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 : #include <linux/highmem.h>
10 : #include <linux/hrtimer_api.h>
11 : #include <linux/ktime_api.h>
12 : #include <linux/sched/signal.h>
13 : #include <linux/syscalls_api.h>
14 : #include <linux/debug_locks.h>
15 : #include <linux/prefetch.h>
16 : #include <linux/capability.h>
17 : #include <linux/pgtable_api.h>
18 : #include <linux/wait_bit.h>
19 : #include <linux/jiffies.h>
20 : #include <linux/spinlock_api.h>
21 : #include <linux/cpumask_api.h>
22 : #include <linux/lockdep_api.h>
23 : #include <linux/hardirq.h>
24 : #include <linux/softirq.h>
25 : #include <linux/refcount_api.h>
26 : #include <linux/topology.h>
27 : #include <linux/sched/clock.h>
28 : #include <linux/sched/cond_resched.h>
29 : #include <linux/sched/cputime.h>
30 : #include <linux/sched/debug.h>
31 : #include <linux/sched/hotplug.h>
32 : #include <linux/sched/init.h>
33 : #include <linux/sched/isolation.h>
34 : #include <linux/sched/loadavg.h>
35 : #include <linux/sched/mm.h>
36 : #include <linux/sched/nohz.h>
37 : #include <linux/sched/rseq_api.h>
38 : #include <linux/sched/rt.h>
39 :
40 : #include <linux/blkdev.h>
41 : #include <linux/context_tracking.h>
42 : #include <linux/cpuset.h>
43 : #include <linux/delayacct.h>
44 : #include <linux/init_task.h>
45 : #include <linux/interrupt.h>
46 : #include <linux/ioprio.h>
47 : #include <linux/kallsyms.h>
48 : #include <linux/kcov.h>
49 : #include <linux/kprobes.h>
50 : #include <linux/llist_api.h>
51 : #include <linux/mmu_context.h>
52 : #include <linux/mmzone.h>
53 : #include <linux/mutex_api.h>
54 : #include <linux/nmi.h>
55 : #include <linux/nospec.h>
56 : #include <linux/perf_event_api.h>
57 : #include <linux/profile.h>
58 : #include <linux/psi.h>
59 : #include <linux/rcuwait_api.h>
60 : #include <linux/sched/wake_q.h>
61 : #include <linux/scs.h>
62 : #include <linux/slab.h>
63 : #include <linux/syscalls.h>
64 : #include <linux/vtime.h>
65 : #include <linux/wait_api.h>
66 : #include <linux/workqueue_api.h>
67 :
68 : #ifdef CONFIG_PREEMPT_DYNAMIC
69 : # ifdef CONFIG_GENERIC_ENTRY
70 : # include <linux/entry-common.h>
71 : # endif
72 : #endif
73 :
74 : #include <uapi/linux/sched/types.h>
75 :
76 : #include <asm/irq_regs.h>
77 : #include <asm/switch_to.h>
78 : #include <asm/tlb.h>
79 :
80 : #define CREATE_TRACE_POINTS
81 : #include <linux/sched/rseq_api.h>
82 : #include <trace/events/sched.h>
83 : #undef CREATE_TRACE_POINTS
84 :
85 : #include "sched.h"
86 : #include "stats.h"
87 : #include "autogroup.h"
88 :
89 : #include "autogroup.h"
90 : #include "pelt.h"
91 : #include "smp.h"
92 : #include "stats.h"
93 :
94 : #include "../workqueue_internal.h"
95 : #include "../../io_uring/io-wq.h"
96 : #include "../smpboot.h"
97 :
98 : /*
99 : * Export tracepoints that act as a bare tracehook (ie: have no trace event
100 : * associated with them) to allow external modules to probe them.
101 : */
102 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
103 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
104 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
105 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
106 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
107 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
108 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
109 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
110 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
111 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
112 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
113 :
114 : DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 :
116 : #ifdef CONFIG_SCHED_DEBUG
117 : /*
118 : * Debugging: various feature bits
119 : *
120 : * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
121 : * sysctl_sched_features, defined in sched.h, to allow constants propagation
122 : * at compile time and compiler optimization based on features default.
123 : */
124 : #define SCHED_FEAT(name, enabled) \
125 : (1UL << __SCHED_FEAT_##name) * enabled |
126 : const_debug unsigned int sysctl_sched_features =
127 : #include "features.h"
128 : 0;
129 : #undef SCHED_FEAT
130 :
131 : /*
132 : * Print a warning if need_resched is set for the given duration (if
133 : * LATENCY_WARN is enabled).
134 : *
135 : * If sysctl_resched_latency_warn_once is set, only one warning will be shown
136 : * per boot.
137 : */
138 : __read_mostly int sysctl_resched_latency_warn_ms = 100;
139 : __read_mostly int sysctl_resched_latency_warn_once = 1;
140 : #endif /* CONFIG_SCHED_DEBUG */
141 :
142 : /*
143 : * Number of tasks to iterate in a single balance run.
144 : * Limited because this is done with IRQs disabled.
145 : */
146 : const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
147 :
148 : __read_mostly int scheduler_running;
149 :
150 : #ifdef CONFIG_SCHED_CORE
151 :
152 : DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
153 :
154 : /* kernel prio, less is more */
155 : static inline int __task_prio(const struct task_struct *p)
156 : {
157 : if (p->sched_class == &stop_sched_class) /* trumps deadline */
158 : return -2;
159 :
160 : if (rt_prio(p->prio)) /* includes deadline */
161 : return p->prio; /* [-1, 99] */
162 :
163 : if (p->sched_class == &idle_sched_class)
164 : return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
165 :
166 : return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
167 : }
168 :
169 : /*
170 : * l(a,b)
171 : * le(a,b) := !l(b,a)
172 : * g(a,b) := l(b,a)
173 : * ge(a,b) := !l(a,b)
174 : */
175 :
176 : /* real prio, less is less */
177 : static inline bool prio_less(const struct task_struct *a,
178 : const struct task_struct *b, bool in_fi)
179 : {
180 :
181 : int pa = __task_prio(a), pb = __task_prio(b);
182 :
183 : if (-pa < -pb)
184 : return true;
185 :
186 : if (-pb < -pa)
187 : return false;
188 :
189 : if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
190 : return !dl_time_before(a->dl.deadline, b->dl.deadline);
191 :
192 : if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
193 : return cfs_prio_less(a, b, in_fi);
194 :
195 : return false;
196 : }
197 :
198 : static inline bool __sched_core_less(const struct task_struct *a,
199 : const struct task_struct *b)
200 : {
201 : if (a->core_cookie < b->core_cookie)
202 : return true;
203 :
204 : if (a->core_cookie > b->core_cookie)
205 : return false;
206 :
207 : /* flip prio, so high prio is leftmost */
208 : if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
209 : return true;
210 :
211 : return false;
212 : }
213 :
214 : #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
215 :
216 : static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
217 : {
218 : return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
219 : }
220 :
221 : static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
222 : {
223 : const struct task_struct *p = __node_2_sc(node);
224 : unsigned long cookie = (unsigned long)key;
225 :
226 : if (cookie < p->core_cookie)
227 : return -1;
228 :
229 : if (cookie > p->core_cookie)
230 : return 1;
231 :
232 : return 0;
233 : }
234 :
235 : void sched_core_enqueue(struct rq *rq, struct task_struct *p)
236 : {
237 : rq->core->core_task_seq++;
238 :
239 : if (!p->core_cookie)
240 : return;
241 :
242 : rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
243 : }
244 :
245 : void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
246 : {
247 : rq->core->core_task_seq++;
248 :
249 : if (sched_core_enqueued(p)) {
250 : rb_erase(&p->core_node, &rq->core_tree);
251 : RB_CLEAR_NODE(&p->core_node);
252 : }
253 :
254 : /*
255 : * Migrating the last task off the cpu, with the cpu in forced idle
256 : * state. Reschedule to create an accounting edge for forced idle,
257 : * and re-examine whether the core is still in forced idle state.
258 : */
259 : if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
260 : rq->core->core_forceidle_count && rq->curr == rq->idle)
261 : resched_curr(rq);
262 : }
263 :
264 : /*
265 : * Find left-most (aka, highest priority) task matching @cookie.
266 : */
267 : static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
268 : {
269 : struct rb_node *node;
270 :
271 : node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
272 : /*
273 : * The idle task always matches any cookie!
274 : */
275 : if (!node)
276 : return idle_sched_class.pick_task(rq);
277 :
278 : return __node_2_sc(node);
279 : }
280 :
281 : static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
282 : {
283 : struct rb_node *node = &p->core_node;
284 :
285 : node = rb_next(node);
286 : if (!node)
287 : return NULL;
288 :
289 : p = container_of(node, struct task_struct, core_node);
290 : if (p->core_cookie != cookie)
291 : return NULL;
292 :
293 : return p;
294 : }
295 :
296 : /*
297 : * Magic required such that:
298 : *
299 : * raw_spin_rq_lock(rq);
300 : * ...
301 : * raw_spin_rq_unlock(rq);
302 : *
303 : * ends up locking and unlocking the _same_ lock, and all CPUs
304 : * always agree on what rq has what lock.
305 : *
306 : * XXX entirely possible to selectively enable cores, don't bother for now.
307 : */
308 :
309 : static DEFINE_MUTEX(sched_core_mutex);
310 : static atomic_t sched_core_count;
311 : static struct cpumask sched_core_mask;
312 :
313 : static void sched_core_lock(int cpu, unsigned long *flags)
314 : {
315 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
316 : int t, i = 0;
317 :
318 : local_irq_save(*flags);
319 : for_each_cpu(t, smt_mask)
320 : raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
321 : }
322 :
323 : static void sched_core_unlock(int cpu, unsigned long *flags)
324 : {
325 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
326 : int t;
327 :
328 : for_each_cpu(t, smt_mask)
329 : raw_spin_unlock(&cpu_rq(t)->__lock);
330 : local_irq_restore(*flags);
331 : }
332 :
333 : static void __sched_core_flip(bool enabled)
334 : {
335 : unsigned long flags;
336 : int cpu, t;
337 :
338 : cpus_read_lock();
339 :
340 : /*
341 : * Toggle the online cores, one by one.
342 : */
343 : cpumask_copy(&sched_core_mask, cpu_online_mask);
344 : for_each_cpu(cpu, &sched_core_mask) {
345 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
346 :
347 : sched_core_lock(cpu, &flags);
348 :
349 : for_each_cpu(t, smt_mask)
350 : cpu_rq(t)->core_enabled = enabled;
351 :
352 : cpu_rq(cpu)->core->core_forceidle_start = 0;
353 :
354 : sched_core_unlock(cpu, &flags);
355 :
356 : cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
357 : }
358 :
359 : /*
360 : * Toggle the offline CPUs.
361 : */
362 : for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
363 : cpu_rq(cpu)->core_enabled = enabled;
364 :
365 : cpus_read_unlock();
366 : }
367 :
368 : static void sched_core_assert_empty(void)
369 : {
370 : int cpu;
371 :
372 : for_each_possible_cpu(cpu)
373 : WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
374 : }
375 :
376 : static void __sched_core_enable(void)
377 : {
378 : static_branch_enable(&__sched_core_enabled);
379 : /*
380 : * Ensure all previous instances of raw_spin_rq_*lock() have finished
381 : * and future ones will observe !sched_core_disabled().
382 : */
383 : synchronize_rcu();
384 : __sched_core_flip(true);
385 : sched_core_assert_empty();
386 : }
387 :
388 : static void __sched_core_disable(void)
389 : {
390 : sched_core_assert_empty();
391 : __sched_core_flip(false);
392 : static_branch_disable(&__sched_core_enabled);
393 : }
394 :
395 : void sched_core_get(void)
396 : {
397 : if (atomic_inc_not_zero(&sched_core_count))
398 : return;
399 :
400 : mutex_lock(&sched_core_mutex);
401 : if (!atomic_read(&sched_core_count))
402 : __sched_core_enable();
403 :
404 : smp_mb__before_atomic();
405 : atomic_inc(&sched_core_count);
406 : mutex_unlock(&sched_core_mutex);
407 : }
408 :
409 : static void __sched_core_put(struct work_struct *work)
410 : {
411 : if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
412 : __sched_core_disable();
413 : mutex_unlock(&sched_core_mutex);
414 : }
415 : }
416 :
417 : void sched_core_put(void)
418 : {
419 : static DECLARE_WORK(_work, __sched_core_put);
420 :
421 : /*
422 : * "There can be only one"
423 : *
424 : * Either this is the last one, or we don't actually need to do any
425 : * 'work'. If it is the last *again*, we rely on
426 : * WORK_STRUCT_PENDING_BIT.
427 : */
428 : if (!atomic_add_unless(&sched_core_count, -1, 1))
429 : schedule_work(&_work);
430 : }
431 :
432 : #else /* !CONFIG_SCHED_CORE */
433 :
434 : static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
435 : static inline void
436 : sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
437 :
438 : #endif /* CONFIG_SCHED_CORE */
439 :
440 : /*
441 : * Serialization rules:
442 : *
443 : * Lock order:
444 : *
445 : * p->pi_lock
446 : * rq->lock
447 : * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
448 : *
449 : * rq1->lock
450 : * rq2->lock where: rq1 < rq2
451 : *
452 : * Regular state:
453 : *
454 : * Normal scheduling state is serialized by rq->lock. __schedule() takes the
455 : * local CPU's rq->lock, it optionally removes the task from the runqueue and
456 : * always looks at the local rq data structures to find the most eligible task
457 : * to run next.
458 : *
459 : * Task enqueue is also under rq->lock, possibly taken from another CPU.
460 : * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
461 : * the local CPU to avoid bouncing the runqueue state around [ see
462 : * ttwu_queue_wakelist() ]
463 : *
464 : * Task wakeup, specifically wakeups that involve migration, are horribly
465 : * complicated to avoid having to take two rq->locks.
466 : *
467 : * Special state:
468 : *
469 : * System-calls and anything external will use task_rq_lock() which acquires
470 : * both p->pi_lock and rq->lock. As a consequence the state they change is
471 : * stable while holding either lock:
472 : *
473 : * - sched_setaffinity()/
474 : * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
475 : * - set_user_nice(): p->se.load, p->*prio
476 : * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
477 : * p->se.load, p->rt_priority,
478 : * p->dl.dl_{runtime, deadline, period, flags, bw, density}
479 : * - sched_setnuma(): p->numa_preferred_nid
480 : * - sched_move_task(): p->sched_task_group
481 : * - uclamp_update_active() p->uclamp*
482 : *
483 : * p->state <- TASK_*:
484 : *
485 : * is changed locklessly using set_current_state(), __set_current_state() or
486 : * set_special_state(), see their respective comments, or by
487 : * try_to_wake_up(). This latter uses p->pi_lock to serialize against
488 : * concurrent self.
489 : *
490 : * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
491 : *
492 : * is set by activate_task() and cleared by deactivate_task(), under
493 : * rq->lock. Non-zero indicates the task is runnable, the special
494 : * ON_RQ_MIGRATING state is used for migration without holding both
495 : * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
496 : *
497 : * p->on_cpu <- { 0, 1 }:
498 : *
499 : * is set by prepare_task() and cleared by finish_task() such that it will be
500 : * set before p is scheduled-in and cleared after p is scheduled-out, both
501 : * under rq->lock. Non-zero indicates the task is running on its CPU.
502 : *
503 : * [ The astute reader will observe that it is possible for two tasks on one
504 : * CPU to have ->on_cpu = 1 at the same time. ]
505 : *
506 : * task_cpu(p): is changed by set_task_cpu(), the rules are:
507 : *
508 : * - Don't call set_task_cpu() on a blocked task:
509 : *
510 : * We don't care what CPU we're not running on, this simplifies hotplug,
511 : * the CPU assignment of blocked tasks isn't required to be valid.
512 : *
513 : * - for try_to_wake_up(), called under p->pi_lock:
514 : *
515 : * This allows try_to_wake_up() to only take one rq->lock, see its comment.
516 : *
517 : * - for migration called under rq->lock:
518 : * [ see task_on_rq_migrating() in task_rq_lock() ]
519 : *
520 : * o move_queued_task()
521 : * o detach_task()
522 : *
523 : * - for migration called under double_rq_lock():
524 : *
525 : * o __migrate_swap_task()
526 : * o push_rt_task() / pull_rt_task()
527 : * o push_dl_task() / pull_dl_task()
528 : * o dl_task_offline_migration()
529 : *
530 : */
531 :
532 348 : void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
533 : {
534 : raw_spinlock_t *lock;
535 :
536 : /* Matches synchronize_rcu() in __sched_core_enable() */
537 7891 : preempt_disable();
538 : if (sched_core_disabled()) {
539 7891 : raw_spin_lock_nested(&rq->__lock, subclass);
540 : /* preempt_count *MUST* be > 1 */
541 7891 : preempt_enable_no_resched();
542 : return;
543 : }
544 :
545 : for (;;) {
546 : lock = __rq_lockp(rq);
547 : raw_spin_lock_nested(lock, subclass);
548 : if (likely(lock == __rq_lockp(rq))) {
549 : /* preempt_count *MUST* be > 1 */
550 : preempt_enable_no_resched();
551 : return;
552 : }
553 : raw_spin_unlock(lock);
554 : }
555 : }
556 :
557 0 : bool raw_spin_rq_trylock(struct rq *rq)
558 : {
559 : raw_spinlock_t *lock;
560 : bool ret;
561 :
562 : /* Matches synchronize_rcu() in __sched_core_enable() */
563 0 : preempt_disable();
564 : if (sched_core_disabled()) {
565 0 : ret = raw_spin_trylock(&rq->__lock);
566 0 : preempt_enable();
567 : return ret;
568 : }
569 :
570 : for (;;) {
571 : lock = __rq_lockp(rq);
572 : ret = raw_spin_trylock(lock);
573 : if (!ret || (likely(lock == __rq_lockp(rq)))) {
574 : preempt_enable();
575 : return ret;
576 : }
577 : raw_spin_unlock(lock);
578 : }
579 : }
580 :
581 348 : void raw_spin_rq_unlock(struct rq *rq)
582 : {
583 7891 : raw_spin_unlock(rq_lockp(rq));
584 348 : }
585 :
586 : #ifdef CONFIG_SMP
587 : /*
588 : * double_rq_lock - safely lock two runqueues
589 : */
590 : void double_rq_lock(struct rq *rq1, struct rq *rq2)
591 : {
592 : lockdep_assert_irqs_disabled();
593 :
594 : if (rq_order_less(rq2, rq1))
595 : swap(rq1, rq2);
596 :
597 : raw_spin_rq_lock(rq1);
598 : if (__rq_lockp(rq1) != __rq_lockp(rq2))
599 : raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
600 :
601 : double_rq_clock_clear_update(rq1, rq2);
602 : }
603 : #endif
604 :
605 : /*
606 : * __task_rq_lock - lock the rq @p resides on.
607 : */
608 0 : struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
609 : __acquires(rq->lock)
610 : {
611 : struct rq *rq;
612 :
613 : lockdep_assert_held(&p->pi_lock);
614 :
615 : for (;;) {
616 348 : rq = task_rq(p);
617 348 : raw_spin_rq_lock(rq);
618 696 : if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
619 348 : rq_pin_lock(rq, rf);
620 0 : return rq;
621 : }
622 : raw_spin_rq_unlock(rq);
623 :
624 0 : while (unlikely(task_on_rq_migrating(p)))
625 : cpu_relax();
626 : }
627 : }
628 :
629 : /*
630 : * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
631 : */
632 351 : struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
633 : __acquires(p->pi_lock)
634 : __acquires(rq->lock)
635 : {
636 : struct rq *rq;
637 :
638 : for (;;) {
639 351 : raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
640 351 : rq = task_rq(p);
641 351 : raw_spin_rq_lock(rq);
642 : /*
643 : * move_queued_task() task_rq_lock()
644 : *
645 : * ACQUIRE (rq->lock)
646 : * [S] ->on_rq = MIGRATING [L] rq = task_rq()
647 : * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
648 : * [S] ->cpu = new_cpu [L] task_rq()
649 : * [L] ->on_rq
650 : * RELEASE (rq->lock)
651 : *
652 : * If we observe the old CPU in task_rq_lock(), the acquire of
653 : * the old rq->lock will fully serialize against the stores.
654 : *
655 : * If we observe the new CPU in task_rq_lock(), the address
656 : * dependency headed by '[L] rq = task_rq()' and the acquire
657 : * will pair with the WMB to ensure we then also see migrating.
658 : */
659 702 : if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
660 351 : rq_pin_lock(rq, rf);
661 351 : return rq;
662 : }
663 0 : raw_spin_rq_unlock(rq);
664 0 : raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
665 :
666 0 : while (unlikely(task_on_rq_migrating(p)))
667 : cpu_relax();
668 : }
669 : }
670 :
671 : /*
672 : * RQ-clock updating methods:
673 : */
674 :
675 : static void update_rq_clock_task(struct rq *rq, s64 delta)
676 : {
677 : /*
678 : * In theory, the compile should just see 0 here, and optimize out the call
679 : * to sched_rt_avg_update. But I don't trust it...
680 : */
681 7126 : s64 __maybe_unused steal = 0, irq_delta = 0;
682 :
683 : #ifdef CONFIG_IRQ_TIME_ACCOUNTING
684 : irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
685 :
686 : /*
687 : * Since irq_time is only updated on {soft,}irq_exit, we might run into
688 : * this case when a previous update_rq_clock() happened inside a
689 : * {soft,}irq region.
690 : *
691 : * When this happens, we stop ->clock_task and only update the
692 : * prev_irq_time stamp to account for the part that fit, so that a next
693 : * update will consume the rest. This ensures ->clock_task is
694 : * monotonic.
695 : *
696 : * It does however cause some slight miss-attribution of {soft,}irq
697 : * time, a more accurate solution would be to update the irq_time using
698 : * the current rq->clock timestamp, except that would require using
699 : * atomic ops.
700 : */
701 : if (irq_delta > delta)
702 : irq_delta = delta;
703 :
704 : rq->prev_irq_time += irq_delta;
705 : delta -= irq_delta;
706 : psi_account_irqtime(rq->curr, irq_delta);
707 : #endif
708 : #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
709 : if (static_key_false((¶virt_steal_rq_enabled))) {
710 : steal = paravirt_steal_clock(cpu_of(rq));
711 : steal -= rq->prev_steal_time_rq;
712 :
713 : if (unlikely(steal > delta))
714 : steal = delta;
715 :
716 : rq->prev_steal_time_rq += steal;
717 : delta -= steal;
718 : }
719 : #endif
720 :
721 7126 : rq->clock_task += delta;
722 :
723 : #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
724 : if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
725 : update_irq_load_avg(rq, irq_delta + steal);
726 : #endif
727 7126 : update_rq_clock_pelt(rq, delta);
728 : }
729 :
730 7890 : void update_rq_clock(struct rq *rq)
731 : {
732 : s64 delta;
733 :
734 7890 : lockdep_assert_rq_held(rq);
735 :
736 7890 : if (rq->clock_update_flags & RQCF_ACT_SKIP)
737 : return;
738 :
739 : #ifdef CONFIG_SCHED_DEBUG
740 7126 : if (sched_feat(WARN_DOUBLE_CLOCK))
741 0 : SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
742 7126 : rq->clock_update_flags |= RQCF_UPDATED;
743 : #endif
744 :
745 7126 : delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
746 7126 : if (delta < 0)
747 : return;
748 7126 : rq->clock += delta;
749 7126 : update_rq_clock_task(rq, delta);
750 : }
751 :
752 : #ifdef CONFIG_SCHED_HRTICK
753 : /*
754 : * Use HR-timers to deliver accurate preemption points.
755 : */
756 :
757 : static void hrtick_clear(struct rq *rq)
758 : {
759 : if (hrtimer_active(&rq->hrtick_timer))
760 : hrtimer_cancel(&rq->hrtick_timer);
761 : }
762 :
763 : /*
764 : * High-resolution timer tick.
765 : * Runs from hardirq context with interrupts disabled.
766 : */
767 : static enum hrtimer_restart hrtick(struct hrtimer *timer)
768 : {
769 : struct rq *rq = container_of(timer, struct rq, hrtick_timer);
770 : struct rq_flags rf;
771 :
772 : WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
773 :
774 : rq_lock(rq, &rf);
775 : update_rq_clock(rq);
776 : rq->curr->sched_class->task_tick(rq, rq->curr, 1);
777 : rq_unlock(rq, &rf);
778 :
779 : return HRTIMER_NORESTART;
780 : }
781 :
782 : #ifdef CONFIG_SMP
783 :
784 : static void __hrtick_restart(struct rq *rq)
785 : {
786 : struct hrtimer *timer = &rq->hrtick_timer;
787 : ktime_t time = rq->hrtick_time;
788 :
789 : hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
790 : }
791 :
792 : /*
793 : * called from hardirq (IPI) context
794 : */
795 : static void __hrtick_start(void *arg)
796 : {
797 : struct rq *rq = arg;
798 : struct rq_flags rf;
799 :
800 : rq_lock(rq, &rf);
801 : __hrtick_restart(rq);
802 : rq_unlock(rq, &rf);
803 : }
804 :
805 : /*
806 : * Called to set the hrtick timer state.
807 : *
808 : * called with rq->lock held and irqs disabled
809 : */
810 : void hrtick_start(struct rq *rq, u64 delay)
811 : {
812 : struct hrtimer *timer = &rq->hrtick_timer;
813 : s64 delta;
814 :
815 : /*
816 : * Don't schedule slices shorter than 10000ns, that just
817 : * doesn't make sense and can cause timer DoS.
818 : */
819 : delta = max_t(s64, delay, 10000LL);
820 : rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
821 :
822 : if (rq == this_rq())
823 : __hrtick_restart(rq);
824 : else
825 : smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
826 : }
827 :
828 : #else
829 : /*
830 : * Called to set the hrtick timer state.
831 : *
832 : * called with rq->lock held and irqs disabled
833 : */
834 : void hrtick_start(struct rq *rq, u64 delay)
835 : {
836 : /*
837 : * Don't schedule slices shorter than 10000ns, that just
838 : * doesn't make sense. Rely on vruntime for fairness.
839 : */
840 : delay = max_t(u64, delay, 10000LL);
841 : hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
842 : HRTIMER_MODE_REL_PINNED_HARD);
843 : }
844 :
845 : #endif /* CONFIG_SMP */
846 :
847 : static void hrtick_rq_init(struct rq *rq)
848 : {
849 : #ifdef CONFIG_SMP
850 : INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
851 : #endif
852 : hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
853 : rq->hrtick_timer.function = hrtick;
854 : }
855 : #else /* CONFIG_SCHED_HRTICK */
856 : static inline void hrtick_clear(struct rq *rq)
857 : {
858 : }
859 :
860 : static inline void hrtick_rq_init(struct rq *rq)
861 : {
862 : }
863 : #endif /* CONFIG_SCHED_HRTICK */
864 :
865 : /*
866 : * cmpxchg based fetch_or, macro so it works for different integer types
867 : */
868 : #define fetch_or(ptr, mask) \
869 : ({ \
870 : typeof(ptr) _ptr = (ptr); \
871 : typeof(mask) _mask = (mask); \
872 : typeof(*_ptr) _val = *_ptr; \
873 : \
874 : do { \
875 : } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
876 : _val; \
877 : })
878 :
879 : #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
880 : /*
881 : * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
882 : * this avoids any races wrt polling state changes and thereby avoids
883 : * spurious IPIs.
884 : */
885 : static inline bool set_nr_and_not_polling(struct task_struct *p)
886 : {
887 : struct thread_info *ti = task_thread_info(p);
888 : return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
889 : }
890 :
891 : /*
892 : * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
893 : *
894 : * If this returns true, then the idle task promises to call
895 : * sched_ttwu_pending() and reschedule soon.
896 : */
897 : static bool set_nr_if_polling(struct task_struct *p)
898 : {
899 : struct thread_info *ti = task_thread_info(p);
900 : typeof(ti->flags) val = READ_ONCE(ti->flags);
901 :
902 : for (;;) {
903 : if (!(val & _TIF_POLLING_NRFLAG))
904 : return false;
905 : if (val & _TIF_NEED_RESCHED)
906 : return true;
907 : if (try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED))
908 : break;
909 : }
910 : return true;
911 : }
912 :
913 : #else
914 : static inline bool set_nr_and_not_polling(struct task_struct *p)
915 : {
916 : set_tsk_need_resched(p);
917 : return true;
918 : }
919 :
920 : #ifdef CONFIG_SMP
921 : static inline bool set_nr_if_polling(struct task_struct *p)
922 : {
923 : return false;
924 : }
925 : #endif
926 : #endif
927 :
928 : static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
929 : {
930 0 : struct wake_q_node *node = &task->wake_q;
931 :
932 : /*
933 : * Atomically grab the task, if ->wake_q is !nil already it means
934 : * it's already queued (either by us or someone else) and will get the
935 : * wakeup due to that.
936 : *
937 : * In order to ensure that a pending wakeup will observe our pending
938 : * state, even in the failed case, an explicit smp_mb() must be used.
939 : */
940 0 : smp_mb__before_atomic();
941 0 : if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
942 : return false;
943 :
944 : /*
945 : * The head is context local, there can be no concurrency.
946 : */
947 0 : *head->lastp = node;
948 0 : head->lastp = &node->next;
949 : return true;
950 : }
951 :
952 : /**
953 : * wake_q_add() - queue a wakeup for 'later' waking.
954 : * @head: the wake_q_head to add @task to
955 : * @task: the task to queue for 'later' wakeup
956 : *
957 : * Queue a task for later wakeup, most likely by the wake_up_q() call in the
958 : * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
959 : * instantly.
960 : *
961 : * This function must be used as-if it were wake_up_process(); IOW the task
962 : * must be ready to be woken at this location.
963 : */
964 0 : void wake_q_add(struct wake_q_head *head, struct task_struct *task)
965 : {
966 0 : if (__wake_q_add(head, task))
967 : get_task_struct(task);
968 0 : }
969 :
970 : /**
971 : * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
972 : * @head: the wake_q_head to add @task to
973 : * @task: the task to queue for 'later' wakeup
974 : *
975 : * Queue a task for later wakeup, most likely by the wake_up_q() call in the
976 : * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
977 : * instantly.
978 : *
979 : * This function must be used as-if it were wake_up_process(); IOW the task
980 : * must be ready to be woken at this location.
981 : *
982 : * This function is essentially a task-safe equivalent to wake_q_add(). Callers
983 : * that already hold reference to @task can call the 'safe' version and trust
984 : * wake_q to do the right thing depending whether or not the @task is already
985 : * queued for wakeup.
986 : */
987 0 : void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
988 : {
989 0 : if (!__wake_q_add(head, task))
990 0 : put_task_struct(task);
991 0 : }
992 :
993 0 : void wake_up_q(struct wake_q_head *head)
994 : {
995 0 : struct wake_q_node *node = head->first;
996 :
997 0 : while (node != WAKE_Q_TAIL) {
998 : struct task_struct *task;
999 :
1000 0 : task = container_of(node, struct task_struct, wake_q);
1001 : /* Task can safely be re-inserted now: */
1002 0 : node = node->next;
1003 0 : task->wake_q.next = NULL;
1004 :
1005 : /*
1006 : * wake_up_process() executes a full barrier, which pairs with
1007 : * the queueing in wake_q_add() so as not to miss wakeups.
1008 : */
1009 0 : wake_up_process(task);
1010 0 : put_task_struct(task);
1011 : }
1012 0 : }
1013 :
1014 : /*
1015 : * resched_curr - mark rq's current task 'to be rescheduled now'.
1016 : *
1017 : * On UP this means the setting of the need_resched flag, on SMP it
1018 : * might also involve a cross-CPU call to trigger the scheduler on
1019 : * the target CPU.
1020 : */
1021 908 : void resched_curr(struct rq *rq)
1022 : {
1023 908 : struct task_struct *curr = rq->curr;
1024 : int cpu;
1025 :
1026 1816 : lockdep_assert_rq_held(rq);
1027 :
1028 908 : if (test_tsk_need_resched(curr))
1029 : return;
1030 :
1031 770 : cpu = cpu_of(rq);
1032 :
1033 : if (cpu == smp_processor_id()) {
1034 : set_tsk_need_resched(curr);
1035 : set_preempt_need_resched();
1036 : return;
1037 : }
1038 :
1039 : if (set_nr_and_not_polling(curr))
1040 : smp_send_reschedule(cpu);
1041 : else
1042 : trace_sched_wake_idle_without_ipi(cpu);
1043 : }
1044 :
1045 0 : void resched_cpu(int cpu)
1046 : {
1047 0 : struct rq *rq = cpu_rq(cpu);
1048 : unsigned long flags;
1049 :
1050 0 : raw_spin_rq_lock_irqsave(rq, flags);
1051 0 : if (cpu_online(cpu) || cpu == smp_processor_id())
1052 0 : resched_curr(rq);
1053 0 : raw_spin_rq_unlock_irqrestore(rq, flags);
1054 0 : }
1055 :
1056 : #ifdef CONFIG_SMP
1057 : #ifdef CONFIG_NO_HZ_COMMON
1058 : /*
1059 : * In the semi idle case, use the nearest busy CPU for migrating timers
1060 : * from an idle CPU. This is good for power-savings.
1061 : *
1062 : * We don't do similar optimization for completely idle system, as
1063 : * selecting an idle CPU will add more delays to the timers than intended
1064 : * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1065 : */
1066 : int get_nohz_timer_target(void)
1067 : {
1068 : int i, cpu = smp_processor_id(), default_cpu = -1;
1069 : struct sched_domain *sd;
1070 : const struct cpumask *hk_mask;
1071 :
1072 : if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1073 : if (!idle_cpu(cpu))
1074 : return cpu;
1075 : default_cpu = cpu;
1076 : }
1077 :
1078 : hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1079 :
1080 : rcu_read_lock();
1081 : for_each_domain(cpu, sd) {
1082 : for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1083 : if (cpu == i)
1084 : continue;
1085 :
1086 : if (!idle_cpu(i)) {
1087 : cpu = i;
1088 : goto unlock;
1089 : }
1090 : }
1091 : }
1092 :
1093 : if (default_cpu == -1)
1094 : default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1095 : cpu = default_cpu;
1096 : unlock:
1097 : rcu_read_unlock();
1098 : return cpu;
1099 : }
1100 :
1101 : /*
1102 : * When add_timer_on() enqueues a timer into the timer wheel of an
1103 : * idle CPU then this timer might expire before the next timer event
1104 : * which is scheduled to wake up that CPU. In case of a completely
1105 : * idle system the next event might even be infinite time into the
1106 : * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1107 : * leaves the inner idle loop so the newly added timer is taken into
1108 : * account when the CPU goes back to idle and evaluates the timer
1109 : * wheel for the next timer event.
1110 : */
1111 : static void wake_up_idle_cpu(int cpu)
1112 : {
1113 : struct rq *rq = cpu_rq(cpu);
1114 :
1115 : if (cpu == smp_processor_id())
1116 : return;
1117 :
1118 : if (set_nr_and_not_polling(rq->idle))
1119 : smp_send_reschedule(cpu);
1120 : else
1121 : trace_sched_wake_idle_without_ipi(cpu);
1122 : }
1123 :
1124 : static bool wake_up_full_nohz_cpu(int cpu)
1125 : {
1126 : /*
1127 : * We just need the target to call irq_exit() and re-evaluate
1128 : * the next tick. The nohz full kick at least implies that.
1129 : * If needed we can still optimize that later with an
1130 : * empty IRQ.
1131 : */
1132 : if (cpu_is_offline(cpu))
1133 : return true; /* Don't try to wake offline CPUs. */
1134 : if (tick_nohz_full_cpu(cpu)) {
1135 : if (cpu != smp_processor_id() ||
1136 : tick_nohz_tick_stopped())
1137 : tick_nohz_full_kick_cpu(cpu);
1138 : return true;
1139 : }
1140 :
1141 : return false;
1142 : }
1143 :
1144 : /*
1145 : * Wake up the specified CPU. If the CPU is going offline, it is the
1146 : * caller's responsibility to deal with the lost wakeup, for example,
1147 : * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1148 : */
1149 : void wake_up_nohz_cpu(int cpu)
1150 : {
1151 : if (!wake_up_full_nohz_cpu(cpu))
1152 : wake_up_idle_cpu(cpu);
1153 : }
1154 :
1155 : static void nohz_csd_func(void *info)
1156 : {
1157 : struct rq *rq = info;
1158 : int cpu = cpu_of(rq);
1159 : unsigned int flags;
1160 :
1161 : /*
1162 : * Release the rq::nohz_csd.
1163 : */
1164 : flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1165 : WARN_ON(!(flags & NOHZ_KICK_MASK));
1166 :
1167 : rq->idle_balance = idle_cpu(cpu);
1168 : if (rq->idle_balance && !need_resched()) {
1169 : rq->nohz_idle_balance = flags;
1170 : raise_softirq_irqoff(SCHED_SOFTIRQ);
1171 : }
1172 : }
1173 :
1174 : #endif /* CONFIG_NO_HZ_COMMON */
1175 :
1176 : #ifdef CONFIG_NO_HZ_FULL
1177 : bool sched_can_stop_tick(struct rq *rq)
1178 : {
1179 : int fifo_nr_running;
1180 :
1181 : /* Deadline tasks, even if single, need the tick */
1182 : if (rq->dl.dl_nr_running)
1183 : return false;
1184 :
1185 : /*
1186 : * If there are more than one RR tasks, we need the tick to affect the
1187 : * actual RR behaviour.
1188 : */
1189 : if (rq->rt.rr_nr_running) {
1190 : if (rq->rt.rr_nr_running == 1)
1191 : return true;
1192 : else
1193 : return false;
1194 : }
1195 :
1196 : /*
1197 : * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1198 : * forced preemption between FIFO tasks.
1199 : */
1200 : fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1201 : if (fifo_nr_running)
1202 : return true;
1203 :
1204 : /*
1205 : * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1206 : * if there's more than one we need the tick for involuntary
1207 : * preemption.
1208 : */
1209 : if (rq->nr_running > 1)
1210 : return false;
1211 :
1212 : return true;
1213 : }
1214 : #endif /* CONFIG_NO_HZ_FULL */
1215 : #endif /* CONFIG_SMP */
1216 :
1217 : #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1218 : (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1219 : /*
1220 : * Iterate task_group tree rooted at *from, calling @down when first entering a
1221 : * node and @up when leaving it for the final time.
1222 : *
1223 : * Caller must hold rcu_lock or sufficient equivalent.
1224 : */
1225 : int walk_tg_tree_from(struct task_group *from,
1226 : tg_visitor down, tg_visitor up, void *data)
1227 : {
1228 : struct task_group *parent, *child;
1229 : int ret;
1230 :
1231 : parent = from;
1232 :
1233 : down:
1234 : ret = (*down)(parent, data);
1235 : if (ret)
1236 : goto out;
1237 : list_for_each_entry_rcu(child, &parent->children, siblings) {
1238 : parent = child;
1239 : goto down;
1240 :
1241 : up:
1242 : continue;
1243 : }
1244 : ret = (*up)(parent, data);
1245 : if (ret || parent == from)
1246 : goto out;
1247 :
1248 : child = parent;
1249 : parent = parent->parent;
1250 : if (parent)
1251 : goto up;
1252 : out:
1253 : return ret;
1254 : }
1255 :
1256 : int tg_nop(struct task_group *tg, void *data)
1257 : {
1258 : return 0;
1259 : }
1260 : #endif
1261 :
1262 6 : static void set_load_weight(struct task_struct *p, bool update_load)
1263 : {
1264 6 : int prio = p->static_prio - MAX_RT_PRIO;
1265 6 : struct load_weight *load = &p->se.load;
1266 :
1267 : /*
1268 : * SCHED_IDLE tasks get minimal weight:
1269 : */
1270 12 : if (task_has_idle_policy(p)) {
1271 0 : load->weight = scale_load(WEIGHT_IDLEPRIO);
1272 0 : load->inv_weight = WMULT_IDLEPRIO;
1273 0 : return;
1274 : }
1275 :
1276 : /*
1277 : * SCHED_OTHER tasks have to update their load when changing their
1278 : * weight
1279 : */
1280 6 : if (update_load && p->sched_class == &fair_sched_class) {
1281 5 : reweight_task(p, prio);
1282 : } else {
1283 1 : load->weight = scale_load(sched_prio_to_weight[prio]);
1284 1 : load->inv_weight = sched_prio_to_wmult[prio];
1285 : }
1286 : }
1287 :
1288 : #ifdef CONFIG_UCLAMP_TASK
1289 : /*
1290 : * Serializes updates of utilization clamp values
1291 : *
1292 : * The (slow-path) user-space triggers utilization clamp value updates which
1293 : * can require updates on (fast-path) scheduler's data structures used to
1294 : * support enqueue/dequeue operations.
1295 : * While the per-CPU rq lock protects fast-path update operations, user-space
1296 : * requests are serialized using a mutex to reduce the risk of conflicting
1297 : * updates or API abuses.
1298 : */
1299 : static DEFINE_MUTEX(uclamp_mutex);
1300 :
1301 : /* Max allowed minimum utilization */
1302 : static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1303 :
1304 : /* Max allowed maximum utilization */
1305 : static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1306 :
1307 : /*
1308 : * By default RT tasks run at the maximum performance point/capacity of the
1309 : * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1310 : * SCHED_CAPACITY_SCALE.
1311 : *
1312 : * This knob allows admins to change the default behavior when uclamp is being
1313 : * used. In battery powered devices, particularly, running at the maximum
1314 : * capacity and frequency will increase energy consumption and shorten the
1315 : * battery life.
1316 : *
1317 : * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1318 : *
1319 : * This knob will not override the system default sched_util_clamp_min defined
1320 : * above.
1321 : */
1322 : static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1323 :
1324 : /* All clamps are required to be less or equal than these values */
1325 : static struct uclamp_se uclamp_default[UCLAMP_CNT];
1326 :
1327 : /*
1328 : * This static key is used to reduce the uclamp overhead in the fast path. It
1329 : * primarily disables the call to uclamp_rq_{inc, dec}() in
1330 : * enqueue/dequeue_task().
1331 : *
1332 : * This allows users to continue to enable uclamp in their kernel config with
1333 : * minimum uclamp overhead in the fast path.
1334 : *
1335 : * As soon as userspace modifies any of the uclamp knobs, the static key is
1336 : * enabled, since we have an actual users that make use of uclamp
1337 : * functionality.
1338 : *
1339 : * The knobs that would enable this static key are:
1340 : *
1341 : * * A task modifying its uclamp value with sched_setattr().
1342 : * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1343 : * * An admin modifying the cgroup cpu.uclamp.{min, max}
1344 : */
1345 : DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1346 :
1347 : /* Integer rounded range for each bucket */
1348 : #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1349 :
1350 : #define for_each_clamp_id(clamp_id) \
1351 : for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1352 :
1353 : static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1354 : {
1355 : return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1356 : }
1357 :
1358 : static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1359 : {
1360 : if (clamp_id == UCLAMP_MIN)
1361 : return 0;
1362 : return SCHED_CAPACITY_SCALE;
1363 : }
1364 :
1365 : static inline void uclamp_se_set(struct uclamp_se *uc_se,
1366 : unsigned int value, bool user_defined)
1367 : {
1368 : uc_se->value = value;
1369 : uc_se->bucket_id = uclamp_bucket_id(value);
1370 : uc_se->user_defined = user_defined;
1371 : }
1372 :
1373 : static inline unsigned int
1374 : uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1375 : unsigned int clamp_value)
1376 : {
1377 : /*
1378 : * Avoid blocked utilization pushing up the frequency when we go
1379 : * idle (which drops the max-clamp) by retaining the last known
1380 : * max-clamp.
1381 : */
1382 : if (clamp_id == UCLAMP_MAX) {
1383 : rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1384 : return clamp_value;
1385 : }
1386 :
1387 : return uclamp_none(UCLAMP_MIN);
1388 : }
1389 :
1390 : static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1391 : unsigned int clamp_value)
1392 : {
1393 : /* Reset max-clamp retention only on idle exit */
1394 : if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1395 : return;
1396 :
1397 : uclamp_rq_set(rq, clamp_id, clamp_value);
1398 : }
1399 :
1400 : static inline
1401 : unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1402 : unsigned int clamp_value)
1403 : {
1404 : struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1405 : int bucket_id = UCLAMP_BUCKETS - 1;
1406 :
1407 : /*
1408 : * Since both min and max clamps are max aggregated, find the
1409 : * top most bucket with tasks in.
1410 : */
1411 : for ( ; bucket_id >= 0; bucket_id--) {
1412 : if (!bucket[bucket_id].tasks)
1413 : continue;
1414 : return bucket[bucket_id].value;
1415 : }
1416 :
1417 : /* No tasks -- default clamp values */
1418 : return uclamp_idle_value(rq, clamp_id, clamp_value);
1419 : }
1420 :
1421 : static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1422 : {
1423 : unsigned int default_util_min;
1424 : struct uclamp_se *uc_se;
1425 :
1426 : lockdep_assert_held(&p->pi_lock);
1427 :
1428 : uc_se = &p->uclamp_req[UCLAMP_MIN];
1429 :
1430 : /* Only sync if user didn't override the default */
1431 : if (uc_se->user_defined)
1432 : return;
1433 :
1434 : default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1435 : uclamp_se_set(uc_se, default_util_min, false);
1436 : }
1437 :
1438 : static void uclamp_update_util_min_rt_default(struct task_struct *p)
1439 : {
1440 : struct rq_flags rf;
1441 : struct rq *rq;
1442 :
1443 : if (!rt_task(p))
1444 : return;
1445 :
1446 : /* Protect updates to p->uclamp_* */
1447 : rq = task_rq_lock(p, &rf);
1448 : __uclamp_update_util_min_rt_default(p);
1449 : task_rq_unlock(rq, p, &rf);
1450 : }
1451 :
1452 : static inline struct uclamp_se
1453 : uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1454 : {
1455 : /* Copy by value as we could modify it */
1456 : struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1457 : #ifdef CONFIG_UCLAMP_TASK_GROUP
1458 : unsigned int tg_min, tg_max, value;
1459 :
1460 : /*
1461 : * Tasks in autogroups or root task group will be
1462 : * restricted by system defaults.
1463 : */
1464 : if (task_group_is_autogroup(task_group(p)))
1465 : return uc_req;
1466 : if (task_group(p) == &root_task_group)
1467 : return uc_req;
1468 :
1469 : tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1470 : tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1471 : value = uc_req.value;
1472 : value = clamp(value, tg_min, tg_max);
1473 : uclamp_se_set(&uc_req, value, false);
1474 : #endif
1475 :
1476 : return uc_req;
1477 : }
1478 :
1479 : /*
1480 : * The effective clamp bucket index of a task depends on, by increasing
1481 : * priority:
1482 : * - the task specific clamp value, when explicitly requested from userspace
1483 : * - the task group effective clamp value, for tasks not either in the root
1484 : * group or in an autogroup
1485 : * - the system default clamp value, defined by the sysadmin
1486 : */
1487 : static inline struct uclamp_se
1488 : uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1489 : {
1490 : struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1491 : struct uclamp_se uc_max = uclamp_default[clamp_id];
1492 :
1493 : /* System default restrictions always apply */
1494 : if (unlikely(uc_req.value > uc_max.value))
1495 : return uc_max;
1496 :
1497 : return uc_req;
1498 : }
1499 :
1500 : unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1501 : {
1502 : struct uclamp_se uc_eff;
1503 :
1504 : /* Task currently refcounted: use back-annotated (effective) value */
1505 : if (p->uclamp[clamp_id].active)
1506 : return (unsigned long)p->uclamp[clamp_id].value;
1507 :
1508 : uc_eff = uclamp_eff_get(p, clamp_id);
1509 :
1510 : return (unsigned long)uc_eff.value;
1511 : }
1512 :
1513 : /*
1514 : * When a task is enqueued on a rq, the clamp bucket currently defined by the
1515 : * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1516 : * updates the rq's clamp value if required.
1517 : *
1518 : * Tasks can have a task-specific value requested from user-space, track
1519 : * within each bucket the maximum value for tasks refcounted in it.
1520 : * This "local max aggregation" allows to track the exact "requested" value
1521 : * for each bucket when all its RUNNABLE tasks require the same clamp.
1522 : */
1523 : static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1524 : enum uclamp_id clamp_id)
1525 : {
1526 : struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1527 : struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1528 : struct uclamp_bucket *bucket;
1529 :
1530 : lockdep_assert_rq_held(rq);
1531 :
1532 : /* Update task effective clamp */
1533 : p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1534 :
1535 : bucket = &uc_rq->bucket[uc_se->bucket_id];
1536 : bucket->tasks++;
1537 : uc_se->active = true;
1538 :
1539 : uclamp_idle_reset(rq, clamp_id, uc_se->value);
1540 :
1541 : /*
1542 : * Local max aggregation: rq buckets always track the max
1543 : * "requested" clamp value of its RUNNABLE tasks.
1544 : */
1545 : if (bucket->tasks == 1 || uc_se->value > bucket->value)
1546 : bucket->value = uc_se->value;
1547 :
1548 : if (uc_se->value > uclamp_rq_get(rq, clamp_id))
1549 : uclamp_rq_set(rq, clamp_id, uc_se->value);
1550 : }
1551 :
1552 : /*
1553 : * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1554 : * is released. If this is the last task reference counting the rq's max
1555 : * active clamp value, then the rq's clamp value is updated.
1556 : *
1557 : * Both refcounted tasks and rq's cached clamp values are expected to be
1558 : * always valid. If it's detected they are not, as defensive programming,
1559 : * enforce the expected state and warn.
1560 : */
1561 : static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1562 : enum uclamp_id clamp_id)
1563 : {
1564 : struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1565 : struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1566 : struct uclamp_bucket *bucket;
1567 : unsigned int bkt_clamp;
1568 : unsigned int rq_clamp;
1569 :
1570 : lockdep_assert_rq_held(rq);
1571 :
1572 : /*
1573 : * If sched_uclamp_used was enabled after task @p was enqueued,
1574 : * we could end up with unbalanced call to uclamp_rq_dec_id().
1575 : *
1576 : * In this case the uc_se->active flag should be false since no uclamp
1577 : * accounting was performed at enqueue time and we can just return
1578 : * here.
1579 : *
1580 : * Need to be careful of the following enqueue/dequeue ordering
1581 : * problem too
1582 : *
1583 : * enqueue(taskA)
1584 : * // sched_uclamp_used gets enabled
1585 : * enqueue(taskB)
1586 : * dequeue(taskA)
1587 : * // Must not decrement bucket->tasks here
1588 : * dequeue(taskB)
1589 : *
1590 : * where we could end up with stale data in uc_se and
1591 : * bucket[uc_se->bucket_id].
1592 : *
1593 : * The following check here eliminates the possibility of such race.
1594 : */
1595 : if (unlikely(!uc_se->active))
1596 : return;
1597 :
1598 : bucket = &uc_rq->bucket[uc_se->bucket_id];
1599 :
1600 : SCHED_WARN_ON(!bucket->tasks);
1601 : if (likely(bucket->tasks))
1602 : bucket->tasks--;
1603 :
1604 : uc_se->active = false;
1605 :
1606 : /*
1607 : * Keep "local max aggregation" simple and accept to (possibly)
1608 : * overboost some RUNNABLE tasks in the same bucket.
1609 : * The rq clamp bucket value is reset to its base value whenever
1610 : * there are no more RUNNABLE tasks refcounting it.
1611 : */
1612 : if (likely(bucket->tasks))
1613 : return;
1614 :
1615 : rq_clamp = uclamp_rq_get(rq, clamp_id);
1616 : /*
1617 : * Defensive programming: this should never happen. If it happens,
1618 : * e.g. due to future modification, warn and fixup the expected value.
1619 : */
1620 : SCHED_WARN_ON(bucket->value > rq_clamp);
1621 : if (bucket->value >= rq_clamp) {
1622 : bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1623 : uclamp_rq_set(rq, clamp_id, bkt_clamp);
1624 : }
1625 : }
1626 :
1627 : static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1628 : {
1629 : enum uclamp_id clamp_id;
1630 :
1631 : /*
1632 : * Avoid any overhead until uclamp is actually used by the userspace.
1633 : *
1634 : * The condition is constructed such that a NOP is generated when
1635 : * sched_uclamp_used is disabled.
1636 : */
1637 : if (!static_branch_unlikely(&sched_uclamp_used))
1638 : return;
1639 :
1640 : if (unlikely(!p->sched_class->uclamp_enabled))
1641 : return;
1642 :
1643 : for_each_clamp_id(clamp_id)
1644 : uclamp_rq_inc_id(rq, p, clamp_id);
1645 :
1646 : /* Reset clamp idle holding when there is one RUNNABLE task */
1647 : if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1648 : rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1649 : }
1650 :
1651 : static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1652 : {
1653 : enum uclamp_id clamp_id;
1654 :
1655 : /*
1656 : * Avoid any overhead until uclamp is actually used by the userspace.
1657 : *
1658 : * The condition is constructed such that a NOP is generated when
1659 : * sched_uclamp_used is disabled.
1660 : */
1661 : if (!static_branch_unlikely(&sched_uclamp_used))
1662 : return;
1663 :
1664 : if (unlikely(!p->sched_class->uclamp_enabled))
1665 : return;
1666 :
1667 : for_each_clamp_id(clamp_id)
1668 : uclamp_rq_dec_id(rq, p, clamp_id);
1669 : }
1670 :
1671 : static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1672 : enum uclamp_id clamp_id)
1673 : {
1674 : if (!p->uclamp[clamp_id].active)
1675 : return;
1676 :
1677 : uclamp_rq_dec_id(rq, p, clamp_id);
1678 : uclamp_rq_inc_id(rq, p, clamp_id);
1679 :
1680 : /*
1681 : * Make sure to clear the idle flag if we've transiently reached 0
1682 : * active tasks on rq.
1683 : */
1684 : if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1685 : rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1686 : }
1687 :
1688 : static inline void
1689 : uclamp_update_active(struct task_struct *p)
1690 : {
1691 : enum uclamp_id clamp_id;
1692 : struct rq_flags rf;
1693 : struct rq *rq;
1694 :
1695 : /*
1696 : * Lock the task and the rq where the task is (or was) queued.
1697 : *
1698 : * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1699 : * price to pay to safely serialize util_{min,max} updates with
1700 : * enqueues, dequeues and migration operations.
1701 : * This is the same locking schema used by __set_cpus_allowed_ptr().
1702 : */
1703 : rq = task_rq_lock(p, &rf);
1704 :
1705 : /*
1706 : * Setting the clamp bucket is serialized by task_rq_lock().
1707 : * If the task is not yet RUNNABLE and its task_struct is not
1708 : * affecting a valid clamp bucket, the next time it's enqueued,
1709 : * it will already see the updated clamp bucket value.
1710 : */
1711 : for_each_clamp_id(clamp_id)
1712 : uclamp_rq_reinc_id(rq, p, clamp_id);
1713 :
1714 : task_rq_unlock(rq, p, &rf);
1715 : }
1716 :
1717 : #ifdef CONFIG_UCLAMP_TASK_GROUP
1718 : static inline void
1719 : uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1720 : {
1721 : struct css_task_iter it;
1722 : struct task_struct *p;
1723 :
1724 : css_task_iter_start(css, 0, &it);
1725 : while ((p = css_task_iter_next(&it)))
1726 : uclamp_update_active(p);
1727 : css_task_iter_end(&it);
1728 : }
1729 :
1730 : static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1731 : #endif
1732 :
1733 : #ifdef CONFIG_SYSCTL
1734 : #ifdef CONFIG_UCLAMP_TASK
1735 : #ifdef CONFIG_UCLAMP_TASK_GROUP
1736 : static void uclamp_update_root_tg(void)
1737 : {
1738 : struct task_group *tg = &root_task_group;
1739 :
1740 : uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1741 : sysctl_sched_uclamp_util_min, false);
1742 : uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1743 : sysctl_sched_uclamp_util_max, false);
1744 :
1745 : rcu_read_lock();
1746 : cpu_util_update_eff(&root_task_group.css);
1747 : rcu_read_unlock();
1748 : }
1749 : #else
1750 : static void uclamp_update_root_tg(void) { }
1751 : #endif
1752 :
1753 : static void uclamp_sync_util_min_rt_default(void)
1754 : {
1755 : struct task_struct *g, *p;
1756 :
1757 : /*
1758 : * copy_process() sysctl_uclamp
1759 : * uclamp_min_rt = X;
1760 : * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1761 : * // link thread smp_mb__after_spinlock()
1762 : * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1763 : * sched_post_fork() for_each_process_thread()
1764 : * __uclamp_sync_rt() __uclamp_sync_rt()
1765 : *
1766 : * Ensures that either sched_post_fork() will observe the new
1767 : * uclamp_min_rt or for_each_process_thread() will observe the new
1768 : * task.
1769 : */
1770 : read_lock(&tasklist_lock);
1771 : smp_mb__after_spinlock();
1772 : read_unlock(&tasklist_lock);
1773 :
1774 : rcu_read_lock();
1775 : for_each_process_thread(g, p)
1776 : uclamp_update_util_min_rt_default(p);
1777 : rcu_read_unlock();
1778 : }
1779 :
1780 : static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1781 : void *buffer, size_t *lenp, loff_t *ppos)
1782 : {
1783 : bool update_root_tg = false;
1784 : int old_min, old_max, old_min_rt;
1785 : int result;
1786 :
1787 : mutex_lock(&uclamp_mutex);
1788 : old_min = sysctl_sched_uclamp_util_min;
1789 : old_max = sysctl_sched_uclamp_util_max;
1790 : old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1791 :
1792 : result = proc_dointvec(table, write, buffer, lenp, ppos);
1793 : if (result)
1794 : goto undo;
1795 : if (!write)
1796 : goto done;
1797 :
1798 : if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1799 : sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1800 : sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1801 :
1802 : result = -EINVAL;
1803 : goto undo;
1804 : }
1805 :
1806 : if (old_min != sysctl_sched_uclamp_util_min) {
1807 : uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1808 : sysctl_sched_uclamp_util_min, false);
1809 : update_root_tg = true;
1810 : }
1811 : if (old_max != sysctl_sched_uclamp_util_max) {
1812 : uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1813 : sysctl_sched_uclamp_util_max, false);
1814 : update_root_tg = true;
1815 : }
1816 :
1817 : if (update_root_tg) {
1818 : static_branch_enable(&sched_uclamp_used);
1819 : uclamp_update_root_tg();
1820 : }
1821 :
1822 : if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1823 : static_branch_enable(&sched_uclamp_used);
1824 : uclamp_sync_util_min_rt_default();
1825 : }
1826 :
1827 : /*
1828 : * We update all RUNNABLE tasks only when task groups are in use.
1829 : * Otherwise, keep it simple and do just a lazy update at each next
1830 : * task enqueue time.
1831 : */
1832 :
1833 : goto done;
1834 :
1835 : undo:
1836 : sysctl_sched_uclamp_util_min = old_min;
1837 : sysctl_sched_uclamp_util_max = old_max;
1838 : sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1839 : done:
1840 : mutex_unlock(&uclamp_mutex);
1841 :
1842 : return result;
1843 : }
1844 : #endif
1845 : #endif
1846 :
1847 : static int uclamp_validate(struct task_struct *p,
1848 : const struct sched_attr *attr)
1849 : {
1850 : int util_min = p->uclamp_req[UCLAMP_MIN].value;
1851 : int util_max = p->uclamp_req[UCLAMP_MAX].value;
1852 :
1853 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1854 : util_min = attr->sched_util_min;
1855 :
1856 : if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1857 : return -EINVAL;
1858 : }
1859 :
1860 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1861 : util_max = attr->sched_util_max;
1862 :
1863 : if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1864 : return -EINVAL;
1865 : }
1866 :
1867 : if (util_min != -1 && util_max != -1 && util_min > util_max)
1868 : return -EINVAL;
1869 :
1870 : /*
1871 : * We have valid uclamp attributes; make sure uclamp is enabled.
1872 : *
1873 : * We need to do that here, because enabling static branches is a
1874 : * blocking operation which obviously cannot be done while holding
1875 : * scheduler locks.
1876 : */
1877 : static_branch_enable(&sched_uclamp_used);
1878 :
1879 : return 0;
1880 : }
1881 :
1882 : static bool uclamp_reset(const struct sched_attr *attr,
1883 : enum uclamp_id clamp_id,
1884 : struct uclamp_se *uc_se)
1885 : {
1886 : /* Reset on sched class change for a non user-defined clamp value. */
1887 : if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1888 : !uc_se->user_defined)
1889 : return true;
1890 :
1891 : /* Reset on sched_util_{min,max} == -1. */
1892 : if (clamp_id == UCLAMP_MIN &&
1893 : attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1894 : attr->sched_util_min == -1) {
1895 : return true;
1896 : }
1897 :
1898 : if (clamp_id == UCLAMP_MAX &&
1899 : attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1900 : attr->sched_util_max == -1) {
1901 : return true;
1902 : }
1903 :
1904 : return false;
1905 : }
1906 :
1907 : static void __setscheduler_uclamp(struct task_struct *p,
1908 : const struct sched_attr *attr)
1909 : {
1910 : enum uclamp_id clamp_id;
1911 :
1912 : for_each_clamp_id(clamp_id) {
1913 : struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1914 : unsigned int value;
1915 :
1916 : if (!uclamp_reset(attr, clamp_id, uc_se))
1917 : continue;
1918 :
1919 : /*
1920 : * RT by default have a 100% boost value that could be modified
1921 : * at runtime.
1922 : */
1923 : if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1924 : value = sysctl_sched_uclamp_util_min_rt_default;
1925 : else
1926 : value = uclamp_none(clamp_id);
1927 :
1928 : uclamp_se_set(uc_se, value, false);
1929 :
1930 : }
1931 :
1932 : if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1933 : return;
1934 :
1935 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1936 : attr->sched_util_min != -1) {
1937 : uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1938 : attr->sched_util_min, true);
1939 : }
1940 :
1941 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1942 : attr->sched_util_max != -1) {
1943 : uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1944 : attr->sched_util_max, true);
1945 : }
1946 : }
1947 :
1948 : static void uclamp_fork(struct task_struct *p)
1949 : {
1950 : enum uclamp_id clamp_id;
1951 :
1952 : /*
1953 : * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1954 : * as the task is still at its early fork stages.
1955 : */
1956 : for_each_clamp_id(clamp_id)
1957 : p->uclamp[clamp_id].active = false;
1958 :
1959 : if (likely(!p->sched_reset_on_fork))
1960 : return;
1961 :
1962 : for_each_clamp_id(clamp_id) {
1963 : uclamp_se_set(&p->uclamp_req[clamp_id],
1964 : uclamp_none(clamp_id), false);
1965 : }
1966 : }
1967 :
1968 : static void uclamp_post_fork(struct task_struct *p)
1969 : {
1970 : uclamp_update_util_min_rt_default(p);
1971 : }
1972 :
1973 : static void __init init_uclamp_rq(struct rq *rq)
1974 : {
1975 : enum uclamp_id clamp_id;
1976 : struct uclamp_rq *uc_rq = rq->uclamp;
1977 :
1978 : for_each_clamp_id(clamp_id) {
1979 : uc_rq[clamp_id] = (struct uclamp_rq) {
1980 : .value = uclamp_none(clamp_id)
1981 : };
1982 : }
1983 :
1984 : rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1985 : }
1986 :
1987 : static void __init init_uclamp(void)
1988 : {
1989 : struct uclamp_se uc_max = {};
1990 : enum uclamp_id clamp_id;
1991 : int cpu;
1992 :
1993 : for_each_possible_cpu(cpu)
1994 : init_uclamp_rq(cpu_rq(cpu));
1995 :
1996 : for_each_clamp_id(clamp_id) {
1997 : uclamp_se_set(&init_task.uclamp_req[clamp_id],
1998 : uclamp_none(clamp_id), false);
1999 : }
2000 :
2001 : /* System defaults allow max clamp values for both indexes */
2002 : uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2003 : for_each_clamp_id(clamp_id) {
2004 : uclamp_default[clamp_id] = uc_max;
2005 : #ifdef CONFIG_UCLAMP_TASK_GROUP
2006 : root_task_group.uclamp_req[clamp_id] = uc_max;
2007 : root_task_group.uclamp[clamp_id] = uc_max;
2008 : #endif
2009 : }
2010 : }
2011 :
2012 : #else /* CONFIG_UCLAMP_TASK */
2013 : static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2014 : static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2015 : static inline int uclamp_validate(struct task_struct *p,
2016 : const struct sched_attr *attr)
2017 : {
2018 : return -EOPNOTSUPP;
2019 : }
2020 : static void __setscheduler_uclamp(struct task_struct *p,
2021 : const struct sched_attr *attr) { }
2022 : static inline void uclamp_fork(struct task_struct *p) { }
2023 : static inline void uclamp_post_fork(struct task_struct *p) { }
2024 : static inline void init_uclamp(void) { }
2025 : #endif /* CONFIG_UCLAMP_TASK */
2026 :
2027 0 : bool sched_task_on_rq(struct task_struct *p)
2028 : {
2029 0 : return task_on_rq_queued(p);
2030 : }
2031 :
2032 0 : unsigned long get_wchan(struct task_struct *p)
2033 : {
2034 0 : unsigned long ip = 0;
2035 : unsigned int state;
2036 :
2037 0 : if (!p || p == current)
2038 : return 0;
2039 :
2040 : /* Only get wchan if task is blocked and we can keep it that way. */
2041 0 : raw_spin_lock_irq(&p->pi_lock);
2042 0 : state = READ_ONCE(p->__state);
2043 0 : smp_rmb(); /* see try_to_wake_up() */
2044 0 : if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2045 0 : ip = __get_wchan(p);
2046 0 : raw_spin_unlock_irq(&p->pi_lock);
2047 :
2048 0 : return ip;
2049 : }
2050 :
2051 2176 : static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2052 : {
2053 2176 : if (!(flags & ENQUEUE_NOCLOCK))
2054 0 : update_rq_clock(rq);
2055 :
2056 : if (!(flags & ENQUEUE_RESTORE)) {
2057 : sched_info_enqueue(rq, p);
2058 : psi_enqueue(p, (flags & ENQUEUE_WAKEUP) && !(flags & ENQUEUE_MIGRATED));
2059 : }
2060 :
2061 2180 : uclamp_rq_inc(rq, p);
2062 2180 : p->sched_class->enqueue_task(rq, p, flags);
2063 :
2064 2180 : if (sched_core_enabled(rq))
2065 : sched_core_enqueue(rq, p);
2066 2176 : }
2067 :
2068 2174 : static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2069 : {
2070 2178 : if (sched_core_enabled(rq))
2071 : sched_core_dequeue(rq, p, flags);
2072 :
2073 2174 : if (!(flags & DEQUEUE_NOCLOCK))
2074 0 : update_rq_clock(rq);
2075 :
2076 : if (!(flags & DEQUEUE_SAVE)) {
2077 : sched_info_dequeue(rq, p);
2078 : psi_dequeue(p, flags & DEQUEUE_SLEEP);
2079 : }
2080 :
2081 2178 : uclamp_rq_dec(rq, p);
2082 2178 : p->sched_class->dequeue_task(rq, p, flags);
2083 2174 : }
2084 :
2085 0 : void activate_task(struct rq *rq, struct task_struct *p, int flags)
2086 : {
2087 2176 : enqueue_task(rq, p, flags);
2088 :
2089 2176 : p->on_rq = TASK_ON_RQ_QUEUED;
2090 0 : }
2091 :
2092 0 : void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2093 : {
2094 2174 : p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2095 :
2096 2174 : dequeue_task(rq, p, flags);
2097 0 : }
2098 :
2099 : static inline int __normal_prio(int policy, int rt_prio, int nice)
2100 : {
2101 : int prio;
2102 :
2103 5 : if (dl_policy(policy))
2104 : prio = MAX_DL_PRIO - 1;
2105 5 : else if (rt_policy(policy))
2106 0 : prio = MAX_RT_PRIO - 1 - rt_prio;
2107 : else
2108 0 : prio = NICE_TO_PRIO(nice);
2109 :
2110 : return prio;
2111 : }
2112 :
2113 : /*
2114 : * Calculate the expected normal priority: i.e. priority
2115 : * without taking RT-inheritance into account. Might be
2116 : * boosted by interactivity modifiers. Changes upon fork,
2117 : * setprio syscalls, and whenever the interactivity
2118 : * estimator recalculates.
2119 : */
2120 : static inline int normal_prio(struct task_struct *p)
2121 : {
2122 10 : return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2123 : }
2124 :
2125 : /*
2126 : * Calculate the current priority, i.e. the priority
2127 : * taken into account by the scheduler. This value might
2128 : * be boosted by RT tasks, or might be boosted by
2129 : * interactivity modifiers. Will be RT if the task got
2130 : * RT-boosted. If not then it returns p->normal_prio.
2131 : */
2132 : static int effective_prio(struct task_struct *p)
2133 : {
2134 10 : p->normal_prio = normal_prio(p);
2135 : /*
2136 : * If we are RT tasks or we were boosted to RT priority,
2137 : * keep the priority unchanged. Otherwise, update priority
2138 : * to the normal priority:
2139 : */
2140 10 : if (!rt_prio(p->prio))
2141 : return p->normal_prio;
2142 : return p->prio;
2143 : }
2144 :
2145 : /**
2146 : * task_curr - is this task currently executing on a CPU?
2147 : * @p: the task in question.
2148 : *
2149 : * Return: 1 if the task is currently executing. 0 otherwise.
2150 : */
2151 0 : inline int task_curr(const struct task_struct *p)
2152 : {
2153 0 : return cpu_curr(task_cpu(p)) == p;
2154 : }
2155 :
2156 : /*
2157 : * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2158 : * use the balance_callback list if you want balancing.
2159 : *
2160 : * this means any call to check_class_changed() must be followed by a call to
2161 : * balance_callback().
2162 : */
2163 0 : static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2164 : const struct sched_class *prev_class,
2165 : int oldprio)
2166 : {
2167 0 : if (prev_class != p->sched_class) {
2168 0 : if (prev_class->switched_from)
2169 0 : prev_class->switched_from(rq, p);
2170 :
2171 0 : p->sched_class->switched_to(rq, p);
2172 0 : } else if (oldprio != p->prio || dl_task(p))
2173 0 : p->sched_class->prio_changed(rq, p, oldprio);
2174 0 : }
2175 :
2176 2176 : void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2177 : {
2178 2176 : if (p->sched_class == rq->curr->sched_class)
2179 2173 : rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2180 3 : else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2181 3 : resched_curr(rq);
2182 :
2183 : /*
2184 : * A queue event has occurred, and we're going to schedule. In
2185 : * this case, we can save a useless back to back clock update.
2186 : */
2187 4352 : if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2188 1115 : rq_clock_skip_update(rq);
2189 2176 : }
2190 :
2191 : #ifdef CONFIG_SMP
2192 :
2193 : static void
2194 : __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
2195 :
2196 : static int __set_cpus_allowed_ptr(struct task_struct *p,
2197 : struct affinity_context *ctx);
2198 :
2199 : static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2200 : {
2201 : struct affinity_context ac = {
2202 : .new_mask = cpumask_of(rq->cpu),
2203 : .flags = SCA_MIGRATE_DISABLE,
2204 : };
2205 :
2206 : if (likely(!p->migration_disabled))
2207 : return;
2208 :
2209 : if (p->cpus_ptr != &p->cpus_mask)
2210 : return;
2211 :
2212 : /*
2213 : * Violates locking rules! see comment in __do_set_cpus_allowed().
2214 : */
2215 : __do_set_cpus_allowed(p, &ac);
2216 : }
2217 :
2218 : void migrate_disable(void)
2219 : {
2220 : struct task_struct *p = current;
2221 :
2222 : if (p->migration_disabled) {
2223 : p->migration_disabled++;
2224 : return;
2225 : }
2226 :
2227 : preempt_disable();
2228 : this_rq()->nr_pinned++;
2229 : p->migration_disabled = 1;
2230 : preempt_enable();
2231 : }
2232 : EXPORT_SYMBOL_GPL(migrate_disable);
2233 :
2234 : void migrate_enable(void)
2235 : {
2236 : struct task_struct *p = current;
2237 : struct affinity_context ac = {
2238 : .new_mask = &p->cpus_mask,
2239 : .flags = SCA_MIGRATE_ENABLE,
2240 : };
2241 :
2242 : if (p->migration_disabled > 1) {
2243 : p->migration_disabled--;
2244 : return;
2245 : }
2246 :
2247 : if (WARN_ON_ONCE(!p->migration_disabled))
2248 : return;
2249 :
2250 : /*
2251 : * Ensure stop_task runs either before or after this, and that
2252 : * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2253 : */
2254 : preempt_disable();
2255 : if (p->cpus_ptr != &p->cpus_mask)
2256 : __set_cpus_allowed_ptr(p, &ac);
2257 : /*
2258 : * Mustn't clear migration_disabled() until cpus_ptr points back at the
2259 : * regular cpus_mask, otherwise things that race (eg.
2260 : * select_fallback_rq) get confused.
2261 : */
2262 : barrier();
2263 : p->migration_disabled = 0;
2264 : this_rq()->nr_pinned--;
2265 : preempt_enable();
2266 : }
2267 : EXPORT_SYMBOL_GPL(migrate_enable);
2268 :
2269 : static inline bool rq_has_pinned_tasks(struct rq *rq)
2270 : {
2271 : return rq->nr_pinned;
2272 : }
2273 :
2274 : /*
2275 : * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2276 : * __set_cpus_allowed_ptr() and select_fallback_rq().
2277 : */
2278 : static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2279 : {
2280 : /* When not in the task's cpumask, no point in looking further. */
2281 : if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2282 : return false;
2283 :
2284 : /* migrate_disabled() must be allowed to finish. */
2285 : if (is_migration_disabled(p))
2286 : return cpu_online(cpu);
2287 :
2288 : /* Non kernel threads are not allowed during either online or offline. */
2289 : if (!(p->flags & PF_KTHREAD))
2290 : return cpu_active(cpu) && task_cpu_possible(cpu, p);
2291 :
2292 : /* KTHREAD_IS_PER_CPU is always allowed. */
2293 : if (kthread_is_per_cpu(p))
2294 : return cpu_online(cpu);
2295 :
2296 : /* Regular kernel threads don't get to stay during offline. */
2297 : if (cpu_dying(cpu))
2298 : return false;
2299 :
2300 : /* But are allowed during online. */
2301 : return cpu_online(cpu);
2302 : }
2303 :
2304 : /*
2305 : * This is how migration works:
2306 : *
2307 : * 1) we invoke migration_cpu_stop() on the target CPU using
2308 : * stop_one_cpu().
2309 : * 2) stopper starts to run (implicitly forcing the migrated thread
2310 : * off the CPU)
2311 : * 3) it checks whether the migrated task is still in the wrong runqueue.
2312 : * 4) if it's in the wrong runqueue then the migration thread removes
2313 : * it and puts it into the right queue.
2314 : * 5) stopper completes and stop_one_cpu() returns and the migration
2315 : * is done.
2316 : */
2317 :
2318 : /*
2319 : * move_queued_task - move a queued task to new rq.
2320 : *
2321 : * Returns (locked) new rq. Old rq's lock is released.
2322 : */
2323 : static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2324 : struct task_struct *p, int new_cpu)
2325 : {
2326 : lockdep_assert_rq_held(rq);
2327 :
2328 : deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2329 : set_task_cpu(p, new_cpu);
2330 : rq_unlock(rq, rf);
2331 :
2332 : rq = cpu_rq(new_cpu);
2333 :
2334 : rq_lock(rq, rf);
2335 : WARN_ON_ONCE(task_cpu(p) != new_cpu);
2336 : activate_task(rq, p, 0);
2337 : check_preempt_curr(rq, p, 0);
2338 :
2339 : return rq;
2340 : }
2341 :
2342 : struct migration_arg {
2343 : struct task_struct *task;
2344 : int dest_cpu;
2345 : struct set_affinity_pending *pending;
2346 : };
2347 :
2348 : /*
2349 : * @refs: number of wait_for_completion()
2350 : * @stop_pending: is @stop_work in use
2351 : */
2352 : struct set_affinity_pending {
2353 : refcount_t refs;
2354 : unsigned int stop_pending;
2355 : struct completion done;
2356 : struct cpu_stop_work stop_work;
2357 : struct migration_arg arg;
2358 : };
2359 :
2360 : /*
2361 : * Move (not current) task off this CPU, onto the destination CPU. We're doing
2362 : * this because either it can't run here any more (set_cpus_allowed()
2363 : * away from this CPU, or CPU going down), or because we're
2364 : * attempting to rebalance this task on exec (sched_exec).
2365 : *
2366 : * So we race with normal scheduler movements, but that's OK, as long
2367 : * as the task is no longer on this CPU.
2368 : */
2369 : static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2370 : struct task_struct *p, int dest_cpu)
2371 : {
2372 : /* Affinity changed (again). */
2373 : if (!is_cpu_allowed(p, dest_cpu))
2374 : return rq;
2375 :
2376 : update_rq_clock(rq);
2377 : rq = move_queued_task(rq, rf, p, dest_cpu);
2378 :
2379 : return rq;
2380 : }
2381 :
2382 : /*
2383 : * migration_cpu_stop - this will be executed by a highprio stopper thread
2384 : * and performs thread migration by bumping thread off CPU then
2385 : * 'pushing' onto another runqueue.
2386 : */
2387 : static int migration_cpu_stop(void *data)
2388 : {
2389 : struct migration_arg *arg = data;
2390 : struct set_affinity_pending *pending = arg->pending;
2391 : struct task_struct *p = arg->task;
2392 : struct rq *rq = this_rq();
2393 : bool complete = false;
2394 : struct rq_flags rf;
2395 :
2396 : /*
2397 : * The original target CPU might have gone down and we might
2398 : * be on another CPU but it doesn't matter.
2399 : */
2400 : local_irq_save(rf.flags);
2401 : /*
2402 : * We need to explicitly wake pending tasks before running
2403 : * __migrate_task() such that we will not miss enforcing cpus_ptr
2404 : * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2405 : */
2406 : flush_smp_call_function_queue();
2407 :
2408 : raw_spin_lock(&p->pi_lock);
2409 : rq_lock(rq, &rf);
2410 :
2411 : /*
2412 : * If we were passed a pending, then ->stop_pending was set, thus
2413 : * p->migration_pending must have remained stable.
2414 : */
2415 : WARN_ON_ONCE(pending && pending != p->migration_pending);
2416 :
2417 : /*
2418 : * If task_rq(p) != rq, it cannot be migrated here, because we're
2419 : * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2420 : * we're holding p->pi_lock.
2421 : */
2422 : if (task_rq(p) == rq) {
2423 : if (is_migration_disabled(p))
2424 : goto out;
2425 :
2426 : if (pending) {
2427 : p->migration_pending = NULL;
2428 : complete = true;
2429 :
2430 : if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2431 : goto out;
2432 : }
2433 :
2434 : if (task_on_rq_queued(p))
2435 : rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2436 : else
2437 : p->wake_cpu = arg->dest_cpu;
2438 :
2439 : /*
2440 : * XXX __migrate_task() can fail, at which point we might end
2441 : * up running on a dodgy CPU, AFAICT this can only happen
2442 : * during CPU hotplug, at which point we'll get pushed out
2443 : * anyway, so it's probably not a big deal.
2444 : */
2445 :
2446 : } else if (pending) {
2447 : /*
2448 : * This happens when we get migrated between migrate_enable()'s
2449 : * preempt_enable() and scheduling the stopper task. At that
2450 : * point we're a regular task again and not current anymore.
2451 : *
2452 : * A !PREEMPT kernel has a giant hole here, which makes it far
2453 : * more likely.
2454 : */
2455 :
2456 : /*
2457 : * The task moved before the stopper got to run. We're holding
2458 : * ->pi_lock, so the allowed mask is stable - if it got
2459 : * somewhere allowed, we're done.
2460 : */
2461 : if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2462 : p->migration_pending = NULL;
2463 : complete = true;
2464 : goto out;
2465 : }
2466 :
2467 : /*
2468 : * When migrate_enable() hits a rq mis-match we can't reliably
2469 : * determine is_migration_disabled() and so have to chase after
2470 : * it.
2471 : */
2472 : WARN_ON_ONCE(!pending->stop_pending);
2473 : task_rq_unlock(rq, p, &rf);
2474 : stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2475 : &pending->arg, &pending->stop_work);
2476 : return 0;
2477 : }
2478 : out:
2479 : if (pending)
2480 : pending->stop_pending = false;
2481 : task_rq_unlock(rq, p, &rf);
2482 :
2483 : if (complete)
2484 : complete_all(&pending->done);
2485 :
2486 : return 0;
2487 : }
2488 :
2489 : int push_cpu_stop(void *arg)
2490 : {
2491 : struct rq *lowest_rq = NULL, *rq = this_rq();
2492 : struct task_struct *p = arg;
2493 :
2494 : raw_spin_lock_irq(&p->pi_lock);
2495 : raw_spin_rq_lock(rq);
2496 :
2497 : if (task_rq(p) != rq)
2498 : goto out_unlock;
2499 :
2500 : if (is_migration_disabled(p)) {
2501 : p->migration_flags |= MDF_PUSH;
2502 : goto out_unlock;
2503 : }
2504 :
2505 : p->migration_flags &= ~MDF_PUSH;
2506 :
2507 : if (p->sched_class->find_lock_rq)
2508 : lowest_rq = p->sched_class->find_lock_rq(p, rq);
2509 :
2510 : if (!lowest_rq)
2511 : goto out_unlock;
2512 :
2513 : // XXX validate p is still the highest prio task
2514 : if (task_rq(p) == rq) {
2515 : deactivate_task(rq, p, 0);
2516 : set_task_cpu(p, lowest_rq->cpu);
2517 : activate_task(lowest_rq, p, 0);
2518 : resched_curr(lowest_rq);
2519 : }
2520 :
2521 : double_unlock_balance(rq, lowest_rq);
2522 :
2523 : out_unlock:
2524 : rq->push_busy = false;
2525 : raw_spin_rq_unlock(rq);
2526 : raw_spin_unlock_irq(&p->pi_lock);
2527 :
2528 : put_task_struct(p);
2529 : return 0;
2530 : }
2531 :
2532 : /*
2533 : * sched_class::set_cpus_allowed must do the below, but is not required to
2534 : * actually call this function.
2535 : */
2536 : void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
2537 : {
2538 : if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2539 : p->cpus_ptr = ctx->new_mask;
2540 : return;
2541 : }
2542 :
2543 : cpumask_copy(&p->cpus_mask, ctx->new_mask);
2544 : p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
2545 :
2546 : /*
2547 : * Swap in a new user_cpus_ptr if SCA_USER flag set
2548 : */
2549 : if (ctx->flags & SCA_USER)
2550 : swap(p->user_cpus_ptr, ctx->user_mask);
2551 : }
2552 :
2553 : static void
2554 : __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
2555 : {
2556 : struct rq *rq = task_rq(p);
2557 : bool queued, running;
2558 :
2559 : /*
2560 : * This here violates the locking rules for affinity, since we're only
2561 : * supposed to change these variables while holding both rq->lock and
2562 : * p->pi_lock.
2563 : *
2564 : * HOWEVER, it magically works, because ttwu() is the only code that
2565 : * accesses these variables under p->pi_lock and only does so after
2566 : * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2567 : * before finish_task().
2568 : *
2569 : * XXX do further audits, this smells like something putrid.
2570 : */
2571 : if (ctx->flags & SCA_MIGRATE_DISABLE)
2572 : SCHED_WARN_ON(!p->on_cpu);
2573 : else
2574 : lockdep_assert_held(&p->pi_lock);
2575 :
2576 : queued = task_on_rq_queued(p);
2577 : running = task_current(rq, p);
2578 :
2579 : if (queued) {
2580 : /*
2581 : * Because __kthread_bind() calls this on blocked tasks without
2582 : * holding rq->lock.
2583 : */
2584 : lockdep_assert_rq_held(rq);
2585 : dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2586 : }
2587 : if (running)
2588 : put_prev_task(rq, p);
2589 :
2590 : p->sched_class->set_cpus_allowed(p, ctx);
2591 :
2592 : if (queued)
2593 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2594 : if (running)
2595 : set_next_task(rq, p);
2596 : }
2597 :
2598 : /*
2599 : * Used for kthread_bind() and select_fallback_rq(), in both cases the user
2600 : * affinity (if any) should be destroyed too.
2601 : */
2602 : void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2603 : {
2604 : struct affinity_context ac = {
2605 : .new_mask = new_mask,
2606 : .user_mask = NULL,
2607 : .flags = SCA_USER, /* clear the user requested mask */
2608 : };
2609 : union cpumask_rcuhead {
2610 : cpumask_t cpumask;
2611 : struct rcu_head rcu;
2612 : };
2613 :
2614 : __do_set_cpus_allowed(p, &ac);
2615 :
2616 : /*
2617 : * Because this is called with p->pi_lock held, it is not possible
2618 : * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2619 : * kfree_rcu().
2620 : */
2621 : kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
2622 : }
2623 :
2624 : static cpumask_t *alloc_user_cpus_ptr(int node)
2625 : {
2626 : /*
2627 : * See do_set_cpus_allowed() above for the rcu_head usage.
2628 : */
2629 : int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2630 :
2631 : return kmalloc_node(size, GFP_KERNEL, node);
2632 : }
2633 :
2634 : int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2635 : int node)
2636 : {
2637 : cpumask_t *user_mask;
2638 : unsigned long flags;
2639 :
2640 : /*
2641 : * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2642 : * may differ by now due to racing.
2643 : */
2644 : dst->user_cpus_ptr = NULL;
2645 :
2646 : /*
2647 : * This check is racy and losing the race is a valid situation.
2648 : * It is not worth the extra overhead of taking the pi_lock on
2649 : * every fork/clone.
2650 : */
2651 : if (data_race(!src->user_cpus_ptr))
2652 : return 0;
2653 :
2654 : user_mask = alloc_user_cpus_ptr(node);
2655 : if (!user_mask)
2656 : return -ENOMEM;
2657 :
2658 : /*
2659 : * Use pi_lock to protect content of user_cpus_ptr
2660 : *
2661 : * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2662 : * do_set_cpus_allowed().
2663 : */
2664 : raw_spin_lock_irqsave(&src->pi_lock, flags);
2665 : if (src->user_cpus_ptr) {
2666 : swap(dst->user_cpus_ptr, user_mask);
2667 : cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2668 : }
2669 : raw_spin_unlock_irqrestore(&src->pi_lock, flags);
2670 :
2671 : if (unlikely(user_mask))
2672 : kfree(user_mask);
2673 :
2674 : return 0;
2675 : }
2676 :
2677 : static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2678 : {
2679 : struct cpumask *user_mask = NULL;
2680 :
2681 : swap(p->user_cpus_ptr, user_mask);
2682 :
2683 : return user_mask;
2684 : }
2685 :
2686 : void release_user_cpus_ptr(struct task_struct *p)
2687 : {
2688 : kfree(clear_user_cpus_ptr(p));
2689 : }
2690 :
2691 : /*
2692 : * This function is wildly self concurrent; here be dragons.
2693 : *
2694 : *
2695 : * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2696 : * designated task is enqueued on an allowed CPU. If that task is currently
2697 : * running, we have to kick it out using the CPU stopper.
2698 : *
2699 : * Migrate-Disable comes along and tramples all over our nice sandcastle.
2700 : * Consider:
2701 : *
2702 : * Initial conditions: P0->cpus_mask = [0, 1]
2703 : *
2704 : * P0@CPU0 P1
2705 : *
2706 : * migrate_disable();
2707 : * <preempted>
2708 : * set_cpus_allowed_ptr(P0, [1]);
2709 : *
2710 : * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2711 : * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2712 : * This means we need the following scheme:
2713 : *
2714 : * P0@CPU0 P1
2715 : *
2716 : * migrate_disable();
2717 : * <preempted>
2718 : * set_cpus_allowed_ptr(P0, [1]);
2719 : * <blocks>
2720 : * <resumes>
2721 : * migrate_enable();
2722 : * __set_cpus_allowed_ptr();
2723 : * <wakes local stopper>
2724 : * `--> <woken on migration completion>
2725 : *
2726 : * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2727 : * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2728 : * task p are serialized by p->pi_lock, which we can leverage: the one that
2729 : * should come into effect at the end of the Migrate-Disable region is the last
2730 : * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2731 : * but we still need to properly signal those waiting tasks at the appropriate
2732 : * moment.
2733 : *
2734 : * This is implemented using struct set_affinity_pending. The first
2735 : * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2736 : * setup an instance of that struct and install it on the targeted task_struct.
2737 : * Any and all further callers will reuse that instance. Those then wait for
2738 : * a completion signaled at the tail of the CPU stopper callback (1), triggered
2739 : * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2740 : *
2741 : *
2742 : * (1) In the cases covered above. There is one more where the completion is
2743 : * signaled within affine_move_task() itself: when a subsequent affinity request
2744 : * occurs after the stopper bailed out due to the targeted task still being
2745 : * Migrate-Disable. Consider:
2746 : *
2747 : * Initial conditions: P0->cpus_mask = [0, 1]
2748 : *
2749 : * CPU0 P1 P2
2750 : * <P0>
2751 : * migrate_disable();
2752 : * <preempted>
2753 : * set_cpus_allowed_ptr(P0, [1]);
2754 : * <blocks>
2755 : * <migration/0>
2756 : * migration_cpu_stop()
2757 : * is_migration_disabled()
2758 : * <bails>
2759 : * set_cpus_allowed_ptr(P0, [0, 1]);
2760 : * <signal completion>
2761 : * <awakes>
2762 : *
2763 : * Note that the above is safe vs a concurrent migrate_enable(), as any
2764 : * pending affinity completion is preceded by an uninstallation of
2765 : * p->migration_pending done with p->pi_lock held.
2766 : */
2767 : static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2768 : int dest_cpu, unsigned int flags)
2769 : __releases(rq->lock)
2770 : __releases(p->pi_lock)
2771 : {
2772 : struct set_affinity_pending my_pending = { }, *pending = NULL;
2773 : bool stop_pending, complete = false;
2774 :
2775 : /* Can the task run on the task's current CPU? If so, we're done */
2776 : if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2777 : struct task_struct *push_task = NULL;
2778 :
2779 : if ((flags & SCA_MIGRATE_ENABLE) &&
2780 : (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2781 : rq->push_busy = true;
2782 : push_task = get_task_struct(p);
2783 : }
2784 :
2785 : /*
2786 : * If there are pending waiters, but no pending stop_work,
2787 : * then complete now.
2788 : */
2789 : pending = p->migration_pending;
2790 : if (pending && !pending->stop_pending) {
2791 : p->migration_pending = NULL;
2792 : complete = true;
2793 : }
2794 :
2795 : task_rq_unlock(rq, p, rf);
2796 :
2797 : if (push_task) {
2798 : stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2799 : p, &rq->push_work);
2800 : }
2801 :
2802 : if (complete)
2803 : complete_all(&pending->done);
2804 :
2805 : return 0;
2806 : }
2807 :
2808 : if (!(flags & SCA_MIGRATE_ENABLE)) {
2809 : /* serialized by p->pi_lock */
2810 : if (!p->migration_pending) {
2811 : /* Install the request */
2812 : refcount_set(&my_pending.refs, 1);
2813 : init_completion(&my_pending.done);
2814 : my_pending.arg = (struct migration_arg) {
2815 : .task = p,
2816 : .dest_cpu = dest_cpu,
2817 : .pending = &my_pending,
2818 : };
2819 :
2820 : p->migration_pending = &my_pending;
2821 : } else {
2822 : pending = p->migration_pending;
2823 : refcount_inc(&pending->refs);
2824 : /*
2825 : * Affinity has changed, but we've already installed a
2826 : * pending. migration_cpu_stop() *must* see this, else
2827 : * we risk a completion of the pending despite having a
2828 : * task on a disallowed CPU.
2829 : *
2830 : * Serialized by p->pi_lock, so this is safe.
2831 : */
2832 : pending->arg.dest_cpu = dest_cpu;
2833 : }
2834 : }
2835 : pending = p->migration_pending;
2836 : /*
2837 : * - !MIGRATE_ENABLE:
2838 : * we'll have installed a pending if there wasn't one already.
2839 : *
2840 : * - MIGRATE_ENABLE:
2841 : * we're here because the current CPU isn't matching anymore,
2842 : * the only way that can happen is because of a concurrent
2843 : * set_cpus_allowed_ptr() call, which should then still be
2844 : * pending completion.
2845 : *
2846 : * Either way, we really should have a @pending here.
2847 : */
2848 : if (WARN_ON_ONCE(!pending)) {
2849 : task_rq_unlock(rq, p, rf);
2850 : return -EINVAL;
2851 : }
2852 :
2853 : if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2854 : /*
2855 : * MIGRATE_ENABLE gets here because 'p == current', but for
2856 : * anything else we cannot do is_migration_disabled(), punt
2857 : * and have the stopper function handle it all race-free.
2858 : */
2859 : stop_pending = pending->stop_pending;
2860 : if (!stop_pending)
2861 : pending->stop_pending = true;
2862 :
2863 : if (flags & SCA_MIGRATE_ENABLE)
2864 : p->migration_flags &= ~MDF_PUSH;
2865 :
2866 : task_rq_unlock(rq, p, rf);
2867 :
2868 : if (!stop_pending) {
2869 : stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2870 : &pending->arg, &pending->stop_work);
2871 : }
2872 :
2873 : if (flags & SCA_MIGRATE_ENABLE)
2874 : return 0;
2875 : } else {
2876 :
2877 : if (!is_migration_disabled(p)) {
2878 : if (task_on_rq_queued(p))
2879 : rq = move_queued_task(rq, rf, p, dest_cpu);
2880 :
2881 : if (!pending->stop_pending) {
2882 : p->migration_pending = NULL;
2883 : complete = true;
2884 : }
2885 : }
2886 : task_rq_unlock(rq, p, rf);
2887 :
2888 : if (complete)
2889 : complete_all(&pending->done);
2890 : }
2891 :
2892 : wait_for_completion(&pending->done);
2893 :
2894 : if (refcount_dec_and_test(&pending->refs))
2895 : wake_up_var(&pending->refs); /* No UaF, just an address */
2896 :
2897 : /*
2898 : * Block the original owner of &pending until all subsequent callers
2899 : * have seen the completion and decremented the refcount
2900 : */
2901 : wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2902 :
2903 : /* ARGH */
2904 : WARN_ON_ONCE(my_pending.stop_pending);
2905 :
2906 : return 0;
2907 : }
2908 :
2909 : /*
2910 : * Called with both p->pi_lock and rq->lock held; drops both before returning.
2911 : */
2912 : static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2913 : struct affinity_context *ctx,
2914 : struct rq *rq,
2915 : struct rq_flags *rf)
2916 : __releases(rq->lock)
2917 : __releases(p->pi_lock)
2918 : {
2919 : const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2920 : const struct cpumask *cpu_valid_mask = cpu_active_mask;
2921 : bool kthread = p->flags & PF_KTHREAD;
2922 : unsigned int dest_cpu;
2923 : int ret = 0;
2924 :
2925 : update_rq_clock(rq);
2926 :
2927 : if (kthread || is_migration_disabled(p)) {
2928 : /*
2929 : * Kernel threads are allowed on online && !active CPUs,
2930 : * however, during cpu-hot-unplug, even these might get pushed
2931 : * away if not KTHREAD_IS_PER_CPU.
2932 : *
2933 : * Specifically, migration_disabled() tasks must not fail the
2934 : * cpumask_any_and_distribute() pick below, esp. so on
2935 : * SCA_MIGRATE_ENABLE, otherwise we'll not call
2936 : * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2937 : */
2938 : cpu_valid_mask = cpu_online_mask;
2939 : }
2940 :
2941 : if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
2942 : ret = -EINVAL;
2943 : goto out;
2944 : }
2945 :
2946 : /*
2947 : * Must re-check here, to close a race against __kthread_bind(),
2948 : * sched_setaffinity() is not guaranteed to observe the flag.
2949 : */
2950 : if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2951 : ret = -EINVAL;
2952 : goto out;
2953 : }
2954 :
2955 : if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
2956 : if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) {
2957 : if (ctx->flags & SCA_USER)
2958 : swap(p->user_cpus_ptr, ctx->user_mask);
2959 : goto out;
2960 : }
2961 :
2962 : if (WARN_ON_ONCE(p == current &&
2963 : is_migration_disabled(p) &&
2964 : !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
2965 : ret = -EBUSY;
2966 : goto out;
2967 : }
2968 : }
2969 :
2970 : /*
2971 : * Picking a ~random cpu helps in cases where we are changing affinity
2972 : * for groups of tasks (ie. cpuset), so that load balancing is not
2973 : * immediately required to distribute the tasks within their new mask.
2974 : */
2975 : dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
2976 : if (dest_cpu >= nr_cpu_ids) {
2977 : ret = -EINVAL;
2978 : goto out;
2979 : }
2980 :
2981 : __do_set_cpus_allowed(p, ctx);
2982 :
2983 : return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
2984 :
2985 : out:
2986 : task_rq_unlock(rq, p, rf);
2987 :
2988 : return ret;
2989 : }
2990 :
2991 : /*
2992 : * Change a given task's CPU affinity. Migrate the thread to a
2993 : * proper CPU and schedule it away if the CPU it's executing on
2994 : * is removed from the allowed bitmask.
2995 : *
2996 : * NOTE: the caller must have a valid reference to the task, the
2997 : * task must not exit() & deallocate itself prematurely. The
2998 : * call is not atomic; no spinlocks may be held.
2999 : */
3000 : static int __set_cpus_allowed_ptr(struct task_struct *p,
3001 : struct affinity_context *ctx)
3002 : {
3003 : struct rq_flags rf;
3004 : struct rq *rq;
3005 :
3006 : rq = task_rq_lock(p, &rf);
3007 : /*
3008 : * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3009 : * flags are set.
3010 : */
3011 : if (p->user_cpus_ptr &&
3012 : !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
3013 : cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
3014 : ctx->new_mask = rq->scratch_mask;
3015 :
3016 : return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
3017 : }
3018 :
3019 : int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3020 : {
3021 : struct affinity_context ac = {
3022 : .new_mask = new_mask,
3023 : .flags = 0,
3024 : };
3025 :
3026 : return __set_cpus_allowed_ptr(p, &ac);
3027 : }
3028 : EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
3029 :
3030 : /*
3031 : * Change a given task's CPU affinity to the intersection of its current
3032 : * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3033 : * If user_cpus_ptr is defined, use it as the basis for restricting CPU
3034 : * affinity or use cpu_online_mask instead.
3035 : *
3036 : * If the resulting mask is empty, leave the affinity unchanged and return
3037 : * -EINVAL.
3038 : */
3039 : static int restrict_cpus_allowed_ptr(struct task_struct *p,
3040 : struct cpumask *new_mask,
3041 : const struct cpumask *subset_mask)
3042 : {
3043 : struct affinity_context ac = {
3044 : .new_mask = new_mask,
3045 : .flags = 0,
3046 : };
3047 : struct rq_flags rf;
3048 : struct rq *rq;
3049 : int err;
3050 :
3051 : rq = task_rq_lock(p, &rf);
3052 :
3053 : /*
3054 : * Forcefully restricting the affinity of a deadline task is
3055 : * likely to cause problems, so fail and noisily override the
3056 : * mask entirely.
3057 : */
3058 : if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
3059 : err = -EPERM;
3060 : goto err_unlock;
3061 : }
3062 :
3063 : if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
3064 : err = -EINVAL;
3065 : goto err_unlock;
3066 : }
3067 :
3068 : return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
3069 :
3070 : err_unlock:
3071 : task_rq_unlock(rq, p, &rf);
3072 : return err;
3073 : }
3074 :
3075 : /*
3076 : * Restrict the CPU affinity of task @p so that it is a subset of
3077 : * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
3078 : * old affinity mask. If the resulting mask is empty, we warn and walk
3079 : * up the cpuset hierarchy until we find a suitable mask.
3080 : */
3081 : void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3082 : {
3083 : cpumask_var_t new_mask;
3084 : const struct cpumask *override_mask = task_cpu_possible_mask(p);
3085 :
3086 : alloc_cpumask_var(&new_mask, GFP_KERNEL);
3087 :
3088 : /*
3089 : * __migrate_task() can fail silently in the face of concurrent
3090 : * offlining of the chosen destination CPU, so take the hotplug
3091 : * lock to ensure that the migration succeeds.
3092 : */
3093 : cpus_read_lock();
3094 : if (!cpumask_available(new_mask))
3095 : goto out_set_mask;
3096 :
3097 : if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3098 : goto out_free_mask;
3099 :
3100 : /*
3101 : * We failed to find a valid subset of the affinity mask for the
3102 : * task, so override it based on its cpuset hierarchy.
3103 : */
3104 : cpuset_cpus_allowed(p, new_mask);
3105 : override_mask = new_mask;
3106 :
3107 : out_set_mask:
3108 : if (printk_ratelimit()) {
3109 : printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3110 : task_pid_nr(p), p->comm,
3111 : cpumask_pr_args(override_mask));
3112 : }
3113 :
3114 : WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3115 : out_free_mask:
3116 : cpus_read_unlock();
3117 : free_cpumask_var(new_mask);
3118 : }
3119 :
3120 : static int
3121 : __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
3122 :
3123 : /*
3124 : * Restore the affinity of a task @p which was previously restricted by a
3125 : * call to force_compatible_cpus_allowed_ptr().
3126 : *
3127 : * It is the caller's responsibility to serialise this with any calls to
3128 : * force_compatible_cpus_allowed_ptr(@p).
3129 : */
3130 : void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3131 : {
3132 : struct affinity_context ac = {
3133 : .new_mask = task_user_cpus(p),
3134 : .flags = 0,
3135 : };
3136 : int ret;
3137 :
3138 : /*
3139 : * Try to restore the old affinity mask with __sched_setaffinity().
3140 : * Cpuset masking will be done there too.
3141 : */
3142 : ret = __sched_setaffinity(p, &ac);
3143 : WARN_ON_ONCE(ret);
3144 : }
3145 :
3146 : void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3147 : {
3148 : #ifdef CONFIG_SCHED_DEBUG
3149 : unsigned int state = READ_ONCE(p->__state);
3150 :
3151 : /*
3152 : * We should never call set_task_cpu() on a blocked task,
3153 : * ttwu() will sort out the placement.
3154 : */
3155 : WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3156 :
3157 : /*
3158 : * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3159 : * because schedstat_wait_{start,end} rebase migrating task's wait_start
3160 : * time relying on p->on_rq.
3161 : */
3162 : WARN_ON_ONCE(state == TASK_RUNNING &&
3163 : p->sched_class == &fair_sched_class &&
3164 : (p->on_rq && !task_on_rq_migrating(p)));
3165 :
3166 : #ifdef CONFIG_LOCKDEP
3167 : /*
3168 : * The caller should hold either p->pi_lock or rq->lock, when changing
3169 : * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3170 : *
3171 : * sched_move_task() holds both and thus holding either pins the cgroup,
3172 : * see task_group().
3173 : *
3174 : * Furthermore, all task_rq users should acquire both locks, see
3175 : * task_rq_lock().
3176 : */
3177 : WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3178 : lockdep_is_held(__rq_lockp(task_rq(p)))));
3179 : #endif
3180 : /*
3181 : * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3182 : */
3183 : WARN_ON_ONCE(!cpu_online(new_cpu));
3184 :
3185 : WARN_ON_ONCE(is_migration_disabled(p));
3186 : #endif
3187 :
3188 : trace_sched_migrate_task(p, new_cpu);
3189 :
3190 : if (task_cpu(p) != new_cpu) {
3191 : if (p->sched_class->migrate_task_rq)
3192 : p->sched_class->migrate_task_rq(p, new_cpu);
3193 : p->se.nr_migrations++;
3194 : rseq_migrate(p);
3195 : perf_event_task_migrate(p);
3196 : }
3197 :
3198 : __set_task_cpu(p, new_cpu);
3199 : }
3200 :
3201 : #ifdef CONFIG_NUMA_BALANCING
3202 : static void __migrate_swap_task(struct task_struct *p, int cpu)
3203 : {
3204 : if (task_on_rq_queued(p)) {
3205 : struct rq *src_rq, *dst_rq;
3206 : struct rq_flags srf, drf;
3207 :
3208 : src_rq = task_rq(p);
3209 : dst_rq = cpu_rq(cpu);
3210 :
3211 : rq_pin_lock(src_rq, &srf);
3212 : rq_pin_lock(dst_rq, &drf);
3213 :
3214 : deactivate_task(src_rq, p, 0);
3215 : set_task_cpu(p, cpu);
3216 : activate_task(dst_rq, p, 0);
3217 : check_preempt_curr(dst_rq, p, 0);
3218 :
3219 : rq_unpin_lock(dst_rq, &drf);
3220 : rq_unpin_lock(src_rq, &srf);
3221 :
3222 : } else {
3223 : /*
3224 : * Task isn't running anymore; make it appear like we migrated
3225 : * it before it went to sleep. This means on wakeup we make the
3226 : * previous CPU our target instead of where it really is.
3227 : */
3228 : p->wake_cpu = cpu;
3229 : }
3230 : }
3231 :
3232 : struct migration_swap_arg {
3233 : struct task_struct *src_task, *dst_task;
3234 : int src_cpu, dst_cpu;
3235 : };
3236 :
3237 : static int migrate_swap_stop(void *data)
3238 : {
3239 : struct migration_swap_arg *arg = data;
3240 : struct rq *src_rq, *dst_rq;
3241 : int ret = -EAGAIN;
3242 :
3243 : if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3244 : return -EAGAIN;
3245 :
3246 : src_rq = cpu_rq(arg->src_cpu);
3247 : dst_rq = cpu_rq(arg->dst_cpu);
3248 :
3249 : double_raw_lock(&arg->src_task->pi_lock,
3250 : &arg->dst_task->pi_lock);
3251 : double_rq_lock(src_rq, dst_rq);
3252 :
3253 : if (task_cpu(arg->dst_task) != arg->dst_cpu)
3254 : goto unlock;
3255 :
3256 : if (task_cpu(arg->src_task) != arg->src_cpu)
3257 : goto unlock;
3258 :
3259 : if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3260 : goto unlock;
3261 :
3262 : if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3263 : goto unlock;
3264 :
3265 : __migrate_swap_task(arg->src_task, arg->dst_cpu);
3266 : __migrate_swap_task(arg->dst_task, arg->src_cpu);
3267 :
3268 : ret = 0;
3269 :
3270 : unlock:
3271 : double_rq_unlock(src_rq, dst_rq);
3272 : raw_spin_unlock(&arg->dst_task->pi_lock);
3273 : raw_spin_unlock(&arg->src_task->pi_lock);
3274 :
3275 : return ret;
3276 : }
3277 :
3278 : /*
3279 : * Cross migrate two tasks
3280 : */
3281 : int migrate_swap(struct task_struct *cur, struct task_struct *p,
3282 : int target_cpu, int curr_cpu)
3283 : {
3284 : struct migration_swap_arg arg;
3285 : int ret = -EINVAL;
3286 :
3287 : arg = (struct migration_swap_arg){
3288 : .src_task = cur,
3289 : .src_cpu = curr_cpu,
3290 : .dst_task = p,
3291 : .dst_cpu = target_cpu,
3292 : };
3293 :
3294 : if (arg.src_cpu == arg.dst_cpu)
3295 : goto out;
3296 :
3297 : /*
3298 : * These three tests are all lockless; this is OK since all of them
3299 : * will be re-checked with proper locks held further down the line.
3300 : */
3301 : if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3302 : goto out;
3303 :
3304 : if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3305 : goto out;
3306 :
3307 : if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3308 : goto out;
3309 :
3310 : trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3311 : ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3312 :
3313 : out:
3314 : return ret;
3315 : }
3316 : #endif /* CONFIG_NUMA_BALANCING */
3317 :
3318 : /*
3319 : * wait_task_inactive - wait for a thread to unschedule.
3320 : *
3321 : * Wait for the thread to block in any of the states set in @match_state.
3322 : * If it changes, i.e. @p might have woken up, then return zero. When we
3323 : * succeed in waiting for @p to be off its CPU, we return a positive number
3324 : * (its total switch count). If a second call a short while later returns the
3325 : * same number, the caller can be sure that @p has remained unscheduled the
3326 : * whole time.
3327 : *
3328 : * The caller must ensure that the task *will* unschedule sometime soon,
3329 : * else this function might spin for a *long* time. This function can't
3330 : * be called with interrupts off, or it may introduce deadlock with
3331 : * smp_call_function() if an IPI is sent by the same process we are
3332 : * waiting to become inactive.
3333 : */
3334 : unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3335 : {
3336 : int running, queued;
3337 : struct rq_flags rf;
3338 : unsigned long ncsw;
3339 : struct rq *rq;
3340 :
3341 : for (;;) {
3342 : /*
3343 : * We do the initial early heuristics without holding
3344 : * any task-queue locks at all. We'll only try to get
3345 : * the runqueue lock when things look like they will
3346 : * work out!
3347 : */
3348 : rq = task_rq(p);
3349 :
3350 : /*
3351 : * If the task is actively running on another CPU
3352 : * still, just relax and busy-wait without holding
3353 : * any locks.
3354 : *
3355 : * NOTE! Since we don't hold any locks, it's not
3356 : * even sure that "rq" stays as the right runqueue!
3357 : * But we don't care, since "task_on_cpu()" will
3358 : * return false if the runqueue has changed and p
3359 : * is actually now running somewhere else!
3360 : */
3361 : while (task_on_cpu(rq, p)) {
3362 : if (!(READ_ONCE(p->__state) & match_state))
3363 : return 0;
3364 : cpu_relax();
3365 : }
3366 :
3367 : /*
3368 : * Ok, time to look more closely! We need the rq
3369 : * lock now, to be *sure*. If we're wrong, we'll
3370 : * just go back and repeat.
3371 : */
3372 : rq = task_rq_lock(p, &rf);
3373 : trace_sched_wait_task(p);
3374 : running = task_on_cpu(rq, p);
3375 : queued = task_on_rq_queued(p);
3376 : ncsw = 0;
3377 : if (READ_ONCE(p->__state) & match_state)
3378 : ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3379 : task_rq_unlock(rq, p, &rf);
3380 :
3381 : /*
3382 : * If it changed from the expected state, bail out now.
3383 : */
3384 : if (unlikely(!ncsw))
3385 : break;
3386 :
3387 : /*
3388 : * Was it really running after all now that we
3389 : * checked with the proper locks actually held?
3390 : *
3391 : * Oops. Go back and try again..
3392 : */
3393 : if (unlikely(running)) {
3394 : cpu_relax();
3395 : continue;
3396 : }
3397 :
3398 : /*
3399 : * It's not enough that it's not actively running,
3400 : * it must be off the runqueue _entirely_, and not
3401 : * preempted!
3402 : *
3403 : * So if it was still runnable (but just not actively
3404 : * running right now), it's preempted, and we should
3405 : * yield - it could be a while.
3406 : */
3407 : if (unlikely(queued)) {
3408 : ktime_t to = NSEC_PER_SEC / HZ;
3409 :
3410 : set_current_state(TASK_UNINTERRUPTIBLE);
3411 : schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3412 : continue;
3413 : }
3414 :
3415 : /*
3416 : * Ahh, all good. It wasn't running, and it wasn't
3417 : * runnable, which means that it will never become
3418 : * running in the future either. We're all done!
3419 : */
3420 : break;
3421 : }
3422 :
3423 : return ncsw;
3424 : }
3425 :
3426 : /***
3427 : * kick_process - kick a running thread to enter/exit the kernel
3428 : * @p: the to-be-kicked thread
3429 : *
3430 : * Cause a process which is running on another CPU to enter
3431 : * kernel-mode, without any delay. (to get signals handled.)
3432 : *
3433 : * NOTE: this function doesn't have to take the runqueue lock,
3434 : * because all it wants to ensure is that the remote task enters
3435 : * the kernel. If the IPI races and the task has been migrated
3436 : * to another CPU then no harm is done and the purpose has been
3437 : * achieved as well.
3438 : */
3439 : void kick_process(struct task_struct *p)
3440 : {
3441 : int cpu;
3442 :
3443 : preempt_disable();
3444 : cpu = task_cpu(p);
3445 : if ((cpu != smp_processor_id()) && task_curr(p))
3446 : smp_send_reschedule(cpu);
3447 : preempt_enable();
3448 : }
3449 : EXPORT_SYMBOL_GPL(kick_process);
3450 :
3451 : /*
3452 : * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3453 : *
3454 : * A few notes on cpu_active vs cpu_online:
3455 : *
3456 : * - cpu_active must be a subset of cpu_online
3457 : *
3458 : * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3459 : * see __set_cpus_allowed_ptr(). At this point the newly online
3460 : * CPU isn't yet part of the sched domains, and balancing will not
3461 : * see it.
3462 : *
3463 : * - on CPU-down we clear cpu_active() to mask the sched domains and
3464 : * avoid the load balancer to place new tasks on the to be removed
3465 : * CPU. Existing tasks will remain running there and will be taken
3466 : * off.
3467 : *
3468 : * This means that fallback selection must not select !active CPUs.
3469 : * And can assume that any active CPU must be online. Conversely
3470 : * select_task_rq() below may allow selection of !active CPUs in order
3471 : * to satisfy the above rules.
3472 : */
3473 : static int select_fallback_rq(int cpu, struct task_struct *p)
3474 : {
3475 : int nid = cpu_to_node(cpu);
3476 : const struct cpumask *nodemask = NULL;
3477 : enum { cpuset, possible, fail } state = cpuset;
3478 : int dest_cpu;
3479 :
3480 : /*
3481 : * If the node that the CPU is on has been offlined, cpu_to_node()
3482 : * will return -1. There is no CPU on the node, and we should
3483 : * select the CPU on the other node.
3484 : */
3485 : if (nid != -1) {
3486 : nodemask = cpumask_of_node(nid);
3487 :
3488 : /* Look for allowed, online CPU in same node. */
3489 : for_each_cpu(dest_cpu, nodemask) {
3490 : if (is_cpu_allowed(p, dest_cpu))
3491 : return dest_cpu;
3492 : }
3493 : }
3494 :
3495 : for (;;) {
3496 : /* Any allowed, online CPU? */
3497 : for_each_cpu(dest_cpu, p->cpus_ptr) {
3498 : if (!is_cpu_allowed(p, dest_cpu))
3499 : continue;
3500 :
3501 : goto out;
3502 : }
3503 :
3504 : /* No more Mr. Nice Guy. */
3505 : switch (state) {
3506 : case cpuset:
3507 : if (cpuset_cpus_allowed_fallback(p)) {
3508 : state = possible;
3509 : break;
3510 : }
3511 : fallthrough;
3512 : case possible:
3513 : /*
3514 : * XXX When called from select_task_rq() we only
3515 : * hold p->pi_lock and again violate locking order.
3516 : *
3517 : * More yuck to audit.
3518 : */
3519 : do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3520 : state = fail;
3521 : break;
3522 : case fail:
3523 : BUG();
3524 : break;
3525 : }
3526 : }
3527 :
3528 : out:
3529 : if (state != cpuset) {
3530 : /*
3531 : * Don't tell them about moving exiting tasks or
3532 : * kernel threads (both mm NULL), since they never
3533 : * leave kernel.
3534 : */
3535 : if (p->mm && printk_ratelimit()) {
3536 : printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3537 : task_pid_nr(p), p->comm, cpu);
3538 : }
3539 : }
3540 :
3541 : return dest_cpu;
3542 : }
3543 :
3544 : /*
3545 : * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3546 : */
3547 : static inline
3548 : int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3549 : {
3550 : lockdep_assert_held(&p->pi_lock);
3551 :
3552 : if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3553 : cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3554 : else
3555 : cpu = cpumask_any(p->cpus_ptr);
3556 :
3557 : /*
3558 : * In order not to call set_task_cpu() on a blocking task we need
3559 : * to rely on ttwu() to place the task on a valid ->cpus_ptr
3560 : * CPU.
3561 : *
3562 : * Since this is common to all placement strategies, this lives here.
3563 : *
3564 : * [ this allows ->select_task() to simply return task_cpu(p) and
3565 : * not worry about this generic constraint ]
3566 : */
3567 : if (unlikely(!is_cpu_allowed(p, cpu)))
3568 : cpu = select_fallback_rq(task_cpu(p), p);
3569 :
3570 : return cpu;
3571 : }
3572 :
3573 : void sched_set_stop_task(int cpu, struct task_struct *stop)
3574 : {
3575 : static struct lock_class_key stop_pi_lock;
3576 : struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3577 : struct task_struct *old_stop = cpu_rq(cpu)->stop;
3578 :
3579 : if (stop) {
3580 : /*
3581 : * Make it appear like a SCHED_FIFO task, its something
3582 : * userspace knows about and won't get confused about.
3583 : *
3584 : * Also, it will make PI more or less work without too
3585 : * much confusion -- but then, stop work should not
3586 : * rely on PI working anyway.
3587 : */
3588 : sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3589 :
3590 : stop->sched_class = &stop_sched_class;
3591 :
3592 : /*
3593 : * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3594 : * adjust the effective priority of a task. As a result,
3595 : * rt_mutex_setprio() can trigger (RT) balancing operations,
3596 : * which can then trigger wakeups of the stop thread to push
3597 : * around the current task.
3598 : *
3599 : * The stop task itself will never be part of the PI-chain, it
3600 : * never blocks, therefore that ->pi_lock recursion is safe.
3601 : * Tell lockdep about this by placing the stop->pi_lock in its
3602 : * own class.
3603 : */
3604 : lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3605 : }
3606 :
3607 : cpu_rq(cpu)->stop = stop;
3608 :
3609 : if (old_stop) {
3610 : /*
3611 : * Reset it back to a normal scheduling class so that
3612 : * it can die in pieces.
3613 : */
3614 : old_stop->sched_class = &rt_sched_class;
3615 : }
3616 : }
3617 :
3618 : #else /* CONFIG_SMP */
3619 :
3620 : static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3621 : struct affinity_context *ctx)
3622 : {
3623 0 : return set_cpus_allowed_ptr(p, ctx->new_mask);
3624 : }
3625 :
3626 : static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3627 :
3628 : static inline bool rq_has_pinned_tasks(struct rq *rq)
3629 : {
3630 : return false;
3631 : }
3632 :
3633 : static inline cpumask_t *alloc_user_cpus_ptr(int node)
3634 : {
3635 : return NULL;
3636 : }
3637 :
3638 : #endif /* !CONFIG_SMP */
3639 :
3640 : static void
3641 : ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3642 : {
3643 : struct rq *rq;
3644 :
3645 : if (!schedstat_enabled())
3646 : return;
3647 :
3648 : rq = this_rq();
3649 :
3650 : #ifdef CONFIG_SMP
3651 : if (cpu == rq->cpu) {
3652 : __schedstat_inc(rq->ttwu_local);
3653 : __schedstat_inc(p->stats.nr_wakeups_local);
3654 : } else {
3655 : struct sched_domain *sd;
3656 :
3657 : __schedstat_inc(p->stats.nr_wakeups_remote);
3658 : rcu_read_lock();
3659 : for_each_domain(rq->cpu, sd) {
3660 : if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3661 : __schedstat_inc(sd->ttwu_wake_remote);
3662 : break;
3663 : }
3664 : }
3665 : rcu_read_unlock();
3666 : }
3667 :
3668 : if (wake_flags & WF_MIGRATED)
3669 : __schedstat_inc(p->stats.nr_wakeups_migrate);
3670 : #endif /* CONFIG_SMP */
3671 :
3672 : __schedstat_inc(rq->ttwu_count);
3673 : __schedstat_inc(p->stats.nr_wakeups);
3674 :
3675 : if (wake_flags & WF_SYNC)
3676 : __schedstat_inc(p->stats.nr_wakeups_sync);
3677 : }
3678 :
3679 : /*
3680 : * Mark the task runnable.
3681 : */
3682 : static inline void ttwu_do_wakeup(struct task_struct *p)
3683 : {
3684 1828 : WRITE_ONCE(p->__state, TASK_RUNNING);
3685 1828 : trace_sched_wakeup(p);
3686 : }
3687 :
3688 : static void
3689 1828 : ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3690 : struct rq_flags *rf)
3691 : {
3692 1828 : int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3693 :
3694 1828 : lockdep_assert_rq_held(rq);
3695 :
3696 1828 : if (p->sched_contributes_to_load)
3697 1029 : rq->nr_uninterruptible--;
3698 :
3699 : #ifdef CONFIG_SMP
3700 : if (wake_flags & WF_MIGRATED)
3701 : en_flags |= ENQUEUE_MIGRATED;
3702 : else
3703 : #endif
3704 1828 : if (p->in_iowait) {
3705 0 : delayacct_blkio_end(p);
3706 0 : atomic_dec(&task_rq(p)->nr_iowait);
3707 : }
3708 :
3709 1828 : activate_task(rq, p, en_flags);
3710 1828 : check_preempt_curr(rq, p, wake_flags);
3711 :
3712 1828 : ttwu_do_wakeup(p);
3713 :
3714 : #ifdef CONFIG_SMP
3715 : if (p->sched_class->task_woken) {
3716 : /*
3717 : * Our task @p is fully woken up and running; so it's safe to
3718 : * drop the rq->lock, hereafter rq is only used for statistics.
3719 : */
3720 : rq_unpin_lock(rq, rf);
3721 : p->sched_class->task_woken(rq, p);
3722 : rq_repin_lock(rq, rf);
3723 : }
3724 :
3725 : if (rq->idle_stamp) {
3726 : u64 delta = rq_clock(rq) - rq->idle_stamp;
3727 : u64 max = 2*rq->max_idle_balance_cost;
3728 :
3729 : update_avg(&rq->avg_idle, delta);
3730 :
3731 : if (rq->avg_idle > max)
3732 : rq->avg_idle = max;
3733 :
3734 : rq->wake_stamp = jiffies;
3735 : rq->wake_avg_idle = rq->avg_idle / 2;
3736 :
3737 : rq->idle_stamp = 0;
3738 : }
3739 : #endif
3740 1828 : }
3741 :
3742 : /*
3743 : * Consider @p being inside a wait loop:
3744 : *
3745 : * for (;;) {
3746 : * set_current_state(TASK_UNINTERRUPTIBLE);
3747 : *
3748 : * if (CONDITION)
3749 : * break;
3750 : *
3751 : * schedule();
3752 : * }
3753 : * __set_current_state(TASK_RUNNING);
3754 : *
3755 : * between set_current_state() and schedule(). In this case @p is still
3756 : * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3757 : * an atomic manner.
3758 : *
3759 : * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3760 : * then schedule() must still happen and p->state can be changed to
3761 : * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3762 : * need to do a full wakeup with enqueue.
3763 : *
3764 : * Returns: %true when the wakeup is done,
3765 : * %false otherwise.
3766 : */
3767 0 : static int ttwu_runnable(struct task_struct *p, int wake_flags)
3768 : {
3769 : struct rq_flags rf;
3770 : struct rq *rq;
3771 0 : int ret = 0;
3772 :
3773 0 : rq = __task_rq_lock(p, &rf);
3774 0 : if (task_on_rq_queued(p)) {
3775 0 : if (!task_on_cpu(rq, p)) {
3776 : /*
3777 : * When on_rq && !on_cpu the task is preempted, see if
3778 : * it should preempt the task that is current now.
3779 : */
3780 0 : update_rq_clock(rq);
3781 0 : check_preempt_curr(rq, p, wake_flags);
3782 : }
3783 0 : ttwu_do_wakeup(p);
3784 0 : ret = 1;
3785 : }
3786 0 : __task_rq_unlock(rq, &rf);
3787 :
3788 0 : return ret;
3789 : }
3790 :
3791 : #ifdef CONFIG_SMP
3792 : void sched_ttwu_pending(void *arg)
3793 : {
3794 : struct llist_node *llist = arg;
3795 : struct rq *rq = this_rq();
3796 : struct task_struct *p, *t;
3797 : struct rq_flags rf;
3798 :
3799 : if (!llist)
3800 : return;
3801 :
3802 : rq_lock_irqsave(rq, &rf);
3803 : update_rq_clock(rq);
3804 :
3805 : llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3806 : if (WARN_ON_ONCE(p->on_cpu))
3807 : smp_cond_load_acquire(&p->on_cpu, !VAL);
3808 :
3809 : if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3810 : set_task_cpu(p, cpu_of(rq));
3811 :
3812 : ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3813 : }
3814 :
3815 : /*
3816 : * Must be after enqueueing at least once task such that
3817 : * idle_cpu() does not observe a false-negative -- if it does,
3818 : * it is possible for select_idle_siblings() to stack a number
3819 : * of tasks on this CPU during that window.
3820 : *
3821 : * It is ok to clear ttwu_pending when another task pending.
3822 : * We will receive IPI after local irq enabled and then enqueue it.
3823 : * Since now nr_running > 0, idle_cpu() will always get correct result.
3824 : */
3825 : WRITE_ONCE(rq->ttwu_pending, 0);
3826 : rq_unlock_irqrestore(rq, &rf);
3827 : }
3828 :
3829 : void send_call_function_single_ipi(int cpu)
3830 : {
3831 : struct rq *rq = cpu_rq(cpu);
3832 :
3833 : if (!set_nr_if_polling(rq->idle))
3834 : arch_send_call_function_single_ipi(cpu);
3835 : else
3836 : trace_sched_wake_idle_without_ipi(cpu);
3837 : }
3838 :
3839 : /*
3840 : * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3841 : * necessary. The wakee CPU on receipt of the IPI will queue the task
3842 : * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3843 : * of the wakeup instead of the waker.
3844 : */
3845 : static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3846 : {
3847 : struct rq *rq = cpu_rq(cpu);
3848 :
3849 : p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3850 :
3851 : WRITE_ONCE(rq->ttwu_pending, 1);
3852 : __smp_call_single_queue(cpu, &p->wake_entry.llist);
3853 : }
3854 :
3855 : void wake_up_if_idle(int cpu)
3856 : {
3857 : struct rq *rq = cpu_rq(cpu);
3858 : struct rq_flags rf;
3859 :
3860 : rcu_read_lock();
3861 :
3862 : if (!is_idle_task(rcu_dereference(rq->curr)))
3863 : goto out;
3864 :
3865 : rq_lock_irqsave(rq, &rf);
3866 : if (is_idle_task(rq->curr))
3867 : resched_curr(rq);
3868 : /* Else CPU is not idle, do nothing here: */
3869 : rq_unlock_irqrestore(rq, &rf);
3870 :
3871 : out:
3872 : rcu_read_unlock();
3873 : }
3874 :
3875 : bool cpus_share_cache(int this_cpu, int that_cpu)
3876 : {
3877 : if (this_cpu == that_cpu)
3878 : return true;
3879 :
3880 : return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3881 : }
3882 :
3883 : static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3884 : {
3885 : /*
3886 : * Do not complicate things with the async wake_list while the CPU is
3887 : * in hotplug state.
3888 : */
3889 : if (!cpu_active(cpu))
3890 : return false;
3891 :
3892 : /* Ensure the task will still be allowed to run on the CPU. */
3893 : if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3894 : return false;
3895 :
3896 : /*
3897 : * If the CPU does not share cache, then queue the task on the
3898 : * remote rqs wakelist to avoid accessing remote data.
3899 : */
3900 : if (!cpus_share_cache(smp_processor_id(), cpu))
3901 : return true;
3902 :
3903 : if (cpu == smp_processor_id())
3904 : return false;
3905 :
3906 : /*
3907 : * If the wakee cpu is idle, or the task is descheduling and the
3908 : * only running task on the CPU, then use the wakelist to offload
3909 : * the task activation to the idle (or soon-to-be-idle) CPU as
3910 : * the current CPU is likely busy. nr_running is checked to
3911 : * avoid unnecessary task stacking.
3912 : *
3913 : * Note that we can only get here with (wakee) p->on_rq=0,
3914 : * p->on_cpu can be whatever, we've done the dequeue, so
3915 : * the wakee has been accounted out of ->nr_running.
3916 : */
3917 : if (!cpu_rq(cpu)->nr_running)
3918 : return true;
3919 :
3920 : return false;
3921 : }
3922 :
3923 : static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3924 : {
3925 : if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
3926 : sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3927 : __ttwu_queue_wakelist(p, cpu, wake_flags);
3928 : return true;
3929 : }
3930 :
3931 : return false;
3932 : }
3933 :
3934 : #else /* !CONFIG_SMP */
3935 :
3936 : static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3937 : {
3938 : return false;
3939 : }
3940 :
3941 : #endif /* CONFIG_SMP */
3942 :
3943 1828 : static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3944 : {
3945 1828 : struct rq *rq = cpu_rq(cpu);
3946 : struct rq_flags rf;
3947 :
3948 1828 : if (ttwu_queue_wakelist(p, cpu, wake_flags))
3949 : return;
3950 :
3951 1828 : rq_lock(rq, &rf);
3952 1828 : update_rq_clock(rq);
3953 1828 : ttwu_do_activate(rq, p, wake_flags, &rf);
3954 1828 : rq_unlock(rq, &rf);
3955 : }
3956 :
3957 : /*
3958 : * Invoked from try_to_wake_up() to check whether the task can be woken up.
3959 : *
3960 : * The caller holds p::pi_lock if p != current or has preemption
3961 : * disabled when p == current.
3962 : *
3963 : * The rules of PREEMPT_RT saved_state:
3964 : *
3965 : * The related locking code always holds p::pi_lock when updating
3966 : * p::saved_state, which means the code is fully serialized in both cases.
3967 : *
3968 : * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3969 : * bits set. This allows to distinguish all wakeup scenarios.
3970 : */
3971 : static __always_inline
3972 : bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3973 : {
3974 : if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3975 : WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3976 : state != TASK_RTLOCK_WAIT);
3977 : }
3978 :
3979 2199 : if (READ_ONCE(p->__state) & state) {
3980 1828 : *success = 1;
3981 : return true;
3982 : }
3983 :
3984 : #ifdef CONFIG_PREEMPT_RT
3985 : /*
3986 : * Saved state preserves the task state across blocking on
3987 : * an RT lock. If the state matches, set p::saved_state to
3988 : * TASK_RUNNING, but do not wake the task because it waits
3989 : * for a lock wakeup. Also indicate success because from
3990 : * the regular waker's point of view this has succeeded.
3991 : *
3992 : * After acquiring the lock the task will restore p::__state
3993 : * from p::saved_state which ensures that the regular
3994 : * wakeup is not lost. The restore will also set
3995 : * p::saved_state to TASK_RUNNING so any further tests will
3996 : * not result in false positives vs. @success
3997 : */
3998 : if (p->saved_state & state) {
3999 : p->saved_state = TASK_RUNNING;
4000 : *success = 1;
4001 : }
4002 : #endif
4003 : return false;
4004 : }
4005 :
4006 : /*
4007 : * Notes on Program-Order guarantees on SMP systems.
4008 : *
4009 : * MIGRATION
4010 : *
4011 : * The basic program-order guarantee on SMP systems is that when a task [t]
4012 : * migrates, all its activity on its old CPU [c0] happens-before any subsequent
4013 : * execution on its new CPU [c1].
4014 : *
4015 : * For migration (of runnable tasks) this is provided by the following means:
4016 : *
4017 : * A) UNLOCK of the rq(c0)->lock scheduling out task t
4018 : * B) migration for t is required to synchronize *both* rq(c0)->lock and
4019 : * rq(c1)->lock (if not at the same time, then in that order).
4020 : * C) LOCK of the rq(c1)->lock scheduling in task
4021 : *
4022 : * Release/acquire chaining guarantees that B happens after A and C after B.
4023 : * Note: the CPU doing B need not be c0 or c1
4024 : *
4025 : * Example:
4026 : *
4027 : * CPU0 CPU1 CPU2
4028 : *
4029 : * LOCK rq(0)->lock
4030 : * sched-out X
4031 : * sched-in Y
4032 : * UNLOCK rq(0)->lock
4033 : *
4034 : * LOCK rq(0)->lock // orders against CPU0
4035 : * dequeue X
4036 : * UNLOCK rq(0)->lock
4037 : *
4038 : * LOCK rq(1)->lock
4039 : * enqueue X
4040 : * UNLOCK rq(1)->lock
4041 : *
4042 : * LOCK rq(1)->lock // orders against CPU2
4043 : * sched-out Z
4044 : * sched-in X
4045 : * UNLOCK rq(1)->lock
4046 : *
4047 : *
4048 : * BLOCKING -- aka. SLEEP + WAKEUP
4049 : *
4050 : * For blocking we (obviously) need to provide the same guarantee as for
4051 : * migration. However the means are completely different as there is no lock
4052 : * chain to provide order. Instead we do:
4053 : *
4054 : * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
4055 : * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
4056 : *
4057 : * Example:
4058 : *
4059 : * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
4060 : *
4061 : * LOCK rq(0)->lock LOCK X->pi_lock
4062 : * dequeue X
4063 : * sched-out X
4064 : * smp_store_release(X->on_cpu, 0);
4065 : *
4066 : * smp_cond_load_acquire(&X->on_cpu, !VAL);
4067 : * X->state = WAKING
4068 : * set_task_cpu(X,2)
4069 : *
4070 : * LOCK rq(2)->lock
4071 : * enqueue X
4072 : * X->state = RUNNING
4073 : * UNLOCK rq(2)->lock
4074 : *
4075 : * LOCK rq(2)->lock // orders against CPU1
4076 : * sched-out Z
4077 : * sched-in X
4078 : * UNLOCK rq(2)->lock
4079 : *
4080 : * UNLOCK X->pi_lock
4081 : * UNLOCK rq(0)->lock
4082 : *
4083 : *
4084 : * However, for wakeups there is a second guarantee we must provide, namely we
4085 : * must ensure that CONDITION=1 done by the caller can not be reordered with
4086 : * accesses to the task state; see try_to_wake_up() and set_current_state().
4087 : */
4088 :
4089 : /**
4090 : * try_to_wake_up - wake up a thread
4091 : * @p: the thread to be awakened
4092 : * @state: the mask of task states that can be woken
4093 : * @wake_flags: wake modifier flags (WF_*)
4094 : *
4095 : * Conceptually does:
4096 : *
4097 : * If (@state & @p->state) @p->state = TASK_RUNNING.
4098 : *
4099 : * If the task was not queued/runnable, also place it back on a runqueue.
4100 : *
4101 : * This function is atomic against schedule() which would dequeue the task.
4102 : *
4103 : * It issues a full memory barrier before accessing @p->state, see the comment
4104 : * with set_current_state().
4105 : *
4106 : * Uses p->pi_lock to serialize against concurrent wake-ups.
4107 : *
4108 : * Relies on p->pi_lock stabilizing:
4109 : * - p->sched_class
4110 : * - p->cpus_ptr
4111 : * - p->sched_task_group
4112 : * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4113 : *
4114 : * Tries really hard to only take one task_rq(p)->lock for performance.
4115 : * Takes rq->lock in:
4116 : * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4117 : * - ttwu_queue() -- new rq, for enqueue of the task;
4118 : * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4119 : *
4120 : * As a consequence we race really badly with just about everything. See the
4121 : * many memory barriers and their comments for details.
4122 : *
4123 : * Return: %true if @p->state changes (an actual wakeup was done),
4124 : * %false otherwise.
4125 : */
4126 : static int
4127 2199 : try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4128 : {
4129 : unsigned long flags;
4130 2199 : int cpu, success = 0;
4131 :
4132 2199 : preempt_disable();
4133 2199 : if (p == current) {
4134 : /*
4135 : * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4136 : * == smp_processor_id()'. Together this means we can special
4137 : * case the whole 'p->on_rq && ttwu_runnable()' case below
4138 : * without taking any locks.
4139 : *
4140 : * In particular:
4141 : * - we rely on Program-Order guarantees for all the ordering,
4142 : * - we're serialized against set_special_state() by virtue of
4143 : * it disabling IRQs (this allows not taking ->pi_lock).
4144 : */
4145 35 : if (!ttwu_state_match(p, state, &success))
4146 : goto out;
4147 :
4148 0 : trace_sched_waking(p);
4149 : ttwu_do_wakeup(p);
4150 : goto out;
4151 : }
4152 :
4153 : /*
4154 : * If we are going to wake up a thread waiting for CONDITION we
4155 : * need to ensure that CONDITION=1 done by the caller can not be
4156 : * reordered with p->state check below. This pairs with smp_store_mb()
4157 : * in set_current_state() that the waiting thread does.
4158 : */
4159 2164 : raw_spin_lock_irqsave(&p->pi_lock, flags);
4160 : smp_mb__after_spinlock();
4161 2164 : if (!ttwu_state_match(p, state, &success))
4162 : goto unlock;
4163 :
4164 1828 : trace_sched_waking(p);
4165 :
4166 : /*
4167 : * Ensure we load p->on_rq _after_ p->state, otherwise it would
4168 : * be possible to, falsely, observe p->on_rq == 0 and get stuck
4169 : * in smp_cond_load_acquire() below.
4170 : *
4171 : * sched_ttwu_pending() try_to_wake_up()
4172 : * STORE p->on_rq = 1 LOAD p->state
4173 : * UNLOCK rq->lock
4174 : *
4175 : * __schedule() (switch to task 'p')
4176 : * LOCK rq->lock smp_rmb();
4177 : * smp_mb__after_spinlock();
4178 : * UNLOCK rq->lock
4179 : *
4180 : * [task p]
4181 : * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4182 : *
4183 : * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4184 : * __schedule(). See the comment for smp_mb__after_spinlock().
4185 : *
4186 : * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4187 : */
4188 1828 : smp_rmb();
4189 1828 : if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4190 : goto unlock;
4191 :
4192 : #ifdef CONFIG_SMP
4193 : /*
4194 : * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4195 : * possible to, falsely, observe p->on_cpu == 0.
4196 : *
4197 : * One must be running (->on_cpu == 1) in order to remove oneself
4198 : * from the runqueue.
4199 : *
4200 : * __schedule() (switch to task 'p') try_to_wake_up()
4201 : * STORE p->on_cpu = 1 LOAD p->on_rq
4202 : * UNLOCK rq->lock
4203 : *
4204 : * __schedule() (put 'p' to sleep)
4205 : * LOCK rq->lock smp_rmb();
4206 : * smp_mb__after_spinlock();
4207 : * STORE p->on_rq = 0 LOAD p->on_cpu
4208 : *
4209 : * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4210 : * __schedule(). See the comment for smp_mb__after_spinlock().
4211 : *
4212 : * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4213 : * schedule()'s deactivate_task() has 'happened' and p will no longer
4214 : * care about it's own p->state. See the comment in __schedule().
4215 : */
4216 : smp_acquire__after_ctrl_dep();
4217 :
4218 : /*
4219 : * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4220 : * == 0), which means we need to do an enqueue, change p->state to
4221 : * TASK_WAKING such that we can unlock p->pi_lock before doing the
4222 : * enqueue, such as ttwu_queue_wakelist().
4223 : */
4224 : WRITE_ONCE(p->__state, TASK_WAKING);
4225 :
4226 : /*
4227 : * If the owning (remote) CPU is still in the middle of schedule() with
4228 : * this task as prev, considering queueing p on the remote CPUs wake_list
4229 : * which potentially sends an IPI instead of spinning on p->on_cpu to
4230 : * let the waker make forward progress. This is safe because IRQs are
4231 : * disabled and the IPI will deliver after on_cpu is cleared.
4232 : *
4233 : * Ensure we load task_cpu(p) after p->on_cpu:
4234 : *
4235 : * set_task_cpu(p, cpu);
4236 : * STORE p->cpu = @cpu
4237 : * __schedule() (switch to task 'p')
4238 : * LOCK rq->lock
4239 : * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4240 : * STORE p->on_cpu = 1 LOAD p->cpu
4241 : *
4242 : * to ensure we observe the correct CPU on which the task is currently
4243 : * scheduling.
4244 : */
4245 : if (smp_load_acquire(&p->on_cpu) &&
4246 : ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4247 : goto unlock;
4248 :
4249 : /*
4250 : * If the owning (remote) CPU is still in the middle of schedule() with
4251 : * this task as prev, wait until it's done referencing the task.
4252 : *
4253 : * Pairs with the smp_store_release() in finish_task().
4254 : *
4255 : * This ensures that tasks getting woken will be fully ordered against
4256 : * their previous state and preserve Program Order.
4257 : */
4258 : smp_cond_load_acquire(&p->on_cpu, !VAL);
4259 :
4260 : cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4261 : if (task_cpu(p) != cpu) {
4262 : if (p->in_iowait) {
4263 : delayacct_blkio_end(p);
4264 : atomic_dec(&task_rq(p)->nr_iowait);
4265 : }
4266 :
4267 : wake_flags |= WF_MIGRATED;
4268 : psi_ttwu_dequeue(p);
4269 : set_task_cpu(p, cpu);
4270 : }
4271 : #else
4272 1828 : cpu = task_cpu(p);
4273 : #endif /* CONFIG_SMP */
4274 :
4275 1828 : ttwu_queue(p, cpu, wake_flags);
4276 : unlock:
4277 4328 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4278 : out:
4279 : if (success)
4280 : ttwu_stat(p, task_cpu(p), wake_flags);
4281 2199 : preempt_enable();
4282 :
4283 2199 : return success;
4284 : }
4285 :
4286 : static bool __task_needs_rq_lock(struct task_struct *p)
4287 : {
4288 0 : unsigned int state = READ_ONCE(p->__state);
4289 :
4290 : /*
4291 : * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4292 : * the task is blocked. Make sure to check @state since ttwu() can drop
4293 : * locks at the end, see ttwu_queue_wakelist().
4294 : */
4295 0 : if (state == TASK_RUNNING || state == TASK_WAKING)
4296 : return true;
4297 :
4298 : /*
4299 : * Ensure we load p->on_rq after p->__state, otherwise it would be
4300 : * possible to, falsely, observe p->on_rq == 0.
4301 : *
4302 : * See try_to_wake_up() for a longer comment.
4303 : */
4304 0 : smp_rmb();
4305 0 : if (p->on_rq)
4306 : return true;
4307 :
4308 : #ifdef CONFIG_SMP
4309 : /*
4310 : * Ensure the task has finished __schedule() and will not be referenced
4311 : * anymore. Again, see try_to_wake_up() for a longer comment.
4312 : */
4313 : smp_rmb();
4314 : smp_cond_load_acquire(&p->on_cpu, !VAL);
4315 : #endif
4316 :
4317 : return false;
4318 : }
4319 :
4320 : /**
4321 : * task_call_func - Invoke a function on task in fixed state
4322 : * @p: Process for which the function is to be invoked, can be @current.
4323 : * @func: Function to invoke.
4324 : * @arg: Argument to function.
4325 : *
4326 : * Fix the task in it's current state by avoiding wakeups and or rq operations
4327 : * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4328 : * to work out what the state is, if required. Given that @func can be invoked
4329 : * with a runqueue lock held, it had better be quite lightweight.
4330 : *
4331 : * Returns:
4332 : * Whatever @func returns
4333 : */
4334 0 : int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4335 : {
4336 0 : struct rq *rq = NULL;
4337 : struct rq_flags rf;
4338 : int ret;
4339 :
4340 0 : raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4341 :
4342 0 : if (__task_needs_rq_lock(p))
4343 0 : rq = __task_rq_lock(p, &rf);
4344 :
4345 : /*
4346 : * At this point the task is pinned; either:
4347 : * - blocked and we're holding off wakeups (pi->lock)
4348 : * - woken, and we're holding off enqueue (rq->lock)
4349 : * - queued, and we're holding off schedule (rq->lock)
4350 : * - running, and we're holding off de-schedule (rq->lock)
4351 : *
4352 : * The called function (@func) can use: task_curr(), p->on_rq and
4353 : * p->__state to differentiate between these states.
4354 : */
4355 0 : ret = func(p, arg);
4356 :
4357 0 : if (rq)
4358 : rq_unlock(rq, &rf);
4359 :
4360 0 : raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4361 0 : return ret;
4362 : }
4363 :
4364 : /**
4365 : * cpu_curr_snapshot - Return a snapshot of the currently running task
4366 : * @cpu: The CPU on which to snapshot the task.
4367 : *
4368 : * Returns the task_struct pointer of the task "currently" running on
4369 : * the specified CPU. If the same task is running on that CPU throughout,
4370 : * the return value will be a pointer to that task's task_struct structure.
4371 : * If the CPU did any context switches even vaguely concurrently with the
4372 : * execution of this function, the return value will be a pointer to the
4373 : * task_struct structure of a randomly chosen task that was running on
4374 : * that CPU somewhere around the time that this function was executing.
4375 : *
4376 : * If the specified CPU was offline, the return value is whatever it
4377 : * is, perhaps a pointer to the task_struct structure of that CPU's idle
4378 : * task, but there is no guarantee. Callers wishing a useful return
4379 : * value must take some action to ensure that the specified CPU remains
4380 : * online throughout.
4381 : *
4382 : * This function executes full memory barriers before and after fetching
4383 : * the pointer, which permits the caller to confine this function's fetch
4384 : * with respect to the caller's accesses to other shared variables.
4385 : */
4386 0 : struct task_struct *cpu_curr_snapshot(int cpu)
4387 : {
4388 : struct task_struct *t;
4389 :
4390 0 : smp_mb(); /* Pairing determined by caller's synchronization design. */
4391 0 : t = rcu_dereference(cpu_curr(cpu));
4392 0 : smp_mb(); /* Pairing determined by caller's synchronization design. */
4393 0 : return t;
4394 : }
4395 :
4396 : /**
4397 : * wake_up_process - Wake up a specific process
4398 : * @p: The process to be woken up.
4399 : *
4400 : * Attempt to wake up the nominated process and move it to the set of runnable
4401 : * processes.
4402 : *
4403 : * Return: 1 if the process was woken up, 0 if it was already running.
4404 : *
4405 : * This function executes a full memory barrier before accessing the task state.
4406 : */
4407 2198 : int wake_up_process(struct task_struct *p)
4408 : {
4409 2198 : return try_to_wake_up(p, TASK_NORMAL, 0);
4410 : }
4411 : EXPORT_SYMBOL(wake_up_process);
4412 :
4413 1 : int wake_up_state(struct task_struct *p, unsigned int state)
4414 : {
4415 1 : return try_to_wake_up(p, state, 0);
4416 : }
4417 :
4418 : /*
4419 : * Perform scheduler related setup for a newly forked process p.
4420 : * p is forked by current.
4421 : *
4422 : * __sched_fork() is basic setup used by init_idle() too:
4423 : */
4424 349 : static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4425 : {
4426 349 : p->on_rq = 0;
4427 :
4428 349 : p->se.on_rq = 0;
4429 349 : p->se.exec_start = 0;
4430 349 : p->se.sum_exec_runtime = 0;
4431 349 : p->se.prev_sum_exec_runtime = 0;
4432 349 : p->se.nr_migrations = 0;
4433 349 : p->se.vruntime = 0;
4434 698 : INIT_LIST_HEAD(&p->se.group_node);
4435 :
4436 : #ifdef CONFIG_FAIR_GROUP_SCHED
4437 : p->se.cfs_rq = NULL;
4438 : #endif
4439 :
4440 : #ifdef CONFIG_SCHEDSTATS
4441 : /* Even if schedstat is disabled, there should not be garbage */
4442 : memset(&p->stats, 0, sizeof(p->stats));
4443 : #endif
4444 :
4445 349 : RB_CLEAR_NODE(&p->dl.rb_node);
4446 349 : init_dl_task_timer(&p->dl);
4447 349 : init_dl_inactive_task_timer(&p->dl);
4448 349 : __dl_clear_params(p);
4449 :
4450 698 : INIT_LIST_HEAD(&p->rt.run_list);
4451 349 : p->rt.timeout = 0;
4452 349 : p->rt.time_slice = sched_rr_timeslice;
4453 349 : p->rt.on_rq = 0;
4454 349 : p->rt.on_list = 0;
4455 :
4456 : #ifdef CONFIG_PREEMPT_NOTIFIERS
4457 : INIT_HLIST_HEAD(&p->preempt_notifiers);
4458 : #endif
4459 :
4460 : #ifdef CONFIG_COMPACTION
4461 349 : p->capture_control = NULL;
4462 : #endif
4463 349 : init_numa_balancing(clone_flags, p);
4464 : #ifdef CONFIG_SMP
4465 : p->wake_entry.u_flags = CSD_TYPE_TTWU;
4466 : p->migration_pending = NULL;
4467 : #endif
4468 349 : }
4469 :
4470 : DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4471 :
4472 : #ifdef CONFIG_NUMA_BALANCING
4473 :
4474 : int sysctl_numa_balancing_mode;
4475 :
4476 : static void __set_numabalancing_state(bool enabled)
4477 : {
4478 : if (enabled)
4479 : static_branch_enable(&sched_numa_balancing);
4480 : else
4481 : static_branch_disable(&sched_numa_balancing);
4482 : }
4483 :
4484 : void set_numabalancing_state(bool enabled)
4485 : {
4486 : if (enabled)
4487 : sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4488 : else
4489 : sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4490 : __set_numabalancing_state(enabled);
4491 : }
4492 :
4493 : #ifdef CONFIG_PROC_SYSCTL
4494 : static void reset_memory_tiering(void)
4495 : {
4496 : struct pglist_data *pgdat;
4497 :
4498 : for_each_online_pgdat(pgdat) {
4499 : pgdat->nbp_threshold = 0;
4500 : pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
4501 : pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
4502 : }
4503 : }
4504 :
4505 : static int sysctl_numa_balancing(struct ctl_table *table, int write,
4506 : void *buffer, size_t *lenp, loff_t *ppos)
4507 : {
4508 : struct ctl_table t;
4509 : int err;
4510 : int state = sysctl_numa_balancing_mode;
4511 :
4512 : if (write && !capable(CAP_SYS_ADMIN))
4513 : return -EPERM;
4514 :
4515 : t = *table;
4516 : t.data = &state;
4517 : err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4518 : if (err < 0)
4519 : return err;
4520 : if (write) {
4521 : if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4522 : (state & NUMA_BALANCING_MEMORY_TIERING))
4523 : reset_memory_tiering();
4524 : sysctl_numa_balancing_mode = state;
4525 : __set_numabalancing_state(state);
4526 : }
4527 : return err;
4528 : }
4529 : #endif
4530 : #endif
4531 :
4532 : #ifdef CONFIG_SCHEDSTATS
4533 :
4534 : DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4535 :
4536 : static void set_schedstats(bool enabled)
4537 : {
4538 : if (enabled)
4539 : static_branch_enable(&sched_schedstats);
4540 : else
4541 : static_branch_disable(&sched_schedstats);
4542 : }
4543 :
4544 : void force_schedstat_enabled(void)
4545 : {
4546 : if (!schedstat_enabled()) {
4547 : pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4548 : static_branch_enable(&sched_schedstats);
4549 : }
4550 : }
4551 :
4552 : static int __init setup_schedstats(char *str)
4553 : {
4554 : int ret = 0;
4555 : if (!str)
4556 : goto out;
4557 :
4558 : if (!strcmp(str, "enable")) {
4559 : set_schedstats(true);
4560 : ret = 1;
4561 : } else if (!strcmp(str, "disable")) {
4562 : set_schedstats(false);
4563 : ret = 1;
4564 : }
4565 : out:
4566 : if (!ret)
4567 : pr_warn("Unable to parse schedstats=\n");
4568 :
4569 : return ret;
4570 : }
4571 : __setup("schedstats=", setup_schedstats);
4572 :
4573 : #ifdef CONFIG_PROC_SYSCTL
4574 : static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4575 : size_t *lenp, loff_t *ppos)
4576 : {
4577 : struct ctl_table t;
4578 : int err;
4579 : int state = static_branch_likely(&sched_schedstats);
4580 :
4581 : if (write && !capable(CAP_SYS_ADMIN))
4582 : return -EPERM;
4583 :
4584 : t = *table;
4585 : t.data = &state;
4586 : err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4587 : if (err < 0)
4588 : return err;
4589 : if (write)
4590 : set_schedstats(state);
4591 : return err;
4592 : }
4593 : #endif /* CONFIG_PROC_SYSCTL */
4594 : #endif /* CONFIG_SCHEDSTATS */
4595 :
4596 : #ifdef CONFIG_SYSCTL
4597 : static struct ctl_table sched_core_sysctls[] = {
4598 : #ifdef CONFIG_SCHEDSTATS
4599 : {
4600 : .procname = "sched_schedstats",
4601 : .data = NULL,
4602 : .maxlen = sizeof(unsigned int),
4603 : .mode = 0644,
4604 : .proc_handler = sysctl_schedstats,
4605 : .extra1 = SYSCTL_ZERO,
4606 : .extra2 = SYSCTL_ONE,
4607 : },
4608 : #endif /* CONFIG_SCHEDSTATS */
4609 : #ifdef CONFIG_UCLAMP_TASK
4610 : {
4611 : .procname = "sched_util_clamp_min",
4612 : .data = &sysctl_sched_uclamp_util_min,
4613 : .maxlen = sizeof(unsigned int),
4614 : .mode = 0644,
4615 : .proc_handler = sysctl_sched_uclamp_handler,
4616 : },
4617 : {
4618 : .procname = "sched_util_clamp_max",
4619 : .data = &sysctl_sched_uclamp_util_max,
4620 : .maxlen = sizeof(unsigned int),
4621 : .mode = 0644,
4622 : .proc_handler = sysctl_sched_uclamp_handler,
4623 : },
4624 : {
4625 : .procname = "sched_util_clamp_min_rt_default",
4626 : .data = &sysctl_sched_uclamp_util_min_rt_default,
4627 : .maxlen = sizeof(unsigned int),
4628 : .mode = 0644,
4629 : .proc_handler = sysctl_sched_uclamp_handler,
4630 : },
4631 : #endif /* CONFIG_UCLAMP_TASK */
4632 : #ifdef CONFIG_NUMA_BALANCING
4633 : {
4634 : .procname = "numa_balancing",
4635 : .data = NULL, /* filled in by handler */
4636 : .maxlen = sizeof(unsigned int),
4637 : .mode = 0644,
4638 : .proc_handler = sysctl_numa_balancing,
4639 : .extra1 = SYSCTL_ZERO,
4640 : .extra2 = SYSCTL_FOUR,
4641 : },
4642 : #endif /* CONFIG_NUMA_BALANCING */
4643 : {}
4644 : };
4645 1 : static int __init sched_core_sysctl_init(void)
4646 : {
4647 1 : register_sysctl_init("kernel", sched_core_sysctls);
4648 1 : return 0;
4649 : }
4650 : late_initcall(sched_core_sysctl_init);
4651 : #endif /* CONFIG_SYSCTL */
4652 :
4653 : /*
4654 : * fork()/clone()-time setup:
4655 : */
4656 348 : int sched_fork(unsigned long clone_flags, struct task_struct *p)
4657 : {
4658 348 : __sched_fork(clone_flags, p);
4659 : /*
4660 : * We mark the process as NEW here. This guarantees that
4661 : * nobody will actually run it, and a signal or other external
4662 : * event cannot wake it up and insert it on the runqueue either.
4663 : */
4664 348 : p->__state = TASK_NEW;
4665 :
4666 : /*
4667 : * Make sure we do not leak PI boosting priority to the child.
4668 : */
4669 348 : p->prio = current->normal_prio;
4670 :
4671 348 : uclamp_fork(p);
4672 :
4673 : /*
4674 : * Revert to default priority/policy on fork if requested.
4675 : */
4676 348 : if (unlikely(p->sched_reset_on_fork)) {
4677 0 : if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4678 0 : p->policy = SCHED_NORMAL;
4679 0 : p->static_prio = NICE_TO_PRIO(0);
4680 0 : p->rt_priority = 0;
4681 0 : } else if (PRIO_TO_NICE(p->static_prio) < 0)
4682 0 : p->static_prio = NICE_TO_PRIO(0);
4683 :
4684 0 : p->prio = p->normal_prio = p->static_prio;
4685 0 : set_load_weight(p, false);
4686 :
4687 : /*
4688 : * We don't need the reset flag anymore after the fork. It has
4689 : * fulfilled its duty:
4690 : */
4691 0 : p->sched_reset_on_fork = 0;
4692 : }
4693 :
4694 696 : if (dl_prio(p->prio))
4695 : return -EAGAIN;
4696 696 : else if (rt_prio(p->prio))
4697 0 : p->sched_class = &rt_sched_class;
4698 : else
4699 348 : p->sched_class = &fair_sched_class;
4700 :
4701 348 : init_entity_runnable_average(&p->se);
4702 :
4703 :
4704 : #ifdef CONFIG_SCHED_INFO
4705 : if (likely(sched_info_on()))
4706 : memset(&p->sched_info, 0, sizeof(p->sched_info));
4707 : #endif
4708 : #if defined(CONFIG_SMP)
4709 : p->on_cpu = 0;
4710 : #endif
4711 348 : init_task_preempt_count(p);
4712 : #ifdef CONFIG_SMP
4713 : plist_node_init(&p->pushable_tasks, MAX_PRIO);
4714 : RB_CLEAR_NODE(&p->pushable_dl_tasks);
4715 : #endif
4716 348 : return 0;
4717 : }
4718 :
4719 348 : void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4720 : {
4721 : unsigned long flags;
4722 :
4723 : /*
4724 : * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4725 : * required yet, but lockdep gets upset if rules are violated.
4726 : */
4727 348 : raw_spin_lock_irqsave(&p->pi_lock, flags);
4728 : #ifdef CONFIG_CGROUP_SCHED
4729 : if (1) {
4730 : struct task_group *tg;
4731 : tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4732 : struct task_group, css);
4733 : tg = autogroup_task_group(p, tg);
4734 : p->sched_task_group = tg;
4735 : }
4736 : #endif
4737 348 : rseq_migrate(p);
4738 : /*
4739 : * We're setting the CPU for the first time, we don't migrate,
4740 : * so use __set_task_cpu().
4741 : */
4742 348 : __set_task_cpu(p, smp_processor_id());
4743 348 : if (p->sched_class->task_fork)
4744 348 : p->sched_class->task_fork(p);
4745 696 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4746 348 : }
4747 :
4748 348 : void sched_post_fork(struct task_struct *p)
4749 : {
4750 348 : uclamp_post_fork(p);
4751 348 : }
4752 :
4753 3 : unsigned long to_ratio(u64 period, u64 runtime)
4754 : {
4755 3 : if (runtime == RUNTIME_INF)
4756 : return BW_UNIT;
4757 :
4758 : /*
4759 : * Doing this here saves a lot of checks in all
4760 : * the calling paths, and returning zero seems
4761 : * safe for them anyway.
4762 : */
4763 3 : if (period == 0)
4764 : return 0;
4765 :
4766 6 : return div64_u64(runtime << BW_SHIFT, period);
4767 : }
4768 :
4769 : /*
4770 : * wake_up_new_task - wake up a newly created task for the first time.
4771 : *
4772 : * This function will do some initial scheduler statistics housekeeping
4773 : * that must be done for every newly created context, then puts the task
4774 : * on the runqueue and wakes it.
4775 : */
4776 348 : void wake_up_new_task(struct task_struct *p)
4777 : {
4778 : struct rq_flags rf;
4779 : struct rq *rq;
4780 :
4781 348 : raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4782 348 : WRITE_ONCE(p->__state, TASK_RUNNING);
4783 : #ifdef CONFIG_SMP
4784 : /*
4785 : * Fork balancing, do it here and not earlier because:
4786 : * - cpus_ptr can change in the fork path
4787 : * - any previously selected CPU might disappear through hotplug
4788 : *
4789 : * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4790 : * as we're not fully set-up yet.
4791 : */
4792 : p->recent_used_cpu = task_cpu(p);
4793 : rseq_migrate(p);
4794 : __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4795 : #endif
4796 348 : rq = __task_rq_lock(p, &rf);
4797 348 : update_rq_clock(rq);
4798 348 : post_init_entity_util_avg(p);
4799 :
4800 348 : activate_task(rq, p, ENQUEUE_NOCLOCK);
4801 348 : trace_sched_wakeup_new(p);
4802 348 : check_preempt_curr(rq, p, WF_FORK);
4803 : #ifdef CONFIG_SMP
4804 : if (p->sched_class->task_woken) {
4805 : /*
4806 : * Nothing relies on rq->lock after this, so it's fine to
4807 : * drop it.
4808 : */
4809 : rq_unpin_lock(rq, &rf);
4810 : p->sched_class->task_woken(rq, p);
4811 : rq_repin_lock(rq, &rf);
4812 : }
4813 : #endif
4814 696 : task_rq_unlock(rq, p, &rf);
4815 348 : }
4816 :
4817 : #ifdef CONFIG_PREEMPT_NOTIFIERS
4818 :
4819 : static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4820 :
4821 : void preempt_notifier_inc(void)
4822 : {
4823 : static_branch_inc(&preempt_notifier_key);
4824 : }
4825 : EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4826 :
4827 : void preempt_notifier_dec(void)
4828 : {
4829 : static_branch_dec(&preempt_notifier_key);
4830 : }
4831 : EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4832 :
4833 : /**
4834 : * preempt_notifier_register - tell me when current is being preempted & rescheduled
4835 : * @notifier: notifier struct to register
4836 : */
4837 : void preempt_notifier_register(struct preempt_notifier *notifier)
4838 : {
4839 : if (!static_branch_unlikely(&preempt_notifier_key))
4840 : WARN(1, "registering preempt_notifier while notifiers disabled\n");
4841 :
4842 : hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4843 : }
4844 : EXPORT_SYMBOL_GPL(preempt_notifier_register);
4845 :
4846 : /**
4847 : * preempt_notifier_unregister - no longer interested in preemption notifications
4848 : * @notifier: notifier struct to unregister
4849 : *
4850 : * This is *not* safe to call from within a preemption notifier.
4851 : */
4852 : void preempt_notifier_unregister(struct preempt_notifier *notifier)
4853 : {
4854 : hlist_del(¬ifier->link);
4855 : }
4856 : EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4857 :
4858 : static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4859 : {
4860 : struct preempt_notifier *notifier;
4861 :
4862 : hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4863 : notifier->ops->sched_in(notifier, raw_smp_processor_id());
4864 : }
4865 :
4866 : static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4867 : {
4868 : if (static_branch_unlikely(&preempt_notifier_key))
4869 : __fire_sched_in_preempt_notifiers(curr);
4870 : }
4871 :
4872 : static void
4873 : __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4874 : struct task_struct *next)
4875 : {
4876 : struct preempt_notifier *notifier;
4877 :
4878 : hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4879 : notifier->ops->sched_out(notifier, next);
4880 : }
4881 :
4882 : static __always_inline void
4883 : fire_sched_out_preempt_notifiers(struct task_struct *curr,
4884 : struct task_struct *next)
4885 : {
4886 : if (static_branch_unlikely(&preempt_notifier_key))
4887 : __fire_sched_out_preempt_notifiers(curr, next);
4888 : }
4889 :
4890 : #else /* !CONFIG_PREEMPT_NOTIFIERS */
4891 :
4892 : static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4893 : {
4894 : }
4895 :
4896 : static inline void
4897 : fire_sched_out_preempt_notifiers(struct task_struct *curr,
4898 : struct task_struct *next)
4899 : {
4900 : }
4901 :
4902 : #endif /* CONFIG_PREEMPT_NOTIFIERS */
4903 :
4904 : static inline void prepare_task(struct task_struct *next)
4905 : {
4906 : #ifdef CONFIG_SMP
4907 : /*
4908 : * Claim the task as running, we do this before switching to it
4909 : * such that any running task will have this set.
4910 : *
4911 : * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4912 : * its ordering comment.
4913 : */
4914 : WRITE_ONCE(next->on_cpu, 1);
4915 : #endif
4916 : }
4917 :
4918 : static inline void finish_task(struct task_struct *prev)
4919 : {
4920 : #ifdef CONFIG_SMP
4921 : /*
4922 : * This must be the very last reference to @prev from this CPU. After
4923 : * p->on_cpu is cleared, the task can be moved to a different CPU. We
4924 : * must ensure this doesn't happen until the switch is completely
4925 : * finished.
4926 : *
4927 : * In particular, the load of prev->state in finish_task_switch() must
4928 : * happen before this.
4929 : *
4930 : * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4931 : */
4932 : smp_store_release(&prev->on_cpu, 0);
4933 : #endif
4934 : }
4935 :
4936 : #ifdef CONFIG_SMP
4937 :
4938 : static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
4939 : {
4940 : void (*func)(struct rq *rq);
4941 : struct balance_callback *next;
4942 :
4943 : lockdep_assert_rq_held(rq);
4944 :
4945 : while (head) {
4946 : func = (void (*)(struct rq *))head->func;
4947 : next = head->next;
4948 : head->next = NULL;
4949 : head = next;
4950 :
4951 : func(rq);
4952 : }
4953 : }
4954 :
4955 : static void balance_push(struct rq *rq);
4956 :
4957 : /*
4958 : * balance_push_callback is a right abuse of the callback interface and plays
4959 : * by significantly different rules.
4960 : *
4961 : * Where the normal balance_callback's purpose is to be ran in the same context
4962 : * that queued it (only later, when it's safe to drop rq->lock again),
4963 : * balance_push_callback is specifically targeted at __schedule().
4964 : *
4965 : * This abuse is tolerated because it places all the unlikely/odd cases behind
4966 : * a single test, namely: rq->balance_callback == NULL.
4967 : */
4968 : struct balance_callback balance_push_callback = {
4969 : .next = NULL,
4970 : .func = balance_push,
4971 : };
4972 :
4973 : static inline struct balance_callback *
4974 : __splice_balance_callbacks(struct rq *rq, bool split)
4975 : {
4976 : struct balance_callback *head = rq->balance_callback;
4977 :
4978 : if (likely(!head))
4979 : return NULL;
4980 :
4981 : lockdep_assert_rq_held(rq);
4982 : /*
4983 : * Must not take balance_push_callback off the list when
4984 : * splice_balance_callbacks() and balance_callbacks() are not
4985 : * in the same rq->lock section.
4986 : *
4987 : * In that case it would be possible for __schedule() to interleave
4988 : * and observe the list empty.
4989 : */
4990 : if (split && head == &balance_push_callback)
4991 : head = NULL;
4992 : else
4993 : rq->balance_callback = NULL;
4994 :
4995 : return head;
4996 : }
4997 :
4998 : static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
4999 : {
5000 : return __splice_balance_callbacks(rq, true);
5001 : }
5002 :
5003 : static void __balance_callbacks(struct rq *rq)
5004 : {
5005 : do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
5006 : }
5007 :
5008 : static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5009 : {
5010 : unsigned long flags;
5011 :
5012 : if (unlikely(head)) {
5013 : raw_spin_rq_lock_irqsave(rq, flags);
5014 : do_balance_callbacks(rq, head);
5015 : raw_spin_rq_unlock_irqrestore(rq, flags);
5016 : }
5017 : }
5018 :
5019 : #else
5020 :
5021 : static inline void __balance_callbacks(struct rq *rq)
5022 : {
5023 : }
5024 :
5025 : static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5026 : {
5027 : return NULL;
5028 : }
5029 :
5030 : static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5031 : {
5032 : }
5033 :
5034 : #endif
5035 :
5036 : static inline void
5037 : prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
5038 : {
5039 : /*
5040 : * Since the runqueue lock will be released by the next
5041 : * task (which is an invalid locking op but in the case
5042 : * of the scheduler it's an obvious special-case), so we
5043 : * do an early lockdep release here:
5044 : */
5045 2264 : rq_unpin_lock(rq, rf);
5046 : spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
5047 : #ifdef CONFIG_DEBUG_SPINLOCK
5048 : /* this is a valid case when another task releases the spinlock */
5049 : rq_lockp(rq)->owner = next;
5050 : #endif
5051 : }
5052 :
5053 : static inline void finish_lock_switch(struct rq *rq)
5054 : {
5055 : /*
5056 : * If we are tracking spinlock dependencies then we have to
5057 : * fix up the runqueue lock - which gets 'carried over' from
5058 : * prev into current:
5059 : */
5060 : spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
5061 2264 : __balance_callbacks(rq);
5062 2264 : raw_spin_rq_unlock_irq(rq);
5063 : }
5064 :
5065 : /*
5066 : * NOP if the arch has not defined these:
5067 : */
5068 :
5069 : #ifndef prepare_arch_switch
5070 : # define prepare_arch_switch(next) do { } while (0)
5071 : #endif
5072 :
5073 : #ifndef finish_arch_post_lock_switch
5074 : # define finish_arch_post_lock_switch() do { } while (0)
5075 : #endif
5076 :
5077 : static inline void kmap_local_sched_out(void)
5078 : {
5079 : #ifdef CONFIG_KMAP_LOCAL
5080 : if (unlikely(current->kmap_ctrl.idx))
5081 : __kmap_local_sched_out();
5082 : #endif
5083 : }
5084 :
5085 : static inline void kmap_local_sched_in(void)
5086 : {
5087 : #ifdef CONFIG_KMAP_LOCAL
5088 : if (unlikely(current->kmap_ctrl.idx))
5089 : __kmap_local_sched_in();
5090 : #endif
5091 : }
5092 :
5093 : /**
5094 : * prepare_task_switch - prepare to switch tasks
5095 : * @rq: the runqueue preparing to switch
5096 : * @prev: the current task that is being switched out
5097 : * @next: the task we are going to switch to.
5098 : *
5099 : * This is called with the rq lock held and interrupts off. It must
5100 : * be paired with a subsequent finish_task_switch after the context
5101 : * switch.
5102 : *
5103 : * prepare_task_switch sets up locking and calls architecture specific
5104 : * hooks.
5105 : */
5106 : static inline void
5107 : prepare_task_switch(struct rq *rq, struct task_struct *prev,
5108 : struct task_struct *next)
5109 : {
5110 : kcov_prepare_switch(prev);
5111 : sched_info_switch(rq, prev, next);
5112 : perf_event_task_sched_out(prev, next);
5113 : rseq_preempt(prev);
5114 : switch_mm_cid(prev, next);
5115 : fire_sched_out_preempt_notifiers(prev, next);
5116 : kmap_local_sched_out();
5117 : prepare_task(next);
5118 : prepare_arch_switch(next);
5119 : }
5120 :
5121 : /**
5122 : * finish_task_switch - clean up after a task-switch
5123 : * @prev: the thread we just switched away from.
5124 : *
5125 : * finish_task_switch must be called after the context switch, paired
5126 : * with a prepare_task_switch call before the context switch.
5127 : * finish_task_switch will reconcile locking set up by prepare_task_switch,
5128 : * and do any other architecture-specific cleanup actions.
5129 : *
5130 : * Note that we may have delayed dropping an mm in context_switch(). If
5131 : * so, we finish that here outside of the runqueue lock. (Doing it
5132 : * with the lock held can cause deadlocks; see schedule() for
5133 : * details.)
5134 : *
5135 : * The context switch have flipped the stack from under us and restored the
5136 : * local variables which were saved when this task called schedule() in the
5137 : * past. prev == current is still correct but we need to recalculate this_rq
5138 : * because prev may have moved to another CPU.
5139 : */
5140 2264 : static struct rq *finish_task_switch(struct task_struct *prev)
5141 : __releases(rq->lock)
5142 : {
5143 2264 : struct rq *rq = this_rq();
5144 2264 : struct mm_struct *mm = rq->prev_mm;
5145 : unsigned int prev_state;
5146 :
5147 : /*
5148 : * The previous task will have left us with a preempt_count of 2
5149 : * because it left us after:
5150 : *
5151 : * schedule()
5152 : * preempt_disable(); // 1
5153 : * __schedule()
5154 : * raw_spin_lock_irq(&rq->lock) // 2
5155 : *
5156 : * Also, see FORK_PREEMPT_COUNT.
5157 : */
5158 2264 : if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5159 : "corrupted preempt_count: %s/%d/0x%x\n",
5160 : current->comm, current->pid, preempt_count()))
5161 : preempt_count_set(FORK_PREEMPT_COUNT);
5162 :
5163 2264 : rq->prev_mm = NULL;
5164 :
5165 : /*
5166 : * A task struct has one reference for the use as "current".
5167 : * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5168 : * schedule one last time. The schedule call will never return, and
5169 : * the scheduled task must drop that reference.
5170 : *
5171 : * We must observe prev->state before clearing prev->on_cpu (in
5172 : * finish_task), otherwise a concurrent wakeup can get prev
5173 : * running on another CPU and we could rave with its RUNNING -> DEAD
5174 : * transition, resulting in a double drop.
5175 : */
5176 2264 : prev_state = READ_ONCE(prev->__state);
5177 2264 : vtime_task_switch(prev);
5178 2264 : perf_event_task_sched_in(prev, current);
5179 2264 : finish_task(prev);
5180 : tick_nohz_task_switch();
5181 2264 : finish_lock_switch(rq);
5182 : finish_arch_post_lock_switch();
5183 2264 : kcov_finish_switch(current);
5184 : /*
5185 : * kmap_local_sched_out() is invoked with rq::lock held and
5186 : * interrupts disabled. There is no requirement for that, but the
5187 : * sched out code does not have an interrupt enabled section.
5188 : * Restoring the maps on sched in does not require interrupts being
5189 : * disabled either.
5190 : */
5191 : kmap_local_sched_in();
5192 :
5193 2264 : fire_sched_in_preempt_notifiers(current);
5194 : /*
5195 : * When switching through a kernel thread, the loop in
5196 : * membarrier_{private,global}_expedited() may have observed that
5197 : * kernel thread and not issued an IPI. It is therefore possible to
5198 : * schedule between user->kernel->user threads without passing though
5199 : * switch_mm(). Membarrier requires a barrier after storing to
5200 : * rq->curr, before returning to userspace, so provide them here:
5201 : *
5202 : * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5203 : * provided by mmdrop(),
5204 : * - a sync_core for SYNC_CORE.
5205 : */
5206 2264 : if (mm) {
5207 0 : membarrier_mm_sync_core_before_usermode(mm);
5208 : mmdrop_sched(mm);
5209 : }
5210 2264 : if (unlikely(prev_state == TASK_DEAD)) {
5211 333 : if (prev->sched_class->task_dead)
5212 0 : prev->sched_class->task_dead(prev);
5213 :
5214 : /* Task is done with its stack. */
5215 333 : put_task_stack(prev);
5216 :
5217 333 : put_task_struct_rcu_user(prev);
5218 : }
5219 :
5220 2264 : return rq;
5221 : }
5222 :
5223 : /**
5224 : * schedule_tail - first thing a freshly forked thread must call.
5225 : * @prev: the thread we just switched away from.
5226 : */
5227 348 : asmlinkage __visible void schedule_tail(struct task_struct *prev)
5228 : __releases(rq->lock)
5229 : {
5230 : /*
5231 : * New tasks start with FORK_PREEMPT_COUNT, see there and
5232 : * finish_task_switch() for details.
5233 : *
5234 : * finish_task_switch() will drop rq->lock() and lower preempt_count
5235 : * and the preempt_enable() will end up enabling preemption (on
5236 : * PREEMPT_COUNT kernels).
5237 : */
5238 :
5239 348 : finish_task_switch(prev);
5240 348 : preempt_enable();
5241 :
5242 348 : if (current->set_child_tid)
5243 0 : put_user(task_pid_vnr(current), current->set_child_tid);
5244 :
5245 348 : calculate_sigpending();
5246 348 : }
5247 :
5248 : /*
5249 : * context_switch - switch to the new MM and the new thread's register state.
5250 : */
5251 : static __always_inline struct rq *
5252 : context_switch(struct rq *rq, struct task_struct *prev,
5253 : struct task_struct *next, struct rq_flags *rf)
5254 : {
5255 2264 : prepare_task_switch(rq, prev, next);
5256 :
5257 : /*
5258 : * For paravirt, this is coupled with an exit in switch_to to
5259 : * combine the page table reload and the switch backend into
5260 : * one hypercall.
5261 : */
5262 : arch_start_context_switch(prev);
5263 :
5264 : /*
5265 : * kernel -> kernel lazy + transfer active
5266 : * user -> kernel lazy + mmgrab() active
5267 : *
5268 : * kernel -> user switch + mmdrop() active
5269 : * user -> user switch
5270 : */
5271 2264 : if (!next->mm) { // to kernel
5272 2264 : enter_lazy_tlb(prev->active_mm, next);
5273 :
5274 2264 : next->active_mm = prev->active_mm;
5275 2264 : if (prev->mm) // from user
5276 0 : mmgrab(prev->active_mm);
5277 : else
5278 2264 : prev->active_mm = NULL;
5279 : } else { // to user
5280 0 : membarrier_switch_mm(rq, prev->active_mm, next->mm);
5281 : /*
5282 : * sys_membarrier() requires an smp_mb() between setting
5283 : * rq->curr / membarrier_switch_mm() and returning to userspace.
5284 : *
5285 : * The below provides this either through switch_mm(), or in
5286 : * case 'prev->active_mm == next->mm' through
5287 : * finish_task_switch()'s mmdrop().
5288 : */
5289 0 : switch_mm_irqs_off(prev->active_mm, next->mm, next);
5290 0 : lru_gen_use_mm(next->mm);
5291 :
5292 0 : if (!prev->mm) { // from kernel
5293 : /* will mmdrop() in finish_task_switch(). */
5294 0 : rq->prev_mm = prev->active_mm;
5295 0 : prev->active_mm = NULL;
5296 : }
5297 : }
5298 :
5299 2264 : rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5300 :
5301 4528 : prepare_lock_switch(rq, next, rf);
5302 :
5303 : /* Here we just switch the register state and the stack. */
5304 2264 : switch_to(prev, next, prev);
5305 1916 : barrier();
5306 :
5307 1916 : return finish_task_switch(prev);
5308 : }
5309 :
5310 : /*
5311 : * nr_running and nr_context_switches:
5312 : *
5313 : * externally visible scheduler statistics: current number of runnable
5314 : * threads, total number of context switches performed since bootup.
5315 : */
5316 0 : unsigned int nr_running(void)
5317 : {
5318 0 : unsigned int i, sum = 0;
5319 :
5320 0 : for_each_online_cpu(i)
5321 0 : sum += cpu_rq(i)->nr_running;
5322 :
5323 0 : return sum;
5324 : }
5325 :
5326 : /*
5327 : * Check if only the current task is running on the CPU.
5328 : *
5329 : * Caution: this function does not check that the caller has disabled
5330 : * preemption, thus the result might have a time-of-check-to-time-of-use
5331 : * race. The caller is responsible to use it correctly, for example:
5332 : *
5333 : * - from a non-preemptible section (of course)
5334 : *
5335 : * - from a thread that is bound to a single CPU
5336 : *
5337 : * - in a loop with very short iterations (e.g. a polling loop)
5338 : */
5339 0 : bool single_task_running(void)
5340 : {
5341 0 : return raw_rq()->nr_running == 1;
5342 : }
5343 : EXPORT_SYMBOL(single_task_running);
5344 :
5345 0 : unsigned long long nr_context_switches_cpu(int cpu)
5346 : {
5347 0 : return cpu_rq(cpu)->nr_switches;
5348 : }
5349 :
5350 0 : unsigned long long nr_context_switches(void)
5351 : {
5352 : int i;
5353 0 : unsigned long long sum = 0;
5354 :
5355 0 : for_each_possible_cpu(i)
5356 0 : sum += cpu_rq(i)->nr_switches;
5357 :
5358 0 : return sum;
5359 : }
5360 :
5361 : /*
5362 : * Consumers of these two interfaces, like for example the cpuidle menu
5363 : * governor, are using nonsensical data. Preferring shallow idle state selection
5364 : * for a CPU that has IO-wait which might not even end up running the task when
5365 : * it does become runnable.
5366 : */
5367 :
5368 0 : unsigned int nr_iowait_cpu(int cpu)
5369 : {
5370 0 : return atomic_read(&cpu_rq(cpu)->nr_iowait);
5371 : }
5372 :
5373 : /*
5374 : * IO-wait accounting, and how it's mostly bollocks (on SMP).
5375 : *
5376 : * The idea behind IO-wait account is to account the idle time that we could
5377 : * have spend running if it were not for IO. That is, if we were to improve the
5378 : * storage performance, we'd have a proportional reduction in IO-wait time.
5379 : *
5380 : * This all works nicely on UP, where, when a task blocks on IO, we account
5381 : * idle time as IO-wait, because if the storage were faster, it could've been
5382 : * running and we'd not be idle.
5383 : *
5384 : * This has been extended to SMP, by doing the same for each CPU. This however
5385 : * is broken.
5386 : *
5387 : * Imagine for instance the case where two tasks block on one CPU, only the one
5388 : * CPU will have IO-wait accounted, while the other has regular idle. Even
5389 : * though, if the storage were faster, both could've ran at the same time,
5390 : * utilising both CPUs.
5391 : *
5392 : * This means, that when looking globally, the current IO-wait accounting on
5393 : * SMP is a lower bound, by reason of under accounting.
5394 : *
5395 : * Worse, since the numbers are provided per CPU, they are sometimes
5396 : * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5397 : * associated with any one particular CPU, it can wake to another CPU than it
5398 : * blocked on. This means the per CPU IO-wait number is meaningless.
5399 : *
5400 : * Task CPU affinities can make all that even more 'interesting'.
5401 : */
5402 :
5403 0 : unsigned int nr_iowait(void)
5404 : {
5405 0 : unsigned int i, sum = 0;
5406 :
5407 0 : for_each_possible_cpu(i)
5408 0 : sum += nr_iowait_cpu(i);
5409 :
5410 0 : return sum;
5411 : }
5412 :
5413 : #ifdef CONFIG_SMP
5414 :
5415 : /*
5416 : * sched_exec - execve() is a valuable balancing opportunity, because at
5417 : * this point the task has the smallest effective memory and cache footprint.
5418 : */
5419 : void sched_exec(void)
5420 : {
5421 : struct task_struct *p = current;
5422 : unsigned long flags;
5423 : int dest_cpu;
5424 :
5425 : raw_spin_lock_irqsave(&p->pi_lock, flags);
5426 : dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5427 : if (dest_cpu == smp_processor_id())
5428 : goto unlock;
5429 :
5430 : if (likely(cpu_active(dest_cpu))) {
5431 : struct migration_arg arg = { p, dest_cpu };
5432 :
5433 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5434 : stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5435 : return;
5436 : }
5437 : unlock:
5438 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5439 : }
5440 :
5441 : #endif
5442 :
5443 : DEFINE_PER_CPU(struct kernel_stat, kstat);
5444 : DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5445 :
5446 : EXPORT_PER_CPU_SYMBOL(kstat);
5447 : EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5448 :
5449 : /*
5450 : * The function fair_sched_class.update_curr accesses the struct curr
5451 : * and its field curr->exec_start; when called from task_sched_runtime(),
5452 : * we observe a high rate of cache misses in practice.
5453 : * Prefetching this data results in improved performance.
5454 : */
5455 : static inline void prefetch_curr_exec_start(struct task_struct *p)
5456 : {
5457 : #ifdef CONFIG_FAIR_GROUP_SCHED
5458 : struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5459 : #else
5460 0 : struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5461 : #endif
5462 0 : prefetch(curr);
5463 0 : prefetch(&curr->exec_start);
5464 : }
5465 :
5466 : /*
5467 : * Return accounted runtime for the task.
5468 : * In case the task is currently running, return the runtime plus current's
5469 : * pending runtime that have not been accounted yet.
5470 : */
5471 0 : unsigned long long task_sched_runtime(struct task_struct *p)
5472 : {
5473 : struct rq_flags rf;
5474 : struct rq *rq;
5475 : u64 ns;
5476 :
5477 : #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5478 : /*
5479 : * 64-bit doesn't need locks to atomically read a 64-bit value.
5480 : * So we have a optimization chance when the task's delta_exec is 0.
5481 : * Reading ->on_cpu is racy, but this is ok.
5482 : *
5483 : * If we race with it leaving CPU, we'll take a lock. So we're correct.
5484 : * If we race with it entering CPU, unaccounted time is 0. This is
5485 : * indistinguishable from the read occurring a few cycles earlier.
5486 : * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5487 : * been accounted, so we're correct here as well.
5488 : */
5489 : if (!p->on_cpu || !task_on_rq_queued(p))
5490 : return p->se.sum_exec_runtime;
5491 : #endif
5492 :
5493 0 : rq = task_rq_lock(p, &rf);
5494 : /*
5495 : * Must be ->curr _and_ ->on_rq. If dequeued, we would
5496 : * project cycles that may never be accounted to this
5497 : * thread, breaking clock_gettime().
5498 : */
5499 0 : if (task_current(rq, p) && task_on_rq_queued(p)) {
5500 0 : prefetch_curr_exec_start(p);
5501 0 : update_rq_clock(rq);
5502 0 : p->sched_class->update_curr(rq);
5503 : }
5504 0 : ns = p->se.sum_exec_runtime;
5505 0 : task_rq_unlock(rq, p, &rf);
5506 :
5507 0 : return ns;
5508 : }
5509 :
5510 : #ifdef CONFIG_SCHED_DEBUG
5511 0 : static u64 cpu_resched_latency(struct rq *rq)
5512 : {
5513 0 : int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5514 0 : u64 resched_latency, now = rq_clock(rq);
5515 : static bool warned_once;
5516 :
5517 0 : if (sysctl_resched_latency_warn_once && warned_once)
5518 : return 0;
5519 :
5520 0 : if (!need_resched() || !latency_warn_ms)
5521 : return 0;
5522 :
5523 0 : if (system_state == SYSTEM_BOOTING)
5524 : return 0;
5525 :
5526 0 : if (!rq->last_seen_need_resched_ns) {
5527 0 : rq->last_seen_need_resched_ns = now;
5528 0 : rq->ticks_without_resched = 0;
5529 0 : return 0;
5530 : }
5531 :
5532 0 : rq->ticks_without_resched++;
5533 0 : resched_latency = now - rq->last_seen_need_resched_ns;
5534 0 : if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5535 : return 0;
5536 :
5537 0 : warned_once = true;
5538 :
5539 0 : return resched_latency;
5540 : }
5541 :
5542 0 : static int __init setup_resched_latency_warn_ms(char *str)
5543 : {
5544 : long val;
5545 :
5546 0 : if ((kstrtol(str, 0, &val))) {
5547 0 : pr_warn("Unable to set resched_latency_warn_ms\n");
5548 0 : return 1;
5549 : }
5550 :
5551 0 : sysctl_resched_latency_warn_ms = val;
5552 0 : return 1;
5553 : }
5554 : __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5555 : #else
5556 : static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5557 : #endif /* CONFIG_SCHED_DEBUG */
5558 :
5559 : /*
5560 : * This function gets called by the timer code, with HZ frequency.
5561 : * We call it with interrupts disabled.
5562 : */
5563 2751 : void scheduler_tick(void)
5564 : {
5565 2751 : int cpu = smp_processor_id();
5566 2751 : struct rq *rq = cpu_rq(cpu);
5567 2751 : struct task_struct *curr = rq->curr;
5568 : struct rq_flags rf;
5569 : unsigned long thermal_pressure;
5570 : u64 resched_latency;
5571 :
5572 2751 : if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5573 : arch_scale_freq_tick();
5574 :
5575 2751 : sched_clock_tick();
5576 :
5577 2751 : rq_lock(rq, &rf);
5578 :
5579 2751 : update_rq_clock(rq);
5580 2751 : thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5581 2751 : update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5582 2751 : curr->sched_class->task_tick(rq, curr, 0);
5583 2751 : if (sched_feat(LATENCY_WARN))
5584 0 : resched_latency = cpu_resched_latency(rq);
5585 2751 : calc_global_load_tick(rq);
5586 2751 : sched_core_tick(rq);
5587 :
5588 2751 : rq_unlock(rq, &rf);
5589 :
5590 2751 : if (sched_feat(LATENCY_WARN) && resched_latency)
5591 0 : resched_latency_warn(cpu, resched_latency);
5592 :
5593 : perf_event_task_tick();
5594 :
5595 : #ifdef CONFIG_SMP
5596 : rq->idle_balance = idle_cpu(cpu);
5597 : trigger_load_balance(rq);
5598 : #endif
5599 2751 : }
5600 :
5601 : #ifdef CONFIG_NO_HZ_FULL
5602 :
5603 : struct tick_work {
5604 : int cpu;
5605 : atomic_t state;
5606 : struct delayed_work work;
5607 : };
5608 : /* Values for ->state, see diagram below. */
5609 : #define TICK_SCHED_REMOTE_OFFLINE 0
5610 : #define TICK_SCHED_REMOTE_OFFLINING 1
5611 : #define TICK_SCHED_REMOTE_RUNNING 2
5612 :
5613 : /*
5614 : * State diagram for ->state:
5615 : *
5616 : *
5617 : * TICK_SCHED_REMOTE_OFFLINE
5618 : * | ^
5619 : * | |
5620 : * | | sched_tick_remote()
5621 : * | |
5622 : * | |
5623 : * +--TICK_SCHED_REMOTE_OFFLINING
5624 : * | ^
5625 : * | |
5626 : * sched_tick_start() | | sched_tick_stop()
5627 : * | |
5628 : * V |
5629 : * TICK_SCHED_REMOTE_RUNNING
5630 : *
5631 : *
5632 : * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5633 : * and sched_tick_start() are happy to leave the state in RUNNING.
5634 : */
5635 :
5636 : static struct tick_work __percpu *tick_work_cpu;
5637 :
5638 : static void sched_tick_remote(struct work_struct *work)
5639 : {
5640 : struct delayed_work *dwork = to_delayed_work(work);
5641 : struct tick_work *twork = container_of(dwork, struct tick_work, work);
5642 : int cpu = twork->cpu;
5643 : struct rq *rq = cpu_rq(cpu);
5644 : struct task_struct *curr;
5645 : struct rq_flags rf;
5646 : u64 delta;
5647 : int os;
5648 :
5649 : /*
5650 : * Handle the tick only if it appears the remote CPU is running in full
5651 : * dynticks mode. The check is racy by nature, but missing a tick or
5652 : * having one too much is no big deal because the scheduler tick updates
5653 : * statistics and checks timeslices in a time-independent way, regardless
5654 : * of when exactly it is running.
5655 : */
5656 : if (!tick_nohz_tick_stopped_cpu(cpu))
5657 : goto out_requeue;
5658 :
5659 : rq_lock_irq(rq, &rf);
5660 : curr = rq->curr;
5661 : if (cpu_is_offline(cpu))
5662 : goto out_unlock;
5663 :
5664 : update_rq_clock(rq);
5665 :
5666 : if (!is_idle_task(curr)) {
5667 : /*
5668 : * Make sure the next tick runs within a reasonable
5669 : * amount of time.
5670 : */
5671 : delta = rq_clock_task(rq) - curr->se.exec_start;
5672 : WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5673 : }
5674 : curr->sched_class->task_tick(rq, curr, 0);
5675 :
5676 : calc_load_nohz_remote(rq);
5677 : out_unlock:
5678 : rq_unlock_irq(rq, &rf);
5679 : out_requeue:
5680 :
5681 : /*
5682 : * Run the remote tick once per second (1Hz). This arbitrary
5683 : * frequency is large enough to avoid overload but short enough
5684 : * to keep scheduler internal stats reasonably up to date. But
5685 : * first update state to reflect hotplug activity if required.
5686 : */
5687 : os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5688 : WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5689 : if (os == TICK_SCHED_REMOTE_RUNNING)
5690 : queue_delayed_work(system_unbound_wq, dwork, HZ);
5691 : }
5692 :
5693 : static void sched_tick_start(int cpu)
5694 : {
5695 : int os;
5696 : struct tick_work *twork;
5697 :
5698 : if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5699 : return;
5700 :
5701 : WARN_ON_ONCE(!tick_work_cpu);
5702 :
5703 : twork = per_cpu_ptr(tick_work_cpu, cpu);
5704 : os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5705 : WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5706 : if (os == TICK_SCHED_REMOTE_OFFLINE) {
5707 : twork->cpu = cpu;
5708 : INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5709 : queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5710 : }
5711 : }
5712 :
5713 : #ifdef CONFIG_HOTPLUG_CPU
5714 : static void sched_tick_stop(int cpu)
5715 : {
5716 : struct tick_work *twork;
5717 : int os;
5718 :
5719 : if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5720 : return;
5721 :
5722 : WARN_ON_ONCE(!tick_work_cpu);
5723 :
5724 : twork = per_cpu_ptr(tick_work_cpu, cpu);
5725 : /* There cannot be competing actions, but don't rely on stop-machine. */
5726 : os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5727 : WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5728 : /* Don't cancel, as this would mess up the state machine. */
5729 : }
5730 : #endif /* CONFIG_HOTPLUG_CPU */
5731 :
5732 : int __init sched_tick_offload_init(void)
5733 : {
5734 : tick_work_cpu = alloc_percpu(struct tick_work);
5735 : BUG_ON(!tick_work_cpu);
5736 : return 0;
5737 : }
5738 :
5739 : #else /* !CONFIG_NO_HZ_FULL */
5740 : static inline void sched_tick_start(int cpu) { }
5741 : static inline void sched_tick_stop(int cpu) { }
5742 : #endif
5743 :
5744 : #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5745 : defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5746 : /*
5747 : * If the value passed in is equal to the current preempt count
5748 : * then we just disabled preemption. Start timing the latency.
5749 : */
5750 : static inline void preempt_latency_start(int val)
5751 : {
5752 : if (preempt_count() == val) {
5753 : unsigned long ip = get_lock_parent_ip();
5754 : #ifdef CONFIG_DEBUG_PREEMPT
5755 : current->preempt_disable_ip = ip;
5756 : #endif
5757 : trace_preempt_off(CALLER_ADDR0, ip);
5758 : }
5759 : }
5760 :
5761 : void preempt_count_add(int val)
5762 : {
5763 : #ifdef CONFIG_DEBUG_PREEMPT
5764 : /*
5765 : * Underflow?
5766 : */
5767 : if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5768 : return;
5769 : #endif
5770 : __preempt_count_add(val);
5771 : #ifdef CONFIG_DEBUG_PREEMPT
5772 : /*
5773 : * Spinlock count overflowing soon?
5774 : */
5775 : DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5776 : PREEMPT_MASK - 10);
5777 : #endif
5778 : preempt_latency_start(val);
5779 : }
5780 : EXPORT_SYMBOL(preempt_count_add);
5781 : NOKPROBE_SYMBOL(preempt_count_add);
5782 :
5783 : /*
5784 : * If the value passed in equals to the current preempt count
5785 : * then we just enabled preemption. Stop timing the latency.
5786 : */
5787 : static inline void preempt_latency_stop(int val)
5788 : {
5789 : if (preempt_count() == val)
5790 : trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5791 : }
5792 :
5793 : void preempt_count_sub(int val)
5794 : {
5795 : #ifdef CONFIG_DEBUG_PREEMPT
5796 : /*
5797 : * Underflow?
5798 : */
5799 : if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5800 : return;
5801 : /*
5802 : * Is the spinlock portion underflowing?
5803 : */
5804 : if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5805 : !(preempt_count() & PREEMPT_MASK)))
5806 : return;
5807 : #endif
5808 :
5809 : preempt_latency_stop(val);
5810 : __preempt_count_sub(val);
5811 : }
5812 : EXPORT_SYMBOL(preempt_count_sub);
5813 : NOKPROBE_SYMBOL(preempt_count_sub);
5814 :
5815 : #else
5816 : static inline void preempt_latency_start(int val) { }
5817 : static inline void preempt_latency_stop(int val) { }
5818 : #endif
5819 :
5820 : static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5821 : {
5822 : #ifdef CONFIG_DEBUG_PREEMPT
5823 : return p->preempt_disable_ip;
5824 : #else
5825 : return 0;
5826 : #endif
5827 : }
5828 :
5829 : /*
5830 : * Print scheduling while atomic bug:
5831 : */
5832 0 : static noinline void __schedule_bug(struct task_struct *prev)
5833 : {
5834 : /* Save this before calling printk(), since that will clobber it */
5835 0 : unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5836 :
5837 0 : if (oops_in_progress)
5838 : return;
5839 :
5840 0 : printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5841 : prev->comm, prev->pid, preempt_count());
5842 :
5843 0 : debug_show_held_locks(prev);
5844 : print_modules();
5845 0 : if (irqs_disabled())
5846 : print_irqtrace_events(prev);
5847 : if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5848 : && in_atomic_preempt_off()) {
5849 : pr_err("Preemption disabled at:");
5850 : print_ip_sym(KERN_ERR, preempt_disable_ip);
5851 : }
5852 0 : check_panic_on_warn("scheduling while atomic");
5853 :
5854 0 : dump_stack();
5855 0 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5856 : }
5857 :
5858 : /*
5859 : * Various schedule()-time debugging checks and statistics:
5860 : */
5861 : static inline void schedule_debug(struct task_struct *prev, bool preempt)
5862 : {
5863 : #ifdef CONFIG_SCHED_STACK_END_CHECK
5864 : if (task_stack_end_corrupted(prev))
5865 : panic("corrupted stack end detected inside scheduler\n");
5866 :
5867 : if (task_scs_end_corrupted(prev))
5868 : panic("corrupted shadow stack detected inside scheduler\n");
5869 : #endif
5870 :
5871 : #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5872 : if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5873 : printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5874 : prev->comm, prev->pid, prev->non_block_count);
5875 : dump_stack();
5876 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5877 : }
5878 : #endif
5879 :
5880 2264 : if (unlikely(in_atomic_preempt_off())) {
5881 0 : __schedule_bug(prev);
5882 : preempt_count_set(PREEMPT_DISABLED);
5883 : }
5884 : rcu_sleep_check();
5885 2264 : SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5886 :
5887 2264 : profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5888 :
5889 : schedstat_inc(this_rq()->sched_count);
5890 : }
5891 :
5892 : static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5893 : struct rq_flags *rf)
5894 : {
5895 : #ifdef CONFIG_SMP
5896 : const struct sched_class *class;
5897 : /*
5898 : * We must do the balancing pass before put_prev_task(), such
5899 : * that when we release the rq->lock the task is in the same
5900 : * state as before we took rq->lock.
5901 : *
5902 : * We can terminate the balance pass as soon as we know there is
5903 : * a runnable task of @class priority or higher.
5904 : */
5905 : for_class_range(class, prev->sched_class, &idle_sched_class) {
5906 : if (class->balance(rq, prev, rf))
5907 : break;
5908 : }
5909 : #endif
5910 :
5911 0 : put_prev_task(rq, prev);
5912 : }
5913 :
5914 : /*
5915 : * Pick up the highest-prio task:
5916 : */
5917 : static inline struct task_struct *
5918 2264 : __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5919 : {
5920 : const struct sched_class *class;
5921 : struct task_struct *p;
5922 :
5923 : /*
5924 : * Optimization: we know that if all tasks are in the fair class we can
5925 : * call that function directly, but only if the @prev task wasn't of a
5926 : * higher scheduling class, because otherwise those lose the
5927 : * opportunity to pull in more work from other CPUs.
5928 : */
5929 2264 : if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
5930 : rq->nr_running == rq->cfs.h_nr_running)) {
5931 :
5932 2264 : p = pick_next_task_fair(rq, prev, rf);
5933 2264 : if (unlikely(p == RETRY_TASK))
5934 : goto restart;
5935 :
5936 : /* Assume the next prioritized class is idle_sched_class */
5937 2264 : if (!p) {
5938 1 : put_prev_task(rq, prev);
5939 1 : p = pick_next_task_idle(rq);
5940 : }
5941 :
5942 : return p;
5943 : }
5944 :
5945 : restart:
5946 0 : put_prev_task_balance(rq, prev, rf);
5947 :
5948 0 : for_each_class(class) {
5949 0 : p = class->pick_next_task(rq);
5950 0 : if (p)
5951 : return p;
5952 : }
5953 :
5954 0 : BUG(); /* The idle class should always have a runnable task. */
5955 : }
5956 :
5957 : #ifdef CONFIG_SCHED_CORE
5958 : static inline bool is_task_rq_idle(struct task_struct *t)
5959 : {
5960 : return (task_rq(t)->idle == t);
5961 : }
5962 :
5963 : static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5964 : {
5965 : return is_task_rq_idle(a) || (a->core_cookie == cookie);
5966 : }
5967 :
5968 : static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5969 : {
5970 : if (is_task_rq_idle(a) || is_task_rq_idle(b))
5971 : return true;
5972 :
5973 : return a->core_cookie == b->core_cookie;
5974 : }
5975 :
5976 : static inline struct task_struct *pick_task(struct rq *rq)
5977 : {
5978 : const struct sched_class *class;
5979 : struct task_struct *p;
5980 :
5981 : for_each_class(class) {
5982 : p = class->pick_task(rq);
5983 : if (p)
5984 : return p;
5985 : }
5986 :
5987 : BUG(); /* The idle class should always have a runnable task. */
5988 : }
5989 :
5990 : extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5991 :
5992 : static void queue_core_balance(struct rq *rq);
5993 :
5994 : static struct task_struct *
5995 : pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5996 : {
5997 : struct task_struct *next, *p, *max = NULL;
5998 : const struct cpumask *smt_mask;
5999 : bool fi_before = false;
6000 : bool core_clock_updated = (rq == rq->core);
6001 : unsigned long cookie;
6002 : int i, cpu, occ = 0;
6003 : struct rq *rq_i;
6004 : bool need_sync;
6005 :
6006 : if (!sched_core_enabled(rq))
6007 : return __pick_next_task(rq, prev, rf);
6008 :
6009 : cpu = cpu_of(rq);
6010 :
6011 : /* Stopper task is switching into idle, no need core-wide selection. */
6012 : if (cpu_is_offline(cpu)) {
6013 : /*
6014 : * Reset core_pick so that we don't enter the fastpath when
6015 : * coming online. core_pick would already be migrated to
6016 : * another cpu during offline.
6017 : */
6018 : rq->core_pick = NULL;
6019 : return __pick_next_task(rq, prev, rf);
6020 : }
6021 :
6022 : /*
6023 : * If there were no {en,de}queues since we picked (IOW, the task
6024 : * pointers are all still valid), and we haven't scheduled the last
6025 : * pick yet, do so now.
6026 : *
6027 : * rq->core_pick can be NULL if no selection was made for a CPU because
6028 : * it was either offline or went offline during a sibling's core-wide
6029 : * selection. In this case, do a core-wide selection.
6030 : */
6031 : if (rq->core->core_pick_seq == rq->core->core_task_seq &&
6032 : rq->core->core_pick_seq != rq->core_sched_seq &&
6033 : rq->core_pick) {
6034 : WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
6035 :
6036 : next = rq->core_pick;
6037 : if (next != prev) {
6038 : put_prev_task(rq, prev);
6039 : set_next_task(rq, next);
6040 : }
6041 :
6042 : rq->core_pick = NULL;
6043 : goto out;
6044 : }
6045 :
6046 : put_prev_task_balance(rq, prev, rf);
6047 :
6048 : smt_mask = cpu_smt_mask(cpu);
6049 : need_sync = !!rq->core->core_cookie;
6050 :
6051 : /* reset state */
6052 : rq->core->core_cookie = 0UL;
6053 : if (rq->core->core_forceidle_count) {
6054 : if (!core_clock_updated) {
6055 : update_rq_clock(rq->core);
6056 : core_clock_updated = true;
6057 : }
6058 : sched_core_account_forceidle(rq);
6059 : /* reset after accounting force idle */
6060 : rq->core->core_forceidle_start = 0;
6061 : rq->core->core_forceidle_count = 0;
6062 : rq->core->core_forceidle_occupation = 0;
6063 : need_sync = true;
6064 : fi_before = true;
6065 : }
6066 :
6067 : /*
6068 : * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6069 : *
6070 : * @task_seq guards the task state ({en,de}queues)
6071 : * @pick_seq is the @task_seq we did a selection on
6072 : * @sched_seq is the @pick_seq we scheduled
6073 : *
6074 : * However, preemptions can cause multiple picks on the same task set.
6075 : * 'Fix' this by also increasing @task_seq for every pick.
6076 : */
6077 : rq->core->core_task_seq++;
6078 :
6079 : /*
6080 : * Optimize for common case where this CPU has no cookies
6081 : * and there are no cookied tasks running on siblings.
6082 : */
6083 : if (!need_sync) {
6084 : next = pick_task(rq);
6085 : if (!next->core_cookie) {
6086 : rq->core_pick = NULL;
6087 : /*
6088 : * For robustness, update the min_vruntime_fi for
6089 : * unconstrained picks as well.
6090 : */
6091 : WARN_ON_ONCE(fi_before);
6092 : task_vruntime_update(rq, next, false);
6093 : goto out_set_next;
6094 : }
6095 : }
6096 :
6097 : /*
6098 : * For each thread: do the regular task pick and find the max prio task
6099 : * amongst them.
6100 : *
6101 : * Tie-break prio towards the current CPU
6102 : */
6103 : for_each_cpu_wrap(i, smt_mask, cpu) {
6104 : rq_i = cpu_rq(i);
6105 :
6106 : /*
6107 : * Current cpu always has its clock updated on entrance to
6108 : * pick_next_task(). If the current cpu is not the core,
6109 : * the core may also have been updated above.
6110 : */
6111 : if (i != cpu && (rq_i != rq->core || !core_clock_updated))
6112 : update_rq_clock(rq_i);
6113 :
6114 : p = rq_i->core_pick = pick_task(rq_i);
6115 : if (!max || prio_less(max, p, fi_before))
6116 : max = p;
6117 : }
6118 :
6119 : cookie = rq->core->core_cookie = max->core_cookie;
6120 :
6121 : /*
6122 : * For each thread: try and find a runnable task that matches @max or
6123 : * force idle.
6124 : */
6125 : for_each_cpu(i, smt_mask) {
6126 : rq_i = cpu_rq(i);
6127 : p = rq_i->core_pick;
6128 :
6129 : if (!cookie_equals(p, cookie)) {
6130 : p = NULL;
6131 : if (cookie)
6132 : p = sched_core_find(rq_i, cookie);
6133 : if (!p)
6134 : p = idle_sched_class.pick_task(rq_i);
6135 : }
6136 :
6137 : rq_i->core_pick = p;
6138 :
6139 : if (p == rq_i->idle) {
6140 : if (rq_i->nr_running) {
6141 : rq->core->core_forceidle_count++;
6142 : if (!fi_before)
6143 : rq->core->core_forceidle_seq++;
6144 : }
6145 : } else {
6146 : occ++;
6147 : }
6148 : }
6149 :
6150 : if (schedstat_enabled() && rq->core->core_forceidle_count) {
6151 : rq->core->core_forceidle_start = rq_clock(rq->core);
6152 : rq->core->core_forceidle_occupation = occ;
6153 : }
6154 :
6155 : rq->core->core_pick_seq = rq->core->core_task_seq;
6156 : next = rq->core_pick;
6157 : rq->core_sched_seq = rq->core->core_pick_seq;
6158 :
6159 : /* Something should have been selected for current CPU */
6160 : WARN_ON_ONCE(!next);
6161 :
6162 : /*
6163 : * Reschedule siblings
6164 : *
6165 : * NOTE: L1TF -- at this point we're no longer running the old task and
6166 : * sending an IPI (below) ensures the sibling will no longer be running
6167 : * their task. This ensures there is no inter-sibling overlap between
6168 : * non-matching user state.
6169 : */
6170 : for_each_cpu(i, smt_mask) {
6171 : rq_i = cpu_rq(i);
6172 :
6173 : /*
6174 : * An online sibling might have gone offline before a task
6175 : * could be picked for it, or it might be offline but later
6176 : * happen to come online, but its too late and nothing was
6177 : * picked for it. That's Ok - it will pick tasks for itself,
6178 : * so ignore it.
6179 : */
6180 : if (!rq_i->core_pick)
6181 : continue;
6182 :
6183 : /*
6184 : * Update for new !FI->FI transitions, or if continuing to be in !FI:
6185 : * fi_before fi update?
6186 : * 0 0 1
6187 : * 0 1 1
6188 : * 1 0 1
6189 : * 1 1 0
6190 : */
6191 : if (!(fi_before && rq->core->core_forceidle_count))
6192 : task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6193 :
6194 : rq_i->core_pick->core_occupation = occ;
6195 :
6196 : if (i == cpu) {
6197 : rq_i->core_pick = NULL;
6198 : continue;
6199 : }
6200 :
6201 : /* Did we break L1TF mitigation requirements? */
6202 : WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6203 :
6204 : if (rq_i->curr == rq_i->core_pick) {
6205 : rq_i->core_pick = NULL;
6206 : continue;
6207 : }
6208 :
6209 : resched_curr(rq_i);
6210 : }
6211 :
6212 : out_set_next:
6213 : set_next_task(rq, next);
6214 : out:
6215 : if (rq->core->core_forceidle_count && next == rq->idle)
6216 : queue_core_balance(rq);
6217 :
6218 : return next;
6219 : }
6220 :
6221 : static bool try_steal_cookie(int this, int that)
6222 : {
6223 : struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6224 : struct task_struct *p;
6225 : unsigned long cookie;
6226 : bool success = false;
6227 :
6228 : local_irq_disable();
6229 : double_rq_lock(dst, src);
6230 :
6231 : cookie = dst->core->core_cookie;
6232 : if (!cookie)
6233 : goto unlock;
6234 :
6235 : if (dst->curr != dst->idle)
6236 : goto unlock;
6237 :
6238 : p = sched_core_find(src, cookie);
6239 : if (p == src->idle)
6240 : goto unlock;
6241 :
6242 : do {
6243 : if (p == src->core_pick || p == src->curr)
6244 : goto next;
6245 :
6246 : if (!is_cpu_allowed(p, this))
6247 : goto next;
6248 :
6249 : if (p->core_occupation > dst->idle->core_occupation)
6250 : goto next;
6251 :
6252 : deactivate_task(src, p, 0);
6253 : set_task_cpu(p, this);
6254 : activate_task(dst, p, 0);
6255 :
6256 : resched_curr(dst);
6257 :
6258 : success = true;
6259 : break;
6260 :
6261 : next:
6262 : p = sched_core_next(p, cookie);
6263 : } while (p);
6264 :
6265 : unlock:
6266 : double_rq_unlock(dst, src);
6267 : local_irq_enable();
6268 :
6269 : return success;
6270 : }
6271 :
6272 : static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6273 : {
6274 : int i;
6275 :
6276 : for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
6277 : if (i == cpu)
6278 : continue;
6279 :
6280 : if (need_resched())
6281 : break;
6282 :
6283 : if (try_steal_cookie(cpu, i))
6284 : return true;
6285 : }
6286 :
6287 : return false;
6288 : }
6289 :
6290 : static void sched_core_balance(struct rq *rq)
6291 : {
6292 : struct sched_domain *sd;
6293 : int cpu = cpu_of(rq);
6294 :
6295 : preempt_disable();
6296 : rcu_read_lock();
6297 : raw_spin_rq_unlock_irq(rq);
6298 : for_each_domain(cpu, sd) {
6299 : if (need_resched())
6300 : break;
6301 :
6302 : if (steal_cookie_task(cpu, sd))
6303 : break;
6304 : }
6305 : raw_spin_rq_lock_irq(rq);
6306 : rcu_read_unlock();
6307 : preempt_enable();
6308 : }
6309 :
6310 : static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
6311 :
6312 : static void queue_core_balance(struct rq *rq)
6313 : {
6314 : if (!sched_core_enabled(rq))
6315 : return;
6316 :
6317 : if (!rq->core->core_cookie)
6318 : return;
6319 :
6320 : if (!rq->nr_running) /* not forced idle */
6321 : return;
6322 :
6323 : queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6324 : }
6325 :
6326 : static void sched_core_cpu_starting(unsigned int cpu)
6327 : {
6328 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6329 : struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6330 : unsigned long flags;
6331 : int t;
6332 :
6333 : sched_core_lock(cpu, &flags);
6334 :
6335 : WARN_ON_ONCE(rq->core != rq);
6336 :
6337 : /* if we're the first, we'll be our own leader */
6338 : if (cpumask_weight(smt_mask) == 1)
6339 : goto unlock;
6340 :
6341 : /* find the leader */
6342 : for_each_cpu(t, smt_mask) {
6343 : if (t == cpu)
6344 : continue;
6345 : rq = cpu_rq(t);
6346 : if (rq->core == rq) {
6347 : core_rq = rq;
6348 : break;
6349 : }
6350 : }
6351 :
6352 : if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6353 : goto unlock;
6354 :
6355 : /* install and validate core_rq */
6356 : for_each_cpu(t, smt_mask) {
6357 : rq = cpu_rq(t);
6358 :
6359 : if (t == cpu)
6360 : rq->core = core_rq;
6361 :
6362 : WARN_ON_ONCE(rq->core != core_rq);
6363 : }
6364 :
6365 : unlock:
6366 : sched_core_unlock(cpu, &flags);
6367 : }
6368 :
6369 : static void sched_core_cpu_deactivate(unsigned int cpu)
6370 : {
6371 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6372 : struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6373 : unsigned long flags;
6374 : int t;
6375 :
6376 : sched_core_lock(cpu, &flags);
6377 :
6378 : /* if we're the last man standing, nothing to do */
6379 : if (cpumask_weight(smt_mask) == 1) {
6380 : WARN_ON_ONCE(rq->core != rq);
6381 : goto unlock;
6382 : }
6383 :
6384 : /* if we're not the leader, nothing to do */
6385 : if (rq->core != rq)
6386 : goto unlock;
6387 :
6388 : /* find a new leader */
6389 : for_each_cpu(t, smt_mask) {
6390 : if (t == cpu)
6391 : continue;
6392 : core_rq = cpu_rq(t);
6393 : break;
6394 : }
6395 :
6396 : if (WARN_ON_ONCE(!core_rq)) /* impossible */
6397 : goto unlock;
6398 :
6399 : /* copy the shared state to the new leader */
6400 : core_rq->core_task_seq = rq->core_task_seq;
6401 : core_rq->core_pick_seq = rq->core_pick_seq;
6402 : core_rq->core_cookie = rq->core_cookie;
6403 : core_rq->core_forceidle_count = rq->core_forceidle_count;
6404 : core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6405 : core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6406 :
6407 : /*
6408 : * Accounting edge for forced idle is handled in pick_next_task().
6409 : * Don't need another one here, since the hotplug thread shouldn't
6410 : * have a cookie.
6411 : */
6412 : core_rq->core_forceidle_start = 0;
6413 :
6414 : /* install new leader */
6415 : for_each_cpu(t, smt_mask) {
6416 : rq = cpu_rq(t);
6417 : rq->core = core_rq;
6418 : }
6419 :
6420 : unlock:
6421 : sched_core_unlock(cpu, &flags);
6422 : }
6423 :
6424 : static inline void sched_core_cpu_dying(unsigned int cpu)
6425 : {
6426 : struct rq *rq = cpu_rq(cpu);
6427 :
6428 : if (rq->core != rq)
6429 : rq->core = rq;
6430 : }
6431 :
6432 : #else /* !CONFIG_SCHED_CORE */
6433 :
6434 : static inline void sched_core_cpu_starting(unsigned int cpu) {}
6435 : static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6436 : static inline void sched_core_cpu_dying(unsigned int cpu) {}
6437 :
6438 : static struct task_struct *
6439 : pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6440 : {
6441 2264 : return __pick_next_task(rq, prev, rf);
6442 : }
6443 :
6444 : #endif /* CONFIG_SCHED_CORE */
6445 :
6446 : /*
6447 : * Constants for the sched_mode argument of __schedule().
6448 : *
6449 : * The mode argument allows RT enabled kernels to differentiate a
6450 : * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6451 : * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6452 : * optimize the AND operation out and just check for zero.
6453 : */
6454 : #define SM_NONE 0x0
6455 : #define SM_PREEMPT 0x1
6456 : #define SM_RTLOCK_WAIT 0x2
6457 :
6458 : #ifndef CONFIG_PREEMPT_RT
6459 : # define SM_MASK_PREEMPT (~0U)
6460 : #else
6461 : # define SM_MASK_PREEMPT SM_PREEMPT
6462 : #endif
6463 :
6464 : /*
6465 : * __schedule() is the main scheduler function.
6466 : *
6467 : * The main means of driving the scheduler and thus entering this function are:
6468 : *
6469 : * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6470 : *
6471 : * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6472 : * paths. For example, see arch/x86/entry_64.S.
6473 : *
6474 : * To drive preemption between tasks, the scheduler sets the flag in timer
6475 : * interrupt handler scheduler_tick().
6476 : *
6477 : * 3. Wakeups don't really cause entry into schedule(). They add a
6478 : * task to the run-queue and that's it.
6479 : *
6480 : * Now, if the new task added to the run-queue preempts the current
6481 : * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6482 : * called on the nearest possible occasion:
6483 : *
6484 : * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6485 : *
6486 : * - in syscall or exception context, at the next outmost
6487 : * preempt_enable(). (this might be as soon as the wake_up()'s
6488 : * spin_unlock()!)
6489 : *
6490 : * - in IRQ context, return from interrupt-handler to
6491 : * preemptible context
6492 : *
6493 : * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6494 : * then at the next:
6495 : *
6496 : * - cond_resched() call
6497 : * - explicit schedule() call
6498 : * - return from syscall or exception to user-space
6499 : * - return from interrupt-handler to user-space
6500 : *
6501 : * WARNING: must be called with preemption disabled!
6502 : */
6503 2264 : static void __sched notrace __schedule(unsigned int sched_mode)
6504 : {
6505 : struct task_struct *prev, *next;
6506 : unsigned long *switch_count;
6507 : unsigned long prev_state;
6508 : struct rq_flags rf;
6509 : struct rq *rq;
6510 : int cpu;
6511 :
6512 2264 : cpu = smp_processor_id();
6513 2264 : rq = cpu_rq(cpu);
6514 2264 : prev = rq->curr;
6515 :
6516 4528 : schedule_debug(prev, !!sched_mode);
6517 :
6518 : if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6519 : hrtick_clear(rq);
6520 :
6521 : local_irq_disable();
6522 2264 : rcu_note_context_switch(!!sched_mode);
6523 :
6524 : /*
6525 : * Make sure that signal_pending_state()->signal_pending() below
6526 : * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6527 : * done by the caller to avoid the race with signal_wake_up():
6528 : *
6529 : * __set_current_state(@state) signal_wake_up()
6530 : * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6531 : * wake_up_state(p, state)
6532 : * LOCK rq->lock LOCK p->pi_state
6533 : * smp_mb__after_spinlock() smp_mb__after_spinlock()
6534 : * if (signal_pending_state()) if (p->state & @state)
6535 : *
6536 : * Also, the membarrier system call requires a full memory barrier
6537 : * after coming from user-space, before storing to rq->curr.
6538 : */
6539 2264 : rq_lock(rq, &rf);
6540 : smp_mb__after_spinlock();
6541 :
6542 : /* Promote REQ to ACT */
6543 2264 : rq->clock_update_flags <<= 1;
6544 2264 : update_rq_clock(rq);
6545 :
6546 2264 : switch_count = &prev->nivcsw;
6547 :
6548 : /*
6549 : * We must load prev->state once (task_struct::state is volatile), such
6550 : * that we form a control dependency vs deactivate_task() below.
6551 : */
6552 2264 : prev_state = READ_ONCE(prev->__state);
6553 2264 : if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6554 2174 : if (signal_pending_state(prev_state, prev)) {
6555 0 : WRITE_ONCE(prev->__state, TASK_RUNNING);
6556 : } else {
6557 2174 : prev->sched_contributes_to_load =
6558 : (prev_state & TASK_UNINTERRUPTIBLE) &&
6559 2174 : !(prev_state & TASK_NOLOAD) &&
6560 : !(prev_state & TASK_FROZEN);
6561 :
6562 2174 : if (prev->sched_contributes_to_load)
6563 1029 : rq->nr_uninterruptible++;
6564 :
6565 : /*
6566 : * __schedule() ttwu()
6567 : * prev_state = prev->state; if (p->on_rq && ...)
6568 : * if (prev_state) goto out;
6569 : * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6570 : * p->state = TASK_WAKING
6571 : *
6572 : * Where __schedule() and ttwu() have matching control dependencies.
6573 : *
6574 : * After this, schedule() must not care about p->state any more.
6575 : */
6576 2174 : deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6577 :
6578 2174 : if (prev->in_iowait) {
6579 0 : atomic_inc(&rq->nr_iowait);
6580 : delayacct_blkio_start();
6581 : }
6582 : }
6583 2174 : switch_count = &prev->nvcsw;
6584 : }
6585 :
6586 2264 : next = pick_next_task(rq, prev, &rf);
6587 2264 : clear_tsk_need_resched(prev);
6588 : clear_preempt_need_resched();
6589 : #ifdef CONFIG_SCHED_DEBUG
6590 2264 : rq->last_seen_need_resched_ns = 0;
6591 : #endif
6592 :
6593 2264 : if (likely(prev != next)) {
6594 2264 : rq->nr_switches++;
6595 : /*
6596 : * RCU users of rcu_dereference(rq->curr) may not see
6597 : * changes to task_struct made by pick_next_task().
6598 : */
6599 2264 : RCU_INIT_POINTER(rq->curr, next);
6600 : /*
6601 : * The membarrier system call requires each architecture
6602 : * to have a full memory barrier after updating
6603 : * rq->curr, before returning to user-space.
6604 : *
6605 : * Here are the schemes providing that barrier on the
6606 : * various architectures:
6607 : * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6608 : * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6609 : * - finish_lock_switch() for weakly-ordered
6610 : * architectures where spin_unlock is a full barrier,
6611 : * - switch_to() for arm64 (weakly-ordered, spin_unlock
6612 : * is a RELEASE barrier),
6613 : */
6614 2264 : ++*switch_count;
6615 :
6616 2264 : migrate_disable_switch(rq, prev);
6617 2264 : psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6618 :
6619 2264 : trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6620 :
6621 : /* Also unlocks the rq: */
6622 1916 : rq = context_switch(rq, prev, next, &rf);
6623 : } else {
6624 0 : rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6625 :
6626 0 : rq_unpin_lock(rq, &rf);
6627 0 : __balance_callbacks(rq);
6628 0 : raw_spin_rq_unlock_irq(rq);
6629 : }
6630 1916 : }
6631 :
6632 333 : void __noreturn do_task_dead(void)
6633 : {
6634 : /* Causes final put_task_struct in finish_task_switch(): */
6635 1665 : set_special_state(TASK_DEAD);
6636 :
6637 : /* Tell freezer to ignore us: */
6638 333 : current->flags |= PF_NOFREEZE;
6639 :
6640 333 : __schedule(SM_NONE);
6641 0 : BUG();
6642 :
6643 : /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6644 : for (;;)
6645 : cpu_relax();
6646 : }
6647 :
6648 1842 : static inline void sched_submit_work(struct task_struct *tsk)
6649 : {
6650 : unsigned int task_flags;
6651 :
6652 1842 : if (task_is_running(tsk))
6653 : return;
6654 :
6655 1841 : task_flags = tsk->flags;
6656 : /*
6657 : * If a worker goes to sleep, notify and ask workqueue whether it
6658 : * wants to wake up a task to maintain concurrency.
6659 : */
6660 1841 : if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6661 61 : if (task_flags & PF_WQ_WORKER)
6662 61 : wq_worker_sleeping(tsk);
6663 : else
6664 0 : io_wq_worker_sleeping(tsk);
6665 : }
6666 :
6667 : /*
6668 : * spinlock and rwlock must not flush block requests. This will
6669 : * deadlock if the callback attempts to acquire a lock which is
6670 : * already acquired.
6671 : */
6672 1841 : SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6673 :
6674 : /*
6675 : * If we are going to sleep and we have plugged IO queued,
6676 : * make sure to submit it to avoid deadlocks.
6677 : */
6678 1841 : blk_flush_plug(tsk->plug, true);
6679 : }
6680 :
6681 1827 : static void sched_update_worker(struct task_struct *tsk)
6682 : {
6683 1827 : if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6684 52 : if (tsk->flags & PF_WQ_WORKER)
6685 52 : wq_worker_running(tsk);
6686 : else
6687 0 : io_wq_worker_running(tsk);
6688 : }
6689 1827 : }
6690 :
6691 1842 : asmlinkage __visible void __sched schedule(void)
6692 : {
6693 1842 : struct task_struct *tsk = current;
6694 :
6695 1842 : sched_submit_work(tsk);
6696 : do {
6697 1842 : preempt_disable();
6698 1842 : __schedule(SM_NONE);
6699 1827 : sched_preempt_enable_no_resched();
6700 1827 : } while (need_resched());
6701 1827 : sched_update_worker(tsk);
6702 1827 : }
6703 : EXPORT_SYMBOL(schedule);
6704 :
6705 : /*
6706 : * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6707 : * state (have scheduled out non-voluntarily) by making sure that all
6708 : * tasks have either left the run queue or have gone into user space.
6709 : * As idle tasks do not do either, they must not ever be preempted
6710 : * (schedule out non-voluntarily).
6711 : *
6712 : * schedule_idle() is similar to schedule_preempt_disable() except that it
6713 : * never enables preemption because it does not call sched_submit_work().
6714 : */
6715 0 : void __sched schedule_idle(void)
6716 : {
6717 : /*
6718 : * As this skips calling sched_submit_work(), which the idle task does
6719 : * regardless because that function is a nop when the task is in a
6720 : * TASK_RUNNING state, make sure this isn't used someplace that the
6721 : * current task can be in any other state. Note, idle is always in the
6722 : * TASK_RUNNING state.
6723 : */
6724 0 : WARN_ON_ONCE(current->__state);
6725 : do {
6726 0 : __schedule(SM_NONE);
6727 0 : } while (need_resched());
6728 0 : }
6729 :
6730 : #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6731 : asmlinkage __visible void __sched schedule_user(void)
6732 : {
6733 : /*
6734 : * If we come here after a random call to set_need_resched(),
6735 : * or we have been woken up remotely but the IPI has not yet arrived,
6736 : * we haven't yet exited the RCU idle mode. Do it here manually until
6737 : * we find a better solution.
6738 : *
6739 : * NB: There are buggy callers of this function. Ideally we
6740 : * should warn if prev_state != CONTEXT_USER, but that will trigger
6741 : * too frequently to make sense yet.
6742 : */
6743 : enum ctx_state prev_state = exception_enter();
6744 : schedule();
6745 : exception_exit(prev_state);
6746 : }
6747 : #endif
6748 :
6749 : /**
6750 : * schedule_preempt_disabled - called with preemption disabled
6751 : *
6752 : * Returns with preemption disabled. Note: preempt_count must be 1
6753 : */
6754 348 : void __sched schedule_preempt_disabled(void)
6755 : {
6756 348 : sched_preempt_enable_no_resched();
6757 348 : schedule();
6758 347 : preempt_disable();
6759 347 : }
6760 :
6761 : #ifdef CONFIG_PREEMPT_RT
6762 : void __sched notrace schedule_rtlock(void)
6763 : {
6764 : do {
6765 : preempt_disable();
6766 : __schedule(SM_RTLOCK_WAIT);
6767 : sched_preempt_enable_no_resched();
6768 : } while (need_resched());
6769 : }
6770 : NOKPROBE_SYMBOL(schedule_rtlock);
6771 : #endif
6772 :
6773 : static void __sched notrace preempt_schedule_common(void)
6774 : {
6775 : do {
6776 : /*
6777 : * Because the function tracer can trace preempt_count_sub()
6778 : * and it also uses preempt_enable/disable_notrace(), if
6779 : * NEED_RESCHED is set, the preempt_enable_notrace() called
6780 : * by the function tracer will call this function again and
6781 : * cause infinite recursion.
6782 : *
6783 : * Preemption must be disabled here before the function
6784 : * tracer can trace. Break up preempt_disable() into two
6785 : * calls. One to disable preemption without fear of being
6786 : * traced. The other to still record the preemption latency,
6787 : * which can also be traced by the function tracer.
6788 : */
6789 89 : preempt_disable_notrace();
6790 89 : preempt_latency_start(1);
6791 89 : __schedule(SM_PREEMPT);
6792 89 : preempt_latency_stop(1);
6793 89 : preempt_enable_no_resched_notrace();
6794 :
6795 : /*
6796 : * Check again in case we missed a preemption opportunity
6797 : * between schedule and now.
6798 : */
6799 89 : } while (need_resched());
6800 : }
6801 :
6802 : #ifdef CONFIG_PREEMPTION
6803 : /*
6804 : * This is the entry point to schedule() from in-kernel preemption
6805 : * off of preempt_enable.
6806 : */
6807 : asmlinkage __visible void __sched notrace preempt_schedule(void)
6808 : {
6809 : /*
6810 : * If there is a non-zero preempt_count or interrupts are disabled,
6811 : * we do not want to preempt the current task. Just return..
6812 : */
6813 : if (likely(!preemptible()))
6814 : return;
6815 : preempt_schedule_common();
6816 : }
6817 : NOKPROBE_SYMBOL(preempt_schedule);
6818 : EXPORT_SYMBOL(preempt_schedule);
6819 :
6820 : #ifdef CONFIG_PREEMPT_DYNAMIC
6821 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6822 : #ifndef preempt_schedule_dynamic_enabled
6823 : #define preempt_schedule_dynamic_enabled preempt_schedule
6824 : #define preempt_schedule_dynamic_disabled NULL
6825 : #endif
6826 : DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6827 : EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6828 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6829 : static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6830 : void __sched notrace dynamic_preempt_schedule(void)
6831 : {
6832 : if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6833 : return;
6834 : preempt_schedule();
6835 : }
6836 : NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6837 : EXPORT_SYMBOL(dynamic_preempt_schedule);
6838 : #endif
6839 : #endif
6840 :
6841 : /**
6842 : * preempt_schedule_notrace - preempt_schedule called by tracing
6843 : *
6844 : * The tracing infrastructure uses preempt_enable_notrace to prevent
6845 : * recursion and tracing preempt enabling caused by the tracing
6846 : * infrastructure itself. But as tracing can happen in areas coming
6847 : * from userspace or just about to enter userspace, a preempt enable
6848 : * can occur before user_exit() is called. This will cause the scheduler
6849 : * to be called when the system is still in usermode.
6850 : *
6851 : * To prevent this, the preempt_enable_notrace will use this function
6852 : * instead of preempt_schedule() to exit user context if needed before
6853 : * calling the scheduler.
6854 : */
6855 : asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6856 : {
6857 : enum ctx_state prev_ctx;
6858 :
6859 : if (likely(!preemptible()))
6860 : return;
6861 :
6862 : do {
6863 : /*
6864 : * Because the function tracer can trace preempt_count_sub()
6865 : * and it also uses preempt_enable/disable_notrace(), if
6866 : * NEED_RESCHED is set, the preempt_enable_notrace() called
6867 : * by the function tracer will call this function again and
6868 : * cause infinite recursion.
6869 : *
6870 : * Preemption must be disabled here before the function
6871 : * tracer can trace. Break up preempt_disable() into two
6872 : * calls. One to disable preemption without fear of being
6873 : * traced. The other to still record the preemption latency,
6874 : * which can also be traced by the function tracer.
6875 : */
6876 : preempt_disable_notrace();
6877 : preempt_latency_start(1);
6878 : /*
6879 : * Needs preempt disabled in case user_exit() is traced
6880 : * and the tracer calls preempt_enable_notrace() causing
6881 : * an infinite recursion.
6882 : */
6883 : prev_ctx = exception_enter();
6884 : __schedule(SM_PREEMPT);
6885 : exception_exit(prev_ctx);
6886 :
6887 : preempt_latency_stop(1);
6888 : preempt_enable_no_resched_notrace();
6889 : } while (need_resched());
6890 : }
6891 : EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6892 :
6893 : #ifdef CONFIG_PREEMPT_DYNAMIC
6894 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6895 : #ifndef preempt_schedule_notrace_dynamic_enabled
6896 : #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6897 : #define preempt_schedule_notrace_dynamic_disabled NULL
6898 : #endif
6899 : DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6900 : EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6901 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6902 : static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6903 : void __sched notrace dynamic_preempt_schedule_notrace(void)
6904 : {
6905 : if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6906 : return;
6907 : preempt_schedule_notrace();
6908 : }
6909 : NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6910 : EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6911 : #endif
6912 : #endif
6913 :
6914 : #endif /* CONFIG_PREEMPTION */
6915 :
6916 : /*
6917 : * This is the entry point to schedule() from kernel preemption
6918 : * off of irq context.
6919 : * Note, that this is called and return with irqs disabled. This will
6920 : * protect us against recursive calling from irq.
6921 : */
6922 0 : asmlinkage __visible void __sched preempt_schedule_irq(void)
6923 : {
6924 : enum ctx_state prev_state;
6925 :
6926 : /* Catch callers which need to be fixed */
6927 0 : BUG_ON(preempt_count() || !irqs_disabled());
6928 :
6929 : prev_state = exception_enter();
6930 :
6931 : do {
6932 0 : preempt_disable();
6933 : local_irq_enable();
6934 0 : __schedule(SM_PREEMPT);
6935 : local_irq_disable();
6936 0 : sched_preempt_enable_no_resched();
6937 0 : } while (need_resched());
6938 :
6939 : exception_exit(prev_state);
6940 0 : }
6941 :
6942 0 : int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6943 : void *key)
6944 : {
6945 0 : WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6946 0 : return try_to_wake_up(curr->private, mode, wake_flags);
6947 : }
6948 : EXPORT_SYMBOL(default_wake_function);
6949 :
6950 : static void __setscheduler_prio(struct task_struct *p, int prio)
6951 : {
6952 0 : if (dl_prio(prio))
6953 0 : p->sched_class = &dl_sched_class;
6954 0 : else if (rt_prio(prio))
6955 0 : p->sched_class = &rt_sched_class;
6956 : else
6957 0 : p->sched_class = &fair_sched_class;
6958 :
6959 0 : p->prio = prio;
6960 : }
6961 :
6962 : #ifdef CONFIG_RT_MUTEXES
6963 :
6964 : static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6965 : {
6966 0 : if (pi_task)
6967 0 : prio = min(prio, pi_task->prio);
6968 :
6969 : return prio;
6970 : }
6971 :
6972 : static inline int rt_effective_prio(struct task_struct *p, int prio)
6973 : {
6974 0 : struct task_struct *pi_task = rt_mutex_get_top_task(p);
6975 :
6976 0 : return __rt_effective_prio(pi_task, prio);
6977 : }
6978 :
6979 : /*
6980 : * rt_mutex_setprio - set the current priority of a task
6981 : * @p: task to boost
6982 : * @pi_task: donor task
6983 : *
6984 : * This function changes the 'effective' priority of a task. It does
6985 : * not touch ->normal_prio like __setscheduler().
6986 : *
6987 : * Used by the rt_mutex code to implement priority inheritance
6988 : * logic. Call site only calls if the priority of the task changed.
6989 : */
6990 0 : void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6991 : {
6992 0 : int prio, oldprio, queued, running, queue_flag =
6993 : DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6994 : const struct sched_class *prev_class;
6995 : struct rq_flags rf;
6996 : struct rq *rq;
6997 :
6998 : /* XXX used to be waiter->prio, not waiter->task->prio */
6999 0 : prio = __rt_effective_prio(pi_task, p->normal_prio);
7000 :
7001 : /*
7002 : * If nothing changed; bail early.
7003 : */
7004 0 : if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
7005 : return;
7006 :
7007 0 : rq = __task_rq_lock(p, &rf);
7008 0 : update_rq_clock(rq);
7009 : /*
7010 : * Set under pi_lock && rq->lock, such that the value can be used under
7011 : * either lock.
7012 : *
7013 : * Note that there is loads of tricky to make this pointer cache work
7014 : * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
7015 : * ensure a task is de-boosted (pi_task is set to NULL) before the
7016 : * task is allowed to run again (and can exit). This ensures the pointer
7017 : * points to a blocked task -- which guarantees the task is present.
7018 : */
7019 0 : p->pi_top_task = pi_task;
7020 :
7021 : /*
7022 : * For FIFO/RR we only need to set prio, if that matches we're done.
7023 : */
7024 0 : if (prio == p->prio && !dl_prio(prio))
7025 : goto out_unlock;
7026 :
7027 : /*
7028 : * Idle task boosting is a nono in general. There is one
7029 : * exception, when PREEMPT_RT and NOHZ is active:
7030 : *
7031 : * The idle task calls get_next_timer_interrupt() and holds
7032 : * the timer wheel base->lock on the CPU and another CPU wants
7033 : * to access the timer (probably to cancel it). We can safely
7034 : * ignore the boosting request, as the idle CPU runs this code
7035 : * with interrupts disabled and will complete the lock
7036 : * protected section without being interrupted. So there is no
7037 : * real need to boost.
7038 : */
7039 0 : if (unlikely(p == rq->idle)) {
7040 0 : WARN_ON(p != rq->curr);
7041 0 : WARN_ON(p->pi_blocked_on);
7042 : goto out_unlock;
7043 : }
7044 :
7045 0 : trace_sched_pi_setprio(p, pi_task);
7046 0 : oldprio = p->prio;
7047 :
7048 0 : if (oldprio == prio)
7049 0 : queue_flag &= ~DEQUEUE_MOVE;
7050 :
7051 0 : prev_class = p->sched_class;
7052 0 : queued = task_on_rq_queued(p);
7053 0 : running = task_current(rq, p);
7054 0 : if (queued)
7055 0 : dequeue_task(rq, p, queue_flag);
7056 0 : if (running)
7057 0 : put_prev_task(rq, p);
7058 :
7059 : /*
7060 : * Boosting condition are:
7061 : * 1. -rt task is running and holds mutex A
7062 : * --> -dl task blocks on mutex A
7063 : *
7064 : * 2. -dl task is running and holds mutex A
7065 : * --> -dl task blocks on mutex A and could preempt the
7066 : * running task
7067 : */
7068 0 : if (dl_prio(prio)) {
7069 0 : if (!dl_prio(p->normal_prio) ||
7070 0 : (pi_task && dl_prio(pi_task->prio) &&
7071 0 : dl_entity_preempt(&pi_task->dl, &p->dl))) {
7072 0 : p->dl.pi_se = pi_task->dl.pi_se;
7073 0 : queue_flag |= ENQUEUE_REPLENISH;
7074 : } else {
7075 0 : p->dl.pi_se = &p->dl;
7076 : }
7077 0 : } else if (rt_prio(prio)) {
7078 0 : if (dl_prio(oldprio))
7079 0 : p->dl.pi_se = &p->dl;
7080 0 : if (oldprio < prio)
7081 0 : queue_flag |= ENQUEUE_HEAD;
7082 : } else {
7083 0 : if (dl_prio(oldprio))
7084 0 : p->dl.pi_se = &p->dl;
7085 0 : if (rt_prio(oldprio))
7086 0 : p->rt.timeout = 0;
7087 : }
7088 :
7089 0 : __setscheduler_prio(p, prio);
7090 :
7091 0 : if (queued)
7092 0 : enqueue_task(rq, p, queue_flag);
7093 0 : if (running)
7094 : set_next_task(rq, p);
7095 :
7096 0 : check_class_changed(rq, p, prev_class, oldprio);
7097 : out_unlock:
7098 : /* Avoid rq from going away on us: */
7099 0 : preempt_disable();
7100 :
7101 0 : rq_unpin_lock(rq, &rf);
7102 0 : __balance_callbacks(rq);
7103 0 : raw_spin_rq_unlock(rq);
7104 :
7105 0 : preempt_enable();
7106 : }
7107 : #else
7108 : static inline int rt_effective_prio(struct task_struct *p, int prio)
7109 : {
7110 : return prio;
7111 : }
7112 : #endif
7113 :
7114 9 : void set_user_nice(struct task_struct *p, long nice)
7115 : {
7116 : bool queued, running;
7117 : int old_prio;
7118 : struct rq_flags rf;
7119 : struct rq *rq;
7120 :
7121 18 : if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
7122 4 : return;
7123 : /*
7124 : * We have to be careful, if called from sys_setpriority(),
7125 : * the task might be in the middle of scheduling on another CPU.
7126 : */
7127 5 : rq = task_rq_lock(p, &rf);
7128 5 : update_rq_clock(rq);
7129 :
7130 : /*
7131 : * The RT priorities are set via sched_setscheduler(), but we still
7132 : * allow the 'normal' nice value to be set - but as expected
7133 : * it won't have any effect on scheduling until the task is
7134 : * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7135 : */
7136 15 : if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7137 0 : p->static_prio = NICE_TO_PRIO(nice);
7138 0 : goto out_unlock;
7139 : }
7140 5 : queued = task_on_rq_queued(p);
7141 5 : running = task_current(rq, p);
7142 5 : if (queued)
7143 : dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7144 5 : if (running)
7145 4 : put_prev_task(rq, p);
7146 :
7147 5 : p->static_prio = NICE_TO_PRIO(nice);
7148 5 : set_load_weight(p, true);
7149 5 : old_prio = p->prio;
7150 5 : p->prio = effective_prio(p);
7151 :
7152 5 : if (queued)
7153 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7154 5 : if (running)
7155 : set_next_task(rq, p);
7156 :
7157 : /*
7158 : * If the task increased its priority or is running and
7159 : * lowered its priority, then reschedule its CPU:
7160 : */
7161 5 : p->sched_class->prio_changed(rq, p, old_prio);
7162 :
7163 : out_unlock:
7164 10 : task_rq_unlock(rq, p, &rf);
7165 : }
7166 : EXPORT_SYMBOL(set_user_nice);
7167 :
7168 : /*
7169 : * is_nice_reduction - check if nice value is an actual reduction
7170 : *
7171 : * Similar to can_nice() but does not perform a capability check.
7172 : *
7173 : * @p: task
7174 : * @nice: nice value
7175 : */
7176 : static bool is_nice_reduction(const struct task_struct *p, const int nice)
7177 : {
7178 : /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7179 0 : int nice_rlim = nice_to_rlimit(nice);
7180 :
7181 0 : return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7182 : }
7183 :
7184 : /*
7185 : * can_nice - check if a task can reduce its nice value
7186 : * @p: task
7187 : * @nice: nice value
7188 : */
7189 0 : int can_nice(const struct task_struct *p, const int nice)
7190 : {
7191 0 : return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7192 : }
7193 :
7194 : #ifdef __ARCH_WANT_SYS_NICE
7195 :
7196 : /*
7197 : * sys_nice - change the priority of the current process.
7198 : * @increment: priority increment
7199 : *
7200 : * sys_setpriority is a more generic, but much slower function that
7201 : * does similar things.
7202 : */
7203 0 : SYSCALL_DEFINE1(nice, int, increment)
7204 : {
7205 : long nice, retval;
7206 :
7207 : /*
7208 : * Setpriority might change our priority at the same moment.
7209 : * We don't have to worry. Conceptually one call occurs first
7210 : * and we have a single winner.
7211 : */
7212 0 : increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7213 0 : nice = task_nice(current) + increment;
7214 :
7215 0 : nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7216 0 : if (increment < 0 && !can_nice(current, nice))
7217 : return -EPERM;
7218 :
7219 0 : retval = security_task_setnice(current, nice);
7220 0 : if (retval)
7221 : return retval;
7222 :
7223 0 : set_user_nice(current, nice);
7224 0 : return 0;
7225 : }
7226 :
7227 : #endif
7228 :
7229 : /**
7230 : * task_prio - return the priority value of a given task.
7231 : * @p: the task in question.
7232 : *
7233 : * Return: The priority value as seen by users in /proc.
7234 : *
7235 : * sched policy return value kernel prio user prio/nice
7236 : *
7237 : * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7238 : * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7239 : * deadline -101 -1 0
7240 : */
7241 0 : int task_prio(const struct task_struct *p)
7242 : {
7243 0 : return p->prio - MAX_RT_PRIO;
7244 : }
7245 :
7246 : /**
7247 : * idle_cpu - is a given CPU idle currently?
7248 : * @cpu: the processor in question.
7249 : *
7250 : * Return: 1 if the CPU is currently idle. 0 otherwise.
7251 : */
7252 0 : int idle_cpu(int cpu)
7253 : {
7254 0 : struct rq *rq = cpu_rq(cpu);
7255 :
7256 0 : if (rq->curr != rq->idle)
7257 : return 0;
7258 :
7259 0 : if (rq->nr_running)
7260 : return 0;
7261 :
7262 : #ifdef CONFIG_SMP
7263 : if (rq->ttwu_pending)
7264 : return 0;
7265 : #endif
7266 :
7267 0 : return 1;
7268 : }
7269 :
7270 : /**
7271 : * available_idle_cpu - is a given CPU idle for enqueuing work.
7272 : * @cpu: the CPU in question.
7273 : *
7274 : * Return: 1 if the CPU is currently idle. 0 otherwise.
7275 : */
7276 0 : int available_idle_cpu(int cpu)
7277 : {
7278 0 : if (!idle_cpu(cpu))
7279 : return 0;
7280 :
7281 0 : if (vcpu_is_preempted(cpu))
7282 : return 0;
7283 :
7284 0 : return 1;
7285 : }
7286 :
7287 : /**
7288 : * idle_task - return the idle task for a given CPU.
7289 : * @cpu: the processor in question.
7290 : *
7291 : * Return: The idle task for the CPU @cpu.
7292 : */
7293 0 : struct task_struct *idle_task(int cpu)
7294 : {
7295 0 : return cpu_rq(cpu)->idle;
7296 : }
7297 :
7298 : #ifdef CONFIG_SMP
7299 : /*
7300 : * This function computes an effective utilization for the given CPU, to be
7301 : * used for frequency selection given the linear relation: f = u * f_max.
7302 : *
7303 : * The scheduler tracks the following metrics:
7304 : *
7305 : * cpu_util_{cfs,rt,dl,irq}()
7306 : * cpu_bw_dl()
7307 : *
7308 : * Where the cfs,rt and dl util numbers are tracked with the same metric and
7309 : * synchronized windows and are thus directly comparable.
7310 : *
7311 : * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7312 : * which excludes things like IRQ and steal-time. These latter are then accrued
7313 : * in the irq utilization.
7314 : *
7315 : * The DL bandwidth number otoh is not a measured metric but a value computed
7316 : * based on the task model parameters and gives the minimal utilization
7317 : * required to meet deadlines.
7318 : */
7319 : unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7320 : enum cpu_util_type type,
7321 : struct task_struct *p)
7322 : {
7323 : unsigned long dl_util, util, irq, max;
7324 : struct rq *rq = cpu_rq(cpu);
7325 :
7326 : max = arch_scale_cpu_capacity(cpu);
7327 :
7328 : if (!uclamp_is_used() &&
7329 : type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7330 : return max;
7331 : }
7332 :
7333 : /*
7334 : * Early check to see if IRQ/steal time saturates the CPU, can be
7335 : * because of inaccuracies in how we track these -- see
7336 : * update_irq_load_avg().
7337 : */
7338 : irq = cpu_util_irq(rq);
7339 : if (unlikely(irq >= max))
7340 : return max;
7341 :
7342 : /*
7343 : * Because the time spend on RT/DL tasks is visible as 'lost' time to
7344 : * CFS tasks and we use the same metric to track the effective
7345 : * utilization (PELT windows are synchronized) we can directly add them
7346 : * to obtain the CPU's actual utilization.
7347 : *
7348 : * CFS and RT utilization can be boosted or capped, depending on
7349 : * utilization clamp constraints requested by currently RUNNABLE
7350 : * tasks.
7351 : * When there are no CFS RUNNABLE tasks, clamps are released and
7352 : * frequency will be gracefully reduced with the utilization decay.
7353 : */
7354 : util = util_cfs + cpu_util_rt(rq);
7355 : if (type == FREQUENCY_UTIL)
7356 : util = uclamp_rq_util_with(rq, util, p);
7357 :
7358 : dl_util = cpu_util_dl(rq);
7359 :
7360 : /*
7361 : * For frequency selection we do not make cpu_util_dl() a permanent part
7362 : * of this sum because we want to use cpu_bw_dl() later on, but we need
7363 : * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7364 : * that we select f_max when there is no idle time.
7365 : *
7366 : * NOTE: numerical errors or stop class might cause us to not quite hit
7367 : * saturation when we should -- something for later.
7368 : */
7369 : if (util + dl_util >= max)
7370 : return max;
7371 :
7372 : /*
7373 : * OTOH, for energy computation we need the estimated running time, so
7374 : * include util_dl and ignore dl_bw.
7375 : */
7376 : if (type == ENERGY_UTIL)
7377 : util += dl_util;
7378 :
7379 : /*
7380 : * There is still idle time; further improve the number by using the
7381 : * irq metric. Because IRQ/steal time is hidden from the task clock we
7382 : * need to scale the task numbers:
7383 : *
7384 : * max - irq
7385 : * U' = irq + --------- * U
7386 : * max
7387 : */
7388 : util = scale_irq_capacity(util, irq, max);
7389 : util += irq;
7390 :
7391 : /*
7392 : * Bandwidth required by DEADLINE must always be granted while, for
7393 : * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7394 : * to gracefully reduce the frequency when no tasks show up for longer
7395 : * periods of time.
7396 : *
7397 : * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7398 : * bw_dl as requested freq. However, cpufreq is not yet ready for such
7399 : * an interface. So, we only do the latter for now.
7400 : */
7401 : if (type == FREQUENCY_UTIL)
7402 : util += cpu_bw_dl(rq);
7403 :
7404 : return min(max, util);
7405 : }
7406 :
7407 : unsigned long sched_cpu_util(int cpu)
7408 : {
7409 : return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
7410 : }
7411 : #endif /* CONFIG_SMP */
7412 :
7413 : /**
7414 : * find_process_by_pid - find a process with a matching PID value.
7415 : * @pid: the pid in question.
7416 : *
7417 : * The task of @pid, if found. %NULL otherwise.
7418 : */
7419 : static struct task_struct *find_process_by_pid(pid_t pid)
7420 : {
7421 0 : return pid ? find_task_by_vpid(pid) : current;
7422 : }
7423 :
7424 : /*
7425 : * sched_setparam() passes in -1 for its policy, to let the functions
7426 : * it calls know not to change it.
7427 : */
7428 : #define SETPARAM_POLICY -1
7429 :
7430 0 : static void __setscheduler_params(struct task_struct *p,
7431 : const struct sched_attr *attr)
7432 : {
7433 0 : int policy = attr->sched_policy;
7434 :
7435 0 : if (policy == SETPARAM_POLICY)
7436 0 : policy = p->policy;
7437 :
7438 0 : p->policy = policy;
7439 :
7440 0 : if (dl_policy(policy))
7441 0 : __setparam_dl(p, attr);
7442 0 : else if (fair_policy(policy))
7443 0 : p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7444 :
7445 : /*
7446 : * __sched_setscheduler() ensures attr->sched_priority == 0 when
7447 : * !rt_policy. Always setting this ensures that things like
7448 : * getparam()/getattr() don't report silly values for !rt tasks.
7449 : */
7450 0 : p->rt_priority = attr->sched_priority;
7451 0 : p->normal_prio = normal_prio(p);
7452 0 : set_load_weight(p, true);
7453 0 : }
7454 :
7455 : /*
7456 : * Check the target process has a UID that matches the current process's:
7457 : */
7458 : static bool check_same_owner(struct task_struct *p)
7459 : {
7460 0 : const struct cred *cred = current_cred(), *pcred;
7461 : bool match;
7462 :
7463 : rcu_read_lock();
7464 0 : pcred = __task_cred(p);
7465 0 : match = (uid_eq(cred->euid, pcred->euid) ||
7466 0 : uid_eq(cred->euid, pcred->uid));
7467 : rcu_read_unlock();
7468 : return match;
7469 : }
7470 :
7471 : /*
7472 : * Allow unprivileged RT tasks to decrease priority.
7473 : * Only issue a capable test if needed and only once to avoid an audit
7474 : * event on permitted non-privileged operations:
7475 : */
7476 0 : static int user_check_sched_setscheduler(struct task_struct *p,
7477 : const struct sched_attr *attr,
7478 : int policy, int reset_on_fork)
7479 : {
7480 0 : if (fair_policy(policy)) {
7481 0 : if (attr->sched_nice < task_nice(p) &&
7482 0 : !is_nice_reduction(p, attr->sched_nice))
7483 : goto req_priv;
7484 : }
7485 :
7486 0 : if (rt_policy(policy)) {
7487 0 : unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7488 :
7489 : /* Can't set/change the rt policy: */
7490 0 : if (policy != p->policy && !rlim_rtprio)
7491 : goto req_priv;
7492 :
7493 : /* Can't increase priority: */
7494 0 : if (attr->sched_priority > p->rt_priority &&
7495 0 : attr->sched_priority > rlim_rtprio)
7496 : goto req_priv;
7497 : }
7498 :
7499 : /*
7500 : * Can't set/change SCHED_DEADLINE policy at all for now
7501 : * (safest behavior); in the future we would like to allow
7502 : * unprivileged DL tasks to increase their relative deadline
7503 : * or reduce their runtime (both ways reducing utilization)
7504 : */
7505 0 : if (dl_policy(policy))
7506 : goto req_priv;
7507 :
7508 : /*
7509 : * Treat SCHED_IDLE as nice 20. Only allow a switch to
7510 : * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7511 : */
7512 0 : if (task_has_idle_policy(p) && !idle_policy(policy)) {
7513 0 : if (!is_nice_reduction(p, task_nice(p)))
7514 : goto req_priv;
7515 : }
7516 :
7517 : /* Can't change other user's priorities: */
7518 0 : if (!check_same_owner(p))
7519 : goto req_priv;
7520 :
7521 : /* Normal users shall not reset the sched_reset_on_fork flag: */
7522 0 : if (p->sched_reset_on_fork && !reset_on_fork)
7523 : goto req_priv;
7524 :
7525 : return 0;
7526 :
7527 : req_priv:
7528 0 : if (!capable(CAP_SYS_NICE))
7529 : return -EPERM;
7530 :
7531 : return 0;
7532 : }
7533 :
7534 346 : static int __sched_setscheduler(struct task_struct *p,
7535 : const struct sched_attr *attr,
7536 : bool user, bool pi)
7537 : {
7538 346 : int oldpolicy = -1, policy = attr->sched_policy;
7539 : int retval, oldprio, newprio, queued, running;
7540 : const struct sched_class *prev_class;
7541 : struct balance_callback *head;
7542 : struct rq_flags rf;
7543 : int reset_on_fork;
7544 346 : int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7545 : struct rq *rq;
7546 :
7547 : /* The pi code expects interrupts enabled */
7548 692 : BUG_ON(pi && in_interrupt());
7549 : recheck:
7550 : /* Double check policy once rq lock held: */
7551 346 : if (policy < 0) {
7552 0 : reset_on_fork = p->sched_reset_on_fork;
7553 0 : policy = oldpolicy = p->policy;
7554 : } else {
7555 346 : reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7556 :
7557 346 : if (!valid_policy(policy))
7558 : return -EINVAL;
7559 : }
7560 :
7561 346 : if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7562 : return -EINVAL;
7563 :
7564 : /*
7565 : * Valid priorities for SCHED_FIFO and SCHED_RR are
7566 : * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7567 : * SCHED_BATCH and SCHED_IDLE is 0.
7568 : */
7569 346 : if (attr->sched_priority > MAX_RT_PRIO-1)
7570 : return -EINVAL;
7571 692 : if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7572 346 : (rt_policy(policy) != (attr->sched_priority != 0)))
7573 : return -EINVAL;
7574 :
7575 346 : if (user) {
7576 0 : retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7577 0 : if (retval)
7578 : return retval;
7579 :
7580 0 : if (attr->sched_flags & SCHED_FLAG_SUGOV)
7581 : return -EINVAL;
7582 :
7583 0 : retval = security_task_setscheduler(p);
7584 0 : if (retval)
7585 : return retval;
7586 : }
7587 :
7588 : /* Update task specific "requested" clamps */
7589 346 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7590 : retval = uclamp_validate(p, attr);
7591 : if (retval)
7592 : return retval;
7593 : }
7594 :
7595 : if (pi)
7596 : cpuset_read_lock();
7597 :
7598 : /*
7599 : * Make sure no PI-waiters arrive (or leave) while we are
7600 : * changing the priority of the task:
7601 : *
7602 : * To be able to change p->policy safely, the appropriate
7603 : * runqueue lock must be held.
7604 : */
7605 346 : rq = task_rq_lock(p, &rf);
7606 346 : update_rq_clock(rq);
7607 :
7608 : /*
7609 : * Changing the policy of the stop threads its a very bad idea:
7610 : */
7611 346 : if (p == rq->stop) {
7612 : retval = -EINVAL;
7613 : goto unlock;
7614 : }
7615 :
7616 : /*
7617 : * If not changing anything there's no need to proceed further,
7618 : * but store a possible modification of reset_on_fork.
7619 : */
7620 346 : if (unlikely(policy == p->policy)) {
7621 692 : if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7622 : goto change;
7623 346 : if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7624 : goto change;
7625 346 : if (dl_policy(policy) && dl_param_changed(p, attr))
7626 : goto change;
7627 346 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7628 : goto change;
7629 :
7630 346 : p->sched_reset_on_fork = reset_on_fork;
7631 346 : retval = 0;
7632 346 : goto unlock;
7633 : }
7634 : change:
7635 :
7636 : if (user) {
7637 : #ifdef CONFIG_RT_GROUP_SCHED
7638 : /*
7639 : * Do not allow realtime tasks into groups that have no runtime
7640 : * assigned.
7641 : */
7642 : if (rt_bandwidth_enabled() && rt_policy(policy) &&
7643 : task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7644 : !task_group_is_autogroup(task_group(p))) {
7645 : retval = -EPERM;
7646 : goto unlock;
7647 : }
7648 : #endif
7649 : #ifdef CONFIG_SMP
7650 : if (dl_bandwidth_enabled() && dl_policy(policy) &&
7651 : !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7652 : cpumask_t *span = rq->rd->span;
7653 :
7654 : /*
7655 : * Don't allow tasks with an affinity mask smaller than
7656 : * the entire root_domain to become SCHED_DEADLINE. We
7657 : * will also fail if there's no bandwidth available.
7658 : */
7659 : if (!cpumask_subset(span, p->cpus_ptr) ||
7660 : rq->rd->dl_bw.bw == 0) {
7661 : retval = -EPERM;
7662 : goto unlock;
7663 : }
7664 : }
7665 : #endif
7666 : }
7667 :
7668 : /* Re-check policy now with rq lock held: */
7669 0 : if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7670 0 : policy = oldpolicy = -1;
7671 0 : task_rq_unlock(rq, p, &rf);
7672 : if (pi)
7673 : cpuset_read_unlock();
7674 : goto recheck;
7675 : }
7676 :
7677 : /*
7678 : * If setscheduling to SCHED_DEADLINE (or changing the parameters
7679 : * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7680 : * is available.
7681 : */
7682 0 : if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7683 : retval = -EBUSY;
7684 : goto unlock;
7685 : }
7686 :
7687 0 : p->sched_reset_on_fork = reset_on_fork;
7688 0 : oldprio = p->prio;
7689 :
7690 0 : newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7691 0 : if (pi) {
7692 : /*
7693 : * Take priority boosted tasks into account. If the new
7694 : * effective priority is unchanged, we just store the new
7695 : * normal parameters and do not touch the scheduler class and
7696 : * the runqueue. This will be done when the task deboost
7697 : * itself.
7698 : */
7699 0 : newprio = rt_effective_prio(p, newprio);
7700 0 : if (newprio == oldprio)
7701 0 : queue_flags &= ~DEQUEUE_MOVE;
7702 : }
7703 :
7704 0 : queued = task_on_rq_queued(p);
7705 0 : running = task_current(rq, p);
7706 0 : if (queued)
7707 0 : dequeue_task(rq, p, queue_flags);
7708 0 : if (running)
7709 0 : put_prev_task(rq, p);
7710 :
7711 0 : prev_class = p->sched_class;
7712 :
7713 0 : if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7714 0 : __setscheduler_params(p, attr);
7715 0 : __setscheduler_prio(p, newprio);
7716 : }
7717 0 : __setscheduler_uclamp(p, attr);
7718 :
7719 0 : if (queued) {
7720 : /*
7721 : * We enqueue to tail when the priority of a task is
7722 : * increased (user space view).
7723 : */
7724 0 : if (oldprio < p->prio)
7725 0 : queue_flags |= ENQUEUE_HEAD;
7726 :
7727 0 : enqueue_task(rq, p, queue_flags);
7728 : }
7729 0 : if (running)
7730 : set_next_task(rq, p);
7731 :
7732 0 : check_class_changed(rq, p, prev_class, oldprio);
7733 :
7734 : /* Avoid rq from going away on us: */
7735 0 : preempt_disable();
7736 0 : head = splice_balance_callbacks(rq);
7737 0 : task_rq_unlock(rq, p, &rf);
7738 :
7739 0 : if (pi) {
7740 : cpuset_read_unlock();
7741 0 : rt_mutex_adjust_pi(p);
7742 : }
7743 :
7744 : /* Run balance callbacks after we've adjusted the PI chain: */
7745 0 : balance_callbacks(rq, head);
7746 0 : preempt_enable();
7747 :
7748 0 : return 0;
7749 :
7750 : unlock:
7751 692 : task_rq_unlock(rq, p, &rf);
7752 : if (pi)
7753 : cpuset_read_unlock();
7754 346 : return retval;
7755 : }
7756 :
7757 346 : static int _sched_setscheduler(struct task_struct *p, int policy,
7758 : const struct sched_param *param, bool check)
7759 : {
7760 1038 : struct sched_attr attr = {
7761 : .sched_policy = policy,
7762 346 : .sched_priority = param->sched_priority,
7763 346 : .sched_nice = PRIO_TO_NICE(p->static_prio),
7764 : };
7765 :
7766 : /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7767 346 : if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7768 0 : attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7769 0 : policy &= ~SCHED_RESET_ON_FORK;
7770 0 : attr.sched_policy = policy;
7771 : }
7772 :
7773 346 : return __sched_setscheduler(p, &attr, check, true);
7774 : }
7775 : /**
7776 : * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7777 : * @p: the task in question.
7778 : * @policy: new policy.
7779 : * @param: structure containing the new RT priority.
7780 : *
7781 : * Use sched_set_fifo(), read its comment.
7782 : *
7783 : * Return: 0 on success. An error code otherwise.
7784 : *
7785 : * NOTE that the task may be already dead.
7786 : */
7787 0 : int sched_setscheduler(struct task_struct *p, int policy,
7788 : const struct sched_param *param)
7789 : {
7790 0 : return _sched_setscheduler(p, policy, param, true);
7791 : }
7792 :
7793 0 : int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7794 : {
7795 0 : return __sched_setscheduler(p, attr, true, true);
7796 : }
7797 :
7798 0 : int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7799 : {
7800 0 : return __sched_setscheduler(p, attr, false, true);
7801 : }
7802 : EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7803 :
7804 : /**
7805 : * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7806 : * @p: the task in question.
7807 : * @policy: new policy.
7808 : * @param: structure containing the new RT priority.
7809 : *
7810 : * Just like sched_setscheduler, only don't bother checking if the
7811 : * current context has permission. For example, this is needed in
7812 : * stop_machine(): we create temporary high priority worker threads,
7813 : * but our caller might not have that capability.
7814 : *
7815 : * Return: 0 on success. An error code otherwise.
7816 : */
7817 346 : int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7818 : const struct sched_param *param)
7819 : {
7820 346 : return _sched_setscheduler(p, policy, param, false);
7821 : }
7822 :
7823 : /*
7824 : * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7825 : * incapable of resource management, which is the one thing an OS really should
7826 : * be doing.
7827 : *
7828 : * This is of course the reason it is limited to privileged users only.
7829 : *
7830 : * Worse still; it is fundamentally impossible to compose static priority
7831 : * workloads. You cannot take two correctly working static prio workloads
7832 : * and smash them together and still expect them to work.
7833 : *
7834 : * For this reason 'all' FIFO tasks the kernel creates are basically at:
7835 : *
7836 : * MAX_RT_PRIO / 2
7837 : *
7838 : * The administrator _MUST_ configure the system, the kernel simply doesn't
7839 : * know enough information to make a sensible choice.
7840 : */
7841 0 : void sched_set_fifo(struct task_struct *p)
7842 : {
7843 0 : struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7844 0 : WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7845 0 : }
7846 : EXPORT_SYMBOL_GPL(sched_set_fifo);
7847 :
7848 : /*
7849 : * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7850 : */
7851 0 : void sched_set_fifo_low(struct task_struct *p)
7852 : {
7853 0 : struct sched_param sp = { .sched_priority = 1 };
7854 0 : WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7855 0 : }
7856 : EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7857 :
7858 0 : void sched_set_normal(struct task_struct *p, int nice)
7859 : {
7860 0 : struct sched_attr attr = {
7861 : .sched_policy = SCHED_NORMAL,
7862 : .sched_nice = nice,
7863 : };
7864 0 : WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7865 0 : }
7866 : EXPORT_SYMBOL_GPL(sched_set_normal);
7867 :
7868 : static int
7869 0 : do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7870 : {
7871 : struct sched_param lparam;
7872 : struct task_struct *p;
7873 : int retval;
7874 :
7875 0 : if (!param || pid < 0)
7876 : return -EINVAL;
7877 0 : if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7878 : return -EFAULT;
7879 :
7880 : rcu_read_lock();
7881 0 : retval = -ESRCH;
7882 0 : p = find_process_by_pid(pid);
7883 0 : if (likely(p))
7884 : get_task_struct(p);
7885 : rcu_read_unlock();
7886 :
7887 0 : if (likely(p)) {
7888 0 : retval = sched_setscheduler(p, policy, &lparam);
7889 0 : put_task_struct(p);
7890 : }
7891 :
7892 : return retval;
7893 : }
7894 :
7895 : /*
7896 : * Mimics kernel/events/core.c perf_copy_attr().
7897 : */
7898 0 : static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7899 : {
7900 : u32 size;
7901 : int ret;
7902 :
7903 : /* Zero the full structure, so that a short copy will be nice: */
7904 0 : memset(attr, 0, sizeof(*attr));
7905 :
7906 0 : ret = get_user(size, &uattr->size);
7907 0 : if (ret)
7908 : return ret;
7909 :
7910 : /* ABI compatibility quirk: */
7911 0 : if (!size)
7912 0 : size = SCHED_ATTR_SIZE_VER0;
7913 0 : if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7914 : goto err_size;
7915 :
7916 0 : ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7917 0 : if (ret) {
7918 0 : if (ret == -E2BIG)
7919 : goto err_size;
7920 : return ret;
7921 : }
7922 :
7923 0 : if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7924 : size < SCHED_ATTR_SIZE_VER1)
7925 : return -EINVAL;
7926 :
7927 : /*
7928 : * XXX: Do we want to be lenient like existing syscalls; or do we want
7929 : * to be strict and return an error on out-of-bounds values?
7930 : */
7931 0 : attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7932 :
7933 0 : return 0;
7934 :
7935 : err_size:
7936 0 : put_user(sizeof(*attr), &uattr->size);
7937 : return -E2BIG;
7938 : }
7939 :
7940 0 : static void get_params(struct task_struct *p, struct sched_attr *attr)
7941 : {
7942 0 : if (task_has_dl_policy(p))
7943 0 : __getparam_dl(p, attr);
7944 0 : else if (task_has_rt_policy(p))
7945 0 : attr->sched_priority = p->rt_priority;
7946 : else
7947 0 : attr->sched_nice = task_nice(p);
7948 0 : }
7949 :
7950 : /**
7951 : * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7952 : * @pid: the pid in question.
7953 : * @policy: new policy.
7954 : * @param: structure containing the new RT priority.
7955 : *
7956 : * Return: 0 on success. An error code otherwise.
7957 : */
7958 0 : SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7959 : {
7960 0 : if (policy < 0)
7961 : return -EINVAL;
7962 :
7963 0 : return do_sched_setscheduler(pid, policy, param);
7964 : }
7965 :
7966 : /**
7967 : * sys_sched_setparam - set/change the RT priority of a thread
7968 : * @pid: the pid in question.
7969 : * @param: structure containing the new RT priority.
7970 : *
7971 : * Return: 0 on success. An error code otherwise.
7972 : */
7973 0 : SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7974 : {
7975 0 : return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7976 : }
7977 :
7978 : /**
7979 : * sys_sched_setattr - same as above, but with extended sched_attr
7980 : * @pid: the pid in question.
7981 : * @uattr: structure containing the extended parameters.
7982 : * @flags: for future extension.
7983 : */
7984 0 : SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7985 : unsigned int, flags)
7986 : {
7987 : struct sched_attr attr;
7988 : struct task_struct *p;
7989 : int retval;
7990 :
7991 0 : if (!uattr || pid < 0 || flags)
7992 : return -EINVAL;
7993 :
7994 0 : retval = sched_copy_attr(uattr, &attr);
7995 0 : if (retval)
7996 0 : return retval;
7997 :
7998 0 : if ((int)attr.sched_policy < 0)
7999 : return -EINVAL;
8000 0 : if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
8001 0 : attr.sched_policy = SETPARAM_POLICY;
8002 :
8003 : rcu_read_lock();
8004 0 : retval = -ESRCH;
8005 0 : p = find_process_by_pid(pid);
8006 0 : if (likely(p))
8007 : get_task_struct(p);
8008 : rcu_read_unlock();
8009 :
8010 0 : if (likely(p)) {
8011 0 : if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
8012 0 : get_params(p, &attr);
8013 0 : retval = sched_setattr(p, &attr);
8014 0 : put_task_struct(p);
8015 : }
8016 :
8017 0 : return retval;
8018 : }
8019 :
8020 : /**
8021 : * sys_sched_getscheduler - get the policy (scheduling class) of a thread
8022 : * @pid: the pid in question.
8023 : *
8024 : * Return: On success, the policy of the thread. Otherwise, a negative error
8025 : * code.
8026 : */
8027 0 : SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
8028 : {
8029 : struct task_struct *p;
8030 : int retval;
8031 :
8032 0 : if (pid < 0)
8033 : return -EINVAL;
8034 :
8035 0 : retval = -ESRCH;
8036 : rcu_read_lock();
8037 0 : p = find_process_by_pid(pid);
8038 0 : if (p) {
8039 0 : retval = security_task_getscheduler(p);
8040 : if (!retval)
8041 0 : retval = p->policy
8042 0 : | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
8043 : }
8044 : rcu_read_unlock();
8045 0 : return retval;
8046 : }
8047 :
8048 : /**
8049 : * sys_sched_getparam - get the RT priority of a thread
8050 : * @pid: the pid in question.
8051 : * @param: structure containing the RT priority.
8052 : *
8053 : * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8054 : * code.
8055 : */
8056 0 : SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
8057 : {
8058 0 : struct sched_param lp = { .sched_priority = 0 };
8059 : struct task_struct *p;
8060 : int retval;
8061 :
8062 0 : if (!param || pid < 0)
8063 : return -EINVAL;
8064 :
8065 : rcu_read_lock();
8066 0 : p = find_process_by_pid(pid);
8067 0 : retval = -ESRCH;
8068 0 : if (!p)
8069 : goto out_unlock;
8070 :
8071 0 : retval = security_task_getscheduler(p);
8072 : if (retval)
8073 : goto out_unlock;
8074 :
8075 0 : if (task_has_rt_policy(p))
8076 0 : lp.sched_priority = p->rt_priority;
8077 0 : rcu_read_unlock();
8078 :
8079 : /*
8080 : * This one might sleep, we cannot do it with a spinlock held ...
8081 : */
8082 0 : retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
8083 :
8084 0 : return retval;
8085 :
8086 : out_unlock:
8087 : rcu_read_unlock();
8088 0 : return retval;
8089 : }
8090 :
8091 : /*
8092 : * Copy the kernel size attribute structure (which might be larger
8093 : * than what user-space knows about) to user-space.
8094 : *
8095 : * Note that all cases are valid: user-space buffer can be larger or
8096 : * smaller than the kernel-space buffer. The usual case is that both
8097 : * have the same size.
8098 : */
8099 : static int
8100 0 : sched_attr_copy_to_user(struct sched_attr __user *uattr,
8101 : struct sched_attr *kattr,
8102 : unsigned int usize)
8103 : {
8104 0 : unsigned int ksize = sizeof(*kattr);
8105 :
8106 0 : if (!access_ok(uattr, usize))
8107 : return -EFAULT;
8108 :
8109 : /*
8110 : * sched_getattr() ABI forwards and backwards compatibility:
8111 : *
8112 : * If usize == ksize then we just copy everything to user-space and all is good.
8113 : *
8114 : * If usize < ksize then we only copy as much as user-space has space for,
8115 : * this keeps ABI compatibility as well. We skip the rest.
8116 : *
8117 : * If usize > ksize then user-space is using a newer version of the ABI,
8118 : * which part the kernel doesn't know about. Just ignore it - tooling can
8119 : * detect the kernel's knowledge of attributes from the attr->size value
8120 : * which is set to ksize in this case.
8121 : */
8122 0 : kattr->size = min(usize, ksize);
8123 :
8124 0 : if (copy_to_user(uattr, kattr, kattr->size))
8125 : return -EFAULT;
8126 :
8127 0 : return 0;
8128 : }
8129 :
8130 : /**
8131 : * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8132 : * @pid: the pid in question.
8133 : * @uattr: structure containing the extended parameters.
8134 : * @usize: sizeof(attr) for fwd/bwd comp.
8135 : * @flags: for future extension.
8136 : */
8137 0 : SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8138 : unsigned int, usize, unsigned int, flags)
8139 : {
8140 0 : struct sched_attr kattr = { };
8141 : struct task_struct *p;
8142 : int retval;
8143 :
8144 0 : if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8145 0 : usize < SCHED_ATTR_SIZE_VER0 || flags)
8146 : return -EINVAL;
8147 :
8148 : rcu_read_lock();
8149 0 : p = find_process_by_pid(pid);
8150 0 : retval = -ESRCH;
8151 0 : if (!p)
8152 : goto out_unlock;
8153 :
8154 0 : retval = security_task_getscheduler(p);
8155 : if (retval)
8156 : goto out_unlock;
8157 :
8158 0 : kattr.sched_policy = p->policy;
8159 0 : if (p->sched_reset_on_fork)
8160 0 : kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8161 0 : get_params(p, &kattr);
8162 0 : kattr.sched_flags &= SCHED_FLAG_ALL;
8163 :
8164 : #ifdef CONFIG_UCLAMP_TASK
8165 : /*
8166 : * This could race with another potential updater, but this is fine
8167 : * because it'll correctly read the old or the new value. We don't need
8168 : * to guarantee who wins the race as long as it doesn't return garbage.
8169 : */
8170 : kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8171 : kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8172 : #endif
8173 :
8174 : rcu_read_unlock();
8175 :
8176 0 : return sched_attr_copy_to_user(uattr, &kattr, usize);
8177 :
8178 : out_unlock:
8179 : rcu_read_unlock();
8180 0 : return retval;
8181 : }
8182 :
8183 : #ifdef CONFIG_SMP
8184 : int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8185 : {
8186 : int ret = 0;
8187 :
8188 : /*
8189 : * If the task isn't a deadline task or admission control is
8190 : * disabled then we don't care about affinity changes.
8191 : */
8192 : if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8193 : return 0;
8194 :
8195 : /*
8196 : * Since bandwidth control happens on root_domain basis,
8197 : * if admission test is enabled, we only admit -deadline
8198 : * tasks allowed to run on all the CPUs in the task's
8199 : * root_domain.
8200 : */
8201 : rcu_read_lock();
8202 : if (!cpumask_subset(task_rq(p)->rd->span, mask))
8203 : ret = -EBUSY;
8204 : rcu_read_unlock();
8205 : return ret;
8206 : }
8207 : #endif
8208 :
8209 : static int
8210 0 : __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
8211 : {
8212 : int retval;
8213 : cpumask_var_t cpus_allowed, new_mask;
8214 :
8215 0 : if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8216 : return -ENOMEM;
8217 :
8218 0 : if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8219 : retval = -ENOMEM;
8220 : goto out_free_cpus_allowed;
8221 : }
8222 :
8223 0 : cpuset_cpus_allowed(p, cpus_allowed);
8224 0 : cpumask_and(new_mask, ctx->new_mask, cpus_allowed);
8225 :
8226 0 : ctx->new_mask = new_mask;
8227 0 : ctx->flags |= SCA_CHECK;
8228 :
8229 0 : retval = dl_task_check_affinity(p, new_mask);
8230 : if (retval)
8231 : goto out_free_new_mask;
8232 :
8233 0 : retval = __set_cpus_allowed_ptr(p, ctx);
8234 0 : if (retval)
8235 : goto out_free_new_mask;
8236 :
8237 0 : cpuset_cpus_allowed(p, cpus_allowed);
8238 0 : if (!cpumask_subset(new_mask, cpus_allowed)) {
8239 : /*
8240 : * We must have raced with a concurrent cpuset update.
8241 : * Just reset the cpumask to the cpuset's cpus_allowed.
8242 : */
8243 0 : cpumask_copy(new_mask, cpus_allowed);
8244 :
8245 : /*
8246 : * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8247 : * will restore the previous user_cpus_ptr value.
8248 : *
8249 : * In the unlikely event a previous user_cpus_ptr exists,
8250 : * we need to further restrict the mask to what is allowed
8251 : * by that old user_cpus_ptr.
8252 : */
8253 0 : if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
8254 0 : bool empty = !cpumask_and(new_mask, new_mask,
8255 0 : ctx->user_mask);
8256 :
8257 0 : if (WARN_ON_ONCE(empty))
8258 : cpumask_copy(new_mask, cpus_allowed);
8259 : }
8260 0 : __set_cpus_allowed_ptr(p, ctx);
8261 : retval = -EINVAL;
8262 : }
8263 :
8264 : out_free_new_mask:
8265 0 : free_cpumask_var(new_mask);
8266 : out_free_cpus_allowed:
8267 0 : free_cpumask_var(cpus_allowed);
8268 : return retval;
8269 : }
8270 :
8271 0 : long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8272 : {
8273 : struct affinity_context ac;
8274 : struct cpumask *user_mask;
8275 : struct task_struct *p;
8276 : int retval;
8277 :
8278 : rcu_read_lock();
8279 :
8280 0 : p = find_process_by_pid(pid);
8281 0 : if (!p) {
8282 : rcu_read_unlock();
8283 0 : return -ESRCH;
8284 : }
8285 :
8286 : /* Prevent p going away */
8287 0 : get_task_struct(p);
8288 : rcu_read_unlock();
8289 :
8290 0 : if (p->flags & PF_NO_SETAFFINITY) {
8291 : retval = -EINVAL;
8292 : goto out_put_task;
8293 : }
8294 :
8295 0 : if (!check_same_owner(p)) {
8296 : rcu_read_lock();
8297 0 : if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8298 : rcu_read_unlock();
8299 0 : retval = -EPERM;
8300 0 : goto out_put_task;
8301 : }
8302 : rcu_read_unlock();
8303 : }
8304 :
8305 0 : retval = security_task_setscheduler(p);
8306 0 : if (retval)
8307 : goto out_put_task;
8308 :
8309 : /*
8310 : * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8311 : * alloc_user_cpus_ptr() returns NULL.
8312 : */
8313 0 : user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
8314 : if (user_mask) {
8315 : cpumask_copy(user_mask, in_mask);
8316 : } else if (IS_ENABLED(CONFIG_SMP)) {
8317 : retval = -ENOMEM;
8318 : goto out_put_task;
8319 : }
8320 :
8321 0 : ac = (struct affinity_context){
8322 : .new_mask = in_mask,
8323 : .user_mask = user_mask,
8324 : .flags = SCA_USER,
8325 : };
8326 :
8327 0 : retval = __sched_setaffinity(p, &ac);
8328 0 : kfree(ac.user_mask);
8329 :
8330 : out_put_task:
8331 0 : put_task_struct(p);
8332 0 : return retval;
8333 : }
8334 :
8335 0 : static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8336 : struct cpumask *new_mask)
8337 : {
8338 0 : if (len < cpumask_size())
8339 : cpumask_clear(new_mask);
8340 0 : else if (len > cpumask_size())
8341 0 : len = cpumask_size();
8342 :
8343 0 : return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8344 : }
8345 :
8346 : /**
8347 : * sys_sched_setaffinity - set the CPU affinity of a process
8348 : * @pid: pid of the process
8349 : * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8350 : * @user_mask_ptr: user-space pointer to the new CPU mask
8351 : *
8352 : * Return: 0 on success. An error code otherwise.
8353 : */
8354 0 : SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8355 : unsigned long __user *, user_mask_ptr)
8356 : {
8357 : cpumask_var_t new_mask;
8358 : int retval;
8359 :
8360 0 : if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8361 : return -ENOMEM;
8362 :
8363 0 : retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8364 0 : if (retval == 0)
8365 0 : retval = sched_setaffinity(pid, new_mask);
8366 0 : free_cpumask_var(new_mask);
8367 0 : return retval;
8368 : }
8369 :
8370 0 : long sched_getaffinity(pid_t pid, struct cpumask *mask)
8371 : {
8372 : struct task_struct *p;
8373 : unsigned long flags;
8374 : int retval;
8375 :
8376 : rcu_read_lock();
8377 :
8378 0 : retval = -ESRCH;
8379 0 : p = find_process_by_pid(pid);
8380 0 : if (!p)
8381 : goto out_unlock;
8382 :
8383 0 : retval = security_task_getscheduler(p);
8384 : if (retval)
8385 : goto out_unlock;
8386 :
8387 0 : raw_spin_lock_irqsave(&p->pi_lock, flags);
8388 0 : cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8389 0 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8390 :
8391 : out_unlock:
8392 : rcu_read_unlock();
8393 :
8394 0 : return retval;
8395 : }
8396 :
8397 : /**
8398 : * sys_sched_getaffinity - get the CPU affinity of a process
8399 : * @pid: pid of the process
8400 : * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8401 : * @user_mask_ptr: user-space pointer to hold the current CPU mask
8402 : *
8403 : * Return: size of CPU mask copied to user_mask_ptr on success. An
8404 : * error code otherwise.
8405 : */
8406 0 : SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8407 : unsigned long __user *, user_mask_ptr)
8408 : {
8409 : int ret;
8410 : cpumask_var_t mask;
8411 :
8412 0 : if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8413 : return -EINVAL;
8414 0 : if (len & (sizeof(unsigned long)-1))
8415 : return -EINVAL;
8416 :
8417 0 : if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8418 : return -ENOMEM;
8419 :
8420 0 : ret = sched_getaffinity(pid, mask);
8421 0 : if (ret == 0) {
8422 0 : unsigned int retlen = min(len, cpumask_size());
8423 :
8424 0 : if (copy_to_user(user_mask_ptr, mask, retlen))
8425 : ret = -EFAULT;
8426 : else
8427 0 : ret = retlen;
8428 : }
8429 0 : free_cpumask_var(mask);
8430 :
8431 0 : return ret;
8432 : }
8433 :
8434 0 : static void do_sched_yield(void)
8435 : {
8436 : struct rq_flags rf;
8437 : struct rq *rq;
8438 :
8439 0 : rq = this_rq_lock_irq(&rf);
8440 :
8441 : schedstat_inc(rq->yld_count);
8442 0 : current->sched_class->yield_task(rq);
8443 :
8444 0 : preempt_disable();
8445 0 : rq_unlock_irq(rq, &rf);
8446 0 : sched_preempt_enable_no_resched();
8447 :
8448 0 : schedule();
8449 0 : }
8450 :
8451 : /**
8452 : * sys_sched_yield - yield the current processor to other threads.
8453 : *
8454 : * This function yields the current CPU to other tasks. If there are no
8455 : * other threads running on this CPU then this function will return.
8456 : *
8457 : * Return: 0.
8458 : */
8459 0 : SYSCALL_DEFINE0(sched_yield)
8460 : {
8461 0 : do_sched_yield();
8462 0 : return 0;
8463 : }
8464 :
8465 : #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8466 406935 : int __sched __cond_resched(void)
8467 : {
8468 406935 : if (should_resched(0)) {
8469 : preempt_schedule_common();
8470 : return 1;
8471 : }
8472 : /*
8473 : * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8474 : * whether the current CPU is in an RCU read-side critical section,
8475 : * so the tick can report quiescent states even for CPUs looping
8476 : * in kernel context. In contrast, in non-preemptible kernels,
8477 : * RCU readers leave no in-memory hints, which means that CPU-bound
8478 : * processes executing in kernel context might never report an
8479 : * RCU quiescent state. Therefore, the following code causes
8480 : * cond_resched() to report a quiescent state, but only when RCU
8481 : * is in urgent need of one.
8482 : */
8483 : #ifndef CONFIG_PREEMPT_RCU
8484 : rcu_all_qs();
8485 : #endif
8486 406846 : return 0;
8487 : }
8488 : EXPORT_SYMBOL(__cond_resched);
8489 : #endif
8490 :
8491 : #ifdef CONFIG_PREEMPT_DYNAMIC
8492 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8493 : #define cond_resched_dynamic_enabled __cond_resched
8494 : #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8495 : DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8496 : EXPORT_STATIC_CALL_TRAMP(cond_resched);
8497 :
8498 : #define might_resched_dynamic_enabled __cond_resched
8499 : #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8500 : DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8501 : EXPORT_STATIC_CALL_TRAMP(might_resched);
8502 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8503 : static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8504 : int __sched dynamic_cond_resched(void)
8505 : {
8506 : if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8507 : return 0;
8508 : return __cond_resched();
8509 : }
8510 : EXPORT_SYMBOL(dynamic_cond_resched);
8511 :
8512 : static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8513 : int __sched dynamic_might_resched(void)
8514 : {
8515 : if (!static_branch_unlikely(&sk_dynamic_might_resched))
8516 : return 0;
8517 : return __cond_resched();
8518 : }
8519 : EXPORT_SYMBOL(dynamic_might_resched);
8520 : #endif
8521 : #endif
8522 :
8523 : /*
8524 : * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8525 : * call schedule, and on return reacquire the lock.
8526 : *
8527 : * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8528 : * operations here to prevent schedule() from being called twice (once via
8529 : * spin_unlock(), once by hand).
8530 : */
8531 5 : int __cond_resched_lock(spinlock_t *lock)
8532 : {
8533 5 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
8534 5 : int ret = 0;
8535 :
8536 : lockdep_assert_held(lock);
8537 :
8538 5 : if (spin_needbreak(lock) || resched) {
8539 3 : spin_unlock(lock);
8540 3 : if (!_cond_resched())
8541 : cpu_relax();
8542 3 : ret = 1;
8543 : spin_lock(lock);
8544 : }
8545 5 : return ret;
8546 : }
8547 : EXPORT_SYMBOL(__cond_resched_lock);
8548 :
8549 0 : int __cond_resched_rwlock_read(rwlock_t *lock)
8550 : {
8551 0 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
8552 0 : int ret = 0;
8553 :
8554 : lockdep_assert_held_read(lock);
8555 :
8556 0 : if (rwlock_needbreak(lock) || resched) {
8557 0 : read_unlock(lock);
8558 0 : if (!_cond_resched())
8559 : cpu_relax();
8560 0 : ret = 1;
8561 0 : read_lock(lock);
8562 : }
8563 0 : return ret;
8564 : }
8565 : EXPORT_SYMBOL(__cond_resched_rwlock_read);
8566 :
8567 0 : int __cond_resched_rwlock_write(rwlock_t *lock)
8568 : {
8569 0 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
8570 0 : int ret = 0;
8571 :
8572 : lockdep_assert_held_write(lock);
8573 :
8574 0 : if (rwlock_needbreak(lock) || resched) {
8575 0 : write_unlock(lock);
8576 0 : if (!_cond_resched())
8577 : cpu_relax();
8578 0 : ret = 1;
8579 0 : write_lock(lock);
8580 : }
8581 0 : return ret;
8582 : }
8583 : EXPORT_SYMBOL(__cond_resched_rwlock_write);
8584 :
8585 : #ifdef CONFIG_PREEMPT_DYNAMIC
8586 :
8587 : #ifdef CONFIG_GENERIC_ENTRY
8588 : #include <linux/entry-common.h>
8589 : #endif
8590 :
8591 : /*
8592 : * SC:cond_resched
8593 : * SC:might_resched
8594 : * SC:preempt_schedule
8595 : * SC:preempt_schedule_notrace
8596 : * SC:irqentry_exit_cond_resched
8597 : *
8598 : *
8599 : * NONE:
8600 : * cond_resched <- __cond_resched
8601 : * might_resched <- RET0
8602 : * preempt_schedule <- NOP
8603 : * preempt_schedule_notrace <- NOP
8604 : * irqentry_exit_cond_resched <- NOP
8605 : *
8606 : * VOLUNTARY:
8607 : * cond_resched <- __cond_resched
8608 : * might_resched <- __cond_resched
8609 : * preempt_schedule <- NOP
8610 : * preempt_schedule_notrace <- NOP
8611 : * irqentry_exit_cond_resched <- NOP
8612 : *
8613 : * FULL:
8614 : * cond_resched <- RET0
8615 : * might_resched <- RET0
8616 : * preempt_schedule <- preempt_schedule
8617 : * preempt_schedule_notrace <- preempt_schedule_notrace
8618 : * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8619 : */
8620 :
8621 : enum {
8622 : preempt_dynamic_undefined = -1,
8623 : preempt_dynamic_none,
8624 : preempt_dynamic_voluntary,
8625 : preempt_dynamic_full,
8626 : };
8627 :
8628 : int preempt_dynamic_mode = preempt_dynamic_undefined;
8629 :
8630 : int sched_dynamic_mode(const char *str)
8631 : {
8632 : if (!strcmp(str, "none"))
8633 : return preempt_dynamic_none;
8634 :
8635 : if (!strcmp(str, "voluntary"))
8636 : return preempt_dynamic_voluntary;
8637 :
8638 : if (!strcmp(str, "full"))
8639 : return preempt_dynamic_full;
8640 :
8641 : return -EINVAL;
8642 : }
8643 :
8644 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8645 : #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8646 : #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8647 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8648 : #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8649 : #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8650 : #else
8651 : #error "Unsupported PREEMPT_DYNAMIC mechanism"
8652 : #endif
8653 :
8654 : void sched_dynamic_update(int mode)
8655 : {
8656 : /*
8657 : * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8658 : * the ZERO state, which is invalid.
8659 : */
8660 : preempt_dynamic_enable(cond_resched);
8661 : preempt_dynamic_enable(might_resched);
8662 : preempt_dynamic_enable(preempt_schedule);
8663 : preempt_dynamic_enable(preempt_schedule_notrace);
8664 : preempt_dynamic_enable(irqentry_exit_cond_resched);
8665 :
8666 : switch (mode) {
8667 : case preempt_dynamic_none:
8668 : preempt_dynamic_enable(cond_resched);
8669 : preempt_dynamic_disable(might_resched);
8670 : preempt_dynamic_disable(preempt_schedule);
8671 : preempt_dynamic_disable(preempt_schedule_notrace);
8672 : preempt_dynamic_disable(irqentry_exit_cond_resched);
8673 : pr_info("Dynamic Preempt: none\n");
8674 : break;
8675 :
8676 : case preempt_dynamic_voluntary:
8677 : preempt_dynamic_enable(cond_resched);
8678 : preempt_dynamic_enable(might_resched);
8679 : preempt_dynamic_disable(preempt_schedule);
8680 : preempt_dynamic_disable(preempt_schedule_notrace);
8681 : preempt_dynamic_disable(irqentry_exit_cond_resched);
8682 : pr_info("Dynamic Preempt: voluntary\n");
8683 : break;
8684 :
8685 : case preempt_dynamic_full:
8686 : preempt_dynamic_disable(cond_resched);
8687 : preempt_dynamic_disable(might_resched);
8688 : preempt_dynamic_enable(preempt_schedule);
8689 : preempt_dynamic_enable(preempt_schedule_notrace);
8690 : preempt_dynamic_enable(irqentry_exit_cond_resched);
8691 : pr_info("Dynamic Preempt: full\n");
8692 : break;
8693 : }
8694 :
8695 : preempt_dynamic_mode = mode;
8696 : }
8697 :
8698 : static int __init setup_preempt_mode(char *str)
8699 : {
8700 : int mode = sched_dynamic_mode(str);
8701 : if (mode < 0) {
8702 : pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8703 : return 0;
8704 : }
8705 :
8706 : sched_dynamic_update(mode);
8707 : return 1;
8708 : }
8709 : __setup("preempt=", setup_preempt_mode);
8710 :
8711 : static void __init preempt_dynamic_init(void)
8712 : {
8713 : if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8714 : if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8715 : sched_dynamic_update(preempt_dynamic_none);
8716 : } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8717 : sched_dynamic_update(preempt_dynamic_voluntary);
8718 : } else {
8719 : /* Default static call setting, nothing to do */
8720 : WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8721 : preempt_dynamic_mode = preempt_dynamic_full;
8722 : pr_info("Dynamic Preempt: full\n");
8723 : }
8724 : }
8725 : }
8726 :
8727 : #define PREEMPT_MODEL_ACCESSOR(mode) \
8728 : bool preempt_model_##mode(void) \
8729 : { \
8730 : WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8731 : return preempt_dynamic_mode == preempt_dynamic_##mode; \
8732 : } \
8733 : EXPORT_SYMBOL_GPL(preempt_model_##mode)
8734 :
8735 : PREEMPT_MODEL_ACCESSOR(none);
8736 : PREEMPT_MODEL_ACCESSOR(voluntary);
8737 : PREEMPT_MODEL_ACCESSOR(full);
8738 :
8739 : #else /* !CONFIG_PREEMPT_DYNAMIC */
8740 :
8741 : static inline void preempt_dynamic_init(void) { }
8742 :
8743 : #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8744 :
8745 : /**
8746 : * yield - yield the current processor to other threads.
8747 : *
8748 : * Do not ever use this function, there's a 99% chance you're doing it wrong.
8749 : *
8750 : * The scheduler is at all times free to pick the calling task as the most
8751 : * eligible task to run, if removing the yield() call from your code breaks
8752 : * it, it's already broken.
8753 : *
8754 : * Typical broken usage is:
8755 : *
8756 : * while (!event)
8757 : * yield();
8758 : *
8759 : * where one assumes that yield() will let 'the other' process run that will
8760 : * make event true. If the current task is a SCHED_FIFO task that will never
8761 : * happen. Never use yield() as a progress guarantee!!
8762 : *
8763 : * If you want to use yield() to wait for something, use wait_event().
8764 : * If you want to use yield() to be 'nice' for others, use cond_resched().
8765 : * If you still want to use yield(), do not!
8766 : */
8767 0 : void __sched yield(void)
8768 : {
8769 0 : set_current_state(TASK_RUNNING);
8770 0 : do_sched_yield();
8771 0 : }
8772 : EXPORT_SYMBOL(yield);
8773 :
8774 : /**
8775 : * yield_to - yield the current processor to another thread in
8776 : * your thread group, or accelerate that thread toward the
8777 : * processor it's on.
8778 : * @p: target task
8779 : * @preempt: whether task preemption is allowed or not
8780 : *
8781 : * It's the caller's job to ensure that the target task struct
8782 : * can't go away on us before we can do any checks.
8783 : *
8784 : * Return:
8785 : * true (>0) if we indeed boosted the target task.
8786 : * false (0) if we failed to boost the target.
8787 : * -ESRCH if there's no task to yield to.
8788 : */
8789 0 : int __sched yield_to(struct task_struct *p, bool preempt)
8790 : {
8791 0 : struct task_struct *curr = current;
8792 : struct rq *rq, *p_rq;
8793 : unsigned long flags;
8794 0 : int yielded = 0;
8795 :
8796 0 : local_irq_save(flags);
8797 0 : rq = this_rq();
8798 :
8799 : again:
8800 0 : p_rq = task_rq(p);
8801 : /*
8802 : * If we're the only runnable task on the rq and target rq also
8803 : * has only one task, there's absolutely no point in yielding.
8804 : */
8805 0 : if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8806 : yielded = -ESRCH;
8807 : goto out_irq;
8808 : }
8809 :
8810 0 : double_rq_lock(rq, p_rq);
8811 0 : if (task_rq(p) != p_rq) {
8812 : double_rq_unlock(rq, p_rq);
8813 : goto again;
8814 : }
8815 :
8816 0 : if (!curr->sched_class->yield_to_task)
8817 : goto out_unlock;
8818 :
8819 0 : if (curr->sched_class != p->sched_class)
8820 : goto out_unlock;
8821 :
8822 0 : if (task_on_cpu(p_rq, p) || !task_is_running(p))
8823 : goto out_unlock;
8824 :
8825 0 : yielded = curr->sched_class->yield_to_task(rq, p);
8826 : if (yielded) {
8827 : schedstat_inc(rq->yld_count);
8828 : /*
8829 : * Make p's CPU reschedule; pick_next_entity takes care of
8830 : * fairness.
8831 : */
8832 : if (preempt && rq != p_rq)
8833 : resched_curr(p_rq);
8834 : }
8835 :
8836 : out_unlock:
8837 0 : double_rq_unlock(rq, p_rq);
8838 : out_irq:
8839 0 : local_irq_restore(flags);
8840 :
8841 0 : if (yielded > 0)
8842 0 : schedule();
8843 :
8844 0 : return yielded;
8845 : }
8846 : EXPORT_SYMBOL_GPL(yield_to);
8847 :
8848 0 : int io_schedule_prepare(void)
8849 : {
8850 0 : int old_iowait = current->in_iowait;
8851 :
8852 0 : current->in_iowait = 1;
8853 0 : blk_flush_plug(current->plug, true);
8854 0 : return old_iowait;
8855 : }
8856 :
8857 0 : void io_schedule_finish(int token)
8858 : {
8859 0 : current->in_iowait = token;
8860 0 : }
8861 :
8862 : /*
8863 : * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8864 : * that process accounting knows that this is a task in IO wait state.
8865 : */
8866 0 : long __sched io_schedule_timeout(long timeout)
8867 : {
8868 : int token;
8869 : long ret;
8870 :
8871 0 : token = io_schedule_prepare();
8872 0 : ret = schedule_timeout(timeout);
8873 0 : io_schedule_finish(token);
8874 :
8875 0 : return ret;
8876 : }
8877 : EXPORT_SYMBOL(io_schedule_timeout);
8878 :
8879 0 : void __sched io_schedule(void)
8880 : {
8881 : int token;
8882 :
8883 0 : token = io_schedule_prepare();
8884 0 : schedule();
8885 0 : io_schedule_finish(token);
8886 0 : }
8887 : EXPORT_SYMBOL(io_schedule);
8888 :
8889 : /**
8890 : * sys_sched_get_priority_max - return maximum RT priority.
8891 : * @policy: scheduling class.
8892 : *
8893 : * Return: On success, this syscall returns the maximum
8894 : * rt_priority that can be used by a given scheduling class.
8895 : * On failure, a negative error code is returned.
8896 : */
8897 0 : SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8898 : {
8899 0 : int ret = -EINVAL;
8900 :
8901 : switch (policy) {
8902 : case SCHED_FIFO:
8903 : case SCHED_RR:
8904 0 : ret = MAX_RT_PRIO-1;
8905 : break;
8906 : case SCHED_DEADLINE:
8907 : case SCHED_NORMAL:
8908 : case SCHED_BATCH:
8909 : case SCHED_IDLE:
8910 : ret = 0;
8911 : break;
8912 : }
8913 0 : return ret;
8914 : }
8915 :
8916 : /**
8917 : * sys_sched_get_priority_min - return minimum RT priority.
8918 : * @policy: scheduling class.
8919 : *
8920 : * Return: On success, this syscall returns the minimum
8921 : * rt_priority that can be used by a given scheduling class.
8922 : * On failure, a negative error code is returned.
8923 : */
8924 0 : SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8925 : {
8926 0 : int ret = -EINVAL;
8927 :
8928 : switch (policy) {
8929 : case SCHED_FIFO:
8930 : case SCHED_RR:
8931 0 : ret = 1;
8932 : break;
8933 : case SCHED_DEADLINE:
8934 : case SCHED_NORMAL:
8935 : case SCHED_BATCH:
8936 : case SCHED_IDLE:
8937 : ret = 0;
8938 : }
8939 0 : return ret;
8940 : }
8941 :
8942 0 : static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8943 : {
8944 : struct task_struct *p;
8945 : unsigned int time_slice;
8946 : struct rq_flags rf;
8947 : struct rq *rq;
8948 : int retval;
8949 :
8950 0 : if (pid < 0)
8951 : return -EINVAL;
8952 :
8953 0 : retval = -ESRCH;
8954 : rcu_read_lock();
8955 0 : p = find_process_by_pid(pid);
8956 0 : if (!p)
8957 : goto out_unlock;
8958 :
8959 0 : retval = security_task_getscheduler(p);
8960 : if (retval)
8961 : goto out_unlock;
8962 :
8963 0 : rq = task_rq_lock(p, &rf);
8964 0 : time_slice = 0;
8965 0 : if (p->sched_class->get_rr_interval)
8966 0 : time_slice = p->sched_class->get_rr_interval(rq, p);
8967 0 : task_rq_unlock(rq, p, &rf);
8968 :
8969 : rcu_read_unlock();
8970 0 : jiffies_to_timespec64(time_slice, t);
8971 0 : return 0;
8972 :
8973 : out_unlock:
8974 : rcu_read_unlock();
8975 0 : return retval;
8976 : }
8977 :
8978 : /**
8979 : * sys_sched_rr_get_interval - return the default timeslice of a process.
8980 : * @pid: pid of the process.
8981 : * @interval: userspace pointer to the timeslice value.
8982 : *
8983 : * this syscall writes the default timeslice value of a given process
8984 : * into the user-space timespec buffer. A value of '0' means infinity.
8985 : *
8986 : * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8987 : * an error code.
8988 : */
8989 0 : SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8990 : struct __kernel_timespec __user *, interval)
8991 : {
8992 : struct timespec64 t;
8993 0 : int retval = sched_rr_get_interval(pid, &t);
8994 :
8995 0 : if (retval == 0)
8996 0 : retval = put_timespec64(&t, interval);
8997 :
8998 0 : return retval;
8999 : }
9000 :
9001 : #ifdef CONFIG_COMPAT_32BIT_TIME
9002 : SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
9003 : struct old_timespec32 __user *, interval)
9004 : {
9005 : struct timespec64 t;
9006 : int retval = sched_rr_get_interval(pid, &t);
9007 :
9008 : if (retval == 0)
9009 : retval = put_old_timespec32(&t, interval);
9010 : return retval;
9011 : }
9012 : #endif
9013 :
9014 0 : void sched_show_task(struct task_struct *p)
9015 : {
9016 0 : unsigned long free = 0;
9017 : int ppid;
9018 :
9019 0 : if (!try_get_task_stack(p))
9020 : return;
9021 :
9022 0 : pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
9023 :
9024 0 : if (task_is_running(p))
9025 0 : pr_cont(" running task ");
9026 : #ifdef CONFIG_DEBUG_STACK_USAGE
9027 : free = stack_not_used(p);
9028 : #endif
9029 0 : ppid = 0;
9030 : rcu_read_lock();
9031 0 : if (pid_alive(p))
9032 0 : ppid = task_pid_nr(rcu_dereference(p->real_parent));
9033 : rcu_read_unlock();
9034 0 : pr_cont(" stack:%-5lu pid:%-5d ppid:%-6d flags:0x%08lx\n",
9035 : free, task_pid_nr(p), ppid,
9036 : read_task_thread_flags(p));
9037 :
9038 0 : print_worker_info(KERN_INFO, p);
9039 0 : print_stop_info(KERN_INFO, p);
9040 0 : show_stack(p, NULL, KERN_INFO);
9041 0 : put_task_stack(p);
9042 : }
9043 : EXPORT_SYMBOL_GPL(sched_show_task);
9044 :
9045 : static inline bool
9046 : state_filter_match(unsigned long state_filter, struct task_struct *p)
9047 : {
9048 0 : unsigned int state = READ_ONCE(p->__state);
9049 :
9050 : /* no filter, everything matches */
9051 0 : if (!state_filter)
9052 : return true;
9053 :
9054 : /* filter, but doesn't match */
9055 0 : if (!(state & state_filter))
9056 : return false;
9057 :
9058 : /*
9059 : * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9060 : * TASK_KILLABLE).
9061 : */
9062 0 : if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
9063 : return false;
9064 :
9065 : return true;
9066 : }
9067 :
9068 :
9069 0 : void show_state_filter(unsigned int state_filter)
9070 : {
9071 : struct task_struct *g, *p;
9072 :
9073 : rcu_read_lock();
9074 0 : for_each_process_thread(g, p) {
9075 : /*
9076 : * reset the NMI-timeout, listing all files on a slow
9077 : * console might take a lot of time:
9078 : * Also, reset softlockup watchdogs on all CPUs, because
9079 : * another CPU might be blocked waiting for us to process
9080 : * an IPI.
9081 : */
9082 : touch_nmi_watchdog();
9083 : touch_all_softlockup_watchdogs();
9084 0 : if (state_filter_match(state_filter, p))
9085 0 : sched_show_task(p);
9086 : }
9087 :
9088 : #ifdef CONFIG_SCHED_DEBUG
9089 0 : if (!state_filter)
9090 0 : sysrq_sched_debug_show();
9091 : #endif
9092 : rcu_read_unlock();
9093 : /*
9094 : * Only show locks if all tasks are dumped:
9095 : */
9096 : if (!state_filter)
9097 : debug_show_all_locks();
9098 0 : }
9099 :
9100 : /**
9101 : * init_idle - set up an idle thread for a given CPU
9102 : * @idle: task in question
9103 : * @cpu: CPU the idle task belongs to
9104 : *
9105 : * NOTE: this function does not set the idle thread's NEED_RESCHED
9106 : * flag, to make booting more robust.
9107 : */
9108 1 : void __init init_idle(struct task_struct *idle, int cpu)
9109 : {
9110 : #ifdef CONFIG_SMP
9111 : struct affinity_context ac = (struct affinity_context) {
9112 : .new_mask = cpumask_of(cpu),
9113 : .flags = 0,
9114 : };
9115 : #endif
9116 1 : struct rq *rq = cpu_rq(cpu);
9117 : unsigned long flags;
9118 :
9119 1 : __sched_fork(0, idle);
9120 :
9121 1 : raw_spin_lock_irqsave(&idle->pi_lock, flags);
9122 1 : raw_spin_rq_lock(rq);
9123 :
9124 1 : idle->__state = TASK_RUNNING;
9125 1 : idle->se.exec_start = sched_clock();
9126 : /*
9127 : * PF_KTHREAD should already be set at this point; regardless, make it
9128 : * look like a proper per-CPU kthread.
9129 : */
9130 1 : idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
9131 1 : kthread_set_per_cpu(idle, cpu);
9132 :
9133 : #ifdef CONFIG_SMP
9134 : /*
9135 : * It's possible that init_idle() gets called multiple times on a task,
9136 : * in that case do_set_cpus_allowed() will not do the right thing.
9137 : *
9138 : * And since this is boot we can forgo the serialization.
9139 : */
9140 : set_cpus_allowed_common(idle, &ac);
9141 : #endif
9142 : /*
9143 : * We're having a chicken and egg problem, even though we are
9144 : * holding rq->lock, the CPU isn't yet set to this CPU so the
9145 : * lockdep check in task_group() will fail.
9146 : *
9147 : * Similar case to sched_fork(). / Alternatively we could
9148 : * use task_rq_lock() here and obtain the other rq->lock.
9149 : *
9150 : * Silence PROVE_RCU
9151 : */
9152 : rcu_read_lock();
9153 1 : __set_task_cpu(idle, cpu);
9154 : rcu_read_unlock();
9155 :
9156 1 : rq->idle = idle;
9157 1 : rcu_assign_pointer(rq->curr, idle);
9158 1 : idle->on_rq = TASK_ON_RQ_QUEUED;
9159 : #ifdef CONFIG_SMP
9160 : idle->on_cpu = 1;
9161 : #endif
9162 1 : raw_spin_rq_unlock(rq);
9163 2 : raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
9164 :
9165 : /* Set the preempt count _outside_ the spinlocks! */
9166 1 : init_idle_preempt_count(idle, cpu);
9167 :
9168 : /*
9169 : * The idle tasks have their own, simple scheduling class:
9170 : */
9171 1 : idle->sched_class = &idle_sched_class;
9172 1 : ftrace_graph_init_idle_task(idle, cpu);
9173 1 : vtime_init_idle(idle, cpu);
9174 : #ifdef CONFIG_SMP
9175 : sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9176 : #endif
9177 1 : }
9178 :
9179 : #ifdef CONFIG_SMP
9180 :
9181 : int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9182 : const struct cpumask *trial)
9183 : {
9184 : int ret = 1;
9185 :
9186 : if (cpumask_empty(cur))
9187 : return ret;
9188 :
9189 : ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9190 :
9191 : return ret;
9192 : }
9193 :
9194 : int task_can_attach(struct task_struct *p,
9195 : const struct cpumask *cs_effective_cpus)
9196 : {
9197 : int ret = 0;
9198 :
9199 : /*
9200 : * Kthreads which disallow setaffinity shouldn't be moved
9201 : * to a new cpuset; we don't want to change their CPU
9202 : * affinity and isolating such threads by their set of
9203 : * allowed nodes is unnecessary. Thus, cpusets are not
9204 : * applicable for such threads. This prevents checking for
9205 : * success of set_cpus_allowed_ptr() on all attached tasks
9206 : * before cpus_mask may be changed.
9207 : */
9208 : if (p->flags & PF_NO_SETAFFINITY) {
9209 : ret = -EINVAL;
9210 : goto out;
9211 : }
9212 :
9213 : if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
9214 : cs_effective_cpus)) {
9215 : int cpu = cpumask_any_and(cpu_active_mask, cs_effective_cpus);
9216 :
9217 : if (unlikely(cpu >= nr_cpu_ids))
9218 : return -EINVAL;
9219 : ret = dl_cpu_busy(cpu, p);
9220 : }
9221 :
9222 : out:
9223 : return ret;
9224 : }
9225 :
9226 : bool sched_smp_initialized __read_mostly;
9227 :
9228 : #ifdef CONFIG_NUMA_BALANCING
9229 : /* Migrate current task p to target_cpu */
9230 : int migrate_task_to(struct task_struct *p, int target_cpu)
9231 : {
9232 : struct migration_arg arg = { p, target_cpu };
9233 : int curr_cpu = task_cpu(p);
9234 :
9235 : if (curr_cpu == target_cpu)
9236 : return 0;
9237 :
9238 : if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9239 : return -EINVAL;
9240 :
9241 : /* TODO: This is not properly updating schedstats */
9242 :
9243 : trace_sched_move_numa(p, curr_cpu, target_cpu);
9244 : return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9245 : }
9246 :
9247 : /*
9248 : * Requeue a task on a given node and accurately track the number of NUMA
9249 : * tasks on the runqueues
9250 : */
9251 : void sched_setnuma(struct task_struct *p, int nid)
9252 : {
9253 : bool queued, running;
9254 : struct rq_flags rf;
9255 : struct rq *rq;
9256 :
9257 : rq = task_rq_lock(p, &rf);
9258 : queued = task_on_rq_queued(p);
9259 : running = task_current(rq, p);
9260 :
9261 : if (queued)
9262 : dequeue_task(rq, p, DEQUEUE_SAVE);
9263 : if (running)
9264 : put_prev_task(rq, p);
9265 :
9266 : p->numa_preferred_nid = nid;
9267 :
9268 : if (queued)
9269 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9270 : if (running)
9271 : set_next_task(rq, p);
9272 : task_rq_unlock(rq, p, &rf);
9273 : }
9274 : #endif /* CONFIG_NUMA_BALANCING */
9275 :
9276 : #ifdef CONFIG_HOTPLUG_CPU
9277 : /*
9278 : * Ensure that the idle task is using init_mm right before its CPU goes
9279 : * offline.
9280 : */
9281 : void idle_task_exit(void)
9282 : {
9283 : struct mm_struct *mm = current->active_mm;
9284 :
9285 : BUG_ON(cpu_online(smp_processor_id()));
9286 : BUG_ON(current != this_rq()->idle);
9287 :
9288 : if (mm != &init_mm) {
9289 : switch_mm(mm, &init_mm, current);
9290 : finish_arch_post_lock_switch();
9291 : }
9292 :
9293 : /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9294 : }
9295 :
9296 : static int __balance_push_cpu_stop(void *arg)
9297 : {
9298 : struct task_struct *p = arg;
9299 : struct rq *rq = this_rq();
9300 : struct rq_flags rf;
9301 : int cpu;
9302 :
9303 : raw_spin_lock_irq(&p->pi_lock);
9304 : rq_lock(rq, &rf);
9305 :
9306 : update_rq_clock(rq);
9307 :
9308 : if (task_rq(p) == rq && task_on_rq_queued(p)) {
9309 : cpu = select_fallback_rq(rq->cpu, p);
9310 : rq = __migrate_task(rq, &rf, p, cpu);
9311 : }
9312 :
9313 : rq_unlock(rq, &rf);
9314 : raw_spin_unlock_irq(&p->pi_lock);
9315 :
9316 : put_task_struct(p);
9317 :
9318 : return 0;
9319 : }
9320 :
9321 : static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9322 :
9323 : /*
9324 : * Ensure we only run per-cpu kthreads once the CPU goes !active.
9325 : *
9326 : * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9327 : * effective when the hotplug motion is down.
9328 : */
9329 : static void balance_push(struct rq *rq)
9330 : {
9331 : struct task_struct *push_task = rq->curr;
9332 :
9333 : lockdep_assert_rq_held(rq);
9334 :
9335 : /*
9336 : * Ensure the thing is persistent until balance_push_set(.on = false);
9337 : */
9338 : rq->balance_callback = &balance_push_callback;
9339 :
9340 : /*
9341 : * Only active while going offline and when invoked on the outgoing
9342 : * CPU.
9343 : */
9344 : if (!cpu_dying(rq->cpu) || rq != this_rq())
9345 : return;
9346 :
9347 : /*
9348 : * Both the cpu-hotplug and stop task are in this case and are
9349 : * required to complete the hotplug process.
9350 : */
9351 : if (kthread_is_per_cpu(push_task) ||
9352 : is_migration_disabled(push_task)) {
9353 :
9354 : /*
9355 : * If this is the idle task on the outgoing CPU try to wake
9356 : * up the hotplug control thread which might wait for the
9357 : * last task to vanish. The rcuwait_active() check is
9358 : * accurate here because the waiter is pinned on this CPU
9359 : * and can't obviously be running in parallel.
9360 : *
9361 : * On RT kernels this also has to check whether there are
9362 : * pinned and scheduled out tasks on the runqueue. They
9363 : * need to leave the migrate disabled section first.
9364 : */
9365 : if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9366 : rcuwait_active(&rq->hotplug_wait)) {
9367 : raw_spin_rq_unlock(rq);
9368 : rcuwait_wake_up(&rq->hotplug_wait);
9369 : raw_spin_rq_lock(rq);
9370 : }
9371 : return;
9372 : }
9373 :
9374 : get_task_struct(push_task);
9375 : /*
9376 : * Temporarily drop rq->lock such that we can wake-up the stop task.
9377 : * Both preemption and IRQs are still disabled.
9378 : */
9379 : raw_spin_rq_unlock(rq);
9380 : stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9381 : this_cpu_ptr(&push_work));
9382 : /*
9383 : * At this point need_resched() is true and we'll take the loop in
9384 : * schedule(). The next pick is obviously going to be the stop task
9385 : * which kthread_is_per_cpu() and will push this task away.
9386 : */
9387 : raw_spin_rq_lock(rq);
9388 : }
9389 :
9390 : static void balance_push_set(int cpu, bool on)
9391 : {
9392 : struct rq *rq = cpu_rq(cpu);
9393 : struct rq_flags rf;
9394 :
9395 : rq_lock_irqsave(rq, &rf);
9396 : if (on) {
9397 : WARN_ON_ONCE(rq->balance_callback);
9398 : rq->balance_callback = &balance_push_callback;
9399 : } else if (rq->balance_callback == &balance_push_callback) {
9400 : rq->balance_callback = NULL;
9401 : }
9402 : rq_unlock_irqrestore(rq, &rf);
9403 : }
9404 :
9405 : /*
9406 : * Invoked from a CPUs hotplug control thread after the CPU has been marked
9407 : * inactive. All tasks which are not per CPU kernel threads are either
9408 : * pushed off this CPU now via balance_push() or placed on a different CPU
9409 : * during wakeup. Wait until the CPU is quiescent.
9410 : */
9411 : static void balance_hotplug_wait(void)
9412 : {
9413 : struct rq *rq = this_rq();
9414 :
9415 : rcuwait_wait_event(&rq->hotplug_wait,
9416 : rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9417 : TASK_UNINTERRUPTIBLE);
9418 : }
9419 :
9420 : #else
9421 :
9422 : static inline void balance_push(struct rq *rq)
9423 : {
9424 : }
9425 :
9426 : static inline void balance_push_set(int cpu, bool on)
9427 : {
9428 : }
9429 :
9430 : static inline void balance_hotplug_wait(void)
9431 : {
9432 : }
9433 :
9434 : #endif /* CONFIG_HOTPLUG_CPU */
9435 :
9436 : void set_rq_online(struct rq *rq)
9437 : {
9438 : if (!rq->online) {
9439 : const struct sched_class *class;
9440 :
9441 : cpumask_set_cpu(rq->cpu, rq->rd->online);
9442 : rq->online = 1;
9443 :
9444 : for_each_class(class) {
9445 : if (class->rq_online)
9446 : class->rq_online(rq);
9447 : }
9448 : }
9449 : }
9450 :
9451 : void set_rq_offline(struct rq *rq)
9452 : {
9453 : if (rq->online) {
9454 : const struct sched_class *class;
9455 :
9456 : for_each_class(class) {
9457 : if (class->rq_offline)
9458 : class->rq_offline(rq);
9459 : }
9460 :
9461 : cpumask_clear_cpu(rq->cpu, rq->rd->online);
9462 : rq->online = 0;
9463 : }
9464 : }
9465 :
9466 : /*
9467 : * used to mark begin/end of suspend/resume:
9468 : */
9469 : static int num_cpus_frozen;
9470 :
9471 : /*
9472 : * Update cpusets according to cpu_active mask. If cpusets are
9473 : * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9474 : * around partition_sched_domains().
9475 : *
9476 : * If we come here as part of a suspend/resume, don't touch cpusets because we
9477 : * want to restore it back to its original state upon resume anyway.
9478 : */
9479 : static void cpuset_cpu_active(void)
9480 : {
9481 : if (cpuhp_tasks_frozen) {
9482 : /*
9483 : * num_cpus_frozen tracks how many CPUs are involved in suspend
9484 : * resume sequence. As long as this is not the last online
9485 : * operation in the resume sequence, just build a single sched
9486 : * domain, ignoring cpusets.
9487 : */
9488 : partition_sched_domains(1, NULL, NULL);
9489 : if (--num_cpus_frozen)
9490 : return;
9491 : /*
9492 : * This is the last CPU online operation. So fall through and
9493 : * restore the original sched domains by considering the
9494 : * cpuset configurations.
9495 : */
9496 : cpuset_force_rebuild();
9497 : }
9498 : cpuset_update_active_cpus();
9499 : }
9500 :
9501 : static int cpuset_cpu_inactive(unsigned int cpu)
9502 : {
9503 : if (!cpuhp_tasks_frozen) {
9504 : int ret = dl_cpu_busy(cpu, NULL);
9505 :
9506 : if (ret)
9507 : return ret;
9508 : cpuset_update_active_cpus();
9509 : } else {
9510 : num_cpus_frozen++;
9511 : partition_sched_domains(1, NULL, NULL);
9512 : }
9513 : return 0;
9514 : }
9515 :
9516 : int sched_cpu_activate(unsigned int cpu)
9517 : {
9518 : struct rq *rq = cpu_rq(cpu);
9519 : struct rq_flags rf;
9520 :
9521 : /*
9522 : * Clear the balance_push callback and prepare to schedule
9523 : * regular tasks.
9524 : */
9525 : balance_push_set(cpu, false);
9526 :
9527 : #ifdef CONFIG_SCHED_SMT
9528 : /*
9529 : * When going up, increment the number of cores with SMT present.
9530 : */
9531 : if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9532 : static_branch_inc_cpuslocked(&sched_smt_present);
9533 : #endif
9534 : set_cpu_active(cpu, true);
9535 :
9536 : if (sched_smp_initialized) {
9537 : sched_update_numa(cpu, true);
9538 : sched_domains_numa_masks_set(cpu);
9539 : cpuset_cpu_active();
9540 : }
9541 :
9542 : /*
9543 : * Put the rq online, if not already. This happens:
9544 : *
9545 : * 1) In the early boot process, because we build the real domains
9546 : * after all CPUs have been brought up.
9547 : *
9548 : * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9549 : * domains.
9550 : */
9551 : rq_lock_irqsave(rq, &rf);
9552 : if (rq->rd) {
9553 : BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9554 : set_rq_online(rq);
9555 : }
9556 : rq_unlock_irqrestore(rq, &rf);
9557 :
9558 : return 0;
9559 : }
9560 :
9561 : int sched_cpu_deactivate(unsigned int cpu)
9562 : {
9563 : struct rq *rq = cpu_rq(cpu);
9564 : struct rq_flags rf;
9565 : int ret;
9566 :
9567 : /*
9568 : * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9569 : * load balancing when not active
9570 : */
9571 : nohz_balance_exit_idle(rq);
9572 :
9573 : set_cpu_active(cpu, false);
9574 :
9575 : /*
9576 : * From this point forward, this CPU will refuse to run any task that
9577 : * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9578 : * push those tasks away until this gets cleared, see
9579 : * sched_cpu_dying().
9580 : */
9581 : balance_push_set(cpu, true);
9582 :
9583 : /*
9584 : * We've cleared cpu_active_mask / set balance_push, wait for all
9585 : * preempt-disabled and RCU users of this state to go away such that
9586 : * all new such users will observe it.
9587 : *
9588 : * Specifically, we rely on ttwu to no longer target this CPU, see
9589 : * ttwu_queue_cond() and is_cpu_allowed().
9590 : *
9591 : * Do sync before park smpboot threads to take care the rcu boost case.
9592 : */
9593 : synchronize_rcu();
9594 :
9595 : rq_lock_irqsave(rq, &rf);
9596 : if (rq->rd) {
9597 : update_rq_clock(rq);
9598 : BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9599 : set_rq_offline(rq);
9600 : }
9601 : rq_unlock_irqrestore(rq, &rf);
9602 :
9603 : #ifdef CONFIG_SCHED_SMT
9604 : /*
9605 : * When going down, decrement the number of cores with SMT present.
9606 : */
9607 : if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9608 : static_branch_dec_cpuslocked(&sched_smt_present);
9609 :
9610 : sched_core_cpu_deactivate(cpu);
9611 : #endif
9612 :
9613 : if (!sched_smp_initialized)
9614 : return 0;
9615 :
9616 : sched_update_numa(cpu, false);
9617 : ret = cpuset_cpu_inactive(cpu);
9618 : if (ret) {
9619 : balance_push_set(cpu, false);
9620 : set_cpu_active(cpu, true);
9621 : sched_update_numa(cpu, true);
9622 : return ret;
9623 : }
9624 : sched_domains_numa_masks_clear(cpu);
9625 : return 0;
9626 : }
9627 :
9628 : static void sched_rq_cpu_starting(unsigned int cpu)
9629 : {
9630 : struct rq *rq = cpu_rq(cpu);
9631 :
9632 : rq->calc_load_update = calc_load_update;
9633 : update_max_interval();
9634 : }
9635 :
9636 : int sched_cpu_starting(unsigned int cpu)
9637 : {
9638 : sched_core_cpu_starting(cpu);
9639 : sched_rq_cpu_starting(cpu);
9640 : sched_tick_start(cpu);
9641 : return 0;
9642 : }
9643 :
9644 : #ifdef CONFIG_HOTPLUG_CPU
9645 :
9646 : /*
9647 : * Invoked immediately before the stopper thread is invoked to bring the
9648 : * CPU down completely. At this point all per CPU kthreads except the
9649 : * hotplug thread (current) and the stopper thread (inactive) have been
9650 : * either parked or have been unbound from the outgoing CPU. Ensure that
9651 : * any of those which might be on the way out are gone.
9652 : *
9653 : * If after this point a bound task is being woken on this CPU then the
9654 : * responsible hotplug callback has failed to do it's job.
9655 : * sched_cpu_dying() will catch it with the appropriate fireworks.
9656 : */
9657 : int sched_cpu_wait_empty(unsigned int cpu)
9658 : {
9659 : balance_hotplug_wait();
9660 : return 0;
9661 : }
9662 :
9663 : /*
9664 : * Since this CPU is going 'away' for a while, fold any nr_active delta we
9665 : * might have. Called from the CPU stopper task after ensuring that the
9666 : * stopper is the last running task on the CPU, so nr_active count is
9667 : * stable. We need to take the teardown thread which is calling this into
9668 : * account, so we hand in adjust = 1 to the load calculation.
9669 : *
9670 : * Also see the comment "Global load-average calculations".
9671 : */
9672 : static void calc_load_migrate(struct rq *rq)
9673 : {
9674 : long delta = calc_load_fold_active(rq, 1);
9675 :
9676 : if (delta)
9677 : atomic_long_add(delta, &calc_load_tasks);
9678 : }
9679 :
9680 : static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9681 : {
9682 : struct task_struct *g, *p;
9683 : int cpu = cpu_of(rq);
9684 :
9685 : lockdep_assert_rq_held(rq);
9686 :
9687 : printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9688 : for_each_process_thread(g, p) {
9689 : if (task_cpu(p) != cpu)
9690 : continue;
9691 :
9692 : if (!task_on_rq_queued(p))
9693 : continue;
9694 :
9695 : printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9696 : }
9697 : }
9698 :
9699 : int sched_cpu_dying(unsigned int cpu)
9700 : {
9701 : struct rq *rq = cpu_rq(cpu);
9702 : struct rq_flags rf;
9703 :
9704 : /* Handle pending wakeups and then migrate everything off */
9705 : sched_tick_stop(cpu);
9706 :
9707 : rq_lock_irqsave(rq, &rf);
9708 : if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9709 : WARN(true, "Dying CPU not properly vacated!");
9710 : dump_rq_tasks(rq, KERN_WARNING);
9711 : }
9712 : rq_unlock_irqrestore(rq, &rf);
9713 :
9714 : calc_load_migrate(rq);
9715 : update_max_interval();
9716 : hrtick_clear(rq);
9717 : sched_core_cpu_dying(cpu);
9718 : return 0;
9719 : }
9720 : #endif
9721 :
9722 : void __init sched_init_smp(void)
9723 : {
9724 : sched_init_numa(NUMA_NO_NODE);
9725 :
9726 : /*
9727 : * There's no userspace yet to cause hotplug operations; hence all the
9728 : * CPU masks are stable and all blatant races in the below code cannot
9729 : * happen.
9730 : */
9731 : mutex_lock(&sched_domains_mutex);
9732 : sched_init_domains(cpu_active_mask);
9733 : mutex_unlock(&sched_domains_mutex);
9734 :
9735 : /* Move init over to a non-isolated CPU */
9736 : if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9737 : BUG();
9738 : current->flags &= ~PF_NO_SETAFFINITY;
9739 : sched_init_granularity();
9740 :
9741 : init_sched_rt_class();
9742 : init_sched_dl_class();
9743 :
9744 : sched_smp_initialized = true;
9745 : }
9746 :
9747 : static int __init migration_init(void)
9748 : {
9749 : sched_cpu_starting(smp_processor_id());
9750 : return 0;
9751 : }
9752 : early_initcall(migration_init);
9753 :
9754 : #else
9755 1 : void __init sched_init_smp(void)
9756 : {
9757 1 : sched_init_granularity();
9758 1 : }
9759 : #endif /* CONFIG_SMP */
9760 :
9761 0 : int in_sched_functions(unsigned long addr)
9762 : {
9763 0 : return in_lock_functions(addr) ||
9764 0 : (addr >= (unsigned long)__sched_text_start
9765 0 : && addr < (unsigned long)__sched_text_end);
9766 : }
9767 :
9768 : #ifdef CONFIG_CGROUP_SCHED
9769 : /*
9770 : * Default task group.
9771 : * Every task in system belongs to this group at bootup.
9772 : */
9773 : struct task_group root_task_group;
9774 : LIST_HEAD(task_groups);
9775 :
9776 : /* Cacheline aligned slab cache for task_group */
9777 : static struct kmem_cache *task_group_cache __read_mostly;
9778 : #endif
9779 :
9780 1 : void __init sched_init(void)
9781 : {
9782 1 : unsigned long ptr = 0;
9783 : int i;
9784 :
9785 : /* Make sure the linker didn't screw up */
9786 1 : BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9787 : &fair_sched_class != &rt_sched_class + 1 ||
9788 : &rt_sched_class != &dl_sched_class + 1);
9789 : #ifdef CONFIG_SMP
9790 : BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9791 : #endif
9792 :
9793 1 : wait_bit_init();
9794 :
9795 : #ifdef CONFIG_FAIR_GROUP_SCHED
9796 : ptr += 2 * nr_cpu_ids * sizeof(void **);
9797 : #endif
9798 : #ifdef CONFIG_RT_GROUP_SCHED
9799 : ptr += 2 * nr_cpu_ids * sizeof(void **);
9800 : #endif
9801 : if (ptr) {
9802 : ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9803 :
9804 : #ifdef CONFIG_FAIR_GROUP_SCHED
9805 : root_task_group.se = (struct sched_entity **)ptr;
9806 : ptr += nr_cpu_ids * sizeof(void **);
9807 :
9808 : root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9809 : ptr += nr_cpu_ids * sizeof(void **);
9810 :
9811 : root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9812 : init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9813 : #endif /* CONFIG_FAIR_GROUP_SCHED */
9814 : #ifdef CONFIG_RT_GROUP_SCHED
9815 : root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9816 : ptr += nr_cpu_ids * sizeof(void **);
9817 :
9818 : root_task_group.rt_rq = (struct rt_rq **)ptr;
9819 : ptr += nr_cpu_ids * sizeof(void **);
9820 :
9821 : #endif /* CONFIG_RT_GROUP_SCHED */
9822 : }
9823 :
9824 1 : init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9825 :
9826 : #ifdef CONFIG_SMP
9827 : init_defrootdomain();
9828 : #endif
9829 :
9830 : #ifdef CONFIG_RT_GROUP_SCHED
9831 : init_rt_bandwidth(&root_task_group.rt_bandwidth,
9832 : global_rt_period(), global_rt_runtime());
9833 : #endif /* CONFIG_RT_GROUP_SCHED */
9834 :
9835 : #ifdef CONFIG_CGROUP_SCHED
9836 : task_group_cache = KMEM_CACHE(task_group, 0);
9837 :
9838 : list_add(&root_task_group.list, &task_groups);
9839 : INIT_LIST_HEAD(&root_task_group.children);
9840 : INIT_LIST_HEAD(&root_task_group.siblings);
9841 : autogroup_init(&init_task);
9842 : #endif /* CONFIG_CGROUP_SCHED */
9843 :
9844 2 : for_each_possible_cpu(i) {
9845 : struct rq *rq;
9846 :
9847 1 : rq = cpu_rq(i);
9848 : raw_spin_lock_init(&rq->__lock);
9849 1 : rq->nr_running = 0;
9850 1 : rq->calc_load_active = 0;
9851 1 : rq->calc_load_update = jiffies + LOAD_FREQ;
9852 1 : init_cfs_rq(&rq->cfs);
9853 1 : init_rt_rq(&rq->rt);
9854 1 : init_dl_rq(&rq->dl);
9855 : #ifdef CONFIG_FAIR_GROUP_SCHED
9856 : INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9857 : rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9858 : /*
9859 : * How much CPU bandwidth does root_task_group get?
9860 : *
9861 : * In case of task-groups formed thr' the cgroup filesystem, it
9862 : * gets 100% of the CPU resources in the system. This overall
9863 : * system CPU resource is divided among the tasks of
9864 : * root_task_group and its child task-groups in a fair manner,
9865 : * based on each entity's (task or task-group's) weight
9866 : * (se->load.weight).
9867 : *
9868 : * In other words, if root_task_group has 10 tasks of weight
9869 : * 1024) and two child groups A0 and A1 (of weight 1024 each),
9870 : * then A0's share of the CPU resource is:
9871 : *
9872 : * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9873 : *
9874 : * We achieve this by letting root_task_group's tasks sit
9875 : * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9876 : */
9877 : init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9878 : #endif /* CONFIG_FAIR_GROUP_SCHED */
9879 :
9880 1 : rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9881 : #ifdef CONFIG_RT_GROUP_SCHED
9882 : init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9883 : #endif
9884 : #ifdef CONFIG_SMP
9885 : rq->sd = NULL;
9886 : rq->rd = NULL;
9887 : rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9888 : rq->balance_callback = &balance_push_callback;
9889 : rq->active_balance = 0;
9890 : rq->next_balance = jiffies;
9891 : rq->push_cpu = 0;
9892 : rq->cpu = i;
9893 : rq->online = 0;
9894 : rq->idle_stamp = 0;
9895 : rq->avg_idle = 2*sysctl_sched_migration_cost;
9896 : rq->wake_stamp = jiffies;
9897 : rq->wake_avg_idle = rq->avg_idle;
9898 : rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9899 :
9900 : INIT_LIST_HEAD(&rq->cfs_tasks);
9901 :
9902 : rq_attach_root(rq, &def_root_domain);
9903 : #ifdef CONFIG_NO_HZ_COMMON
9904 : rq->last_blocked_load_update_tick = jiffies;
9905 : atomic_set(&rq->nohz_flags, 0);
9906 :
9907 : INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9908 : #endif
9909 : #ifdef CONFIG_HOTPLUG_CPU
9910 : rcuwait_init(&rq->hotplug_wait);
9911 : #endif
9912 : #endif /* CONFIG_SMP */
9913 1 : hrtick_rq_init(rq);
9914 2 : atomic_set(&rq->nr_iowait, 0);
9915 :
9916 : #ifdef CONFIG_SCHED_CORE
9917 : rq->core = rq;
9918 : rq->core_pick = NULL;
9919 : rq->core_enabled = 0;
9920 : rq->core_tree = RB_ROOT;
9921 : rq->core_forceidle_count = 0;
9922 : rq->core_forceidle_occupation = 0;
9923 : rq->core_forceidle_start = 0;
9924 :
9925 : rq->core_cookie = 0UL;
9926 : #endif
9927 2 : zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
9928 : }
9929 :
9930 1 : set_load_weight(&init_task, false);
9931 :
9932 : /*
9933 : * The boot idle thread does lazy MMU switching as well:
9934 : */
9935 1 : mmgrab(&init_mm);
9936 1 : enter_lazy_tlb(&init_mm, current);
9937 :
9938 : /*
9939 : * The idle task doesn't need the kthread struct to function, but it
9940 : * is dressed up as a per-CPU kthread and thus needs to play the part
9941 : * if we want to avoid special-casing it in code that deals with per-CPU
9942 : * kthreads.
9943 : */
9944 1 : WARN_ON(!set_kthread_struct(current));
9945 :
9946 : /*
9947 : * Make us the idle thread. Technically, schedule() should not be
9948 : * called from this thread, however somewhere below it might be,
9949 : * but because we are the idle thread, we just pick up running again
9950 : * when this runqueue becomes "idle".
9951 : */
9952 1 : init_idle(current, smp_processor_id());
9953 :
9954 1 : calc_load_update = jiffies + LOAD_FREQ;
9955 :
9956 : #ifdef CONFIG_SMP
9957 : idle_thread_set_boot_cpu();
9958 : balance_push_set(smp_processor_id(), false);
9959 : #endif
9960 1 : init_sched_fair_class();
9961 :
9962 : psi_init();
9963 :
9964 : init_uclamp();
9965 :
9966 : preempt_dynamic_init();
9967 :
9968 1 : scheduler_running = 1;
9969 1 : }
9970 :
9971 : #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9972 :
9973 : void __might_sleep(const char *file, int line)
9974 : {
9975 : unsigned int state = get_current_state();
9976 : /*
9977 : * Blocking primitives will set (and therefore destroy) current->state,
9978 : * since we will exit with TASK_RUNNING make sure we enter with it,
9979 : * otherwise we will destroy state.
9980 : */
9981 : WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9982 : "do not call blocking ops when !TASK_RUNNING; "
9983 : "state=%x set at [<%p>] %pS\n", state,
9984 : (void *)current->task_state_change,
9985 : (void *)current->task_state_change);
9986 :
9987 : __might_resched(file, line, 0);
9988 : }
9989 : EXPORT_SYMBOL(__might_sleep);
9990 :
9991 : static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9992 : {
9993 : if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9994 : return;
9995 :
9996 : if (preempt_count() == preempt_offset)
9997 : return;
9998 :
9999 : pr_err("Preemption disabled at:");
10000 : print_ip_sym(KERN_ERR, ip);
10001 : }
10002 :
10003 : static inline bool resched_offsets_ok(unsigned int offsets)
10004 : {
10005 : unsigned int nested = preempt_count();
10006 :
10007 : nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
10008 :
10009 : return nested == offsets;
10010 : }
10011 :
10012 : void __might_resched(const char *file, int line, unsigned int offsets)
10013 : {
10014 : /* Ratelimiting timestamp: */
10015 : static unsigned long prev_jiffy;
10016 :
10017 : unsigned long preempt_disable_ip;
10018 :
10019 : /* WARN_ON_ONCE() by default, no rate limit required: */
10020 : rcu_sleep_check();
10021 :
10022 : if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
10023 : !is_idle_task(current) && !current->non_block_count) ||
10024 : system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10025 : oops_in_progress)
10026 : return;
10027 :
10028 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10029 : return;
10030 : prev_jiffy = jiffies;
10031 :
10032 : /* Save this before calling printk(), since that will clobber it: */
10033 : preempt_disable_ip = get_preempt_disable_ip(current);
10034 :
10035 : pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10036 : file, line);
10037 : pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10038 : in_atomic(), irqs_disabled(), current->non_block_count,
10039 : current->pid, current->comm);
10040 : pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10041 : offsets & MIGHT_RESCHED_PREEMPT_MASK);
10042 :
10043 : if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
10044 : pr_err("RCU nest depth: %d, expected: %u\n",
10045 : rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
10046 : }
10047 :
10048 : if (task_stack_end_corrupted(current))
10049 : pr_emerg("Thread overran stack, or stack corrupted\n");
10050 :
10051 : debug_show_held_locks(current);
10052 : if (irqs_disabled())
10053 : print_irqtrace_events(current);
10054 :
10055 : print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
10056 : preempt_disable_ip);
10057 :
10058 : dump_stack();
10059 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10060 : }
10061 : EXPORT_SYMBOL(__might_resched);
10062 :
10063 : void __cant_sleep(const char *file, int line, int preempt_offset)
10064 : {
10065 : static unsigned long prev_jiffy;
10066 :
10067 : if (irqs_disabled())
10068 : return;
10069 :
10070 : if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10071 : return;
10072 :
10073 : if (preempt_count() > preempt_offset)
10074 : return;
10075 :
10076 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10077 : return;
10078 : prev_jiffy = jiffies;
10079 :
10080 : printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10081 : printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10082 : in_atomic(), irqs_disabled(),
10083 : current->pid, current->comm);
10084 :
10085 : debug_show_held_locks(current);
10086 : dump_stack();
10087 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10088 : }
10089 : EXPORT_SYMBOL_GPL(__cant_sleep);
10090 :
10091 : #ifdef CONFIG_SMP
10092 : void __cant_migrate(const char *file, int line)
10093 : {
10094 : static unsigned long prev_jiffy;
10095 :
10096 : if (irqs_disabled())
10097 : return;
10098 :
10099 : if (is_migration_disabled(current))
10100 : return;
10101 :
10102 : if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10103 : return;
10104 :
10105 : if (preempt_count() > 0)
10106 : return;
10107 :
10108 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10109 : return;
10110 : prev_jiffy = jiffies;
10111 :
10112 : pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10113 : pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10114 : in_atomic(), irqs_disabled(), is_migration_disabled(current),
10115 : current->pid, current->comm);
10116 :
10117 : debug_show_held_locks(current);
10118 : dump_stack();
10119 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10120 : }
10121 : EXPORT_SYMBOL_GPL(__cant_migrate);
10122 : #endif
10123 : #endif
10124 :
10125 : #ifdef CONFIG_MAGIC_SYSRQ
10126 : void normalize_rt_tasks(void)
10127 : {
10128 : struct task_struct *g, *p;
10129 : struct sched_attr attr = {
10130 : .sched_policy = SCHED_NORMAL,
10131 : };
10132 :
10133 : read_lock(&tasklist_lock);
10134 : for_each_process_thread(g, p) {
10135 : /*
10136 : * Only normalize user tasks:
10137 : */
10138 : if (p->flags & PF_KTHREAD)
10139 : continue;
10140 :
10141 : p->se.exec_start = 0;
10142 : schedstat_set(p->stats.wait_start, 0);
10143 : schedstat_set(p->stats.sleep_start, 0);
10144 : schedstat_set(p->stats.block_start, 0);
10145 :
10146 : if (!dl_task(p) && !rt_task(p)) {
10147 : /*
10148 : * Renice negative nice level userspace
10149 : * tasks back to 0:
10150 : */
10151 : if (task_nice(p) < 0)
10152 : set_user_nice(p, 0);
10153 : continue;
10154 : }
10155 :
10156 : __sched_setscheduler(p, &attr, false, false);
10157 : }
10158 : read_unlock(&tasklist_lock);
10159 : }
10160 :
10161 : #endif /* CONFIG_MAGIC_SYSRQ */
10162 :
10163 : #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
10164 : /*
10165 : * These functions are only useful for the IA64 MCA handling, or kdb.
10166 : *
10167 : * They can only be called when the whole system has been
10168 : * stopped - every CPU needs to be quiescent, and no scheduling
10169 : * activity can take place. Using them for anything else would
10170 : * be a serious bug, and as a result, they aren't even visible
10171 : * under any other configuration.
10172 : */
10173 :
10174 : /**
10175 : * curr_task - return the current task for a given CPU.
10176 : * @cpu: the processor in question.
10177 : *
10178 : * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10179 : *
10180 : * Return: The current task for @cpu.
10181 : */
10182 : struct task_struct *curr_task(int cpu)
10183 : {
10184 : return cpu_curr(cpu);
10185 : }
10186 :
10187 : #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10188 :
10189 : #ifdef CONFIG_IA64
10190 : /**
10191 : * ia64_set_curr_task - set the current task for a given CPU.
10192 : * @cpu: the processor in question.
10193 : * @p: the task pointer to set.
10194 : *
10195 : * Description: This function must only be used when non-maskable interrupts
10196 : * are serviced on a separate stack. It allows the architecture to switch the
10197 : * notion of the current task on a CPU in a non-blocking manner. This function
10198 : * must be called with all CPU's synchronized, and interrupts disabled, the
10199 : * and caller must save the original value of the current task (see
10200 : * curr_task() above) and restore that value before reenabling interrupts and
10201 : * re-starting the system.
10202 : *
10203 : * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10204 : */
10205 : void ia64_set_curr_task(int cpu, struct task_struct *p)
10206 : {
10207 : cpu_curr(cpu) = p;
10208 : }
10209 :
10210 : #endif
10211 :
10212 : #ifdef CONFIG_CGROUP_SCHED
10213 : /* task_group_lock serializes the addition/removal of task groups */
10214 : static DEFINE_SPINLOCK(task_group_lock);
10215 :
10216 : static inline void alloc_uclamp_sched_group(struct task_group *tg,
10217 : struct task_group *parent)
10218 : {
10219 : #ifdef CONFIG_UCLAMP_TASK_GROUP
10220 : enum uclamp_id clamp_id;
10221 :
10222 : for_each_clamp_id(clamp_id) {
10223 : uclamp_se_set(&tg->uclamp_req[clamp_id],
10224 : uclamp_none(clamp_id), false);
10225 : tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10226 : }
10227 : #endif
10228 : }
10229 :
10230 : static void sched_free_group(struct task_group *tg)
10231 : {
10232 : free_fair_sched_group(tg);
10233 : free_rt_sched_group(tg);
10234 : autogroup_free(tg);
10235 : kmem_cache_free(task_group_cache, tg);
10236 : }
10237 :
10238 : static void sched_free_group_rcu(struct rcu_head *rcu)
10239 : {
10240 : sched_free_group(container_of(rcu, struct task_group, rcu));
10241 : }
10242 :
10243 : static void sched_unregister_group(struct task_group *tg)
10244 : {
10245 : unregister_fair_sched_group(tg);
10246 : unregister_rt_sched_group(tg);
10247 : /*
10248 : * We have to wait for yet another RCU grace period to expire, as
10249 : * print_cfs_stats() might run concurrently.
10250 : */
10251 : call_rcu(&tg->rcu, sched_free_group_rcu);
10252 : }
10253 :
10254 : /* allocate runqueue etc for a new task group */
10255 : struct task_group *sched_create_group(struct task_group *parent)
10256 : {
10257 : struct task_group *tg;
10258 :
10259 : tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10260 : if (!tg)
10261 : return ERR_PTR(-ENOMEM);
10262 :
10263 : if (!alloc_fair_sched_group(tg, parent))
10264 : goto err;
10265 :
10266 : if (!alloc_rt_sched_group(tg, parent))
10267 : goto err;
10268 :
10269 : alloc_uclamp_sched_group(tg, parent);
10270 :
10271 : return tg;
10272 :
10273 : err:
10274 : sched_free_group(tg);
10275 : return ERR_PTR(-ENOMEM);
10276 : }
10277 :
10278 : void sched_online_group(struct task_group *tg, struct task_group *parent)
10279 : {
10280 : unsigned long flags;
10281 :
10282 : spin_lock_irqsave(&task_group_lock, flags);
10283 : list_add_rcu(&tg->list, &task_groups);
10284 :
10285 : /* Root should already exist: */
10286 : WARN_ON(!parent);
10287 :
10288 : tg->parent = parent;
10289 : INIT_LIST_HEAD(&tg->children);
10290 : list_add_rcu(&tg->siblings, &parent->children);
10291 : spin_unlock_irqrestore(&task_group_lock, flags);
10292 :
10293 : online_fair_sched_group(tg);
10294 : }
10295 :
10296 : /* rcu callback to free various structures associated with a task group */
10297 : static void sched_unregister_group_rcu(struct rcu_head *rhp)
10298 : {
10299 : /* Now it should be safe to free those cfs_rqs: */
10300 : sched_unregister_group(container_of(rhp, struct task_group, rcu));
10301 : }
10302 :
10303 : void sched_destroy_group(struct task_group *tg)
10304 : {
10305 : /* Wait for possible concurrent references to cfs_rqs complete: */
10306 : call_rcu(&tg->rcu, sched_unregister_group_rcu);
10307 : }
10308 :
10309 : void sched_release_group(struct task_group *tg)
10310 : {
10311 : unsigned long flags;
10312 :
10313 : /*
10314 : * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10315 : * sched_cfs_period_timer()).
10316 : *
10317 : * For this to be effective, we have to wait for all pending users of
10318 : * this task group to leave their RCU critical section to ensure no new
10319 : * user will see our dying task group any more. Specifically ensure
10320 : * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10321 : *
10322 : * We therefore defer calling unregister_fair_sched_group() to
10323 : * sched_unregister_group() which is guarantied to get called only after the
10324 : * current RCU grace period has expired.
10325 : */
10326 : spin_lock_irqsave(&task_group_lock, flags);
10327 : list_del_rcu(&tg->list);
10328 : list_del_rcu(&tg->siblings);
10329 : spin_unlock_irqrestore(&task_group_lock, flags);
10330 : }
10331 :
10332 : static void sched_change_group(struct task_struct *tsk)
10333 : {
10334 : struct task_group *tg;
10335 :
10336 : /*
10337 : * All callers are synchronized by task_rq_lock(); we do not use RCU
10338 : * which is pointless here. Thus, we pass "true" to task_css_check()
10339 : * to prevent lockdep warnings.
10340 : */
10341 : tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10342 : struct task_group, css);
10343 : tg = autogroup_task_group(tsk, tg);
10344 : tsk->sched_task_group = tg;
10345 :
10346 : #ifdef CONFIG_FAIR_GROUP_SCHED
10347 : if (tsk->sched_class->task_change_group)
10348 : tsk->sched_class->task_change_group(tsk);
10349 : else
10350 : #endif
10351 : set_task_rq(tsk, task_cpu(tsk));
10352 : }
10353 :
10354 : /*
10355 : * Change task's runqueue when it moves between groups.
10356 : *
10357 : * The caller of this function should have put the task in its new group by
10358 : * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10359 : * its new group.
10360 : */
10361 : void sched_move_task(struct task_struct *tsk)
10362 : {
10363 : int queued, running, queue_flags =
10364 : DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10365 : struct rq_flags rf;
10366 : struct rq *rq;
10367 :
10368 : rq = task_rq_lock(tsk, &rf);
10369 : update_rq_clock(rq);
10370 :
10371 : running = task_current(rq, tsk);
10372 : queued = task_on_rq_queued(tsk);
10373 :
10374 : if (queued)
10375 : dequeue_task(rq, tsk, queue_flags);
10376 : if (running)
10377 : put_prev_task(rq, tsk);
10378 :
10379 : sched_change_group(tsk);
10380 :
10381 : if (queued)
10382 : enqueue_task(rq, tsk, queue_flags);
10383 : if (running) {
10384 : set_next_task(rq, tsk);
10385 : /*
10386 : * After changing group, the running task may have joined a
10387 : * throttled one but it's still the running task. Trigger a
10388 : * resched to make sure that task can still run.
10389 : */
10390 : resched_curr(rq);
10391 : }
10392 :
10393 : task_rq_unlock(rq, tsk, &rf);
10394 : }
10395 :
10396 : static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10397 : {
10398 : return css ? container_of(css, struct task_group, css) : NULL;
10399 : }
10400 :
10401 : static struct cgroup_subsys_state *
10402 : cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10403 : {
10404 : struct task_group *parent = css_tg(parent_css);
10405 : struct task_group *tg;
10406 :
10407 : if (!parent) {
10408 : /* This is early initialization for the top cgroup */
10409 : return &root_task_group.css;
10410 : }
10411 :
10412 : tg = sched_create_group(parent);
10413 : if (IS_ERR(tg))
10414 : return ERR_PTR(-ENOMEM);
10415 :
10416 : return &tg->css;
10417 : }
10418 :
10419 : /* Expose task group only after completing cgroup initialization */
10420 : static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10421 : {
10422 : struct task_group *tg = css_tg(css);
10423 : struct task_group *parent = css_tg(css->parent);
10424 :
10425 : if (parent)
10426 : sched_online_group(tg, parent);
10427 :
10428 : #ifdef CONFIG_UCLAMP_TASK_GROUP
10429 : /* Propagate the effective uclamp value for the new group */
10430 : mutex_lock(&uclamp_mutex);
10431 : rcu_read_lock();
10432 : cpu_util_update_eff(css);
10433 : rcu_read_unlock();
10434 : mutex_unlock(&uclamp_mutex);
10435 : #endif
10436 :
10437 : return 0;
10438 : }
10439 :
10440 : static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10441 : {
10442 : struct task_group *tg = css_tg(css);
10443 :
10444 : sched_release_group(tg);
10445 : }
10446 :
10447 : static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10448 : {
10449 : struct task_group *tg = css_tg(css);
10450 :
10451 : /*
10452 : * Relies on the RCU grace period between css_released() and this.
10453 : */
10454 : sched_unregister_group(tg);
10455 : }
10456 :
10457 : #ifdef CONFIG_RT_GROUP_SCHED
10458 : static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10459 : {
10460 : struct task_struct *task;
10461 : struct cgroup_subsys_state *css;
10462 :
10463 : cgroup_taskset_for_each(task, css, tset) {
10464 : if (!sched_rt_can_attach(css_tg(css), task))
10465 : return -EINVAL;
10466 : }
10467 : return 0;
10468 : }
10469 : #endif
10470 :
10471 : static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10472 : {
10473 : struct task_struct *task;
10474 : struct cgroup_subsys_state *css;
10475 :
10476 : cgroup_taskset_for_each(task, css, tset)
10477 : sched_move_task(task);
10478 : }
10479 :
10480 : #ifdef CONFIG_UCLAMP_TASK_GROUP
10481 : static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10482 : {
10483 : struct cgroup_subsys_state *top_css = css;
10484 : struct uclamp_se *uc_parent = NULL;
10485 : struct uclamp_se *uc_se = NULL;
10486 : unsigned int eff[UCLAMP_CNT];
10487 : enum uclamp_id clamp_id;
10488 : unsigned int clamps;
10489 :
10490 : lockdep_assert_held(&uclamp_mutex);
10491 : SCHED_WARN_ON(!rcu_read_lock_held());
10492 :
10493 : css_for_each_descendant_pre(css, top_css) {
10494 : uc_parent = css_tg(css)->parent
10495 : ? css_tg(css)->parent->uclamp : NULL;
10496 :
10497 : for_each_clamp_id(clamp_id) {
10498 : /* Assume effective clamps matches requested clamps */
10499 : eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10500 : /* Cap effective clamps with parent's effective clamps */
10501 : if (uc_parent &&
10502 : eff[clamp_id] > uc_parent[clamp_id].value) {
10503 : eff[clamp_id] = uc_parent[clamp_id].value;
10504 : }
10505 : }
10506 : /* Ensure protection is always capped by limit */
10507 : eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10508 :
10509 : /* Propagate most restrictive effective clamps */
10510 : clamps = 0x0;
10511 : uc_se = css_tg(css)->uclamp;
10512 : for_each_clamp_id(clamp_id) {
10513 : if (eff[clamp_id] == uc_se[clamp_id].value)
10514 : continue;
10515 : uc_se[clamp_id].value = eff[clamp_id];
10516 : uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10517 : clamps |= (0x1 << clamp_id);
10518 : }
10519 : if (!clamps) {
10520 : css = css_rightmost_descendant(css);
10521 : continue;
10522 : }
10523 :
10524 : /* Immediately update descendants RUNNABLE tasks */
10525 : uclamp_update_active_tasks(css);
10526 : }
10527 : }
10528 :
10529 : /*
10530 : * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10531 : * C expression. Since there is no way to convert a macro argument (N) into a
10532 : * character constant, use two levels of macros.
10533 : */
10534 : #define _POW10(exp) ((unsigned int)1e##exp)
10535 : #define POW10(exp) _POW10(exp)
10536 :
10537 : struct uclamp_request {
10538 : #define UCLAMP_PERCENT_SHIFT 2
10539 : #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10540 : s64 percent;
10541 : u64 util;
10542 : int ret;
10543 : };
10544 :
10545 : static inline struct uclamp_request
10546 : capacity_from_percent(char *buf)
10547 : {
10548 : struct uclamp_request req = {
10549 : .percent = UCLAMP_PERCENT_SCALE,
10550 : .util = SCHED_CAPACITY_SCALE,
10551 : .ret = 0,
10552 : };
10553 :
10554 : buf = strim(buf);
10555 : if (strcmp(buf, "max")) {
10556 : req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10557 : &req.percent);
10558 : if (req.ret)
10559 : return req;
10560 : if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10561 : req.ret = -ERANGE;
10562 : return req;
10563 : }
10564 :
10565 : req.util = req.percent << SCHED_CAPACITY_SHIFT;
10566 : req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10567 : }
10568 :
10569 : return req;
10570 : }
10571 :
10572 : static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10573 : size_t nbytes, loff_t off,
10574 : enum uclamp_id clamp_id)
10575 : {
10576 : struct uclamp_request req;
10577 : struct task_group *tg;
10578 :
10579 : req = capacity_from_percent(buf);
10580 : if (req.ret)
10581 : return req.ret;
10582 :
10583 : static_branch_enable(&sched_uclamp_used);
10584 :
10585 : mutex_lock(&uclamp_mutex);
10586 : rcu_read_lock();
10587 :
10588 : tg = css_tg(of_css(of));
10589 : if (tg->uclamp_req[clamp_id].value != req.util)
10590 : uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10591 :
10592 : /*
10593 : * Because of not recoverable conversion rounding we keep track of the
10594 : * exact requested value
10595 : */
10596 : tg->uclamp_pct[clamp_id] = req.percent;
10597 :
10598 : /* Update effective clamps to track the most restrictive value */
10599 : cpu_util_update_eff(of_css(of));
10600 :
10601 : rcu_read_unlock();
10602 : mutex_unlock(&uclamp_mutex);
10603 :
10604 : return nbytes;
10605 : }
10606 :
10607 : static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10608 : char *buf, size_t nbytes,
10609 : loff_t off)
10610 : {
10611 : return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10612 : }
10613 :
10614 : static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10615 : char *buf, size_t nbytes,
10616 : loff_t off)
10617 : {
10618 : return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10619 : }
10620 :
10621 : static inline void cpu_uclamp_print(struct seq_file *sf,
10622 : enum uclamp_id clamp_id)
10623 : {
10624 : struct task_group *tg;
10625 : u64 util_clamp;
10626 : u64 percent;
10627 : u32 rem;
10628 :
10629 : rcu_read_lock();
10630 : tg = css_tg(seq_css(sf));
10631 : util_clamp = tg->uclamp_req[clamp_id].value;
10632 : rcu_read_unlock();
10633 :
10634 : if (util_clamp == SCHED_CAPACITY_SCALE) {
10635 : seq_puts(sf, "max\n");
10636 : return;
10637 : }
10638 :
10639 : percent = tg->uclamp_pct[clamp_id];
10640 : percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10641 : seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10642 : }
10643 :
10644 : static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10645 : {
10646 : cpu_uclamp_print(sf, UCLAMP_MIN);
10647 : return 0;
10648 : }
10649 :
10650 : static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10651 : {
10652 : cpu_uclamp_print(sf, UCLAMP_MAX);
10653 : return 0;
10654 : }
10655 : #endif /* CONFIG_UCLAMP_TASK_GROUP */
10656 :
10657 : #ifdef CONFIG_FAIR_GROUP_SCHED
10658 : static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10659 : struct cftype *cftype, u64 shareval)
10660 : {
10661 : if (shareval > scale_load_down(ULONG_MAX))
10662 : shareval = MAX_SHARES;
10663 : return sched_group_set_shares(css_tg(css), scale_load(shareval));
10664 : }
10665 :
10666 : static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10667 : struct cftype *cft)
10668 : {
10669 : struct task_group *tg = css_tg(css);
10670 :
10671 : return (u64) scale_load_down(tg->shares);
10672 : }
10673 :
10674 : #ifdef CONFIG_CFS_BANDWIDTH
10675 : static DEFINE_MUTEX(cfs_constraints_mutex);
10676 :
10677 : const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10678 : static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10679 : /* More than 203 days if BW_SHIFT equals 20. */
10680 : static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10681 :
10682 : static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10683 :
10684 : static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10685 : u64 burst)
10686 : {
10687 : int i, ret = 0, runtime_enabled, runtime_was_enabled;
10688 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10689 :
10690 : if (tg == &root_task_group)
10691 : return -EINVAL;
10692 :
10693 : /*
10694 : * Ensure we have at some amount of bandwidth every period. This is
10695 : * to prevent reaching a state of large arrears when throttled via
10696 : * entity_tick() resulting in prolonged exit starvation.
10697 : */
10698 : if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10699 : return -EINVAL;
10700 :
10701 : /*
10702 : * Likewise, bound things on the other side by preventing insane quota
10703 : * periods. This also allows us to normalize in computing quota
10704 : * feasibility.
10705 : */
10706 : if (period > max_cfs_quota_period)
10707 : return -EINVAL;
10708 :
10709 : /*
10710 : * Bound quota to defend quota against overflow during bandwidth shift.
10711 : */
10712 : if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10713 : return -EINVAL;
10714 :
10715 : if (quota != RUNTIME_INF && (burst > quota ||
10716 : burst + quota > max_cfs_runtime))
10717 : return -EINVAL;
10718 :
10719 : /*
10720 : * Prevent race between setting of cfs_rq->runtime_enabled and
10721 : * unthrottle_offline_cfs_rqs().
10722 : */
10723 : cpus_read_lock();
10724 : mutex_lock(&cfs_constraints_mutex);
10725 : ret = __cfs_schedulable(tg, period, quota);
10726 : if (ret)
10727 : goto out_unlock;
10728 :
10729 : runtime_enabled = quota != RUNTIME_INF;
10730 : runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10731 : /*
10732 : * If we need to toggle cfs_bandwidth_used, off->on must occur
10733 : * before making related changes, and on->off must occur afterwards
10734 : */
10735 : if (runtime_enabled && !runtime_was_enabled)
10736 : cfs_bandwidth_usage_inc();
10737 : raw_spin_lock_irq(&cfs_b->lock);
10738 : cfs_b->period = ns_to_ktime(period);
10739 : cfs_b->quota = quota;
10740 : cfs_b->burst = burst;
10741 :
10742 : __refill_cfs_bandwidth_runtime(cfs_b);
10743 :
10744 : /* Restart the period timer (if active) to handle new period expiry: */
10745 : if (runtime_enabled)
10746 : start_cfs_bandwidth(cfs_b);
10747 :
10748 : raw_spin_unlock_irq(&cfs_b->lock);
10749 :
10750 : for_each_online_cpu(i) {
10751 : struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10752 : struct rq *rq = cfs_rq->rq;
10753 : struct rq_flags rf;
10754 :
10755 : rq_lock_irq(rq, &rf);
10756 : cfs_rq->runtime_enabled = runtime_enabled;
10757 : cfs_rq->runtime_remaining = 0;
10758 :
10759 : if (cfs_rq->throttled)
10760 : unthrottle_cfs_rq(cfs_rq);
10761 : rq_unlock_irq(rq, &rf);
10762 : }
10763 : if (runtime_was_enabled && !runtime_enabled)
10764 : cfs_bandwidth_usage_dec();
10765 : out_unlock:
10766 : mutex_unlock(&cfs_constraints_mutex);
10767 : cpus_read_unlock();
10768 :
10769 : return ret;
10770 : }
10771 :
10772 : static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10773 : {
10774 : u64 quota, period, burst;
10775 :
10776 : period = ktime_to_ns(tg->cfs_bandwidth.period);
10777 : burst = tg->cfs_bandwidth.burst;
10778 : if (cfs_quota_us < 0)
10779 : quota = RUNTIME_INF;
10780 : else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10781 : quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10782 : else
10783 : return -EINVAL;
10784 :
10785 : return tg_set_cfs_bandwidth(tg, period, quota, burst);
10786 : }
10787 :
10788 : static long tg_get_cfs_quota(struct task_group *tg)
10789 : {
10790 : u64 quota_us;
10791 :
10792 : if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10793 : return -1;
10794 :
10795 : quota_us = tg->cfs_bandwidth.quota;
10796 : do_div(quota_us, NSEC_PER_USEC);
10797 :
10798 : return quota_us;
10799 : }
10800 :
10801 : static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10802 : {
10803 : u64 quota, period, burst;
10804 :
10805 : if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10806 : return -EINVAL;
10807 :
10808 : period = (u64)cfs_period_us * NSEC_PER_USEC;
10809 : quota = tg->cfs_bandwidth.quota;
10810 : burst = tg->cfs_bandwidth.burst;
10811 :
10812 : return tg_set_cfs_bandwidth(tg, period, quota, burst);
10813 : }
10814 :
10815 : static long tg_get_cfs_period(struct task_group *tg)
10816 : {
10817 : u64 cfs_period_us;
10818 :
10819 : cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10820 : do_div(cfs_period_us, NSEC_PER_USEC);
10821 :
10822 : return cfs_period_us;
10823 : }
10824 :
10825 : static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10826 : {
10827 : u64 quota, period, burst;
10828 :
10829 : if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10830 : return -EINVAL;
10831 :
10832 : burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10833 : period = ktime_to_ns(tg->cfs_bandwidth.period);
10834 : quota = tg->cfs_bandwidth.quota;
10835 :
10836 : return tg_set_cfs_bandwidth(tg, period, quota, burst);
10837 : }
10838 :
10839 : static long tg_get_cfs_burst(struct task_group *tg)
10840 : {
10841 : u64 burst_us;
10842 :
10843 : burst_us = tg->cfs_bandwidth.burst;
10844 : do_div(burst_us, NSEC_PER_USEC);
10845 :
10846 : return burst_us;
10847 : }
10848 :
10849 : static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10850 : struct cftype *cft)
10851 : {
10852 : return tg_get_cfs_quota(css_tg(css));
10853 : }
10854 :
10855 : static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10856 : struct cftype *cftype, s64 cfs_quota_us)
10857 : {
10858 : return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10859 : }
10860 :
10861 : static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10862 : struct cftype *cft)
10863 : {
10864 : return tg_get_cfs_period(css_tg(css));
10865 : }
10866 :
10867 : static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10868 : struct cftype *cftype, u64 cfs_period_us)
10869 : {
10870 : return tg_set_cfs_period(css_tg(css), cfs_period_us);
10871 : }
10872 :
10873 : static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10874 : struct cftype *cft)
10875 : {
10876 : return tg_get_cfs_burst(css_tg(css));
10877 : }
10878 :
10879 : static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10880 : struct cftype *cftype, u64 cfs_burst_us)
10881 : {
10882 : return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10883 : }
10884 :
10885 : struct cfs_schedulable_data {
10886 : struct task_group *tg;
10887 : u64 period, quota;
10888 : };
10889 :
10890 : /*
10891 : * normalize group quota/period to be quota/max_period
10892 : * note: units are usecs
10893 : */
10894 : static u64 normalize_cfs_quota(struct task_group *tg,
10895 : struct cfs_schedulable_data *d)
10896 : {
10897 : u64 quota, period;
10898 :
10899 : if (tg == d->tg) {
10900 : period = d->period;
10901 : quota = d->quota;
10902 : } else {
10903 : period = tg_get_cfs_period(tg);
10904 : quota = tg_get_cfs_quota(tg);
10905 : }
10906 :
10907 : /* note: these should typically be equivalent */
10908 : if (quota == RUNTIME_INF || quota == -1)
10909 : return RUNTIME_INF;
10910 :
10911 : return to_ratio(period, quota);
10912 : }
10913 :
10914 : static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10915 : {
10916 : struct cfs_schedulable_data *d = data;
10917 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10918 : s64 quota = 0, parent_quota = -1;
10919 :
10920 : if (!tg->parent) {
10921 : quota = RUNTIME_INF;
10922 : } else {
10923 : struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10924 :
10925 : quota = normalize_cfs_quota(tg, d);
10926 : parent_quota = parent_b->hierarchical_quota;
10927 :
10928 : /*
10929 : * Ensure max(child_quota) <= parent_quota. On cgroup2,
10930 : * always take the min. On cgroup1, only inherit when no
10931 : * limit is set:
10932 : */
10933 : if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10934 : quota = min(quota, parent_quota);
10935 : } else {
10936 : if (quota == RUNTIME_INF)
10937 : quota = parent_quota;
10938 : else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10939 : return -EINVAL;
10940 : }
10941 : }
10942 : cfs_b->hierarchical_quota = quota;
10943 :
10944 : return 0;
10945 : }
10946 :
10947 : static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10948 : {
10949 : int ret;
10950 : struct cfs_schedulable_data data = {
10951 : .tg = tg,
10952 : .period = period,
10953 : .quota = quota,
10954 : };
10955 :
10956 : if (quota != RUNTIME_INF) {
10957 : do_div(data.period, NSEC_PER_USEC);
10958 : do_div(data.quota, NSEC_PER_USEC);
10959 : }
10960 :
10961 : rcu_read_lock();
10962 : ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10963 : rcu_read_unlock();
10964 :
10965 : return ret;
10966 : }
10967 :
10968 : static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10969 : {
10970 : struct task_group *tg = css_tg(seq_css(sf));
10971 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10972 :
10973 : seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10974 : seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10975 : seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10976 :
10977 : if (schedstat_enabled() && tg != &root_task_group) {
10978 : struct sched_statistics *stats;
10979 : u64 ws = 0;
10980 : int i;
10981 :
10982 : for_each_possible_cpu(i) {
10983 : stats = __schedstats_from_se(tg->se[i]);
10984 : ws += schedstat_val(stats->wait_sum);
10985 : }
10986 :
10987 : seq_printf(sf, "wait_sum %llu\n", ws);
10988 : }
10989 :
10990 : seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10991 : seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10992 :
10993 : return 0;
10994 : }
10995 : #endif /* CONFIG_CFS_BANDWIDTH */
10996 : #endif /* CONFIG_FAIR_GROUP_SCHED */
10997 :
10998 : #ifdef CONFIG_RT_GROUP_SCHED
10999 : static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
11000 : struct cftype *cft, s64 val)
11001 : {
11002 : return sched_group_set_rt_runtime(css_tg(css), val);
11003 : }
11004 :
11005 : static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
11006 : struct cftype *cft)
11007 : {
11008 : return sched_group_rt_runtime(css_tg(css));
11009 : }
11010 :
11011 : static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
11012 : struct cftype *cftype, u64 rt_period_us)
11013 : {
11014 : return sched_group_set_rt_period(css_tg(css), rt_period_us);
11015 : }
11016 :
11017 : static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11018 : struct cftype *cft)
11019 : {
11020 : return sched_group_rt_period(css_tg(css));
11021 : }
11022 : #endif /* CONFIG_RT_GROUP_SCHED */
11023 :
11024 : #ifdef CONFIG_FAIR_GROUP_SCHED
11025 : static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11026 : struct cftype *cft)
11027 : {
11028 : return css_tg(css)->idle;
11029 : }
11030 :
11031 : static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11032 : struct cftype *cft, s64 idle)
11033 : {
11034 : return sched_group_set_idle(css_tg(css), idle);
11035 : }
11036 : #endif
11037 :
11038 : static struct cftype cpu_legacy_files[] = {
11039 : #ifdef CONFIG_FAIR_GROUP_SCHED
11040 : {
11041 : .name = "shares",
11042 : .read_u64 = cpu_shares_read_u64,
11043 : .write_u64 = cpu_shares_write_u64,
11044 : },
11045 : {
11046 : .name = "idle",
11047 : .read_s64 = cpu_idle_read_s64,
11048 : .write_s64 = cpu_idle_write_s64,
11049 : },
11050 : #endif
11051 : #ifdef CONFIG_CFS_BANDWIDTH
11052 : {
11053 : .name = "cfs_quota_us",
11054 : .read_s64 = cpu_cfs_quota_read_s64,
11055 : .write_s64 = cpu_cfs_quota_write_s64,
11056 : },
11057 : {
11058 : .name = "cfs_period_us",
11059 : .read_u64 = cpu_cfs_period_read_u64,
11060 : .write_u64 = cpu_cfs_period_write_u64,
11061 : },
11062 : {
11063 : .name = "cfs_burst_us",
11064 : .read_u64 = cpu_cfs_burst_read_u64,
11065 : .write_u64 = cpu_cfs_burst_write_u64,
11066 : },
11067 : {
11068 : .name = "stat",
11069 : .seq_show = cpu_cfs_stat_show,
11070 : },
11071 : #endif
11072 : #ifdef CONFIG_RT_GROUP_SCHED
11073 : {
11074 : .name = "rt_runtime_us",
11075 : .read_s64 = cpu_rt_runtime_read,
11076 : .write_s64 = cpu_rt_runtime_write,
11077 : },
11078 : {
11079 : .name = "rt_period_us",
11080 : .read_u64 = cpu_rt_period_read_uint,
11081 : .write_u64 = cpu_rt_period_write_uint,
11082 : },
11083 : #endif
11084 : #ifdef CONFIG_UCLAMP_TASK_GROUP
11085 : {
11086 : .name = "uclamp.min",
11087 : .flags = CFTYPE_NOT_ON_ROOT,
11088 : .seq_show = cpu_uclamp_min_show,
11089 : .write = cpu_uclamp_min_write,
11090 : },
11091 : {
11092 : .name = "uclamp.max",
11093 : .flags = CFTYPE_NOT_ON_ROOT,
11094 : .seq_show = cpu_uclamp_max_show,
11095 : .write = cpu_uclamp_max_write,
11096 : },
11097 : #endif
11098 : { } /* Terminate */
11099 : };
11100 :
11101 : static int cpu_extra_stat_show(struct seq_file *sf,
11102 : struct cgroup_subsys_state *css)
11103 : {
11104 : #ifdef CONFIG_CFS_BANDWIDTH
11105 : {
11106 : struct task_group *tg = css_tg(css);
11107 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11108 : u64 throttled_usec, burst_usec;
11109 :
11110 : throttled_usec = cfs_b->throttled_time;
11111 : do_div(throttled_usec, NSEC_PER_USEC);
11112 : burst_usec = cfs_b->burst_time;
11113 : do_div(burst_usec, NSEC_PER_USEC);
11114 :
11115 : seq_printf(sf, "nr_periods %d\n"
11116 : "nr_throttled %d\n"
11117 : "throttled_usec %llu\n"
11118 : "nr_bursts %d\n"
11119 : "burst_usec %llu\n",
11120 : cfs_b->nr_periods, cfs_b->nr_throttled,
11121 : throttled_usec, cfs_b->nr_burst, burst_usec);
11122 : }
11123 : #endif
11124 : return 0;
11125 : }
11126 :
11127 : #ifdef CONFIG_FAIR_GROUP_SCHED
11128 : static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11129 : struct cftype *cft)
11130 : {
11131 : struct task_group *tg = css_tg(css);
11132 : u64 weight = scale_load_down(tg->shares);
11133 :
11134 : return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11135 : }
11136 :
11137 : static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11138 : struct cftype *cft, u64 weight)
11139 : {
11140 : /*
11141 : * cgroup weight knobs should use the common MIN, DFL and MAX
11142 : * values which are 1, 100 and 10000 respectively. While it loses
11143 : * a bit of range on both ends, it maps pretty well onto the shares
11144 : * value used by scheduler and the round-trip conversions preserve
11145 : * the original value over the entire range.
11146 : */
11147 : if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11148 : return -ERANGE;
11149 :
11150 : weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11151 :
11152 : return sched_group_set_shares(css_tg(css), scale_load(weight));
11153 : }
11154 :
11155 : static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11156 : struct cftype *cft)
11157 : {
11158 : unsigned long weight = scale_load_down(css_tg(css)->shares);
11159 : int last_delta = INT_MAX;
11160 : int prio, delta;
11161 :
11162 : /* find the closest nice value to the current weight */
11163 : for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11164 : delta = abs(sched_prio_to_weight[prio] - weight);
11165 : if (delta >= last_delta)
11166 : break;
11167 : last_delta = delta;
11168 : }
11169 :
11170 : return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11171 : }
11172 :
11173 : static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11174 : struct cftype *cft, s64 nice)
11175 : {
11176 : unsigned long weight;
11177 : int idx;
11178 :
11179 : if (nice < MIN_NICE || nice > MAX_NICE)
11180 : return -ERANGE;
11181 :
11182 : idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11183 : idx = array_index_nospec(idx, 40);
11184 : weight = sched_prio_to_weight[idx];
11185 :
11186 : return sched_group_set_shares(css_tg(css), scale_load(weight));
11187 : }
11188 : #endif
11189 :
11190 : static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11191 : long period, long quota)
11192 : {
11193 : if (quota < 0)
11194 : seq_puts(sf, "max");
11195 : else
11196 : seq_printf(sf, "%ld", quota);
11197 :
11198 : seq_printf(sf, " %ld\n", period);
11199 : }
11200 :
11201 : /* caller should put the current value in *@periodp before calling */
11202 : static int __maybe_unused cpu_period_quota_parse(char *buf,
11203 : u64 *periodp, u64 *quotap)
11204 : {
11205 : char tok[21]; /* U64_MAX */
11206 :
11207 : if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11208 : return -EINVAL;
11209 :
11210 : *periodp *= NSEC_PER_USEC;
11211 :
11212 : if (sscanf(tok, "%llu", quotap))
11213 : *quotap *= NSEC_PER_USEC;
11214 : else if (!strcmp(tok, "max"))
11215 : *quotap = RUNTIME_INF;
11216 : else
11217 : return -EINVAL;
11218 :
11219 : return 0;
11220 : }
11221 :
11222 : #ifdef CONFIG_CFS_BANDWIDTH
11223 : static int cpu_max_show(struct seq_file *sf, void *v)
11224 : {
11225 : struct task_group *tg = css_tg(seq_css(sf));
11226 :
11227 : cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11228 : return 0;
11229 : }
11230 :
11231 : static ssize_t cpu_max_write(struct kernfs_open_file *of,
11232 : char *buf, size_t nbytes, loff_t off)
11233 : {
11234 : struct task_group *tg = css_tg(of_css(of));
11235 : u64 period = tg_get_cfs_period(tg);
11236 : u64 burst = tg_get_cfs_burst(tg);
11237 : u64 quota;
11238 : int ret;
11239 :
11240 : ret = cpu_period_quota_parse(buf, &period, "a);
11241 : if (!ret)
11242 : ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11243 : return ret ?: nbytes;
11244 : }
11245 : #endif
11246 :
11247 : static struct cftype cpu_files[] = {
11248 : #ifdef CONFIG_FAIR_GROUP_SCHED
11249 : {
11250 : .name = "weight",
11251 : .flags = CFTYPE_NOT_ON_ROOT,
11252 : .read_u64 = cpu_weight_read_u64,
11253 : .write_u64 = cpu_weight_write_u64,
11254 : },
11255 : {
11256 : .name = "weight.nice",
11257 : .flags = CFTYPE_NOT_ON_ROOT,
11258 : .read_s64 = cpu_weight_nice_read_s64,
11259 : .write_s64 = cpu_weight_nice_write_s64,
11260 : },
11261 : {
11262 : .name = "idle",
11263 : .flags = CFTYPE_NOT_ON_ROOT,
11264 : .read_s64 = cpu_idle_read_s64,
11265 : .write_s64 = cpu_idle_write_s64,
11266 : },
11267 : #endif
11268 : #ifdef CONFIG_CFS_BANDWIDTH
11269 : {
11270 : .name = "max",
11271 : .flags = CFTYPE_NOT_ON_ROOT,
11272 : .seq_show = cpu_max_show,
11273 : .write = cpu_max_write,
11274 : },
11275 : {
11276 : .name = "max.burst",
11277 : .flags = CFTYPE_NOT_ON_ROOT,
11278 : .read_u64 = cpu_cfs_burst_read_u64,
11279 : .write_u64 = cpu_cfs_burst_write_u64,
11280 : },
11281 : #endif
11282 : #ifdef CONFIG_UCLAMP_TASK_GROUP
11283 : {
11284 : .name = "uclamp.min",
11285 : .flags = CFTYPE_NOT_ON_ROOT,
11286 : .seq_show = cpu_uclamp_min_show,
11287 : .write = cpu_uclamp_min_write,
11288 : },
11289 : {
11290 : .name = "uclamp.max",
11291 : .flags = CFTYPE_NOT_ON_ROOT,
11292 : .seq_show = cpu_uclamp_max_show,
11293 : .write = cpu_uclamp_max_write,
11294 : },
11295 : #endif
11296 : { } /* terminate */
11297 : };
11298 :
11299 : struct cgroup_subsys cpu_cgrp_subsys = {
11300 : .css_alloc = cpu_cgroup_css_alloc,
11301 : .css_online = cpu_cgroup_css_online,
11302 : .css_released = cpu_cgroup_css_released,
11303 : .css_free = cpu_cgroup_css_free,
11304 : .css_extra_stat_show = cpu_extra_stat_show,
11305 : #ifdef CONFIG_RT_GROUP_SCHED
11306 : .can_attach = cpu_cgroup_can_attach,
11307 : #endif
11308 : .attach = cpu_cgroup_attach,
11309 : .legacy_cftypes = cpu_legacy_files,
11310 : .dfl_cftypes = cpu_files,
11311 : .early_init = true,
11312 : .threaded = true,
11313 : };
11314 :
11315 : #endif /* CONFIG_CGROUP_SCHED */
11316 :
11317 0 : void dump_cpu_task(int cpu)
11318 : {
11319 0 : if (cpu == smp_processor_id() && in_hardirq()) {
11320 : struct pt_regs *regs;
11321 :
11322 0 : regs = get_irq_regs();
11323 0 : if (regs) {
11324 0 : show_regs(regs);
11325 0 : return;
11326 : }
11327 : }
11328 :
11329 0 : if (trigger_single_cpu_backtrace(cpu))
11330 : return;
11331 :
11332 0 : pr_info("Task dump for CPU %d:\n", cpu);
11333 0 : sched_show_task(cpu_curr(cpu));
11334 : }
11335 :
11336 : /*
11337 : * Nice levels are multiplicative, with a gentle 10% change for every
11338 : * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11339 : * nice 1, it will get ~10% less CPU time than another CPU-bound task
11340 : * that remained on nice 0.
11341 : *
11342 : * The "10% effect" is relative and cumulative: from _any_ nice level,
11343 : * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11344 : * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11345 : * If a task goes up by ~10% and another task goes down by ~10% then
11346 : * the relative distance between them is ~25%.)
11347 : */
11348 : const int sched_prio_to_weight[40] = {
11349 : /* -20 */ 88761, 71755, 56483, 46273, 36291,
11350 : /* -15 */ 29154, 23254, 18705, 14949, 11916,
11351 : /* -10 */ 9548, 7620, 6100, 4904, 3906,
11352 : /* -5 */ 3121, 2501, 1991, 1586, 1277,
11353 : /* 0 */ 1024, 820, 655, 526, 423,
11354 : /* 5 */ 335, 272, 215, 172, 137,
11355 : /* 10 */ 110, 87, 70, 56, 45,
11356 : /* 15 */ 36, 29, 23, 18, 15,
11357 : };
11358 :
11359 : /*
11360 : * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11361 : *
11362 : * In cases where the weight does not change often, we can use the
11363 : * precalculated inverse to speed up arithmetics by turning divisions
11364 : * into multiplications:
11365 : */
11366 : const u32 sched_prio_to_wmult[40] = {
11367 : /* -20 */ 48388, 59856, 76040, 92818, 118348,
11368 : /* -15 */ 147320, 184698, 229616, 287308, 360437,
11369 : /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11370 : /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11371 : /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11372 : /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11373 : /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11374 : /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11375 : };
11376 :
11377 0 : void call_trace_sched_update_nr_running(struct rq *rq, int count)
11378 : {
11379 0 : trace_sched_update_nr_running_tp(rq, count);
11380 0 : }
11381 :
11382 : #ifdef CONFIG_SCHED_MM_CID
11383 : void sched_mm_cid_exit_signals(struct task_struct *t)
11384 : {
11385 : struct mm_struct *mm = t->mm;
11386 : unsigned long flags;
11387 :
11388 : if (!mm)
11389 : return;
11390 : local_irq_save(flags);
11391 : mm_cid_put(mm, t->mm_cid);
11392 : t->mm_cid = -1;
11393 : t->mm_cid_active = 0;
11394 : local_irq_restore(flags);
11395 : }
11396 :
11397 : void sched_mm_cid_before_execve(struct task_struct *t)
11398 : {
11399 : struct mm_struct *mm = t->mm;
11400 : unsigned long flags;
11401 :
11402 : if (!mm)
11403 : return;
11404 : local_irq_save(flags);
11405 : mm_cid_put(mm, t->mm_cid);
11406 : t->mm_cid = -1;
11407 : t->mm_cid_active = 0;
11408 : local_irq_restore(flags);
11409 : }
11410 :
11411 : void sched_mm_cid_after_execve(struct task_struct *t)
11412 : {
11413 : struct mm_struct *mm = t->mm;
11414 : unsigned long flags;
11415 :
11416 : if (!mm)
11417 : return;
11418 : local_irq_save(flags);
11419 : t->mm_cid = mm_cid_get(mm);
11420 : t->mm_cid_active = 1;
11421 : local_irq_restore(flags);
11422 : rseq_set_notify_resume(t);
11423 : }
11424 :
11425 : void sched_mm_cid_fork(struct task_struct *t)
11426 : {
11427 : WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
11428 : t->mm_cid_active = 1;
11429 : }
11430 : #endif
|
