271 : struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
205 : __acquires(p->pi_lock)
206 : __acquires(rq->lock)
207 : {
208 271 : struct rq *rq;
209 :
210 271 : for (;;) {
211 271 : raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
212 271 : rq = task_rq(p);
213 271 : raw_spin_lock(&rq->lock);
214 : /*
215 : * move_queued_task() task_rq_lock()
216 : *
217 : * ACQUIRE (rq->lock)
218 : * [S] ->on_rq = MIGRATING [L] rq = task_rq()
219 : * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
220 : * [S] ->cpu = new_cpu [L] task_rq()
221 : * [L] ->on_rq
222 : * RELEASE (rq->lock)
223 : *
224 : * If we observe the old CPU in task_rq_lock(), the acquire of
225 : * the old rq->lock will fully serialize against the stores.
226 : *
227 : * If we observe the new CPU in task_rq_lock(), the address
228 : * dependency headed by '[L] rq = task_rq()' and the acquire
229 : * will pair with the WMB to ensure we then also see migrating.
230 : */
231 271 : if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
232 271 : rq_pin_lock(rq, rf);
233 271 : return rq;
234 : }
235 0 : raw_spin_unlock(&rq->lock);
236 0 : raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
237 :
238 0 : while (unlikely(task_on_rq_migrating(p)))
239 0 : cpu_relax();
240 : }
241 : }
242 :
243 : /*
244 : * RQ-clock updating methods:
245 : */
246 :
247 72902 : static void update_rq_clock_task(struct rq *rq, s64 delta)
248 : {
249 : /*
250 : * In theory, the compile should just see 0 here, and optimize out the call
251 : * to sched_rt_avg_update. But I don't trust it...
252 : */
253 72902 : s64 __maybe_unused steal = 0, irq_delta = 0;
254 :
255 : #ifdef CONFIG_IRQ_TIME_ACCOUNTING
256 : irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
257 :
258 : /*
259 : * Since irq_time is only updated on {soft,}irq_exit, we might run into
260 : * this case when a previous update_rq_clock() happened inside a
261 : * {soft,}irq region.
262 : *
263 : * When this happens, we stop ->clock_task and only update the
264 : * prev_irq_time stamp to account for the part that fit, so that a next
265 : * update will consume the rest. This ensures ->clock_task is
266 : * monotonic.
267 : *
268 : * It does however cause some slight miss-attribution of {soft,}irq
269 : * time, a more accurate solution would be to update the irq_time using
270 : * the current rq->clock timestamp, except that would require using
271 : * atomic ops.
272 : */
273 : if (irq_delta > delta)
274 : irq_delta = delta;
275 :
276 : rq->prev_irq_time += irq_delta;
277 : delta -= irq_delta;
278 : #endif
279 : #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
280 72902 : if (static_key_false((¶virt_steal_rq_enabled))) {
281 71456 : steal = paravirt_steal_clock(cpu_of(rq));
282 71441 : steal -= rq->prev_steal_time_rq;
283 :
284 71441 : if (unlikely(steal > delta))
285 112 : steal = delta;
286 :
287 71441 : rq->prev_steal_time_rq += steal;
288 71441 : delta -= steal;
289 : }
290 : #endif
291 :
292 72786 : rq->clock_task += delta;
293 :
294 : #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
295 72786 : if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
296 7253 : update_irq_load_avg(rq, irq_delta + steal);
297 : #endif
298 72767 : update_rq_clock_pelt(rq, delta);
299 72939 : }
300 :
301 84069 : void update_rq_clock(struct rq *rq)
302 : {
303 84069 : s64 delta;
304 :
305 169119 : lockdep_assert_held(&rq->lock);
306 :
307 84055 : if (rq->clock_update_flags & RQCF_ACT_SKIP)
308 : return;
309 :
310 : #ifdef CONFIG_SCHED_DEBUG
311 : if (sched_feat(WARN_DOUBLE_CLOCK))
312 : SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
313 : rq->clock_update_flags |= RQCF_UPDATED;
314 : #endif
315 :
316 72942 : delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
317 72950 : if (delta < 0)
318 : return;
319 72950 : rq->clock += delta;
320 72950 : update_rq_clock_task(rq, delta);
321 : }
322 :
323 : #ifdef CONFIG_SCHED_HRTICK
324 : /*
325 : * Use HR-timers to deliver accurate preemption points.
326 : */
327 :
328 : static void hrtick_clear(struct rq *rq)
329 : {
330 : if (hrtimer_active(&rq->hrtick_timer))
331 : hrtimer_cancel(&rq->hrtick_timer);
332 : }
333 :
334 : /*
335 : * High-resolution timer tick.
336 : * Runs from hardirq context with interrupts disabled.
337 : */
338 : static enum hrtimer_restart hrtick(struct hrtimer *timer)
339 : {
340 : struct rq *rq = container_of(timer, struct rq, hrtick_timer);
341 : struct rq_flags rf;
342 :
343 : WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
344 :
345 : rq_lock(rq, &rf);
346 : update_rq_clock(rq);
347 : rq->curr->sched_class->task_tick(rq, rq->curr, 1);
348 : rq_unlock(rq, &rf);
349 :
350 : return HRTIMER_NORESTART;
351 : }
352 :
353 : #ifdef CONFIG_SMP
354 :
355 : static void __hrtick_restart(struct rq *rq)
356 : {
357 : struct hrtimer *timer = &rq->hrtick_timer;
358 : ktime_t time = rq->hrtick_time;
359 :
360 : hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
361 : }
362 :
363 : /*
364 : * called from hardirq (IPI) context
365 : */
366 : static void __hrtick_start(void *arg)
367 : {
368 : struct rq *rq = arg;
369 : struct rq_flags rf;
370 :
371 : rq_lock(rq, &rf);
372 : __hrtick_restart(rq);
373 : rq_unlock(rq, &rf);
374 : }
375 :
376 : /*
377 : * Called to set the hrtick timer state.
378 : *
379 : * called with rq->lock held and irqs disabled
380 : */
381 : void hrtick_start(struct rq *rq, u64 delay)
382 : {
383 : struct hrtimer *timer = &rq->hrtick_timer;
384 : s64 delta;
385 :
386 : /*
387 : * Don't schedule slices shorter than 10000ns, that just
388 : * doesn't make sense and can cause timer DoS.
389 : */
390 : delta = max_t(s64, delay, 10000LL);
391 : rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
392 :
393 : if (rq == this_rq())
394 : __hrtick_restart(rq);
395 : else
396 : smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
397 : }
398 :
399 : #else
400 : /*
401 : * Called to set the hrtick timer state.
402 : *
403 : * called with rq->lock held and irqs disabled
404 : */
405 : void hrtick_start(struct rq *rq, u64 delay)
406 : {
407 : /*
408 : * Don't schedule slices shorter than 10000ns, that just
409 : * doesn't make sense. Rely on vruntime for fairness.
410 : */
411 : delay = max_t(u64, delay, 10000LL);
412 : hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
413 : HRTIMER_MODE_REL_PINNED_HARD);
414 : }
415 :
416 : #endif /* CONFIG_SMP */
417 :
418 : static void hrtick_rq_init(struct rq *rq)
419 : {
420 : #ifdef CONFIG_SMP
421 : INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
422 : #endif
423 : hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
424 : rq->hrtick_timer.function = hrtick;
425 : }
426 : #else /* CONFIG_SCHED_HRTICK */
427 0 : static inline void hrtick_clear(struct rq *rq)
428 : {
429 0 : }
430 :
431 4 : static inline void hrtick_rq_init(struct rq *rq)
432 : {
433 4 : }
434 : #endif /* CONFIG_SCHED_HRTICK */
435 :
436 : /*
437 : * cmpxchg based fetch_or, macro so it works for different integer types
438 : */
439 : #define fetch_or(ptr, mask) \
440 : ({ \
441 : typeof(ptr) _ptr = (ptr); \
442 : typeof(mask) _mask = (mask); \
443 : typeof(*_ptr) _old, _val = *_ptr; \
444 : \
445 : for (;;) { \
446 : _old = cmpxchg(_ptr, _val, _val | _mask); \
447 : if (_old == _val) \
448 : break; \
449 : _val = _old; \
450 : } \
451 : _old; \
452 : })
453 :
454 : #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
455 : /*
456 : * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
457 : * this avoids any races wrt polling state changes and thereby avoids
458 : * spurious IPIs.
459 : */
460 746 : static bool set_nr_and_not_polling(struct task_struct *p)
461 : {
462 746 : struct thread_info *ti = task_thread_info(p);
463 746 : return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
464 : }
465 :
466 : /*
467 : * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
468 : *
469 : * If this returns true, then the idle task promises to call
470 : * sched_ttwu_pending() and reschedule soon.
471 : */
472 7103 : static bool set_nr_if_polling(struct task_struct *p)
473 : {
474 7103 : struct thread_info *ti = task_thread_info(p);
475 7103 : typeof(ti->flags) old, val = READ_ONCE(ti->flags);
476 :
477 7103 : for (;;) {
478 7103 : if (!(val & _TIF_POLLING_NRFLAG))
479 : return false;
480 28 : if (val & _TIF_NEED_RESCHED)
481 : return true;
482 28 : old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
483 28 : if (old == val)
484 : break;
485 : val = old;
486 : }
487 : return true;
488 : }
489 :
490 : #else
491 : static bool set_nr_and_not_polling(struct task_struct *p)
492 : {
493 : set_tsk_need_resched(p);
494 : return true;
495 : }
496 :
497 : #ifdef CONFIG_SMP
498 : static bool set_nr_if_polling(struct task_struct *p)
499 : {
500 : return false;
501 : }
502 : #endif
503 : #endif
504 :
505 620 : static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
506 : {
507 620 : struct wake_q_node *node = &task->wake_q;
508 :
509 : /*
510 : * Atomically grab the task, if ->wake_q is !nil already it means
511 : * it's already queued (either by us or someone else) and will get the
512 : * wakeup due to that.
513 : *
514 : * In order to ensure that a pending wakeup will observe our pending
515 : * state, even in the failed case, an explicit smp_mb() must be used.
516 : */
517 620 : smp_mb__before_atomic();
518 620 : if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
519 : return false;
520 :
521 : /*
522 : * The head is context local, there can be no concurrency.
523 : */
524 620 : *head->lastp = node;
525 620 : head->lastp = &node->next;
526 620 : return true;
527 : }
528 :
529 : /**
530 : * wake_q_add() - queue a wakeup for 'later' waking.
531 : * @head: the wake_q_head to add @task to
532 : * @task: the task to queue for 'later' wakeup
533 : *
534 : * Queue a task for later wakeup, most likely by the wake_up_q() call in the
535 : * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
536 : * instantly.
537 : *
538 : * This function must be used as-if it were wake_up_process(); IOW the task
539 : * must be ready to be woken at this location.
540 : */
541 382 : void wake_q_add(struct wake_q_head *head, struct task_struct *task)
542 : {
543 382 : if (__wake_q_add(head, task))
544 382 : get_task_struct(task);
545 382 : }
546 :
547 : /**
548 : * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
549 : * @head: the wake_q_head to add @task to
550 : * @task: the task to queue for 'later' wakeup
551 : *
552 : * Queue a task for later wakeup, most likely by the wake_up_q() call in the
553 : * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
554 : * instantly.
555 : *
556 : * This function must be used as-if it were wake_up_process(); IOW the task
557 : * must be ready to be woken at this location.
558 : *
559 : * This function is essentially a task-safe equivalent to wake_q_add(). Callers
560 : * that already hold reference to @task can call the 'safe' version and trust
561 : * wake_q to do the right thing depending whether or not the @task is already
562 : * queued for wakeup.
563 : */
564 238 : void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
565 : {
566 238 : if (!__wake_q_add(head, task))
567 0 : put_task_struct(task);
568 238 : }
569 :
570 703 : void wake_up_q(struct wake_q_head *head)
571 : {
572 703 : struct wake_q_node *node = head->first;
573 :
574 1323 : while (node != WAKE_Q_TAIL) {
575 620 : struct task_struct *task;
576 :
577 620 : task = container_of(node, struct task_struct, wake_q);
578 620 : BUG_ON(!task);
579 : /* Task can safely be re-inserted now: */
580 620 : node = node->next;
581 620 : task->wake_q.next = NULL;
582 :
583 : /*
584 : * wake_up_process() executes a full barrier, which pairs with
585 : * the queueing in wake_q_add() so as not to miss wakeups.
586 : */
587 1240 : wake_up_process(task);
588 620 : put_task_struct(task);
589 : }
590 703 : }
591 :
592 : /*
593 : * resched_curr - mark rq's current task 'to be rescheduled now'.
594 : *
595 : * On UP this means the setting of the need_resched flag, on SMP it
596 : * might also involve a cross-CPU call to trigger the scheduler on
597 : * the target CPU.
598 : */
599 13562 : void resched_curr(struct rq *rq)
600 : {
601 13562 : struct task_struct *curr = rq->curr;
602 13562 : int cpu;
603 :
604 27140 : lockdep_assert_held(&rq->lock);
605 :
606 13569 : if (test_tsk_need_resched(curr))
607 : return;
608 :
609 12600 : cpu = cpu_of(rq);
610 :
611 12600 : if (cpu == smp_processor_id()) {
612 11862 : set_tsk_need_resched(curr);
613 11868 : set_preempt_need_resched();
614 11868 : return;
615 : }
616 :
617 738 : if (set_nr_and_not_polling(curr))
618 738 : smp_send_reschedule(cpu);
619 : else
620 0 : trace_sched_wake_idle_without_ipi(cpu);
621 : }
622 :
623 0 : void resched_cpu(int cpu)
624 : {
625 0 : struct rq *rq = cpu_rq(cpu);
626 0 : unsigned long flags;
627 :
628 0 : raw_spin_lock_irqsave(&rq->lock, flags);
629 0 : if (cpu_online(cpu) || cpu == smp_processor_id())
630 0 : resched_curr(rq);
631 0 : raw_spin_unlock_irqrestore(&rq->lock, flags);
632 0 : }
633 :
634 : #ifdef CONFIG_SMP
635 : #ifdef CONFIG_NO_HZ_COMMON
636 : /*
637 : * In the semi idle case, use the nearest busy CPU for migrating timers
638 : * from an idle CPU. This is good for power-savings.
639 : *
640 : * We don't do similar optimization for completely idle system, as
641 : * selecting an idle CPU will add more delays to the timers than intended
642 : * (as that CPU's timer base may not be uptodate wrt jiffies etc).
643 : */
644 5340 : int get_nohz_timer_target(void)
645 : {
646 5340 : int i, cpu = smp_processor_id(), default_cpu = -1;
647 5340 : struct sched_domain *sd;
648 :
649 5340 : if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
650 5340 : if (!idle_cpu(cpu))
651 : return cpu;
652 53 : default_cpu = cpu;
653 : }
654 :
655 53 : rcu_read_lock();
656 167 : for_each_domain(cpu, sd) {
657 115 : for_each_cpu_and(i, sched_domain_span(sd),
658 : housekeeping_cpumask(HK_FLAG_TIMER)) {
659 107 : if (cpu == i)
660 22 : continue;
661 :
662 85 : if (!idle_cpu(i)) {
663 45 : cpu = i;
664 45 : goto unlock;
665 : }
666 : }
667 : }
668 :
669 8 : if (default_cpu == -1)
670 0 : default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
671 : cpu = default_cpu;
672 53 : unlock:
673 53 : rcu_read_unlock();
674 53 : return cpu;
675 : }
676 :
677 : /*
678 : * When add_timer_on() enqueues a timer into the timer wheel of an
679 : * idle CPU then this timer might expire before the next timer event
680 : * which is scheduled to wake up that CPU. In case of a completely
681 : * idle system the next event might even be infinite time into the
682 : * future. wake_up_idle_cpu() ensures that the CPU is woken up and
683 : * leaves the inner idle loop so the newly added timer is taken into
684 : * account when the CPU goes back to idle and evaluates the timer
685 : * wheel for the next timer event.
686 : */
687 9 : static void wake_up_idle_cpu(int cpu)
688 : {
689 9 : struct rq *rq = cpu_rq(cpu);
690 :
691 9 : if (cpu == smp_processor_id())
692 : return;
693 :
694 8 : if (set_nr_and_not_polling(rq->idle))
695 8 : smp_send_reschedule(cpu);
696 : else
697 0 : trace_sched_wake_idle_without_ipi(cpu);
698 : }
699 :
700 9 : static bool wake_up_full_nohz_cpu(int cpu)
701 : {
702 : /*
703 : * We just need the target to call irq_exit() and re-evaluate
704 : * the next tick. The nohz full kick at least implies that.
705 : * If needed we can still optimize that later with an
706 : * empty IRQ.
707 : */
708 9 : if (cpu_is_offline(cpu))
709 0 : return true; /* Don't try to wake offline CPUs. */
710 9 : if (tick_nohz_full_cpu(cpu)) {
711 : if (cpu != smp_processor_id() ||
712 : tick_nohz_tick_stopped())
713 : tick_nohz_full_kick_cpu(cpu);
714 : return true;
715 : }
716 :
717 : return false;
718 : }
719 :
720 : /*
721 : * Wake up the specified CPU. If the CPU is going offline, it is the
722 : * caller's responsibility to deal with the lost wakeup, for example,
723 : * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
724 : */
725 9 : void wake_up_nohz_cpu(int cpu)
726 : {
727 9 : if (!wake_up_full_nohz_cpu(cpu))
728 9 : wake_up_idle_cpu(cpu);
729 9 : }
730 :
731 220 : static void nohz_csd_func(void *info)
732 : {
733 220 : struct rq *rq = info;
734 220 : int cpu = cpu_of(rq);
735 220 : unsigned int flags;
736 :
737 : /*
738 : * Release the rq::nohz_csd.
739 : */
740 220 : flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
741 220 : WARN_ON(!(flags & NOHZ_KICK_MASK));
742 :
743 220 : rq->idle_balance = idle_cpu(cpu);
744 431 : if (rq->idle_balance && !need_resched()) {
745 211 : rq->nohz_idle_balance = flags;
746 211 : raise_softirq_irqoff(SCHED_SOFTIRQ);
747 : }
748 220 : }
749 :
750 : #endif /* CONFIG_NO_HZ_COMMON */
751 :
752 : #ifdef CONFIG_NO_HZ_FULL
753 : bool sched_can_stop_tick(struct rq *rq)
754 : {
755 : int fifo_nr_running;
756 :
757 : /* Deadline tasks, even if single, need the tick */
758 : if (rq->dl.dl_nr_running)
759 : return false;
760 :
761 : /*
762 : * If there are more than one RR tasks, we need the tick to affect the
763 : * actual RR behaviour.
764 : */
765 : if (rq->rt.rr_nr_running) {
766 : if (rq->rt.rr_nr_running == 1)
767 : return true;
768 : else
769 : return false;
770 : }
771 :
772 : /*
773 : * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
774 : * forced preemption between FIFO tasks.
775 : */
776 : fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
777 : if (fifo_nr_running)
778 : return true;
779 :
780 : /*
781 : * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
782 : * if there's more than one we need the tick for involuntary
783 : * preemption.
784 : */
785 : if (rq->nr_running > 1)
786 : return false;
787 :
788 : return true;
789 : }
790 : #endif /* CONFIG_NO_HZ_FULL */
791 : #endif /* CONFIG_SMP */
792 :
793 : #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
794 : (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
795 : /*
796 : * Iterate task_group tree rooted at *from, calling @down when first entering a
797 : * node and @up when leaving it for the final time.
798 : *
799 : * Caller must hold rcu_lock or sufficient equivalent.
800 : */
801 : int walk_tg_tree_from(struct task_group *from,
802 : tg_visitor down, tg_visitor up, void *data)
803 : {
804 : struct task_group *parent, *child;
805 : int ret;
806 :
807 : parent = from;
808 :
809 : down:
810 : ret = (*down)(parent, data);
811 : if (ret)
812 : goto out;
813 : list_for_each_entry_rcu(child, &parent->children, siblings) {
814 : parent = child;
815 : goto down;
816 :
817 : up:
818 : continue;
819 : }
820 : ret = (*up)(parent, data);
821 : if (ret || parent == from)
822 : goto out;
823 :
824 : child = parent;
825 : parent = parent->parent;
826 : if (parent)
827 : goto up;
828 : out:
829 : return ret;
830 : }
831 :
832 : int tg_nop(struct task_group *tg, void *data)
833 : {
834 : return 0;
835 : }
836 : #endif
837 :
838 26 : static void set_load_weight(struct task_struct *p, bool update_load)
839 : {
840 26 : int prio = p->static_prio - MAX_RT_PRIO;
841 26 : struct load_weight *load = &p->se.load;
842 :
843 : /*
844 : * SCHED_IDLE tasks get minimal weight:
845 : */
846 26 : if (task_has_idle_policy(p)) {
847 0 : load->weight = scale_load(WEIGHT_IDLEPRIO);
848 0 : load->inv_weight = WMULT_IDLEPRIO;
849 0 : return;
850 : }
851 :
852 : /*
853 : * SCHED_OTHER tasks have to update their load when changing their
854 : * weight
855 : */
856 26 : if (update_load && p->sched_class == &fair_sched_class) {
857 25 : reweight_task(p, prio);
858 : } else {
859 1 : load->weight = scale_load(sched_prio_to_weight[prio]);
860 1 : load->inv_weight = sched_prio_to_wmult[prio];
861 : }
862 : }
863 :
864 : #ifdef CONFIG_UCLAMP_TASK
865 : /*
866 : * Serializes updates of utilization clamp values
867 : *
868 : * The (slow-path) user-space triggers utilization clamp value updates which
869 : * can require updates on (fast-path) scheduler's data structures used to
870 : * support enqueue/dequeue operations.
871 : * While the per-CPU rq lock protects fast-path update operations, user-space
872 : * requests are serialized using a mutex to reduce the risk of conflicting
873 : * updates or API abuses.
874 : */
875 : static DEFINE_MUTEX(uclamp_mutex);
876 :
877 : /* Max allowed minimum utilization */
878 : unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
879 :
880 : /* Max allowed maximum utilization */
881 : unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
882 :
883 : /*
884 : * By default RT tasks run at the maximum performance point/capacity of the
885 : * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
886 : * SCHED_CAPACITY_SCALE.
887 : *
888 : * This knob allows admins to change the default behavior when uclamp is being
889 : * used. In battery powered devices, particularly, running at the maximum
890 : * capacity and frequency will increase energy consumption and shorten the
891 : * battery life.
892 : *
893 : * This knob only affects RT tasks that their uclamp_se->user_defined == false.
894 : *
895 : * This knob will not override the system default sched_util_clamp_min defined
896 : * above.
897 : */
898 : unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
899 :
900 : /* All clamps are required to be less or equal than these values */
901 : static struct uclamp_se uclamp_default[UCLAMP_CNT];
902 :
903 : /*
904 : * This static key is used to reduce the uclamp overhead in the fast path. It
905 : * primarily disables the call to uclamp_rq_{inc, dec}() in
906 : * enqueue/dequeue_task().
907 : *
908 : * This allows users to continue to enable uclamp in their kernel config with
909 : * minimum uclamp overhead in the fast path.
910 : *
911 : * As soon as userspace modifies any of the uclamp knobs, the static key is
912 : * enabled, since we have an actual users that make use of uclamp
913 : * functionality.
914 : *
915 : * The knobs that would enable this static key are:
916 : *
917 : * * A task modifying its uclamp value with sched_setattr().
918 : * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
919 : * * An admin modifying the cgroup cpu.uclamp.{min, max}
920 : */
921 : DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
922 :
923 : /* Integer rounded range for each bucket */
924 : #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
925 :
926 : #define for_each_clamp_id(clamp_id) \
927 : for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
928 :
929 : static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
930 : {
931 : return clamp_value / UCLAMP_BUCKET_DELTA;
932 : }
933 :
934 : static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
935 : {
936 : if (clamp_id == UCLAMP_MIN)
937 : return 0;
938 : return SCHED_CAPACITY_SCALE;
939 : }
940 :
941 : static inline void uclamp_se_set(struct uclamp_se *uc_se,
942 : unsigned int value, bool user_defined)
943 : {
944 : uc_se->value = value;
945 : uc_se->bucket_id = uclamp_bucket_id(value);
946 : uc_se->user_defined = user_defined;
947 : }
948 :
949 : static inline unsigned int
950 : uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
951 : unsigned int clamp_value)
952 : {
953 : /*
954 : * Avoid blocked utilization pushing up the frequency when we go
955 : * idle (which drops the max-clamp) by retaining the last known
956 : * max-clamp.
957 : */
958 : if (clamp_id == UCLAMP_MAX) {
959 : rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
960 : return clamp_value;
961 : }
962 :
963 : return uclamp_none(UCLAMP_MIN);
964 : }
965 :
966 : static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
967 : unsigned int clamp_value)
968 : {
969 : /* Reset max-clamp retention only on idle exit */
970 : if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
971 : return;
972 :
973 : WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
974 : }
975 :
976 : static inline
977 : unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
978 : unsigned int clamp_value)
979 : {
980 : struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
981 : int bucket_id = UCLAMP_BUCKETS - 1;
982 :
983 : /*
984 : * Since both min and max clamps are max aggregated, find the
985 : * top most bucket with tasks in.
986 : */
987 : for ( ; bucket_id >= 0; bucket_id--) {
988 : if (!bucket[bucket_id].tasks)
989 : continue;
990 : return bucket[bucket_id].value;
991 : }
992 :
993 : /* No tasks -- default clamp values */
994 : return uclamp_idle_value(rq, clamp_id, clamp_value);
995 : }
996 :
997 : static void __uclamp_update_util_min_rt_default(struct task_struct *p)
998 : {
999 : unsigned int default_util_min;
1000 : struct uclamp_se *uc_se;
1001 :
1002 : lockdep_assert_held(&p->pi_lock);
1003 :
1004 : uc_se = &p->uclamp_req[UCLAMP_MIN];
1005 :
1006 : /* Only sync if user didn't override the default */
1007 : if (uc_se->user_defined)
1008 : return;
1009 :
1010 : default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1011 : uclamp_se_set(uc_se, default_util_min, false);
1012 : }
1013 :
1014 : static void uclamp_update_util_min_rt_default(struct task_struct *p)
1015 : {
1016 : struct rq_flags rf;
1017 : struct rq *rq;
1018 :
1019 : if (!rt_task(p))
1020 : return;
1021 :
1022 : /* Protect updates to p->uclamp_* */
1023 : rq = task_rq_lock(p, &rf);
1024 : __uclamp_update_util_min_rt_default(p);
1025 : task_rq_unlock(rq, p, &rf);
1026 : }
1027 :
1028 : static void uclamp_sync_util_min_rt_default(void)
1029 : {
1030 : struct task_struct *g, *p;
1031 :
1032 : /*
1033 : * copy_process() sysctl_uclamp
1034 : * uclamp_min_rt = X;
1035 : * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1036 : * // link thread smp_mb__after_spinlock()
1037 : * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1038 : * sched_post_fork() for_each_process_thread()
1039 : * __uclamp_sync_rt() __uclamp_sync_rt()
1040 : *
1041 : * Ensures that either sched_post_fork() will observe the new
1042 : * uclamp_min_rt or for_each_process_thread() will observe the new
1043 : * task.
1044 : */
1045 : read_lock(&tasklist_lock);
1046 : smp_mb__after_spinlock();
1047 : read_unlock(&tasklist_lock);
1048 :
1049 : rcu_read_lock();
1050 : for_each_process_thread(g, p)
1051 : uclamp_update_util_min_rt_default(p);
1052 : rcu_read_unlock();
1053 : }
1054 :
1055 : static inline struct uclamp_se
1056 : uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1057 : {
1058 : struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1059 : #ifdef CONFIG_UCLAMP_TASK_GROUP
1060 : struct uclamp_se uc_max;
1061 :
1062 : /*
1063 : * Tasks in autogroups or root task group will be
1064 : * restricted by system defaults.
1065 : */
1066 : if (task_group_is_autogroup(task_group(p)))
1067 : return uc_req;
1068 : if (task_group(p) == &root_task_group)
1069 : return uc_req;
1070 :
1071 : uc_max = task_group(p)->uclamp[clamp_id];
1072 : if (uc_req.value > uc_max.value || !uc_req.user_defined)
1073 : return uc_max;
1074 : #endif
1075 :
1076 : return uc_req;
1077 : }
1078 :
1079 : /*
1080 : * The effective clamp bucket index of a task depends on, by increasing
1081 : * priority:
1082 : * - the task specific clamp value, when explicitly requested from userspace
1083 : * - the task group effective clamp value, for tasks not either in the root
1084 : * group or in an autogroup
1085 : * - the system default clamp value, defined by the sysadmin
1086 : */
1087 : static inline struct uclamp_se
1088 : uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1089 : {
1090 : struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1091 : struct uclamp_se uc_max = uclamp_default[clamp_id];
1092 :
1093 : /* System default restrictions always apply */
1094 : if (unlikely(uc_req.value > uc_max.value))
1095 : return uc_max;
1096 :
1097 : return uc_req;
1098 : }
1099 :
1100 : unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1101 : {
1102 : struct uclamp_se uc_eff;
1103 :
1104 : /* Task currently refcounted: use back-annotated (effective) value */
1105 : if (p->uclamp[clamp_id].active)
1106 : return (unsigned long)p->uclamp[clamp_id].value;
1107 :
1108 : uc_eff = uclamp_eff_get(p, clamp_id);
1109 :
1110 : return (unsigned long)uc_eff.value;
1111 : }
1112 :
1113 : /*
1114 : * When a task is enqueued on a rq, the clamp bucket currently defined by the
1115 : * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1116 : * updates the rq's clamp value if required.
1117 : *
1118 : * Tasks can have a task-specific value requested from user-space, track
1119 : * within each bucket the maximum value for tasks refcounted in it.
1120 : * This "local max aggregation" allows to track the exact "requested" value
1121 : * for each bucket when all its RUNNABLE tasks require the same clamp.
1122 : */
1123 : static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1124 : enum uclamp_id clamp_id)
1125 : {
1126 : struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1127 : struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1128 : struct uclamp_bucket *bucket;
1129 :
1130 : lockdep_assert_held(&rq->lock);
1131 :
1132 : /* Update task effective clamp */
1133 : p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1134 :
1135 : bucket = &uc_rq->bucket[uc_se->bucket_id];
1136 : bucket->tasks++;
1137 : uc_se->active = true;
1138 :
1139 : uclamp_idle_reset(rq, clamp_id, uc_se->value);
1140 :
1141 : /*
1142 : * Local max aggregation: rq buckets always track the max
1143 : * "requested" clamp value of its RUNNABLE tasks.
1144 : */
1145 : if (bucket->tasks == 1 || uc_se->value > bucket->value)
1146 : bucket->value = uc_se->value;
1147 :
1148 : if (uc_se->value > READ_ONCE(uc_rq->value))
1149 : WRITE_ONCE(uc_rq->value, uc_se->value);
1150 : }
1151 :
1152 : /*
1153 : * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1154 : * is released. If this is the last task reference counting the rq's max
1155 : * active clamp value, then the rq's clamp value is updated.
1156 : *
1157 : * Both refcounted tasks and rq's cached clamp values are expected to be
1158 : * always valid. If it's detected they are not, as defensive programming,
1159 : * enforce the expected state and warn.
1160 : */
1161 : static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1162 : enum uclamp_id clamp_id)
1163 : {
1164 : struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1165 : struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1166 : struct uclamp_bucket *bucket;
1167 : unsigned int bkt_clamp;
1168 : unsigned int rq_clamp;
1169 :
1170 : lockdep_assert_held(&rq->lock);
1171 :
1172 : /*
1173 : * If sched_uclamp_used was enabled after task @p was enqueued,
1174 : * we could end up with unbalanced call to uclamp_rq_dec_id().
1175 : *
1176 : * In this case the uc_se->active flag should be false since no uclamp
1177 : * accounting was performed at enqueue time and we can just return
1178 : * here.
1179 : *
1180 : * Need to be careful of the following enqueue/dequeue ordering
1181 : * problem too
1182 : *
1183 : * enqueue(taskA)
1184 : * // sched_uclamp_used gets enabled
1185 : * enqueue(taskB)
1186 : * dequeue(taskA)
1187 : * // Must not decrement bucket->tasks here
1188 : * dequeue(taskB)
1189 : *
1190 : * where we could end up with stale data in uc_se and
1191 : * bucket[uc_se->bucket_id].
1192 : *
1193 : * The following check here eliminates the possibility of such race.
1194 : */
1195 : if (unlikely(!uc_se->active))
1196 : return;
1197 :
1198 : bucket = &uc_rq->bucket[uc_se->bucket_id];
1199 :
1200 : SCHED_WARN_ON(!bucket->tasks);
1201 : if (likely(bucket->tasks))
1202 : bucket->tasks--;
1203 :
1204 : uc_se->active = false;
1205 :
1206 : /*
1207 : * Keep "local max aggregation" simple and accept to (possibly)
1208 : * overboost some RUNNABLE tasks in the same bucket.
1209 : * The rq clamp bucket value is reset to its base value whenever
1210 : * there are no more RUNNABLE tasks refcounting it.
1211 : */
1212 : if (likely(bucket->tasks))
1213 : return;
1214 :
1215 : rq_clamp = READ_ONCE(uc_rq->value);
1216 : /*
1217 : * Defensive programming: this should never happen. If it happens,
1218 : * e.g. due to future modification, warn and fixup the expected value.
1219 : */
1220 : SCHED_WARN_ON(bucket->value > rq_clamp);
1221 : if (bucket->value >= rq_clamp) {
1222 : bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1223 : WRITE_ONCE(uc_rq->value, bkt_clamp);
1224 : }
1225 : }
1226 :
1227 : static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1228 : {
1229 : enum uclamp_id clamp_id;
1230 :
1231 : /*
1232 : * Avoid any overhead until uclamp is actually used by the userspace.
1233 : *
1234 : * The condition is constructed such that a NOP is generated when
1235 : * sched_uclamp_used is disabled.
1236 : */
1237 : if (!static_branch_unlikely(&sched_uclamp_used))
1238 : return;
1239 :
1240 : if (unlikely(!p->sched_class->uclamp_enabled))
1241 : return;
1242 :
1243 : for_each_clamp_id(clamp_id)
1244 : uclamp_rq_inc_id(rq, p, clamp_id);
1245 :
1246 : /* Reset clamp idle holding when there is one RUNNABLE task */
1247 : if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1248 : rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1249 : }
1250 :
1251 : static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1252 : {
1253 : enum uclamp_id clamp_id;
1254 :
1255 : /*
1256 : * Avoid any overhead until uclamp is actually used by the userspace.
1257 : *
1258 : * The condition is constructed such that a NOP is generated when
1259 : * sched_uclamp_used is disabled.
1260 : */
1261 : if (!static_branch_unlikely(&sched_uclamp_used))
1262 : return;
1263 :
1264 : if (unlikely(!p->sched_class->uclamp_enabled))
1265 : return;
1266 :
1267 : for_each_clamp_id(clamp_id)
1268 : uclamp_rq_dec_id(rq, p, clamp_id);
1269 : }
1270 :
1271 : static inline void
1272 : uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1273 : {
1274 : struct rq_flags rf;
1275 : struct rq *rq;
1276 :
1277 : /*
1278 : * Lock the task and the rq where the task is (or was) queued.
1279 : *
1280 : * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1281 : * price to pay to safely serialize util_{min,max} updates with
1282 : * enqueues, dequeues and migration operations.
1283 : * This is the same locking schema used by __set_cpus_allowed_ptr().
1284 : */
1285 : rq = task_rq_lock(p, &rf);
1286 :
1287 : /*
1288 : * Setting the clamp bucket is serialized by task_rq_lock().
1289 : * If the task is not yet RUNNABLE and its task_struct is not
1290 : * affecting a valid clamp bucket, the next time it's enqueued,
1291 : * it will already see the updated clamp bucket value.
1292 : */
1293 : if (p->uclamp[clamp_id].active) {
1294 : uclamp_rq_dec_id(rq, p, clamp_id);
1295 : uclamp_rq_inc_id(rq, p, clamp_id);
1296 : }
1297 :
1298 : task_rq_unlock(rq, p, &rf);
1299 : }
1300 :
1301 : #ifdef CONFIG_UCLAMP_TASK_GROUP
1302 : static inline void
1303 : uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1304 : unsigned int clamps)
1305 : {
1306 : enum uclamp_id clamp_id;
1307 : struct css_task_iter it;
1308 : struct task_struct *p;
1309 :
1310 : css_task_iter_start(css, 0, &it);
1311 : while ((p = css_task_iter_next(&it))) {
1312 : for_each_clamp_id(clamp_id) {
1313 : if ((0x1 << clamp_id) & clamps)
1314 : uclamp_update_active(p, clamp_id);
1315 : }
1316 : }
1317 : css_task_iter_end(&it);
1318 : }
1319 :
1320 : static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1321 : static void uclamp_update_root_tg(void)
1322 : {
1323 : struct task_group *tg = &root_task_group;
1324 :
1325 : uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1326 : sysctl_sched_uclamp_util_min, false);
1327 : uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1328 : sysctl_sched_uclamp_util_max, false);
1329 :
1330 : rcu_read_lock();
1331 : cpu_util_update_eff(&root_task_group.css);
1332 : rcu_read_unlock();
1333 : }
1334 : #else
1335 : static void uclamp_update_root_tg(void) { }
1336 : #endif
1337 :
1338 : int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1339 : void *buffer, size_t *lenp, loff_t *ppos)
1340 : {
1341 : bool update_root_tg = false;
1342 : int old_min, old_max, old_min_rt;
1343 : int result;
1344 :
1345 : mutex_lock(&uclamp_mutex);
1346 : old_min = sysctl_sched_uclamp_util_min;
1347 : old_max = sysctl_sched_uclamp_util_max;
1348 : old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1349 :
1350 : result = proc_dointvec(table, write, buffer, lenp, ppos);
1351 : if (result)
1352 : goto undo;
1353 : if (!write)
1354 : goto done;
1355 :
1356 : if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1357 : sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1358 : sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1359 :
1360 : result = -EINVAL;
1361 : goto undo;
1362 : }
1363 :
1364 : if (old_min != sysctl_sched_uclamp_util_min) {
1365 : uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1366 : sysctl_sched_uclamp_util_min, false);
1367 : update_root_tg = true;
1368 : }
1369 : if (old_max != sysctl_sched_uclamp_util_max) {
1370 : uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1371 : sysctl_sched_uclamp_util_max, false);
1372 : update_root_tg = true;
1373 : }
1374 :
1375 : if (update_root_tg) {
1376 : static_branch_enable(&sched_uclamp_used);
1377 : uclamp_update_root_tg();
1378 : }
1379 :
1380 : if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1381 : static_branch_enable(&sched_uclamp_used);
1382 : uclamp_sync_util_min_rt_default();
1383 : }
1384 :
1385 : /*
1386 : * We update all RUNNABLE tasks only when task groups are in use.
1387 : * Otherwise, keep it simple and do just a lazy update at each next
1388 : * task enqueue time.
1389 : */
1390 :
1391 : goto done;
1392 :
1393 : undo:
1394 : sysctl_sched_uclamp_util_min = old_min;
1395 : sysctl_sched_uclamp_util_max = old_max;
1396 : sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1397 : done:
1398 : mutex_unlock(&uclamp_mutex);
1399 :
1400 : return result;
1401 : }
1402 :
1403 : static int uclamp_validate(struct task_struct *p,
1404 : const struct sched_attr *attr)
1405 : {
1406 : int util_min = p->uclamp_req[UCLAMP_MIN].value;
1407 : int util_max = p->uclamp_req[UCLAMP_MAX].value;
1408 :
1409 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1410 : util_min = attr->sched_util_min;
1411 :
1412 : if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1413 : return -EINVAL;
1414 : }
1415 :
1416 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1417 : util_max = attr->sched_util_max;
1418 :
1419 : if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1420 : return -EINVAL;
1421 : }
1422 :
1423 : if (util_min != -1 && util_max != -1 && util_min > util_max)
1424 : return -EINVAL;
1425 :
1426 : /*
1427 : * We have valid uclamp attributes; make sure uclamp is enabled.
1428 : *
1429 : * We need to do that here, because enabling static branches is a
1430 : * blocking operation which obviously cannot be done while holding
1431 : * scheduler locks.
1432 : */
1433 : static_branch_enable(&sched_uclamp_used);
1434 :
1435 : return 0;
1436 : }
1437 :
1438 : static bool uclamp_reset(const struct sched_attr *attr,
1439 : enum uclamp_id clamp_id,
1440 : struct uclamp_se *uc_se)
1441 : {
1442 : /* Reset on sched class change for a non user-defined clamp value. */
1443 : if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1444 : !uc_se->user_defined)
1445 : return true;
1446 :
1447 : /* Reset on sched_util_{min,max} == -1. */
1448 : if (clamp_id == UCLAMP_MIN &&
1449 : attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1450 : attr->sched_util_min == -1) {
1451 : return true;
1452 : }
1453 :
1454 : if (clamp_id == UCLAMP_MAX &&
1455 : attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1456 : attr->sched_util_max == -1) {
1457 : return true;
1458 : }
1459 :
1460 : return false;
1461 : }
1462 :
1463 : static void __setscheduler_uclamp(struct task_struct *p,
1464 : const struct sched_attr *attr)
1465 : {
1466 : enum uclamp_id clamp_id;
1467 :
1468 : for_each_clamp_id(clamp_id) {
1469 : struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1470 : unsigned int value;
1471 :
1472 : if (!uclamp_reset(attr, clamp_id, uc_se))
1473 : continue;
1474 :
1475 : /*
1476 : * RT by default have a 100% boost value that could be modified
1477 : * at runtime.
1478 : */
1479 : if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1480 : value = sysctl_sched_uclamp_util_min_rt_default;
1481 : else
1482 : value = uclamp_none(clamp_id);
1483 :
1484 : uclamp_se_set(uc_se, value, false);
1485 :
1486 : }
1487 :
1488 : if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1489 : return;
1490 :
1491 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1492 : attr->sched_util_min != -1) {
1493 : uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1494 : attr->sched_util_min, true);
1495 : }
1496 :
1497 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1498 : attr->sched_util_max != -1) {
1499 : uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1500 : attr->sched_util_max, true);
1501 : }
1502 : }
1503 :
1504 : static void uclamp_fork(struct task_struct *p)
1505 : {
1506 : enum uclamp_id clamp_id;
1507 :
1508 : /*
1509 : * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1510 : * as the task is still at its early fork stages.
1511 : */
1512 : for_each_clamp_id(clamp_id)
1513 : p->uclamp[clamp_id].active = false;
1514 :
1515 : if (likely(!p->sched_reset_on_fork))
1516 : return;
1517 :
1518 : for_each_clamp_id(clamp_id) {
1519 : uclamp_se_set(&p->uclamp_req[clamp_id],
1520 : uclamp_none(clamp_id), false);
1521 : }
1522 : }
1523 :
1524 : static void uclamp_post_fork(struct task_struct *p)
1525 : {
1526 : uclamp_update_util_min_rt_default(p);
1527 : }
1528 :
1529 : static void __init init_uclamp_rq(struct rq *rq)
1530 : {
1531 : enum uclamp_id clamp_id;
1532 : struct uclamp_rq *uc_rq = rq->uclamp;
1533 :
1534 : for_each_clamp_id(clamp_id) {
1535 : uc_rq[clamp_id] = (struct uclamp_rq) {
1536 : .value = uclamp_none(clamp_id)
1537 : };
1538 : }
1539 :
1540 : rq->uclamp_flags = 0;
1541 : }
1542 :
1543 : static void __init init_uclamp(void)
1544 : {
1545 : struct uclamp_se uc_max = {};
1546 : enum uclamp_id clamp_id;
1547 : int cpu;
1548 :
1549 : for_each_possible_cpu(cpu)
1550 : init_uclamp_rq(cpu_rq(cpu));
1551 :
1552 : for_each_clamp_id(clamp_id) {
1553 : uclamp_se_set(&init_task.uclamp_req[clamp_id],
1554 : uclamp_none(clamp_id), false);
1555 : }
1556 :
1557 : /* System defaults allow max clamp values for both indexes */
1558 : uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1559 : for_each_clamp_id(clamp_id) {
1560 : uclamp_default[clamp_id] = uc_max;
1561 : #ifdef CONFIG_UCLAMP_TASK_GROUP
1562 : root_task_group.uclamp_req[clamp_id] = uc_max;
1563 : root_task_group.uclamp[clamp_id] = uc_max;
1564 : #endif
1565 : }
1566 : }
1567 :
1568 : #else /* CONFIG_UCLAMP_TASK */
1569 15735 : static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1570 15739 : static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1571 : static inline int uclamp_validate(struct task_struct *p,
1572 : const struct sched_attr *attr)
1573 : {
1574 : return -EOPNOTSUPP;
1575 : }
1576 4 : static void __setscheduler_uclamp(struct task_struct *p,
1577 4 : const struct sched_attr *attr) { }
1578 980 : static inline void uclamp_fork(struct task_struct *p) { }
1579 980 : static inline void uclamp_post_fork(struct task_struct *p) { }
1580 1 : static inline void init_uclamp(void) { }
1581 : #endif /* CONFIG_UCLAMP_TASK */
1582 :
1583 15735 : static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1584 : {
1585 15735 : if (!(flags & ENQUEUE_NOCLOCK))
1586 31 : update_rq_clock(rq);
1587 :
1588 15735 : if (!(flags & ENQUEUE_RESTORE)) {
1589 15723 : sched_info_queued(rq, p);
1590 15723 : psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1591 : }
1592 :
1593 15735 : uclamp_rq_inc(rq, p);
1594 15735 : p->sched_class->enqueue_task(rq, p, flags);
1595 15737 : }
1596 :
1597 15738 : static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1598 : {
1599 15738 : if (!(flags & DEQUEUE_NOCLOCK))
1600 0 : update_rq_clock(rq);
1601 :
1602 15738 : if (!(flags & DEQUEUE_SAVE)) {
1603 15722 : sched_info_dequeued(rq, p);
1604 15722 : psi_dequeue(p, flags & DEQUEUE_SLEEP);
1605 : }
1606 :
1607 15739 : uclamp_rq_dec(rq, p);
1608 15739 : p->sched_class->dequeue_task(rq, p, flags);
1609 15737 : }
1610 :
1611 15720 : void activate_task(struct rq *rq, struct task_struct *p, int flags)
1612 : {
1613 853 : enqueue_task(rq, p, flags);
1614 :
1615 15722 : p->on_rq = TASK_ON_RQ_QUEUED;
1616 853 : }
1617 :
1618 15722 : void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1619 : {
1620 853 : p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1621 :
1622 853 : dequeue_task(rq, p, flags);
1623 854 : }
1624 :
1625 : /*
1626 : * __normal_prio - return the priority that is based on the static prio
1627 : */
1628 21 : static inline int __normal_prio(struct task_struct *p)
1629 : {
1630 21 : return p->static_prio;
1631 : }
1632 :
1633 : /*
1634 : * Calculate the expected normal priority: i.e. priority
1635 : * without taking RT-inheritance into account. Might be
1636 : * boosted by interactivity modifiers. Changes upon fork,
1637 : * setprio syscalls, and whenever the interactivity
1638 : * estimator recalculates.
1639 : */
1640 29 : static inline int normal_prio(struct task_struct *p)
1641 : {
1642 29 : int prio;
1643 :
1644 29 : if (task_has_dl_policy(p))
1645 : prio = MAX_DL_PRIO-1;
1646 29 : else if (task_has_rt_policy(p))
1647 8 : prio = MAX_RT_PRIO-1 - p->rt_priority;
1648 : else
1649 21 : prio = __normal_prio(p);
1650 29 : return prio;
1651 : }
1652 :
1653 : /*
1654 : * Calculate the current priority, i.e. the priority
1655 : * taken into account by the scheduler. This value might
1656 : * be boosted by RT tasks, or might be boosted by
1657 : * interactivity modifiers. Will be RT if the task got
1658 : * RT-boosted. If not then it returns p->normal_prio.
1659 : */
1660 21 : static int effective_prio(struct task_struct *p)
1661 : {
1662 21 : p->normal_prio = normal_prio(p);
1663 : /*
1664 : * If we are RT tasks or we were boosted to RT priority,
1665 : * keep the priority unchanged. Otherwise, update priority
1666 : * to the normal priority:
1667 : */
1668 21 : if (!rt_prio(p->prio))
1669 21 : return p->normal_prio;
1670 : return p->prio;
1671 : }
1672 :
1673 : /**
1674 : * task_curr - is this task currently executing on a CPU?
1675 : * @p: the task in question.
1676 : *
1677 : * Return: 1 if the task is currently executing. 0 otherwise.
1678 : */
1679 546 : inline int task_curr(const struct task_struct *p)
1680 : {
1681 60 : return cpu_curr(task_cpu(p)) == p;
1682 : }
1683 :
1684 : /*
1685 : * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1686 : * use the balance_callback list if you want balancing.
1687 : *
1688 : * this means any call to check_class_changed() must be followed by a call to
1689 : * balance_callback().
1690 : */
1691 4 : static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1692 : const struct sched_class *prev_class,
1693 : int oldprio)
1694 : {
1695 4 : if (prev_class != p->sched_class) {
1696 4 : if (prev_class->switched_from)
1697 4 : prev_class->switched_from(rq, p);
1698 :
1699 4 : p->sched_class->switched_to(rq, p);
1700 0 : } else if (oldprio != p->prio || dl_task(p))
1701 0 : p->sched_class->prio_changed(rq, p, oldprio);
1702 4 : }
1703 :
1704 15885 : void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1705 : {
1706 15885 : if (p->sched_class == rq->curr->sched_class)
1707 8352 : rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1708 7533 : else if (p->sched_class > rq->curr->sched_class)
1709 7532 : resched_curr(rq);
1710 :
1711 : /*
1712 : * A queue event has occurred, and we're going to schedule. In
1713 : * this case, we can save a useless back to back clock update.
1714 : */
1715 15886 : if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1716 12988 : rq_clock_skip_update(rq);
1717 15883 : }
1718 :
1719 : #ifdef CONFIG_SMP
1720 :
1721 : static void
1722 : __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
1723 :
1724 : static int __set_cpus_allowed_ptr(struct task_struct *p,
1725 : const struct cpumask *new_mask,
1726 : u32 flags);
1727 :
1728 26513 : static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
1729 : {
1730 26513 : if (likely(!p->migration_disabled))
1731 : return;
1732 :
1733 0 : if (p->cpus_ptr != &p->cpus_mask)
1734 : return;
1735 :
1736 : /*
1737 : * Violates locking rules! see comment in __do_set_cpus_allowed().
1738 : */
1739 0 : __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
1740 : }
1741 :
1742 1068 : void migrate_disable(void)
1743 : {
1744 1068 : struct task_struct *p = current;
1745 :
1746 1068 : if (p->migration_disabled) {
1747 0 : p->migration_disabled++;
1748 0 : return;
1749 : }
1750 :
1751 1068 : preempt_disable();
1752 1068 : this_rq()->nr_pinned++;
1753 1068 : p->migration_disabled = 1;
1754 1068 : preempt_enable();
1755 : }
1756 : EXPORT_SYMBOL_GPL(migrate_disable);
1757 :
1758 1068 : void migrate_enable(void)
1759 : {
1760 1068 : struct task_struct *p = current;
1761 :
1762 1068 : if (p->migration_disabled > 1) {
1763 0 : p->migration_disabled--;
1764 0 : return;
1765 : }
1766 :
1767 : /*
1768 : * Ensure stop_task runs either before or after this, and that
1769 : * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
1770 : */
1771 1068 : preempt_disable();
1772 1068 : if (p->cpus_ptr != &p->cpus_mask)
1773 0 : __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
1774 : /*
1775 : * Mustn't clear migration_disabled() until cpus_ptr points back at the
1776 : * regular cpus_mask, otherwise things that race (eg.
1777 : * select_fallback_rq) get confused.
1778 : */
1779 1068 : barrier();
1780 1068 : p->migration_disabled = 0;
1781 1068 : this_rq()->nr_pinned--;
1782 1068 : preempt_enable();
1783 : }
1784 : EXPORT_SYMBOL_GPL(migrate_enable);
1785 :
1786 0 : static inline bool rq_has_pinned_tasks(struct rq *rq)
1787 : {
1788 0 : return rq->nr_pinned;
1789 : }
1790 :
1791 : /*
1792 : * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1793 : * __set_cpus_allowed_ptr() and select_fallback_rq().
1794 : */
1795 14849 : static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1796 : {
1797 : /* When not in the task's cpumask, no point in looking further. */
1798 14849 : if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1799 : return false;
1800 :
1801 : /* migrate_disabled() must be allowed to finish. */
1802 14850 : if (is_migration_disabled(p))
1803 0 : return cpu_online(cpu);
1804 :
1805 : /* Non kernel threads are not allowed during either online or offline. */
1806 14850 : if (!(p->flags & PF_KTHREAD))
1807 6614 : return cpu_active(cpu);
1808 :
1809 : /* KTHREAD_IS_PER_CPU is always allowed. */
1810 8236 : if (kthread_is_per_cpu(p))
1811 3711 : return cpu_online(cpu);
1812 :
1813 : /* Regular kernel threads don't get to stay during offline. */
1814 4523 : if (cpu_rq(cpu)->balance_push)
1815 : return false;
1816 :
1817 : /* But are allowed during online. */
1818 4525 : return cpu_online(cpu);
1819 : }
1820 :
1821 : /*
1822 : * This is how migration works:
1823 : *
1824 : * 1) we invoke migration_cpu_stop() on the target CPU using
1825 : * stop_one_cpu().
1826 : * 2) stopper starts to run (implicitly forcing the migrated thread
1827 : * off the CPU)
1828 : * 3) it checks whether the migrated task is still in the wrong runqueue.
1829 : * 4) if it's in the wrong runqueue then the migration thread removes
1830 : * it and puts it into the right queue.
1831 : * 5) stopper completes and stop_one_cpu() returns and the migration
1832 : * is done.
1833 : */
1834 :
1835 : /*
1836 : * move_queued_task - move a queued task to new rq.
1837 : *
1838 : * Returns (locked) new rq. Old rq's lock is released.
1839 : */
1840 31 : static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1841 : struct task_struct *p, int new_cpu)
1842 : {
1843 62 : lockdep_assert_held(&rq->lock);
1844 :
1845 31 : deactivate_task(rq, p, DEQUEUE_NOCLOCK);
1846 31 : set_task_cpu(p, new_cpu);
1847 31 : rq_unlock(rq, rf);
1848 :
1849 31 : rq = cpu_rq(new_cpu);
1850 :
1851 31 : rq_lock(rq, rf);
1852 31 : BUG_ON(task_cpu(p) != new_cpu);
1853 31 : activate_task(rq, p, 0);
1854 31 : check_preempt_curr(rq, p, 0);
1855 :
1856 31 : return rq;
1857 : }
1858 :
1859 : struct migration_arg {
1860 : struct task_struct *task;
1861 : int dest_cpu;
1862 : struct set_affinity_pending *pending;
1863 : };
1864 :
1865 : /*
1866 : * @refs: number of wait_for_completion()
1867 : * @stop_pending: is @stop_work in use
1868 : */
1869 : struct set_affinity_pending {
1870 : refcount_t refs;
1871 : unsigned int stop_pending;
1872 : struct completion done;
1873 : struct cpu_stop_work stop_work;
1874 : struct migration_arg arg;
1875 : };
1876 :
1877 : /*
1878 : * Move (not current) task off this CPU, onto the destination CPU. We're doing
1879 : * this because either it can't run here any more (set_cpus_allowed()
1880 : * away from this CPU, or CPU going down), or because we're
1881 : * attempting to rebalance this task on exec (sched_exec).
1882 : *
1883 : * So we race with normal scheduler movements, but that's OK, as long
1884 : * as the task is no longer on this CPU.
1885 : */
1886 31 : static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1887 : struct task_struct *p, int dest_cpu)
1888 : {
1889 : /* Affinity changed (again). */
1890 31 : if (!is_cpu_allowed(p, dest_cpu))
1891 : return rq;
1892 :
1893 31 : update_rq_clock(rq);
1894 31 : rq = move_queued_task(rq, rf, p, dest_cpu);
1895 :
1896 31 : return rq;
1897 : }
1898 :
1899 : /*
1900 : * migration_cpu_stop - this will be executed by a highprio stopper thread
1901 : * and performs thread migration by bumping thread off CPU then
1902 : * 'pushing' onto another runqueue.
1903 : */
1904 31 : static int migration_cpu_stop(void *data)
1905 : {
1906 31 : struct migration_arg *arg = data;
1907 31 : struct set_affinity_pending *pending = arg->pending;
1908 31 : struct task_struct *p = arg->task;
1909 31 : int dest_cpu = arg->dest_cpu;
1910 31 : struct rq *rq = this_rq();
1911 31 : bool complete = false;
1912 31 : struct rq_flags rf;
1913 :
1914 : /*
1915 : * The original target CPU might have gone down and we might
1916 : * be on another CPU but it doesn't matter.
1917 : */
1918 62 : local_irq_save(rf.flags);
1919 : /*
1920 : * We need to explicitly wake pending tasks before running
1921 : * __migrate_task() such that we will not miss enforcing cpus_ptr
1922 : * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1923 : */
1924 31 : flush_smp_call_function_from_idle();
1925 :
1926 31 : raw_spin_lock(&p->pi_lock);
1927 31 : rq_lock(rq, &rf);
1928 :
1929 : /*
1930 : * If task_rq(p) != rq, it cannot be migrated here, because we're
1931 : * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1932 : * we're holding p->pi_lock.
1933 : */
1934 31 : if (task_rq(p) == rq) {
1935 31 : if (is_migration_disabled(p))
1936 0 : goto out;
1937 :
1938 31 : if (pending) {
1939 0 : if (p->migration_pending == pending)
1940 0 : p->migration_pending = NULL;
1941 : complete = true;
1942 : }
1943 :
1944 31 : if (dest_cpu < 0) {
1945 0 : if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
1946 0 : goto out;
1947 :
1948 0 : dest_cpu = cpumask_any_distribute(&p->cpus_mask);
1949 : }
1950 :
1951 31 : if (task_on_rq_queued(p))
1952 31 : rq = __migrate_task(rq, &rf, p, dest_cpu);
1953 : else
1954 0 : p->wake_cpu = dest_cpu;
1955 :
1956 : /*
1957 : * XXX __migrate_task() can fail, at which point we might end
1958 : * up running on a dodgy CPU, AFAICT this can only happen
1959 : * during CPU hotplug, at which point we'll get pushed out
1960 : * anyway, so it's probably not a big deal.
1961 : */
1962 :
1963 0 : } else if (pending) {
1964 : /*
1965 : * This happens when we get migrated between migrate_enable()'s
1966 : * preempt_enable() and scheduling the stopper task. At that
1967 : * point we're a regular task again and not current anymore.
1968 : *
1969 : * A !PREEMPT kernel has a giant hole here, which makes it far
1970 : * more likely.
1971 : */
1972 :
1973 : /*
1974 : * The task moved before the stopper got to run. We're holding
1975 : * ->pi_lock, so the allowed mask is stable - if it got
1976 : * somewhere allowed, we're done.
1977 : */
1978 0 : if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
1979 0 : if (p->migration_pending == pending)
1980 0 : p->migration_pending = NULL;
1981 0 : complete = true;
1982 0 : goto out;
1983 : }
1984 :
1985 : /*
1986 : * When migrate_enable() hits a rq mis-match we can't reliably
1987 : * determine is_migration_disabled() and so have to chase after
1988 : * it.
1989 : */
1990 0 : WARN_ON_ONCE(!pending->stop_pending);
1991 0 : task_rq_unlock(rq, p, &rf);
1992 0 : stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
1993 0 : &pending->arg, &pending->stop_work);
1994 0 : return 0;
1995 : }
1996 0 : out:
1997 31 : if (pending)
1998 0 : pending->stop_pending = false;
1999 31 : task_rq_unlock(rq, p, &rf);
2000 :
2001 31 : if (complete)
2002 0 : complete_all(&pending->done);
2003 :
2004 : return 0;
2005 : }
2006 :
2007 0 : int push_cpu_stop(void *arg)
2008 : {
2009 0 : struct rq *lowest_rq = NULL, *rq = this_rq();
2010 0 : struct task_struct *p = arg;
2011 :
2012 0 : raw_spin_lock_irq(&p->pi_lock);
2013 0 : raw_spin_lock(&rq->lock);
2014 :
2015 0 : if (task_rq(p) != rq)
2016 0 : goto out_unlock;
2017 :
2018 0 : if (is_migration_disabled(p)) {
2019 0 : p->migration_flags |= MDF_PUSH;
2020 0 : goto out_unlock;
2021 : }
2022 :
2023 0 : p->migration_flags &= ~MDF_PUSH;
2024 :
2025 0 : if (p->sched_class->find_lock_rq)
2026 0 : lowest_rq = p->sched_class->find_lock_rq(p, rq);
2027 :
2028 0 : if (!lowest_rq)
2029 0 : goto out_unlock;
2030 :
2031 : // XXX validate p is still the highest prio task
2032 0 : if (task_rq(p) == rq) {
2033 0 : deactivate_task(rq, p, 0);
2034 0 : set_task_cpu(p, lowest_rq->cpu);
2035 0 : activate_task(lowest_rq, p, 0);
2036 0 : resched_curr(lowest_rq);
2037 : }
2038 :
2039 0 : double_unlock_balance(rq, lowest_rq);
2040 :
2041 0 : out_unlock:
2042 0 : rq->push_busy = false;
2043 0 : raw_spin_unlock(&rq->lock);
2044 0 : raw_spin_unlock_irq(&p->pi_lock);
2045 :
2046 0 : put_task_struct(p);
2047 0 : return 0;
2048 : }
2049 :
2050 : /*
2051 : * sched_class::set_cpus_allowed must do the below, but is not required to
2052 : * actually call this function.
2053 : */
2054 88 : void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2055 : {
2056 78 : if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2057 0 : p->cpus_ptr = new_mask;
2058 0 : return;
2059 : }
2060 :
2061 88 : cpumask_copy(&p->cpus_mask, new_mask);
2062 78 : p->nr_cpus_allowed = cpumask_weight(new_mask);
2063 : }
2064 :
2065 : static void
2066 78 : __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2067 : {
2068 78 : struct rq *rq = task_rq(p);
2069 78 : bool queued, running;
2070 :
2071 : /*
2072 : * This here violates the locking rules for affinity, since we're only
2073 : * supposed to change these variables while holding both rq->lock and
2074 : * p->pi_lock.
2075 : *
2076 : * HOWEVER, it magically works, because ttwu() is the only code that
2077 : * accesses these variables under p->pi_lock and only does so after
2078 : * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2079 : * before finish_task().
2080 : *
2081 : * XXX do further audits, this smells like something putrid.
2082 : */
2083 78 : if (flags & SCA_MIGRATE_DISABLE)
2084 : SCHED_WARN_ON(!p->on_cpu);
2085 : else
2086 156 : lockdep_assert_held(&p->pi_lock);
2087 :
2088 78 : queued = task_on_rq_queued(p);
2089 78 : running = task_current(rq, p);
2090 :
2091 78 : if (queued) {
2092 : /*
2093 : * Because __kthread_bind() calls this on blocked tasks without
2094 : * holding rq->lock.
2095 : */
2096 4 : lockdep_assert_held(&rq->lock);
2097 2 : dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2098 : }
2099 78 : if (running)
2100 2 : put_prev_task(rq, p);
2101 :
2102 78 : p->sched_class->set_cpus_allowed(p, new_mask, flags);
2103 :
2104 78 : if (queued)
2105 2 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2106 78 : if (running)
2107 2 : set_next_task(rq, p);
2108 78 : }
2109 :
2110 67 : void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2111 : {
2112 52 : __do_set_cpus_allowed(p, new_mask, 0);
2113 52 : }
2114 :
2115 : /*
2116 : * This function is wildly self concurrent; here be dragons.
2117 : *
2118 : *
2119 : * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2120 : * designated task is enqueued on an allowed CPU. If that task is currently
2121 : * running, we have to kick it out using the CPU stopper.
2122 : *
2123 : * Migrate-Disable comes along and tramples all over our nice sandcastle.
2124 : * Consider:
2125 : *
2126 : * Initial conditions: P0->cpus_mask = [0, 1]
2127 : *
2128 : * P0@CPU0 P1
2129 : *
2130 : * migrate_disable();
2131 : * <preempted>
2132 : * set_cpus_allowed_ptr(P0, [1]);
2133 : *
2134 : * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2135 : * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2136 : * This means we need the following scheme:
2137 : *
2138 : * P0@CPU0 P1
2139 : *
2140 : * migrate_disable();
2141 : * <preempted>
2142 : * set_cpus_allowed_ptr(P0, [1]);
2143 : * <blocks>
2144 : * <resumes>
2145 : * migrate_enable();
2146 : * __set_cpus_allowed_ptr();
2147 : * <wakes local stopper>
2148 : * `--> <woken on migration completion>
2149 : *
2150 : * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2151 : * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2152 : * task p are serialized by p->pi_lock, which we can leverage: the one that
2153 : * should come into effect at the end of the Migrate-Disable region is the last
2154 : * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2155 : * but we still need to properly signal those waiting tasks at the appropriate
2156 : * moment.
2157 : *
2158 : * This is implemented using struct set_affinity_pending. The first
2159 : * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2160 : * setup an instance of that struct and install it on the targeted task_struct.
2161 : * Any and all further callers will reuse that instance. Those then wait for
2162 : * a completion signaled at the tail of the CPU stopper callback (1), triggered
2163 : * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2164 : *
2165 : *
2166 : * (1) In the cases covered above. There is one more where the completion is
2167 : * signaled within affine_move_task() itself: when a subsequent affinity request
2168 : * cancels the need for an active migration. Consider:
2169 : *
2170 : * Initial conditions: P0->cpus_mask = [0, 1]
2171 : *
2172 : * P0@CPU0 P1 P2
2173 : *
2174 : * migrate_disable();
2175 : * <preempted>
2176 : * set_cpus_allowed_ptr(P0, [1]);
2177 : * <blocks>
2178 : * set_cpus_allowed_ptr(P0, [0, 1]);
2179 : * <signal completion>
2180 : * <awakes>
2181 : *
2182 : * Note that the above is safe vs a concurrent migrate_enable(), as any
2183 : * pending affinity completion is preceded by an uninstallation of
2184 : * p->migration_pending done with p->pi_lock held.
2185 : */
2186 11 : static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2187 : int dest_cpu, unsigned int flags)
2188 : {
2189 11 : struct set_affinity_pending my_pending = { }, *pending = NULL;
2190 11 : bool stop_pending, complete = false;
2191 :
2192 : /* Can the task run on the task's current CPU? If so, we're done */
2193 11 : if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2194 5 : struct task_struct *push_task = NULL;
2195 :
2196 5 : if ((flags & SCA_MIGRATE_ENABLE) &&
2197 0 : (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2198 0 : rq->push_busy = true;
2199 0 : push_task = get_task_struct(p);
2200 : }
2201 :
2202 : /*
2203 : * If there are pending waiters, but no pending stop_work,
2204 : * then complete now.
2205 : */
2206 5 : pending = p->migration_pending;
2207 5 : if (pending && !pending->stop_pending) {
2208 0 : p->migration_pending = NULL;
2209 0 : complete = true;
2210 : }
2211 :
2212 5 : task_rq_unlock(rq, p, rf);
2213 :
2214 5 : if (push_task) {
2215 0 : stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2216 : p, &rq->push_work);
2217 : }
2218 :
2219 5 : if (complete)
2220 0 : complete_all(&pending->done);
2221 :
2222 5 : return 0;
2223 : }
2224 :
2225 6 : if (!(flags & SCA_MIGRATE_ENABLE)) {
2226 : /* serialized by p->pi_lock */
2227 6 : if (!p->migration_pending) {
2228 : /* Install the request */
2229 6 : refcount_set(&my_pending.refs, 1);
2230 6 : init_completion(&my_pending.done);
2231 6 : my_pending.arg = (struct migration_arg) {
2232 : .task = p,
2233 : .dest_cpu = -1, /* any */
2234 : .pending = &my_pending,
2235 : };
2236 :
2237 6 : p->migration_pending = &my_pending;
2238 : } else {
2239 0 : pending = p->migration_pending;
2240 0 : refcount_inc(&pending->refs);
2241 : }
2242 : }
2243 6 : pending = p->migration_pending;
2244 : /*
2245 : * - !MIGRATE_ENABLE:
2246 : * we'll have installed a pending if there wasn't one already.
2247 : *
2248 : * - MIGRATE_ENABLE:
2249 : * we're here because the current CPU isn't matching anymore,
2250 : * the only way that can happen is because of a concurrent
2251 : * set_cpus_allowed_ptr() call, which should then still be
2252 : * pending completion.
2253 : *
2254 : * Either way, we really should have a @pending here.
2255 : */
2256 6 : if (WARN_ON_ONCE(!pending)) {
2257 0 : task_rq_unlock(rq, p, rf);
2258 0 : return -EINVAL;
2259 : }
2260 :
2261 6 : if (task_running(rq, p) || p->state == TASK_WAKING) {
2262 : /*
2263 : * MIGRATE_ENABLE gets here because 'p == current', but for
2264 : * anything else we cannot do is_migration_disabled(), punt
2265 : * and have the stopper function handle it all race-free.
2266 : */
2267 0 : stop_pending = pending->stop_pending;
2268 0 : if (!stop_pending)
2269 0 : pending->stop_pending = true;
2270 :
2271 0 : if (flags & SCA_MIGRATE_ENABLE)
2272 0 : p->migration_flags &= ~MDF_PUSH;
2273 :
2274 0 : task_rq_unlock(rq, p, rf);
2275 :
2276 0 : if (!stop_pending) {
2277 0 : stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2278 0 : &pending->arg, &pending->stop_work);
2279 : }
2280 :
2281 0 : if (flags & SCA_MIGRATE_ENABLE)
2282 : return 0;
2283 : } else {
2284 :
2285 6 : if (!is_migration_disabled(p)) {
2286 6 : if (task_on_rq_queued(p))
2287 0 : rq = move_queued_task(rq, rf, p, dest_cpu);
2288 :
2289 6 : if (!pending->stop_pending) {
2290 6 : p->migration_pending = NULL;
2291 6 : complete = true;
2292 : }
2293 : }
2294 6 : task_rq_unlock(rq, p, rf);
2295 :
2296 6 : if (complete)
2297 6 : complete_all(&pending->done);
2298 : }
2299 :
2300 6 : wait_for_completion(&pending->done);
2301 :
2302 6 : if (refcount_dec_and_test(&pending->refs))
2303 6 : wake_up_var(&pending->refs); /* No UaF, just an address */
2304 :
2305 : /*
2306 : * Block the original owner of &pending until all subsequent callers
2307 : * have seen the completion and decremented the refcount
2308 : */
2309 6 : wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2310 :
2311 : /* ARGH */
2312 6 : WARN_ON_ONCE(my_pending.stop_pending);
2313 :
2314 : return 0;
2315 : }
2316 :
2317 : /*
2318 : * Change a given task's CPU affinity. Migrate the thread to a
2319 : * proper CPU and schedule it away if the CPU it's executing on
2320 : * is removed from the allowed bitmask.
2321 : *
2322 : * NOTE: the caller must have a valid reference to the task, the
2323 : * task must not exit() & deallocate itself prematurely. The
2324 : * call is not atomic; no spinlocks may be held.
2325 : */
2326 62 : static int __set_cpus_allowed_ptr(struct task_struct *p,
2327 : const struct cpumask *new_mask,
2328 : u32 flags)
2329 : {
2330 62 : const struct cpumask *cpu_valid_mask = cpu_active_mask;
2331 62 : unsigned int dest_cpu;
2332 62 : struct rq_flags rf;
2333 62 : struct rq *rq;
2334 62 : int ret = 0;
2335 :
2336 62 : rq = task_rq_lock(p, &rf);
2337 62 : update_rq_clock(rq);
2338 :
2339 62 : if (p->flags & PF_KTHREAD || is_migration_disabled(p)) {
2340 : /*
2341 : * Kernel threads are allowed on online && !active CPUs,
2342 : * however, during cpu-hot-unplug, even these might get pushed
2343 : * away if not KTHREAD_IS_PER_CPU.
2344 : *
2345 : * Specifically, migration_disabled() tasks must not fail the
2346 : * cpumask_any_and_distribute() pick below, esp. so on
2347 : * SCA_MIGRATE_ENABLE, otherwise we'll not call
2348 : * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2349 : */
2350 : cpu_valid_mask = cpu_online_mask;
2351 : }
2352 :
2353 : /*
2354 : * Must re-check here, to close a race against __kthread_bind(),
2355 : * sched_setaffinity() is not guaranteed to observe the flag.
2356 : */
2357 62 : if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2358 0 : ret = -EINVAL;
2359 0 : goto out;
2360 : }
2361 :
2362 62 : if (!(flags & SCA_MIGRATE_ENABLE)) {
2363 62 : if (cpumask_equal(&p->cpus_mask, new_mask))
2364 51 : goto out;
2365 :
2366 11 : if (WARN_ON_ONCE(p == current &&
2367 : is_migration_disabled(p) &&
2368 : !cpumask_test_cpu(task_cpu(p), new_mask))) {
2369 0 : ret = -EBUSY;
2370 0 : goto out;
2371 : }
2372 : }
2373 :
2374 : /*
2375 : * Picking a ~random cpu helps in cases where we are changing affinity
2376 : * for groups of tasks (ie. cpuset), so that load balancing is not
2377 : * immediately required to distribute the tasks within their new mask.
2378 : */
2379 11 : dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2380 11 : if (dest_cpu >= nr_cpu_ids) {
2381 0 : ret = -EINVAL;
2382 0 : goto out;
2383 : }
2384 :
2385 11 : __do_set_cpus_allowed(p, new_mask, flags);
2386 :
2387 11 : return affine_move_task(rq, p, &rf, dest_cpu, flags);
2388 :
2389 51 : out:
2390 51 : task_rq_unlock(rq, p, &rf);
2391 :
2392 51 : return ret;
2393 : }
2394 :
2395 62 : int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2396 : {
2397 62 : return __set_cpus_allowed_ptr(p, new_mask, 0);
2398 : }
2399 : EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2400 :
2401 994 : void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2402 : {
2403 : #ifdef CONFIG_SCHED_DEBUG
2404 : /*
2405 : * We should never call set_task_cpu() on a blocked task,
2406 : * ttwu() will sort out the placement.
2407 : */
2408 : WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2409 : !p->on_rq);
2410 :
2411 : /*
2412 : * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2413 : * because schedstat_wait_{start,end} rebase migrating task's wait_start
2414 : * time relying on p->on_rq.
2415 : */
2416 : WARN_ON_ONCE(p->state == TASK_RUNNING &&
2417 : p->sched_class == &fair_sched_class &&
2418 : (p->on_rq && !task_on_rq_migrating(p)));
2419 :
2420 : #ifdef CONFIG_LOCKDEP
2421 : /*
2422 : * The caller should hold either p->pi_lock or rq->lock, when changing
2423 : * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2424 : *
2425 : * sched_move_task() holds both and thus holding either pins the cgroup,
2426 : * see task_group().
2427 : *
2428 : * Furthermore, all task_rq users should acquire both locks, see
2429 : * task_rq_lock().
2430 : */
2431 : WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2432 : lockdep_is_held(&task_rq(p)->lock)));
2433 : #endif
2434 : /*
2435 : * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
2436 : */
2437 : WARN_ON_ONCE(!cpu_online(new_cpu));
2438 :
2439 : WARN_ON_ONCE(is_migration_disabled(p));
2440 : #endif
2441 :
2442 994 : trace_sched_migrate_task(p, new_cpu);
2443 :
2444 995 : if (task_cpu(p) != new_cpu) {
2445 995 : if (p->sched_class->migrate_task_rq)
2446 992 : p->sched_class->migrate_task_rq(p, new_cpu);
2447 995 : p->se.nr_migrations++;
2448 995 : rseq_migrate(p);
2449 995 : perf_event_task_migrate(p);
2450 : }
2451 :
2452 995 : __set_task_cpu(p, new_cpu);
2453 995 : }
2454 :
2455 : #ifdef CONFIG_NUMA_BALANCING
2456 : static void __migrate_swap_task(struct task_struct *p, int cpu)
2457 : {
2458 : if (task_on_rq_queued(p)) {
2459 : struct rq *src_rq, *dst_rq;
2460 : struct rq_flags srf, drf;
2461 :
2462 : src_rq = task_rq(p);
2463 : dst_rq = cpu_rq(cpu);
2464 :
2465 : rq_pin_lock(src_rq, &srf);
2466 : rq_pin_lock(dst_rq, &drf);
2467 :
2468 : deactivate_task(src_rq, p, 0);
2469 : set_task_cpu(p, cpu);
2470 : activate_task(dst_rq, p, 0);
2471 : check_preempt_curr(dst_rq, p, 0);
2472 :
2473 : rq_unpin_lock(dst_rq, &drf);
2474 : rq_unpin_lock(src_rq, &srf);
2475 :
2476 : } else {
2477 : /*
2478 : * Task isn't running anymore; make it appear like we migrated
2479 : * it before it went to sleep. This means on wakeup we make the
2480 : * previous CPU our target instead of where it really is.
2481 : */
2482 : p->wake_cpu = cpu;
2483 : }
2484 : }
2485 :
2486 : struct migration_swap_arg {
2487 : struct task_struct *src_task, *dst_task;
2488 : int src_cpu, dst_cpu;
2489 : };
2490 :
2491 : static int migrate_swap_stop(void *data)
2492 : {
2493 : struct migration_swap_arg *arg = data;
2494 : struct rq *src_rq, *dst_rq;
2495 : int ret = -EAGAIN;
2496 :
2497 : if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
2498 : return -EAGAIN;
2499 :
2500 : src_rq = cpu_rq(arg->src_cpu);
2501 : dst_rq = cpu_rq(arg->dst_cpu);
2502 :
2503 : double_raw_lock(&arg->src_task->pi_lock,
2504 : &arg->dst_task->pi_lock);
2505 : double_rq_lock(src_rq, dst_rq);
2506 :
2507 : if (task_cpu(arg->dst_task) != arg->dst_cpu)
2508 : goto unlock;
2509 :
2510 : if (task_cpu(arg->src_task) != arg->src_cpu)
2511 : goto unlock;
2512 :
2513 : if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
2514 : goto unlock;
2515 :
2516 : if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
2517 : goto unlock;
2518 :
2519 : __migrate_swap_task(arg->src_task, arg->dst_cpu);
2520 : __migrate_swap_task(arg->dst_task, arg->src_cpu);
2521 :
2522 : ret = 0;
2523 :
2524 : unlock:
2525 : double_rq_unlock(src_rq, dst_rq);
2526 : raw_spin_unlock(&arg->dst_task->pi_lock);
2527 : raw_spin_unlock(&arg->src_task->pi_lock);
2528 :
2529 : return ret;
2530 : }
2531 :
2532 : /*
2533 : * Cross migrate two tasks
2534 : */
2535 : int migrate_swap(struct task_struct *cur, struct task_struct *p,
2536 : int target_cpu, int curr_cpu)
2537 : {
2538 : struct migration_swap_arg arg;
2539 : int ret = -EINVAL;
2540 :
2541 : arg = (struct migration_swap_arg){
2542 : .src_task = cur,
2543 : .src_cpu = curr_cpu,
2544 : .dst_task = p,
2545 : .dst_cpu = target_cpu,
2546 : };
2547 :
2548 : if (arg.src_cpu == arg.dst_cpu)
2549 : goto out;
2550 :
2551 : /*
2552 : * These three tests are all lockless; this is OK since all of them
2553 : * will be re-checked with proper locks held further down the line.
2554 : */
2555 : if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
2556 : goto out;
2557 :
2558 : if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
2559 : goto out;
2560 :
2561 : if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
2562 : goto out;
2563 :
2564 : trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
2565 : ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
2566 :
2567 : out:
2568 : return ret;
2569 : }
2570 : #endif /* CONFIG_NUMA_BALANCING */
2571 :
2572 : /*
2573 : * wait_task_inactive - wait for a thread to unschedule.
2574 : *
2575 : * If @match_state is nonzero, it's the @p->state value just checked and
2576 : * not expected to change. If it changes, i.e. @p might have woken up,
2577 : * then return zero. When we succeed in waiting for @p to be off its CPU,
2578 : * we return a positive number (its total switch count). If a second call
2579 : * a short while later returns the same number, the caller can be sure that
2580 : * @p has remained unscheduled the whole time.
2581 : *
2582 : * The caller must ensure that the task *will* unschedule sometime soon,
2583 : * else this function might spin for a *long* time. This function can't
2584 : * be called with interrupts off, or it may introduce deadlock with
2585 : * smp_call_function() if an IPI is sent by the same process we are
2586 : * waiting to become inactive.
2587 : */
2588 84 : unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2589 : {
2590 84 : int running, queued;
2591 84 : struct rq_flags rf;
2592 84 : unsigned long ncsw;
2593 84 : struct rq *rq;
2594 :
2595 84 : for (;;) {
2596 : /*
2597 : * We do the initial early heuristics without holding
2598 : * any task-queue locks at all. We'll only try to get
2599 : * the runqueue lock when things look like they will
2600 : * work out!
2601 : */
2602 84 : rq = task_rq(p);
2603 :
2604 : /*
2605 : * If the task is actively running on another CPU
2606 : * still, just relax and busy-wait without holding
2607 : * any locks.
2608 : *
2609 : * NOTE! Since we don't hold any locks, it's not
2610 : * even sure that "rq" stays as the right runqueue!
2611 : * But we don't care, since "task_running()" will
2612 : * return false if the runqueue has changed and p
2613 : * is actually now running somewhere else!
2614 : */
2615 84 : while (task_running(rq, p)) {
2616 0 : if (match_state && unlikely(p->state != match_state))
2617 : return 0;
2618 0 : cpu_relax();
2619 : }
2620 :
2621 : /*
2622 : * Ok, time to look more closely! We need the rq
2623 : * lock now, to be *sure*. If we're wrong, we'll
2624 : * just go back and repeat.
2625 : */
2626 84 : rq = task_rq_lock(p, &rf);
2627 84 : trace_sched_wait_task(p);
2628 84 : running = task_running(rq, p);
2629 84 : queued = task_on_rq_queued(p);
2630 84 : ncsw = 0;
2631 84 : if (!match_state || p->state == match_state)
2632 84 : ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2633 84 : task_rq_unlock(rq, p, &rf);
2634 :
2635 : /*
2636 : * If it changed from the expected state, bail out now.
2637 : */
2638 84 : if (unlikely(!ncsw))
2639 : break;
2640 :
2641 : /*
2642 : * Was it really running after all now that we
2643 : * checked with the proper locks actually held?
2644 : *
2645 : * Oops. Go back and try again..
2646 : */
2647 84 : if (unlikely(running)) {
2648 0 : cpu_relax();
2649 0 : continue;
2650 : }
2651 :
2652 : /*
2653 : * It's not enough that it's not actively running,
2654 : * it must be off the runqueue _entirely_, and not
2655 : * preempted!
2656 : *
2657 : * So if it was still runnable (but just not actively
2658 : * running right now), it's preempted, and we should
2659 : * yield - it could be a while.
2660 : */
2661 84 : if (unlikely(queued)) {
2662 0 : ktime_t to = NSEC_PER_SEC / HZ;
2663 :
2664 0 : set_current_state(TASK_UNINTERRUPTIBLE);
2665 0 : schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2666 0 : continue;
2667 : }
2668 :
2669 : /*
2670 : * Ahh, all good. It wasn't running, and it wasn't
2671 : * runnable, which means that it will never become
2672 : * running in the future either. We're all done!
2673 : */
2674 : break;
2675 : }
2676 :
2677 : return ncsw;
2678 : }
2679 :
2680 : /***
2681 : * kick_process - kick a running thread to enter/exit the kernel
2682 : * @p: the to-be-kicked thread
2683 : *
2684 : * Cause a process which is running on another CPU to enter
2685 : * kernel-mode, without any delay. (to get signals handled.)
2686 : *
2687 : * NOTE: this function doesn't have to take the runqueue lock,
2688 : * because all it wants to ensure is that the remote task enters
2689 : * the kernel. If the IPI races and the task has been migrated
2690 : * to another CPU then no harm is done and the purpose has been
2691 : * achieved as well.
2692 : */
2693 51704 : void kick_process(struct task_struct *p)
2694 : {
2695 51704 : int cpu;
2696 :
2697 51704 : preempt_disable();
2698 51713 : cpu = task_cpu(p);
2699 51713 : if ((cpu != smp_processor_id()) && task_curr(p))
2700 40 : smp_send_reschedule(cpu);
2701 51713 : preempt_enable();
2702 51712 : }
2703 : EXPORT_SYMBOL_GPL(kick_process);
2704 :
2705 : /*
2706 : * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2707 : *
2708 : * A few notes on cpu_active vs cpu_online:
2709 : *
2710 : * - cpu_active must be a subset of cpu_online
2711 : *
2712 : * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2713 : * see __set_cpus_allowed_ptr(). At this point the newly online
2714 : * CPU isn't yet part of the sched domains, and balancing will not
2715 : * see it.
2716 : *
2717 : * - on CPU-down we clear cpu_active() to mask the sched domains and
2718 : * avoid the load balancer to place new tasks on the to be removed
2719 : * CPU. Existing tasks will remain running there and will be taken
2720 : * off.
2721 : *
2722 : * This means that fallback selection must not select !active CPUs.
2723 : * And can assume that any active CPU must be online. Conversely
2724 : * select_task_rq() below may allow selection of !active CPUs in order
2725 : * to satisfy the above rules.
2726 : */
2727 15 : static int select_fallback_rq(int cpu, struct task_struct *p)
2728 : {
2729 15 : int nid = cpu_to_node(cpu);
2730 15 : const struct cpumask *nodemask = NULL;
2731 15 : enum { cpuset, possible, fail } state = cpuset;
2732 15 : int dest_cpu;
2733 :
2734 : /*
2735 : * If the node that the CPU is on has been offlined, cpu_to_node()
2736 : * will return -1. There is no CPU on the node, and we should
2737 : * select the CPU on the other node.
2738 : */
2739 15 : if (nid != -1) {
2740 15 : nodemask = cpumask_of_node(nid);
2741 :
2742 : /* Look for allowed, online CPU in same node. */
2743 45 : for_each_cpu(dest_cpu, nodemask) {
2744 30 : if (!cpu_active(dest_cpu))
2745 0 : continue;
2746 30 : if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2747 0 : return dest_cpu;
2748 : }
2749 : }
2750 :
2751 30 : for (;;) {
2752 : /* Any allowed, online CPU? */
2753 45 : for_each_cpu(dest_cpu, p->cpus_ptr) {
2754 30 : if (!is_cpu_allowed(p, dest_cpu))
2755 15 : continue;
2756 :
2757 15 : goto out;
2758 : }
2759 :
2760 : /* No more Mr. Nice Guy. */
2761 15 : switch (state) {
2762 : case cpuset:
2763 : if (IS_ENABLED(CONFIG_CPUSETS)) {
2764 : cpuset_cpus_allowed_fallback(p);
2765 : state = possible;
2766 : break;
2767 : }
2768 15 : fallthrough;
2769 : case possible:
2770 : /*
2771 : * XXX When called from select_task_rq() we only
2772 : * hold p->pi_lock and again violate locking order.
2773 : *
2774 : * More yuck to audit.
2775 : */
2776 15 : do_set_cpus_allowed(p, cpu_possible_mask);
2777 15 : state = fail;
2778 15 : break;
2779 :
2780 0 : case fail:
2781 0 : BUG();
2782 : break;
2783 : }
2784 : }
2785 :
2786 15 : out:
2787 15 : if (state != cpuset) {
2788 : /*
2789 : * Don't tell them about moving exiting tasks or
2790 : * kernel threads (both mm NULL), since they never
2791 : * leave kernel.
2792 : */
2793 15 : if (p->mm && printk_ratelimit()) {
2794 0 : printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2795 0 : task_pid_nr(p), p->comm, cpu);
2796 : }
2797 : }
2798 :
2799 : return dest_cpu;
2800 : }
2801 :
2802 : /*
2803 : * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2804 : */
2805 : static inline
2806 14785 : int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
2807 : {
2808 29576 : lockdep_assert_held(&p->pi_lock);
2809 :
2810 14789 : if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
2811 11048 : cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
2812 : else
2813 3741 : cpu = cpumask_any(p->cpus_ptr);
2814 :
2815 : /*
2816 : * In order not to call set_task_cpu() on a blocking task we need
2817 : * to rely on ttwu() to place the task on a valid ->cpus_ptr
2818 : * CPU.
2819 : *
2820 : * Since this is common to all placement strategies, this lives here.
2821 : *
2822 : * [ this allows ->select_task() to simply return task_cpu(p) and
2823 : * not worry about this generic constraint ]
2824 : */
2825 14791 : if (unlikely(!is_cpu_allowed(p, cpu)))
2826 15 : cpu = select_fallback_rq(task_cpu(p), p);
2827 :
2828 14789 : return cpu;
2829 : }
2830 :
2831 4 : void sched_set_stop_task(int cpu, struct task_struct *stop)
2832 : {
2833 4 : static struct lock_class_key stop_pi_lock;
2834 4 : struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2835 4 : struct task_struct *old_stop = cpu_rq(cpu)->stop;
2836 :
2837 4 : if (stop) {
2838 : /*
2839 : * Make it appear like a SCHED_FIFO task, its something
2840 : * userspace knows about and won't get confused about.
2841 : *
2842 : * Also, it will make PI more or less work without too
2843 : * much confusion -- but then, stop work should not
2844 : * rely on PI working anyway.
2845 : */
2846 8 : sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2847 :
2848 4 : stop->sched_class = &stop_sched_class;
2849 :
2850 : /*
2851 : * The PI code calls rt_mutex_setprio() with ->pi_lock held to
2852 : * adjust the effective priority of a task. As a result,
2853 : * rt_mutex_setprio() can trigger (RT) balancing operations,
2854 : * which can then trigger wakeups of the stop thread to push
2855 : * around the current task.
2856 : *
2857 : * The stop task itself will never be part of the PI-chain, it
2858 : * never blocks, therefore that ->pi_lock recursion is safe.
2859 : * Tell lockdep about this by placing the stop->pi_lock in its
2860 : * own class.
2861 : */
2862 4 : lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
2863 : }
2864 :
2865 4 : cpu_rq(cpu)->stop = stop;
2866 :
2867 4 : if (old_stop) {
2868 : /*
2869 : * Reset it back to a normal scheduling class so that
2870 : * it can die in pieces.
2871 : */
2872 0 : old_stop->sched_class = &rt_sched_class;
2873 : }
2874 4 : }
2875 :
2876 : #else /* CONFIG_SMP */
2877 :
2878 : static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2879 : const struct cpumask *new_mask,
2880 : u32 flags)
2881 : {
2882 : return set_cpus_allowed_ptr(p, new_mask);
2883 : }
2884 :
2885 : static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
2886 :
2887 : static inline bool rq_has_pinned_tasks(struct rq *rq)
2888 : {
2889 : return false;
2890 : }
2891 :
2892 : #endif /* !CONFIG_SMP */
2893 :
2894 : static void
2895 14026 : ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2896 : {
2897 : struct rq *rq;
2898 :
2899 : if (!schedstat_enabled())
2900 : return;
2901 :
2902 : rq = this_rq();
2903 :
2904 : #ifdef CONFIG_SMP
2905 : if (cpu == rq->cpu) {
2906 : __schedstat_inc(rq->ttwu_local);
2907 : __schedstat_inc(p->se.statistics.nr_wakeups_local);
2908 : } else {
2909 : struct sched_domain *sd;
2910 :
2911 : __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2912 : rcu_read_lock();
2913 : for_each_domain(rq->cpu, sd) {
2914 : if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2915 : __schedstat_inc(sd->ttwu_wake_remote);
2916 : break;
2917 : }
2918 : }
2919 : rcu_read_unlock();
2920 : }
2921 :
2922 : if (wake_flags & WF_MIGRATED)
2923 : __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2924 : #endif /* CONFIG_SMP */
2925 :
2926 : __schedstat_inc(rq->ttwu_count);
2927 : __schedstat_inc(p->se.statistics.nr_wakeups);
2928 :
2929 : if (wake_flags & WF_SYNC)
2930 : __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2931 : }
2932 :
2933 : /*
2934 : * Mark the task runnable and perform wakeup-preemption.
2935 : */
2936 14023 : static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2937 : struct rq_flags *rf)
2938 : {
2939 14023 : check_preempt_curr(rq, p, wake_flags);
2940 14022 : p->state = TASK_RUNNING;
2941 14022 : trace_sched_wakeup(p);
2942 :
2943 : #ifdef CONFIG_SMP
2944 14017 : if (p->sched_class->task_woken) {
2945 : /*
2946 : * Our task @p is fully woken up and running; so it's safe to
2947 : * drop the rq->lock, hereafter rq is only used for statistics.
2948 : */
2949 0 : rq_unpin_lock(rq, rf);
2950 0 : p->sched_class->task_woken(rq, p);
2951 0 : rq_repin_lock(rq, rf);
2952 : }
2953 :
2954 14017 : if (rq->idle_stamp) {
2955 6461 : u64 delta = rq_clock(rq) - rq->idle_stamp;
2956 6465 : u64 max = 2*rq->max_idle_balance_cost;
2957 :
2958 6465 : update_avg(&rq->avg_idle, delta);
2959 :
2960 6465 : if (rq->avg_idle > max)
2961 3683 : rq->avg_idle = max;
2962 :
2963 6465 : rq->idle_stamp = 0;
2964 : }
2965 : #endif
2966 14021 : }
2967 :
2968 : static void
2969 13856 : ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2970 : struct rq_flags *rf)
2971 : {
2972 13856 : int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2973 :
2974 27716 : lockdep_assert_held(&rq->lock);
2975 :
2976 13859 : if (p->sched_contributes_to_load)
2977 2713 : rq->nr_uninterruptible--;
2978 :
2979 : #ifdef CONFIG_SMP
2980 13859 : if (wake_flags & WF_MIGRATED)
2981 : en_flags |= ENQUEUE_MIGRATED;
2982 : else
2983 : #endif
2984 13751 : if (p->in_iowait) {
2985 1873 : delayacct_blkio_end(p);
2986 1873 : atomic_dec(&task_rq(p)->nr_iowait);
2987 : }
2988 :
2989 13859 : activate_task(rq, p, en_flags);
2990 13861 : ttwu_do_wakeup(rq, p, wake_flags, rf);
2991 13858 : }
2992 :
2993 : /*
2994 : * Consider @p being inside a wait loop:
2995 : *
2996 : * for (;;) {
2997 : * set_current_state(TASK_UNINTERRUPTIBLE);
2998 : *
2999 : * if (CONDITION)
3000 : * break;
3001 : *
3002 : * schedule();
3003 : * }
3004 : * __set_current_state(TASK_RUNNING);
3005 : *
3006 : * between set_current_state() and schedule(). In this case @p is still
3007 : * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3008 : * an atomic manner.
3009 : *
3010 : * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3011 : * then schedule() must still happen and p->state can be changed to
3012 : * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3013 : * need to do a full wakeup with enqueue.
3014 : *
3015 : * Returns: %true when the wakeup is done,
3016 : * %false otherwise.
3017 : */
3018 171 : static int ttwu_runnable(struct task_struct *p, int wake_flags)
3019 : {
3020 171 : struct rq_flags rf;
3021 171 : struct rq *rq;
3022 171 : int ret = 0;
3023 :
3024 171 : rq = __task_rq_lock(p, &rf);
3025 171 : if (task_on_rq_queued(p)) {
3026 : /* check_preempt_curr() may use rq clock */
3027 162 : update_rq_clock(rq);
3028 162 : ttwu_do_wakeup(rq, p, wake_flags, &rf);
3029 162 : ret = 1;
3030 : }
3031 171 : __task_rq_unlock(rq, &rf);
3032 :
3033 171 : return ret;
3034 : }
3035 :
3036 : #ifdef CONFIG_SMP
3037 4170 : void sched_ttwu_pending(void *arg)
3038 : {
3039 4170 : struct llist_node *llist = arg;
3040 4170 : struct rq *rq = this_rq();
3041 4171 : struct task_struct *p, *t;
3042 4171 : struct rq_flags rf;
3043 :
3044 4171 : if (!llist)
3045 0 : return;
3046 :
3047 : /*
3048 : * rq::ttwu_pending racy indication of out-standing wakeups.
3049 : * Races such that false-negatives are possible, since they
3050 : * are shorter lived that false-positives would be.
3051 : */
3052 4171 : WRITE_ONCE(rq->ttwu_pending, 0);
3053 :
3054 4171 : rq_lock_irqsave(rq, &rf);
3055 4174 : update_rq_clock(rq);
3056 :
3057 8359 : llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3058 4185 : if (WARN_ON_ONCE(p->on_cpu))
3059 0 : smp_cond_load_acquire(&p->on_cpu, !VAL);
3060 :
3061 4185 : if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3062 0 : set_task_cpu(p, cpu_of(rq));
3063 :
3064 4185 : ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3065 : }
3066 :
3067 4174 : rq_unlock_irqrestore(rq, &rf);
3068 : }
3069 :
3070 7103 : void send_call_function_single_ipi(int cpu)
3071 : {
3072 7103 : struct rq *rq = cpu_rq(cpu);
3073 :
3074 7103 : if (!set_nr_if_polling(rq->idle))
3075 7075 : arch_send_call_function_single_ipi(cpu);
3076 : else
3077 28 : trace_sched_wake_idle_without_ipi(cpu);
3078 7103 : }
3079 :
3080 : /*
3081 : * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3082 : * necessary. The wakee CPU on receipt of the IPI will queue the task
3083 : * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3084 : * of the wakeup instead of the waker.
3085 : */
3086 4188 : static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3087 : {
3088 4188 : struct rq *rq = cpu_rq(cpu);
3089 :
3090 4188 : p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3091 :
3092 4188 : WRITE_ONCE(rq->ttwu_pending, 1);
3093 4188 : __smp_call_single_queue(cpu, &p->wake_entry.llist);
3094 4188 : }
3095 :
3096 0 : void wake_up_if_idle(int cpu)
3097 : {
3098 0 : struct rq *rq = cpu_rq(cpu);
3099 0 : struct rq_flags rf;
3100 :
3101 0 : rcu_read_lock();
3102 :
3103 0 : if (!is_idle_task(rcu_dereference(rq->curr)))
3104 0 : goto out;
3105 :
3106 0 : if (set_nr_if_polling(rq->idle)) {
3107 0 : trace_sched_wake_idle_without_ipi(cpu);
3108 : } else {
3109 0 : rq_lock_irqsave(rq, &rf);
3110 0 : if (is_idle_task(rq->curr))
3111 0 : smp_send_reschedule(cpu);
3112 : /* Else CPU is not idle, do nothing here: */
3113 0 : rq_unlock_irqrestore(rq, &rf);
3114 : }
3115 :
3116 0 : out:
3117 0 : rcu_read_unlock();
3118 0 : }
3119 :
3120 20953 : bool cpus_share_cache(int this_cpu, int that_cpu)
3121 : {
3122 20953 : return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3123 : }
3124 :
3125 13854 : static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3126 : {
3127 : /*
3128 : * Do not complicate things with the async wake_list while the CPU is
3129 : * in hotplug state.
3130 : */
3131 13854 : if (!cpu_active(cpu))
3132 : return false;
3133 :
3134 : /*
3135 : * If the CPU does not share cache, then queue the task on the
3136 : * remote rqs wakelist to avoid accessing remote data.
3137 : */
3138 13844 : if (!cpus_share_cache(smp_processor_id(), cpu))
3139 : return true;
3140 :
3141 : /*
3142 : * If the task is descheduling and the only running task on the
3143 : * CPU then use the wakelist to offload the task activation to
3144 : * the soon-to-be-idle CPU as the current CPU is likely busy.
3145 : * nr_running is checked to avoid unnecessary task stacking.
3146 : */
3147 9657 : if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3148 1 : return true;
3149 :
3150 : return false;
3151 : }
3152 :
3153 13855 : static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3154 : {
3155 13855 : if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3156 4188 : if (WARN_ON_ONCE(cpu == smp_processor_id()))
3157 : return false;
3158 :
3159 4188 : sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3160 4188 : __ttwu_queue_wakelist(p, cpu, wake_flags);
3161 4188 : return true;
3162 : }
3163 :
3164 : return false;
3165 : }
3166 :
3167 : #else /* !CONFIG_SMP */
3168 :
3169 : static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3170 : {
3171 : return false;
3172 : }
3173 :
3174 : #endif /* CONFIG_SMP */
3175 :
3176 13810 : static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3177 : {
3178 13810 : struct rq *rq = cpu_rq(cpu);
3179 13810 : struct rq_flags rf;
3180 :
3181 13810 : if (ttwu_queue_wakelist(p, cpu, wake_flags))
3182 4140 : return;
3183 :
3184 9668 : rq_lock(rq, &rf);
3185 9673 : update_rq_clock(rq);
3186 9674 : ttwu_do_activate(rq, p, wake_flags, &rf);
3187 19344 : rq_unlock(rq, &rf);
3188 : }
3189 :
3190 : /*
3191 : * Notes on Program-Order guarantees on SMP systems.
3192 : *
3193 : * MIGRATION
3194 : *
3195 : * The basic program-order guarantee on SMP systems is that when a task [t]
3196 : * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3197 : * execution on its new CPU [c1].
3198 : *
3199 : * For migration (of runnable tasks) this is provided by the following means:
3200 : *
3201 : * A) UNLOCK of the rq(c0)->lock scheduling out task t
3202 : * B) migration for t is required to synchronize *both* rq(c0)->lock and
3203 : * rq(c1)->lock (if not at the same time, then in that order).
3204 : * C) LOCK of the rq(c1)->lock scheduling in task
3205 : *
3206 : * Release/acquire chaining guarantees that B happens after A and C after B.
3207 : * Note: the CPU doing B need not be c0 or c1
3208 : *
3209 : * Example:
3210 : *
3211 : * CPU0 CPU1 CPU2
3212 : *
3213 : * LOCK rq(0)->lock
3214 : * sched-out X
3215 : * sched-in Y
3216 : * UNLOCK rq(0)->lock
3217 : *
3218 : * LOCK rq(0)->lock // orders against CPU0
3219 : * dequeue X
3220 : * UNLOCK rq(0)->lock
3221 : *
3222 : * LOCK rq(1)->lock
3223 : * enqueue X
3224 : * UNLOCK rq(1)->lock
3225 : *
3226 : * LOCK rq(1)->lock // orders against CPU2
3227 : * sched-out Z
3228 : * sched-in X
3229 : * UNLOCK rq(1)->lock
3230 : *
3231 : *
3232 : * BLOCKING -- aka. SLEEP + WAKEUP
3233 : *
3234 : * For blocking we (obviously) need to provide the same guarantee as for
3235 : * migration. However the means are completely different as there is no lock
3236 : * chain to provide order. Instead we do:
3237 : *
3238 : * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3239 : * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3240 : *
3241 : * Example:
3242 : *
3243 : * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3244 : *
3245 : * LOCK rq(0)->lock LOCK X->pi_lock
3246 : * dequeue X
3247 : * sched-out X
3248 : * smp_store_release(X->on_cpu, 0);
3249 : *
3250 : * smp_cond_load_acquire(&X->on_cpu, !VAL);
3251 : * X->state = WAKING
3252 : * set_task_cpu(X,2)
3253 : *
3254 : * LOCK rq(2)->lock
3255 : * enqueue X
3256 : * X->state = RUNNING
3257 : * UNLOCK rq(2)->lock
3258 : *
3259 : * LOCK rq(2)->lock // orders against CPU1
3260 : * sched-out Z
3261 : * sched-in X
3262 : * UNLOCK rq(2)->lock
3263 : *
3264 : * UNLOCK X->pi_lock
3265 : * UNLOCK rq(0)->lock
3266 : *
3267 : *
3268 : * However, for wakeups there is a second guarantee we must provide, namely we
3269 : * must ensure that CONDITION=1 done by the caller can not be reordered with
3270 : * accesses to the task state; see try_to_wake_up() and set_current_state().
3271 : */
3272 :
3273 : /**
3274 : * try_to_wake_up - wake up a thread
3275 : * @p: the thread to be awakened
3276 : * @state: the mask of task states that can be woken
3277 : * @wake_flags: wake modifier flags (WF_*)
3278 : *
3279 : * Conceptually does:
3280 : *
3281 : * If (@state & @p->state) @p->state = TASK_RUNNING.
3282 : *
3283 : * If the task was not queued/runnable, also place it back on a runqueue.
3284 : *
3285 : * This function is atomic against schedule() which would dequeue the task.
3286 : *
3287 : * It issues a full memory barrier before accessing @p->state, see the comment
3288 : * with set_current_state().
3289 : *
3290 : * Uses p->pi_lock to serialize against concurrent wake-ups.
3291 : *
3292 : * Relies on p->pi_lock stabilizing:
3293 : * - p->sched_class
3294 : * - p->cpus_ptr
3295 : * - p->sched_task_group
3296 : * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3297 : *
3298 : * Tries really hard to only take one task_rq(p)->lock for performance.
3299 : * Takes rq->lock in:
3300 : * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3301 : * - ttwu_queue() -- new rq, for enqueue of the task;
3302 : * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3303 : *
3304 : * As a consequence we race really badly with just about everything. See the
3305 : * many memory barriers and their comments for details.
3306 : *
3307 : * Return: %true if @p->state changes (an actual wakeup was done),
3308 : * %false otherwise.
3309 : */
3310 : static int
3311 15701 : try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3312 : {
3313 15701 : unsigned long flags;
3314 15701 : int cpu, success = 0;
3315 :
3316 15701 : preempt_disable();
3317 15704 : if (p == current) {
3318 : /*
3319 : * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3320 : * == smp_processor_id()'. Together this means we can special
3321 : * case the whole 'p->on_rq && ttwu_runnable()' case below
3322 : * without taking any locks.
3323 : *
3324 : * In particular:
3325 : * - we rely on Program-Order guarantees for all the ordering,
3326 : * - we're serialized against set_special_state() by virtue of
3327 : * it disabling IRQs (this allows not taking ->pi_lock).
3328 : */
3329 153 : if (!(p->state & state))
3330 150 : goto out;
3331 :
3332 3 : success = 1;
3333 3 : trace_sched_waking(p);
3334 3 : p->state = TASK_RUNNING;
3335 3 : trace_sched_wakeup(p);
3336 3 : goto out;
3337 : }
3338 :
3339 : /*
3340 : * If we are going to wake up a thread waiting for CONDITION we
3341 : * need to ensure that CONDITION=1 done by the caller can not be
3342 : * reordered with p->state check below. This pairs with smp_store_mb()
3343 : * in set_current_state() that the waiting thread does.
3344 : */
3345 15551 : raw_spin_lock_irqsave(&p->pi_lock, flags);
3346 15554 : smp_mb__after_spinlock();
3347 15554 : if (!(p->state & state))
3348 1531 : goto unlock;
3349 :
3350 14023 : trace_sched_waking(p);
3351 :
3352 : /* We're going to change ->state: */
3353 14023 : success = 1;
3354 :
3355 : /*
3356 : * Ensure we load p->on_rq _after_ p->state, otherwise it would
3357 : * be possible to, falsely, observe p->on_rq == 0 and get stuck
3358 : * in smp_cond_load_acquire() below.
3359 : *
3360 : * sched_ttwu_pending() try_to_wake_up()
3361 : * STORE p->on_rq = 1 LOAD p->state
3362 : * UNLOCK rq->lock
3363 : *
3364 : * __schedule() (switch to task 'p')
3365 : * LOCK rq->lock smp_rmb();
3366 : * smp_mb__after_spinlock();
3367 : * UNLOCK rq->lock
3368 : *
3369 : * [task p]
3370 : * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3371 : *
3372 : * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3373 : * __schedule(). See the comment for smp_mb__after_spinlock().
3374 : *
3375 : * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3376 : */
3377 14023 : smp_rmb();
3378 14024 : if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
3379 162 : goto unlock;
3380 :
3381 : #ifdef CONFIG_SMP
3382 : /*
3383 : * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3384 : * possible to, falsely, observe p->on_cpu == 0.
3385 : *
3386 : * One must be running (->on_cpu == 1) in order to remove oneself
3387 : * from the runqueue.
3388 : *
3389 : * __schedule() (switch to task 'p') try_to_wake_up()
3390 : * STORE p->on_cpu = 1 LOAD p->on_rq
3391 : * UNLOCK rq->lock
3392 : *
3393 : * __schedule() (put 'p' to sleep)
3394 : * LOCK rq->lock smp_rmb();
3395 : * smp_mb__after_spinlock();
3396 : * STORE p->on_rq = 0 LOAD p->on_cpu
3397 : *
3398 : * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3399 : * __schedule(). See the comment for smp_mb__after_spinlock().
3400 : *
3401 : * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3402 : * schedule()'s deactivate_task() has 'happened' and p will no longer
3403 : * care about it's own p->state. See the comment in __schedule().
3404 : */
3405 13862 : smp_acquire__after_ctrl_dep();
3406 :
3407 : /*
3408 : * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3409 : * == 0), which means we need to do an enqueue, change p->state to
3410 : * TASK_WAKING such that we can unlock p->pi_lock before doing the
3411 : * enqueue, such as ttwu_queue_wakelist().
3412 : */
3413 13862 : p->state = TASK_WAKING;
3414 :
3415 : /*
3416 : * If the owning (remote) CPU is still in the middle of schedule() with
3417 : * this task as prev, considering queueing p on the remote CPUs wake_list
3418 : * which potentially sends an IPI instead of spinning on p->on_cpu to
3419 : * let the waker make forward progress. This is safe because IRQs are
3420 : * disabled and the IPI will deliver after on_cpu is cleared.
3421 : *
3422 : * Ensure we load task_cpu(p) after p->on_cpu:
3423 : *
3424 : * set_task_cpu(p, cpu);
3425 : * STORE p->cpu = @cpu
3426 : * __schedule() (switch to task 'p')
3427 : * LOCK rq->lock
3428 : * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
3429 : * STORE p->on_cpu = 1 LOAD p->cpu
3430 : *
3431 : * to ensure we observe the correct CPU on which the task is currently
3432 : * scheduling.
3433 : */
3434 13910 : if (smp_load_acquire(&p->on_cpu) &&
3435 48 : ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
3436 48 : goto unlock;
3437 :
3438 : /*
3439 : * If the owning (remote) CPU is still in the middle of schedule() with
3440 : * this task as prev, wait until it's done referencing the task.
3441 : *
3442 : * Pairs with the smp_store_release() in finish_task().
3443 : *
3444 : * This ensures that tasks getting woken will be fully ordered against
3445 : * their previous state and preserve Program Order.
3446 : */
3447 13813 : smp_cond_load_acquire(&p->on_cpu, !VAL);
3448 :
3449 13813 : cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
3450 13810 : if (task_cpu(p) != cpu) {
3451 110 : if (p->in_iowait) {
3452 0 : delayacct_blkio_end(p);
3453 0 : atomic_dec(&task_rq(p)->nr_iowait);
3454 : }
3455 :
3456 110 : wake_flags |= WF_MIGRATED;
3457 110 : psi_ttwu_dequeue(p);
3458 110 : set_task_cpu(p, cpu);
3459 : }
3460 : #else
3461 : cpu = task_cpu(p);
3462 : #endif /* CONFIG_SMP */
3463 :
3464 13810 : ttwu_queue(p, cpu, wake_flags);
3465 15555 : unlock:
3466 15555 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3467 15554 : out:
3468 15707 : if (success)
3469 14026 : ttwu_stat(p, task_cpu(p), wake_flags);
3470 15707 : preempt_enable();
3471 :
3472 15707 : return success;
3473 : }
3474 :
3475 : /**
3476 : * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
3477 : * @p: Process for which the function is to be invoked, can be @current.
3478 : * @func: Function to invoke.
3479 : * @arg: Argument to function.
3480 : *
3481 : * If the specified task can be quickly locked into a definite state
3482 : * (either sleeping or on a given runqueue), arrange to keep it in that
3483 : * state while invoking @func(@arg). This function can use ->on_rq and
3484 : * task_curr() to work out what the state is, if required. Given that
3485 : * @func can be invoked with a runqueue lock held, it had better be quite
3486 : * lightweight.
3487 : *
3488 : * Returns:
3489 : * @false if the task slipped out from under the locks.
3490 : * @true if the task was locked onto a runqueue or is sleeping.
3491 : * However, @func can override this by returning @false.
3492 : */
3493 0 : bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
3494 : {
3495 0 : struct rq_flags rf;
3496 0 : bool ret = false;
3497 0 : struct rq *rq;
3498 :
3499 0 : raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3500 0 : if (p->on_rq) {
3501 0 : rq = __task_rq_lock(p, &rf);
3502 0 : if (task_rq(p) == rq)
3503 0 : ret = func(p, arg);
3504 0 : rq_unlock(rq, &rf);
3505 : } else {
3506 0 : switch (p->state) {
3507 : case TASK_RUNNING:
3508 : case TASK_WAKING:
3509 : break;
3510 0 : default:
3511 0 : smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3512 0 : if (!p->on_rq)
3513 0 : ret = func(p, arg);
3514 : }
3515 : }
3516 0 : raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
3517 0 : return ret;
3518 : }
3519 :
3520 : /**
3521 : * wake_up_process - Wake up a specific process
3522 : * @p: The process to be woken up.
3523 : *
3524 : * Attempt to wake up the nominated process and move it to the set of runnable
3525 : * processes.
3526 : *
3527 : * Return: 1 if the process was woken up, 0 if it was already running.
3528 : *
3529 : * This function executes a full memory barrier before accessing the task state.
3530 : */
3531 9441 : int wake_up_process(struct task_struct *p)
3532 : {
3533 9441 : return try_to_wake_up(p, TASK_NORMAL, 0);
3534 : }
3535 : EXPORT_SYMBOL(wake_up_process);
3536 :
3537 1520 : int wake_up_state(struct task_struct *p, unsigned int state)
3538 : {
3539 1520 : return try_to_wake_up(p, state, 0);
3540 : }
3541 :
3542 : /*
3543 : * Perform scheduler related setup for a newly forked process p.
3544 : * p is forked by current.
3545 : *
3546 : * __sched_fork() is basic setup used by init_idle() too:
3547 : */
3548 990 : static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
3549 : {
3550 990 : p->on_rq = 0;
3551 :
3552 990 : p->se.on_rq = 0;
3553 990 : p->se.exec_start = 0;
3554 990 : p->se.sum_exec_runtime = 0;
3555 990 : p->se.prev_sum_exec_runtime = 0;
3556 990 : p->se.nr_migrations = 0;
3557 990 : p->se.vruntime = 0;
3558 990 : INIT_LIST_HEAD(&p->se.group_node);
3559 :
3560 : #ifdef CONFIG_FAIR_GROUP_SCHED
3561 : p->se.cfs_rq = NULL;
3562 : #endif
3563 :
3564 : #ifdef CONFIG_SCHEDSTATS
3565 : /* Even if schedstat is disabled, there should not be garbage */
3566 : memset(&p->se.statistics, 0, sizeof(p->se.statistics));
3567 : #endif
3568 :
3569 990 : RB_CLEAR_NODE(&p->dl.rb_node);
3570 990 : init_dl_task_timer(&p->dl);
3571 990 : init_dl_inactive_task_timer(&p->dl);
3572 990 : __dl_clear_params(p);
3573 :
3574 990 : INIT_LIST_HEAD(&p->rt.run_list);
3575 990 : p->rt.timeout = 0;
3576 990 : p->rt.time_slice = sched_rr_timeslice;
3577 990 : p->rt.on_rq = 0;
3578 990 : p->rt.on_list = 0;
3579 :
3580 : #ifdef CONFIG_PREEMPT_NOTIFIERS
3581 : INIT_HLIST_HEAD(&p->preempt_notifiers);
3582 : #endif
3583 :
3584 : #ifdef CONFIG_COMPACTION
3585 990 : p->capture_control = NULL;
3586 : #endif
3587 990 : init_numa_balancing(clone_flags, p);
3588 : #ifdef CONFIG_SMP
3589 990 : p->wake_entry.u_flags = CSD_TYPE_TTWU;
3590 990 : p->migration_pending = NULL;
3591 : #endif
3592 990 : }
3593 :
3594 : DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
3595 :
3596 : #ifdef CONFIG_NUMA_BALANCING
3597 :
3598 : void set_numabalancing_state(bool enabled)
3599 : {
3600 : if (enabled)
3601 : static_branch_enable(&sched_numa_balancing);
3602 : else
3603 : static_branch_disable(&sched_numa_balancing);
3604 : }
3605 :
3606 : #ifdef CONFIG_PROC_SYSCTL
3607 : int sysctl_numa_balancing(struct ctl_table *table, int write,
3608 : void *buffer, size_t *lenp, loff_t *ppos)
3609 : {
3610 : struct ctl_table t;
3611 : int err;
3612 : int state = static_branch_likely(&sched_numa_balancing);
3613 :
3614 : if (write && !capable(CAP_SYS_ADMIN))
3615 : return -EPERM;
3616 :
3617 : t = *table;
3618 : t.data = &state;
3619 : err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3620 : if (err < 0)
3621 : return err;
3622 : if (write)
3623 : set_numabalancing_state(state);
3624 : return err;
3625 : }
3626 : #endif
3627 : #endif
3628 :
3629 : #ifdef CONFIG_SCHEDSTATS
3630 :
3631 : DEFINE_STATIC_KEY_FALSE(sched_schedstats);
3632 : static bool __initdata __sched_schedstats = false;
3633 :
3634 : static void set_schedstats(bool enabled)
3635 : {
3636 : if (enabled)
3637 : static_branch_enable(&sched_schedstats);
3638 : else
3639 : static_branch_disable(&sched_schedstats);
3640 : }
3641 :
3642 : void force_schedstat_enabled(void)
3643 : {
3644 : if (!schedstat_enabled()) {
3645 : pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
3646 : static_branch_enable(&sched_schedstats);
3647 : }
3648 : }
3649 :
3650 : static int __init setup_schedstats(char *str)
3651 : {
3652 : int ret = 0;
3653 : if (!str)
3654 : goto out;
3655 :
3656 : /*
3657 : * This code is called before jump labels have been set up, so we can't
3658 : * change the static branch directly just yet. Instead set a temporary
3659 : * variable so init_schedstats() can do it later.
3660 : */
3661 : if (!strcmp(str, "enable")) {
3662 : __sched_schedstats = true;
3663 : ret = 1;
3664 : } else if (!strcmp(str, "disable")) {
3665 : __sched_schedstats = false;
3666 : ret = 1;
3667 : }
3668 : out:
3669 : if (!ret)
3670 : pr_warn("Unable to parse schedstats=\n");
3671 :
3672 : return ret;
3673 : }
3674 : __setup("schedstats=", setup_schedstats);
3675 :
3676 : static void __init init_schedstats(void)
3677 : {
3678 : set_schedstats(__sched_schedstats);
3679 : }
3680 :
3681 : #ifdef CONFIG_PROC_SYSCTL
3682 : int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
3683 : size_t *lenp, loff_t *ppos)
3684 : {
3685 : struct ctl_table t;
3686 : int err;
3687 : int state = static_branch_likely(&sched_schedstats);
3688 :
3689 : if (write && !capable(CAP_SYS_ADMIN))
3690 : return -EPERM;
3691 :
3692 : t = *table;
3693 : t.data = &state;
3694 : err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3695 : if (err < 0)
3696 : return err;
3697 : if (write)
3698 : set_schedstats(state);
3699 : return err;
3700 : }
3701 : #endif /* CONFIG_PROC_SYSCTL */
3702 : #else /* !CONFIG_SCHEDSTATS */
3703 1 : static inline void init_schedstats(void) {}
3704 : #endif /* CONFIG_SCHEDSTATS */
3705 :
3706 : /*
3707 : * fork()/clone()-time setup:
3708 : */
3709 980 : int sched_fork(unsigned long clone_flags, struct task_struct *p)
3710 : {
3711 980 : unsigned long flags;
3712 :
3713 980 : __sched_fork(clone_flags, p);
3714 : /*
3715 : * We mark the process as NEW here. This guarantees that
3716 : * nobody will actually run it, and a signal or other external
3717 : * event cannot wake it up and insert it on the runqueue either.
3718 : */
3719 980 : p->state = TASK_NEW;
3720 :
3721 : /*
3722 : * Make sure we do not leak PI boosting priority to the child.
3723 : */
3724 980 : p->prio = current->normal_prio;
3725 :
3726 980 : uclamp_fork(p);
3727 :
3728 : /*
3729 : * Revert to default priority/policy on fork if requested.
3730 : */
3731 980 : if (unlikely(p->sched_reset_on_fork)) {
3732 0 : if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3733 0 : p->policy = SCHED_NORMAL;
3734 0 : p->static_prio = NICE_TO_PRIO(0);
3735 0 : p->rt_priority = 0;
3736 0 : } else if (PRIO_TO_NICE(p->static_prio) < 0)
3737 0 : p->static_prio = NICE_TO_PRIO(0);
3738 :
3739 0 : p->prio = p->normal_prio = __normal_prio(p);
3740 0 : set_load_weight(p, false);
3741 :
3742 : /*
3743 : * We don't need the reset flag anymore after the fork. It has
3744 : * fulfilled its duty:
3745 : */
3746 0 : p->sched_reset_on_fork = 0;
3747 : }
3748 :
3749 980 : if (dl_prio(p->prio))
3750 : return -EAGAIN;
3751 980 : else if (rt_prio(p->prio))
3752 0 : p->sched_class = &rt_sched_class;
3753 : else
3754 980 : p->sched_class = &fair_sched_class;
3755 :
3756 980 : init_entity_runnable_average(&p->se);
3757 :
3758 : /*
3759 : * The child is not yet in the pid-hash so no cgroup attach races,
3760 : * and the cgroup is pinned to this child due to cgroup_fork()
3761 : * is ran before sched_fork().
3762 : *
3763 : * Silence PROVE_RCU.
3764 : */
3765 980 : raw_spin_lock_irqsave(&p->pi_lock, flags);
3766 980 : rseq_migrate(p);
3767 : /*
3768 : * We're setting the CPU for the first time, we don't migrate,
3769 : * so use __set_task_cpu().
3770 : */
3771 980 : __set_task_cpu(p, smp_processor_id());
3772 980 : if (p->sched_class->task_fork)
3773 980 : p->sched_class->task_fork(p);
3774 980 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3775 :
3776 : #ifdef CONFIG_SCHED_INFO
3777 980 : if (likely(sched_info_on()))
3778 980 : memset(&p->sched_info, 0, sizeof(p->sched_info));
3779 : #endif
3780 : #if defined(CONFIG_SMP)
3781 980 : p->on_cpu = 0;
3782 : #endif
3783 980 : init_task_preempt_count(p);
3784 : #ifdef CONFIG_SMP
3785 980 : plist_node_init(&p->pushable_tasks, MAX_PRIO);
3786 980 : RB_CLEAR_NODE(&p->pushable_dl_tasks);
3787 : #endif
3788 980 : return 0;
3789 : }
3790 :
3791 980 : void sched_post_fork(struct task_struct *p)
3792 : {
3793 980 : uclamp_post_fork(p);
3794 980 : }
3795 :
3796 10 : unsigned long to_ratio(u64 period, u64 runtime)
3797 : {
3798 10 : if (runtime == RUNTIME_INF)
3799 : return BW_UNIT;
3800 :
3801 : /*
3802 : * Doing this here saves a lot of checks in all
3803 : * the calling paths, and returning zero seems
3804 : * safe for them anyway.
3805 : */
3806 10 : if (period == 0)
3807 : return 0;
3808 :
3809 10 : return div64_u64(runtime << BW_SHIFT, period);
3810 : }
3811 :
3812 : /*
3813 : * wake_up_new_task - wake up a newly created task for the first time.
3814 : *
3815 : * This function will do some initial scheduler statistics housekeeping
3816 : * that must be done for every newly created context, then puts the task
3817 : * on the runqueue and wakes it.
3818 : */
3819 977 : void wake_up_new_task(struct task_struct *p)
3820 : {
3821 977 : struct rq_flags rf;
3822 977 : struct rq *rq;
3823 :
3824 977 : raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3825 977 : p->state = TASK_RUNNING;
3826 : #ifdef CONFIG_SMP
3827 : /*
3828 : * Fork balancing, do it here and not earlier because:
3829 : * - cpus_ptr can change in the fork path
3830 : * - any previously selected CPU might disappear through hotplug
3831 : *
3832 : * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3833 : * as we're not fully set-up yet.
3834 : */
3835 977 : p->recent_used_cpu = task_cpu(p);
3836 977 : rseq_migrate(p);
3837 977 : __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
3838 : #endif
3839 977 : rq = __task_rq_lock(p, &rf);
3840 977 : update_rq_clock(rq);
3841 977 : post_init_entity_util_avg(p);
3842 :
3843 977 : activate_task(rq, p, ENQUEUE_NOCLOCK);
3844 977 : trace_sched_wakeup_new(p);
3845 977 : check_preempt_curr(rq, p, WF_FORK);
3846 : #ifdef CONFIG_SMP
3847 977 : if (p->sched_class->task_woken) {
3848 : /*
3849 : * Nothing relies on rq->lock after this, so it's fine to
3850 : * drop it.
3851 : */
3852 0 : rq_unpin_lock(rq, &rf);
3853 0 : p->sched_class->task_woken(rq, p);
3854 0 : rq_repin_lock(rq, &rf);
3855 : }
3856 : #endif
3857 977 : task_rq_unlock(rq, p, &rf);
3858 977 : }
3859 :
3860 : #ifdef CONFIG_PREEMPT_NOTIFIERS
3861 :
3862 : static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
3863 :
3864 : void preempt_notifier_inc(void)
3865 : {
3866 : static_branch_inc(&preempt_notifier_key);
3867 : }
3868 : EXPORT_SYMBOL_GPL(preempt_notifier_inc);
3869 :
3870 : void preempt_notifier_dec(void)
3871 : {
3872 : static_branch_dec(&preempt_notifier_key);
3873 : }
3874 : EXPORT_SYMBOL_GPL(preempt_notifier_dec);
3875 :
3876 : /**
3877 : * preempt_notifier_register - tell me when current is being preempted & rescheduled
3878 : * @notifier: notifier struct to register
3879 : */
3880 : void preempt_notifier_register(struct preempt_notifier *notifier)
3881 : {
3882 : if (!static_branch_unlikely(&preempt_notifier_key))
3883 : WARN(1, "registering preempt_notifier while notifiers disabled\n");
3884 :
3885 : hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3886 : }
3887 : EXPORT_SYMBOL_GPL(preempt_notifier_register);
3888 :
3889 : /**
3890 : * preempt_notifier_unregister - no longer interested in preemption notifications
3891 : * @notifier: notifier struct to unregister
3892 : *
3893 : * This is *not* safe to call from within a preemption notifier.
3894 : */
3895 : void preempt_notifier_unregister(struct preempt_notifier *notifier)
3896 : {
3897 : hlist_del(¬ifier->link);
3898 : }
3899 : EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3900 :
3901 : static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3902 : {
3903 : struct preempt_notifier *notifier;
3904 :
3905 : hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3906 : notifier->ops->sched_in(notifier, raw_smp_processor_id());
3907 : }
3908 :
3909 : static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3910 : {
3911 : if (static_branch_unlikely(&preempt_notifier_key))
3912 : __fire_sched_in_preempt_notifiers(curr);
3913 : }
3914 :
3915 : static void
3916 : __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3917 : struct task_struct *next)
3918 : {
3919 : struct preempt_notifier *notifier;
3920 :
3921 : hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3922 : notifier->ops->sched_out(notifier, next);
3923 : }
3924 :
3925 : static __always_inline void
3926 : fire_sched_out_preempt_notifiers(struct task_struct *curr,
3927 : struct task_struct *next)
3928 : {
3929 : if (static_branch_unlikely(&preempt_notifier_key))
3930 : __fire_sched_out_preempt_notifiers(curr, next);
3931 : }
3932 :
3933 : #else /* !CONFIG_PREEMPT_NOTIFIERS */
3934 :
3935 26515 : static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3936 : {
3937 26515 : }
3938 :
3939 : static inline void
3940 26512 : fire_sched_out_preempt_notifiers(struct task_struct *curr,
3941 : struct task_struct *next)
3942 : {
3943 26512 : }
3944 :
3945 : #endif /* CONFIG_PREEMPT_NOTIFIERS */
3946 :
3947 26512 : static inline void prepare_task(struct task_struct *next)
3948 : {
3949 : #ifdef CONFIG_SMP
3950 : /*
3951 : * Claim the task as running, we do this before switching to it
3952 : * such that any running task will have this set.
3953 : *
3954 : * See the ttwu() WF_ON_CPU case and its ordering comment.
3955 : */
3956 26512 : WRITE_ONCE(next->on_cpu, 1);
3957 : #endif
3958 : }
3959 :
3960 26512 : static inline void finish_task(struct task_struct *prev)
3961 : {
3962 : #ifdef CONFIG_SMP
3963 : /*
3964 : * This must be the very last reference to @prev from this CPU. After
3965 : * p->on_cpu is cleared, the task can be moved to a different CPU. We
3966 : * must ensure this doesn't happen until the switch is completely
3967 : * finished.
3968 : *
3969 : * In particular, the load of prev->state in finish_task_switch() must
3970 : * happen before this.
3971 : *
3972 : * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3973 : */
3974 53025 : smp_store_release(&prev->on_cpu, 0);
3975 : #endif
3976 : }
3977 :
3978 : #ifdef CONFIG_SMP
3979 :
3980 27956 : static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
3981 : {
3982 27956 : void (*func)(struct rq *rq);
3983 27956 : struct callback_head *next;
3984 :
3985 55921 : lockdep_assert_held(&rq->lock);
3986 :
3987 27966 : while (head) {
3988 0 : func = (void (*)(struct rq *))head->func;
3989 0 : next = head->next;
3990 0 : head->next = NULL;
3991 0 : head = next;
3992 :
3993 0 : func(rq);
3994 : }
3995 27966 : }
3996 :
3997 : static void balance_push(struct rq *rq);
3998 :
3999 : struct callback_head balance_push_callback = {
4000 : .next = NULL,
4001 : .func = (void (*)(struct callback_head *))balance_push,
4002 : };
4003 :
4004 27965 : static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4005 : {
4006 27965 : struct callback_head *head = rq->balance_callback;
4007 :
4008 55931 : lockdep_assert_held(&rq->lock);
4009 27966 : if (head)
4010 0 : rq->balance_callback = NULL;
4011 :
4012 27966 : return head;
4013 : }
4014 :
4015 27959 : static void __balance_callbacks(struct rq *rq)
4016 : {
4017 27959 : do_balance_callbacks(rq, splice_balance_callbacks(rq));
4018 27966 : }
4019 :
4020 4 : static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4021 : {
4022 4 : unsigned long flags;
4023 :
4024 4 : if (unlikely(head)) {
4025 0 : raw_spin_lock_irqsave(&rq->lock, flags);
4026 0 : do_balance_callbacks(rq, head);
4027 0 : raw_spin_unlock_irqrestore(&rq->lock, flags);
4028 : }
4029 4 : }
4030 :
4031 : #else
4032 :
4033 : static inline void __balance_callbacks(struct rq *rq)
4034 : {
4035 : }
4036 :
4037 : static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4038 : {
4039 : return NULL;
4040 : }
4041 :
4042 : static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4043 : {
4044 : }
4045 :
4046 : #endif
4047 :
4048 : static inline void
4049 26511 : prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4050 : {
4051 : /*
4052 : * Since the runqueue lock will be released by the next
4053 : * task (which is an invalid locking op but in the case
4054 : * of the scheduler it's an obvious special-case), so we
4055 : * do an early lockdep release here:
4056 : */
4057 26511 : rq_unpin_lock(rq, rf);
4058 26511 : spin_release(&rq->lock.dep_map, _THIS_IP_);
4059 : #ifdef CONFIG_DEBUG_SPINLOCK
4060 : /* this is a valid case when another task releases the spinlock */
4061 26507 : rq->lock.owner = next;
4062 : #endif
4063 26507 : }
4064 :
4065 26509 : static inline void finish_lock_switch(struct rq *rq)
4066 : {
4067 : /*
4068 : * If we are tracking spinlock dependencies then we have to
4069 : * fix up the runqueue lock - which gets 'carried over' from
4070 : * prev into current:
4071 : */
4072 26509 : spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
4073 26508 : __balance_callbacks(rq);
4074 26514 : raw_spin_unlock_irq(&rq->lock);
4075 26514 : }
4076 :
4077 : /*
4078 : * NOP if the arch has not defined these:
4079 : */
4080 :
4081 : #ifndef prepare_arch_switch
4082 : # define prepare_arch_switch(next) do { } while (0)
4083 : #endif
4084 :
4085 : #ifndef finish_arch_post_lock_switch
4086 : # define finish_arch_post_lock_switch() do { } while (0)
4087 : #endif
4088 :
4089 26512 : static inline void kmap_local_sched_out(void)
4090 : {
4091 : #ifdef CONFIG_KMAP_LOCAL
4092 : if (unlikely(current->kmap_ctrl.idx))
4093 : __kmap_local_sched_out();
4094 : #endif
4095 26512 : }
4096 :
4097 26515 : static inline void kmap_local_sched_in(void)
4098 : {
4099 : #ifdef CONFIG_KMAP_LOCAL
4100 : if (unlikely(current->kmap_ctrl.idx))
4101 : __kmap_local_sched_in();
4102 : #endif
4103 26515 : }
4104 :
4105 : /**
4106 : * prepare_task_switch - prepare to switch tasks
4107 : * @rq: the runqueue preparing to switch
4108 : * @prev: the current task that is being switched out
4109 : * @next: the task we are going to switch to.
4110 : *
4111 : * This is called with the rq lock held and interrupts off. It must
4112 : * be paired with a subsequent finish_task_switch after the context
4113 : * switch.
4114 : *
4115 : * prepare_task_switch sets up locking and calls architecture specific
4116 : * hooks.
4117 : */
4118 : static inline void
4119 26511 : prepare_task_switch(struct rq *rq, struct task_struct *prev,
4120 : struct task_struct *next)
4121 : {
4122 26511 : kcov_prepare_switch(prev);
4123 26511 : sched_info_switch(rq, prev, next);
4124 26511 : perf_event_task_sched_out(prev, next);
4125 26512 : rseq_preempt(prev);
4126 26512 : fire_sched_out_preempt_notifiers(prev, next);
4127 26512 : kmap_local_sched_out();
4128 26512 : prepare_task(next);
4129 26512 : prepare_arch_switch(next);
4130 26512 : }
4131 :
4132 : /**
4133 : * finish_task_switch - clean up after a task-switch
4134 : * @prev: the thread we just switched away from.
4135 : *
4136 : * finish_task_switch must be called after the context switch, paired
4137 : * with a prepare_task_switch call before the context switch.
4138 : * finish_task_switch will reconcile locking set up by prepare_task_switch,
4139 : * and do any other architecture-specific cleanup actions.
4140 : *
4141 : * Note that we may have delayed dropping an mm in context_switch(). If
4142 : * so, we finish that here outside of the runqueue lock. (Doing it
4143 : * with the lock held can cause deadlocks; see schedule() for
4144 : * details.)
4145 : *
4146 : * The context switch have flipped the stack from under us and restored the
4147 : * local variables which were saved when this task called schedule() in the
4148 : * past. prev == current is still correct but we need to recalculate this_rq
4149 : * because prev may have moved to another CPU.
4150 : */
4151 26508 : static struct rq *finish_task_switch(struct task_struct *prev)
4152 : __releases(rq->lock)
4153 : {
4154 26508 : struct rq *rq = this_rq();
4155 26509 : struct mm_struct *mm = rq->prev_mm;
4156 26509 : long prev_state;
4157 :
4158 : /*
4159 : * The previous task will have left us with a preempt_count of 2
4160 : * because it left us after:
4161 : *
4162 : * schedule()
4163 : * preempt_disable(); // 1
4164 : * __schedule()
4165 : * raw_spin_lock_irq(&rq->lock) // 2
4166 : *
4167 : * Also, see FORK_PREEMPT_COUNT.
4168 : */
4169 26509 : if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4170 : "corrupted preempt_count: %s/%d/0x%x\n",
4171 : current->comm, current->pid, preempt_count()))
4172 0 : preempt_count_set(FORK_PREEMPT_COUNT);
4173 :
4174 26509 : rq->prev_mm = NULL;
4175 :
4176 : /*
4177 : * A task struct has one reference for the use as "current".
4178 : * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4179 : * schedule one last time. The schedule call will never return, and
4180 : * the scheduled task must drop that reference.
4181 : *
4182 : * We must observe prev->state before clearing prev->on_cpu (in
4183 : * finish_task), otherwise a concurrent wakeup can get prev
4184 : * running on another CPU and we could rave with its RUNNING -> DEAD
4185 : * transition, resulting in a double drop.
4186 : */
4187 26509 : prev_state = prev->state;
4188 26509 : vtime_task_switch(prev);
4189 26509 : perf_event_task_sched_in(prev, current);
4190 26512 : finish_task(prev);
4191 26513 : finish_lock_switch(rq);
4192 26515 : finish_arch_post_lock_switch();
4193 26515 : kcov_finish_switch(current);
4194 : /*
4195 : * kmap_local_sched_out() is invoked with rq::lock held and
4196 : * interrupts disabled. There is no requirement for that, but the
4197 : * sched out code does not have an interrupt enabled section.
4198 : * Restoring the maps on sched in does not require interrupts being
4199 : * disabled either.
4200 : */
4201 26515 : kmap_local_sched_in();
4202 :
4203 26515 : fire_sched_in_preempt_notifiers(current);
4204 : /*
4205 : * When switching through a kernel thread, the loop in
4206 : * membarrier_{private,global}_expedited() may have observed that
4207 : * kernel thread and not issued an IPI. It is therefore possible to
4208 : * schedule between user->kernel->user threads without passing though
4209 : * switch_mm(). Membarrier requires a barrier after storing to
4210 : * rq->curr, before returning to userspace, so provide them here:
4211 : *
4212 : * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4213 : * provided by mmdrop(),
4214 : * - a sync_core for SYNC_CORE.
4215 : */
4216 26515 : if (mm) {
4217 7345 : membarrier_mm_sync_core_before_usermode(mm);
4218 7345 : mmdrop(mm);
4219 : }
4220 26516 : if (unlikely(prev_state == TASK_DEAD)) {
4221 900 : if (prev->sched_class->task_dead)
4222 900 : prev->sched_class->task_dead(prev);
4223 :
4224 : /*
4225 : * Remove function-return probe instances associated with this
4226 : * task and put them back on the free list.
4227 : */
4228 900 : kprobe_flush_task(prev);
4229 :
4230 : /* Task is done with its stack. */
4231 900 : put_task_stack(prev);
4232 :
4233 900 : put_task_struct_rcu_user(prev);
4234 : }
4235 :
4236 26516 : tick_nohz_task_switch();
4237 26516 : return rq;
4238 : }
4239 :
4240 : /**
4241 : * schedule_tail - first thing a freshly forked thread must call.
4242 : * @prev: the thread we just switched away from.
4243 : */
4244 976 : asmlinkage __visible void schedule_tail(struct task_struct *prev)
4245 : __releases(rq->lock)
4246 : {
4247 976 : struct rq *rq;
4248 :
4249 : /*
4250 : * New tasks start with FORK_PREEMPT_COUNT, see there and
4251 : * finish_task_switch() for details.
4252 : *
4253 : * finish_task_switch() will drop rq->lock() and lower preempt_count
4254 : * and the preempt_enable() will end up enabling preemption (on
4255 : * PREEMPT_COUNT kernels).
4256 : */
4257 :
4258 976 : rq = finish_task_switch(prev);
4259 977 : preempt_enable();
4260 :
4261 977 : if (current->set_child_tid)
4262 917 : put_user(task_pid_vnr(current), current->set_child_tid);
4263 :
4264 977 : calculate_sigpending();
4265 977 : }
4266 :
4267 : /*
4268 : * context_switch - switch to the new MM and the new thread's register state.
4269 : */
4270 : static __always_inline struct rq *
4271 26512 : context_switch(struct rq *rq, struct task_struct *prev,
4272 : struct task_struct *next, struct rq_flags *rf)
4273 : {
4274 26512 : prepare_task_switch(rq, prev, next);
4275 :
4276 : /*
4277 : * For paravirt, this is coupled with an exit in switch_to to
4278 : * combine the page table reload and the switch backend into
4279 : * one hypercall.
4280 : */
4281 26512 : arch_start_context_switch(prev);
4282 :
4283 : /*
4284 : * kernel -> kernel lazy + transfer active
4285 : * user -> kernel lazy + mmgrab() active
4286 : *
4287 : * kernel -> user switch + mmdrop() active
4288 : * user -> user switch
4289 : */
4290 26512 : if (!next->mm) { // to kernel
4291 16097 : enter_lazy_tlb(prev->active_mm, next);
4292 :
4293 16097 : next->active_mm = prev->active_mm;
4294 16097 : if (prev->mm) // from user
4295 6446 : mmgrab(prev->active_mm);
4296 : else
4297 9651 : prev->active_mm = NULL;
4298 : } else { // to user
4299 10415 : membarrier_switch_mm(rq, prev->active_mm, next->mm);
4300 : /*
4301 : * sys_membarrier() requires an smp_mb() between setting
4302 : * rq->curr / membarrier_switch_mm() and returning to userspace.
4303 : *
4304 : * The below provides this either through switch_mm(), or in
4305 : * case 'prev->active_mm == next->mm' through
4306 : * finish_task_switch()'s mmdrop().
4307 : */
4308 10415 : switch_mm_irqs_off(prev->active_mm, next->mm, next);
4309 :
4310 10415 : if (!prev->mm) { // from kernel
4311 : /* will mmdrop() in finish_task_switch(). */
4312 7346 : rq->prev_mm = prev->active_mm;
4313 7346 : prev->active_mm = NULL;
4314 : }
4315 : }
4316 :
4317 26512 : rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4318 :
4319 26512 : prepare_lock_switch(rq, next, rf);
4320 :
4321 : /* Here we just switch the register state and the stack. */
4322 26507 : switch_to(prev, next, prev);
4323 25536 : barrier();
4324 :
4325 25537 : return finish_task_switch(prev);
4326 : }
4327 :
4328 : /*
4329 : * nr_running and nr_context_switches:
4330 : *
4331 : * externally visible scheduler statistics: current number of runnable
4332 : * threads, total number of context switches performed since bootup.
4333 : */
4334 0 : unsigned long nr_running(void)
4335 : {
4336 0 : unsigned long i, sum = 0;
4337 :
4338 0 : for_each_online_cpu(i)
4339 0 : sum += cpu_rq(i)->nr_running;
4340 :
4341 0 : return sum;
4342 : }
4343 :
4344 : /*
4345 : * Check if only the current task is running on the CPU.
4346 : *
4347 : * Caution: this function does not check that the caller has disabled
4348 : * preemption, thus the result might have a time-of-check-to-time-of-use
4349 : * race. The caller is responsible to use it correctly, for example:
4350 : *
4351 : * - from a non-preemptible section (of course)
4352 : *
4353 : * - from a thread that is bound to a single CPU
4354 : *
4355 : * - in a loop with very short iterations (e.g. a polling loop)
4356 : */
4357 0 : bool single_task_running(void)
4358 : {
4359 0 : return raw_rq()->nr_running == 1;
4360 : }
4361 : EXPORT_SYMBOL(single_task_running);
4362 :
4363 1 : unsigned long long nr_context_switches(void)
4364 : {
4365 1 : int i;
4366 1 : unsigned long long sum = 0;
4367 :
4368 5 : for_each_possible_cpu(i)
4369 4 : sum += cpu_rq(i)->nr_switches;
4370 :
4371 1 : return sum;
4372 : }
4373 :
4374 : /*
4375 : * Consumers of these two interfaces, like for example the cpuidle menu
4376 : * governor, are using nonsensical data. Preferring shallow idle state selection
4377 : * for a CPU that has IO-wait which might not even end up running the task when
4378 : * it does become runnable.
4379 : */
4380 :
4381 14964 : unsigned long nr_iowait_cpu(int cpu)
4382 : {
4383 14964 : return atomic_read(&cpu_rq(cpu)->nr_iowait);
4384 : }
4385 :
4386 : /*
4387 : * IO-wait accounting, and how it's mostly bollocks (on SMP).
4388 : *
4389 : * The idea behind IO-wait account is to account the idle time that we could
4390 : * have spend running if it were not for IO. That is, if we were to improve the
4391 : * storage performance, we'd have a proportional reduction in IO-wait time.
4392 : *
4393 : * This all works nicely on UP, where, when a task blocks on IO, we account
4394 : * idle time as IO-wait, because if the storage were faster, it could've been
4395 : * running and we'd not be idle.
4396 : *
4397 : * This has been extended to SMP, by doing the same for each CPU. This however
4398 : * is broken.
4399 : *
4400 : * Imagine for instance the case where two tasks block on one CPU, only the one
4401 : * CPU will have IO-wait accounted, while the other has regular idle. Even
4402 : * though, if the storage were faster, both could've ran at the same time,
4403 : * utilising both CPUs.
4404 : *
4405 : * This means, that when looking globally, the current IO-wait accounting on
4406 : * SMP is a lower bound, by reason of under accounting.
4407 : *
4408 : * Worse, since the numbers are provided per CPU, they are sometimes
4409 : * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4410 : * associated with any one particular CPU, it can wake to another CPU than it
4411 : * blocked on. This means the per CPU IO-wait number is meaningless.
4412 : *
4413 : * Task CPU affinities can make all that even more 'interesting'.
4414 : */
4415 :
4416 0 : unsigned long nr_iowait(void)
4417 : {
4418 0 : unsigned long i, sum = 0;
4419 :
4420 0 : for_each_possible_cpu(i)
4421 0 : sum += nr_iowait_cpu(i);
4422 :
4423 0 : return sum;
4424 : }
4425 :
4426 : #ifdef CONFIG_SMP
4427 :
4428 : /*
4429 : * sched_exec - execve() is a valuable balancing opportunity, because at
4430 : * this point the task has the smallest effective memory and cache footprint.
4431 : */
4432 615 : void sched_exec(void)
4433 : {
4434 615 : struct task_struct *p = current;
4435 615 : unsigned long flags;
4436 615 : int dest_cpu;
4437 :
4438 615 : raw_spin_lock_irqsave(&p->pi_lock, flags);
4439 615 : dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
4440 615 : if (dest_cpu == smp_processor_id())
4441 584 : goto unlock;
4442 :
4443 31 : if (likely(cpu_active(dest_cpu))) {
4444 31 : struct migration_arg arg = { p, dest_cpu };
4445 :
4446 31 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4447 31 : stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
4448 31 : return;
4449 : }
4450 0 : unlock:
4451 584 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4452 : }
4453 :
4454 : #endif
4455 :
4456 : DEFINE_PER_CPU(struct kernel_stat, kstat);
4457 : DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
4458 :
4459 : EXPORT_PER_CPU_SYMBOL(kstat);
4460 : EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
4461 :
4462 : /*
4463 : * The function fair_sched_class.update_curr accesses the struct curr
4464 : * and its field curr->exec_start; when called from task_sched_runtime(),
4465 : * we observe a high rate of cache misses in practice.
4466 : * Prefetching this data results in improved performance.
4467 : */
4468 49 : static inline void prefetch_curr_exec_start(struct task_struct *p)
4469 : {
4470 : #ifdef CONFIG_FAIR_GROUP_SCHED
4471 : struct sched_entity *curr = (&p->se)->cfs_rq->curr;
4472 : #else
4473 49 : struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
4474 : #endif
4475 49 : prefetch(curr);
4476 49 : prefetch(&curr->exec_start);
4477 : }
4478 :
4479 : /*
4480 : * Return accounted runtime for the task.
4481 : * In case the task is currently running, return the runtime plus current's
4482 : * pending runtime that have not been accounted yet.
4483 : */
4484 49 : unsigned long long task_sched_runtime(struct task_struct *p)
4485 : {
4486 49 : struct rq_flags rf;
4487 49 : struct rq *rq;
4488 49 : u64 ns;
4489 :
4490 : #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
4491 : /*
4492 : * 64-bit doesn't need locks to atomically read a 64-bit value.
4493 : * So we have a optimization chance when the task's delta_exec is 0.
4494 : * Reading ->on_cpu is racy, but this is ok.
4495 : *
4496 : * If we race with it leaving CPU, we'll take a lock. So we're correct.
4497 : * If we race with it entering CPU, unaccounted time is 0. This is
4498 : * indistinguishable from the read occurring a few cycles earlier.
4499 : * If we see ->on_cpu without ->on_rq, the task is leaving, and has
4500 : * been accounted, so we're correct here as well.
4501 : */
4502 49 : if (!p->on_cpu || !task_on_rq_queued(p))
4503 0 : return p->se.sum_exec_runtime;
4504 : #endif
4505 :
4506 49 : rq = task_rq_lock(p, &rf);
4507 : /*
4508 : * Must be ->curr _and_ ->on_rq. If dequeued, we would
4509 : * project cycles that may never be accounted to this
4510 : * thread, breaking clock_gettime().
4511 : */
4512 49 : if (task_current(rq, p) && task_on_rq_queued(p)) {
4513 49 : prefetch_curr_exec_start(p);
4514 49 : update_rq_clock(rq);
4515 49 : p->sched_class->update_curr(rq);
4516 : }
4517 49 : ns = p->se.sum_exec_runtime;
4518 49 : task_rq_unlock(rq, p, &rf);
4519 :
4520 49 : return ns;
4521 : }
4522 :
4523 : /*
4524 : * This function gets called by the timer code, with HZ frequency.
4525 : * We call it with interrupts disabled.
4526 : */
4527 24506 : void scheduler_tick(void)
4528 : {
4529 24506 : int cpu = smp_processor_id();
4530 24506 : struct rq *rq = cpu_rq(cpu);
4531 24506 : struct task_struct *curr = rq->curr;
4532 24506 : struct rq_flags rf;
4533 24506 : unsigned long thermal_pressure;
4534 :
4535 24506 : arch_scale_freq_tick();
4536 24192 : sched_clock_tick();
4537 :
4538 24363 : rq_lock(rq, &rf);
4539 :
4540 24691 : update_rq_clock(rq);
4541 25079 : thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
4542 25079 : update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
4543 25197 : curr->sched_class->task_tick(rq, curr, 0);
4544 24910 : calc_global_load_tick(rq);
4545 24877 : psi_task_tick(rq);
4546 :
4547 24877 : rq_unlock(rq, &rf);
4548 :
4549 25344 : perf_event_task_tick();
4550 :
4551 : #ifdef CONFIG_SMP
4552 25087 : rq->idle_balance = idle_cpu(cpu);
4553 25087 : trigger_load_balance(rq);
4554 : #endif
4555 24998 : }
4556 :
4557 : #ifdef CONFIG_NO_HZ_FULL
4558 :
4559 : struct tick_work {
4560 : int cpu;
4561 : atomic_t state;
4562 : struct delayed_work work;
4563 : };
4564 : /* Values for ->state, see diagram below. */
4565 : #define TICK_SCHED_REMOTE_OFFLINE 0
4566 : #define TICK_SCHED_REMOTE_OFFLINING 1
4567 : #define TICK_SCHED_REMOTE_RUNNING 2
4568 :
4569 : /*
4570 : * State diagram for ->state:
4571 : *
4572 : *
4573 : * TICK_SCHED_REMOTE_OFFLINE
4574 : * | ^
4575 : * | |
4576 : * | | sched_tick_remote()
4577 : * | |
4578 : * | |
4579 : * +--TICK_SCHED_REMOTE_OFFLINING
4580 : * | ^
4581 : * | |
4582 : * sched_tick_start() | | sched_tick_stop()
4583 : * | |
4584 : * V |
4585 : * TICK_SCHED_REMOTE_RUNNING
4586 : *
4587 : *
4588 : * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
4589 : * and sched_tick_start() are happy to leave the state in RUNNING.
4590 : */
4591 :
4592 : static struct tick_work __percpu *tick_work_cpu;
4593 :
4594 : static void sched_tick_remote(struct work_struct *work)
4595 : {
4596 : struct delayed_work *dwork = to_delayed_work(work);
4597 : struct tick_work *twork = container_of(dwork, struct tick_work, work);
4598 : int cpu = twork->cpu;
4599 : struct rq *rq = cpu_rq(cpu);
4600 : struct task_struct *curr;
4601 : struct rq_flags rf;
4602 : u64 delta;
4603 : int os;
4604 :
4605 : /*
4606 : * Handle the tick only if it appears the remote CPU is running in full
4607 : * dynticks mode. The check is racy by nature, but missing a tick or
4608 : * having one too much is no big deal because the scheduler tick updates
4609 : * statistics and checks timeslices in a time-independent way, regardless
4610 : * of when exactly it is running.
4611 : */
4612 : if (!tick_nohz_tick_stopped_cpu(cpu))
4613 : goto out_requeue;
4614 :
4615 : rq_lock_irq(rq, &rf);
4616 : curr = rq->curr;
4617 : if (cpu_is_offline(cpu))
4618 : goto out_unlock;
4619 :
4620 : update_rq_clock(rq);
4621 :
4622 : if (!is_idle_task(curr)) {
4623 : /*
4624 : * Make sure the next tick runs within a reasonable
4625 : * amount of time.
4626 : */
4627 : delta = rq_clock_task(rq) - curr->se.exec_start;
4628 : WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
4629 : }
4630 : curr->sched_class->task_tick(rq, curr, 0);
4631 :
4632 : calc_load_nohz_remote(rq);
4633 : out_unlock:
4634 : rq_unlock_irq(rq, &rf);
4635 : out_requeue:
4636 :
4637 : /*
4638 : * Run the remote tick once per second (1Hz). This arbitrary
4639 : * frequency is large enough to avoid overload but short enough
4640 : * to keep scheduler internal stats reasonably up to date. But
4641 : * first update state to reflect hotplug activity if required.
4642 : */
4643 : os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
4644 : WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
4645 : if (os == TICK_SCHED_REMOTE_RUNNING)
4646 : queue_delayed_work(system_unbound_wq, dwork, HZ);
4647 : }
4648 :
4649 : static void sched_tick_start(int cpu)
4650 : {
4651 : int os;
4652 : struct tick_work *twork;
4653 :
4654 : if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4655 : return;
4656 :
4657 : WARN_ON_ONCE(!tick_work_cpu);
4658 :
4659 : twork = per_cpu_ptr(tick_work_cpu, cpu);
4660 : os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
4661 : WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
4662 : if (os == TICK_SCHED_REMOTE_OFFLINE) {
4663 : twork->cpu = cpu;
4664 : INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
4665 : queue_delayed_work(system_unbound_wq, &twork->work, HZ);
4666 : }
4667 : }
4668 :
4669 : #ifdef CONFIG_HOTPLUG_CPU
4670 : static void sched_tick_stop(int cpu)
4671 : {
4672 : struct tick_work *twork;
4673 : int os;
4674 :
4675 : if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4676 : return;
4677 :
4678 : WARN_ON_ONCE(!tick_work_cpu);
4679 :
4680 : twork = per_cpu_ptr(tick_work_cpu, cpu);
4681 : /* There cannot be competing actions, but don't rely on stop-machine. */
4682 : os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
4683 : WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
4684 : /* Don't cancel, as this would mess up the state machine. */
4685 : }
4686 : #endif /* CONFIG_HOTPLUG_CPU */
4687 :
4688 : int __init sched_tick_offload_init(void)
4689 : {
4690 : tick_work_cpu = alloc_percpu(struct tick_work);
4691 : BUG_ON(!tick_work_cpu);
4692 : return 0;
4693 : }
4694 :
4695 : #else /* !CONFIG_NO_HZ_FULL */
4696 4 : static inline void sched_tick_start(int cpu) { }
4697 0 : static inline void sched_tick_stop(int cpu) { }
4698 : #endif
4699 :
4700 : #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
4701 : defined(CONFIG_TRACE_PREEMPT_TOGGLE))
4702 : /*
4703 : * If the value passed in is equal to the current preempt count
4704 : * then we just disabled preemption. Start timing the latency.
4705 : */
4706 : static inline void preempt_latency_start(int val)
4707 : {
4708 : if (preempt_count() == val) {
4709 : unsigned long ip = get_lock_parent_ip();
4710 : #ifdef CONFIG_DEBUG_PREEMPT
4711 : current->preempt_disable_ip = ip;
4712 : #endif
4713 : trace_preempt_off(CALLER_ADDR0, ip);
4714 : }
4715 : }
4716 :
4717 : void preempt_count_add(int val)
4718 : {
4719 : #ifdef CONFIG_DEBUG_PREEMPT
4720 : /*
4721 : * Underflow?
4722 : */
4723 : if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4724 : return;
4725 : #endif
4726 : __preempt_count_add(val);
4727 : #ifdef CONFIG_DEBUG_PREEMPT
4728 : /*
4729 : * Spinlock count overflowing soon?
4730 : */
4731 : DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4732 : PREEMPT_MASK - 10);
4733 : #endif
4734 : preempt_latency_start(val);
4735 : }
4736 : EXPORT_SYMBOL(preempt_count_add);
4737 : NOKPROBE_SYMBOL(preempt_count_add);
4738 :
4739 : /*
4740 : * If the value passed in equals to the current preempt count
4741 : * then we just enabled preemption. Stop timing the latency.
4742 : */
4743 : static inline void preempt_latency_stop(int val)
4744 : {
4745 : if (preempt_count() == val)
4746 : trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
4747 : }
4748 :
4749 : void preempt_count_sub(int val)
4750 : {
4751 : #ifdef CONFIG_DEBUG_PREEMPT
4752 : /*
4753 : * Underflow?
4754 : */
4755 : if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4756 : return;
4757 : /*
4758 : * Is the spinlock portion underflowing?
4759 : */
4760 : if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4761 : !(preempt_count() & PREEMPT_MASK)))
4762 : return;
4763 : #endif
4764 :
4765 : preempt_latency_stop(val);
4766 : __preempt_count_sub(val);
4767 : }
4768 : EXPORT_SYMBOL(preempt_count_sub);
4769 : NOKPROBE_SYMBOL(preempt_count_sub);
4770 :
4771 : #else
4772 2543 : static inline void preempt_latency_start(int val) { }
4773 2543 : static inline void preempt_latency_stop(int val) { }
4774 : #endif
4775 :
4776 0 : static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
4777 : {
4778 : #ifdef CONFIG_DEBUG_PREEMPT
4779 : return p->preempt_disable_ip;
4780 : #else
4781 0 : return 0;
4782 : #endif
4783 : }
4784 :
4785 : /*
4786 : * Print scheduling while atomic bug:
4787 : */
4788 0 : static noinline void __schedule_bug(struct task_struct *prev)
4789 : {
4790 : /* Save this before calling printk(), since that will clobber it */
4791 0 : unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
4792 :
4793 0 : if (oops_in_progress)
4794 : return;
4795 :
4796 0 : printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4797 0 : prev->comm, prev->pid, preempt_count());
4798 :
4799 0 : debug_show_held_locks(prev);
4800 0 : print_modules();
4801 0 : if (irqs_disabled())
4802 0 : print_irqtrace_events(prev);
4803 0 : if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
4804 : && in_atomic_preempt_off()) {
4805 : pr_err("Preemption disabled at:");
4806 : print_ip_sym(KERN_ERR, preempt_disable_ip);
4807 : }
4808 0 : if (panic_on_warn)
4809 0 : panic("scheduling while atomic\n");
4810 :
4811 0 : dump_stack();
4812 0 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4813 : }
4814 :
4815 : /*
4816 : * Various schedule()-time debugging checks and statistics:
4817 : */
4818 27963 : static inline void schedule_debug(struct task_struct *prev, bool preempt)
4819 : {
4820 : #ifdef CONFIG_SCHED_STACK_END_CHECK
4821 : if (task_stack_end_corrupted(prev))
4822 : panic("corrupted stack end detected inside scheduler\n");
4823 :
4824 : if (task_scs_end_corrupted(prev))
4825 : panic("corrupted shadow stack detected inside scheduler\n");
4826 : #endif
4827 :
4828 : #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
4829 27963 : if (!preempt && prev->state && prev->non_block_count) {
4830 0 : printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
4831 0 : prev->comm, prev->pid, prev->non_block_count);
4832 0 : dump_stack();
4833 0 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4834 : }
4835 : #endif
4836 :
4837 27963 : if (unlikely(in_atomic_preempt_off())) {
4838 0 : __schedule_bug(prev);
4839 0 : preempt_count_set(PREEMPT_DISABLED);
4840 : }
4841 83897 : rcu_sleep_check();
4842 27966 : SCHED_WARN_ON(ct_state() == CONTEXT_USER);
4843 :
4844 27966 : profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4845 :
4846 27967 : schedstat_inc(this_rq()->sched_count);
4847 27967 : }
4848 :
4849 79 : static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
4850 : struct rq_flags *rf)
4851 : {
4852 : #ifdef CONFIG_SMP
4853 79 : const struct sched_class *class;
4854 : /*
4855 : * We must do the balancing pass before put_prev_task(), such
4856 : * that when we release the rq->lock the task is in the same
4857 : * state as before we took rq->lock.
4858 : *
4859 : * We can terminate the balance pass as soon as we know there is
4860 : * a runnable task of @class priority or higher.
4861 : */
4862 204 : for_class_range(class, prev->sched_class, &idle_sched_class) {
4863 193 : if (class->balance(rq, prev, rf))
4864 : break;
4865 : }
4866 : #endif
4867 :
4868 78 : put_prev_task(rq, prev);
4869 78 : }
4870 :
4871 : /*
4872 : * Pick up the highest-prio task:
4873 : */
4874 : static inline struct task_struct *
4875 27962 : pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
4876 : {
4877 27962 : const struct sched_class *class;
4878 27962 : struct task_struct *p;
4879 :
4880 : /*
4881 : * Optimization: we know that if all tasks are in the fair class we can
4882 : * call that function directly, but only if the @prev task wasn't of a
4883 : * higher scheduling class, because otherwise those lose the
4884 : * opportunity to pull in more work from other CPUs.
4885 : */
4886 27962 : if (likely(prev->sched_class <= &fair_sched_class &&
4887 : rq->nr_running == rq->cfs.h_nr_running)) {
4888 :
4889 27883 : p = pick_next_task_fair(rq, prev, rf);
4890 27884 : if (unlikely(p == RETRY_TASK))
4891 0 : goto restart;
4892 :
4893 : /* Assumes fair_sched_class->next == idle_sched_class */
4894 27884 : if (!p) {
4895 6926 : put_prev_task(rq, prev);
4896 6927 : p = pick_next_task_idle(rq);
4897 : }
4898 :
4899 27884 : return p;
4900 : }
4901 :
4902 79 : restart:
4903 79 : put_prev_task_balance(rq, prev, rf);
4904 :
4905 281 : for_each_class(class) {
4906 202 : p = class->pick_next_task(rq);
4907 203 : if (p)
4908 79 : return p;
4909 : }
4910 :
4911 : /* The idle class should always have a runnable task: */
4912 0 : BUG();
4913 : }
4914 :
4915 : /*
4916 : * __schedule() is the main scheduler function.
4917 : *
4918 : * The main means of driving the scheduler and thus entering this function are:
4919 : *
4920 : * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4921 : *
4922 : * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4923 : * paths. For example, see arch/x86/entry_64.S.
4924 : *
4925 : * To drive preemption between tasks, the scheduler sets the flag in timer
4926 : * interrupt handler scheduler_tick().
4927 : *
4928 : * 3. Wakeups don't really cause entry into schedule(). They add a
4929 : * task to the run-queue and that's it.
4930 : *
4931 : * Now, if the new task added to the run-queue preempts the current
4932 : * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4933 : * called on the nearest possible occasion:
4934 : *
4935 : * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4936 : *
4937 : * - in syscall or exception context, at the next outmost
4938 : * preempt_enable(). (this might be as soon as the wake_up()'s
4939 : * spin_unlock()!)
4940 : *
4941 : * - in IRQ context, return from interrupt-handler to
4942 : * preemptible context
4943 : *
4944 : * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4945 : * then at the next:
4946 : *
4947 : * - cond_resched() call
4948 : * - explicit schedule() call
4949 : * - return from syscall or exception to user-space
4950 : * - return from interrupt-handler to user-space
4951 : *
4952 : * WARNING: must be called with preemption disabled!
4953 : */
4954 27961 : static void __sched notrace __schedule(bool preempt)
4955 : {
4956 27961 : struct task_struct *prev, *next;
4957 27961 : unsigned long *switch_count;
4958 27961 : unsigned long prev_state;
4959 27961 : struct rq_flags rf;
4960 27961 : struct rq *rq;
4961 27961 : int cpu;
4962 :
4963 27961 : cpu = smp_processor_id();
4964 27961 : rq = cpu_rq(cpu);
4965 27961 : prev = rq->curr;
4966 :
4967 27961 : schedule_debug(prev, preempt);
4968 :
4969 27967 : if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
4970 : hrtick_clear(rq);
4971 :
4972 27967 : local_irq_disable();
4973 27967 : rcu_note_context_switch(preempt);
4974 :
4975 : /*
4976 : * Make sure that signal_pending_state()->signal_pending() below
4977 : * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4978 : * done by the caller to avoid the race with signal_wake_up():
4979 : *
4980 : * __set_current_state(@state) signal_wake_up()
4981 : * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
4982 : * wake_up_state(p, state)
4983 : * LOCK rq->lock LOCK p->pi_state
4984 : * smp_mb__after_spinlock() smp_mb__after_spinlock()
4985 : * if (signal_pending_state()) if (p->state & @state)
4986 : *
4987 : * Also, the membarrier system call requires a full memory barrier
4988 : * after coming from user-space, before storing to rq->curr.
4989 : */
4990 27964 : rq_lock(rq, &rf);
4991 27965 : smp_mb__after_spinlock();
4992 :
4993 : /* Promote REQ to ACT */
4994 27965 : rq->clock_update_flags <<= 1;
4995 27965 : update_rq_clock(rq);
4996 :
4997 27963 : switch_count = &prev->nivcsw;
4998 :
4999 : /*
5000 : * We must load prev->state once (task_struct::state is volatile), such
5001 : * that:
5002 : *
5003 : * - we form a control dependency vs deactivate_task() below.
5004 : * - ptrace_{,un}freeze_traced() can change ->state underneath us.
5005 : */
5006 27963 : prev_state = prev->state;
5007 27963 : if (!preempt && prev_state) {
5008 14838 : if (signal_pending_state(prev_state, prev)) {
5009 0 : prev->state = TASK_RUNNING;
5010 : } else {
5011 29676 : prev->sched_contributes_to_load =
5012 14838 : (prev_state & TASK_UNINTERRUPTIBLE) &&
5013 14838 : !(prev_state & TASK_NOLOAD) &&
5014 2713 : !(prev->flags & PF_FROZEN);
5015 :
5016 14838 : if (prev->sched_contributes_to_load)
5017 2713 : rq->nr_uninterruptible++;
5018 :
5019 : /*
5020 : * __schedule() ttwu()
5021 : * prev_state = prev->state; if (p->on_rq && ...)
5022 : * if (prev_state) goto out;
5023 : * p->on_rq = 0; smp_acquire__after_ctrl_dep();
5024 : * p->state = TASK_WAKING
5025 : *
5026 : * Where __schedule() and ttwu() have matching control dependencies.
5027 : *
5028 : * After this, schedule() must not care about p->state any more.
5029 : */
5030 14838 : deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
5031 :
5032 14837 : if (prev->in_iowait) {
5033 1873 : atomic_inc(&rq->nr_iowait);
5034 1873 : delayacct_blkio_start();
5035 : }
5036 : }
5037 14837 : switch_count = &prev->nvcsw;
5038 : }
5039 :
5040 27962 : next = pick_next_task(rq, prev, &rf);
5041 27960 : clear_tsk_need_resched(prev);
5042 27967 : clear_preempt_need_resched();
5043 :
5044 27967 : if (likely(prev != next)) {
5045 26515 : rq->nr_switches++;
5046 : /*
5047 : * RCU users of rcu_dereference(rq->curr) may not see
5048 : * changes to task_struct made by pick_next_task().
5049 : */
5050 26515 : RCU_INIT_POINTER(rq->curr, next);
5051 : /*
5052 : * The membarrier system call requires each architecture
5053 : * to have a full memory barrier after updating
5054 : * rq->curr, before returning to user-space.
5055 : *
5056 : * Here are the schemes providing that barrier on the
5057 : * various architectures:
5058 : * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
5059 : * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
5060 : * - finish_lock_switch() for weakly-ordered
5061 : * architectures where spin_unlock is a full barrier,
5062 : * - switch_to() for arm64 (weakly-ordered, spin_unlock
5063 : * is a RELEASE barrier),
5064 : */
5065 26515 : ++*switch_count;
5066 :
5067 26515 : migrate_disable_switch(rq, prev);
5068 26512 : psi_sched_switch(prev, next, !task_on_rq_queued(prev));
5069 :
5070 26512 : trace_sched_switch(preempt, prev, next);
5071 :
5072 : /* Also unlocks the rq: */
5073 52049 : rq = context_switch(rq, prev, next, &rf);
5074 : } else {
5075 1452 : rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5076 :
5077 1452 : rq_unpin_lock(rq, &rf);
5078 1452 : __balance_callbacks(rq);
5079 1452 : raw_spin_unlock_irq(&rq->lock);
5080 : }
5081 26990 : }
5082 :
5083 900 : void __noreturn do_task_dead(void)
5084 : {
5085 : /* Causes final put_task_struct in finish_task_switch(): */
5086 900 : set_special_state(TASK_DEAD);
5087 :
5088 : /* Tell freezer to ignore us: */
5089 900 : current->flags |= PF_NOFREEZE;
5090 :
5091 900 : __schedule(false);
5092 0 : BUG();
5093 :
5094 : /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
5095 : for (;;)
5096 : cpu_relax();
5097 : }
5098 :
5099 17578 : static inline void sched_submit_work(struct task_struct *tsk)
5100 : {
5101 17578 : unsigned int task_flags;
5102 :
5103 17578 : if (!tsk->state)
5104 : return;
5105 :
5106 13944 : task_flags = tsk->flags;
5107 : /*
5108 : * If a worker went to sleep, notify and ask workqueue whether
5109 : * it wants to wake up a task to maintain concurrency.
5110 : * As this function is called inside the schedule() context,
5111 : * we disable preemption to avoid it calling schedule() again
5112 : * in the possible wakeup of a kworker and because wq_worker_sleeping()
5113 : * requires it.
5114 : */
5115 13944 : if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
5116 1908 : preempt_disable();
5117 1908 : if (task_flags & PF_WQ_WORKER)
5118 1908 : wq_worker_sleeping(tsk);
5119 : else
5120 1908 : io_wq_worker_sleeping(tsk);
5121 1908 : preempt_enable_no_resched();
5122 : }
5123 :
5124 13944 : if (tsk_is_pi_blocked(tsk))
5125 : return;
5126 :
5127 : /*
5128 : * If we are going to sleep and we have plugged IO queued,
5129 : * make sure to submit it to avoid deadlocks.
5130 : */
5131 13958 : if (blk_needs_flush_plug(tsk))
5132 0 : blk_schedule_flush_plug(tsk);
5133 : }
5134 :
5135 17498 : static void sched_update_worker(struct task_struct *tsk)
5136 : {
5137 17498 : if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
5138 1879 : if (tsk->flags & PF_WQ_WORKER)
5139 1879 : wq_worker_running(tsk);
5140 : else
5141 17499 : io_wq_worker_running(tsk);
5142 : }
5143 17499 : }
5144 :
5145 17578 : asmlinkage __visible void __sched schedule(void)
5146 : {
5147 17578 : struct task_struct *tsk = current;
5148 :
5149 17578 : sched_submit_work(tsk);
5150 17590 : do {
5151 17590 : preempt_disable();
5152 17590 : __schedule(false);
5153 17510 : sched_preempt_enable_no_resched();
5154 17510 : } while (need_resched());
5155 17499 : sched_update_worker(tsk);
5156 17498 : }
5157 : EXPORT_SYMBOL(schedule);
5158 :
5159 : /*
5160 : * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
5161 : * state (have scheduled out non-voluntarily) by making sure that all
5162 : * tasks have either left the run queue or have gone into user space.
5163 : * As idle tasks do not do either, they must not ever be preempted
5164 : * (schedule out non-voluntarily).
5165 : *
5166 : * schedule_idle() is similar to schedule_preempt_disable() except that it
5167 : * never enables preemption because it does not call sched_submit_work().
5168 : */
5169 6839 : void __sched schedule_idle(void)
5170 : {
5171 : /*
5172 : * As this skips calling sched_submit_work(), which the idle task does
5173 : * regardless because that function is a nop when the task is in a
5174 : * TASK_RUNNING state, make sure this isn't used someplace that the
5175 : * current task can be in any other state. Note, idle is always in the
5176 : * TASK_RUNNING state.
5177 : */
5178 6839 : WARN_ON_ONCE(current->state);
5179 6932 : do {
5180 6932 : __schedule(false);
5181 6933 : } while (need_resched());
5182 6840 : }
5183 :
5184 : #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
5185 : asmlinkage __visible void __sched schedule_user(void)
5186 : {
5187 : /*
5188 : * If we come here after a random call to set_need_resched(),
5189 : * or we have been woken up remotely but the IPI has not yet arrived,
5190 : * we haven't yet exited the RCU idle mode. Do it here manually until
5191 : * we find a better solution.
5192 : *
5193 : * NB: There are buggy callers of this function. Ideally we
5194 : * should warn if prev_state != CONTEXT_USER, but that will trigger
5195 : * too frequently to make sense yet.
5196 : */
5197 : enum ctx_state prev_state = exception_enter();
5198 : schedule();
5199 : exception_exit(prev_state);
5200 : }
5201 : #endif
5202 :
5203 : /**
5204 : * schedule_preempt_disabled - called with preemption disabled
5205 : *
5206 : * Returns with preemption disabled. Note: preempt_count must be 1
5207 : */
5208 94 : void __sched schedule_preempt_disabled(void)
5209 : {
5210 94 : sched_preempt_enable_no_resched();
5211 94 : schedule();
5212 94 : preempt_disable();
5213 94 : }
5214 :
5215 2537 : static void __sched notrace preempt_schedule_common(void)
5216 : {
5217 2543 : do {
5218 : /*
5219 : * Because the function tracer can trace preempt_count_sub()
5220 : * and it also uses preempt_enable/disable_notrace(), if
5221 : * NEED_RESCHED is set, the preempt_enable_notrace() called
5222 : * by the function tracer will call this function again and
5223 : * cause infinite recursion.
5224 : *
5225 : * Preemption must be disabled here before the function
5226 : * tracer can trace. Break up preempt_disable() into two
5227 : * calls. One to disable preemption without fear of being
5228 : * traced. The other to still record the preemption latency,
5229 : * which can also be traced by the function tracer.
5230 : */
5231 2543 : preempt_disable_notrace();
5232 2543 : preempt_latency_start(1);
5233 2543 : __schedule(true);
5234 2543 : preempt_latency_stop(1);
5235 2543 : preempt_enable_no_resched_notrace();
5236 :
5237 : /*
5238 : * Check again in case we missed a preemption opportunity
5239 : * between schedule and now.
5240 : */
5241 2543 : } while (need_resched());
5242 2537 : }
5243 :
5244 : #ifdef CONFIG_PREEMPTION
5245 : /*
5246 : * This is the entry point to schedule() from in-kernel preemption
5247 : * off of preempt_enable.
5248 : */
5249 : asmlinkage __visible void __sched notrace preempt_schedule(void)
5250 : {
5251 : /*
5252 : * If there is a non-zero preempt_count or interrupts are disabled,
5253 : * we do not want to preempt the current task. Just return..
5254 : */
5255 : if (likely(!preemptible()))
5256 : return;
5257 :
5258 : preempt_schedule_common();
5259 : }
5260 : NOKPROBE_SYMBOL(preempt_schedule);
5261 : EXPORT_SYMBOL(preempt_schedule);
5262 :
5263 : #ifdef CONFIG_PREEMPT_DYNAMIC
5264 : DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
5265 : EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
5266 : #endif
5267 :
5268 :
5269 : /**
5270 : * preempt_schedule_notrace - preempt_schedule called by tracing
5271 : *
5272 : * The tracing infrastructure uses preempt_enable_notrace to prevent
5273 : * recursion and tracing preempt enabling caused by the tracing
5274 : * infrastructure itself. But as tracing can happen in areas coming
5275 : * from userspace or just about to enter userspace, a preempt enable
5276 : * can occur before user_exit() is called. This will cause the scheduler
5277 : * to be called when the system is still in usermode.
5278 : *
5279 : * To prevent this, the preempt_enable_notrace will use this function
5280 : * instead of preempt_schedule() to exit user context if needed before
5281 : * calling the scheduler.
5282 : */
5283 : asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
5284 : {
5285 : enum ctx_state prev_ctx;
5286 :
5287 : if (likely(!preemptible()))
5288 : return;
5289 :
5290 : do {
5291 : /*
5292 : * Because the function tracer can trace preempt_count_sub()
5293 : * and it also uses preempt_enable/disable_notrace(), if
5294 : * NEED_RESCHED is set, the preempt_enable_notrace() called
5295 : * by the function tracer will call this function again and
5296 : * cause infinite recursion.
5297 : *
5298 : * Preemption must be disabled here before the function
5299 : * tracer can trace. Break up preempt_disable() into two
5300 : * calls. One to disable preemption without fear of being
5301 : * traced. The other to still record the preemption latency,
5302 : * which can also be traced by the function tracer.
5303 : */
5304 : preempt_disable_notrace();
5305 : preempt_latency_start(1);
5306 : /*
5307 : * Needs preempt disabled in case user_exit() is traced
5308 : * and the tracer calls preempt_enable_notrace() causing
5309 : * an infinite recursion.
5310 : */
5311 : prev_ctx = exception_enter();
5312 : __schedule(true);
5313 : exception_exit(prev_ctx);
5314 :
5315 : preempt_latency_stop(1);
5316 : preempt_enable_no_resched_notrace();
5317 : } while (need_resched());
5318 : }
5319 : EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
5320 :
5321 : #ifdef CONFIG_PREEMPT_DYNAMIC
5322 : DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
5323 : EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
5324 : #endif
5325 :
5326 : #endif /* CONFIG_PREEMPTION */
5327 :
5328 : #ifdef CONFIG_PREEMPT_DYNAMIC
5329 :
5330 : #include <linux/entry-common.h>
5331 :
5332 : /*
5333 : * SC:cond_resched
5334 : * SC:might_resched
5335 : * SC:preempt_schedule
5336 : * SC:preempt_schedule_notrace
5337 : * SC:irqentry_exit_cond_resched
5338 : *
5339 : *
5340 : * NONE:
5341 : * cond_resched <- __cond_resched
5342 : * might_resched <- RET0
5343 : * preempt_schedule <- NOP
5344 : * preempt_schedule_notrace <- NOP
5345 : * irqentry_exit_cond_resched <- NOP
5346 : *
5347 : * VOLUNTARY:
5348 : * cond_resched <- __cond_resched
5349 : * might_resched <- __cond_resched
5350 : * preempt_schedule <- NOP
5351 : * preempt_schedule_notrace <- NOP
5352 : * irqentry_exit_cond_resched <- NOP
5353 : *
5354 : * FULL:
5355 : * cond_resched <- RET0
5356 : * might_resched <- RET0
5357 : * preempt_schedule <- preempt_schedule
5358 : * preempt_schedule_notrace <- preempt_schedule_notrace
5359 : * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
5360 : */
5361 :
5362 : enum {
5363 : preempt_dynamic_none = 0,
5364 : preempt_dynamic_voluntary,
5365 : preempt_dynamic_full,
5366 : };
5367 :
5368 : static int preempt_dynamic_mode = preempt_dynamic_full;
5369 :
5370 : static int sched_dynamic_mode(const char *str)
5371 : {
5372 : if (!strcmp(str, "none"))
5373 : return 0;
5374 :
5375 : if (!strcmp(str, "voluntary"))
5376 : return 1;
5377 :
5378 : if (!strcmp(str, "full"))
5379 : return 2;
5380 :
5381 : return -1;
5382 : }
5383 :
5384 : static void sched_dynamic_update(int mode)
5385 : {
5386 : /*
5387 : * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
5388 : * the ZERO state, which is invalid.
5389 : */
5390 : static_call_update(cond_resched, __cond_resched);
5391 : static_call_update(might_resched, __cond_resched);
5392 : static_call_update(preempt_schedule, __preempt_schedule_func);
5393 : static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
5394 : static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
5395 :
5396 : switch (mode) {
5397 : case preempt_dynamic_none:
5398 : static_call_update(cond_resched, __cond_resched);
5399 : static_call_update(might_resched, (typeof(&__cond_resched)) __static_call_return0);
5400 : static_call_update(preempt_schedule, (typeof(&preempt_schedule)) NULL);
5401 : static_call_update(preempt_schedule_notrace, (typeof(&preempt_schedule_notrace)) NULL);
5402 : static_call_update(irqentry_exit_cond_resched, (typeof(&irqentry_exit_cond_resched)) NULL);
5403 : pr_info("Dynamic Preempt: none\n");
5404 : break;
5405 :
5406 : case preempt_dynamic_voluntary:
5407 : static_call_update(cond_resched, __cond_resched);
5408 : static_call_update(might_resched, __cond_resched);
5409 : static_call_update(preempt_schedule, (typeof(&preempt_schedule)) NULL);
5410 : static_call_update(preempt_schedule_notrace, (typeof(&preempt_schedule_notrace)) NULL);
5411 : static_call_update(irqentry_exit_cond_resched, (typeof(&irqentry_exit_cond_resched)) NULL);
5412 : pr_info("Dynamic Preempt: voluntary\n");
5413 : break;
5414 :
5415 : case preempt_dynamic_full:
5416 : static_call_update(cond_resched, (typeof(&__cond_resched)) __static_call_return0);
5417 : static_call_update(might_resched, (typeof(&__cond_resched)) __static_call_return0);
5418 : static_call_update(preempt_schedule, __preempt_schedule_func);
5419 : static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
5420 : static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
5421 : pr_info("Dynamic Preempt: full\n");
5422 : break;
5423 : }
5424 :
5425 : preempt_dynamic_mode = mode;
5426 : }
5427 :
5428 : static int __init setup_preempt_mode(char *str)
5429 : {
5430 : int mode = sched_dynamic_mode(str);
5431 : if (mode < 0) {
5432 : pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
5433 : return 1;
5434 : }
5435 :
5436 : sched_dynamic_update(mode);
5437 : return 0;
5438 : }
5439 : __setup("preempt=", setup_preempt_mode);
5440 :
5441 : #ifdef CONFIG_SCHED_DEBUG
5442 :
5443 : static ssize_t sched_dynamic_write(struct file *filp, const char __user *ubuf,
5444 : size_t cnt, loff_t *ppos)
5445 : {
5446 : char buf[16];
5447 : int mode;
5448 :
5449 : if (cnt > 15)
5450 : cnt = 15;
5451 :
5452 : if (copy_from_user(&buf, ubuf, cnt))
5453 : return -EFAULT;
5454 :
5455 : buf[cnt] = 0;
5456 : mode = sched_dynamic_mode(strstrip(buf));
5457 : if (mode < 0)
5458 : return mode;
5459 :
5460 : sched_dynamic_update(mode);
5461 :
5462 : *ppos += cnt;
5463 :
5464 : return cnt;
5465 : }
5466 :
5467 : static int sched_dynamic_show(struct seq_file *m, void *v)
5468 : {
5469 : static const char * preempt_modes[] = {
5470 : "none", "voluntary", "full"
5471 : };
5472 : int i;
5473 :
5474 : for (i = 0; i < ARRAY_SIZE(preempt_modes); i++) {
5475 : if (preempt_dynamic_mode == i)
5476 : seq_puts(m, "(");
5477 : seq_puts(m, preempt_modes[i]);
5478 : if (preempt_dynamic_mode == i)
5479 : seq_puts(m, ")");
5480 :
5481 : seq_puts(m, " ");
5482 : }
5483 :
5484 : seq_puts(m, "\n");
5485 : return 0;
5486 : }
5487 :
5488 : static int sched_dynamic_open(struct inode *inode, struct file *filp)
5489 : {
5490 : return single_open(filp, sched_dynamic_show, NULL);
5491 : }
5492 :
5493 : static const struct file_operations sched_dynamic_fops = {
5494 : .open = sched_dynamic_open,
5495 : .write = sched_dynamic_write,
5496 : .read = seq_read,
5497 : .llseek = seq_lseek,
5498 : .release = single_release,
5499 : };
5500 :
5501 : static __init int sched_init_debug_dynamic(void)
5502 : {
5503 : debugfs_create_file("sched_preempt", 0644, NULL, NULL, &sched_dynamic_fops);
5504 : return 0;
5505 : }
5506 : late_initcall(sched_init_debug_dynamic);
5507 :
5508 : #endif /* CONFIG_SCHED_DEBUG */
5509 : #endif /* CONFIG_PREEMPT_DYNAMIC */
5510 :
5511 :
5512 : /*
5513 : * This is the entry point to schedule() from kernel preemption
5514 : * off of irq context.
5515 : * Note, that this is called and return with irqs disabled. This will
5516 : * protect us against recursive calling from irq.
5517 : */
5518 0 : asmlinkage __visible void __sched preempt_schedule_irq(void)
5519 : {
5520 0 : enum ctx_state prev_state;
5521 :
5522 : /* Catch callers which need to be fixed */
5523 0 : BUG_ON(preempt_count() || !irqs_disabled());
5524 :
5525 0 : prev_state = exception_enter();
5526 :
5527 0 : do {
5528 0 : preempt_disable();
5529 0 : local_irq_enable();
5530 0 : __schedule(true);
5531 0 : local_irq_disable();
5532 0 : sched_preempt_enable_no_resched();
5533 0 : } while (need_resched());
5534 :
5535 0 : exception_exit(prev_state);
5536 0 : }
5537 :
5538 4745 : int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
5539 : void *key)
5540 : {
5541 4745 : WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
5542 4745 : return try_to_wake_up(curr->private, mode, wake_flags);
5543 : }
5544 : EXPORT_SYMBOL(default_wake_function);
5545 :
5546 : #ifdef CONFIG_RT_MUTEXES
5547 :
5548 8 : static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
5549 : {
5550 8 : if (pi_task)
5551 0 : prio = min(prio, pi_task->prio);
5552 :
5553 8 : return prio;
5554 : }
5555 :
5556 8 : static inline int rt_effective_prio(struct task_struct *p, int prio)
5557 : {
5558 8 : struct task_struct *pi_task = rt_mutex_get_top_task(p);
5559 :
5560 8 : return __rt_effective_prio(pi_task, prio);
5561 : }
5562 :
5563 : /*
5564 : * rt_mutex_setprio - set the current priority of a task
5565 : * @p: task to boost
5566 : * @pi_task: donor task
5567 : *
5568 : * This function changes the 'effective' priority of a task. It does
5569 : * not touch ->normal_prio like __setscheduler().
5570 : *
5571 : * Used by the rt_mutex code to implement priority inheritance
5572 : * logic. Call site only calls if the priority of the task changed.
5573 : */
5574 0 : void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
5575 : {
5576 0 : int prio, oldprio, queued, running, queue_flag =
5577 : DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
5578 0 : const struct sched_class *prev_class;
5579 0 : struct rq_flags rf;
5580 0 : struct rq *rq;
5581 :
5582 : /* XXX used to be waiter->prio, not waiter->task->prio */
5583 0 : prio = __rt_effective_prio(pi_task, p->normal_prio);
5584 :
5585 : /*
5586 : * If nothing changed; bail early.
5587 : */
5588 0 : if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
5589 0 : return;
5590 :
5591 0 : rq = __task_rq_lock(p, &rf);
5592 0 : update_rq_clock(rq);
5593 : /*
5594 : * Set under pi_lock && rq->lock, such that the value can be used under
5595 : * either lock.
5596 : *
5597 : * Note that there is loads of tricky to make this pointer cache work
5598 : * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
5599 : * ensure a task is de-boosted (pi_task is set to NULL) before the
5600 : * task is allowed to run again (and can exit). This ensures the pointer
5601 : * points to a blocked task -- which guarantees the task is present.
5602 : */
5603 0 : p->pi_top_task = pi_task;
5604 :
5605 : /*
5606 : * For FIFO/RR we only need to set prio, if that matches we're done.
5607 : */
5608 0 : if (prio == p->prio && !dl_prio(prio))
5609 0 : goto out_unlock;
5610 :
5611 : /*
5612 : * Idle task boosting is a nono in general. There is one
5613 : * exception, when PREEMPT_RT and NOHZ is active:
5614 : *
5615 : * The idle task calls get_next_timer_interrupt() and holds
5616 : * the timer wheel base->lock on the CPU and another CPU wants
5617 : * to access the timer (probably to cancel it). We can safely
5618 : * ignore the boosting request, as the idle CPU runs this code
5619 : * with interrupts disabled and will complete the lock
5620 : * protected section without being interrupted. So there is no
5621 : * real need to boost.
5622 : */
5623 0 : if (unlikely(p == rq->idle)) {
5624 0 : WARN_ON(p != rq->curr);
5625 0 : WARN_ON(p->pi_blocked_on);
5626 0 : goto out_unlock;
5627 : }
5628 :
5629 0 : trace_sched_pi_setprio(p, pi_task);
5630 0 : oldprio = p->prio;
5631 :
5632 0 : if (oldprio == prio)
5633 0 : queue_flag &= ~DEQUEUE_MOVE;
5634 :
5635 0 : prev_class = p->sched_class;
5636 0 : queued = task_on_rq_queued(p);
5637 0 : running = task_current(rq, p);
5638 0 : if (queued)
5639 0 : dequeue_task(rq, p, queue_flag);
5640 0 : if (running)
5641 0 : put_prev_task(rq, p);
5642 :
5643 : /*
5644 : * Boosting condition are:
5645 : * 1. -rt task is running and holds mutex A
5646 : * --> -dl task blocks on mutex A
5647 : *
5648 : * 2. -dl task is running and holds mutex A
5649 : * --> -dl task blocks on mutex A and could preempt the
5650 : * running task
5651 : */
5652 0 : if (dl_prio(prio)) {
5653 0 : if (!dl_prio(p->normal_prio) ||
5654 0 : (pi_task && dl_prio(pi_task->prio) &&
5655 0 : dl_entity_preempt(&pi_task->dl, &p->dl))) {
5656 0 : p->dl.pi_se = pi_task->dl.pi_se;
5657 0 : queue_flag |= ENQUEUE_REPLENISH;
5658 : } else {
5659 0 : p->dl.pi_se = &p->dl;
5660 : }
5661 0 : p->sched_class = &dl_sched_class;
5662 0 : } else if (rt_prio(prio)) {
5663 0 : if (dl_prio(oldprio))
5664 0 : p->dl.pi_se = &p->dl;
5665 0 : if (oldprio < prio)
5666 0 : queue_flag |= ENQUEUE_HEAD;
5667 0 : p->sched_class = &rt_sched_class;
5668 : } else {
5669 0 : if (dl_prio(oldprio))
5670 0 : p->dl.pi_se = &p->dl;
5671 0 : if (rt_prio(oldprio))
5672 0 : p->rt.timeout = 0;
5673 0 : p->sched_class = &fair_sched_class;
5674 : }
5675 :
5676 0 : p->prio = prio;
5677 :
5678 0 : if (queued)
5679 0 : enqueue_task(rq, p, queue_flag);
5680 0 : if (running)
5681 0 : set_next_task(rq, p);
5682 :
5683 0 : check_class_changed(rq, p, prev_class, oldprio);
5684 0 : out_unlock:
5685 : /* Avoid rq from going away on us: */
5686 0 : preempt_disable();
5687 :
5688 0 : rq_unpin_lock(rq, &rf);
5689 0 : __balance_callbacks(rq);
5690 0 : raw_spin_unlock(&rq->lock);
5691 :
5692 0 : preempt_enable();
5693 : }
5694 : #else
5695 : static inline int rt_effective_prio(struct task_struct *p, int prio)
5696 : {
5697 : return prio;
5698 : }
5699 : #endif
5700 :
5701 39 : void set_user_nice(struct task_struct *p, long nice)
5702 : {
5703 39 : bool queued, running;
5704 39 : int old_prio;
5705 39 : struct rq_flags rf;
5706 39 : struct rq *rq;
5707 :
5708 39 : if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
5709 18 : return;
5710 : /*
5711 : * We have to be careful, if called from sys_setpriority(),
5712 : * the task might be in the middle of scheduling on another CPU.
5713 : */
5714 21 : rq = task_rq_lock(p, &rf);
5715 21 : update_rq_clock(rq);
5716 :
5717 : /*
5718 : * The RT priorities are set via sched_setscheduler(), but we still
5719 : * allow the 'normal' nice value to be set - but as expected
5720 : * it won't have any effect on scheduling until the task is
5721 : * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
5722 : */
5723 21 : if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
5724 0 : p->static_prio = NICE_TO_PRIO(nice);
5725 0 : goto out_unlock;
5726 : }
5727 21 : queued = task_on_rq_queued(p);
5728 21 : running = task_current(rq, p);
5729 21 : if (queued)
5730 13 : dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
5731 21 : if (running)
5732 13 : put_prev_task(rq, p);
5733 :
5734 21 : p->static_prio = NICE_TO_PRIO(nice);
5735 21 : set_load_weight(p, true);
5736 21 : old_prio = p->prio;
5737 21 : p->prio = effective_prio(p);
5738 :
5739 21 : if (queued)
5740 13 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5741 21 : if (running)
5742 13 : set_next_task(rq, p);
5743 :
5744 : /*
5745 : * If the task increased its priority or is running and
5746 : * lowered its priority, then reschedule its CPU:
5747 : */
5748 21 : p->sched_class->prio_changed(rq, p, old_prio);
5749 :
5750 21 : out_unlock:
5751 21 : task_rq_unlock(rq, p, &rf);
5752 : }
5753 : EXPORT_SYMBOL(set_user_nice);
5754 :
5755 : /*
5756 : * can_nice - check if a task can reduce its nice value
5757 : * @p: task
5758 : * @nice: nice value
5759 : */
5760 0 : int can_nice(const struct task_struct *p, const int nice)
5761 : {
5762 : /* Convert nice value [19,-20] to rlimit style value [1,40]: */
5763 0 : int nice_rlim = nice_to_rlimit(nice);
5764 :
5765 0 : return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5766 0 : capable(CAP_SYS_NICE));
5767 : }
5768 :
5769 : #ifdef __ARCH_WANT_SYS_NICE
5770 :
5771 : /*
5772 : * sys_nice - change the priority of the current process.
5773 : * @increment: priority increment
5774 : *
5775 : * sys_setpriority is a more generic, but much slower function that
5776 : * does similar things.
5777 : */
5778 0 : SYSCALL_DEFINE1(nice, int, increment)
5779 : {
5780 0 : long nice, retval;
5781 :
5782 : /*
5783 : * Setpriority might change our priority at the same moment.
5784 : * We don't have to worry. Conceptually one call occurs first
5785 : * and we have a single winner.
5786 : */
5787 0 : increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
5788 0 : nice = task_nice(current) + increment;
5789 :
5790 0 : nice = clamp_val(nice, MIN_NICE, MAX_NICE);
5791 0 : if (increment < 0 && !can_nice(current, nice))
5792 : return -EPERM;
5793 :
5794 0 : retval = security_task_setnice(current, nice);
5795 0 : if (retval)
5796 : return retval;
5797 :
5798 0 : set_user_nice(current, nice);
5799 0 : return 0;
5800 : }
5801 :
5802 : #endif
5803 :
5804 : /**
5805 : * task_prio - return the priority value of a given task.
5806 : * @p: the task in question.
5807 : *
5808 : * Return: The priority value as seen by users in /proc.
5809 : *
5810 : * sched policy return value kernel prio user prio/nice
5811 : *
5812 : * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
5813 : * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
5814 : * deadline -101 -1 0
5815 : */
5816 139 : int task_prio(const struct task_struct *p)
5817 : {
5818 139 : return p->prio - MAX_RT_PRIO;
5819 : }
5820 :
5821 : /**
5822 : * idle_cpu - is a given CPU idle currently?
5823 : * @cpu: the processor in question.
5824 : *
5825 : * Return: 1 if the CPU is currently idle. 0 otherwise.
5826 : */
5827 102866 : int idle_cpu(int cpu)
5828 : {
5829 102866 : struct rq *rq = cpu_rq(cpu);
5830 :
5831 102866 : if (rq->curr != rq->idle)
5832 : return 0;
5833 :
5834 54767 : if (rq->nr_running)
5835 : return 0;
5836 :
5837 : #ifdef CONFIG_SMP
5838 46497 : if (rq->ttwu_pending)
5839 209 : return 0;
5840 : #endif
5841 :
5842 : return 1;
5843 : }
5844 :
5845 : /**
5846 : * available_idle_cpu - is a given CPU idle for enqueuing work.
5847 : * @cpu: the CPU in question.
5848 : *
5849 : * Return: 1 if the CPU is currently idle. 0 otherwise.
5850 : */
5851 10602 : int available_idle_cpu(int cpu)
5852 : {
5853 10602 : if (!idle_cpu(cpu))
5854 : return 0;
5855 :
5856 5387 : if (vcpu_is_preempted(cpu))
5857 2031 : return 0;
5858 :
5859 : return 1;
5860 : }
5861 :
5862 : /**
5863 : * idle_task - return the idle task for a given CPU.
5864 : * @cpu: the processor in question.
5865 : *
5866 : * Return: The idle task for the CPU @cpu.
5867 : */
5868 0 : struct task_struct *idle_task(int cpu)
5869 : {
5870 0 : return cpu_rq(cpu)->idle;
5871 : }
5872 :
5873 : #ifdef CONFIG_SMP
5874 : /*
5875 : * This function computes an effective utilization for the given CPU, to be
5876 : * used for frequency selection given the linear relation: f = u * f_max.
5877 : *
5878 : * The scheduler tracks the following metrics:
5879 : *
5880 : * cpu_util_{cfs,rt,dl,irq}()
5881 : * cpu_bw_dl()
5882 : *
5883 : * Where the cfs,rt and dl util numbers are tracked with the same metric and
5884 : * synchronized windows and are thus directly comparable.
5885 : *
5886 : * The cfs,rt,dl utilization are the running times measured with rq->clock_task
5887 : * which excludes things like IRQ and steal-time. These latter are then accrued
5888 : * in the irq utilization.
5889 : *
5890 : * The DL bandwidth number otoh is not a measured metric but a value computed
5891 : * based on the task model parameters and gives the minimal utilization
5892 : * required to meet deadlines.
5893 : */
5894 0 : unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
5895 : unsigned long max, enum cpu_util_type type,
5896 : struct task_struct *p)
5897 : {
5898 0 : unsigned long dl_util, util, irq;
5899 0 : struct rq *rq = cpu_rq(cpu);
5900 :
5901 0 : if (!uclamp_is_used() &&
5902 0 : type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
5903 : return max;
5904 : }
5905 :
5906 : /*
5907 : * Early check to see if IRQ/steal time saturates the CPU, can be
5908 : * because of inaccuracies in how we track these -- see
5909 : * update_irq_load_avg().
5910 : */
5911 0 : irq = cpu_util_irq(rq);
5912 0 : if (unlikely(irq >= max))
5913 : return max;
5914 :
5915 : /*
5916 : * Because the time spend on RT/DL tasks is visible as 'lost' time to
5917 : * CFS tasks and we use the same metric to track the effective
5918 : * utilization (PELT windows are synchronized) we can directly add them
5919 : * to obtain the CPU's actual utilization.
5920 : *
5921 : * CFS and RT utilization can be boosted or capped, depending on
5922 : * utilization clamp constraints requested by currently RUNNABLE
5923 : * tasks.
5924 : * When there are no CFS RUNNABLE tasks, clamps are released and
5925 : * frequency will be gracefully reduced with the utilization decay.
5926 : */
5927 0 : util = util_cfs + cpu_util_rt(rq);
5928 0 : if (type == FREQUENCY_UTIL)
5929 0 : util = uclamp_rq_util_with(rq, util, p);
5930 :
5931 0 : dl_util = cpu_util_dl(rq);
5932 :
5933 : /*
5934 : * For frequency selection we do not make cpu_util_dl() a permanent part
5935 : * of this sum because we want to use cpu_bw_dl() later on, but we need
5936 : * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
5937 : * that we select f_max when there is no idle time.
5938 : *
5939 : * NOTE: numerical errors or stop class might cause us to not quite hit
5940 : * saturation when we should -- something for later.
5941 : */
5942 0 : if (util + dl_util >= max)
5943 : return max;
5944 :
5945 : /*
5946 : * OTOH, for energy computation we need the estimated running time, so
5947 : * include util_dl and ignore dl_bw.
5948 : */
5949 0 : if (type == ENERGY_UTIL)
5950 0 : util += dl_util;
5951 :
5952 : /*
5953 : * There is still idle time; further improve the number by using the
5954 : * irq metric. Because IRQ/steal time is hidden from the task clock we
5955 : * need to scale the task numbers:
5956 : *
5957 : * max - irq
5958 : * U' = irq + --------- * U
5959 : * max
5960 : */
5961 0 : util = scale_irq_capacity(util, irq, max);
5962 0 : util += irq;
5963 :
5964 : /*
5965 : * Bandwidth required by DEADLINE must always be granted while, for
5966 : * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
5967 : * to gracefully reduce the frequency when no tasks show up for longer
5968 : * periods of time.
5969 : *
5970 : * Ideally we would like to set bw_dl as min/guaranteed freq and util +
5971 : * bw_dl as requested freq. However, cpufreq is not yet ready for such
5972 : * an interface. So, we only do the latter for now.
5973 : */
5974 0 : if (type == FREQUENCY_UTIL)
5975 0 : util += cpu_bw_dl(rq);
5976 :
5977 0 : return min(max, util);
5978 : }
5979 :
5980 0 : unsigned long sched_cpu_util(int cpu, unsigned long max)
5981 : {
5982 0 : return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
5983 : ENERGY_UTIL, NULL);
5984 : }
5985 : #endif /* CONFIG_SMP */
5986 :
5987 : /**
5988 : * find_process_by_pid - find a process with a matching PID value.
5989 : * @pid: the pid in question.
5990 : *
5991 : * The task of @pid, if found. %NULL otherwise.
5992 : */
5993 8 : static struct task_struct *find_process_by_pid(pid_t pid)
5994 : {
5995 8 : return pid ? find_task_by_vpid(pid) : current;
5996 : }
5997 :
5998 : /*
5999 : * sched_setparam() passes in -1 for its policy, to let the functions
6000 : * it calls know not to change it.
6001 : */
6002 : #define SETPARAM_POLICY -1
6003 :
6004 4 : static void __setscheduler_params(struct task_struct *p,
6005 : const struct sched_attr *attr)
6006 : {
6007 4 : int policy = attr->sched_policy;
6008 :
6009 4 : if (policy == SETPARAM_POLICY)
6010 0 : policy = p->policy;
6011 :
6012 4 : p->policy = policy;
6013 :
6014 4 : if (dl_policy(policy))
6015 0 : __setparam_dl(p, attr);
6016 4 : else if (fair_policy(policy))
6017 0 : p->static_prio = NICE_TO_PRIO(attr->sched_nice);
6018 :
6019 : /*
6020 : * __sched_setscheduler() ensures attr->sched_priority == 0 when
6021 : * !rt_policy. Always setting this ensures that things like
6022 : * getparam()/getattr() don't report silly values for !rt tasks.
6023 : */
6024 4 : p->rt_priority = attr->sched_priority;
6025 4 : p->normal_prio = normal_prio(p);
6026 4 : set_load_weight(p, true);
6027 4 : }
6028 :
6029 : /* Actually do priority change: must hold pi & rq lock. */
6030 4 : static void __setscheduler(struct rq *rq, struct task_struct *p,
6031 : const struct sched_attr *attr, bool keep_boost)
6032 : {
6033 : /*
6034 : * If params can't change scheduling class changes aren't allowed
6035 : * either.
6036 : */
6037 4 : if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
6038 : return;
6039 :
6040 4 : __setscheduler_params(p, attr);
6041 :
6042 : /*
6043 : * Keep a potential priority boosting if called from
6044 : * sched_setscheduler().
6045 : */
6046 4 : p->prio = normal_prio(p);
6047 4 : if (keep_boost)
6048 4 : p->prio = rt_effective_prio(p, p->prio);
6049 :
6050 4 : if (dl_prio(p->prio))
6051 0 : p->sched_class = &dl_sched_class;
6052 4 : else if (rt_prio(p->prio))
6053 4 : p->sched_class = &rt_sched_class;
6054 : else
6055 0 : p->sched_class = &fair_sched_class;
6056 : }
6057 :
6058 : /*
6059 : * Check the target process has a UID that matches the current process's:
6060 : */
6061 0 : static bool check_same_owner(struct task_struct *p)
6062 : {
6063 0 : const struct cred *cred = current_cred(), *pcred;
6064 0 : bool match;
6065 :
6066 0 : rcu_read_lock();
6067 0 : pcred = __task_cred(p);
6068 0 : match = (uid_eq(cred->euid, pcred->euid) ||
6069 0 : uid_eq(cred->euid, pcred->uid));
6070 0 : rcu_read_unlock();
6071 0 : return match;
6072 : }
6073 :
6074 55 : static int __sched_setscheduler(struct task_struct *p,
6075 : const struct sched_attr *attr,
6076 : bool user, bool pi)
6077 : {
6078 55 : int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
6079 55 : MAX_RT_PRIO - 1 - attr->sched_priority;
6080 55 : int retval, oldprio, oldpolicy = -1, queued, running;
6081 55 : int new_effective_prio, policy = attr->sched_policy;
6082 55 : const struct sched_class *prev_class;
6083 55 : struct callback_head *head;
6084 55 : struct rq_flags rf;
6085 55 : int reset_on_fork;
6086 55 : int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6087 55 : struct rq *rq;
6088 :
6089 : /* The pi code expects interrupts enabled */
6090 55 : BUG_ON(pi && in_interrupt());
6091 55 : recheck:
6092 : /* Double check policy once rq lock held: */
6093 55 : if (policy < 0) {
6094 0 : reset_on_fork = p->sched_reset_on_fork;
6095 0 : policy = oldpolicy = p->policy;
6096 : } else {
6097 55 : reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
6098 :
6099 55 : if (!valid_policy(policy))
6100 : return -EINVAL;
6101 : }
6102 :
6103 55 : if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
6104 : return -EINVAL;
6105 :
6106 : /*
6107 : * Valid priorities for SCHED_FIFO and SCHED_RR are
6108 : * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
6109 : * SCHED_BATCH and SCHED_IDLE is 0.
6110 : */
6111 55 : if (attr->sched_priority > MAX_RT_PRIO-1)
6112 : return -EINVAL;
6113 55 : if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
6114 55 : (rt_policy(policy) != (attr->sched_priority != 0)))
6115 0 : return -EINVAL;
6116 :
6117 : /*
6118 : * Allow unprivileged RT tasks to decrease priority:
6119 : */
6120 55 : if (user && !capable(CAP_SYS_NICE)) {
6121 0 : if (fair_policy(policy)) {
6122 0 : if (attr->sched_nice < task_nice(p) &&
6123 0 : !can_nice(p, attr->sched_nice))
6124 : return -EPERM;
6125 : }
6126 :
6127 0 : if (rt_policy(policy)) {
6128 0 : unsigned long rlim_rtprio =
6129 0 : task_rlimit(p, RLIMIT_RTPRIO);
6130 :
6131 : /* Can't set/change the rt policy: */
6132 0 : if (policy != p->policy && !rlim_rtprio)
6133 : return -EPERM;
6134 :
6135 : /* Can't increase priority: */
6136 0 : if (attr->sched_priority > p->rt_priority &&
6137 0 : attr->sched_priority > rlim_rtprio)
6138 : return -EPERM;
6139 : }
6140 :
6141 : /*
6142 : * Can't set/change SCHED_DEADLINE policy at all for now
6143 : * (safest behavior); in the future we would like to allow
6144 : * unprivileged DL tasks to increase their relative deadline
6145 : * or reduce their runtime (both ways reducing utilization)
6146 : */
6147 0 : if (dl_policy(policy))
6148 : return -EPERM;
6149 :
6150 : /*
6151 : * Treat SCHED_IDLE as nice 20. Only allow a switch to
6152 : * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
6153 : */
6154 0 : if (task_has_idle_policy(p) && !idle_policy(policy)) {
6155 0 : if (!can_nice(p, task_nice(p)))
6156 : return -EPERM;
6157 : }
6158 :
6159 : /* Can't change other user's priorities: */
6160 0 : if (!check_same_owner(p))
6161 : return -EPERM;
6162 :
6163 : /* Normal users shall not reset the sched_reset_on_fork flag: */
6164 0 : if (p->sched_reset_on_fork && !reset_on_fork)
6165 : return -EPERM;
6166 : }
6167 :
6168 55 : if (user) {
6169 3 : if (attr->sched_flags & SCHED_FLAG_SUGOV)
6170 : return -EINVAL;
6171 :
6172 3 : retval = security_task_setscheduler(p);
6173 3 : if (retval)
6174 : return retval;
6175 : }
6176 :
6177 : /* Update task specific "requested" clamps */
6178 55 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
6179 55 : retval = uclamp_validate(p, attr);
6180 : if (retval)
6181 : return retval;
6182 : }
6183 :
6184 55 : if (pi)
6185 : cpuset_read_lock();
6186 :
6187 : /*
6188 : * Make sure no PI-waiters arrive (or leave) while we are
6189 : * changing the priority of the task:
6190 : *
6191 : * To be able to change p->policy safely, the appropriate
6192 : * runqueue lock must be held.
6193 : */
6194 55 : rq = task_rq_lock(p, &rf);
6195 55 : update_rq_clock(rq);
6196 :
6197 : /*
6198 : * Changing the policy of the stop threads its a very bad idea:
6199 : */
6200 55 : if (p == rq->stop) {
6201 0 : retval = -EINVAL;
6202 0 : goto unlock;
6203 : }
6204 :
6205 : /*
6206 : * If not changing anything there's no need to proceed further,
6207 : * but store a possible modification of reset_on_fork.
6208 : */
6209 55 : if (unlikely(policy == p->policy)) {
6210 51 : if (fair_policy(policy) && attr->sched_nice != task_nice(p))
6211 0 : goto change;
6212 51 : if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
6213 0 : goto change;
6214 51 : if (dl_policy(policy) && dl_param_changed(p, attr))
6215 0 : goto change;
6216 51 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
6217 0 : goto change;
6218 :
6219 51 : p->sched_reset_on_fork = reset_on_fork;
6220 51 : retval = 0;
6221 51 : goto unlock;
6222 : }
6223 4 : change:
6224 :
6225 4 : if (user) {
6226 : #ifdef CONFIG_RT_GROUP_SCHED
6227 : /*
6228 : * Do not allow realtime tasks into groups that have no runtime
6229 : * assigned.
6230 : */
6231 : if (rt_bandwidth_enabled() && rt_policy(policy) &&
6232 : task_group(p)->rt_bandwidth.rt_runtime == 0 &&
6233 : !task_group_is_autogroup(task_group(p))) {
6234 : retval = -EPERM;
6235 : goto unlock;
6236 : }
6237 : #endif
6238 : #ifdef CONFIG_SMP
6239 0 : if (dl_bandwidth_enabled() && dl_policy(policy) &&
6240 0 : !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
6241 0 : cpumask_t *span = rq->rd->span;
6242 :
6243 : /*
6244 : * Don't allow tasks with an affinity mask smaller than
6245 : * the entire root_domain to become SCHED_DEADLINE. We
6246 : * will also fail if there's no bandwidth available.
6247 : */
6248 0 : if (!cpumask_subset(span, p->cpus_ptr) ||
6249 0 : rq->rd->dl_bw.bw == 0) {
6250 0 : retval = -EPERM;
6251 0 : goto unlock;
6252 : }
6253 : }
6254 : #endif
6255 : }
6256 :
6257 : /* Re-check policy now with rq lock held: */
6258 4 : if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6259 0 : policy = oldpolicy = -1;
6260 0 : task_rq_unlock(rq, p, &rf);
6261 0 : if (pi)
6262 : cpuset_read_unlock();
6263 0 : goto recheck;
6264 : }
6265 :
6266 : /*
6267 : * If setscheduling to SCHED_DEADLINE (or changing the parameters
6268 : * of a SCHED_DEADLINE task) we need to check if enough bandwidth
6269 : * is available.
6270 : */
6271 4 : if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
6272 0 : retval = -EBUSY;
6273 0 : goto unlock;
6274 : }
6275 :
6276 4 : p->sched_reset_on_fork = reset_on_fork;
6277 4 : oldprio = p->prio;
6278 :
6279 4 : if (pi) {
6280 : /*
6281 : * Take priority boosted tasks into account. If the new
6282 : * effective priority is unchanged, we just store the new
6283 : * normal parameters and do not touch the scheduler class and
6284 : * the runqueue. This will be done when the task deboost
6285 : * itself.
6286 : */
6287 4 : new_effective_prio = rt_effective_prio(p, newprio);
6288 4 : if (new_effective_prio == oldprio)
6289 0 : queue_flags &= ~DEQUEUE_MOVE;
6290 : }
6291 :
6292 4 : queued = task_on_rq_queued(p);
6293 4 : running = task_current(rq, p);
6294 4 : if (queued)
6295 0 : dequeue_task(rq, p, queue_flags);
6296 4 : if (running)
6297 0 : put_prev_task(rq, p);
6298 :
6299 4 : prev_class = p->sched_class;
6300 :
6301 4 : __setscheduler(rq, p, attr, pi);
6302 4 : __setscheduler_uclamp(p, attr);
6303 :
6304 4 : if (queued) {
6305 : /*
6306 : * We enqueue to tail when the priority of a task is
6307 : * increased (user space view).
6308 : */
6309 0 : if (oldprio < p->prio)
6310 0 : queue_flags |= ENQUEUE_HEAD;
6311 :
6312 0 : enqueue_task(rq, p, queue_flags);
6313 : }
6314 4 : if (running)
6315 0 : set_next_task(rq, p);
6316 :
6317 4 : check_class_changed(rq, p, prev_class, oldprio);
6318 :
6319 : /* Avoid rq from going away on us: */
6320 4 : preempt_disable();
6321 4 : head = splice_balance_callbacks(rq);
6322 4 : task_rq_unlock(rq, p, &rf);
6323 :
6324 4 : if (pi) {
6325 4 : cpuset_read_unlock();
6326 4 : rt_mutex_adjust_pi(p);
6327 : }
6328 :
6329 : /* Run balance callbacks after we've adjusted the PI chain: */
6330 4 : balance_callbacks(rq, head);
6331 4 : preempt_enable();
6332 :
6333 4 : return 0;
6334 :
6335 51 : unlock:
6336 51 : task_rq_unlock(rq, p, &rf);
6337 51 : if (pi)
6338 : cpuset_read_unlock();
6339 51 : return retval;
6340 : }
6341 :
6342 55 : static int _sched_setscheduler(struct task_struct *p, int policy,
6343 : const struct sched_param *param, bool check)
6344 : {
6345 55 : struct sched_attr attr = {
6346 : .sched_policy = policy,
6347 55 : .sched_priority = param->sched_priority,
6348 55 : .sched_nice = PRIO_TO_NICE(p->static_prio),
6349 : };
6350 :
6351 : /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
6352 55 : if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
6353 0 : attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
6354 0 : policy &= ~SCHED_RESET_ON_FORK;
6355 0 : attr.sched_policy = policy;
6356 : }
6357 :
6358 55 : return __sched_setscheduler(p, &attr, check, true);
6359 : }
6360 : /**
6361 : * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6362 : * @p: the task in question.
6363 : * @policy: new policy.
6364 : * @param: structure containing the new RT priority.
6365 : *
6366 : * Use sched_set_fifo(), read its comment.
6367 : *
6368 : * Return: 0 on success. An error code otherwise.
6369 : *
6370 : * NOTE that the task may be already dead.
6371 : */
6372 3 : int sched_setscheduler(struct task_struct *p, int policy,
6373 : const struct sched_param *param)
6374 : {
6375 0 : return _sched_setscheduler(p, policy, param, true);
6376 : }
6377 :
6378 0 : int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
6379 : {
6380 0 : return __sched_setscheduler(p, attr, true, true);
6381 : }
6382 :
6383 0 : int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
6384 : {
6385 0 : return __sched_setscheduler(p, attr, false, true);
6386 : }
6387 :
6388 : /**
6389 : * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6390 : * @p: the task in question.
6391 : * @policy: new policy.
6392 : * @param: structure containing the new RT priority.
6393 : *
6394 : * Just like sched_setscheduler, only don't bother checking if the
6395 : * current context has permission. For example, this is needed in
6396 : * stop_machine(): we create temporary high priority worker threads,
6397 : * but our caller might not have that capability.
6398 : *
6399 : * Return: 0 on success. An error code otherwise.
6400 : */
6401 52 : int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6402 : const struct sched_param *param)
6403 : {
6404 52 : return _sched_setscheduler(p, policy, param, false);
6405 : }
6406 :
6407 : /*
6408 : * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
6409 : * incapable of resource management, which is the one thing an OS really should
6410 : * be doing.
6411 : *
6412 : * This is of course the reason it is limited to privileged users only.
6413 : *
6414 : * Worse still; it is fundamentally impossible to compose static priority
6415 : * workloads. You cannot take two correctly working static prio workloads
6416 : * and smash them together and still expect them to work.
6417 : *
6418 : * For this reason 'all' FIFO tasks the kernel creates are basically at:
6419 : *
6420 : * MAX_RT_PRIO / 2
6421 : *
6422 : * The administrator _MUST_ configure the system, the kernel simply doesn't
6423 : * know enough information to make a sensible choice.
6424 : */
6425 0 : void sched_set_fifo(struct task_struct *p)
6426 : {
6427 0 : struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
6428 0 : WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
6429 0 : }
6430 : EXPORT_SYMBOL_GPL(sched_set_fifo);
6431 :
6432 : /*
6433 : * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
6434 : */
6435 0 : void sched_set_fifo_low(struct task_struct *p)
6436 : {
6437 0 : struct sched_param sp = { .sched_priority = 1 };
6438 0 : WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
6439 0 : }
6440 : EXPORT_SYMBOL_GPL(sched_set_fifo_low);
6441 :
6442 0 : void sched_set_normal(struct task_struct *p, int nice)
6443 : {
6444 0 : struct sched_attr attr = {
6445 : .sched_policy = SCHED_NORMAL,
6446 : .sched_nice = nice,
6447 : };
6448 0 : WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
6449 0 : }
6450 : EXPORT_SYMBOL_GPL(sched_set_normal);
6451 :
6452 : static int
6453 3 : do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6454 : {
6455 3 : struct sched_param lparam;
6456 3 : struct task_struct *p;
6457 3 : int retval;
6458 :
6459 3 : if (!param || pid < 0)
6460 : return -EINVAL;
6461 3 : if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6462 : return -EFAULT;
6463 :
6464 3 : rcu_read_lock();
6465 3 : retval = -ESRCH;
6466 3 : p = find_process_by_pid(pid);
6467 3 : if (likely(p))
6468 3 : get_task_struct(p);
6469 3 : rcu_read_unlock();
6470 :
6471 3 : if (likely(p)) {
6472 3 : retval = sched_setscheduler(p, policy, &lparam);
6473 3 : put_task_struct(p);
6474 : }
6475 :
6476 : return retval;
6477 : }
6478 :
6479 : /*
6480 : * Mimics kernel/events/core.c perf_copy_attr().
6481 : */
6482 0 : static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
6483 : {
6484 0 : u32 size;
6485 0 : int ret;
6486 :
6487 : /* Zero the full structure, so that a short copy will be nice: */
6488 0 : memset(attr, 0, sizeof(*attr));
6489 :
6490 0 : ret = get_user(size, &uattr->size);
6491 0 : if (ret)
6492 : return ret;
6493 :
6494 : /* ABI compatibility quirk: */
6495 0 : if (!size)
6496 0 : size = SCHED_ATTR_SIZE_VER0;
6497 0 : if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
6498 0 : goto err_size;
6499 :
6500 0 : ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
6501 0 : if (ret) {
6502 0 : if (ret == -E2BIG)
6503 0 : goto err_size;
6504 : return ret;
6505 : }
6506 :
6507 0 : if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
6508 : size < SCHED_ATTR_SIZE_VER1)
6509 : return -EINVAL;
6510 :
6511 : /*
6512 : * XXX: Do we want to be lenient like existing syscalls; or do we want
6513 : * to be strict and return an error on out-of-bounds values?
6514 : */
6515 0 : attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
6516 :
6517 0 : return 0;
6518 :
6519 0 : err_size:
6520 0 : put_user(sizeof(*attr), &uattr->size);
6521 0 : return -E2BIG;
6522 : }
6523 :
6524 : /**
6525 : * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6526 : * @pid: the pid in question.
6527 : * @policy: new policy.
6528 : * @param: structure containing the new RT priority.
6529 : *
6530 : * Return: 0 on success. An error code otherwise.
6531 : */
6532 6 : SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
6533 : {
6534 3 : if (policy < 0)
6535 : return -EINVAL;
6536 :
6537 3 : return do_sched_setscheduler(pid, policy, param);
6538 : }
6539 :
6540 : /**
6541 : * sys_sched_setparam - set/change the RT priority of a thread
6542 : * @pid: the pid in question.
6543 : * @param: structure containing the new RT priority.
6544 : *
6545 : * Return: 0 on success. An error code otherwise.
6546 : */
6547 0 : SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6548 : {
6549 0 : return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
6550 : }
6551 :
6552 : /**
6553 : * sys_sched_setattr - same as above, but with extended sched_attr
6554 : * @pid: the pid in question.
6555 : * @uattr: structure containing the extended parameters.
6556 : * @flags: for future extension.
6557 : */
6558 0 : SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
6559 : unsigned int, flags)
6560 : {
6561 0 : struct sched_attr attr;
6562 0 : struct task_struct *p;
6563 0 : int retval;
6564 :
6565 0 : if (!uattr || pid < 0 || flags)
6566 : return -EINVAL;
6567 :
6568 0 : retval = sched_copy_attr(uattr, &attr);
6569 0 : if (retval)
6570 0 : return retval;
6571 :
6572 0 : if ((int)attr.sched_policy < 0)
6573 : return -EINVAL;
6574 0 : if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
6575 0 : attr.sched_policy = SETPARAM_POLICY;
6576 :
6577 0 : rcu_read_lock();
6578 0 : retval = -ESRCH;
6579 0 : p = find_process_by_pid(pid);
6580 0 : if (likely(p))
6581 0 : get_task_struct(p);
6582 0 : rcu_read_unlock();
6583 :
6584 0 : if (likely(p)) {
6585 0 : retval = sched_setattr(p, &attr);
6586 0 : put_task_struct(p);
6587 : }
6588 :
6589 0 : return retval;
6590 : }
6591 :
6592 : /**
6593 : * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6594 : * @pid: the pid in question.
6595 : *
6596 : * Return: On success, the policy of the thread. Otherwise, a negative error
6597 : * code.
6598 : */
6599 4 : SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6600 : {
6601 2 : struct task_struct *p;
6602 2 : int retval;
6603 :
6604 2 : if (pid < 0)
6605 : return -EINVAL;
6606 :
6607 2 : retval = -ESRCH;
6608 2 : rcu_read_lock();
6609 2 : p = find_process_by_pid(pid);
6610 2 : if (p) {
6611 2 : retval = security_task_getscheduler(p);
6612 2 : if (!retval)
6613 2 : retval = p->policy
6614 2 : | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6615 : }
6616 2 : rcu_read_unlock();
6617 2 : return retval;
6618 : }
6619 :
6620 : /**
6621 : * sys_sched_getparam - get the RT priority of a thread
6622 : * @pid: the pid in question.
6623 : * @param: structure containing the RT priority.
6624 : *
6625 : * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
6626 : * code.
6627 : */
6628 4 : SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6629 : {
6630 2 : struct sched_param lp = { .sched_priority = 0 };
6631 2 : struct task_struct *p;
6632 2 : int retval;
6633 :
6634 2 : if (!param || pid < 0)
6635 : return -EINVAL;
6636 :
6637 2 : rcu_read_lock();
6638 2 : p = find_process_by_pid(pid);
6639 2 : retval = -ESRCH;
6640 2 : if (!p)
6641 0 : goto out_unlock;
6642 :
6643 2 : retval = security_task_getscheduler(p);
6644 2 : if (retval)
6645 0 : goto out_unlock;
6646 :
6647 2 : if (task_has_rt_policy(p))
6648 0 : lp.sched_priority = p->rt_priority;
6649 2 : rcu_read_unlock();
6650 :
6651 : /*
6652 : * This one might sleep, we cannot do it with a spinlock held ...
6653 : */
6654 2 : retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6655 :
6656 2 : return retval;
6657 :
6658 0 : out_unlock:
6659 0 : rcu_read_unlock();
6660 0 : return retval;
6661 : }
6662 :
6663 : /*
6664 : * Copy the kernel size attribute structure (which might be larger
6665 : * than what user-space knows about) to user-space.
6666 : *
6667 : * Note that all cases are valid: user-space buffer can be larger or
6668 : * smaller than the kernel-space buffer. The usual case is that both
6669 : * have the same size.
6670 : */
6671 : static int
6672 0 : sched_attr_copy_to_user(struct sched_attr __user *uattr,
6673 : struct sched_attr *kattr,
6674 : unsigned int usize)
6675 : {
6676 0 : unsigned int ksize = sizeof(*kattr);
6677 :
6678 0 : if (!access_ok(uattr, usize))
6679 : return -EFAULT;
6680 :
6681 : /*
6682 : * sched_getattr() ABI forwards and backwards compatibility:
6683 : *
6684 : * If usize == ksize then we just copy everything to user-space and all is good.
6685 : *
6686 : * If usize < ksize then we only copy as much as user-space has space for,
6687 : * this keeps ABI compatibility as well. We skip the rest.
6688 : *
6689 : * If usize > ksize then user-space is using a newer version of the ABI,
6690 : * which part the kernel doesn't know about. Just ignore it - tooling can
6691 : * detect the kernel's knowledge of attributes from the attr->size value
6692 : * which is set to ksize in this case.
6693 : */
6694 0 : kattr->size = min(usize, ksize);
6695 :
6696 0 : if (copy_to_user(uattr, kattr, kattr->size))
6697 0 : return -EFAULT;
6698 :
6699 : return 0;
6700 : }
6701 :
6702 : /**
6703 : * sys_sched_getattr - similar to sched_getparam, but with sched_attr
6704 : * @pid: the pid in question.
6705 : * @uattr: structure containing the extended parameters.
6706 : * @usize: sizeof(attr) for fwd/bwd comp.
6707 : * @flags: for future extension.
6708 : */
6709 0 : SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
6710 : unsigned int, usize, unsigned int, flags)
6711 : {
6712 0 : struct sched_attr kattr = { };
6713 0 : struct task_struct *p;
6714 0 : int retval;
6715 :
6716 0 : if (!uattr || pid < 0 || usize > PAGE_SIZE ||
6717 0 : usize < SCHED_ATTR_SIZE_VER0 || flags)
6718 : return -EINVAL;
6719 :
6720 0 : rcu_read_lock();
6721 0 : p = find_process_by_pid(pid);
6722 0 : retval = -ESRCH;
6723 0 : if (!p)
6724 0 : goto out_unlock;
6725 :
6726 0 : retval = security_task_getscheduler(p);
6727 0 : if (retval)
6728 0 : goto out_unlock;
6729 :
6730 0 : kattr.sched_policy = p->policy;
6731 0 : if (p->sched_reset_on_fork)
6732 0 : kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
6733 0 : if (task_has_dl_policy(p))
6734 0 : __getparam_dl(p, &kattr);
6735 0 : else if (task_has_rt_policy(p))
6736 0 : kattr.sched_priority = p->rt_priority;
6737 : else
6738 0 : kattr.sched_nice = task_nice(p);
6739 :
6740 : #ifdef CONFIG_UCLAMP_TASK
6741 : /*
6742 : * This could race with another potential updater, but this is fine
6743 : * because it'll correctly read the old or the new value. We don't need
6744 : * to guarantee who wins the race as long as it doesn't return garbage.
6745 : */
6746 : kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
6747 : kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
6748 : #endif
6749 :
6750 0 : rcu_read_unlock();
6751 :
6752 0 : return sched_attr_copy_to_user(uattr, &kattr, usize);
6753 :
6754 0 : out_unlock:
6755 0 : rcu_read_unlock();
6756 0 : return retval;
6757 : }
6758 :
6759 0 : long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6760 : {
6761 0 : cpumask_var_t cpus_allowed, new_mask;
6762 0 : struct task_struct *p;
6763 0 : int retval;
6764 :
6765 0 : rcu_read_lock();
6766 :
6767 0 : p = find_process_by_pid(pid);
6768 0 : if (!p) {
6769 0 : rcu_read_unlock();
6770 0 : return -ESRCH;
6771 : }
6772 :
6773 : /* Prevent p going away */
6774 0 : get_task_struct(p);
6775 0 : rcu_read_unlock();
6776 :
6777 0 : if (p->flags & PF_NO_SETAFFINITY) {
6778 0 : retval = -EINVAL;
6779 0 : goto out_put_task;
6780 : }
6781 0 : if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6782 : retval = -ENOMEM;
6783 : goto out_put_task;
6784 : }
6785 0 : if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6786 : retval = -ENOMEM;
6787 : goto out_free_cpus_allowed;
6788 : }
6789 0 : retval = -EPERM;
6790 0 : if (!check_same_owner(p)) {
6791 0 : rcu_read_lock();
6792 0 : if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
6793 0 : rcu_read_unlock();
6794 0 : goto out_free_new_mask;
6795 : }
6796 0 : rcu_read_unlock();
6797 : }
6798 :
6799 0 : retval = security_task_setscheduler(p);
6800 0 : if (retval)
6801 0 : goto out_free_new_mask;
6802 :
6803 :
6804 0 : cpuset_cpus_allowed(p, cpus_allowed);
6805 0 : cpumask_and(new_mask, in_mask, cpus_allowed);
6806 :
6807 : /*
6808 : * Since bandwidth control happens on root_domain basis,
6809 : * if admission test is enabled, we only admit -deadline
6810 : * tasks allowed to run on all the CPUs in the task's
6811 : * root_domain.
6812 : */
6813 : #ifdef CONFIG_SMP
6814 0 : if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
6815 0 : rcu_read_lock();
6816 0 : if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
6817 0 : retval = -EBUSY;
6818 0 : rcu_read_unlock();
6819 0 : goto out_free_new_mask;
6820 : }
6821 0 : rcu_read_unlock();
6822 : }
6823 : #endif
6824 0 : again:
6825 0 : retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK);
6826 :
6827 0 : if (!retval) {
6828 0 : cpuset_cpus_allowed(p, cpus_allowed);
6829 0 : if (!cpumask_subset(new_mask, cpus_allowed)) {
6830 : /*
6831 : * We must have raced with a concurrent cpuset
6832 : * update. Just reset the cpus_allowed to the
6833 : * cpuset's cpus_allowed
6834 : */
6835 0 : cpumask_copy(new_mask, cpus_allowed);
6836 0 : goto again;
6837 : }
6838 : }
6839 0 : out_free_new_mask:
6840 0 : free_cpumask_var(new_mask);
6841 0 : out_free_cpus_allowed:
6842 0 : free_cpumask_var(cpus_allowed);
6843 0 : out_put_task:
6844 0 : put_task_struct(p);
6845 0 : return retval;
6846 : }
6847 :
6848 0 : static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6849 : struct cpumask *new_mask)
6850 : {
6851 0 : if (len < cpumask_size())
6852 0 : cpumask_clear(new_mask);
6853 0 : else if (len > cpumask_size())
6854 0 : len = cpumask_size();
6855 :
6856 0 : return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6857 : }
6858 :
6859 : /**
6860 : * sys_sched_setaffinity - set the CPU affinity of a process
6861 : * @pid: pid of the process
6862 : * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6863 : * @user_mask_ptr: user-space pointer to the new CPU mask
6864 : *
6865 : * Return: 0 on success. An error code otherwise.
6866 : */
6867 0 : SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6868 : unsigned long __user *, user_mask_ptr)
6869 : {
6870 0 : cpumask_var_t new_mask;
6871 0 : int retval;
6872 :
6873 0 : if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6874 : return -ENOMEM;
6875 :
6876 0 : retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6877 0 : if (retval == 0)
6878 0 : retval = sched_setaffinity(pid, new_mask);
6879 0 : free_cpumask_var(new_mask);
6880 0 : return retval;
6881 : }
6882 :
6883 1 : long sched_getaffinity(pid_t pid, struct cpumask *mask)
6884 : {
6885 1 : struct task_struct *p;
6886 1 : unsigned long flags;
6887 1 : int retval;
6888 :
6889 1 : rcu_read_lock();
6890 :
6891 1 : retval = -ESRCH;
6892 1 : p = find_process_by_pid(pid);
6893 1 : if (!p)
6894 0 : goto out_unlock;
6895 :
6896 1 : retval = security_task_getscheduler(p);
6897 1 : if (retval)
6898 0 : goto out_unlock;
6899 :
6900 1 : raw_spin_lock_irqsave(&p->pi_lock, flags);
6901 1 : cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
6902 1 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6903 :
6904 1 : out_unlock:
6905 1 : rcu_read_unlock();
6906 :
6907 1 : return retval;
6908 : }
6909 :
6910 : /**
6911 : * sys_sched_getaffinity - get the CPU affinity of a process
6912 : * @pid: pid of the process
6913 : * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6914 : * @user_mask_ptr: user-space pointer to hold the current CPU mask
6915 : *
6916 : * Return: size of CPU mask copied to user_mask_ptr on success. An
6917 : * error code otherwise.
6918 : */
6919 2 : SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6920 : unsigned long __user *, user_mask_ptr)
6921 : {
6922 1 : int ret;
6923 1 : cpumask_var_t mask;
6924 :
6925 1 : if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6926 : return -EINVAL;
6927 1 : if (len & (sizeof(unsigned long)-1))
6928 : return -EINVAL;
6929 :
6930 1 : if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6931 : return -ENOMEM;
6932 :
6933 1 : ret = sched_getaffinity(pid, mask);
6934 1 : if (ret == 0) {
6935 1 : unsigned int retlen = min(len, cpumask_size());
6936 :
6937 2 : if (copy_to_user(user_mask_ptr, mask, retlen))
6938 : ret = -EFAULT;
6939 : else
6940 1 : ret = retlen;
6941 : }
6942 1 : free_cpumask_var(mask);
6943 :
6944 1 : return ret;
6945 : }
6946 :
6947 0 : static void do_sched_yield(void)
6948 : {
6949 0 : struct rq_flags rf;
6950 0 : struct rq *rq;
6951 :
6952 0 : rq = this_rq_lock_irq(&rf);
6953 :
6954 0 : schedstat_inc(rq->yld_count);
6955 0 : current->sched_class->yield_task(rq);
6956 :
6957 0 : preempt_disable();
6958 0 : rq_unlock_irq(rq, &rf);
6959 0 : sched_preempt_enable_no_resched();
6960 :
6961 0 : schedule();
6962 0 : }
6963 :
6964 : /**
6965 : * sys_sched_yield - yield the current processor to other threads.
6966 : *
6967 : * This function yields the current CPU to other tasks. If there are no
6968 : * other threads running on this CPU then this function will return.
6969 : *
6970 : * Return: 0.
6971 : */
6972 0 : SYSCALL_DEFINE0(sched_yield)
6973 : {
6974 0 : do_sched_yield();
6975 0 : return 0;
6976 : }
6977 :
6978 : #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
6979 486332 : int __sched __cond_resched(void)
6980 : {
6981 486332 : if (should_resched(0)) {
6982 2536 : preempt_schedule_common();
6983 2536 : return 1;
6984 : }
6985 : #ifndef CONFIG_PREEMPT_RCU
6986 483796 : rcu_all_qs();
6987 : #endif
6988 483796 : return 0;
6989 : }
6990 : EXPORT_SYMBOL(__cond_resched);
6991 : #endif
6992 :
6993 : #ifdef CONFIG_PREEMPT_DYNAMIC
6994 : DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
6995 : EXPORT_STATIC_CALL_TRAMP(cond_resched);
6996 :
6997 : DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
6998 : EXPORT_STATIC_CALL_TRAMP(might_resched);
6999 : #endif
7000 :
7001 : /*
7002 : * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
7003 : * call schedule, and on return reacquire the lock.
7004 : *
7005 : * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
7006 : * operations here to prevent schedule() from being called twice (once via
7007 : * spin_unlock(), once by hand).
7008 : */
7009 1656 : int __cond_resched_lock(spinlock_t *lock)
7010 : {
7011 1656 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
7012 1656 : int ret = 0;
7013 :
7014 3312 : lockdep_assert_held(lock);
7015 :
7016 1656 : if (spin_needbreak(lock) || resched) {
7017 1 : spin_unlock(lock);
7018 1 : if (resched)
7019 1 : preempt_schedule_common();
7020 : else
7021 : cpu_relax();
7022 1 : ret = 1;
7023 1 : spin_lock(lock);
7024 : }
7025 1656 : return ret;
7026 : }
7027 : EXPORT_SYMBOL(__cond_resched_lock);
7028 :
7029 0 : int __cond_resched_rwlock_read(rwlock_t *lock)
7030 : {
7031 0 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
7032 0 : int ret = 0;
7033 :
7034 0 : lockdep_assert_held_read(lock);
7035 :
7036 0 : if (rwlock_needbreak(lock) || resched) {
7037 0 : read_unlock(lock);
7038 0 : if (resched)
7039 0 : preempt_schedule_common();
7040 : else
7041 : cpu_relax();
7042 0 : ret = 1;
7043 0 : read_lock(lock);
7044 : }
7045 0 : return ret;
7046 : }
7047 : EXPORT_SYMBOL(__cond_resched_rwlock_read);
7048 :
7049 0 : int __cond_resched_rwlock_write(rwlock_t *lock)
7050 : {
7051 0 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
7052 0 : int ret = 0;
7053 :
7054 0 : lockdep_assert_held_write(lock);
7055 :
7056 0 : if (rwlock_needbreak(lock) || resched) {
7057 0 : write_unlock(lock);
7058 0 : if (resched)
7059 0 : preempt_schedule_common();
7060 : else
7061 : cpu_relax();
7062 0 : ret = 1;
7063 0 : write_lock(lock);
7064 : }
7065 0 : return ret;
7066 : }
7067 : EXPORT_SYMBOL(__cond_resched_rwlock_write);
7068 :
7069 : /**
7070 : * yield - yield the current processor to other threads.
7071 : *
7072 : * Do not ever use this function, there's a 99% chance you're doing it wrong.
7073 : *
7074 : * The scheduler is at all times free to pick the calling task as the most
7075 : * eligible task to run, if removing the yield() call from your code breaks
7076 : * it, it's already broken.
7077 : *
7078 : * Typical broken usage is:
7079 : *
7080 : * while (!event)
7081 : * yield();
7082 : *
7083 : * where one assumes that yield() will let 'the other' process run that will
7084 : * make event true. If the current task is a SCHED_FIFO task that will never
7085 : * happen. Never use yield() as a progress guarantee!!
7086 : *
7087 : * If you want to use yield() to wait for something, use wait_event().
7088 : * If you want to use yield() to be 'nice' for others, use cond_resched().
7089 : * If you still want to use yield(), do not!
7090 : */
7091 0 : void __sched yield(void)
7092 : {
7093 0 : set_current_state(TASK_RUNNING);
7094 0 : do_sched_yield();
7095 0 : }
7096 : EXPORT_SYMBOL(yield);
7097 :
7098 : /**
7099 : * yield_to - yield the current processor to another thread in
7100 : * your thread group, or accelerate that thread toward the
7101 : * processor it's on.
7102 : * @p: target task
7103 : * @preempt: whether task preemption is allowed or not
7104 : *
7105 : * It's the caller's job to ensure that the target task struct
7106 : * can't go away on us before we can do any checks.
7107 : *
7108 : * Return:
7109 : * true (>0) if we indeed boosted the target task.
7110 : * false (0) if we failed to boost the target.
7111 : * -ESRCH if there's no task to yield to.
7112 : */
7113 0 : int __sched yield_to(struct task_struct *p, bool preempt)
7114 : {
7115 0 : struct task_struct *curr = current;
7116 0 : struct rq *rq, *p_rq;
7117 0 : unsigned long flags;
7118 0 : int yielded = 0;
7119 :
7120 0 : local_irq_save(flags);
7121 0 : rq = this_rq();
7122 :
7123 0 : again:
7124 0 : p_rq = task_rq(p);
7125 : /*
7126 : * If we're the only runnable task on the rq and target rq also
7127 : * has only one task, there's absolutely no point in yielding.
7128 : */
7129 0 : if (rq->nr_running == 1 && p_rq->nr_running == 1) {
7130 0 : yielded = -ESRCH;
7131 0 : goto out_irq;
7132 : }
7133 :
7134 0 : double_rq_lock(rq, p_rq);
7135 0 : if (task_rq(p) != p_rq) {
7136 0 : double_rq_unlock(rq, p_rq);
7137 0 : goto again;
7138 : }
7139 :
7140 0 : if (!curr->sched_class->yield_to_task)
7141 0 : goto out_unlock;
7142 :
7143 0 : if (curr->sched_class != p->sched_class)
7144 0 : goto out_unlock;
7145 :
7146 0 : if (task_running(p_rq, p) || p->state)
7147 0 : goto out_unlock;
7148 :
7149 0 : yielded = curr->sched_class->yield_to_task(rq, p);
7150 0 : if (yielded) {
7151 0 : schedstat_inc(rq->yld_count);
7152 : /*
7153 : * Make p's CPU reschedule; pick_next_entity takes care of
7154 : * fairness.
7155 : */
7156 0 : if (preempt && rq != p_rq)
7157 0 : resched_curr(p_rq);
7158 : }
7159 :
7160 0 : out_unlock:
7161 0 : double_rq_unlock(rq, p_rq);
7162 0 : out_irq:
7163 0 : local_irq_restore(flags);
7164 :
7165 0 : if (yielded > 0)
7166 0 : schedule();
7167 :
7168 0 : return yielded;
7169 : }
7170 : EXPORT_SYMBOL_GPL(yield_to);
7171 :
7172 1875 : int io_schedule_prepare(void)
7173 : {
7174 1875 : int old_iowait = current->in_iowait;
7175 :
7176 1875 : current->in_iowait = 1;
7177 1875 : blk_schedule_flush_plug(current);
7178 :
7179 1875 : return old_iowait;
7180 : }
7181 :
7182 1875 : void io_schedule_finish(int token)
7183 : {
7184 1875 : current->in_iowait = token;
7185 1 : }
7186 :
7187 : /*
7188 : * This task is about to go to sleep on IO. Increment rq->nr_iowait so
7189 : * that process accounting knows that this is a task in IO wait state.
7190 : */
7191 74 : long __sched io_schedule_timeout(long timeout)
7192 : {
7193 74 : int token;
7194 74 : long ret;
7195 :
7196 74 : token = io_schedule_prepare();
7197 74 : ret = schedule_timeout(timeout);
7198 74 : io_schedule_finish(token);
7199 :
7200 74 : return ret;
7201 : }
7202 : EXPORT_SYMBOL(io_schedule_timeout);
7203 :
7204 1800 : void __sched io_schedule(void)
7205 : {
7206 1800 : int token;
7207 :
7208 1800 : token = io_schedule_prepare();
7209 1800 : schedule();
7210 1800 : io_schedule_finish(token);
7211 1800 : }
7212 : EXPORT_SYMBOL(io_schedule);
7213 :
7214 : /**
7215 : * sys_sched_get_priority_max - return maximum RT priority.
7216 : * @policy: scheduling class.
7217 : *
7218 : * Return: On success, this syscall returns the maximum
7219 : * rt_priority that can be used by a given scheduling class.
7220 : * On failure, a negative error code is returned.
7221 : */
7222 4 : SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
7223 : {
7224 2 : int ret = -EINVAL;
7225 :
7226 2 : switch (policy) {
7227 0 : case SCHED_FIFO:
7228 : case SCHED_RR:
7229 0 : ret = MAX_RT_PRIO-1;
7230 0 : break;
7231 2 : case SCHED_DEADLINE:
7232 : case SCHED_NORMAL:
7233 : case SCHED_BATCH:
7234 : case SCHED_IDLE:
7235 2 : ret = 0;
7236 2 : break;
7237 : }
7238 2 : return ret;
7239 : }
7240 :
7241 : /**
7242 : * sys_sched_get_priority_min - return minimum RT priority.
7243 : * @policy: scheduling class.
7244 : *
7245 : * Return: On success, this syscall returns the minimum
7246 : * rt_priority that can be used by a given scheduling class.
7247 : * On failure, a negative error code is returned.
7248 : */
7249 4 : SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
7250 : {
7251 2 : int ret = -EINVAL;
7252 :
7253 2 : switch (policy) {
7254 0 : case SCHED_FIFO:
7255 : case SCHED_RR:
7256 0 : ret = 1;
7257 0 : break;
7258 2 : case SCHED_DEADLINE:
7259 : case SCHED_NORMAL:
7260 : case SCHED_BATCH:
7261 : case SCHED_IDLE:
7262 2 : ret = 0;
7263 : }
7264 2 : return ret;
7265 : }
7266 :
7267 0 : static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
7268 : {
7269 0 : struct task_struct *p;
7270 0 : unsigned int time_slice;
7271 0 : struct rq_flags rf;
7272 0 : struct rq *rq;
7273 0 : int retval;
7274 :
7275 0 : if (pid < 0)
7276 : return -EINVAL;
7277 :
7278 0 : retval = -ESRCH;
7279 0 : rcu_read_lock();
7280 0 : p = find_process_by_pid(pid);
7281 0 : if (!p)
7282 0 : goto out_unlock;
7283 :
7284 0 : retval = security_task_getscheduler(p);
7285 0 : if (retval)
7286 0 : goto out_unlock;
7287 :
7288 0 : rq = task_rq_lock(p, &rf);
7289 0 : time_slice = 0;
7290 0 : if (p->sched_class->get_rr_interval)
7291 0 : time_slice = p->sched_class->get_rr_interval(rq, p);
7292 0 : task_rq_unlock(rq, p, &rf);
7293 :
7294 0 : rcu_read_unlock();
7295 0 : jiffies_to_timespec64(time_slice, t);
7296 0 : return 0;
7297 :
7298 0 : out_unlock:
7299 0 : rcu_read_unlock();
7300 0 : return retval;
7301 : }
7302 :
7303 : /**
7304 : * sys_sched_rr_get_interval - return the default timeslice of a process.
7305 : * @pid: pid of the process.
7306 : * @interval: userspace pointer to the timeslice value.
7307 : *
7308 : * this syscall writes the default timeslice value of a given process
7309 : * into the user-space timespec buffer. A value of '0' means infinity.
7310 : *
7311 : * Return: On success, 0 and the timeslice is in @interval. Otherwise,
7312 : * an error code.
7313 : */
7314 0 : SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
7315 : struct __kernel_timespec __user *, interval)
7316 : {
7317 0 : struct timespec64 t;
7318 0 : int retval = sched_rr_get_interval(pid, &t);
7319 :
7320 0 : if (retval == 0)
7321 0 : retval = put_timespec64(&t, interval);
7322 :
7323 0 : return retval;
7324 : }
7325 :
7326 : #ifdef CONFIG_COMPAT_32BIT_TIME
7327 : SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
7328 : struct old_timespec32 __user *, interval)
7329 : {
7330 : struct timespec64 t;
7331 : int retval = sched_rr_get_interval(pid, &t);
7332 :
7333 : if (retval == 0)
7334 : retval = put_old_timespec32(&t, interval);
7335 : return retval;
7336 : }
7337 : #endif
7338 :
7339 0 : void sched_show_task(struct task_struct *p)
7340 : {
7341 0 : unsigned long free = 0;
7342 0 : int ppid;
7343 :
7344 0 : if (!try_get_task_stack(p))
7345 : return;
7346 :
7347 0 : pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
7348 :
7349 0 : if (p->state == TASK_RUNNING)
7350 0 : pr_cont(" running task ");
7351 : #ifdef CONFIG_DEBUG_STACK_USAGE
7352 : free = stack_not_used(p);
7353 : #endif
7354 0 : ppid = 0;
7355 0 : rcu_read_lock();
7356 0 : if (pid_alive(p))
7357 0 : ppid = task_pid_nr(rcu_dereference(p->real_parent));
7358 0 : rcu_read_unlock();
7359 0 : pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
7360 : free, task_pid_nr(p), ppid,
7361 : (unsigned long)task_thread_info(p)->flags);
7362 :
7363 0 : print_worker_info(KERN_INFO, p);
7364 0 : print_stop_info(KERN_INFO, p);
7365 0 : show_stack(p, NULL, KERN_INFO);
7366 0 : put_task_stack(p);
7367 : }
7368 : EXPORT_SYMBOL_GPL(sched_show_task);
7369 :
7370 : static inline bool
7371 0 : state_filter_match(unsigned long state_filter, struct task_struct *p)
7372 : {
7373 : /* no filter, everything matches */
7374 0 : if (!state_filter)
7375 : return true;
7376 :
7377 : /* filter, but doesn't match */
7378 0 : if (!(p->state & state_filter))
7379 : return false;
7380 :
7381 : /*
7382 : * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
7383 : * TASK_KILLABLE).
7384 : */
7385 0 : if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
7386 : return false;
7387 :
7388 : return true;
7389 : }
7390 :
7391 :
7392 0 : void show_state_filter(unsigned long state_filter)
7393 : {
7394 0 : struct task_struct *g, *p;
7395 :
7396 0 : rcu_read_lock();
7397 0 : for_each_process_thread(g, p) {
7398 : /*
7399 : * reset the NMI-timeout, listing all files on a slow
7400 : * console might take a lot of time:
7401 : * Also, reset softlockup watchdogs on all CPUs, because
7402 : * another CPU might be blocked waiting for us to process
7403 : * an IPI.
7404 : */
7405 0 : touch_nmi_watchdog();
7406 0 : touch_all_softlockup_watchdogs();
7407 0 : if (state_filter_match(state_filter, p))
7408 0 : sched_show_task(p);
7409 : }
7410 :
7411 : #ifdef CONFIG_SCHED_DEBUG
7412 : if (!state_filter)
7413 : sysrq_sched_debug_show();
7414 : #endif
7415 0 : rcu_read_unlock();
7416 : /*
7417 : * Only show locks if all tasks are dumped:
7418 : */
7419 0 : if (!state_filter)
7420 0 : debug_show_all_locks();
7421 0 : }
7422 :
7423 : /**
7424 : * init_idle - set up an idle thread for a given CPU
7425 : * @idle: task in question
7426 : * @cpu: CPU the idle task belongs to
7427 : *
7428 : * NOTE: this function does not set the idle thread's NEED_RESCHED
7429 : * flag, to make booting more robust.
7430 : */
7431 10 : void init_idle(struct task_struct *idle, int cpu)
7432 : {
7433 10 : struct rq *rq = cpu_rq(cpu);
7434 10 : unsigned long flags;
7435 :
7436 10 : __sched_fork(0, idle);
7437 :
7438 10 : raw_spin_lock_irqsave(&idle->pi_lock, flags);
7439 10 : raw_spin_lock(&rq->lock);
7440 :
7441 10 : idle->state = TASK_RUNNING;
7442 10 : idle->se.exec_start = sched_clock();
7443 10 : idle->flags |= PF_IDLE;
7444 :
7445 10 : scs_task_reset(idle);
7446 10 : kasan_unpoison_task_stack(idle);
7447 :
7448 : #ifdef CONFIG_SMP
7449 : /*
7450 : * It's possible that init_idle() gets called multiple times on a task,
7451 : * in that case do_set_cpus_allowed() will not do the right thing.
7452 : *
7453 : * And since this is boot we can forgo the serialization.
7454 : */
7455 10 : set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
7456 : #endif
7457 : /*
7458 : * We're having a chicken and egg problem, even though we are
7459 : * holding rq->lock, the CPU isn't yet set to this CPU so the
7460 : * lockdep check in task_group() will fail.
7461 : *
7462 : * Similar case to sched_fork(). / Alternatively we could
7463 : * use task_rq_lock() here and obtain the other rq->lock.
7464 : *
7465 : * Silence PROVE_RCU
7466 : */
7467 10 : rcu_read_lock();
7468 10 : __set_task_cpu(idle, cpu);
7469 10 : rcu_read_unlock();
7470 :
7471 10 : rq->idle = idle;
7472 10 : rcu_assign_pointer(rq->curr, idle);
7473 10 : idle->on_rq = TASK_ON_RQ_QUEUED;
7474 : #ifdef CONFIG_SMP
7475 10 : idle->on_cpu = 1;
7476 : #endif
7477 10 : raw_spin_unlock(&rq->lock);
7478 10 : raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
7479 :
7480 : /* Set the preempt count _outside_ the spinlocks! */
7481 10 : init_idle_preempt_count(idle, cpu);
7482 :
7483 : /*
7484 : * The idle tasks have their own, simple scheduling class:
7485 : */
7486 10 : idle->sched_class = &idle_sched_class;
7487 10 : ftrace_graph_init_idle_task(idle, cpu);
7488 10 : vtime_init_idle(idle, cpu);
7489 : #ifdef CONFIG_SMP
7490 10 : sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
7491 : #endif
7492 10 : }
7493 :
7494 : #ifdef CONFIG_SMP
7495 :
7496 0 : int cpuset_cpumask_can_shrink(const struct cpumask *cur,
7497 : const struct cpumask *trial)
7498 : {
7499 0 : int ret = 1;
7500 :
7501 0 : if (!cpumask_weight(cur))
7502 : return ret;
7503 :
7504 0 : ret = dl_cpuset_cpumask_can_shrink(cur, trial);
7505 :
7506 0 : return ret;
7507 : }
7508 :
7509 0 : int task_can_attach(struct task_struct *p,
7510 : const struct cpumask *cs_cpus_allowed)
7511 : {
7512 0 : int ret = 0;
7513 :
7514 : /*
7515 : * Kthreads which disallow setaffinity shouldn't be moved
7516 : * to a new cpuset; we don't want to change their CPU
7517 : * affinity and isolating such threads by their set of
7518 : * allowed nodes is unnecessary. Thus, cpusets are not
7519 : * applicable for such threads. This prevents checking for
7520 : * success of set_cpus_allowed_ptr() on all attached tasks
7521 : * before cpus_mask may be changed.
7522 : */
7523 0 : if (p->flags & PF_NO_SETAFFINITY) {
7524 0 : ret = -EINVAL;
7525 0 : goto out;
7526 : }
7527 :
7528 0 : if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
7529 : cs_cpus_allowed))
7530 0 : ret = dl_task_can_attach(p, cs_cpus_allowed);
7531 :
7532 0 : out:
7533 0 : return ret;
7534 : }
7535 :
7536 : bool sched_smp_initialized __read_mostly;
7537 :
7538 : #ifdef CONFIG_NUMA_BALANCING
7539 : /* Migrate current task p to target_cpu */
7540 : int migrate_task_to(struct task_struct *p, int target_cpu)
7541 : {
7542 : struct migration_arg arg = { p, target_cpu };
7543 : int curr_cpu = task_cpu(p);
7544 :
7545 : if (curr_cpu == target_cpu)
7546 : return 0;
7547 :
7548 : if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
7549 : return -EINVAL;
7550 :
7551 : /* TODO: This is not properly updating schedstats */
7552 :
7553 : trace_sched_move_numa(p, curr_cpu, target_cpu);
7554 : return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
7555 : }
7556 :
7557 : /*
7558 : * Requeue a task on a given node and accurately track the number of NUMA
7559 : * tasks on the runqueues
7560 : */
7561 : void sched_setnuma(struct task_struct *p, int nid)
7562 : {
7563 : bool queued, running;
7564 : struct rq_flags rf;
7565 : struct rq *rq;
7566 :
7567 : rq = task_rq_lock(p, &rf);
7568 : queued = task_on_rq_queued(p);
7569 : running = task_current(rq, p);
7570 :
7571 : if (queued)
7572 : dequeue_task(rq, p, DEQUEUE_SAVE);
7573 : if (running)
7574 : put_prev_task(rq, p);
7575 :
7576 : p->numa_preferred_nid = nid;
7577 :
7578 : if (queued)
7579 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7580 : if (running)
7581 : set_next_task(rq, p);
7582 : task_rq_unlock(rq, p, &rf);
7583 : }
7584 : #endif /* CONFIG_NUMA_BALANCING */
7585 :
7586 : #ifdef CONFIG_HOTPLUG_CPU
7587 : /*
7588 : * Ensure that the idle task is using init_mm right before its CPU goes
7589 : * offline.
7590 : */
7591 0 : void idle_task_exit(void)
7592 : {
7593 0 : struct mm_struct *mm = current->active_mm;
7594 :
7595 0 : BUG_ON(cpu_online(smp_processor_id()));
7596 0 : BUG_ON(current != this_rq()->idle);
7597 :
7598 0 : if (mm != &init_mm) {
7599 0 : switch_mm(mm, &init_mm, current);
7600 0 : finish_arch_post_lock_switch();
7601 : }
7602 :
7603 : /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
7604 0 : }
7605 :
7606 0 : static int __balance_push_cpu_stop(void *arg)
7607 : {
7608 0 : struct task_struct *p = arg;
7609 0 : struct rq *rq = this_rq();
7610 0 : struct rq_flags rf;
7611 0 : int cpu;
7612 :
7613 0 : raw_spin_lock_irq(&p->pi_lock);
7614 0 : rq_lock(rq, &rf);
7615 :
7616 0 : update_rq_clock(rq);
7617 :
7618 0 : if (task_rq(p) == rq && task_on_rq_queued(p)) {
7619 0 : cpu = select_fallback_rq(rq->cpu, p);
7620 0 : rq = __migrate_task(rq, &rf, p, cpu);
7621 : }
7622 :
7623 0 : rq_unlock(rq, &rf);
7624 0 : raw_spin_unlock_irq(&p->pi_lock);
7625 :
7626 0 : put_task_struct(p);
7627 :
7628 0 : return 0;
7629 : }
7630 :
7631 : static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
7632 :
7633 : /*
7634 : * Ensure we only run per-cpu kthreads once the CPU goes !active.
7635 : */
7636 0 : static void balance_push(struct rq *rq)
7637 : {
7638 0 : struct task_struct *push_task = rq->curr;
7639 :
7640 0 : lockdep_assert_held(&rq->lock);
7641 0 : SCHED_WARN_ON(rq->cpu != smp_processor_id());
7642 : /*
7643 : * Ensure the thing is persistent until balance_push_set(.on = false);
7644 : */
7645 0 : rq->balance_callback = &balance_push_callback;
7646 :
7647 : /*
7648 : * Both the cpu-hotplug and stop task are in this case and are
7649 : * required to complete the hotplug process.
7650 : *
7651 : * XXX: the idle task does not match kthread_is_per_cpu() due to
7652 : * histerical raisins.
7653 : */
7654 0 : if (rq->idle == push_task ||
7655 0 : ((push_task->flags & PF_KTHREAD) && kthread_is_per_cpu(push_task)) ||
7656 0 : is_migration_disabled(push_task)) {
7657 :
7658 : /*
7659 : * If this is the idle task on the outgoing CPU try to wake
7660 : * up the hotplug control thread which might wait for the
7661 : * last task to vanish. The rcuwait_active() check is
7662 : * accurate here because the waiter is pinned on this CPU
7663 : * and can't obviously be running in parallel.
7664 : *
7665 : * On RT kernels this also has to check whether there are
7666 : * pinned and scheduled out tasks on the runqueue. They
7667 : * need to leave the migrate disabled section first.
7668 : */
7669 0 : if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
7670 0 : rcuwait_active(&rq->hotplug_wait)) {
7671 0 : raw_spin_unlock(&rq->lock);
7672 0 : rcuwait_wake_up(&rq->hotplug_wait);
7673 0 : raw_spin_lock(&rq->lock);
7674 : }
7675 0 : return;
7676 : }
7677 :
7678 0 : get_task_struct(push_task);
7679 : /*
7680 : * Temporarily drop rq->lock such that we can wake-up the stop task.
7681 : * Both preemption and IRQs are still disabled.
7682 : */
7683 0 : raw_spin_unlock(&rq->lock);
7684 0 : stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
7685 0 : this_cpu_ptr(&push_work));
7686 : /*
7687 : * At this point need_resched() is true and we'll take the loop in
7688 : * schedule(). The next pick is obviously going to be the stop task
7689 : * which kthread_is_per_cpu() and will push this task away.
7690 : */
7691 0 : raw_spin_lock(&rq->lock);
7692 : }
7693 :
7694 3 : static void balance_push_set(int cpu, bool on)
7695 : {
7696 3 : struct rq *rq = cpu_rq(cpu);
7697 3 : struct rq_flags rf;
7698 :
7699 3 : rq_lock_irqsave(rq, &rf);
7700 3 : rq->balance_push = on;
7701 3 : if (on) {
7702 0 : WARN_ON_ONCE(rq->balance_callback);
7703 0 : rq->balance_callback = &balance_push_callback;
7704 3 : } else if (rq->balance_callback == &balance_push_callback) {
7705 0 : rq->balance_callback = NULL;
7706 : }
7707 3 : rq_unlock_irqrestore(rq, &rf);
7708 3 : }
7709 :
7710 : /*
7711 : * Invoked from a CPUs hotplug control thread after the CPU has been marked
7712 : * inactive. All tasks which are not per CPU kernel threads are either
7713 : * pushed off this CPU now via balance_push() or placed on a different CPU
7714 : * during wakeup. Wait until the CPU is quiescent.
7715 : */
7716 0 : static void balance_hotplug_wait(void)
7717 : {
7718 0 : struct rq *rq = this_rq();
7719 :
7720 0 : rcuwait_wait_event(&rq->hotplug_wait,
7721 : rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
7722 : TASK_UNINTERRUPTIBLE);
7723 0 : }
7724 :
7725 : #else
7726 :
7727 : static inline void balance_push(struct rq *rq)
7728 : {
7729 : }
7730 :
7731 : static inline void balance_push_set(int cpu, bool on)
7732 : {
7733 : }
7734 :
7735 : static inline void balance_hotplug_wait(void)
7736 : {
7737 : }
7738 :
7739 : #endif /* CONFIG_HOTPLUG_CPU */
7740 :
7741 8 : void set_rq_online(struct rq *rq)
7742 : {
7743 8 : if (!rq->online) {
7744 8 : const struct sched_class *class;
7745 :
7746 8 : cpumask_set_cpu(rq->cpu, rq->rd->online);
7747 8 : rq->online = 1;
7748 :
7749 48 : for_each_class(class) {
7750 40 : if (class->rq_online)
7751 24 : class->rq_online(rq);
7752 : }
7753 : }
7754 8 : }
7755 :
7756 4 : void set_rq_offline(struct rq *rq)
7757 : {
7758 4 : if (rq->online) {
7759 : const struct sched_class *class;
7760 :
7761 24 : for_each_class(class) {
7762 20 : if (class->rq_offline)
7763 12 : class->rq_offline(rq);
7764 : }
7765 :
7766 4 : cpumask_clear_cpu(rq->cpu, rq->rd->online);
7767 4 : rq->online = 0;
7768 : }
7769 4 : }
7770 :
7771 : /*
7772 : * used to mark begin/end of suspend/resume:
7773 : */
7774 : static int num_cpus_frozen;
7775 :
7776 : /*
7777 : * Update cpusets according to cpu_active mask. If cpusets are
7778 : * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7779 : * around partition_sched_domains().
7780 : *
7781 : * If we come here as part of a suspend/resume, don't touch cpusets because we
7782 : * want to restore it back to its original state upon resume anyway.
7783 : */
7784 0 : static void cpuset_cpu_active(void)
7785 : {
7786 0 : if (cpuhp_tasks_frozen) {
7787 : /*
7788 : * num_cpus_frozen tracks how many CPUs are involved in suspend
7789 : * resume sequence. As long as this is not the last online
7790 : * operation in the resume sequence, just build a single sched
7791 : * domain, ignoring cpusets.
7792 : */
7793 0 : partition_sched_domains(1, NULL, NULL);
7794 0 : if (--num_cpus_frozen)
7795 : return;
7796 : /*
7797 : * This is the last CPU online operation. So fall through and
7798 : * restore the original sched domains by considering the
7799 : * cpuset configurations.
7800 : */
7801 : cpuset_force_rebuild();
7802 : }
7803 0 : cpuset_update_active_cpus();
7804 : }
7805 :
7806 0 : static int cpuset_cpu_inactive(unsigned int cpu)
7807 : {
7808 0 : if (!cpuhp_tasks_frozen) {
7809 0 : if (dl_cpu_busy(cpu))
7810 : return -EBUSY;
7811 0 : cpuset_update_active_cpus();
7812 : } else {
7813 0 : num_cpus_frozen++;
7814 0 : partition_sched_domains(1, NULL, NULL);
7815 : }
7816 : return 0;
7817 : }
7818 :
7819 3 : int sched_cpu_activate(unsigned int cpu)
7820 : {
7821 3 : struct rq *rq = cpu_rq(cpu);
7822 3 : struct rq_flags rf;
7823 :
7824 : /*
7825 : * Make sure that when the hotplug state machine does a roll-back
7826 : * we clear balance_push. Ideally that would happen earlier...
7827 : */
7828 3 : balance_push_set(cpu, false);
7829 :
7830 : #ifdef CONFIG_SCHED_SMT
7831 : /*
7832 : * When going up, increment the number of cores with SMT present.
7833 : */
7834 3 : if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
7835 0 : static_branch_inc_cpuslocked(&sched_smt_present);
7836 : #endif
7837 3 : set_cpu_active(cpu, true);
7838 :
7839 3 : if (sched_smp_initialized) {
7840 0 : sched_domains_numa_masks_set(cpu);
7841 0 : cpuset_cpu_active();
7842 : }
7843 :
7844 : /*
7845 : * Put the rq online, if not already. This happens:
7846 : *
7847 : * 1) In the early boot process, because we build the real domains
7848 : * after all CPUs have been brought up.
7849 : *
7850 : * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7851 : * domains.
7852 : */
7853 3 : rq_lock_irqsave(rq, &rf);
7854 3 : if (rq->rd) {
7855 3 : BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7856 3 : set_rq_online(rq);
7857 : }
7858 3 : rq_unlock_irqrestore(rq, &rf);
7859 :
7860 3 : return 0;
7861 : }
7862 :
7863 0 : int sched_cpu_deactivate(unsigned int cpu)
7864 : {
7865 0 : struct rq *rq = cpu_rq(cpu);
7866 0 : struct rq_flags rf;
7867 0 : int ret;
7868 :
7869 : /*
7870 : * Remove CPU from nohz.idle_cpus_mask to prevent participating in
7871 : * load balancing when not active
7872 : */
7873 0 : nohz_balance_exit_idle(rq);
7874 :
7875 0 : set_cpu_active(cpu, false);
7876 :
7877 : /*
7878 : * From this point forward, this CPU will refuse to run any task that
7879 : * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
7880 : * push those tasks away until this gets cleared, see
7881 : * sched_cpu_dying().
7882 : */
7883 0 : balance_push_set(cpu, true);
7884 :
7885 : /*
7886 : * We've cleared cpu_active_mask / set balance_push, wait for all
7887 : * preempt-disabled and RCU users of this state to go away such that
7888 : * all new such users will observe it.
7889 : *
7890 : * Specifically, we rely on ttwu to no longer target this CPU, see
7891 : * ttwu_queue_cond() and is_cpu_allowed().
7892 : *
7893 : * Do sync before park smpboot threads to take care the rcu boost case.
7894 : */
7895 0 : synchronize_rcu();
7896 :
7897 0 : rq_lock_irqsave(rq, &rf);
7898 0 : if (rq->rd) {
7899 0 : update_rq_clock(rq);
7900 0 : BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7901 0 : set_rq_offline(rq);
7902 : }
7903 0 : rq_unlock_irqrestore(rq, &rf);
7904 :
7905 : #ifdef CONFIG_SCHED_SMT
7906 : /*
7907 : * When going down, decrement the number of cores with SMT present.
7908 : */
7909 0 : if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
7910 0 : static_branch_dec_cpuslocked(&sched_smt_present);
7911 : #endif
7912 :
7913 0 : if (!sched_smp_initialized)
7914 : return 0;
7915 :
7916 0 : ret = cpuset_cpu_inactive(cpu);
7917 0 : if (ret) {
7918 0 : balance_push_set(cpu, false);
7919 0 : set_cpu_active(cpu, true);
7920 0 : return ret;
7921 : }
7922 0 : sched_domains_numa_masks_clear(cpu);
7923 0 : return 0;
7924 : }
7925 :
7926 4 : static void sched_rq_cpu_starting(unsigned int cpu)
7927 : {
7928 4 : struct rq *rq = cpu_rq(cpu);
7929 :
7930 4 : rq->calc_load_update = calc_load_update;
7931 4 : update_max_interval();
7932 4 : }
7933 :
7934 4 : int sched_cpu_starting(unsigned int cpu)
7935 : {
7936 4 : sched_rq_cpu_starting(cpu);
7937 4 : sched_tick_start(cpu);
7938 4 : return 0;
7939 : }
7940 :
7941 : #ifdef CONFIG_HOTPLUG_CPU
7942 :
7943 : /*
7944 : * Invoked immediately before the stopper thread is invoked to bring the
7945 : * CPU down completely. At this point all per CPU kthreads except the
7946 : * hotplug thread (current) and the stopper thread (inactive) have been
7947 : * either parked or have been unbound from the outgoing CPU. Ensure that
7948 : * any of those which might be on the way out are gone.
7949 : *
7950 : * If after this point a bound task is being woken on this CPU then the
7951 : * responsible hotplug callback has failed to do it's job.
7952 : * sched_cpu_dying() will catch it with the appropriate fireworks.
7953 : */
7954 0 : int sched_cpu_wait_empty(unsigned int cpu)
7955 : {
7956 0 : balance_hotplug_wait();
7957 0 : return 0;
7958 : }
7959 :
7960 : /*
7961 : * Since this CPU is going 'away' for a while, fold any nr_active delta we
7962 : * might have. Called from the CPU stopper task after ensuring that the
7963 : * stopper is the last running task on the CPU, so nr_active count is
7964 : * stable. We need to take the teardown thread which is calling this into
7965 : * account, so we hand in adjust = 1 to the load calculation.
7966 : *
7967 : * Also see the comment "Global load-average calculations".
7968 : */
7969 0 : static void calc_load_migrate(struct rq *rq)
7970 : {
7971 0 : long delta = calc_load_fold_active(rq, 1);
7972 :
7973 0 : if (delta)
7974 0 : atomic_long_add(delta, &calc_load_tasks);
7975 0 : }
7976 :
7977 0 : static void dump_rq_tasks(struct rq *rq, const char *loglvl)
7978 : {
7979 0 : struct task_struct *g, *p;
7980 0 : int cpu = cpu_of(rq);
7981 :
7982 0 : lockdep_assert_held(&rq->lock);
7983 :
7984 0 : printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
7985 0 : for_each_process_thread(g, p) {
7986 0 : if (task_cpu(p) != cpu)
7987 0 : continue;
7988 :
7989 0 : if (!task_on_rq_queued(p))
7990 0 : continue;
7991 :
7992 0 : printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
7993 : }
7994 0 : }
7995 :
7996 0 : int sched_cpu_dying(unsigned int cpu)
7997 : {
7998 0 : struct rq *rq = cpu_rq(cpu);
7999 0 : struct rq_flags rf;
8000 :
8001 : /* Handle pending wakeups and then migrate everything off */
8002 0 : sched_tick_stop(cpu);
8003 :
8004 0 : rq_lock_irqsave(rq, &rf);
8005 0 : if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
8006 0 : WARN(true, "Dying CPU not properly vacated!");
8007 0 : dump_rq_tasks(rq, KERN_WARNING);
8008 : }
8009 0 : rq_unlock_irqrestore(rq, &rf);
8010 :
8011 : /*
8012 : * Now that the CPU is offline, make sure we're welcome
8013 : * to new tasks once we come back up.
8014 : */
8015 0 : balance_push_set(cpu, false);
8016 :
8017 0 : calc_load_migrate(rq);
8018 0 : update_max_interval();
8019 0 : hrtick_clear(rq);
8020 0 : return 0;
8021 : }
8022 : #endif
8023 :
8024 1 : void __init sched_init_smp(void)
8025 : {
8026 1 : sched_init_numa();
8027 :
8028 : /*
8029 : * There's no userspace yet to cause hotplug operations; hence all the
8030 : * CPU masks are stable and all blatant races in the below code cannot
8031 : * happen.
8032 : */
8033 1 : mutex_lock(&sched_domains_mutex);
8034 1 : sched_init_domains(cpu_active_mask);
8035 1 : mutex_unlock(&sched_domains_mutex);
8036 :
8037 : /* Move init over to a non-isolated CPU */
8038 1 : if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
8039 0 : BUG();
8040 1 : sched_init_granularity();
8041 :
8042 1 : init_sched_rt_class();
8043 1 : init_sched_dl_class();
8044 :
8045 1 : sched_smp_initialized = true;
8046 1 : }
8047 :
8048 1 : static int __init migration_init(void)
8049 : {
8050 1 : sched_cpu_starting(smp_processor_id());
8051 1 : return 0;
8052 : }
8053 : early_initcall(migration_init);
8054 :
8055 : #else
8056 : void __init sched_init_smp(void)
8057 : {
8058 : sched_init_granularity();
8059 : }
8060 : #endif /* CONFIG_SMP */
8061 :
8062 117 : int in_sched_functions(unsigned long addr)
8063 : {
8064 117 : return in_lock_functions(addr) ||
8065 117 : (addr >= (unsigned long)__sched_text_start
8066 65 : && addr < (unsigned long)__sched_text_end);
8067 : }
8068 :
8069 : #ifdef CONFIG_CGROUP_SCHED
8070 : /*
8071 : * Default task group.
8072 : * Every task in system belongs to this group at bootup.
8073 : */
8074 : struct task_group root_task_group;
8075 : LIST_HEAD(task_groups);
8076 :
8077 : /* Cacheline aligned slab cache for task_group */
8078 : static struct kmem_cache *task_group_cache __read_mostly;
8079 : #endif
8080 :
8081 : DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
8082 : DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
8083 :
8084 1 : void __init sched_init(void)
8085 : {
8086 1 : unsigned long ptr = 0;
8087 1 : int i;
8088 :
8089 : /* Make sure the linker didn't screw up */
8090 1 : BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
8091 : &fair_sched_class + 1 != &rt_sched_class ||
8092 : &rt_sched_class + 1 != &dl_sched_class);
8093 : #ifdef CONFIG_SMP
8094 1 : BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
8095 : #endif
8096 :
8097 1 : wait_bit_init();
8098 :
8099 : #ifdef CONFIG_FAIR_GROUP_SCHED
8100 : ptr += 2 * nr_cpu_ids * sizeof(void **);
8101 : #endif
8102 : #ifdef CONFIG_RT_GROUP_SCHED
8103 : ptr += 2 * nr_cpu_ids * sizeof(void **);
8104 : #endif
8105 1 : if (ptr) {
8106 : ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
8107 :
8108 : #ifdef CONFIG_FAIR_GROUP_SCHED
8109 : root_task_group.se = (struct sched_entity **)ptr;
8110 : ptr += nr_cpu_ids * sizeof(void **);
8111 :
8112 : root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8113 : ptr += nr_cpu_ids * sizeof(void **);
8114 :
8115 : root_task_group.shares = ROOT_TASK_GROUP_LOAD;
8116 : init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8117 : #endif /* CONFIG_FAIR_GROUP_SCHED */
8118 : #ifdef CONFIG_RT_GROUP_SCHED
8119 : root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8120 : ptr += nr_cpu_ids * sizeof(void **);
8121 :
8122 : root_task_group.rt_rq = (struct rt_rq **)ptr;
8123 : ptr += nr_cpu_ids * sizeof(void **);
8124 :
8125 : #endif /* CONFIG_RT_GROUP_SCHED */
8126 : }
8127 : #ifdef CONFIG_CPUMASK_OFFSTACK
8128 : for_each_possible_cpu(i) {
8129 : per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
8130 : cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8131 : per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
8132 : cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8133 : }
8134 : #endif /* CONFIG_CPUMASK_OFFSTACK */
8135 :
8136 2 : init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
8137 2 : init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
8138 :
8139 : #ifdef CONFIG_SMP
8140 1 : init_defrootdomain();
8141 : #endif
8142 :
8143 : #ifdef CONFIG_RT_GROUP_SCHED
8144 : init_rt_bandwidth(&root_task_group.rt_bandwidth,
8145 : global_rt_period(), global_rt_runtime());
8146 : #endif /* CONFIG_RT_GROUP_SCHED */
8147 :
8148 : #ifdef CONFIG_CGROUP_SCHED
8149 : task_group_cache = KMEM_CACHE(task_group, 0);
8150 :
8151 : list_add(&root_task_group.list, &task_groups);
8152 : INIT_LIST_HEAD(&root_task_group.children);
8153 : INIT_LIST_HEAD(&root_task_group.siblings);
8154 : autogroup_init(&init_task);
8155 : #endif /* CONFIG_CGROUP_SCHED */
8156 :
8157 6 : for_each_possible_cpu(i) {
8158 4 : struct rq *rq;
8159 :
8160 4 : rq = cpu_rq(i);
8161 4 : raw_spin_lock_init(&rq->lock);
8162 4 : rq->nr_running = 0;
8163 4 : rq->calc_load_active = 0;
8164 4 : rq->calc_load_update = jiffies + LOAD_FREQ;
8165 4 : init_cfs_rq(&rq->cfs);
8166 4 : init_rt_rq(&rq->rt);
8167 4 : init_dl_rq(&rq->dl);
8168 : #ifdef CONFIG_FAIR_GROUP_SCHED
8169 : INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8170 : rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
8171 : /*
8172 : * How much CPU bandwidth does root_task_group get?
8173 : *
8174 : * In case of task-groups formed thr' the cgroup filesystem, it
8175 : * gets 100% of the CPU resources in the system. This overall
8176 : * system CPU resource is divided among the tasks of
8177 : * root_task_group and its child task-groups in a fair manner,
8178 : * based on each entity's (task or task-group's) weight
8179 : * (se->load.weight).
8180 : *
8181 : * In other words, if root_task_group has 10 tasks of weight
8182 : * 1024) and two child groups A0 and A1 (of weight 1024 each),
8183 : * then A0's share of the CPU resource is:
8184 : *
8185 : * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8186 : *
8187 : * We achieve this by letting root_task_group's tasks sit
8188 : * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8189 : */
8190 : init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8191 : #endif /* CONFIG_FAIR_GROUP_SCHED */
8192 :
8193 4 : rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8194 : #ifdef CONFIG_RT_GROUP_SCHED
8195 : init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8196 : #endif
8197 : #ifdef CONFIG_SMP
8198 4 : rq->sd = NULL;
8199 4 : rq->rd = NULL;
8200 4 : rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
8201 4 : rq->balance_callback = NULL;
8202 4 : rq->active_balance = 0;
8203 4 : rq->next_balance = jiffies;
8204 4 : rq->push_cpu = 0;
8205 4 : rq->cpu = i;
8206 4 : rq->online = 0;
8207 4 : rq->idle_stamp = 0;
8208 4 : rq->avg_idle = 2*sysctl_sched_migration_cost;
8209 4 : rq->max_idle_balance_cost = sysctl_sched_migration_cost;
8210 :
8211 4 : INIT_LIST_HEAD(&rq->cfs_tasks);
8212 :
8213 4 : rq_attach_root(rq, &def_root_domain);
8214 : #ifdef CONFIG_NO_HZ_COMMON
8215 4 : rq->last_blocked_load_update_tick = jiffies;
8216 4 : atomic_set(&rq->nohz_flags, 0);
8217 :
8218 4 : INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
8219 : #endif
8220 : #ifdef CONFIG_HOTPLUG_CPU
8221 4 : rcuwait_init(&rq->hotplug_wait);
8222 : #endif
8223 : #endif /* CONFIG_SMP */
8224 4 : hrtick_rq_init(rq);
8225 9 : atomic_set(&rq->nr_iowait, 0);
8226 : }
8227 :
8228 1 : set_load_weight(&init_task, false);
8229 :
8230 : /*
8231 : * The boot idle thread does lazy MMU switching as well:
8232 : */
8233 1 : mmgrab(&init_mm);
8234 1 : enter_lazy_tlb(&init_mm, current);
8235 :
8236 : /*
8237 : * Make us the idle thread. Technically, schedule() should not be
8238 : * called from this thread, however somewhere below it might be,
8239 : * but because we are the idle thread, we just pick up running again
8240 : * when this runqueue becomes "idle".
8241 : */
8242 1 : init_idle(current, smp_processor_id());
8243 :
8244 1 : calc_load_update = jiffies + LOAD_FREQ;
8245 :
8246 : #ifdef CONFIG_SMP
8247 1 : idle_thread_set_boot_cpu();
8248 : #endif
8249 1 : init_sched_fair_class();
8250 :
8251 1 : init_schedstats();
8252 :
8253 1 : psi_init();
8254 :
8255 1 : init_uclamp();
8256 :
8257 1 : scheduler_running = 1;
8258 1 : }
8259 :
8260 : #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8261 3128469 : static inline int preempt_count_equals(int preempt_offset)
8262 : {
8263 3128469 : int nested = preempt_count() + rcu_preempt_depth();
8264 :
8265 3128469 : return (nested == preempt_offset);
8266 : }
8267 :
8268 2639150 : void __might_sleep(const char *file, int line, int preempt_offset)
8269 : {
8270 : /*
8271 : * Blocking primitives will set (and therefore destroy) current->state,
8272 : * since we will exit with TASK_RUNNING make sure we enter with it,
8273 : * otherwise we will destroy state.
8274 : */
8275 2639644 : WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
8276 : "do not call blocking ops when !TASK_RUNNING; "
8277 : "state=%lx set at [<%p>] %pS\n",
8278 : current->state,
8279 : (void *)current->task_state_change,
8280 : (void *)current->task_state_change);
8281 :
8282 2639150 : ___might_sleep(file, line, preempt_offset);
8283 2641758 : }
8284 : EXPORT_SYMBOL(__might_sleep);
8285 :
8286 3125595 : void ___might_sleep(const char *file, int line, int preempt_offset)
8287 : {
8288 : /* Ratelimiting timestamp: */
8289 3125595 : static unsigned long prev_jiffy;
8290 :
8291 3125595 : unsigned long preempt_disable_ip;
8292 :
8293 : /* WARN_ON_ONCE() by default, no rate limit required: */
8294 9379512 : rcu_sleep_check();
8295 :
8296 3128469 : if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
8297 3126608 : !is_idle_task(current) && !current->non_block_count) ||
8298 1763 : system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
8299 : oops_in_progress)
8300 3128938 : return;
8301 :
8302 0 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8303 : return;
8304 0 : prev_jiffy = jiffies;
8305 :
8306 : /* Save this before calling printk(), since that will clobber it: */
8307 0 : preempt_disable_ip = get_preempt_disable_ip(current);
8308 :
8309 0 : printk(KERN_ERR
8310 : "BUG: sleeping function called from invalid context at %s:%d\n",
8311 : file, line);
8312 0 : printk(KERN_ERR
8313 : "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
8314 0 : in_atomic(), irqs_disabled(), current->non_block_count,
8315 0 : current->pid, current->comm);
8316 :
8317 0 : if (task_stack_end_corrupted(current))
8318 0 : printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
8319 :
8320 0 : debug_show_held_locks(current);
8321 0 : if (irqs_disabled())
8322 0 : print_irqtrace_events(current);
8323 0 : if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
8324 : && !preempt_count_equals(preempt_offset)) {
8325 : pr_err("Preemption disabled at:");
8326 : print_ip_sym(KERN_ERR, preempt_disable_ip);
8327 : }
8328 0 : dump_stack();
8329 0 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
8330 : }
8331 : EXPORT_SYMBOL(___might_sleep);
8332 :
8333 0 : void __cant_sleep(const char *file, int line, int preempt_offset)
8334 : {
8335 0 : static unsigned long prev_jiffy;
8336 :
8337 0 : if (irqs_disabled())
8338 : return;
8339 :
8340 0 : if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
8341 : return;
8342 :
8343 0 : if (preempt_count() > preempt_offset)
8344 : return;
8345 :
8346 0 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8347 : return;
8348 0 : prev_jiffy = jiffies;
8349 :
8350 0 : printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
8351 0 : printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8352 0 : in_atomic(), irqs_disabled(),
8353 0 : current->pid, current->comm);
8354 :
8355 0 : debug_show_held_locks(current);
8356 0 : dump_stack();
8357 0 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
8358 : }
8359 : EXPORT_SYMBOL_GPL(__cant_sleep);
8360 :
8361 : #ifdef CONFIG_SMP
8362 1068 : void __cant_migrate(const char *file, int line)
8363 : {
8364 1068 : static unsigned long prev_jiffy;
8365 :
8366 1068 : if (irqs_disabled())
8367 : return;
8368 :
8369 1068 : if (is_migration_disabled(current))
8370 : return;
8371 :
8372 0 : if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
8373 : return;
8374 :
8375 0 : if (preempt_count() > 0)
8376 : return;
8377 :
8378 0 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8379 : return;
8380 0 : prev_jiffy = jiffies;
8381 :
8382 0 : pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
8383 0 : pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
8384 : in_atomic(), irqs_disabled(), is_migration_disabled(current),
8385 : current->pid, current->comm);
8386 :
8387 0 : debug_show_held_locks(current);
8388 0 : dump_stack();
8389 0 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
8390 : }
8391 : EXPORT_SYMBOL_GPL(__cant_migrate);
8392 : #endif
8393 : #endif
8394 :
8395 : #ifdef CONFIG_MAGIC_SYSRQ
8396 0 : void normalize_rt_tasks(void)
8397 : {
8398 0 : struct task_struct *g, *p;
8399 0 : struct sched_attr attr = {
8400 : .sched_policy = SCHED_NORMAL,
8401 : };
8402 :
8403 0 : read_lock(&tasklist_lock);
8404 0 : for_each_process_thread(g, p) {
8405 : /*
8406 : * Only normalize user tasks:
8407 : */
8408 0 : if (p->flags & PF_KTHREAD)
8409 0 : continue;
8410 :
8411 0 : p->se.exec_start = 0;
8412 0 : schedstat_set(p->se.statistics.wait_start, 0);
8413 0 : schedstat_set(p->se.statistics.sleep_start, 0);
8414 0 : schedstat_set(p->se.statistics.block_start, 0);
8415 :
8416 0 : if (!dl_task(p) && !rt_task(p)) {
8417 : /*
8418 : * Renice negative nice level userspace
8419 : * tasks back to 0:
8420 : */
8421 0 : if (task_nice(p) < 0)
8422 0 : set_user_nice(p, 0);
8423 0 : continue;
8424 : }
8425 :
8426 0 : __sched_setscheduler(p, &attr, false, false);
8427 : }
8428 0 : read_unlock(&tasklist_lock);
8429 0 : }
8430 :
8431 : #endif /* CONFIG_MAGIC_SYSRQ */
8432 :
8433 : #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8434 : /*
8435 : * These functions are only useful for the IA64 MCA handling, or kdb.
8436 : *
8437 : * They can only be called when the whole system has been
8438 : * stopped - every CPU needs to be quiescent, and no scheduling
8439 : * activity can take place. Using them for anything else would
8440 : * be a serious bug, and as a result, they aren't even visible
8441 : * under any other configuration.
8442 : */
8443 :
8444 : /**
8445 : * curr_task - return the current task for a given CPU.
8446 : * @cpu: the processor in question.
8447 : *
8448 : * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8449 : *
8450 : * Return: The current task for @cpu.
8451 : */
8452 : struct task_struct *curr_task(int cpu)
8453 : {
8454 : return cpu_curr(cpu);
8455 : }
8456 :
8457 : #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8458 :
8459 : #ifdef CONFIG_IA64
8460 : /**
8461 : * ia64_set_curr_task - set the current task for a given CPU.
8462 : * @cpu: the processor in question.
8463 : * @p: the task pointer to set.
8464 : *
8465 : * Description: This function must only be used when non-maskable interrupts
8466 : * are serviced on a separate stack. It allows the architecture to switch the
8467 : * notion of the current task on a CPU in a non-blocking manner. This function
8468 : * must be called with all CPU's synchronized, and interrupts disabled, the
8469 : * and caller must save the original value of the current task (see
8470 : * curr_task() above) and restore that value before reenabling interrupts and
8471 : * re-starting the system.
8472 : *
8473 : * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8474 : */
8475 : void ia64_set_curr_task(int cpu, struct task_struct *p)
8476 : {
8477 : cpu_curr(cpu) = p;
8478 : }
8479 :
8480 : #endif
8481 :
8482 : #ifdef CONFIG_CGROUP_SCHED
8483 : /* task_group_lock serializes the addition/removal of task groups */
8484 : static DEFINE_SPINLOCK(task_group_lock);
8485 :
8486 : static inline void alloc_uclamp_sched_group(struct task_group *tg,
8487 : struct task_group *parent)
8488 : {
8489 : #ifdef CONFIG_UCLAMP_TASK_GROUP
8490 : enum uclamp_id clamp_id;
8491 :
8492 : for_each_clamp_id(clamp_id) {
8493 : uclamp_se_set(&tg->uclamp_req[clamp_id],
8494 : uclamp_none(clamp_id), false);
8495 : tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
8496 : }
8497 : #endif
8498 : }
8499 :
8500 : static void sched_free_group(struct task_group *tg)
8501 : {
8502 : free_fair_sched_group(tg);
8503 : free_rt_sched_group(tg);
8504 : autogroup_free(tg);
8505 : kmem_cache_free(task_group_cache, tg);
8506 : }
8507 :
8508 : /* allocate runqueue etc for a new task group */
8509 : struct task_group *sched_create_group(struct task_group *parent)
8510 : {
8511 : struct task_group *tg;
8512 :
8513 : tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
8514 : if (!tg)
8515 : return ERR_PTR(-ENOMEM);
8516 :
8517 : if (!alloc_fair_sched_group(tg, parent))
8518 : goto err;
8519 :
8520 : if (!alloc_rt_sched_group(tg, parent))
8521 : goto err;
8522 :
8523 : alloc_uclamp_sched_group(tg, parent);
8524 :
8525 : return tg;
8526 :
8527 : err:
8528 : sched_free_group(tg);
8529 : return ERR_PTR(-ENOMEM);
8530 : }
8531 :
8532 : void sched_online_group(struct task_group *tg, struct task_group *parent)
8533 : {
8534 : unsigned long flags;
8535 :
8536 : spin_lock_irqsave(&task_group_lock, flags);
8537 : list_add_rcu(&tg->list, &task_groups);
8538 :
8539 : /* Root should already exist: */
8540 : WARN_ON(!parent);
8541 :
8542 : tg->parent = parent;
8543 : INIT_LIST_HEAD(&tg->children);
8544 : list_add_rcu(&tg->siblings, &parent->children);
8545 : spin_unlock_irqrestore(&task_group_lock, flags);
8546 :
8547 : online_fair_sched_group(tg);
8548 : }
8549 :
8550 : /* rcu callback to free various structures associated with a task group */
8551 : static void sched_free_group_rcu(struct rcu_head *rhp)
8552 : {
8553 : /* Now it should be safe to free those cfs_rqs: */
8554 : sched_free_group(container_of(rhp, struct task_group, rcu));
8555 : }
8556 :
8557 : void sched_destroy_group(struct task_group *tg)
8558 : {
8559 : /* Wait for possible concurrent references to cfs_rqs complete: */
8560 : call_rcu(&tg->rcu, sched_free_group_rcu);
8561 : }
8562 :
8563 : void sched_offline_group(struct task_group *tg)
8564 : {
8565 : unsigned long flags;
8566 :
8567 : /* End participation in shares distribution: */
8568 : unregister_fair_sched_group(tg);
8569 :
8570 : spin_lock_irqsave(&task_group_lock, flags);
8571 : list_del_rcu(&tg->list);
8572 : list_del_rcu(&tg->siblings);
8573 : spin_unlock_irqrestore(&task_group_lock, flags);
8574 : }
8575 :
8576 : static void sched_change_group(struct task_struct *tsk, int type)
8577 : {
8578 : struct task_group *tg;
8579 :
8580 : /*
8581 : * All callers are synchronized by task_rq_lock(); we do not use RCU
8582 : * which is pointless here. Thus, we pass "true" to task_css_check()
8583 : * to prevent lockdep warnings.
8584 : */
8585 : tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8586 : struct task_group, css);
8587 : tg = autogroup_task_group(tsk, tg);
8588 : tsk->sched_task_group = tg;
8589 :
8590 : #ifdef CONFIG_FAIR_GROUP_SCHED
8591 : if (tsk->sched_class->task_change_group)
8592 : tsk->sched_class->task_change_group(tsk, type);
8593 : else
8594 : #endif
8595 : set_task_rq(tsk, task_cpu(tsk));
8596 : }
8597 :
8598 : /*
8599 : * Change task's runqueue when it moves between groups.
8600 : *
8601 : * The caller of this function should have put the task in its new group by
8602 : * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8603 : * its new group.
8604 : */
8605 : void sched_move_task(struct task_struct *tsk)
8606 : {
8607 : int queued, running, queue_flags =
8608 : DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
8609 : struct rq_flags rf;
8610 : struct rq *rq;
8611 :
8612 : rq = task_rq_lock(tsk, &rf);
8613 : update_rq_clock(rq);
8614 :
8615 : running = task_current(rq, tsk);
8616 : queued = task_on_rq_queued(tsk);
8617 :
8618 : if (queued)
8619 : dequeue_task(rq, tsk, queue_flags);
8620 : if (running)
8621 : put_prev_task(rq, tsk);
8622 :
8623 : sched_change_group(tsk, TASK_MOVE_GROUP);
8624 :
8625 : if (queued)
8626 : enqueue_task(rq, tsk, queue_flags);
8627 : if (running) {
8628 : set_next_task(rq, tsk);
8629 : /*
8630 : * After changing group, the running task may have joined a
8631 : * throttled one but it's still the running task. Trigger a
8632 : * resched to make sure that task can still run.
8633 : */
8634 : resched_curr(rq);
8635 : }
8636 :
8637 : task_rq_unlock(rq, tsk, &rf);
8638 : }
8639 :
8640 : static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8641 : {
8642 : return css ? container_of(css, struct task_group, css) : NULL;
8643 : }
8644 :
8645 : static struct cgroup_subsys_state *
8646 : cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8647 : {
8648 : struct task_group *parent = css_tg(parent_css);
8649 : struct task_group *tg;
8650 :
8651 : if (!parent) {
8652 : /* This is early initialization for the top cgroup */
8653 : return &root_task_group.css;
8654 : }
8655 :
8656 : tg = sched_create_group(parent);
8657 : if (IS_ERR(tg))
8658 : return ERR_PTR(-ENOMEM);
8659 :
8660 : return &tg->css;
8661 : }
8662 :
8663 : /* Expose task group only after completing cgroup initialization */
8664 : static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8665 : {
8666 : struct task_group *tg = css_tg(css);
8667 : struct task_group *parent = css_tg(css->parent);
8668 :
8669 : if (parent)
8670 : sched_online_group(tg, parent);
8671 :
8672 : #ifdef CONFIG_UCLAMP_TASK_GROUP
8673 : /* Propagate the effective uclamp value for the new group */
8674 : cpu_util_update_eff(css);
8675 : #endif
8676 :
8677 : return 0;
8678 : }
8679 :
8680 : static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8681 : {
8682 : struct task_group *tg = css_tg(css);
8683 :
8684 : sched_offline_group(tg);
8685 : }
8686 :
8687 : static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8688 : {
8689 : struct task_group *tg = css_tg(css);
8690 :
8691 : /*
8692 : * Relies on the RCU grace period between css_released() and this.
8693 : */
8694 : sched_free_group(tg);
8695 : }
8696 :
8697 : /*
8698 : * This is called before wake_up_new_task(), therefore we really only
8699 : * have to set its group bits, all the other stuff does not apply.
8700 : */
8701 : static void cpu_cgroup_fork(struct task_struct *task)
8702 : {
8703 : struct rq_flags rf;
8704 : struct rq *rq;
8705 :
8706 : rq = task_rq_lock(task, &rf);
8707 :
8708 : update_rq_clock(rq);
8709 : sched_change_group(task, TASK_SET_GROUP);
8710 :
8711 : task_rq_unlock(rq, task, &rf);
8712 : }
8713 :
8714 : static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8715 : {
8716 : struct task_struct *task;
8717 : struct cgroup_subsys_state *css;
8718 : int ret = 0;
8719 :
8720 : cgroup_taskset_for_each(task, css, tset) {
8721 : #ifdef CONFIG_RT_GROUP_SCHED
8722 : if (!sched_rt_can_attach(css_tg(css), task))
8723 : return -EINVAL;
8724 : #endif
8725 : /*
8726 : * Serialize against wake_up_new_task() such that if it's
8727 : * running, we're sure to observe its full state.
8728 : */
8729 : raw_spin_lock_irq(&task->pi_lock);
8730 : /*
8731 : * Avoid calling sched_move_task() before wake_up_new_task()
8732 : * has happened. This would lead to problems with PELT, due to
8733 : * move wanting to detach+attach while we're not attached yet.
8734 : */
8735 : if (task->state == TASK_NEW)
8736 : ret = -EINVAL;
8737 : raw_spin_unlock_irq(&task->pi_lock);
8738 :
8739 : if (ret)
8740 : break;
8741 : }
8742 : return ret;
8743 : }
8744 :
8745 : static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8746 : {
8747 : struct task_struct *task;
8748 : struct cgroup_subsys_state *css;
8749 :
8750 : cgroup_taskset_for_each(task, css, tset)
8751 : sched_move_task(task);
8752 : }
8753 :
8754 : #ifdef CONFIG_UCLAMP_TASK_GROUP
8755 : static void cpu_util_update_eff(struct cgroup_subsys_state *css)
8756 : {
8757 : struct cgroup_subsys_state *top_css = css;
8758 : struct uclamp_se *uc_parent = NULL;
8759 : struct uclamp_se *uc_se = NULL;
8760 : unsigned int eff[UCLAMP_CNT];
8761 : enum uclamp_id clamp_id;
8762 : unsigned int clamps;
8763 :
8764 : css_for_each_descendant_pre(css, top_css) {
8765 : uc_parent = css_tg(css)->parent
8766 : ? css_tg(css)->parent->uclamp : NULL;
8767 :
8768 : for_each_clamp_id(clamp_id) {
8769 : /* Assume effective clamps matches requested clamps */
8770 : eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
8771 : /* Cap effective clamps with parent's effective clamps */
8772 : if (uc_parent &&
8773 : eff[clamp_id] > uc_parent[clamp_id].value) {
8774 : eff[clamp_id] = uc_parent[clamp_id].value;
8775 : }
8776 : }
8777 : /* Ensure protection is always capped by limit */
8778 : eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
8779 :
8780 : /* Propagate most restrictive effective clamps */
8781 : clamps = 0x0;
8782 : uc_se = css_tg(css)->uclamp;
8783 : for_each_clamp_id(clamp_id) {
8784 : if (eff[clamp_id] == uc_se[clamp_id].value)
8785 : continue;
8786 : uc_se[clamp_id].value = eff[clamp_id];
8787 : uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
8788 : clamps |= (0x1 << clamp_id);
8789 : }
8790 : if (!clamps) {
8791 : css = css_rightmost_descendant(css);
8792 : continue;
8793 : }
8794 :
8795 : /* Immediately update descendants RUNNABLE tasks */
8796 : uclamp_update_active_tasks(css, clamps);
8797 : }
8798 : }
8799 :
8800 : /*
8801 : * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
8802 : * C expression. Since there is no way to convert a macro argument (N) into a
8803 : * character constant, use two levels of macros.
8804 : */
8805 : #define _POW10(exp) ((unsigned int)1e##exp)
8806 : #define POW10(exp) _POW10(exp)
8807 :
8808 : struct uclamp_request {
8809 : #define UCLAMP_PERCENT_SHIFT 2
8810 : #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
8811 : s64 percent;
8812 : u64 util;
8813 : int ret;
8814 : };
8815 :
8816 : static inline struct uclamp_request
8817 : capacity_from_percent(char *buf)
8818 : {
8819 : struct uclamp_request req = {
8820 : .percent = UCLAMP_PERCENT_SCALE,
8821 : .util = SCHED_CAPACITY_SCALE,
8822 : .ret = 0,
8823 : };
8824 :
8825 : buf = strim(buf);
8826 : if (strcmp(buf, "max")) {
8827 : req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
8828 : &req.percent);
8829 : if (req.ret)
8830 : return req;
8831 : if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
8832 : req.ret = -ERANGE;
8833 : return req;
8834 : }
8835 :
8836 : req.util = req.percent << SCHED_CAPACITY_SHIFT;
8837 : req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
8838 : }
8839 :
8840 : return req;
8841 : }
8842 :
8843 : static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
8844 : size_t nbytes, loff_t off,
8845 : enum uclamp_id clamp_id)
8846 : {
8847 : struct uclamp_request req;
8848 : struct task_group *tg;
8849 :
8850 : req = capacity_from_percent(buf);
8851 : if (req.ret)
8852 : return req.ret;
8853 :
8854 : static_branch_enable(&sched_uclamp_used);
8855 :
8856 : mutex_lock(&uclamp_mutex);
8857 : rcu_read_lock();
8858 :
8859 : tg = css_tg(of_css(of));
8860 : if (tg->uclamp_req[clamp_id].value != req.util)
8861 : uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
8862 :
8863 : /*
8864 : * Because of not recoverable conversion rounding we keep track of the
8865 : * exact requested value
8866 : */
8867 : tg->uclamp_pct[clamp_id] = req.percent;
8868 :
8869 : /* Update effective clamps to track the most restrictive value */
8870 : cpu_util_update_eff(of_css(of));
8871 :
8872 : rcu_read_unlock();
8873 : mutex_unlock(&uclamp_mutex);
8874 :
8875 : return nbytes;
8876 : }
8877 :
8878 : static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
8879 : char *buf, size_t nbytes,
8880 : loff_t off)
8881 : {
8882 : return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
8883 : }
8884 :
8885 : static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
8886 : char *buf, size_t nbytes,
8887 : loff_t off)
8888 : {
8889 : return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
8890 : }
8891 :
8892 : static inline void cpu_uclamp_print(struct seq_file *sf,
8893 : enum uclamp_id clamp_id)
8894 : {
8895 : struct task_group *tg;
8896 : u64 util_clamp;
8897 : u64 percent;
8898 : u32 rem;
8899 :
8900 : rcu_read_lock();
8901 : tg = css_tg(seq_css(sf));
8902 : util_clamp = tg->uclamp_req[clamp_id].value;
8903 : rcu_read_unlock();
8904 :
8905 : if (util_clamp == SCHED_CAPACITY_SCALE) {
8906 : seq_puts(sf, "max\n");
8907 : return;
8908 : }
8909 :
8910 : percent = tg->uclamp_pct[clamp_id];
8911 : percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
8912 : seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
8913 : }
8914 :
8915 : static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
8916 : {
8917 : cpu_uclamp_print(sf, UCLAMP_MIN);
8918 : return 0;
8919 : }
8920 :
8921 : static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
8922 : {
8923 : cpu_uclamp_print(sf, UCLAMP_MAX);
8924 : return 0;
8925 : }
8926 : #endif /* CONFIG_UCLAMP_TASK_GROUP */
8927 :
8928 : #ifdef CONFIG_FAIR_GROUP_SCHED
8929 : static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8930 : struct cftype *cftype, u64 shareval)
8931 : {
8932 : if (shareval > scale_load_down(ULONG_MAX))
8933 : shareval = MAX_SHARES;
8934 : return sched_group_set_shares(css_tg(css), scale_load(shareval));
8935 : }
8936 :
8937 : static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8938 : struct cftype *cft)
8939 : {
8940 : struct task_group *tg = css_tg(css);
8941 :
8942 : return (u64) scale_load_down(tg->shares);
8943 : }
8944 :
8945 : #ifdef CONFIG_CFS_BANDWIDTH
8946 : static DEFINE_MUTEX(cfs_constraints_mutex);
8947 :
8948 : const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8949 : static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8950 : /* More than 203 days if BW_SHIFT equals 20. */
8951 : static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
8952 :
8953 : static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8954 :
8955 : static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8956 : {
8957 : int i, ret = 0, runtime_enabled, runtime_was_enabled;
8958 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8959 :
8960 : if (tg == &root_task_group)
8961 : return -EINVAL;
8962 :
8963 : /*
8964 : * Ensure we have at some amount of bandwidth every period. This is
8965 : * to prevent reaching a state of large arrears when throttled via
8966 : * entity_tick() resulting in prolonged exit starvation.
8967 : */
8968 : if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8969 : return -EINVAL;
8970 :
8971 : /*
8972 : * Likewise, bound things on the otherside by preventing insane quota
8973 : * periods. This also allows us to normalize in computing quota
8974 : * feasibility.
8975 : */
8976 : if (period > max_cfs_quota_period)
8977 : return -EINVAL;
8978 :
8979 : /*
8980 : * Bound quota to defend quota against overflow during bandwidth shift.
8981 : */
8982 : if (quota != RUNTIME_INF && quota > max_cfs_runtime)
8983 : return -EINVAL;
8984 :
8985 : /*
8986 : * Prevent race between setting of cfs_rq->runtime_enabled and
8987 : * unthrottle_offline_cfs_rqs().
8988 : */
8989 : get_online_cpus();
8990 : mutex_lock(&cfs_constraints_mutex);
8991 : ret = __cfs_schedulable(tg, period, quota);
8992 : if (ret)
8993 : goto out_unlock;
8994 :
8995 : runtime_enabled = quota != RUNTIME_INF;
8996 : runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8997 : /*
8998 : * If we need to toggle cfs_bandwidth_used, off->on must occur
8999 : * before making related changes, and on->off must occur afterwards
9000 : */
9001 : if (runtime_enabled && !runtime_was_enabled)
9002 : cfs_bandwidth_usage_inc();
9003 : raw_spin_lock_irq(&cfs_b->lock);
9004 : cfs_b->period = ns_to_ktime(period);
9005 : cfs_b->quota = quota;
9006 :
9007 : __refill_cfs_bandwidth_runtime(cfs_b);
9008 :
9009 : /* Restart the period timer (if active) to handle new period expiry: */
9010 : if (runtime_enabled)
9011 : start_cfs_bandwidth(cfs_b);
9012 :
9013 : raw_spin_unlock_irq(&cfs_b->lock);
9014 :
9015 : for_each_online_cpu(i) {
9016 : struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9017 : struct rq *rq = cfs_rq->rq;
9018 : struct rq_flags rf;
9019 :
9020 : rq_lock_irq(rq, &rf);
9021 : cfs_rq->runtime_enabled = runtime_enabled;
9022 : cfs_rq->runtime_remaining = 0;
9023 :
9024 : if (cfs_rq->throttled)
9025 : unthrottle_cfs_rq(cfs_rq);
9026 : rq_unlock_irq(rq, &rf);
9027 : }
9028 : if (runtime_was_enabled && !runtime_enabled)
9029 : cfs_bandwidth_usage_dec();
9030 : out_unlock:
9031 : mutex_unlock(&cfs_constraints_mutex);
9032 : put_online_cpus();
9033 :
9034 : return ret;
9035 : }
9036 :
9037 : static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9038 : {
9039 : u64 quota, period;
9040 :
9041 : period = ktime_to_ns(tg->cfs_bandwidth.period);
9042 : if (cfs_quota_us < 0)
9043 : quota = RUNTIME_INF;
9044 : else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
9045 : quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9046 : else
9047 : return -EINVAL;
9048 :
9049 : return tg_set_cfs_bandwidth(tg, period, quota);
9050 : }
9051 :
9052 : static long tg_get_cfs_quota(struct task_group *tg)
9053 : {
9054 : u64 quota_us;
9055 :
9056 : if (tg->cfs_bandwidth.quota == RUNTIME_INF)
9057 : return -1;
9058 :
9059 : quota_us = tg->cfs_bandwidth.quota;
9060 : do_div(quota_us, NSEC_PER_USEC);
9061 :
9062 : return quota_us;
9063 : }
9064 :
9065 : static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9066 : {
9067 : u64 quota, period;
9068 :
9069 : if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
9070 : return -EINVAL;
9071 :
9072 : period = (u64)cfs_period_us * NSEC_PER_USEC;
9073 : quota = tg->cfs_bandwidth.quota;
9074 :
9075 : return tg_set_cfs_bandwidth(tg, period, quota);
9076 : }
9077 :
9078 : static long tg_get_cfs_period(struct task_group *tg)
9079 : {
9080 : u64 cfs_period_us;
9081 :
9082 : cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
9083 : do_div(cfs_period_us, NSEC_PER_USEC);
9084 :
9085 : return cfs_period_us;
9086 : }
9087 :
9088 : static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
9089 : struct cftype *cft)
9090 : {
9091 : return tg_get_cfs_quota(css_tg(css));
9092 : }
9093 :
9094 : static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
9095 : struct cftype *cftype, s64 cfs_quota_us)
9096 : {
9097 : return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
9098 : }
9099 :
9100 : static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
9101 : struct cftype *cft)
9102 : {
9103 : return tg_get_cfs_period(css_tg(css));
9104 : }
9105 :
9106 : static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
9107 : struct cftype *cftype, u64 cfs_period_us)
9108 : {
9109 : return tg_set_cfs_period(css_tg(css), cfs_period_us);
9110 : }
9111 :
9112 : struct cfs_schedulable_data {
9113 : struct task_group *tg;
9114 : u64 period, quota;
9115 : };
9116 :
9117 : /*
9118 : * normalize group quota/period to be quota/max_period
9119 : * note: units are usecs
9120 : */
9121 : static u64 normalize_cfs_quota(struct task_group *tg,
9122 : struct cfs_schedulable_data *d)
9123 : {
9124 : u64 quota, period;
9125 :
9126 : if (tg == d->tg) {
9127 : period = d->period;
9128 : quota = d->quota;
9129 : } else {
9130 : period = tg_get_cfs_period(tg);
9131 : quota = tg_get_cfs_quota(tg);
9132 : }
9133 :
9134 : /* note: these should typically be equivalent */
9135 : if (quota == RUNTIME_INF || quota == -1)
9136 : return RUNTIME_INF;
9137 :
9138 : return to_ratio(period, quota);
9139 : }
9140 :
9141 : static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9142 : {
9143 : struct cfs_schedulable_data *d = data;
9144 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9145 : s64 quota = 0, parent_quota = -1;
9146 :
9147 : if (!tg->parent) {
9148 : quota = RUNTIME_INF;
9149 : } else {
9150 : struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
9151 :
9152 : quota = normalize_cfs_quota(tg, d);
9153 : parent_quota = parent_b->hierarchical_quota;
9154 :
9155 : /*
9156 : * Ensure max(child_quota) <= parent_quota. On cgroup2,
9157 : * always take the min. On cgroup1, only inherit when no
9158 : * limit is set:
9159 : */
9160 : if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
9161 : quota = min(quota, parent_quota);
9162 : } else {
9163 : if (quota == RUNTIME_INF)
9164 : quota = parent_quota;
9165 : else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9166 : return -EINVAL;
9167 : }
9168 : }
9169 : cfs_b->hierarchical_quota = quota;
9170 :
9171 : return 0;
9172 : }
9173 :
9174 : static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9175 : {
9176 : int ret;
9177 : struct cfs_schedulable_data data = {
9178 : .tg = tg,
9179 : .period = period,
9180 : .quota = quota,
9181 : };
9182 :
9183 : if (quota != RUNTIME_INF) {
9184 : do_div(data.period, NSEC_PER_USEC);
9185 : do_div(data.quota, NSEC_PER_USEC);
9186 : }
9187 :
9188 : rcu_read_lock();
9189 : ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9190 : rcu_read_unlock();
9191 :
9192 : return ret;
9193 : }
9194 :
9195 : static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
9196 : {
9197 : struct task_group *tg = css_tg(seq_css(sf));
9198 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9199 :
9200 : seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
9201 : seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
9202 : seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
9203 :
9204 : if (schedstat_enabled() && tg != &root_task_group) {
9205 : u64 ws = 0;
9206 : int i;
9207 :
9208 : for_each_possible_cpu(i)
9209 : ws += schedstat_val(tg->se[i]->statistics.wait_sum);
9210 :
9211 : seq_printf(sf, "wait_sum %llu\n", ws);
9212 : }
9213 :
9214 : return 0;
9215 : }
9216 : #endif /* CONFIG_CFS_BANDWIDTH */
9217 : #endif /* CONFIG_FAIR_GROUP_SCHED */
9218 :
9219 : #ifdef CONFIG_RT_GROUP_SCHED
9220 : static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
9221 : struct cftype *cft, s64 val)
9222 : {
9223 : return sched_group_set_rt_runtime(css_tg(css), val);
9224 : }
9225 :
9226 : static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
9227 : struct cftype *cft)
9228 : {
9229 : return sched_group_rt_runtime(css_tg(css));
9230 : }
9231 :
9232 : static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
9233 : struct cftype *cftype, u64 rt_period_us)
9234 : {
9235 : return sched_group_set_rt_period(css_tg(css), rt_period_us);
9236 : }
9237 :
9238 : static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
9239 : struct cftype *cft)
9240 : {
9241 : return sched_group_rt_period(css_tg(css));
9242 : }
9243 : #endif /* CONFIG_RT_GROUP_SCHED */
9244 :
9245 : static struct cftype cpu_legacy_files[] = {
9246 : #ifdef CONFIG_FAIR_GROUP_SCHED
9247 : {
9248 : .name = "shares",
9249 : .read_u64 = cpu_shares_read_u64,
9250 : .write_u64 = cpu_shares_write_u64,
9251 : },
9252 : #endif
9253 : #ifdef CONFIG_CFS_BANDWIDTH
9254 : {
9255 : .name = "cfs_quota_us",
9256 : .read_s64 = cpu_cfs_quota_read_s64,
9257 : .write_s64 = cpu_cfs_quota_write_s64,
9258 : },
9259 : {
9260 : .name = "cfs_period_us",
9261 : .read_u64 = cpu_cfs_period_read_u64,
9262 : .write_u64 = cpu_cfs_period_write_u64,
9263 : },
9264 : {
9265 : .name = "stat",
9266 : .seq_show = cpu_cfs_stat_show,
9267 : },
9268 : #endif
9269 : #ifdef CONFIG_RT_GROUP_SCHED
9270 : {
9271 : .name = "rt_runtime_us",
9272 : .read_s64 = cpu_rt_runtime_read,
9273 : .write_s64 = cpu_rt_runtime_write,
9274 : },
9275 : {
9276 : .name = "rt_period_us",
9277 : .read_u64 = cpu_rt_period_read_uint,
9278 : .write_u64 = cpu_rt_period_write_uint,
9279 : },
9280 : #endif
9281 : #ifdef CONFIG_UCLAMP_TASK_GROUP
9282 : {
9283 : .name = "uclamp.min",
9284 : .flags = CFTYPE_NOT_ON_ROOT,
9285 : .seq_show = cpu_uclamp_min_show,
9286 : .write = cpu_uclamp_min_write,
9287 : },
9288 : {
9289 : .name = "uclamp.max",
9290 : .flags = CFTYPE_NOT_ON_ROOT,
9291 : .seq_show = cpu_uclamp_max_show,
9292 : .write = cpu_uclamp_max_write,
9293 : },
9294 : #endif
9295 : { } /* Terminate */
9296 : };
9297 :
9298 : static int cpu_extra_stat_show(struct seq_file *sf,
9299 : struct cgroup_subsys_state *css)
9300 : {
9301 : #ifdef CONFIG_CFS_BANDWIDTH
9302 : {
9303 : struct task_group *tg = css_tg(css);
9304 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9305 : u64 throttled_usec;
9306 :
9307 : throttled_usec = cfs_b->throttled_time;
9308 : do_div(throttled_usec, NSEC_PER_USEC);
9309 :
9310 : seq_printf(sf, "nr_periods %d\n"
9311 : "nr_throttled %d\n"
9312 : "throttled_usec %llu\n",
9313 : cfs_b->nr_periods, cfs_b->nr_throttled,
9314 : throttled_usec);
9315 : }
9316 : #endif
9317 : return 0;
9318 : }
9319 :
9320 : #ifdef CONFIG_FAIR_GROUP_SCHED
9321 : static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
9322 : struct cftype *cft)
9323 : {
9324 : struct task_group *tg = css_tg(css);
9325 : u64 weight = scale_load_down(tg->shares);
9326 :
9327 : return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
9328 : }
9329 :
9330 : static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
9331 : struct cftype *cft, u64 weight)
9332 : {
9333 : /*
9334 : * cgroup weight knobs should use the common MIN, DFL and MAX
9335 : * values which are 1, 100 and 10000 respectively. While it loses
9336 : * a bit of range on both ends, it maps pretty well onto the shares
9337 : * value used by scheduler and the round-trip conversions preserve
9338 : * the original value over the entire range.
9339 : */
9340 : if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
9341 : return -ERANGE;
9342 :
9343 : weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
9344 :
9345 : return sched_group_set_shares(css_tg(css), scale_load(weight));
9346 : }
9347 :
9348 : static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
9349 : struct cftype *cft)
9350 : {
9351 : unsigned long weight = scale_load_down(css_tg(css)->shares);
9352 : int last_delta = INT_MAX;
9353 : int prio, delta;
9354 :
9355 : /* find the closest nice value to the current weight */
9356 : for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
9357 : delta = abs(sched_prio_to_weight[prio] - weight);
9358 : if (delta >= last_delta)
9359 : break;
9360 : last_delta = delta;
9361 : }
9362 :
9363 : return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
9364 : }
9365 :
9366 : static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
9367 : struct cftype *cft, s64 nice)
9368 : {
9369 : unsigned long weight;
9370 : int idx;
9371 :
9372 : if (nice < MIN_NICE || nice > MAX_NICE)
9373 : return -ERANGE;
9374 :
9375 : idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
9376 : idx = array_index_nospec(idx, 40);
9377 : weight = sched_prio_to_weight[idx];
9378 :
9379 : return sched_group_set_shares(css_tg(css), scale_load(weight));
9380 : }
9381 : #endif
9382 :
9383 : static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
9384 : long period, long quota)
9385 : {
9386 : if (quota < 0)
9387 : seq_puts(sf, "max");
9388 : else
9389 : seq_printf(sf, "%ld", quota);
9390 :
9391 : seq_printf(sf, " %ld\n", period);
9392 : }
9393 :
9394 : /* caller should put the current value in *@periodp before calling */
9395 : static int __maybe_unused cpu_period_quota_parse(char *buf,
9396 : u64 *periodp, u64 *quotap)
9397 : {
9398 : char tok[21]; /* U64_MAX */
9399 :
9400 : if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
9401 : return -EINVAL;
9402 :
9403 : *periodp *= NSEC_PER_USEC;
9404 :
9405 : if (sscanf(tok, "%llu", quotap))
9406 : *quotap *= NSEC_PER_USEC;
9407 : else if (!strcmp(tok, "max"))
9408 : *quotap = RUNTIME_INF;
9409 : else
9410 : return -EINVAL;
9411 :
9412 : return 0;
9413 : }
9414 :
9415 : #ifdef CONFIG_CFS_BANDWIDTH
9416 : static int cpu_max_show(struct seq_file *sf, void *v)
9417 : {
9418 : struct task_group *tg = css_tg(seq_css(sf));
9419 :
9420 : cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
9421 : return 0;
9422 : }
9423 :
9424 : static ssize_t cpu_max_write(struct kernfs_open_file *of,
9425 : char *buf, size_t nbytes, loff_t off)
9426 : {
9427 : struct task_group *tg = css_tg(of_css(of));
9428 : u64 period = tg_get_cfs_period(tg);
9429 : u64 quota;
9430 : int ret;
9431 :
9432 : ret = cpu_period_quota_parse(buf, &period, "a);
9433 : if (!ret)
9434 : ret = tg_set_cfs_bandwidth(tg, period, quota);
9435 : return ret ?: nbytes;
9436 : }
9437 : #endif
9438 :
9439 : static struct cftype cpu_files[] = {
9440 : #ifdef CONFIG_FAIR_GROUP_SCHED
9441 : {
9442 : .name = "weight",
9443 : .flags = CFTYPE_NOT_ON_ROOT,
9444 : .read_u64 = cpu_weight_read_u64,
9445 : .write_u64 = cpu_weight_write_u64,
9446 : },
9447 : {
9448 : .name = "weight.nice",
9449 : .flags = CFTYPE_NOT_ON_ROOT,
9450 : .read_s64 = cpu_weight_nice_read_s64,
9451 : .write_s64 = cpu_weight_nice_write_s64,
9452 : },
9453 : #endif
9454 : #ifdef CONFIG_CFS_BANDWIDTH
9455 : {
9456 : .name = "max",
9457 : .flags = CFTYPE_NOT_ON_ROOT,
9458 : .seq_show = cpu_max_show,
9459 : .write = cpu_max_write,
9460 : },
9461 : #endif
9462 : #ifdef CONFIG_UCLAMP_TASK_GROUP
9463 : {
9464 : .name = "uclamp.min",
9465 : .flags = CFTYPE_NOT_ON_ROOT,
9466 : .seq_show = cpu_uclamp_min_show,
9467 : .write = cpu_uclamp_min_write,
9468 : },
9469 : {
9470 : .name = "uclamp.max",
9471 : .flags = CFTYPE_NOT_ON_ROOT,
9472 : .seq_show = cpu_uclamp_max_show,
9473 : .write = cpu_uclamp_max_write,
9474 : },
9475 : #endif
9476 : { } /* terminate */
9477 : };
9478 :
9479 : struct cgroup_subsys cpu_cgrp_subsys = {
9480 : .css_alloc = cpu_cgroup_css_alloc,
9481 : .css_online = cpu_cgroup_css_online,
9482 : .css_released = cpu_cgroup_css_released,
9483 : .css_free = cpu_cgroup_css_free,
9484 : .css_extra_stat_show = cpu_extra_stat_show,
9485 : .fork = cpu_cgroup_fork,
9486 : .can_attach = cpu_cgroup_can_attach,
9487 : .attach = cpu_cgroup_attach,
9488 : .legacy_cftypes = cpu_legacy_files,
9489 : .dfl_cftypes = cpu_files,
9490 : .early_init = true,
9491 : .threaded = true,
9492 : };
9493 :
9494 : #endif /* CONFIG_CGROUP_SCHED */
9495 :
9496 0 : void dump_cpu_task(int cpu)
9497 : {
9498 0 : pr_info("Task dump for CPU %d:\n", cpu);
9499 0 : sched_show_task(cpu_curr(cpu));
9500 0 : }
9501 :
9502 : /*
9503 : * Nice levels are multiplicative, with a gentle 10% change for every
9504 : * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
9505 : * nice 1, it will get ~10% less CPU time than another CPU-bound task
9506 : * that remained on nice 0.
9507 : *
9508 : * The "10% effect" is relative and cumulative: from _any_ nice level,
9509 : * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
9510 : * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
9511 : * If a task goes up by ~10% and another task goes down by ~10% then
9512 : * the relative distance between them is ~25%.)
9513 : */
9514 : const int sched_prio_to_weight[40] = {
9515 : /* -20 */ 88761, 71755, 56483, 46273, 36291,
9516 : /* -15 */ 29154, 23254, 18705, 14949, 11916,
9517 : /* -10 */ 9548, 7620, 6100, 4904, 3906,
9518 : /* -5 */ 3121, 2501, 1991, 1586, 1277,
9519 : /* 0 */ 1024, 820, 655, 526, 423,
9520 : /* 5 */ 335, 272, 215, 172, 137,
9521 : /* 10 */ 110, 87, 70, 56, 45,
9522 : /* 15 */ 36, 29, 23, 18, 15,
9523 : };
9524 :
9525 : /*
9526 : * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
9527 : *
9528 : * In cases where the weight does not change often, we can use the
9529 : * precalculated inverse to speed up arithmetics by turning divisions
9530 : * into multiplications:
9531 : */
9532 : const u32 sched_prio_to_wmult[40] = {
9533 : /* -20 */ 48388, 59856, 76040, 92818, 118348,
9534 : /* -15 */ 147320, 184698, 229616, 287308, 360437,
9535 : /* -10 */ 449829, 563644, 704093, 875809, 1099582,
9536 : /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
9537 : /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
9538 : /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
9539 : /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
9540 : /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
9541 : };
9542 :
9543 0 : void call_trace_sched_update_nr_running(struct rq *rq, int count)
9544 : {
9545 0 : trace_sched_update_nr_running_tp(rq, count);
9546 0 : }
|
