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 : #include <trace/events/ipi.h>
84 : #undef CREATE_TRACE_POINTS
85 :
86 : #include "sched.h"
87 : #include "stats.h"
88 : #include "autogroup.h"
89 :
90 : #include "autogroup.h"
91 : #include "pelt.h"
92 : #include "smp.h"
93 : #include "stats.h"
94 :
95 : #include "../workqueue_internal.h"
96 : #include "../../io_uring/io-wq.h"
97 : #include "../smpboot.h"
98 :
99 : EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpu);
100 : EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpumask);
101 :
102 : /*
103 : * Export tracepoints that act as a bare tracehook (ie: have no trace event
104 : * associated with them) to allow external modules to probe them.
105 : */
106 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
107 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
108 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
109 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
110 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
111 : EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
112 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
113 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
114 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
115 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
116 : EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
117 :
118 : DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
119 :
120 : #ifdef CONFIG_SCHED_DEBUG
121 : /*
122 : * Debugging: various feature bits
123 : *
124 : * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
125 : * sysctl_sched_features, defined in sched.h, to allow constants propagation
126 : * at compile time and compiler optimization based on features default.
127 : */
128 : #define SCHED_FEAT(name, enabled) \
129 : (1UL << __SCHED_FEAT_##name) * enabled |
130 : const_debug unsigned int sysctl_sched_features =
131 : #include "features.h"
132 : 0;
133 : #undef SCHED_FEAT
134 :
135 : /*
136 : * Print a warning if need_resched is set for the given duration (if
137 : * LATENCY_WARN is enabled).
138 : *
139 : * If sysctl_resched_latency_warn_once is set, only one warning will be shown
140 : * per boot.
141 : */
142 : __read_mostly int sysctl_resched_latency_warn_ms = 100;
143 : __read_mostly int sysctl_resched_latency_warn_once = 1;
144 : #endif /* CONFIG_SCHED_DEBUG */
145 :
146 : /*
147 : * Number of tasks to iterate in a single balance run.
148 : * Limited because this is done with IRQs disabled.
149 : */
150 : const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
151 :
152 : __read_mostly int scheduler_running;
153 :
154 : #ifdef CONFIG_SCHED_CORE
155 :
156 : DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
157 :
158 : /* kernel prio, less is more */
159 : static inline int __task_prio(const struct task_struct *p)
160 : {
161 : if (p->sched_class == &stop_sched_class) /* trumps deadline */
162 : return -2;
163 :
164 : if (rt_prio(p->prio)) /* includes deadline */
165 : return p->prio; /* [-1, 99] */
166 :
167 : if (p->sched_class == &idle_sched_class)
168 : return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
169 :
170 : return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
171 : }
172 :
173 : /*
174 : * l(a,b)
175 : * le(a,b) := !l(b,a)
176 : * g(a,b) := l(b,a)
177 : * ge(a,b) := !l(a,b)
178 : */
179 :
180 : /* real prio, less is less */
181 : static inline bool prio_less(const struct task_struct *a,
182 : const struct task_struct *b, bool in_fi)
183 : {
184 :
185 : int pa = __task_prio(a), pb = __task_prio(b);
186 :
187 : if (-pa < -pb)
188 : return true;
189 :
190 : if (-pb < -pa)
191 : return false;
192 :
193 : if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
194 : return !dl_time_before(a->dl.deadline, b->dl.deadline);
195 :
196 : if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
197 : return cfs_prio_less(a, b, in_fi);
198 :
199 : return false;
200 : }
201 :
202 : static inline bool __sched_core_less(const struct task_struct *a,
203 : const struct task_struct *b)
204 : {
205 : if (a->core_cookie < b->core_cookie)
206 : return true;
207 :
208 : if (a->core_cookie > b->core_cookie)
209 : return false;
210 :
211 : /* flip prio, so high prio is leftmost */
212 : if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
213 : return true;
214 :
215 : return false;
216 : }
217 :
218 : #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
219 :
220 : static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
221 : {
222 : return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
223 : }
224 :
225 : static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
226 : {
227 : const struct task_struct *p = __node_2_sc(node);
228 : unsigned long cookie = (unsigned long)key;
229 :
230 : if (cookie < p->core_cookie)
231 : return -1;
232 :
233 : if (cookie > p->core_cookie)
234 : return 1;
235 :
236 : return 0;
237 : }
238 :
239 : void sched_core_enqueue(struct rq *rq, struct task_struct *p)
240 : {
241 : rq->core->core_task_seq++;
242 :
243 : if (!p->core_cookie)
244 : return;
245 :
246 : rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
247 : }
248 :
249 : void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
250 : {
251 : rq->core->core_task_seq++;
252 :
253 : if (sched_core_enqueued(p)) {
254 : rb_erase(&p->core_node, &rq->core_tree);
255 : RB_CLEAR_NODE(&p->core_node);
256 : }
257 :
258 : /*
259 : * Migrating the last task off the cpu, with the cpu in forced idle
260 : * state. Reschedule to create an accounting edge for forced idle,
261 : * and re-examine whether the core is still in forced idle state.
262 : */
263 : if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
264 : rq->core->core_forceidle_count && rq->curr == rq->idle)
265 : resched_curr(rq);
266 : }
267 :
268 : static int sched_task_is_throttled(struct task_struct *p, int cpu)
269 : {
270 : if (p->sched_class->task_is_throttled)
271 : return p->sched_class->task_is_throttled(p, cpu);
272 :
273 : return 0;
274 : }
275 :
276 : static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
277 : {
278 : struct rb_node *node = &p->core_node;
279 : int cpu = task_cpu(p);
280 :
281 : do {
282 : node = rb_next(node);
283 : if (!node)
284 : return NULL;
285 :
286 : p = __node_2_sc(node);
287 : if (p->core_cookie != cookie)
288 : return NULL;
289 :
290 : } while (sched_task_is_throttled(p, cpu));
291 :
292 : return p;
293 : }
294 :
295 : /*
296 : * Find left-most (aka, highest priority) and unthrottled task matching @cookie.
297 : * If no suitable task is found, NULL will be returned.
298 : */
299 : static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
300 : {
301 : struct task_struct *p;
302 : struct rb_node *node;
303 :
304 : node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
305 : if (!node)
306 : return NULL;
307 :
308 : p = __node_2_sc(node);
309 : if (!sched_task_is_throttled(p, rq->cpu))
310 : return p;
311 :
312 : return sched_core_next(p, cookie);
313 : }
314 :
315 : /*
316 : * Magic required such that:
317 : *
318 : * raw_spin_rq_lock(rq);
319 : * ...
320 : * raw_spin_rq_unlock(rq);
321 : *
322 : * ends up locking and unlocking the _same_ lock, and all CPUs
323 : * always agree on what rq has what lock.
324 : *
325 : * XXX entirely possible to selectively enable cores, don't bother for now.
326 : */
327 :
328 : static DEFINE_MUTEX(sched_core_mutex);
329 : static atomic_t sched_core_count;
330 : static struct cpumask sched_core_mask;
331 :
332 : static void sched_core_lock(int cpu, unsigned long *flags)
333 : {
334 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
335 : int t, i = 0;
336 :
337 : local_irq_save(*flags);
338 : for_each_cpu(t, smt_mask)
339 : raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
340 : }
341 :
342 : static void sched_core_unlock(int cpu, unsigned long *flags)
343 : {
344 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
345 : int t;
346 :
347 : for_each_cpu(t, smt_mask)
348 : raw_spin_unlock(&cpu_rq(t)->__lock);
349 : local_irq_restore(*flags);
350 : }
351 :
352 : static void __sched_core_flip(bool enabled)
353 : {
354 : unsigned long flags;
355 : int cpu, t;
356 :
357 : cpus_read_lock();
358 :
359 : /*
360 : * Toggle the online cores, one by one.
361 : */
362 : cpumask_copy(&sched_core_mask, cpu_online_mask);
363 : for_each_cpu(cpu, &sched_core_mask) {
364 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
365 :
366 : sched_core_lock(cpu, &flags);
367 :
368 : for_each_cpu(t, smt_mask)
369 : cpu_rq(t)->core_enabled = enabled;
370 :
371 : cpu_rq(cpu)->core->core_forceidle_start = 0;
372 :
373 : sched_core_unlock(cpu, &flags);
374 :
375 : cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
376 : }
377 :
378 : /*
379 : * Toggle the offline CPUs.
380 : */
381 : for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
382 : cpu_rq(cpu)->core_enabled = enabled;
383 :
384 : cpus_read_unlock();
385 : }
386 :
387 : static void sched_core_assert_empty(void)
388 : {
389 : int cpu;
390 :
391 : for_each_possible_cpu(cpu)
392 : WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
393 : }
394 :
395 : static void __sched_core_enable(void)
396 : {
397 : static_branch_enable(&__sched_core_enabled);
398 : /*
399 : * Ensure all previous instances of raw_spin_rq_*lock() have finished
400 : * and future ones will observe !sched_core_disabled().
401 : */
402 : synchronize_rcu();
403 : __sched_core_flip(true);
404 : sched_core_assert_empty();
405 : }
406 :
407 : static void __sched_core_disable(void)
408 : {
409 : sched_core_assert_empty();
410 : __sched_core_flip(false);
411 : static_branch_disable(&__sched_core_enabled);
412 : }
413 :
414 : void sched_core_get(void)
415 : {
416 : if (atomic_inc_not_zero(&sched_core_count))
417 : return;
418 :
419 : mutex_lock(&sched_core_mutex);
420 : if (!atomic_read(&sched_core_count))
421 : __sched_core_enable();
422 :
423 : smp_mb__before_atomic();
424 : atomic_inc(&sched_core_count);
425 : mutex_unlock(&sched_core_mutex);
426 : }
427 :
428 : static void __sched_core_put(struct work_struct *work)
429 : {
430 : if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
431 : __sched_core_disable();
432 : mutex_unlock(&sched_core_mutex);
433 : }
434 : }
435 :
436 : void sched_core_put(void)
437 : {
438 : static DECLARE_WORK(_work, __sched_core_put);
439 :
440 : /*
441 : * "There can be only one"
442 : *
443 : * Either this is the last one, or we don't actually need to do any
444 : * 'work'. If it is the last *again*, we rely on
445 : * WORK_STRUCT_PENDING_BIT.
446 : */
447 : if (!atomic_add_unless(&sched_core_count, -1, 1))
448 : schedule_work(&_work);
449 : }
450 :
451 : #else /* !CONFIG_SCHED_CORE */
452 :
453 : static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
454 : static inline void
455 : sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
456 :
457 : #endif /* CONFIG_SCHED_CORE */
458 :
459 : /*
460 : * Serialization rules:
461 : *
462 : * Lock order:
463 : *
464 : * p->pi_lock
465 : * rq->lock
466 : * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
467 : *
468 : * rq1->lock
469 : * rq2->lock where: rq1 < rq2
470 : *
471 : * Regular state:
472 : *
473 : * Normal scheduling state is serialized by rq->lock. __schedule() takes the
474 : * local CPU's rq->lock, it optionally removes the task from the runqueue and
475 : * always looks at the local rq data structures to find the most eligible task
476 : * to run next.
477 : *
478 : * Task enqueue is also under rq->lock, possibly taken from another CPU.
479 : * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
480 : * the local CPU to avoid bouncing the runqueue state around [ see
481 : * ttwu_queue_wakelist() ]
482 : *
483 : * Task wakeup, specifically wakeups that involve migration, are horribly
484 : * complicated to avoid having to take two rq->locks.
485 : *
486 : * Special state:
487 : *
488 : * System-calls and anything external will use task_rq_lock() which acquires
489 : * both p->pi_lock and rq->lock. As a consequence the state they change is
490 : * stable while holding either lock:
491 : *
492 : * - sched_setaffinity()/
493 : * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
494 : * - set_user_nice(): p->se.load, p->*prio
495 : * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
496 : * p->se.load, p->rt_priority,
497 : * p->dl.dl_{runtime, deadline, period, flags, bw, density}
498 : * - sched_setnuma(): p->numa_preferred_nid
499 : * - sched_move_task(): p->sched_task_group
500 : * - uclamp_update_active() p->uclamp*
501 : *
502 : * p->state <- TASK_*:
503 : *
504 : * is changed locklessly using set_current_state(), __set_current_state() or
505 : * set_special_state(), see their respective comments, or by
506 : * try_to_wake_up(). This latter uses p->pi_lock to serialize against
507 : * concurrent self.
508 : *
509 : * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
510 : *
511 : * is set by activate_task() and cleared by deactivate_task(), under
512 : * rq->lock. Non-zero indicates the task is runnable, the special
513 : * ON_RQ_MIGRATING state is used for migration without holding both
514 : * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
515 : *
516 : * p->on_cpu <- { 0, 1 }:
517 : *
518 : * is set by prepare_task() and cleared by finish_task() such that it will be
519 : * set before p is scheduled-in and cleared after p is scheduled-out, both
520 : * under rq->lock. Non-zero indicates the task is running on its CPU.
521 : *
522 : * [ The astute reader will observe that it is possible for two tasks on one
523 : * CPU to have ->on_cpu = 1 at the same time. ]
524 : *
525 : * task_cpu(p): is changed by set_task_cpu(), the rules are:
526 : *
527 : * - Don't call set_task_cpu() on a blocked task:
528 : *
529 : * We don't care what CPU we're not running on, this simplifies hotplug,
530 : * the CPU assignment of blocked tasks isn't required to be valid.
531 : *
532 : * - for try_to_wake_up(), called under p->pi_lock:
533 : *
534 : * This allows try_to_wake_up() to only take one rq->lock, see its comment.
535 : *
536 : * - for migration called under rq->lock:
537 : * [ see task_on_rq_migrating() in task_rq_lock() ]
538 : *
539 : * o move_queued_task()
540 : * o detach_task()
541 : *
542 : * - for migration called under double_rq_lock():
543 : *
544 : * o __migrate_swap_task()
545 : * o push_rt_task() / pull_rt_task()
546 : * o push_dl_task() / pull_dl_task()
547 : * o dl_task_offline_migration()
548 : *
549 : */
550 :
551 382 : void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
552 : {
553 : raw_spinlock_t *lock;
554 :
555 : /* Matches synchronize_rcu() in __sched_core_enable() */
556 8666 : preempt_disable();
557 : if (sched_core_disabled()) {
558 8666 : raw_spin_lock_nested(&rq->__lock, subclass);
559 : /* preempt_count *MUST* be > 1 */
560 8666 : preempt_enable_no_resched();
561 : return;
562 : }
563 :
564 : for (;;) {
565 : lock = __rq_lockp(rq);
566 : raw_spin_lock_nested(lock, subclass);
567 : if (likely(lock == __rq_lockp(rq))) {
568 : /* preempt_count *MUST* be > 1 */
569 : preempt_enable_no_resched();
570 : return;
571 : }
572 : raw_spin_unlock(lock);
573 : }
574 : }
575 :
576 0 : bool raw_spin_rq_trylock(struct rq *rq)
577 : {
578 : raw_spinlock_t *lock;
579 : bool ret;
580 :
581 : /* Matches synchronize_rcu() in __sched_core_enable() */
582 0 : preempt_disable();
583 : if (sched_core_disabled()) {
584 0 : ret = raw_spin_trylock(&rq->__lock);
585 0 : preempt_enable();
586 : return ret;
587 : }
588 :
589 : for (;;) {
590 : lock = __rq_lockp(rq);
591 : ret = raw_spin_trylock(lock);
592 : if (!ret || (likely(lock == __rq_lockp(rq)))) {
593 : preempt_enable();
594 : return ret;
595 : }
596 : raw_spin_unlock(lock);
597 : }
598 : }
599 :
600 382 : void raw_spin_rq_unlock(struct rq *rq)
601 : {
602 8666 : raw_spin_unlock(rq_lockp(rq));
603 382 : }
604 :
605 : #ifdef CONFIG_SMP
606 : /*
607 : * double_rq_lock - safely lock two runqueues
608 : */
609 : void double_rq_lock(struct rq *rq1, struct rq *rq2)
610 : {
611 : lockdep_assert_irqs_disabled();
612 :
613 : if (rq_order_less(rq2, rq1))
614 : swap(rq1, rq2);
615 :
616 : raw_spin_rq_lock(rq1);
617 : if (__rq_lockp(rq1) != __rq_lockp(rq2))
618 : raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
619 :
620 : double_rq_clock_clear_update(rq1, rq2);
621 : }
622 : #endif
623 :
624 : /*
625 : * __task_rq_lock - lock the rq @p resides on.
626 : */
627 0 : struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
628 : __acquires(rq->lock)
629 : {
630 : struct rq *rq;
631 :
632 : lockdep_assert_held(&p->pi_lock);
633 :
634 : for (;;) {
635 382 : rq = task_rq(p);
636 382 : raw_spin_rq_lock(rq);
637 764 : if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
638 382 : rq_pin_lock(rq, rf);
639 0 : return rq;
640 : }
641 : raw_spin_rq_unlock(rq);
642 :
643 0 : while (unlikely(task_on_rq_migrating(p)))
644 : cpu_relax();
645 : }
646 : }
647 :
648 : /*
649 : * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
650 : */
651 385 : struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
652 : __acquires(p->pi_lock)
653 : __acquires(rq->lock)
654 : {
655 : struct rq *rq;
656 :
657 : for (;;) {
658 385 : raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
659 385 : rq = task_rq(p);
660 385 : raw_spin_rq_lock(rq);
661 : /*
662 : * move_queued_task() task_rq_lock()
663 : *
664 : * ACQUIRE (rq->lock)
665 : * [S] ->on_rq = MIGRATING [L] rq = task_rq()
666 : * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
667 : * [S] ->cpu = new_cpu [L] task_rq()
668 : * [L] ->on_rq
669 : * RELEASE (rq->lock)
670 : *
671 : * If we observe the old CPU in task_rq_lock(), the acquire of
672 : * the old rq->lock will fully serialize against the stores.
673 : *
674 : * If we observe the new CPU in task_rq_lock(), the address
675 : * dependency headed by '[L] rq = task_rq()' and the acquire
676 : * will pair with the WMB to ensure we then also see migrating.
677 : */
678 770 : if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
679 385 : rq_pin_lock(rq, rf);
680 385 : return rq;
681 : }
682 0 : raw_spin_rq_unlock(rq);
683 0 : raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
684 :
685 0 : while (unlikely(task_on_rq_migrating(p)))
686 : cpu_relax();
687 : }
688 : }
689 :
690 : /*
691 : * RQ-clock updating methods:
692 : */
693 :
694 : static void update_rq_clock_task(struct rq *rq, s64 delta)
695 : {
696 : /*
697 : * In theory, the compile should just see 0 here, and optimize out the call
698 : * to sched_rt_avg_update. But I don't trust it...
699 : */
700 7826 : s64 __maybe_unused steal = 0, irq_delta = 0;
701 :
702 : #ifdef CONFIG_IRQ_TIME_ACCOUNTING
703 : irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
704 :
705 : /*
706 : * Since irq_time is only updated on {soft,}irq_exit, we might run into
707 : * this case when a previous update_rq_clock() happened inside a
708 : * {soft,}irq region.
709 : *
710 : * When this happens, we stop ->clock_task and only update the
711 : * prev_irq_time stamp to account for the part that fit, so that a next
712 : * update will consume the rest. This ensures ->clock_task is
713 : * monotonic.
714 : *
715 : * It does however cause some slight miss-attribution of {soft,}irq
716 : * time, a more accurate solution would be to update the irq_time using
717 : * the current rq->clock timestamp, except that would require using
718 : * atomic ops.
719 : */
720 : if (irq_delta > delta)
721 : irq_delta = delta;
722 :
723 : rq->prev_irq_time += irq_delta;
724 : delta -= irq_delta;
725 : psi_account_irqtime(rq->curr, irq_delta);
726 : delayacct_irq(rq->curr, irq_delta);
727 : #endif
728 : #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
729 : if (static_key_false((¶virt_steal_rq_enabled))) {
730 : steal = paravirt_steal_clock(cpu_of(rq));
731 : steal -= rq->prev_steal_time_rq;
732 :
733 : if (unlikely(steal > delta))
734 : steal = delta;
735 :
736 : rq->prev_steal_time_rq += steal;
737 : delta -= steal;
738 : }
739 : #endif
740 :
741 7826 : rq->clock_task += delta;
742 :
743 : #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
744 : if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
745 : update_irq_load_avg(rq, irq_delta + steal);
746 : #endif
747 7826 : update_rq_clock_pelt(rq, delta);
748 : }
749 :
750 8665 : void update_rq_clock(struct rq *rq)
751 : {
752 : s64 delta;
753 :
754 8665 : lockdep_assert_rq_held(rq);
755 :
756 8665 : if (rq->clock_update_flags & RQCF_ACT_SKIP)
757 : return;
758 :
759 : #ifdef CONFIG_SCHED_DEBUG
760 : if (sched_feat(WARN_DOUBLE_CLOCK))
761 : SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
762 : rq->clock_update_flags |= RQCF_UPDATED;
763 : #endif
764 :
765 7826 : delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
766 7826 : if (delta < 0)
767 : return;
768 7826 : rq->clock += delta;
769 7826 : update_rq_clock_task(rq, delta);
770 : }
771 :
772 : #ifdef CONFIG_SCHED_HRTICK
773 : /*
774 : * Use HR-timers to deliver accurate preemption points.
775 : */
776 :
777 : static void hrtick_clear(struct rq *rq)
778 : {
779 : if (hrtimer_active(&rq->hrtick_timer))
780 : hrtimer_cancel(&rq->hrtick_timer);
781 : }
782 :
783 : /*
784 : * High-resolution timer tick.
785 : * Runs from hardirq context with interrupts disabled.
786 : */
787 : static enum hrtimer_restart hrtick(struct hrtimer *timer)
788 : {
789 : struct rq *rq = container_of(timer, struct rq, hrtick_timer);
790 : struct rq_flags rf;
791 :
792 : WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
793 :
794 : rq_lock(rq, &rf);
795 : update_rq_clock(rq);
796 : rq->curr->sched_class->task_tick(rq, rq->curr, 1);
797 : rq_unlock(rq, &rf);
798 :
799 : return HRTIMER_NORESTART;
800 : }
801 :
802 : #ifdef CONFIG_SMP
803 :
804 : static void __hrtick_restart(struct rq *rq)
805 : {
806 : struct hrtimer *timer = &rq->hrtick_timer;
807 : ktime_t time = rq->hrtick_time;
808 :
809 : hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
810 : }
811 :
812 : /*
813 : * called from hardirq (IPI) context
814 : */
815 : static void __hrtick_start(void *arg)
816 : {
817 : struct rq *rq = arg;
818 : struct rq_flags rf;
819 :
820 : rq_lock(rq, &rf);
821 : __hrtick_restart(rq);
822 : rq_unlock(rq, &rf);
823 : }
824 :
825 : /*
826 : * Called to set the hrtick timer state.
827 : *
828 : * called with rq->lock held and irqs disabled
829 : */
830 : void hrtick_start(struct rq *rq, u64 delay)
831 : {
832 : struct hrtimer *timer = &rq->hrtick_timer;
833 : s64 delta;
834 :
835 : /*
836 : * Don't schedule slices shorter than 10000ns, that just
837 : * doesn't make sense and can cause timer DoS.
838 : */
839 : delta = max_t(s64, delay, 10000LL);
840 : rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
841 :
842 : if (rq == this_rq())
843 : __hrtick_restart(rq);
844 : else
845 : smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
846 : }
847 :
848 : #else
849 : /*
850 : * Called to set the hrtick timer state.
851 : *
852 : * called with rq->lock held and irqs disabled
853 : */
854 : void hrtick_start(struct rq *rq, u64 delay)
855 : {
856 : /*
857 : * Don't schedule slices shorter than 10000ns, that just
858 : * doesn't make sense. Rely on vruntime for fairness.
859 : */
860 : delay = max_t(u64, delay, 10000LL);
861 : hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
862 : HRTIMER_MODE_REL_PINNED_HARD);
863 : }
864 :
865 : #endif /* CONFIG_SMP */
866 :
867 : static void hrtick_rq_init(struct rq *rq)
868 : {
869 : #ifdef CONFIG_SMP
870 : INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
871 : #endif
872 : hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
873 : rq->hrtick_timer.function = hrtick;
874 : }
875 : #else /* CONFIG_SCHED_HRTICK */
876 : static inline void hrtick_clear(struct rq *rq)
877 : {
878 : }
879 :
880 : static inline void hrtick_rq_init(struct rq *rq)
881 : {
882 : }
883 : #endif /* CONFIG_SCHED_HRTICK */
884 :
885 : /*
886 : * cmpxchg based fetch_or, macro so it works for different integer types
887 : */
888 : #define fetch_or(ptr, mask) \
889 : ({ \
890 : typeof(ptr) _ptr = (ptr); \
891 : typeof(mask) _mask = (mask); \
892 : typeof(*_ptr) _val = *_ptr; \
893 : \
894 : do { \
895 : } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
896 : _val; \
897 : })
898 :
899 : #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
900 : /*
901 : * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
902 : * this avoids any races wrt polling state changes and thereby avoids
903 : * spurious IPIs.
904 : */
905 : static inline bool set_nr_and_not_polling(struct task_struct *p)
906 : {
907 : struct thread_info *ti = task_thread_info(p);
908 : return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
909 : }
910 :
911 : /*
912 : * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
913 : *
914 : * If this returns true, then the idle task promises to call
915 : * sched_ttwu_pending() and reschedule soon.
916 : */
917 : static bool set_nr_if_polling(struct task_struct *p)
918 : {
919 : struct thread_info *ti = task_thread_info(p);
920 : typeof(ti->flags) val = READ_ONCE(ti->flags);
921 :
922 : for (;;) {
923 : if (!(val & _TIF_POLLING_NRFLAG))
924 : return false;
925 : if (val & _TIF_NEED_RESCHED)
926 : return true;
927 : if (try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED))
928 : break;
929 : }
930 : return true;
931 : }
932 :
933 : #else
934 : static inline bool set_nr_and_not_polling(struct task_struct *p)
935 : {
936 : set_tsk_need_resched(p);
937 : return true;
938 : }
939 :
940 : #ifdef CONFIG_SMP
941 : static inline bool set_nr_if_polling(struct task_struct *p)
942 : {
943 : return false;
944 : }
945 : #endif
946 : #endif
947 :
948 : static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
949 : {
950 0 : struct wake_q_node *node = &task->wake_q;
951 :
952 : /*
953 : * Atomically grab the task, if ->wake_q is !nil already it means
954 : * it's already queued (either by us or someone else) and will get the
955 : * wakeup due to that.
956 : *
957 : * In order to ensure that a pending wakeup will observe our pending
958 : * state, even in the failed case, an explicit smp_mb() must be used.
959 : */
960 0 : smp_mb__before_atomic();
961 0 : if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
962 : return false;
963 :
964 : /*
965 : * The head is context local, there can be no concurrency.
966 : */
967 0 : *head->lastp = node;
968 0 : head->lastp = &node->next;
969 : return true;
970 : }
971 :
972 : /**
973 : * wake_q_add() - queue a wakeup for 'later' waking.
974 : * @head: the wake_q_head to add @task to
975 : * @task: the task to queue for 'later' wakeup
976 : *
977 : * Queue a task for later wakeup, most likely by the wake_up_q() call in the
978 : * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
979 : * instantly.
980 : *
981 : * This function must be used as-if it were wake_up_process(); IOW the task
982 : * must be ready to be woken at this location.
983 : */
984 0 : void wake_q_add(struct wake_q_head *head, struct task_struct *task)
985 : {
986 0 : if (__wake_q_add(head, task))
987 : get_task_struct(task);
988 0 : }
989 :
990 : /**
991 : * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
992 : * @head: the wake_q_head to add @task to
993 : * @task: the task to queue for 'later' wakeup
994 : *
995 : * Queue a task for later wakeup, most likely by the wake_up_q() call in the
996 : * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
997 : * instantly.
998 : *
999 : * This function must be used as-if it were wake_up_process(); IOW the task
1000 : * must be ready to be woken at this location.
1001 : *
1002 : * This function is essentially a task-safe equivalent to wake_q_add(). Callers
1003 : * that already hold reference to @task can call the 'safe' version and trust
1004 : * wake_q to do the right thing depending whether or not the @task is already
1005 : * queued for wakeup.
1006 : */
1007 0 : void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
1008 : {
1009 0 : if (!__wake_q_add(head, task))
1010 0 : put_task_struct(task);
1011 0 : }
1012 :
1013 0 : void wake_up_q(struct wake_q_head *head)
1014 : {
1015 0 : struct wake_q_node *node = head->first;
1016 :
1017 0 : while (node != WAKE_Q_TAIL) {
1018 : struct task_struct *task;
1019 :
1020 0 : task = container_of(node, struct task_struct, wake_q);
1021 : /* Task can safely be re-inserted now: */
1022 0 : node = node->next;
1023 0 : task->wake_q.next = NULL;
1024 :
1025 : /*
1026 : * wake_up_process() executes a full barrier, which pairs with
1027 : * the queueing in wake_q_add() so as not to miss wakeups.
1028 : */
1029 0 : wake_up_process(task);
1030 0 : put_task_struct(task);
1031 : }
1032 0 : }
1033 :
1034 : /*
1035 : * resched_curr - mark rq's current task 'to be rescheduled now'.
1036 : *
1037 : * On UP this means the setting of the need_resched flag, on SMP it
1038 : * might also involve a cross-CPU call to trigger the scheduler on
1039 : * the target CPU.
1040 : */
1041 1075 : void resched_curr(struct rq *rq)
1042 : {
1043 1075 : struct task_struct *curr = rq->curr;
1044 : int cpu;
1045 :
1046 2150 : lockdep_assert_rq_held(rq);
1047 :
1048 1075 : if (test_tsk_need_resched(curr))
1049 : return;
1050 :
1051 843 : cpu = cpu_of(rq);
1052 :
1053 : if (cpu == smp_processor_id()) {
1054 : set_tsk_need_resched(curr);
1055 : set_preempt_need_resched();
1056 : return;
1057 : }
1058 :
1059 : if (set_nr_and_not_polling(curr))
1060 : smp_send_reschedule(cpu);
1061 : else
1062 : trace_sched_wake_idle_without_ipi(cpu);
1063 : }
1064 :
1065 0 : void resched_cpu(int cpu)
1066 : {
1067 0 : struct rq *rq = cpu_rq(cpu);
1068 : unsigned long flags;
1069 :
1070 0 : raw_spin_rq_lock_irqsave(rq, flags);
1071 0 : if (cpu_online(cpu) || cpu == smp_processor_id())
1072 0 : resched_curr(rq);
1073 0 : raw_spin_rq_unlock_irqrestore(rq, flags);
1074 0 : }
1075 :
1076 : #ifdef CONFIG_SMP
1077 : #ifdef CONFIG_NO_HZ_COMMON
1078 : /*
1079 : * In the semi idle case, use the nearest busy CPU for migrating timers
1080 : * from an idle CPU. This is good for power-savings.
1081 : *
1082 : * We don't do similar optimization for completely idle system, as
1083 : * selecting an idle CPU will add more delays to the timers than intended
1084 : * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1085 : */
1086 : int get_nohz_timer_target(void)
1087 : {
1088 : int i, cpu = smp_processor_id(), default_cpu = -1;
1089 : struct sched_domain *sd;
1090 : const struct cpumask *hk_mask;
1091 :
1092 : if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1093 : if (!idle_cpu(cpu))
1094 : return cpu;
1095 : default_cpu = cpu;
1096 : }
1097 :
1098 : hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1099 :
1100 : rcu_read_lock();
1101 : for_each_domain(cpu, sd) {
1102 : for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1103 : if (cpu == i)
1104 : continue;
1105 :
1106 : if (!idle_cpu(i)) {
1107 : cpu = i;
1108 : goto unlock;
1109 : }
1110 : }
1111 : }
1112 :
1113 : if (default_cpu == -1)
1114 : default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1115 : cpu = default_cpu;
1116 : unlock:
1117 : rcu_read_unlock();
1118 : return cpu;
1119 : }
1120 :
1121 : /*
1122 : * When add_timer_on() enqueues a timer into the timer wheel of an
1123 : * idle CPU then this timer might expire before the next timer event
1124 : * which is scheduled to wake up that CPU. In case of a completely
1125 : * idle system the next event might even be infinite time into the
1126 : * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1127 : * leaves the inner idle loop so the newly added timer is taken into
1128 : * account when the CPU goes back to idle and evaluates the timer
1129 : * wheel for the next timer event.
1130 : */
1131 : static void wake_up_idle_cpu(int cpu)
1132 : {
1133 : struct rq *rq = cpu_rq(cpu);
1134 :
1135 : if (cpu == smp_processor_id())
1136 : return;
1137 :
1138 : if (set_nr_and_not_polling(rq->idle))
1139 : smp_send_reschedule(cpu);
1140 : else
1141 : trace_sched_wake_idle_without_ipi(cpu);
1142 : }
1143 :
1144 : static bool wake_up_full_nohz_cpu(int cpu)
1145 : {
1146 : /*
1147 : * We just need the target to call irq_exit() and re-evaluate
1148 : * the next tick. The nohz full kick at least implies that.
1149 : * If needed we can still optimize that later with an
1150 : * empty IRQ.
1151 : */
1152 : if (cpu_is_offline(cpu))
1153 : return true; /* Don't try to wake offline CPUs. */
1154 : if (tick_nohz_full_cpu(cpu)) {
1155 : if (cpu != smp_processor_id() ||
1156 : tick_nohz_tick_stopped())
1157 : tick_nohz_full_kick_cpu(cpu);
1158 : return true;
1159 : }
1160 :
1161 : return false;
1162 : }
1163 :
1164 : /*
1165 : * Wake up the specified CPU. If the CPU is going offline, it is the
1166 : * caller's responsibility to deal with the lost wakeup, for example,
1167 : * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1168 : */
1169 : void wake_up_nohz_cpu(int cpu)
1170 : {
1171 : if (!wake_up_full_nohz_cpu(cpu))
1172 : wake_up_idle_cpu(cpu);
1173 : }
1174 :
1175 : static void nohz_csd_func(void *info)
1176 : {
1177 : struct rq *rq = info;
1178 : int cpu = cpu_of(rq);
1179 : unsigned int flags;
1180 :
1181 : /*
1182 : * Release the rq::nohz_csd.
1183 : */
1184 : flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1185 : WARN_ON(!(flags & NOHZ_KICK_MASK));
1186 :
1187 : rq->idle_balance = idle_cpu(cpu);
1188 : if (rq->idle_balance && !need_resched()) {
1189 : rq->nohz_idle_balance = flags;
1190 : raise_softirq_irqoff(SCHED_SOFTIRQ);
1191 : }
1192 : }
1193 :
1194 : #endif /* CONFIG_NO_HZ_COMMON */
1195 :
1196 : #ifdef CONFIG_NO_HZ_FULL
1197 : bool sched_can_stop_tick(struct rq *rq)
1198 : {
1199 : int fifo_nr_running;
1200 :
1201 : /* Deadline tasks, even if single, need the tick */
1202 : if (rq->dl.dl_nr_running)
1203 : return false;
1204 :
1205 : /*
1206 : * If there are more than one RR tasks, we need the tick to affect the
1207 : * actual RR behaviour.
1208 : */
1209 : if (rq->rt.rr_nr_running) {
1210 : if (rq->rt.rr_nr_running == 1)
1211 : return true;
1212 : else
1213 : return false;
1214 : }
1215 :
1216 : /*
1217 : * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1218 : * forced preemption between FIFO tasks.
1219 : */
1220 : fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1221 : if (fifo_nr_running)
1222 : return true;
1223 :
1224 : /*
1225 : * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1226 : * if there's more than one we need the tick for involuntary
1227 : * preemption.
1228 : */
1229 : if (rq->nr_running > 1)
1230 : return false;
1231 :
1232 : return true;
1233 : }
1234 : #endif /* CONFIG_NO_HZ_FULL */
1235 : #endif /* CONFIG_SMP */
1236 :
1237 : #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1238 : (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1239 : /*
1240 : * Iterate task_group tree rooted at *from, calling @down when first entering a
1241 : * node and @up when leaving it for the final time.
1242 : *
1243 : * Caller must hold rcu_lock or sufficient equivalent.
1244 : */
1245 : int walk_tg_tree_from(struct task_group *from,
1246 : tg_visitor down, tg_visitor up, void *data)
1247 : {
1248 : struct task_group *parent, *child;
1249 : int ret;
1250 :
1251 : parent = from;
1252 :
1253 : down:
1254 : ret = (*down)(parent, data);
1255 : if (ret)
1256 : goto out;
1257 : list_for_each_entry_rcu(child, &parent->children, siblings) {
1258 : parent = child;
1259 : goto down;
1260 :
1261 : up:
1262 : continue;
1263 : }
1264 : ret = (*up)(parent, data);
1265 : if (ret || parent == from)
1266 : goto out;
1267 :
1268 : child = parent;
1269 : parent = parent->parent;
1270 : if (parent)
1271 : goto up;
1272 : out:
1273 : return ret;
1274 : }
1275 :
1276 : int tg_nop(struct task_group *tg, void *data)
1277 : {
1278 : return 0;
1279 : }
1280 : #endif
1281 :
1282 6 : static void set_load_weight(struct task_struct *p, bool update_load)
1283 : {
1284 6 : int prio = p->static_prio - MAX_RT_PRIO;
1285 6 : struct load_weight *load = &p->se.load;
1286 :
1287 : /*
1288 : * SCHED_IDLE tasks get minimal weight:
1289 : */
1290 12 : if (task_has_idle_policy(p)) {
1291 0 : load->weight = scale_load(WEIGHT_IDLEPRIO);
1292 0 : load->inv_weight = WMULT_IDLEPRIO;
1293 0 : return;
1294 : }
1295 :
1296 : /*
1297 : * SCHED_OTHER tasks have to update their load when changing their
1298 : * weight
1299 : */
1300 6 : if (update_load && p->sched_class == &fair_sched_class) {
1301 5 : reweight_task(p, prio);
1302 : } else {
1303 1 : load->weight = scale_load(sched_prio_to_weight[prio]);
1304 1 : load->inv_weight = sched_prio_to_wmult[prio];
1305 : }
1306 : }
1307 :
1308 : #ifdef CONFIG_UCLAMP_TASK
1309 : /*
1310 : * Serializes updates of utilization clamp values
1311 : *
1312 : * The (slow-path) user-space triggers utilization clamp value updates which
1313 : * can require updates on (fast-path) scheduler's data structures used to
1314 : * support enqueue/dequeue operations.
1315 : * While the per-CPU rq lock protects fast-path update operations, user-space
1316 : * requests are serialized using a mutex to reduce the risk of conflicting
1317 : * updates or API abuses.
1318 : */
1319 : static DEFINE_MUTEX(uclamp_mutex);
1320 :
1321 : /* Max allowed minimum utilization */
1322 : static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1323 :
1324 : /* Max allowed maximum utilization */
1325 : static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1326 :
1327 : /*
1328 : * By default RT tasks run at the maximum performance point/capacity of the
1329 : * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1330 : * SCHED_CAPACITY_SCALE.
1331 : *
1332 : * This knob allows admins to change the default behavior when uclamp is being
1333 : * used. In battery powered devices, particularly, running at the maximum
1334 : * capacity and frequency will increase energy consumption and shorten the
1335 : * battery life.
1336 : *
1337 : * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1338 : *
1339 : * This knob will not override the system default sched_util_clamp_min defined
1340 : * above.
1341 : */
1342 : static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1343 :
1344 : /* All clamps are required to be less or equal than these values */
1345 : static struct uclamp_se uclamp_default[UCLAMP_CNT];
1346 :
1347 : /*
1348 : * This static key is used to reduce the uclamp overhead in the fast path. It
1349 : * primarily disables the call to uclamp_rq_{inc, dec}() in
1350 : * enqueue/dequeue_task().
1351 : *
1352 : * This allows users to continue to enable uclamp in their kernel config with
1353 : * minimum uclamp overhead in the fast path.
1354 : *
1355 : * As soon as userspace modifies any of the uclamp knobs, the static key is
1356 : * enabled, since we have an actual users that make use of uclamp
1357 : * functionality.
1358 : *
1359 : * The knobs that would enable this static key are:
1360 : *
1361 : * * A task modifying its uclamp value with sched_setattr().
1362 : * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1363 : * * An admin modifying the cgroup cpu.uclamp.{min, max}
1364 : */
1365 : DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1366 :
1367 : /* Integer rounded range for each bucket */
1368 : #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1369 :
1370 : #define for_each_clamp_id(clamp_id) \
1371 : for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1372 :
1373 : static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1374 : {
1375 : return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1376 : }
1377 :
1378 : static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1379 : {
1380 : if (clamp_id == UCLAMP_MIN)
1381 : return 0;
1382 : return SCHED_CAPACITY_SCALE;
1383 : }
1384 :
1385 : static inline void uclamp_se_set(struct uclamp_se *uc_se,
1386 : unsigned int value, bool user_defined)
1387 : {
1388 : uc_se->value = value;
1389 : uc_se->bucket_id = uclamp_bucket_id(value);
1390 : uc_se->user_defined = user_defined;
1391 : }
1392 :
1393 : static inline unsigned int
1394 : uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1395 : unsigned int clamp_value)
1396 : {
1397 : /*
1398 : * Avoid blocked utilization pushing up the frequency when we go
1399 : * idle (which drops the max-clamp) by retaining the last known
1400 : * max-clamp.
1401 : */
1402 : if (clamp_id == UCLAMP_MAX) {
1403 : rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1404 : return clamp_value;
1405 : }
1406 :
1407 : return uclamp_none(UCLAMP_MIN);
1408 : }
1409 :
1410 : static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1411 : unsigned int clamp_value)
1412 : {
1413 : /* Reset max-clamp retention only on idle exit */
1414 : if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1415 : return;
1416 :
1417 : uclamp_rq_set(rq, clamp_id, clamp_value);
1418 : }
1419 :
1420 : static inline
1421 : unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1422 : unsigned int clamp_value)
1423 : {
1424 : struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1425 : int bucket_id = UCLAMP_BUCKETS - 1;
1426 :
1427 : /*
1428 : * Since both min and max clamps are max aggregated, find the
1429 : * top most bucket with tasks in.
1430 : */
1431 : for ( ; bucket_id >= 0; bucket_id--) {
1432 : if (!bucket[bucket_id].tasks)
1433 : continue;
1434 : return bucket[bucket_id].value;
1435 : }
1436 :
1437 : /* No tasks -- default clamp values */
1438 : return uclamp_idle_value(rq, clamp_id, clamp_value);
1439 : }
1440 :
1441 : static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1442 : {
1443 : unsigned int default_util_min;
1444 : struct uclamp_se *uc_se;
1445 :
1446 : lockdep_assert_held(&p->pi_lock);
1447 :
1448 : uc_se = &p->uclamp_req[UCLAMP_MIN];
1449 :
1450 : /* Only sync if user didn't override the default */
1451 : if (uc_se->user_defined)
1452 : return;
1453 :
1454 : default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1455 : uclamp_se_set(uc_se, default_util_min, false);
1456 : }
1457 :
1458 : static void uclamp_update_util_min_rt_default(struct task_struct *p)
1459 : {
1460 : struct rq_flags rf;
1461 : struct rq *rq;
1462 :
1463 : if (!rt_task(p))
1464 : return;
1465 :
1466 : /* Protect updates to p->uclamp_* */
1467 : rq = task_rq_lock(p, &rf);
1468 : __uclamp_update_util_min_rt_default(p);
1469 : task_rq_unlock(rq, p, &rf);
1470 : }
1471 :
1472 : static inline struct uclamp_se
1473 : uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1474 : {
1475 : /* Copy by value as we could modify it */
1476 : struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1477 : #ifdef CONFIG_UCLAMP_TASK_GROUP
1478 : unsigned int tg_min, tg_max, value;
1479 :
1480 : /*
1481 : * Tasks in autogroups or root task group will be
1482 : * restricted by system defaults.
1483 : */
1484 : if (task_group_is_autogroup(task_group(p)))
1485 : return uc_req;
1486 : if (task_group(p) == &root_task_group)
1487 : return uc_req;
1488 :
1489 : tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1490 : tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1491 : value = uc_req.value;
1492 : value = clamp(value, tg_min, tg_max);
1493 : uclamp_se_set(&uc_req, value, false);
1494 : #endif
1495 :
1496 : return uc_req;
1497 : }
1498 :
1499 : /*
1500 : * The effective clamp bucket index of a task depends on, by increasing
1501 : * priority:
1502 : * - the task specific clamp value, when explicitly requested from userspace
1503 : * - the task group effective clamp value, for tasks not either in the root
1504 : * group or in an autogroup
1505 : * - the system default clamp value, defined by the sysadmin
1506 : */
1507 : static inline struct uclamp_se
1508 : uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1509 : {
1510 : struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1511 : struct uclamp_se uc_max = uclamp_default[clamp_id];
1512 :
1513 : /* System default restrictions always apply */
1514 : if (unlikely(uc_req.value > uc_max.value))
1515 : return uc_max;
1516 :
1517 : return uc_req;
1518 : }
1519 :
1520 : unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1521 : {
1522 : struct uclamp_se uc_eff;
1523 :
1524 : /* Task currently refcounted: use back-annotated (effective) value */
1525 : if (p->uclamp[clamp_id].active)
1526 : return (unsigned long)p->uclamp[clamp_id].value;
1527 :
1528 : uc_eff = uclamp_eff_get(p, clamp_id);
1529 :
1530 : return (unsigned long)uc_eff.value;
1531 : }
1532 :
1533 : /*
1534 : * When a task is enqueued on a rq, the clamp bucket currently defined by the
1535 : * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1536 : * updates the rq's clamp value if required.
1537 : *
1538 : * Tasks can have a task-specific value requested from user-space, track
1539 : * within each bucket the maximum value for tasks refcounted in it.
1540 : * This "local max aggregation" allows to track the exact "requested" value
1541 : * for each bucket when all its RUNNABLE tasks require the same clamp.
1542 : */
1543 : static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1544 : enum uclamp_id clamp_id)
1545 : {
1546 : struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1547 : struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1548 : struct uclamp_bucket *bucket;
1549 :
1550 : lockdep_assert_rq_held(rq);
1551 :
1552 : /* Update task effective clamp */
1553 : p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1554 :
1555 : bucket = &uc_rq->bucket[uc_se->bucket_id];
1556 : bucket->tasks++;
1557 : uc_se->active = true;
1558 :
1559 : uclamp_idle_reset(rq, clamp_id, uc_se->value);
1560 :
1561 : /*
1562 : * Local max aggregation: rq buckets always track the max
1563 : * "requested" clamp value of its RUNNABLE tasks.
1564 : */
1565 : if (bucket->tasks == 1 || uc_se->value > bucket->value)
1566 : bucket->value = uc_se->value;
1567 :
1568 : if (uc_se->value > uclamp_rq_get(rq, clamp_id))
1569 : uclamp_rq_set(rq, clamp_id, uc_se->value);
1570 : }
1571 :
1572 : /*
1573 : * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1574 : * is released. If this is the last task reference counting the rq's max
1575 : * active clamp value, then the rq's clamp value is updated.
1576 : *
1577 : * Both refcounted tasks and rq's cached clamp values are expected to be
1578 : * always valid. If it's detected they are not, as defensive programming,
1579 : * enforce the expected state and warn.
1580 : */
1581 : static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1582 : enum uclamp_id clamp_id)
1583 : {
1584 : struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1585 : struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1586 : struct uclamp_bucket *bucket;
1587 : unsigned int bkt_clamp;
1588 : unsigned int rq_clamp;
1589 :
1590 : lockdep_assert_rq_held(rq);
1591 :
1592 : /*
1593 : * If sched_uclamp_used was enabled after task @p was enqueued,
1594 : * we could end up with unbalanced call to uclamp_rq_dec_id().
1595 : *
1596 : * In this case the uc_se->active flag should be false since no uclamp
1597 : * accounting was performed at enqueue time and we can just return
1598 : * here.
1599 : *
1600 : * Need to be careful of the following enqueue/dequeue ordering
1601 : * problem too
1602 : *
1603 : * enqueue(taskA)
1604 : * // sched_uclamp_used gets enabled
1605 : * enqueue(taskB)
1606 : * dequeue(taskA)
1607 : * // Must not decrement bucket->tasks here
1608 : * dequeue(taskB)
1609 : *
1610 : * where we could end up with stale data in uc_se and
1611 : * bucket[uc_se->bucket_id].
1612 : *
1613 : * The following check here eliminates the possibility of such race.
1614 : */
1615 : if (unlikely(!uc_se->active))
1616 : return;
1617 :
1618 : bucket = &uc_rq->bucket[uc_se->bucket_id];
1619 :
1620 : SCHED_WARN_ON(!bucket->tasks);
1621 : if (likely(bucket->tasks))
1622 : bucket->tasks--;
1623 :
1624 : uc_se->active = false;
1625 :
1626 : /*
1627 : * Keep "local max aggregation" simple and accept to (possibly)
1628 : * overboost some RUNNABLE tasks in the same bucket.
1629 : * The rq clamp bucket value is reset to its base value whenever
1630 : * there are no more RUNNABLE tasks refcounting it.
1631 : */
1632 : if (likely(bucket->tasks))
1633 : return;
1634 :
1635 : rq_clamp = uclamp_rq_get(rq, clamp_id);
1636 : /*
1637 : * Defensive programming: this should never happen. If it happens,
1638 : * e.g. due to future modification, warn and fixup the expected value.
1639 : */
1640 : SCHED_WARN_ON(bucket->value > rq_clamp);
1641 : if (bucket->value >= rq_clamp) {
1642 : bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1643 : uclamp_rq_set(rq, clamp_id, bkt_clamp);
1644 : }
1645 : }
1646 :
1647 : static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1648 : {
1649 : enum uclamp_id clamp_id;
1650 :
1651 : /*
1652 : * Avoid any overhead until uclamp is actually used by the userspace.
1653 : *
1654 : * The condition is constructed such that a NOP is generated when
1655 : * sched_uclamp_used is disabled.
1656 : */
1657 : if (!static_branch_unlikely(&sched_uclamp_used))
1658 : return;
1659 :
1660 : if (unlikely(!p->sched_class->uclamp_enabled))
1661 : return;
1662 :
1663 : for_each_clamp_id(clamp_id)
1664 : uclamp_rq_inc_id(rq, p, clamp_id);
1665 :
1666 : /* Reset clamp idle holding when there is one RUNNABLE task */
1667 : if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1668 : rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1669 : }
1670 :
1671 : static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1672 : {
1673 : enum uclamp_id clamp_id;
1674 :
1675 : /*
1676 : * Avoid any overhead until uclamp is actually used by the userspace.
1677 : *
1678 : * The condition is constructed such that a NOP is generated when
1679 : * sched_uclamp_used is disabled.
1680 : */
1681 : if (!static_branch_unlikely(&sched_uclamp_used))
1682 : return;
1683 :
1684 : if (unlikely(!p->sched_class->uclamp_enabled))
1685 : return;
1686 :
1687 : for_each_clamp_id(clamp_id)
1688 : uclamp_rq_dec_id(rq, p, clamp_id);
1689 : }
1690 :
1691 : static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1692 : enum uclamp_id clamp_id)
1693 : {
1694 : if (!p->uclamp[clamp_id].active)
1695 : return;
1696 :
1697 : uclamp_rq_dec_id(rq, p, clamp_id);
1698 : uclamp_rq_inc_id(rq, p, clamp_id);
1699 :
1700 : /*
1701 : * Make sure to clear the idle flag if we've transiently reached 0
1702 : * active tasks on rq.
1703 : */
1704 : if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1705 : rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1706 : }
1707 :
1708 : static inline void
1709 : uclamp_update_active(struct task_struct *p)
1710 : {
1711 : enum uclamp_id clamp_id;
1712 : struct rq_flags rf;
1713 : struct rq *rq;
1714 :
1715 : /*
1716 : * Lock the task and the rq where the task is (or was) queued.
1717 : *
1718 : * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1719 : * price to pay to safely serialize util_{min,max} updates with
1720 : * enqueues, dequeues and migration operations.
1721 : * This is the same locking schema used by __set_cpus_allowed_ptr().
1722 : */
1723 : rq = task_rq_lock(p, &rf);
1724 :
1725 : /*
1726 : * Setting the clamp bucket is serialized by task_rq_lock().
1727 : * If the task is not yet RUNNABLE and its task_struct is not
1728 : * affecting a valid clamp bucket, the next time it's enqueued,
1729 : * it will already see the updated clamp bucket value.
1730 : */
1731 : for_each_clamp_id(clamp_id)
1732 : uclamp_rq_reinc_id(rq, p, clamp_id);
1733 :
1734 : task_rq_unlock(rq, p, &rf);
1735 : }
1736 :
1737 : #ifdef CONFIG_UCLAMP_TASK_GROUP
1738 : static inline void
1739 : uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1740 : {
1741 : struct css_task_iter it;
1742 : struct task_struct *p;
1743 :
1744 : css_task_iter_start(css, 0, &it);
1745 : while ((p = css_task_iter_next(&it)))
1746 : uclamp_update_active(p);
1747 : css_task_iter_end(&it);
1748 : }
1749 :
1750 : static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1751 : #endif
1752 :
1753 : #ifdef CONFIG_SYSCTL
1754 : #ifdef CONFIG_UCLAMP_TASK
1755 : #ifdef CONFIG_UCLAMP_TASK_GROUP
1756 : static void uclamp_update_root_tg(void)
1757 : {
1758 : struct task_group *tg = &root_task_group;
1759 :
1760 : uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1761 : sysctl_sched_uclamp_util_min, false);
1762 : uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1763 : sysctl_sched_uclamp_util_max, false);
1764 :
1765 : rcu_read_lock();
1766 : cpu_util_update_eff(&root_task_group.css);
1767 : rcu_read_unlock();
1768 : }
1769 : #else
1770 : static void uclamp_update_root_tg(void) { }
1771 : #endif
1772 :
1773 : static void uclamp_sync_util_min_rt_default(void)
1774 : {
1775 : struct task_struct *g, *p;
1776 :
1777 : /*
1778 : * copy_process() sysctl_uclamp
1779 : * uclamp_min_rt = X;
1780 : * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1781 : * // link thread smp_mb__after_spinlock()
1782 : * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1783 : * sched_post_fork() for_each_process_thread()
1784 : * __uclamp_sync_rt() __uclamp_sync_rt()
1785 : *
1786 : * Ensures that either sched_post_fork() will observe the new
1787 : * uclamp_min_rt or for_each_process_thread() will observe the new
1788 : * task.
1789 : */
1790 : read_lock(&tasklist_lock);
1791 : smp_mb__after_spinlock();
1792 : read_unlock(&tasklist_lock);
1793 :
1794 : rcu_read_lock();
1795 : for_each_process_thread(g, p)
1796 : uclamp_update_util_min_rt_default(p);
1797 : rcu_read_unlock();
1798 : }
1799 :
1800 : static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1801 : void *buffer, size_t *lenp, loff_t *ppos)
1802 : {
1803 : bool update_root_tg = false;
1804 : int old_min, old_max, old_min_rt;
1805 : int result;
1806 :
1807 : mutex_lock(&uclamp_mutex);
1808 : old_min = sysctl_sched_uclamp_util_min;
1809 : old_max = sysctl_sched_uclamp_util_max;
1810 : old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1811 :
1812 : result = proc_dointvec(table, write, buffer, lenp, ppos);
1813 : if (result)
1814 : goto undo;
1815 : if (!write)
1816 : goto done;
1817 :
1818 : if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1819 : sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1820 : sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1821 :
1822 : result = -EINVAL;
1823 : goto undo;
1824 : }
1825 :
1826 : if (old_min != sysctl_sched_uclamp_util_min) {
1827 : uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1828 : sysctl_sched_uclamp_util_min, false);
1829 : update_root_tg = true;
1830 : }
1831 : if (old_max != sysctl_sched_uclamp_util_max) {
1832 : uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1833 : sysctl_sched_uclamp_util_max, false);
1834 : update_root_tg = true;
1835 : }
1836 :
1837 : if (update_root_tg) {
1838 : static_branch_enable(&sched_uclamp_used);
1839 : uclamp_update_root_tg();
1840 : }
1841 :
1842 : if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1843 : static_branch_enable(&sched_uclamp_used);
1844 : uclamp_sync_util_min_rt_default();
1845 : }
1846 :
1847 : /*
1848 : * We update all RUNNABLE tasks only when task groups are in use.
1849 : * Otherwise, keep it simple and do just a lazy update at each next
1850 : * task enqueue time.
1851 : */
1852 :
1853 : goto done;
1854 :
1855 : undo:
1856 : sysctl_sched_uclamp_util_min = old_min;
1857 : sysctl_sched_uclamp_util_max = old_max;
1858 : sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1859 : done:
1860 : mutex_unlock(&uclamp_mutex);
1861 :
1862 : return result;
1863 : }
1864 : #endif
1865 : #endif
1866 :
1867 : static int uclamp_validate(struct task_struct *p,
1868 : const struct sched_attr *attr)
1869 : {
1870 : int util_min = p->uclamp_req[UCLAMP_MIN].value;
1871 : int util_max = p->uclamp_req[UCLAMP_MAX].value;
1872 :
1873 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1874 : util_min = attr->sched_util_min;
1875 :
1876 : if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1877 : return -EINVAL;
1878 : }
1879 :
1880 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1881 : util_max = attr->sched_util_max;
1882 :
1883 : if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1884 : return -EINVAL;
1885 : }
1886 :
1887 : if (util_min != -1 && util_max != -1 && util_min > util_max)
1888 : return -EINVAL;
1889 :
1890 : /*
1891 : * We have valid uclamp attributes; make sure uclamp is enabled.
1892 : *
1893 : * We need to do that here, because enabling static branches is a
1894 : * blocking operation which obviously cannot be done while holding
1895 : * scheduler locks.
1896 : */
1897 : static_branch_enable(&sched_uclamp_used);
1898 :
1899 : return 0;
1900 : }
1901 :
1902 : static bool uclamp_reset(const struct sched_attr *attr,
1903 : enum uclamp_id clamp_id,
1904 : struct uclamp_se *uc_se)
1905 : {
1906 : /* Reset on sched class change for a non user-defined clamp value. */
1907 : if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1908 : !uc_se->user_defined)
1909 : return true;
1910 :
1911 : /* Reset on sched_util_{min,max} == -1. */
1912 : if (clamp_id == UCLAMP_MIN &&
1913 : attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1914 : attr->sched_util_min == -1) {
1915 : return true;
1916 : }
1917 :
1918 : if (clamp_id == UCLAMP_MAX &&
1919 : attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1920 : attr->sched_util_max == -1) {
1921 : return true;
1922 : }
1923 :
1924 : return false;
1925 : }
1926 :
1927 : static void __setscheduler_uclamp(struct task_struct *p,
1928 : const struct sched_attr *attr)
1929 : {
1930 : enum uclamp_id clamp_id;
1931 :
1932 : for_each_clamp_id(clamp_id) {
1933 : struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1934 : unsigned int value;
1935 :
1936 : if (!uclamp_reset(attr, clamp_id, uc_se))
1937 : continue;
1938 :
1939 : /*
1940 : * RT by default have a 100% boost value that could be modified
1941 : * at runtime.
1942 : */
1943 : if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1944 : value = sysctl_sched_uclamp_util_min_rt_default;
1945 : else
1946 : value = uclamp_none(clamp_id);
1947 :
1948 : uclamp_se_set(uc_se, value, false);
1949 :
1950 : }
1951 :
1952 : if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1953 : return;
1954 :
1955 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1956 : attr->sched_util_min != -1) {
1957 : uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1958 : attr->sched_util_min, true);
1959 : }
1960 :
1961 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1962 : attr->sched_util_max != -1) {
1963 : uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1964 : attr->sched_util_max, true);
1965 : }
1966 : }
1967 :
1968 : static void uclamp_fork(struct task_struct *p)
1969 : {
1970 : enum uclamp_id clamp_id;
1971 :
1972 : /*
1973 : * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1974 : * as the task is still at its early fork stages.
1975 : */
1976 : for_each_clamp_id(clamp_id)
1977 : p->uclamp[clamp_id].active = false;
1978 :
1979 : if (likely(!p->sched_reset_on_fork))
1980 : return;
1981 :
1982 : for_each_clamp_id(clamp_id) {
1983 : uclamp_se_set(&p->uclamp_req[clamp_id],
1984 : uclamp_none(clamp_id), false);
1985 : }
1986 : }
1987 :
1988 : static void uclamp_post_fork(struct task_struct *p)
1989 : {
1990 : uclamp_update_util_min_rt_default(p);
1991 : }
1992 :
1993 : static void __init init_uclamp_rq(struct rq *rq)
1994 : {
1995 : enum uclamp_id clamp_id;
1996 : struct uclamp_rq *uc_rq = rq->uclamp;
1997 :
1998 : for_each_clamp_id(clamp_id) {
1999 : uc_rq[clamp_id] = (struct uclamp_rq) {
2000 : .value = uclamp_none(clamp_id)
2001 : };
2002 : }
2003 :
2004 : rq->uclamp_flags = UCLAMP_FLAG_IDLE;
2005 : }
2006 :
2007 : static void __init init_uclamp(void)
2008 : {
2009 : struct uclamp_se uc_max = {};
2010 : enum uclamp_id clamp_id;
2011 : int cpu;
2012 :
2013 : for_each_possible_cpu(cpu)
2014 : init_uclamp_rq(cpu_rq(cpu));
2015 :
2016 : for_each_clamp_id(clamp_id) {
2017 : uclamp_se_set(&init_task.uclamp_req[clamp_id],
2018 : uclamp_none(clamp_id), false);
2019 : }
2020 :
2021 : /* System defaults allow max clamp values for both indexes */
2022 : uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2023 : for_each_clamp_id(clamp_id) {
2024 : uclamp_default[clamp_id] = uc_max;
2025 : #ifdef CONFIG_UCLAMP_TASK_GROUP
2026 : root_task_group.uclamp_req[clamp_id] = uc_max;
2027 : root_task_group.uclamp[clamp_id] = uc_max;
2028 : #endif
2029 : }
2030 : }
2031 :
2032 : #else /* CONFIG_UCLAMP_TASK */
2033 : static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2034 : static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2035 : static inline int uclamp_validate(struct task_struct *p,
2036 : const struct sched_attr *attr)
2037 : {
2038 : return -EOPNOTSUPP;
2039 : }
2040 : static void __setscheduler_uclamp(struct task_struct *p,
2041 : const struct sched_attr *attr) { }
2042 : static inline void uclamp_fork(struct task_struct *p) { }
2043 : static inline void uclamp_post_fork(struct task_struct *p) { }
2044 : static inline void init_uclamp(void) { }
2045 : #endif /* CONFIG_UCLAMP_TASK */
2046 :
2047 0 : bool sched_task_on_rq(struct task_struct *p)
2048 : {
2049 0 : return task_on_rq_queued(p);
2050 : }
2051 :
2052 0 : unsigned long get_wchan(struct task_struct *p)
2053 : {
2054 0 : unsigned long ip = 0;
2055 : unsigned int state;
2056 :
2057 0 : if (!p || p == current)
2058 : return 0;
2059 :
2060 : /* Only get wchan if task is blocked and we can keep it that way. */
2061 0 : raw_spin_lock_irq(&p->pi_lock);
2062 0 : state = READ_ONCE(p->__state);
2063 0 : smp_rmb(); /* see try_to_wake_up() */
2064 0 : if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2065 0 : ip = __get_wchan(p);
2066 0 : raw_spin_unlock_irq(&p->pi_lock);
2067 :
2068 0 : return ip;
2069 : }
2070 :
2071 2442 : static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2072 : {
2073 2442 : if (!(flags & ENQUEUE_NOCLOCK))
2074 0 : update_rq_clock(rq);
2075 :
2076 : if (!(flags & ENQUEUE_RESTORE)) {
2077 : sched_info_enqueue(rq, p);
2078 : psi_enqueue(p, (flags & ENQUEUE_WAKEUP) && !(flags & ENQUEUE_MIGRATED));
2079 : }
2080 :
2081 2446 : uclamp_rq_inc(rq, p);
2082 2446 : p->sched_class->enqueue_task(rq, p, flags);
2083 :
2084 2446 : if (sched_core_enabled(rq))
2085 : sched_core_enqueue(rq, p);
2086 2442 : }
2087 :
2088 2440 : static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2089 : {
2090 2444 : if (sched_core_enabled(rq))
2091 : sched_core_dequeue(rq, p, flags);
2092 :
2093 2440 : if (!(flags & DEQUEUE_NOCLOCK))
2094 0 : update_rq_clock(rq);
2095 :
2096 : if (!(flags & DEQUEUE_SAVE)) {
2097 : sched_info_dequeue(rq, p);
2098 : psi_dequeue(p, flags & DEQUEUE_SLEEP);
2099 : }
2100 :
2101 2444 : uclamp_rq_dec(rq, p);
2102 2444 : p->sched_class->dequeue_task(rq, p, flags);
2103 2440 : }
2104 :
2105 0 : void activate_task(struct rq *rq, struct task_struct *p, int flags)
2106 : {
2107 2442 : if (task_on_rq_migrating(p))
2108 : flags |= ENQUEUE_MIGRATED;
2109 : if (flags & ENQUEUE_MIGRATED)
2110 : sched_mm_cid_migrate_to(rq, p);
2111 :
2112 2442 : enqueue_task(rq, p, flags);
2113 :
2114 2442 : p->on_rq = TASK_ON_RQ_QUEUED;
2115 0 : }
2116 :
2117 0 : void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2118 : {
2119 2440 : p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2120 :
2121 2440 : dequeue_task(rq, p, flags);
2122 0 : }
2123 :
2124 : static inline int __normal_prio(int policy, int rt_prio, int nice)
2125 : {
2126 : int prio;
2127 :
2128 5 : if (dl_policy(policy))
2129 : prio = MAX_DL_PRIO - 1;
2130 5 : else if (rt_policy(policy))
2131 0 : prio = MAX_RT_PRIO - 1 - rt_prio;
2132 : else
2133 0 : prio = NICE_TO_PRIO(nice);
2134 :
2135 : return prio;
2136 : }
2137 :
2138 : /*
2139 : * Calculate the expected normal priority: i.e. priority
2140 : * without taking RT-inheritance into account. Might be
2141 : * boosted by interactivity modifiers. Changes upon fork,
2142 : * setprio syscalls, and whenever the interactivity
2143 : * estimator recalculates.
2144 : */
2145 : static inline int normal_prio(struct task_struct *p)
2146 : {
2147 10 : return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2148 : }
2149 :
2150 : /*
2151 : * Calculate the current priority, i.e. the priority
2152 : * taken into account by the scheduler. This value might
2153 : * be boosted by RT tasks, or might be boosted by
2154 : * interactivity modifiers. Will be RT if the task got
2155 : * RT-boosted. If not then it returns p->normal_prio.
2156 : */
2157 : static int effective_prio(struct task_struct *p)
2158 : {
2159 10 : p->normal_prio = normal_prio(p);
2160 : /*
2161 : * If we are RT tasks or we were boosted to RT priority,
2162 : * keep the priority unchanged. Otherwise, update priority
2163 : * to the normal priority:
2164 : */
2165 10 : if (!rt_prio(p->prio))
2166 : return p->normal_prio;
2167 : return p->prio;
2168 : }
2169 :
2170 : /**
2171 : * task_curr - is this task currently executing on a CPU?
2172 : * @p: the task in question.
2173 : *
2174 : * Return: 1 if the task is currently executing. 0 otherwise.
2175 : */
2176 0 : inline int task_curr(const struct task_struct *p)
2177 : {
2178 0 : return cpu_curr(task_cpu(p)) == p;
2179 : }
2180 :
2181 : /*
2182 : * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2183 : * use the balance_callback list if you want balancing.
2184 : *
2185 : * this means any call to check_class_changed() must be followed by a call to
2186 : * balance_callback().
2187 : */
2188 0 : static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2189 : const struct sched_class *prev_class,
2190 : int oldprio)
2191 : {
2192 0 : if (prev_class != p->sched_class) {
2193 0 : if (prev_class->switched_from)
2194 0 : prev_class->switched_from(rq, p);
2195 :
2196 0 : p->sched_class->switched_to(rq, p);
2197 0 : } else if (oldprio != p->prio || dl_task(p))
2198 0 : p->sched_class->prio_changed(rq, p, oldprio);
2199 0 : }
2200 :
2201 2442 : void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2202 : {
2203 2442 : if (p->sched_class == rq->curr->sched_class)
2204 2439 : rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2205 3 : else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2206 3 : resched_curr(rq);
2207 :
2208 : /*
2209 : * A queue event has occurred, and we're going to schedule. In
2210 : * this case, we can save a useless back to back clock update.
2211 : */
2212 4884 : if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2213 1252 : rq_clock_skip_update(rq);
2214 2442 : }
2215 :
2216 : #ifdef CONFIG_SMP
2217 :
2218 : static void
2219 : __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
2220 :
2221 : static int __set_cpus_allowed_ptr(struct task_struct *p,
2222 : struct affinity_context *ctx);
2223 :
2224 : static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2225 : {
2226 : struct affinity_context ac = {
2227 : .new_mask = cpumask_of(rq->cpu),
2228 : .flags = SCA_MIGRATE_DISABLE,
2229 : };
2230 :
2231 : if (likely(!p->migration_disabled))
2232 : return;
2233 :
2234 : if (p->cpus_ptr != &p->cpus_mask)
2235 : return;
2236 :
2237 : /*
2238 : * Violates locking rules! see comment in __do_set_cpus_allowed().
2239 : */
2240 : __do_set_cpus_allowed(p, &ac);
2241 : }
2242 :
2243 : void migrate_disable(void)
2244 : {
2245 : struct task_struct *p = current;
2246 :
2247 : if (p->migration_disabled) {
2248 : p->migration_disabled++;
2249 : return;
2250 : }
2251 :
2252 : preempt_disable();
2253 : this_rq()->nr_pinned++;
2254 : p->migration_disabled = 1;
2255 : preempt_enable();
2256 : }
2257 : EXPORT_SYMBOL_GPL(migrate_disable);
2258 :
2259 : void migrate_enable(void)
2260 : {
2261 : struct task_struct *p = current;
2262 : struct affinity_context ac = {
2263 : .new_mask = &p->cpus_mask,
2264 : .flags = SCA_MIGRATE_ENABLE,
2265 : };
2266 :
2267 : if (p->migration_disabled > 1) {
2268 : p->migration_disabled--;
2269 : return;
2270 : }
2271 :
2272 : if (WARN_ON_ONCE(!p->migration_disabled))
2273 : return;
2274 :
2275 : /*
2276 : * Ensure stop_task runs either before or after this, and that
2277 : * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2278 : */
2279 : preempt_disable();
2280 : if (p->cpus_ptr != &p->cpus_mask)
2281 : __set_cpus_allowed_ptr(p, &ac);
2282 : /*
2283 : * Mustn't clear migration_disabled() until cpus_ptr points back at the
2284 : * regular cpus_mask, otherwise things that race (eg.
2285 : * select_fallback_rq) get confused.
2286 : */
2287 : barrier();
2288 : p->migration_disabled = 0;
2289 : this_rq()->nr_pinned--;
2290 : preempt_enable();
2291 : }
2292 : EXPORT_SYMBOL_GPL(migrate_enable);
2293 :
2294 : static inline bool rq_has_pinned_tasks(struct rq *rq)
2295 : {
2296 : return rq->nr_pinned;
2297 : }
2298 :
2299 : /*
2300 : * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2301 : * __set_cpus_allowed_ptr() and select_fallback_rq().
2302 : */
2303 : static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2304 : {
2305 : /* When not in the task's cpumask, no point in looking further. */
2306 : if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2307 : return false;
2308 :
2309 : /* migrate_disabled() must be allowed to finish. */
2310 : if (is_migration_disabled(p))
2311 : return cpu_online(cpu);
2312 :
2313 : /* Non kernel threads are not allowed during either online or offline. */
2314 : if (!(p->flags & PF_KTHREAD))
2315 : return cpu_active(cpu) && task_cpu_possible(cpu, p);
2316 :
2317 : /* KTHREAD_IS_PER_CPU is always allowed. */
2318 : if (kthread_is_per_cpu(p))
2319 : return cpu_online(cpu);
2320 :
2321 : /* Regular kernel threads don't get to stay during offline. */
2322 : if (cpu_dying(cpu))
2323 : return false;
2324 :
2325 : /* But are allowed during online. */
2326 : return cpu_online(cpu);
2327 : }
2328 :
2329 : /*
2330 : * This is how migration works:
2331 : *
2332 : * 1) we invoke migration_cpu_stop() on the target CPU using
2333 : * stop_one_cpu().
2334 : * 2) stopper starts to run (implicitly forcing the migrated thread
2335 : * off the CPU)
2336 : * 3) it checks whether the migrated task is still in the wrong runqueue.
2337 : * 4) if it's in the wrong runqueue then the migration thread removes
2338 : * it and puts it into the right queue.
2339 : * 5) stopper completes and stop_one_cpu() returns and the migration
2340 : * is done.
2341 : */
2342 :
2343 : /*
2344 : * move_queued_task - move a queued task to new rq.
2345 : *
2346 : * Returns (locked) new rq. Old rq's lock is released.
2347 : */
2348 : static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2349 : struct task_struct *p, int new_cpu)
2350 : {
2351 : lockdep_assert_rq_held(rq);
2352 :
2353 : deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2354 : set_task_cpu(p, new_cpu);
2355 : rq_unlock(rq, rf);
2356 :
2357 : rq = cpu_rq(new_cpu);
2358 :
2359 : rq_lock(rq, rf);
2360 : WARN_ON_ONCE(task_cpu(p) != new_cpu);
2361 : activate_task(rq, p, 0);
2362 : check_preempt_curr(rq, p, 0);
2363 :
2364 : return rq;
2365 : }
2366 :
2367 : struct migration_arg {
2368 : struct task_struct *task;
2369 : int dest_cpu;
2370 : struct set_affinity_pending *pending;
2371 : };
2372 :
2373 : /*
2374 : * @refs: number of wait_for_completion()
2375 : * @stop_pending: is @stop_work in use
2376 : */
2377 : struct set_affinity_pending {
2378 : refcount_t refs;
2379 : unsigned int stop_pending;
2380 : struct completion done;
2381 : struct cpu_stop_work stop_work;
2382 : struct migration_arg arg;
2383 : };
2384 :
2385 : /*
2386 : * Move (not current) task off this CPU, onto the destination CPU. We're doing
2387 : * this because either it can't run here any more (set_cpus_allowed()
2388 : * away from this CPU, or CPU going down), or because we're
2389 : * attempting to rebalance this task on exec (sched_exec).
2390 : *
2391 : * So we race with normal scheduler movements, but that's OK, as long
2392 : * as the task is no longer on this CPU.
2393 : */
2394 : static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2395 : struct task_struct *p, int dest_cpu)
2396 : {
2397 : /* Affinity changed (again). */
2398 : if (!is_cpu_allowed(p, dest_cpu))
2399 : return rq;
2400 :
2401 : update_rq_clock(rq);
2402 : rq = move_queued_task(rq, rf, p, dest_cpu);
2403 :
2404 : return rq;
2405 : }
2406 :
2407 : /*
2408 : * migration_cpu_stop - this will be executed by a highprio stopper thread
2409 : * and performs thread migration by bumping thread off CPU then
2410 : * 'pushing' onto another runqueue.
2411 : */
2412 : static int migration_cpu_stop(void *data)
2413 : {
2414 : struct migration_arg *arg = data;
2415 : struct set_affinity_pending *pending = arg->pending;
2416 : struct task_struct *p = arg->task;
2417 : struct rq *rq = this_rq();
2418 : bool complete = false;
2419 : struct rq_flags rf;
2420 :
2421 : /*
2422 : * The original target CPU might have gone down and we might
2423 : * be on another CPU but it doesn't matter.
2424 : */
2425 : local_irq_save(rf.flags);
2426 : /*
2427 : * We need to explicitly wake pending tasks before running
2428 : * __migrate_task() such that we will not miss enforcing cpus_ptr
2429 : * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2430 : */
2431 : flush_smp_call_function_queue();
2432 :
2433 : raw_spin_lock(&p->pi_lock);
2434 : rq_lock(rq, &rf);
2435 :
2436 : /*
2437 : * If we were passed a pending, then ->stop_pending was set, thus
2438 : * p->migration_pending must have remained stable.
2439 : */
2440 : WARN_ON_ONCE(pending && pending != p->migration_pending);
2441 :
2442 : /*
2443 : * If task_rq(p) != rq, it cannot be migrated here, because we're
2444 : * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2445 : * we're holding p->pi_lock.
2446 : */
2447 : if (task_rq(p) == rq) {
2448 : if (is_migration_disabled(p))
2449 : goto out;
2450 :
2451 : if (pending) {
2452 : p->migration_pending = NULL;
2453 : complete = true;
2454 :
2455 : if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2456 : goto out;
2457 : }
2458 :
2459 : if (task_on_rq_queued(p))
2460 : rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2461 : else
2462 : p->wake_cpu = arg->dest_cpu;
2463 :
2464 : /*
2465 : * XXX __migrate_task() can fail, at which point we might end
2466 : * up running on a dodgy CPU, AFAICT this can only happen
2467 : * during CPU hotplug, at which point we'll get pushed out
2468 : * anyway, so it's probably not a big deal.
2469 : */
2470 :
2471 : } else if (pending) {
2472 : /*
2473 : * This happens when we get migrated between migrate_enable()'s
2474 : * preempt_enable() and scheduling the stopper task. At that
2475 : * point we're a regular task again and not current anymore.
2476 : *
2477 : * A !PREEMPT kernel has a giant hole here, which makes it far
2478 : * more likely.
2479 : */
2480 :
2481 : /*
2482 : * The task moved before the stopper got to run. We're holding
2483 : * ->pi_lock, so the allowed mask is stable - if it got
2484 : * somewhere allowed, we're done.
2485 : */
2486 : if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2487 : p->migration_pending = NULL;
2488 : complete = true;
2489 : goto out;
2490 : }
2491 :
2492 : /*
2493 : * When migrate_enable() hits a rq mis-match we can't reliably
2494 : * determine is_migration_disabled() and so have to chase after
2495 : * it.
2496 : */
2497 : WARN_ON_ONCE(!pending->stop_pending);
2498 : task_rq_unlock(rq, p, &rf);
2499 : stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2500 : &pending->arg, &pending->stop_work);
2501 : return 0;
2502 : }
2503 : out:
2504 : if (pending)
2505 : pending->stop_pending = false;
2506 : task_rq_unlock(rq, p, &rf);
2507 :
2508 : if (complete)
2509 : complete_all(&pending->done);
2510 :
2511 : return 0;
2512 : }
2513 :
2514 : int push_cpu_stop(void *arg)
2515 : {
2516 : struct rq *lowest_rq = NULL, *rq = this_rq();
2517 : struct task_struct *p = arg;
2518 :
2519 : raw_spin_lock_irq(&p->pi_lock);
2520 : raw_spin_rq_lock(rq);
2521 :
2522 : if (task_rq(p) != rq)
2523 : goto out_unlock;
2524 :
2525 : if (is_migration_disabled(p)) {
2526 : p->migration_flags |= MDF_PUSH;
2527 : goto out_unlock;
2528 : }
2529 :
2530 : p->migration_flags &= ~MDF_PUSH;
2531 :
2532 : if (p->sched_class->find_lock_rq)
2533 : lowest_rq = p->sched_class->find_lock_rq(p, rq);
2534 :
2535 : if (!lowest_rq)
2536 : goto out_unlock;
2537 :
2538 : // XXX validate p is still the highest prio task
2539 : if (task_rq(p) == rq) {
2540 : deactivate_task(rq, p, 0);
2541 : set_task_cpu(p, lowest_rq->cpu);
2542 : activate_task(lowest_rq, p, 0);
2543 : resched_curr(lowest_rq);
2544 : }
2545 :
2546 : double_unlock_balance(rq, lowest_rq);
2547 :
2548 : out_unlock:
2549 : rq->push_busy = false;
2550 : raw_spin_rq_unlock(rq);
2551 : raw_spin_unlock_irq(&p->pi_lock);
2552 :
2553 : put_task_struct(p);
2554 : return 0;
2555 : }
2556 :
2557 : /*
2558 : * sched_class::set_cpus_allowed must do the below, but is not required to
2559 : * actually call this function.
2560 : */
2561 : void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
2562 : {
2563 : if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2564 : p->cpus_ptr = ctx->new_mask;
2565 : return;
2566 : }
2567 :
2568 : cpumask_copy(&p->cpus_mask, ctx->new_mask);
2569 : p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
2570 :
2571 : /*
2572 : * Swap in a new user_cpus_ptr if SCA_USER flag set
2573 : */
2574 : if (ctx->flags & SCA_USER)
2575 : swap(p->user_cpus_ptr, ctx->user_mask);
2576 : }
2577 :
2578 : static void
2579 : __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
2580 : {
2581 : struct rq *rq = task_rq(p);
2582 : bool queued, running;
2583 :
2584 : /*
2585 : * This here violates the locking rules for affinity, since we're only
2586 : * supposed to change these variables while holding both rq->lock and
2587 : * p->pi_lock.
2588 : *
2589 : * HOWEVER, it magically works, because ttwu() is the only code that
2590 : * accesses these variables under p->pi_lock and only does so after
2591 : * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2592 : * before finish_task().
2593 : *
2594 : * XXX do further audits, this smells like something putrid.
2595 : */
2596 : if (ctx->flags & SCA_MIGRATE_DISABLE)
2597 : SCHED_WARN_ON(!p->on_cpu);
2598 : else
2599 : lockdep_assert_held(&p->pi_lock);
2600 :
2601 : queued = task_on_rq_queued(p);
2602 : running = task_current(rq, p);
2603 :
2604 : if (queued) {
2605 : /*
2606 : * Because __kthread_bind() calls this on blocked tasks without
2607 : * holding rq->lock.
2608 : */
2609 : lockdep_assert_rq_held(rq);
2610 : dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2611 : }
2612 : if (running)
2613 : put_prev_task(rq, p);
2614 :
2615 : p->sched_class->set_cpus_allowed(p, ctx);
2616 :
2617 : if (queued)
2618 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2619 : if (running)
2620 : set_next_task(rq, p);
2621 : }
2622 :
2623 : /*
2624 : * Used for kthread_bind() and select_fallback_rq(), in both cases the user
2625 : * affinity (if any) should be destroyed too.
2626 : */
2627 : void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2628 : {
2629 : struct affinity_context ac = {
2630 : .new_mask = new_mask,
2631 : .user_mask = NULL,
2632 : .flags = SCA_USER, /* clear the user requested mask */
2633 : };
2634 : union cpumask_rcuhead {
2635 : cpumask_t cpumask;
2636 : struct rcu_head rcu;
2637 : };
2638 :
2639 : __do_set_cpus_allowed(p, &ac);
2640 :
2641 : /*
2642 : * Because this is called with p->pi_lock held, it is not possible
2643 : * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2644 : * kfree_rcu().
2645 : */
2646 : kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
2647 : }
2648 :
2649 : static cpumask_t *alloc_user_cpus_ptr(int node)
2650 : {
2651 : /*
2652 : * See do_set_cpus_allowed() above for the rcu_head usage.
2653 : */
2654 : int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2655 :
2656 : return kmalloc_node(size, GFP_KERNEL, node);
2657 : }
2658 :
2659 : int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2660 : int node)
2661 : {
2662 : cpumask_t *user_mask;
2663 : unsigned long flags;
2664 :
2665 : /*
2666 : * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2667 : * may differ by now due to racing.
2668 : */
2669 : dst->user_cpus_ptr = NULL;
2670 :
2671 : /*
2672 : * This check is racy and losing the race is a valid situation.
2673 : * It is not worth the extra overhead of taking the pi_lock on
2674 : * every fork/clone.
2675 : */
2676 : if (data_race(!src->user_cpus_ptr))
2677 : return 0;
2678 :
2679 : user_mask = alloc_user_cpus_ptr(node);
2680 : if (!user_mask)
2681 : return -ENOMEM;
2682 :
2683 : /*
2684 : * Use pi_lock to protect content of user_cpus_ptr
2685 : *
2686 : * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2687 : * do_set_cpus_allowed().
2688 : */
2689 : raw_spin_lock_irqsave(&src->pi_lock, flags);
2690 : if (src->user_cpus_ptr) {
2691 : swap(dst->user_cpus_ptr, user_mask);
2692 : cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2693 : }
2694 : raw_spin_unlock_irqrestore(&src->pi_lock, flags);
2695 :
2696 : if (unlikely(user_mask))
2697 : kfree(user_mask);
2698 :
2699 : return 0;
2700 : }
2701 :
2702 : static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2703 : {
2704 : struct cpumask *user_mask = NULL;
2705 :
2706 : swap(p->user_cpus_ptr, user_mask);
2707 :
2708 : return user_mask;
2709 : }
2710 :
2711 : void release_user_cpus_ptr(struct task_struct *p)
2712 : {
2713 : kfree(clear_user_cpus_ptr(p));
2714 : }
2715 :
2716 : /*
2717 : * This function is wildly self concurrent; here be dragons.
2718 : *
2719 : *
2720 : * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2721 : * designated task is enqueued on an allowed CPU. If that task is currently
2722 : * running, we have to kick it out using the CPU stopper.
2723 : *
2724 : * Migrate-Disable comes along and tramples all over our nice sandcastle.
2725 : * Consider:
2726 : *
2727 : * Initial conditions: P0->cpus_mask = [0, 1]
2728 : *
2729 : * P0@CPU0 P1
2730 : *
2731 : * migrate_disable();
2732 : * <preempted>
2733 : * set_cpus_allowed_ptr(P0, [1]);
2734 : *
2735 : * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2736 : * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2737 : * This means we need the following scheme:
2738 : *
2739 : * P0@CPU0 P1
2740 : *
2741 : * migrate_disable();
2742 : * <preempted>
2743 : * set_cpus_allowed_ptr(P0, [1]);
2744 : * <blocks>
2745 : * <resumes>
2746 : * migrate_enable();
2747 : * __set_cpus_allowed_ptr();
2748 : * <wakes local stopper>
2749 : * `--> <woken on migration completion>
2750 : *
2751 : * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2752 : * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2753 : * task p are serialized by p->pi_lock, which we can leverage: the one that
2754 : * should come into effect at the end of the Migrate-Disable region is the last
2755 : * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2756 : * but we still need to properly signal those waiting tasks at the appropriate
2757 : * moment.
2758 : *
2759 : * This is implemented using struct set_affinity_pending. The first
2760 : * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2761 : * setup an instance of that struct and install it on the targeted task_struct.
2762 : * Any and all further callers will reuse that instance. Those then wait for
2763 : * a completion signaled at the tail of the CPU stopper callback (1), triggered
2764 : * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2765 : *
2766 : *
2767 : * (1) In the cases covered above. There is one more where the completion is
2768 : * signaled within affine_move_task() itself: when a subsequent affinity request
2769 : * occurs after the stopper bailed out due to the targeted task still being
2770 : * Migrate-Disable. Consider:
2771 : *
2772 : * Initial conditions: P0->cpus_mask = [0, 1]
2773 : *
2774 : * CPU0 P1 P2
2775 : * <P0>
2776 : * migrate_disable();
2777 : * <preempted>
2778 : * set_cpus_allowed_ptr(P0, [1]);
2779 : * <blocks>
2780 : * <migration/0>
2781 : * migration_cpu_stop()
2782 : * is_migration_disabled()
2783 : * <bails>
2784 : * set_cpus_allowed_ptr(P0, [0, 1]);
2785 : * <signal completion>
2786 : * <awakes>
2787 : *
2788 : * Note that the above is safe vs a concurrent migrate_enable(), as any
2789 : * pending affinity completion is preceded by an uninstallation of
2790 : * p->migration_pending done with p->pi_lock held.
2791 : */
2792 : static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2793 : int dest_cpu, unsigned int flags)
2794 : __releases(rq->lock)
2795 : __releases(p->pi_lock)
2796 : {
2797 : struct set_affinity_pending my_pending = { }, *pending = NULL;
2798 : bool stop_pending, complete = false;
2799 :
2800 : /* Can the task run on the task's current CPU? If so, we're done */
2801 : if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2802 : struct task_struct *push_task = NULL;
2803 :
2804 : if ((flags & SCA_MIGRATE_ENABLE) &&
2805 : (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2806 : rq->push_busy = true;
2807 : push_task = get_task_struct(p);
2808 : }
2809 :
2810 : /*
2811 : * If there are pending waiters, but no pending stop_work,
2812 : * then complete now.
2813 : */
2814 : pending = p->migration_pending;
2815 : if (pending && !pending->stop_pending) {
2816 : p->migration_pending = NULL;
2817 : complete = true;
2818 : }
2819 :
2820 : task_rq_unlock(rq, p, rf);
2821 :
2822 : if (push_task) {
2823 : stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2824 : p, &rq->push_work);
2825 : }
2826 :
2827 : if (complete)
2828 : complete_all(&pending->done);
2829 :
2830 : return 0;
2831 : }
2832 :
2833 : if (!(flags & SCA_MIGRATE_ENABLE)) {
2834 : /* serialized by p->pi_lock */
2835 : if (!p->migration_pending) {
2836 : /* Install the request */
2837 : refcount_set(&my_pending.refs, 1);
2838 : init_completion(&my_pending.done);
2839 : my_pending.arg = (struct migration_arg) {
2840 : .task = p,
2841 : .dest_cpu = dest_cpu,
2842 : .pending = &my_pending,
2843 : };
2844 :
2845 : p->migration_pending = &my_pending;
2846 : } else {
2847 : pending = p->migration_pending;
2848 : refcount_inc(&pending->refs);
2849 : /*
2850 : * Affinity has changed, but we've already installed a
2851 : * pending. migration_cpu_stop() *must* see this, else
2852 : * we risk a completion of the pending despite having a
2853 : * task on a disallowed CPU.
2854 : *
2855 : * Serialized by p->pi_lock, so this is safe.
2856 : */
2857 : pending->arg.dest_cpu = dest_cpu;
2858 : }
2859 : }
2860 : pending = p->migration_pending;
2861 : /*
2862 : * - !MIGRATE_ENABLE:
2863 : * we'll have installed a pending if there wasn't one already.
2864 : *
2865 : * - MIGRATE_ENABLE:
2866 : * we're here because the current CPU isn't matching anymore,
2867 : * the only way that can happen is because of a concurrent
2868 : * set_cpus_allowed_ptr() call, which should then still be
2869 : * pending completion.
2870 : *
2871 : * Either way, we really should have a @pending here.
2872 : */
2873 : if (WARN_ON_ONCE(!pending)) {
2874 : task_rq_unlock(rq, p, rf);
2875 : return -EINVAL;
2876 : }
2877 :
2878 : if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2879 : /*
2880 : * MIGRATE_ENABLE gets here because 'p == current', but for
2881 : * anything else we cannot do is_migration_disabled(), punt
2882 : * and have the stopper function handle it all race-free.
2883 : */
2884 : stop_pending = pending->stop_pending;
2885 : if (!stop_pending)
2886 : pending->stop_pending = true;
2887 :
2888 : if (flags & SCA_MIGRATE_ENABLE)
2889 : p->migration_flags &= ~MDF_PUSH;
2890 :
2891 : task_rq_unlock(rq, p, rf);
2892 :
2893 : if (!stop_pending) {
2894 : stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2895 : &pending->arg, &pending->stop_work);
2896 : }
2897 :
2898 : if (flags & SCA_MIGRATE_ENABLE)
2899 : return 0;
2900 : } else {
2901 :
2902 : if (!is_migration_disabled(p)) {
2903 : if (task_on_rq_queued(p))
2904 : rq = move_queued_task(rq, rf, p, dest_cpu);
2905 :
2906 : if (!pending->stop_pending) {
2907 : p->migration_pending = NULL;
2908 : complete = true;
2909 : }
2910 : }
2911 : task_rq_unlock(rq, p, rf);
2912 :
2913 : if (complete)
2914 : complete_all(&pending->done);
2915 : }
2916 :
2917 : wait_for_completion(&pending->done);
2918 :
2919 : if (refcount_dec_and_test(&pending->refs))
2920 : wake_up_var(&pending->refs); /* No UaF, just an address */
2921 :
2922 : /*
2923 : * Block the original owner of &pending until all subsequent callers
2924 : * have seen the completion and decremented the refcount
2925 : */
2926 : wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2927 :
2928 : /* ARGH */
2929 : WARN_ON_ONCE(my_pending.stop_pending);
2930 :
2931 : return 0;
2932 : }
2933 :
2934 : /*
2935 : * Called with both p->pi_lock and rq->lock held; drops both before returning.
2936 : */
2937 : static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2938 : struct affinity_context *ctx,
2939 : struct rq *rq,
2940 : struct rq_flags *rf)
2941 : __releases(rq->lock)
2942 : __releases(p->pi_lock)
2943 : {
2944 : const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2945 : const struct cpumask *cpu_valid_mask = cpu_active_mask;
2946 : bool kthread = p->flags & PF_KTHREAD;
2947 : unsigned int dest_cpu;
2948 : int ret = 0;
2949 :
2950 : update_rq_clock(rq);
2951 :
2952 : if (kthread || is_migration_disabled(p)) {
2953 : /*
2954 : * Kernel threads are allowed on online && !active CPUs,
2955 : * however, during cpu-hot-unplug, even these might get pushed
2956 : * away if not KTHREAD_IS_PER_CPU.
2957 : *
2958 : * Specifically, migration_disabled() tasks must not fail the
2959 : * cpumask_any_and_distribute() pick below, esp. so on
2960 : * SCA_MIGRATE_ENABLE, otherwise we'll not call
2961 : * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2962 : */
2963 : cpu_valid_mask = cpu_online_mask;
2964 : }
2965 :
2966 : if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
2967 : ret = -EINVAL;
2968 : goto out;
2969 : }
2970 :
2971 : /*
2972 : * Must re-check here, to close a race against __kthread_bind(),
2973 : * sched_setaffinity() is not guaranteed to observe the flag.
2974 : */
2975 : if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2976 : ret = -EINVAL;
2977 : goto out;
2978 : }
2979 :
2980 : if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
2981 : if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) {
2982 : if (ctx->flags & SCA_USER)
2983 : swap(p->user_cpus_ptr, ctx->user_mask);
2984 : goto out;
2985 : }
2986 :
2987 : if (WARN_ON_ONCE(p == current &&
2988 : is_migration_disabled(p) &&
2989 : !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
2990 : ret = -EBUSY;
2991 : goto out;
2992 : }
2993 : }
2994 :
2995 : /*
2996 : * Picking a ~random cpu helps in cases where we are changing affinity
2997 : * for groups of tasks (ie. cpuset), so that load balancing is not
2998 : * immediately required to distribute the tasks within their new mask.
2999 : */
3000 : dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
3001 : if (dest_cpu >= nr_cpu_ids) {
3002 : ret = -EINVAL;
3003 : goto out;
3004 : }
3005 :
3006 : __do_set_cpus_allowed(p, ctx);
3007 :
3008 : return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
3009 :
3010 : out:
3011 : task_rq_unlock(rq, p, rf);
3012 :
3013 : return ret;
3014 : }
3015 :
3016 : /*
3017 : * Change a given task's CPU affinity. Migrate the thread to a
3018 : * proper CPU and schedule it away if the CPU it's executing on
3019 : * is removed from the allowed bitmask.
3020 : *
3021 : * NOTE: the caller must have a valid reference to the task, the
3022 : * task must not exit() & deallocate itself prematurely. The
3023 : * call is not atomic; no spinlocks may be held.
3024 : */
3025 : static int __set_cpus_allowed_ptr(struct task_struct *p,
3026 : struct affinity_context *ctx)
3027 : {
3028 : struct rq_flags rf;
3029 : struct rq *rq;
3030 :
3031 : rq = task_rq_lock(p, &rf);
3032 : /*
3033 : * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3034 : * flags are set.
3035 : */
3036 : if (p->user_cpus_ptr &&
3037 : !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
3038 : cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
3039 : ctx->new_mask = rq->scratch_mask;
3040 :
3041 : return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
3042 : }
3043 :
3044 : int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3045 : {
3046 : struct affinity_context ac = {
3047 : .new_mask = new_mask,
3048 : .flags = 0,
3049 : };
3050 :
3051 : return __set_cpus_allowed_ptr(p, &ac);
3052 : }
3053 : EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
3054 :
3055 : /*
3056 : * Change a given task's CPU affinity to the intersection of its current
3057 : * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3058 : * If user_cpus_ptr is defined, use it as the basis for restricting CPU
3059 : * affinity or use cpu_online_mask instead.
3060 : *
3061 : * If the resulting mask is empty, leave the affinity unchanged and return
3062 : * -EINVAL.
3063 : */
3064 : static int restrict_cpus_allowed_ptr(struct task_struct *p,
3065 : struct cpumask *new_mask,
3066 : const struct cpumask *subset_mask)
3067 : {
3068 : struct affinity_context ac = {
3069 : .new_mask = new_mask,
3070 : .flags = 0,
3071 : };
3072 : struct rq_flags rf;
3073 : struct rq *rq;
3074 : int err;
3075 :
3076 : rq = task_rq_lock(p, &rf);
3077 :
3078 : /*
3079 : * Forcefully restricting the affinity of a deadline task is
3080 : * likely to cause problems, so fail and noisily override the
3081 : * mask entirely.
3082 : */
3083 : if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
3084 : err = -EPERM;
3085 : goto err_unlock;
3086 : }
3087 :
3088 : if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
3089 : err = -EINVAL;
3090 : goto err_unlock;
3091 : }
3092 :
3093 : return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
3094 :
3095 : err_unlock:
3096 : task_rq_unlock(rq, p, &rf);
3097 : return err;
3098 : }
3099 :
3100 : /*
3101 : * Restrict the CPU affinity of task @p so that it is a subset of
3102 : * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
3103 : * old affinity mask. If the resulting mask is empty, we warn and walk
3104 : * up the cpuset hierarchy until we find a suitable mask.
3105 : */
3106 : void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3107 : {
3108 : cpumask_var_t new_mask;
3109 : const struct cpumask *override_mask = task_cpu_possible_mask(p);
3110 :
3111 : alloc_cpumask_var(&new_mask, GFP_KERNEL);
3112 :
3113 : /*
3114 : * __migrate_task() can fail silently in the face of concurrent
3115 : * offlining of the chosen destination CPU, so take the hotplug
3116 : * lock to ensure that the migration succeeds.
3117 : */
3118 : cpus_read_lock();
3119 : if (!cpumask_available(new_mask))
3120 : goto out_set_mask;
3121 :
3122 : if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3123 : goto out_free_mask;
3124 :
3125 : /*
3126 : * We failed to find a valid subset of the affinity mask for the
3127 : * task, so override it based on its cpuset hierarchy.
3128 : */
3129 : cpuset_cpus_allowed(p, new_mask);
3130 : override_mask = new_mask;
3131 :
3132 : out_set_mask:
3133 : if (printk_ratelimit()) {
3134 : printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3135 : task_pid_nr(p), p->comm,
3136 : cpumask_pr_args(override_mask));
3137 : }
3138 :
3139 : WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3140 : out_free_mask:
3141 : cpus_read_unlock();
3142 : free_cpumask_var(new_mask);
3143 : }
3144 :
3145 : static int
3146 : __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
3147 :
3148 : /*
3149 : * Restore the affinity of a task @p which was previously restricted by a
3150 : * call to force_compatible_cpus_allowed_ptr().
3151 : *
3152 : * It is the caller's responsibility to serialise this with any calls to
3153 : * force_compatible_cpus_allowed_ptr(@p).
3154 : */
3155 : void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3156 : {
3157 : struct affinity_context ac = {
3158 : .new_mask = task_user_cpus(p),
3159 : .flags = 0,
3160 : };
3161 : int ret;
3162 :
3163 : /*
3164 : * Try to restore the old affinity mask with __sched_setaffinity().
3165 : * Cpuset masking will be done there too.
3166 : */
3167 : ret = __sched_setaffinity(p, &ac);
3168 : WARN_ON_ONCE(ret);
3169 : }
3170 :
3171 : void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3172 : {
3173 : #ifdef CONFIG_SCHED_DEBUG
3174 : unsigned int state = READ_ONCE(p->__state);
3175 :
3176 : /*
3177 : * We should never call set_task_cpu() on a blocked task,
3178 : * ttwu() will sort out the placement.
3179 : */
3180 : WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3181 :
3182 : /*
3183 : * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3184 : * because schedstat_wait_{start,end} rebase migrating task's wait_start
3185 : * time relying on p->on_rq.
3186 : */
3187 : WARN_ON_ONCE(state == TASK_RUNNING &&
3188 : p->sched_class == &fair_sched_class &&
3189 : (p->on_rq && !task_on_rq_migrating(p)));
3190 :
3191 : #ifdef CONFIG_LOCKDEP
3192 : /*
3193 : * The caller should hold either p->pi_lock or rq->lock, when changing
3194 : * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3195 : *
3196 : * sched_move_task() holds both and thus holding either pins the cgroup,
3197 : * see task_group().
3198 : *
3199 : * Furthermore, all task_rq users should acquire both locks, see
3200 : * task_rq_lock().
3201 : */
3202 : WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3203 : lockdep_is_held(__rq_lockp(task_rq(p)))));
3204 : #endif
3205 : /*
3206 : * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3207 : */
3208 : WARN_ON_ONCE(!cpu_online(new_cpu));
3209 :
3210 : WARN_ON_ONCE(is_migration_disabled(p));
3211 : #endif
3212 :
3213 : trace_sched_migrate_task(p, new_cpu);
3214 :
3215 : if (task_cpu(p) != new_cpu) {
3216 : if (p->sched_class->migrate_task_rq)
3217 : p->sched_class->migrate_task_rq(p, new_cpu);
3218 : p->se.nr_migrations++;
3219 : rseq_migrate(p);
3220 : sched_mm_cid_migrate_from(p);
3221 : perf_event_task_migrate(p);
3222 : }
3223 :
3224 : __set_task_cpu(p, new_cpu);
3225 : }
3226 :
3227 : #ifdef CONFIG_NUMA_BALANCING
3228 : static void __migrate_swap_task(struct task_struct *p, int cpu)
3229 : {
3230 : if (task_on_rq_queued(p)) {
3231 : struct rq *src_rq, *dst_rq;
3232 : struct rq_flags srf, drf;
3233 :
3234 : src_rq = task_rq(p);
3235 : dst_rq = cpu_rq(cpu);
3236 :
3237 : rq_pin_lock(src_rq, &srf);
3238 : rq_pin_lock(dst_rq, &drf);
3239 :
3240 : deactivate_task(src_rq, p, 0);
3241 : set_task_cpu(p, cpu);
3242 : activate_task(dst_rq, p, 0);
3243 : check_preempt_curr(dst_rq, p, 0);
3244 :
3245 : rq_unpin_lock(dst_rq, &drf);
3246 : rq_unpin_lock(src_rq, &srf);
3247 :
3248 : } else {
3249 : /*
3250 : * Task isn't running anymore; make it appear like we migrated
3251 : * it before it went to sleep. This means on wakeup we make the
3252 : * previous CPU our target instead of where it really is.
3253 : */
3254 : p->wake_cpu = cpu;
3255 : }
3256 : }
3257 :
3258 : struct migration_swap_arg {
3259 : struct task_struct *src_task, *dst_task;
3260 : int src_cpu, dst_cpu;
3261 : };
3262 :
3263 : static int migrate_swap_stop(void *data)
3264 : {
3265 : struct migration_swap_arg *arg = data;
3266 : struct rq *src_rq, *dst_rq;
3267 : int ret = -EAGAIN;
3268 :
3269 : if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3270 : return -EAGAIN;
3271 :
3272 : src_rq = cpu_rq(arg->src_cpu);
3273 : dst_rq = cpu_rq(arg->dst_cpu);
3274 :
3275 : double_raw_lock(&arg->src_task->pi_lock,
3276 : &arg->dst_task->pi_lock);
3277 : double_rq_lock(src_rq, dst_rq);
3278 :
3279 : if (task_cpu(arg->dst_task) != arg->dst_cpu)
3280 : goto unlock;
3281 :
3282 : if (task_cpu(arg->src_task) != arg->src_cpu)
3283 : goto unlock;
3284 :
3285 : if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3286 : goto unlock;
3287 :
3288 : if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3289 : goto unlock;
3290 :
3291 : __migrate_swap_task(arg->src_task, arg->dst_cpu);
3292 : __migrate_swap_task(arg->dst_task, arg->src_cpu);
3293 :
3294 : ret = 0;
3295 :
3296 : unlock:
3297 : double_rq_unlock(src_rq, dst_rq);
3298 : raw_spin_unlock(&arg->dst_task->pi_lock);
3299 : raw_spin_unlock(&arg->src_task->pi_lock);
3300 :
3301 : return ret;
3302 : }
3303 :
3304 : /*
3305 : * Cross migrate two tasks
3306 : */
3307 : int migrate_swap(struct task_struct *cur, struct task_struct *p,
3308 : int target_cpu, int curr_cpu)
3309 : {
3310 : struct migration_swap_arg arg;
3311 : int ret = -EINVAL;
3312 :
3313 : arg = (struct migration_swap_arg){
3314 : .src_task = cur,
3315 : .src_cpu = curr_cpu,
3316 : .dst_task = p,
3317 : .dst_cpu = target_cpu,
3318 : };
3319 :
3320 : if (arg.src_cpu == arg.dst_cpu)
3321 : goto out;
3322 :
3323 : /*
3324 : * These three tests are all lockless; this is OK since all of them
3325 : * will be re-checked with proper locks held further down the line.
3326 : */
3327 : if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3328 : goto out;
3329 :
3330 : if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3331 : goto out;
3332 :
3333 : if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3334 : goto out;
3335 :
3336 : trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3337 : ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3338 :
3339 : out:
3340 : return ret;
3341 : }
3342 : #endif /* CONFIG_NUMA_BALANCING */
3343 :
3344 : /*
3345 : * wait_task_inactive - wait for a thread to unschedule.
3346 : *
3347 : * Wait for the thread to block in any of the states set in @match_state.
3348 : * If it changes, i.e. @p might have woken up, then return zero. When we
3349 : * succeed in waiting for @p to be off its CPU, we return a positive number
3350 : * (its total switch count). If a second call a short while later returns the
3351 : * same number, the caller can be sure that @p has remained unscheduled the
3352 : * whole time.
3353 : *
3354 : * The caller must ensure that the task *will* unschedule sometime soon,
3355 : * else this function might spin for a *long* time. This function can't
3356 : * be called with interrupts off, or it may introduce deadlock with
3357 : * smp_call_function() if an IPI is sent by the same process we are
3358 : * waiting to become inactive.
3359 : */
3360 : unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3361 : {
3362 : int running, queued;
3363 : struct rq_flags rf;
3364 : unsigned long ncsw;
3365 : struct rq *rq;
3366 :
3367 : for (;;) {
3368 : /*
3369 : * We do the initial early heuristics without holding
3370 : * any task-queue locks at all. We'll only try to get
3371 : * the runqueue lock when things look like they will
3372 : * work out!
3373 : */
3374 : rq = task_rq(p);
3375 :
3376 : /*
3377 : * If the task is actively running on another CPU
3378 : * still, just relax and busy-wait without holding
3379 : * any locks.
3380 : *
3381 : * NOTE! Since we don't hold any locks, it's not
3382 : * even sure that "rq" stays as the right runqueue!
3383 : * But we don't care, since "task_on_cpu()" will
3384 : * return false if the runqueue has changed and p
3385 : * is actually now running somewhere else!
3386 : */
3387 : while (task_on_cpu(rq, p)) {
3388 : if (!(READ_ONCE(p->__state) & match_state))
3389 : return 0;
3390 : cpu_relax();
3391 : }
3392 :
3393 : /*
3394 : * Ok, time to look more closely! We need the rq
3395 : * lock now, to be *sure*. If we're wrong, we'll
3396 : * just go back and repeat.
3397 : */
3398 : rq = task_rq_lock(p, &rf);
3399 : trace_sched_wait_task(p);
3400 : running = task_on_cpu(rq, p);
3401 : queued = task_on_rq_queued(p);
3402 : ncsw = 0;
3403 : if (READ_ONCE(p->__state) & match_state)
3404 : ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3405 : task_rq_unlock(rq, p, &rf);
3406 :
3407 : /*
3408 : * If it changed from the expected state, bail out now.
3409 : */
3410 : if (unlikely(!ncsw))
3411 : break;
3412 :
3413 : /*
3414 : * Was it really running after all now that we
3415 : * checked with the proper locks actually held?
3416 : *
3417 : * Oops. Go back and try again..
3418 : */
3419 : if (unlikely(running)) {
3420 : cpu_relax();
3421 : continue;
3422 : }
3423 :
3424 : /*
3425 : * It's not enough that it's not actively running,
3426 : * it must be off the runqueue _entirely_, and not
3427 : * preempted!
3428 : *
3429 : * So if it was still runnable (but just not actively
3430 : * running right now), it's preempted, and we should
3431 : * yield - it could be a while.
3432 : */
3433 : if (unlikely(queued)) {
3434 : ktime_t to = NSEC_PER_SEC / HZ;
3435 :
3436 : set_current_state(TASK_UNINTERRUPTIBLE);
3437 : schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3438 : continue;
3439 : }
3440 :
3441 : /*
3442 : * Ahh, all good. It wasn't running, and it wasn't
3443 : * runnable, which means that it will never become
3444 : * running in the future either. We're all done!
3445 : */
3446 : break;
3447 : }
3448 :
3449 : return ncsw;
3450 : }
3451 :
3452 : /***
3453 : * kick_process - kick a running thread to enter/exit the kernel
3454 : * @p: the to-be-kicked thread
3455 : *
3456 : * Cause a process which is running on another CPU to enter
3457 : * kernel-mode, without any delay. (to get signals handled.)
3458 : *
3459 : * NOTE: this function doesn't have to take the runqueue lock,
3460 : * because all it wants to ensure is that the remote task enters
3461 : * the kernel. If the IPI races and the task has been migrated
3462 : * to another CPU then no harm is done and the purpose has been
3463 : * achieved as well.
3464 : */
3465 : void kick_process(struct task_struct *p)
3466 : {
3467 : int cpu;
3468 :
3469 : preempt_disable();
3470 : cpu = task_cpu(p);
3471 : if ((cpu != smp_processor_id()) && task_curr(p))
3472 : smp_send_reschedule(cpu);
3473 : preempt_enable();
3474 : }
3475 : EXPORT_SYMBOL_GPL(kick_process);
3476 :
3477 : /*
3478 : * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3479 : *
3480 : * A few notes on cpu_active vs cpu_online:
3481 : *
3482 : * - cpu_active must be a subset of cpu_online
3483 : *
3484 : * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3485 : * see __set_cpus_allowed_ptr(). At this point the newly online
3486 : * CPU isn't yet part of the sched domains, and balancing will not
3487 : * see it.
3488 : *
3489 : * - on CPU-down we clear cpu_active() to mask the sched domains and
3490 : * avoid the load balancer to place new tasks on the to be removed
3491 : * CPU. Existing tasks will remain running there and will be taken
3492 : * off.
3493 : *
3494 : * This means that fallback selection must not select !active CPUs.
3495 : * And can assume that any active CPU must be online. Conversely
3496 : * select_task_rq() below may allow selection of !active CPUs in order
3497 : * to satisfy the above rules.
3498 : */
3499 : static int select_fallback_rq(int cpu, struct task_struct *p)
3500 : {
3501 : int nid = cpu_to_node(cpu);
3502 : const struct cpumask *nodemask = NULL;
3503 : enum { cpuset, possible, fail } state = cpuset;
3504 : int dest_cpu;
3505 :
3506 : /*
3507 : * If the node that the CPU is on has been offlined, cpu_to_node()
3508 : * will return -1. There is no CPU on the node, and we should
3509 : * select the CPU on the other node.
3510 : */
3511 : if (nid != -1) {
3512 : nodemask = cpumask_of_node(nid);
3513 :
3514 : /* Look for allowed, online CPU in same node. */
3515 : for_each_cpu(dest_cpu, nodemask) {
3516 : if (is_cpu_allowed(p, dest_cpu))
3517 : return dest_cpu;
3518 : }
3519 : }
3520 :
3521 : for (;;) {
3522 : /* Any allowed, online CPU? */
3523 : for_each_cpu(dest_cpu, p->cpus_ptr) {
3524 : if (!is_cpu_allowed(p, dest_cpu))
3525 : continue;
3526 :
3527 : goto out;
3528 : }
3529 :
3530 : /* No more Mr. Nice Guy. */
3531 : switch (state) {
3532 : case cpuset:
3533 : if (cpuset_cpus_allowed_fallback(p)) {
3534 : state = possible;
3535 : break;
3536 : }
3537 : fallthrough;
3538 : case possible:
3539 : /*
3540 : * XXX When called from select_task_rq() we only
3541 : * hold p->pi_lock and again violate locking order.
3542 : *
3543 : * More yuck to audit.
3544 : */
3545 : do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3546 : state = fail;
3547 : break;
3548 : case fail:
3549 : BUG();
3550 : break;
3551 : }
3552 : }
3553 :
3554 : out:
3555 : if (state != cpuset) {
3556 : /*
3557 : * Don't tell them about moving exiting tasks or
3558 : * kernel threads (both mm NULL), since they never
3559 : * leave kernel.
3560 : */
3561 : if (p->mm && printk_ratelimit()) {
3562 : printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3563 : task_pid_nr(p), p->comm, cpu);
3564 : }
3565 : }
3566 :
3567 : return dest_cpu;
3568 : }
3569 :
3570 : /*
3571 : * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3572 : */
3573 : static inline
3574 : int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3575 : {
3576 : lockdep_assert_held(&p->pi_lock);
3577 :
3578 : if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3579 : cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3580 : else
3581 : cpu = cpumask_any(p->cpus_ptr);
3582 :
3583 : /*
3584 : * In order not to call set_task_cpu() on a blocking task we need
3585 : * to rely on ttwu() to place the task on a valid ->cpus_ptr
3586 : * CPU.
3587 : *
3588 : * Since this is common to all placement strategies, this lives here.
3589 : *
3590 : * [ this allows ->select_task() to simply return task_cpu(p) and
3591 : * not worry about this generic constraint ]
3592 : */
3593 : if (unlikely(!is_cpu_allowed(p, cpu)))
3594 : cpu = select_fallback_rq(task_cpu(p), p);
3595 :
3596 : return cpu;
3597 : }
3598 :
3599 : void sched_set_stop_task(int cpu, struct task_struct *stop)
3600 : {
3601 : static struct lock_class_key stop_pi_lock;
3602 : struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3603 : struct task_struct *old_stop = cpu_rq(cpu)->stop;
3604 :
3605 : if (stop) {
3606 : /*
3607 : * Make it appear like a SCHED_FIFO task, its something
3608 : * userspace knows about and won't get confused about.
3609 : *
3610 : * Also, it will make PI more or less work without too
3611 : * much confusion -- but then, stop work should not
3612 : * rely on PI working anyway.
3613 : */
3614 : sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3615 :
3616 : stop->sched_class = &stop_sched_class;
3617 :
3618 : /*
3619 : * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3620 : * adjust the effective priority of a task. As a result,
3621 : * rt_mutex_setprio() can trigger (RT) balancing operations,
3622 : * which can then trigger wakeups of the stop thread to push
3623 : * around the current task.
3624 : *
3625 : * The stop task itself will never be part of the PI-chain, it
3626 : * never blocks, therefore that ->pi_lock recursion is safe.
3627 : * Tell lockdep about this by placing the stop->pi_lock in its
3628 : * own class.
3629 : */
3630 : lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3631 : }
3632 :
3633 : cpu_rq(cpu)->stop = stop;
3634 :
3635 : if (old_stop) {
3636 : /*
3637 : * Reset it back to a normal scheduling class so that
3638 : * it can die in pieces.
3639 : */
3640 : old_stop->sched_class = &rt_sched_class;
3641 : }
3642 : }
3643 :
3644 : #else /* CONFIG_SMP */
3645 :
3646 : static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3647 : struct affinity_context *ctx)
3648 : {
3649 0 : return set_cpus_allowed_ptr(p, ctx->new_mask);
3650 : }
3651 :
3652 : static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3653 :
3654 : static inline bool rq_has_pinned_tasks(struct rq *rq)
3655 : {
3656 : return false;
3657 : }
3658 :
3659 : static inline cpumask_t *alloc_user_cpus_ptr(int node)
3660 : {
3661 : return NULL;
3662 : }
3663 :
3664 : #endif /* !CONFIG_SMP */
3665 :
3666 : static void
3667 : ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3668 : {
3669 : struct rq *rq;
3670 :
3671 : if (!schedstat_enabled())
3672 : return;
3673 :
3674 : rq = this_rq();
3675 :
3676 : #ifdef CONFIG_SMP
3677 : if (cpu == rq->cpu) {
3678 : __schedstat_inc(rq->ttwu_local);
3679 : __schedstat_inc(p->stats.nr_wakeups_local);
3680 : } else {
3681 : struct sched_domain *sd;
3682 :
3683 : __schedstat_inc(p->stats.nr_wakeups_remote);
3684 : rcu_read_lock();
3685 : for_each_domain(rq->cpu, sd) {
3686 : if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3687 : __schedstat_inc(sd->ttwu_wake_remote);
3688 : break;
3689 : }
3690 : }
3691 : rcu_read_unlock();
3692 : }
3693 :
3694 : if (wake_flags & WF_MIGRATED)
3695 : __schedstat_inc(p->stats.nr_wakeups_migrate);
3696 : #endif /* CONFIG_SMP */
3697 :
3698 : __schedstat_inc(rq->ttwu_count);
3699 : __schedstat_inc(p->stats.nr_wakeups);
3700 :
3701 : if (wake_flags & WF_SYNC)
3702 : __schedstat_inc(p->stats.nr_wakeups_sync);
3703 : }
3704 :
3705 : /*
3706 : * Mark the task runnable.
3707 : */
3708 : static inline void ttwu_do_wakeup(struct task_struct *p)
3709 : {
3710 2060 : WRITE_ONCE(p->__state, TASK_RUNNING);
3711 2060 : trace_sched_wakeup(p);
3712 : }
3713 :
3714 : static void
3715 2060 : ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3716 : struct rq_flags *rf)
3717 : {
3718 2060 : int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3719 :
3720 2060 : lockdep_assert_rq_held(rq);
3721 :
3722 2060 : if (p->sched_contributes_to_load)
3723 1153 : rq->nr_uninterruptible--;
3724 :
3725 : #ifdef CONFIG_SMP
3726 : if (wake_flags & WF_MIGRATED)
3727 : en_flags |= ENQUEUE_MIGRATED;
3728 : else
3729 : #endif
3730 2060 : if (p->in_iowait) {
3731 0 : delayacct_blkio_end(p);
3732 0 : atomic_dec(&task_rq(p)->nr_iowait);
3733 : }
3734 :
3735 2060 : activate_task(rq, p, en_flags);
3736 2060 : check_preempt_curr(rq, p, wake_flags);
3737 :
3738 2060 : ttwu_do_wakeup(p);
3739 :
3740 : #ifdef CONFIG_SMP
3741 : if (p->sched_class->task_woken) {
3742 : /*
3743 : * Our task @p is fully woken up and running; so it's safe to
3744 : * drop the rq->lock, hereafter rq is only used for statistics.
3745 : */
3746 : rq_unpin_lock(rq, rf);
3747 : p->sched_class->task_woken(rq, p);
3748 : rq_repin_lock(rq, rf);
3749 : }
3750 :
3751 : if (rq->idle_stamp) {
3752 : u64 delta = rq_clock(rq) - rq->idle_stamp;
3753 : u64 max = 2*rq->max_idle_balance_cost;
3754 :
3755 : update_avg(&rq->avg_idle, delta);
3756 :
3757 : if (rq->avg_idle > max)
3758 : rq->avg_idle = max;
3759 :
3760 : rq->wake_stamp = jiffies;
3761 : rq->wake_avg_idle = rq->avg_idle / 2;
3762 :
3763 : rq->idle_stamp = 0;
3764 : }
3765 : #endif
3766 2060 : }
3767 :
3768 : /*
3769 : * Consider @p being inside a wait loop:
3770 : *
3771 : * for (;;) {
3772 : * set_current_state(TASK_UNINTERRUPTIBLE);
3773 : *
3774 : * if (CONDITION)
3775 : * break;
3776 : *
3777 : * schedule();
3778 : * }
3779 : * __set_current_state(TASK_RUNNING);
3780 : *
3781 : * between set_current_state() and schedule(). In this case @p is still
3782 : * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3783 : * an atomic manner.
3784 : *
3785 : * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3786 : * then schedule() must still happen and p->state can be changed to
3787 : * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3788 : * need to do a full wakeup with enqueue.
3789 : *
3790 : * Returns: %true when the wakeup is done,
3791 : * %false otherwise.
3792 : */
3793 0 : static int ttwu_runnable(struct task_struct *p, int wake_flags)
3794 : {
3795 : struct rq_flags rf;
3796 : struct rq *rq;
3797 0 : int ret = 0;
3798 :
3799 0 : rq = __task_rq_lock(p, &rf);
3800 0 : if (task_on_rq_queued(p)) {
3801 0 : if (!task_on_cpu(rq, p)) {
3802 : /*
3803 : * When on_rq && !on_cpu the task is preempted, see if
3804 : * it should preempt the task that is current now.
3805 : */
3806 0 : update_rq_clock(rq);
3807 0 : check_preempt_curr(rq, p, wake_flags);
3808 : }
3809 0 : ttwu_do_wakeup(p);
3810 0 : ret = 1;
3811 : }
3812 0 : __task_rq_unlock(rq, &rf);
3813 :
3814 0 : return ret;
3815 : }
3816 :
3817 : #ifdef CONFIG_SMP
3818 : void sched_ttwu_pending(void *arg)
3819 : {
3820 : struct llist_node *llist = arg;
3821 : struct rq *rq = this_rq();
3822 : struct task_struct *p, *t;
3823 : struct rq_flags rf;
3824 :
3825 : if (!llist)
3826 : return;
3827 :
3828 : rq_lock_irqsave(rq, &rf);
3829 : update_rq_clock(rq);
3830 :
3831 : llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3832 : if (WARN_ON_ONCE(p->on_cpu))
3833 : smp_cond_load_acquire(&p->on_cpu, !VAL);
3834 :
3835 : if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3836 : set_task_cpu(p, cpu_of(rq));
3837 :
3838 : ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3839 : }
3840 :
3841 : /*
3842 : * Must be after enqueueing at least once task such that
3843 : * idle_cpu() does not observe a false-negative -- if it does,
3844 : * it is possible for select_idle_siblings() to stack a number
3845 : * of tasks on this CPU during that window.
3846 : *
3847 : * It is ok to clear ttwu_pending when another task pending.
3848 : * We will receive IPI after local irq enabled and then enqueue it.
3849 : * Since now nr_running > 0, idle_cpu() will always get correct result.
3850 : */
3851 : WRITE_ONCE(rq->ttwu_pending, 0);
3852 : rq_unlock_irqrestore(rq, &rf);
3853 : }
3854 :
3855 : /*
3856 : * Prepare the scene for sending an IPI for a remote smp_call
3857 : *
3858 : * Returns true if the caller can proceed with sending the IPI.
3859 : * Returns false otherwise.
3860 : */
3861 : bool call_function_single_prep_ipi(int cpu)
3862 : {
3863 : if (set_nr_if_polling(cpu_rq(cpu)->idle)) {
3864 : trace_sched_wake_idle_without_ipi(cpu);
3865 : return false;
3866 : }
3867 :
3868 : return true;
3869 : }
3870 :
3871 : /*
3872 : * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3873 : * necessary. The wakee CPU on receipt of the IPI will queue the task
3874 : * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3875 : * of the wakeup instead of the waker.
3876 : */
3877 : static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3878 : {
3879 : struct rq *rq = cpu_rq(cpu);
3880 :
3881 : p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3882 :
3883 : WRITE_ONCE(rq->ttwu_pending, 1);
3884 : __smp_call_single_queue(cpu, &p->wake_entry.llist);
3885 : }
3886 :
3887 : void wake_up_if_idle(int cpu)
3888 : {
3889 : struct rq *rq = cpu_rq(cpu);
3890 : struct rq_flags rf;
3891 :
3892 : rcu_read_lock();
3893 :
3894 : if (!is_idle_task(rcu_dereference(rq->curr)))
3895 : goto out;
3896 :
3897 : rq_lock_irqsave(rq, &rf);
3898 : if (is_idle_task(rq->curr))
3899 : resched_curr(rq);
3900 : /* Else CPU is not idle, do nothing here: */
3901 : rq_unlock_irqrestore(rq, &rf);
3902 :
3903 : out:
3904 : rcu_read_unlock();
3905 : }
3906 :
3907 : bool cpus_share_cache(int this_cpu, int that_cpu)
3908 : {
3909 : if (this_cpu == that_cpu)
3910 : return true;
3911 :
3912 : return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3913 : }
3914 :
3915 : static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3916 : {
3917 : /*
3918 : * Do not complicate things with the async wake_list while the CPU is
3919 : * in hotplug state.
3920 : */
3921 : if (!cpu_active(cpu))
3922 : return false;
3923 :
3924 : /* Ensure the task will still be allowed to run on the CPU. */
3925 : if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3926 : return false;
3927 :
3928 : /*
3929 : * If the CPU does not share cache, then queue the task on the
3930 : * remote rqs wakelist to avoid accessing remote data.
3931 : */
3932 : if (!cpus_share_cache(smp_processor_id(), cpu))
3933 : return true;
3934 :
3935 : if (cpu == smp_processor_id())
3936 : return false;
3937 :
3938 : /*
3939 : * If the wakee cpu is idle, or the task is descheduling and the
3940 : * only running task on the CPU, then use the wakelist to offload
3941 : * the task activation to the idle (or soon-to-be-idle) CPU as
3942 : * the current CPU is likely busy. nr_running is checked to
3943 : * avoid unnecessary task stacking.
3944 : *
3945 : * Note that we can only get here with (wakee) p->on_rq=0,
3946 : * p->on_cpu can be whatever, we've done the dequeue, so
3947 : * the wakee has been accounted out of ->nr_running.
3948 : */
3949 : if (!cpu_rq(cpu)->nr_running)
3950 : return true;
3951 :
3952 : return false;
3953 : }
3954 :
3955 : static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3956 : {
3957 : if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
3958 : sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3959 : __ttwu_queue_wakelist(p, cpu, wake_flags);
3960 : return true;
3961 : }
3962 :
3963 : return false;
3964 : }
3965 :
3966 : #else /* !CONFIG_SMP */
3967 :
3968 : static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3969 : {
3970 : return false;
3971 : }
3972 :
3973 : #endif /* CONFIG_SMP */
3974 :
3975 : static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3976 : {
3977 2060 : struct rq *rq = cpu_rq(cpu);
3978 : struct rq_flags rf;
3979 :
3980 2060 : if (ttwu_queue_wakelist(p, cpu, wake_flags))
3981 : return;
3982 :
3983 2060 : rq_lock(rq, &rf);
3984 2060 : update_rq_clock(rq);
3985 2060 : ttwu_do_activate(rq, p, wake_flags, &rf);
3986 2060 : rq_unlock(rq, &rf);
3987 : }
3988 :
3989 : /*
3990 : * Invoked from try_to_wake_up() to check whether the task can be woken up.
3991 : *
3992 : * The caller holds p::pi_lock if p != current or has preemption
3993 : * disabled when p == current.
3994 : *
3995 : * The rules of PREEMPT_RT saved_state:
3996 : *
3997 : * The related locking code always holds p::pi_lock when updating
3998 : * p::saved_state, which means the code is fully serialized in both cases.
3999 : *
4000 : * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
4001 : * bits set. This allows to distinguish all wakeup scenarios.
4002 : */
4003 : static __always_inline
4004 : bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
4005 : {
4006 : if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
4007 : WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
4008 : state != TASK_RTLOCK_WAIT);
4009 : }
4010 :
4011 2475 : if (READ_ONCE(p->__state) & state) {
4012 2060 : *success = 1;
4013 : return true;
4014 : }
4015 :
4016 : #ifdef CONFIG_PREEMPT_RT
4017 : /*
4018 : * Saved state preserves the task state across blocking on
4019 : * an RT lock. If the state matches, set p::saved_state to
4020 : * TASK_RUNNING, but do not wake the task because it waits
4021 : * for a lock wakeup. Also indicate success because from
4022 : * the regular waker's point of view this has succeeded.
4023 : *
4024 : * After acquiring the lock the task will restore p::__state
4025 : * from p::saved_state which ensures that the regular
4026 : * wakeup is not lost. The restore will also set
4027 : * p::saved_state to TASK_RUNNING so any further tests will
4028 : * not result in false positives vs. @success
4029 : */
4030 : if (p->saved_state & state) {
4031 : p->saved_state = TASK_RUNNING;
4032 : *success = 1;
4033 : }
4034 : #endif
4035 : return false;
4036 : }
4037 :
4038 : /*
4039 : * Notes on Program-Order guarantees on SMP systems.
4040 : *
4041 : * MIGRATION
4042 : *
4043 : * The basic program-order guarantee on SMP systems is that when a task [t]
4044 : * migrates, all its activity on its old CPU [c0] happens-before any subsequent
4045 : * execution on its new CPU [c1].
4046 : *
4047 : * For migration (of runnable tasks) this is provided by the following means:
4048 : *
4049 : * A) UNLOCK of the rq(c0)->lock scheduling out task t
4050 : * B) migration for t is required to synchronize *both* rq(c0)->lock and
4051 : * rq(c1)->lock (if not at the same time, then in that order).
4052 : * C) LOCK of the rq(c1)->lock scheduling in task
4053 : *
4054 : * Release/acquire chaining guarantees that B happens after A and C after B.
4055 : * Note: the CPU doing B need not be c0 or c1
4056 : *
4057 : * Example:
4058 : *
4059 : * CPU0 CPU1 CPU2
4060 : *
4061 : * LOCK rq(0)->lock
4062 : * sched-out X
4063 : * sched-in Y
4064 : * UNLOCK rq(0)->lock
4065 : *
4066 : * LOCK rq(0)->lock // orders against CPU0
4067 : * dequeue X
4068 : * UNLOCK rq(0)->lock
4069 : *
4070 : * LOCK rq(1)->lock
4071 : * enqueue X
4072 : * UNLOCK rq(1)->lock
4073 : *
4074 : * LOCK rq(1)->lock // orders against CPU2
4075 : * sched-out Z
4076 : * sched-in X
4077 : * UNLOCK rq(1)->lock
4078 : *
4079 : *
4080 : * BLOCKING -- aka. SLEEP + WAKEUP
4081 : *
4082 : * For blocking we (obviously) need to provide the same guarantee as for
4083 : * migration. However the means are completely different as there is no lock
4084 : * chain to provide order. Instead we do:
4085 : *
4086 : * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
4087 : * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
4088 : *
4089 : * Example:
4090 : *
4091 : * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
4092 : *
4093 : * LOCK rq(0)->lock LOCK X->pi_lock
4094 : * dequeue X
4095 : * sched-out X
4096 : * smp_store_release(X->on_cpu, 0);
4097 : *
4098 : * smp_cond_load_acquire(&X->on_cpu, !VAL);
4099 : * X->state = WAKING
4100 : * set_task_cpu(X,2)
4101 : *
4102 : * LOCK rq(2)->lock
4103 : * enqueue X
4104 : * X->state = RUNNING
4105 : * UNLOCK rq(2)->lock
4106 : *
4107 : * LOCK rq(2)->lock // orders against CPU1
4108 : * sched-out Z
4109 : * sched-in X
4110 : * UNLOCK rq(2)->lock
4111 : *
4112 : * UNLOCK X->pi_lock
4113 : * UNLOCK rq(0)->lock
4114 : *
4115 : *
4116 : * However, for wakeups there is a second guarantee we must provide, namely we
4117 : * must ensure that CONDITION=1 done by the caller can not be reordered with
4118 : * accesses to the task state; see try_to_wake_up() and set_current_state().
4119 : */
4120 :
4121 : /**
4122 : * try_to_wake_up - wake up a thread
4123 : * @p: the thread to be awakened
4124 : * @state: the mask of task states that can be woken
4125 : * @wake_flags: wake modifier flags (WF_*)
4126 : *
4127 : * Conceptually does:
4128 : *
4129 : * If (@state & @p->state) @p->state = TASK_RUNNING.
4130 : *
4131 : * If the task was not queued/runnable, also place it back on a runqueue.
4132 : *
4133 : * This function is atomic against schedule() which would dequeue the task.
4134 : *
4135 : * It issues a full memory barrier before accessing @p->state, see the comment
4136 : * with set_current_state().
4137 : *
4138 : * Uses p->pi_lock to serialize against concurrent wake-ups.
4139 : *
4140 : * Relies on p->pi_lock stabilizing:
4141 : * - p->sched_class
4142 : * - p->cpus_ptr
4143 : * - p->sched_task_group
4144 : * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4145 : *
4146 : * Tries really hard to only take one task_rq(p)->lock for performance.
4147 : * Takes rq->lock in:
4148 : * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4149 : * - ttwu_queue() -- new rq, for enqueue of the task;
4150 : * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4151 : *
4152 : * As a consequence we race really badly with just about everything. See the
4153 : * many memory barriers and their comments for details.
4154 : *
4155 : * Return: %true if @p->state changes (an actual wakeup was done),
4156 : * %false otherwise.
4157 : */
4158 : static int
4159 2475 : try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4160 : {
4161 : unsigned long flags;
4162 2475 : int cpu, success = 0;
4163 :
4164 2475 : preempt_disable();
4165 2475 : if (p == current) {
4166 : /*
4167 : * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4168 : * == smp_processor_id()'. Together this means we can special
4169 : * case the whole 'p->on_rq && ttwu_runnable()' case below
4170 : * without taking any locks.
4171 : *
4172 : * In particular:
4173 : * - we rely on Program-Order guarantees for all the ordering,
4174 : * - we're serialized against set_special_state() by virtue of
4175 : * it disabling IRQs (this allows not taking ->pi_lock).
4176 : */
4177 45 : if (!ttwu_state_match(p, state, &success))
4178 : goto out;
4179 :
4180 0 : trace_sched_waking(p);
4181 : ttwu_do_wakeup(p);
4182 : goto out;
4183 : }
4184 :
4185 : /*
4186 : * If we are going to wake up a thread waiting for CONDITION we
4187 : * need to ensure that CONDITION=1 done by the caller can not be
4188 : * reordered with p->state check below. This pairs with smp_store_mb()
4189 : * in set_current_state() that the waiting thread does.
4190 : */
4191 2430 : raw_spin_lock_irqsave(&p->pi_lock, flags);
4192 : smp_mb__after_spinlock();
4193 2430 : if (!ttwu_state_match(p, state, &success))
4194 : goto unlock;
4195 :
4196 2060 : trace_sched_waking(p);
4197 :
4198 : /*
4199 : * Ensure we load p->on_rq _after_ p->state, otherwise it would
4200 : * be possible to, falsely, observe p->on_rq == 0 and get stuck
4201 : * in smp_cond_load_acquire() below.
4202 : *
4203 : * sched_ttwu_pending() try_to_wake_up()
4204 : * STORE p->on_rq = 1 LOAD p->state
4205 : * UNLOCK rq->lock
4206 : *
4207 : * __schedule() (switch to task 'p')
4208 : * LOCK rq->lock smp_rmb();
4209 : * smp_mb__after_spinlock();
4210 : * UNLOCK rq->lock
4211 : *
4212 : * [task p]
4213 : * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4214 : *
4215 : * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4216 : * __schedule(). See the comment for smp_mb__after_spinlock().
4217 : *
4218 : * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4219 : */
4220 2060 : smp_rmb();
4221 2060 : if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4222 : goto unlock;
4223 :
4224 : #ifdef CONFIG_SMP
4225 : /*
4226 : * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4227 : * possible to, falsely, observe p->on_cpu == 0.
4228 : *
4229 : * One must be running (->on_cpu == 1) in order to remove oneself
4230 : * from the runqueue.
4231 : *
4232 : * __schedule() (switch to task 'p') try_to_wake_up()
4233 : * STORE p->on_cpu = 1 LOAD p->on_rq
4234 : * UNLOCK rq->lock
4235 : *
4236 : * __schedule() (put 'p' to sleep)
4237 : * LOCK rq->lock smp_rmb();
4238 : * smp_mb__after_spinlock();
4239 : * STORE p->on_rq = 0 LOAD p->on_cpu
4240 : *
4241 : * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4242 : * __schedule(). See the comment for smp_mb__after_spinlock().
4243 : *
4244 : * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4245 : * schedule()'s deactivate_task() has 'happened' and p will no longer
4246 : * care about it's own p->state. See the comment in __schedule().
4247 : */
4248 : smp_acquire__after_ctrl_dep();
4249 :
4250 : /*
4251 : * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4252 : * == 0), which means we need to do an enqueue, change p->state to
4253 : * TASK_WAKING such that we can unlock p->pi_lock before doing the
4254 : * enqueue, such as ttwu_queue_wakelist().
4255 : */
4256 : WRITE_ONCE(p->__state, TASK_WAKING);
4257 :
4258 : /*
4259 : * If the owning (remote) CPU is still in the middle of schedule() with
4260 : * this task as prev, considering queueing p on the remote CPUs wake_list
4261 : * which potentially sends an IPI instead of spinning on p->on_cpu to
4262 : * let the waker make forward progress. This is safe because IRQs are
4263 : * disabled and the IPI will deliver after on_cpu is cleared.
4264 : *
4265 : * Ensure we load task_cpu(p) after p->on_cpu:
4266 : *
4267 : * set_task_cpu(p, cpu);
4268 : * STORE p->cpu = @cpu
4269 : * __schedule() (switch to task 'p')
4270 : * LOCK rq->lock
4271 : * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4272 : * STORE p->on_cpu = 1 LOAD p->cpu
4273 : *
4274 : * to ensure we observe the correct CPU on which the task is currently
4275 : * scheduling.
4276 : */
4277 : if (smp_load_acquire(&p->on_cpu) &&
4278 : ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4279 : goto unlock;
4280 :
4281 : /*
4282 : * If the owning (remote) CPU is still in the middle of schedule() with
4283 : * this task as prev, wait until it's done referencing the task.
4284 : *
4285 : * Pairs with the smp_store_release() in finish_task().
4286 : *
4287 : * This ensures that tasks getting woken will be fully ordered against
4288 : * their previous state and preserve Program Order.
4289 : */
4290 : smp_cond_load_acquire(&p->on_cpu, !VAL);
4291 :
4292 : cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4293 : if (task_cpu(p) != cpu) {
4294 : if (p->in_iowait) {
4295 : delayacct_blkio_end(p);
4296 : atomic_dec(&task_rq(p)->nr_iowait);
4297 : }
4298 :
4299 : wake_flags |= WF_MIGRATED;
4300 : psi_ttwu_dequeue(p);
4301 : set_task_cpu(p, cpu);
4302 : }
4303 : #else
4304 2060 : cpu = task_cpu(p);
4305 : #endif /* CONFIG_SMP */
4306 :
4307 2060 : ttwu_queue(p, cpu, wake_flags);
4308 : unlock:
4309 4860 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4310 : out:
4311 : if (success)
4312 : ttwu_stat(p, task_cpu(p), wake_flags);
4313 2475 : preempt_enable();
4314 :
4315 2475 : return success;
4316 : }
4317 :
4318 : static bool __task_needs_rq_lock(struct task_struct *p)
4319 : {
4320 0 : unsigned int state = READ_ONCE(p->__state);
4321 :
4322 : /*
4323 : * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4324 : * the task is blocked. Make sure to check @state since ttwu() can drop
4325 : * locks at the end, see ttwu_queue_wakelist().
4326 : */
4327 0 : if (state == TASK_RUNNING || state == TASK_WAKING)
4328 : return true;
4329 :
4330 : /*
4331 : * Ensure we load p->on_rq after p->__state, otherwise it would be
4332 : * possible to, falsely, observe p->on_rq == 0.
4333 : *
4334 : * See try_to_wake_up() for a longer comment.
4335 : */
4336 0 : smp_rmb();
4337 0 : if (p->on_rq)
4338 : return true;
4339 :
4340 : #ifdef CONFIG_SMP
4341 : /*
4342 : * Ensure the task has finished __schedule() and will not be referenced
4343 : * anymore. Again, see try_to_wake_up() for a longer comment.
4344 : */
4345 : smp_rmb();
4346 : smp_cond_load_acquire(&p->on_cpu, !VAL);
4347 : #endif
4348 :
4349 : return false;
4350 : }
4351 :
4352 : /**
4353 : * task_call_func - Invoke a function on task in fixed state
4354 : * @p: Process for which the function is to be invoked, can be @current.
4355 : * @func: Function to invoke.
4356 : * @arg: Argument to function.
4357 : *
4358 : * Fix the task in it's current state by avoiding wakeups and or rq operations
4359 : * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4360 : * to work out what the state is, if required. Given that @func can be invoked
4361 : * with a runqueue lock held, it had better be quite lightweight.
4362 : *
4363 : * Returns:
4364 : * Whatever @func returns
4365 : */
4366 0 : int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4367 : {
4368 0 : struct rq *rq = NULL;
4369 : struct rq_flags rf;
4370 : int ret;
4371 :
4372 0 : raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4373 :
4374 0 : if (__task_needs_rq_lock(p))
4375 : rq = __task_rq_lock(p, &rf);
4376 :
4377 : /*
4378 : * At this point the task is pinned; either:
4379 : * - blocked and we're holding off wakeups (pi->lock)
4380 : * - woken, and we're holding off enqueue (rq->lock)
4381 : * - queued, and we're holding off schedule (rq->lock)
4382 : * - running, and we're holding off de-schedule (rq->lock)
4383 : *
4384 : * The called function (@func) can use: task_curr(), p->on_rq and
4385 : * p->__state to differentiate between these states.
4386 : */
4387 0 : ret = func(p, arg);
4388 :
4389 0 : if (rq)
4390 0 : rq_unlock(rq, &rf);
4391 :
4392 0 : raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4393 0 : return ret;
4394 : }
4395 :
4396 : /**
4397 : * cpu_curr_snapshot - Return a snapshot of the currently running task
4398 : * @cpu: The CPU on which to snapshot the task.
4399 : *
4400 : * Returns the task_struct pointer of the task "currently" running on
4401 : * the specified CPU. If the same task is running on that CPU throughout,
4402 : * the return value will be a pointer to that task's task_struct structure.
4403 : * If the CPU did any context switches even vaguely concurrently with the
4404 : * execution of this function, the return value will be a pointer to the
4405 : * task_struct structure of a randomly chosen task that was running on
4406 : * that CPU somewhere around the time that this function was executing.
4407 : *
4408 : * If the specified CPU was offline, the return value is whatever it
4409 : * is, perhaps a pointer to the task_struct structure of that CPU's idle
4410 : * task, but there is no guarantee. Callers wishing a useful return
4411 : * value must take some action to ensure that the specified CPU remains
4412 : * online throughout.
4413 : *
4414 : * This function executes full memory barriers before and after fetching
4415 : * the pointer, which permits the caller to confine this function's fetch
4416 : * with respect to the caller's accesses to other shared variables.
4417 : */
4418 0 : struct task_struct *cpu_curr_snapshot(int cpu)
4419 : {
4420 : struct task_struct *t;
4421 :
4422 0 : smp_mb(); /* Pairing determined by caller's synchronization design. */
4423 0 : t = rcu_dereference(cpu_curr(cpu));
4424 0 : smp_mb(); /* Pairing determined by caller's synchronization design. */
4425 0 : return t;
4426 : }
4427 :
4428 : /**
4429 : * wake_up_process - Wake up a specific process
4430 : * @p: The process to be woken up.
4431 : *
4432 : * Attempt to wake up the nominated process and move it to the set of runnable
4433 : * processes.
4434 : *
4435 : * Return: 1 if the process was woken up, 0 if it was already running.
4436 : *
4437 : * This function executes a full memory barrier before accessing the task state.
4438 : */
4439 2474 : int wake_up_process(struct task_struct *p)
4440 : {
4441 2474 : return try_to_wake_up(p, TASK_NORMAL, 0);
4442 : }
4443 : EXPORT_SYMBOL(wake_up_process);
4444 :
4445 1 : int wake_up_state(struct task_struct *p, unsigned int state)
4446 : {
4447 1 : return try_to_wake_up(p, state, 0);
4448 : }
4449 :
4450 : /*
4451 : * Perform scheduler related setup for a newly forked process p.
4452 : * p is forked by current.
4453 : *
4454 : * __sched_fork() is basic setup used by init_idle() too:
4455 : */
4456 383 : static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4457 : {
4458 383 : p->on_rq = 0;
4459 :
4460 383 : p->se.on_rq = 0;
4461 383 : p->se.exec_start = 0;
4462 383 : p->se.sum_exec_runtime = 0;
4463 383 : p->se.prev_sum_exec_runtime = 0;
4464 383 : p->se.nr_migrations = 0;
4465 383 : p->se.vruntime = 0;
4466 766 : INIT_LIST_HEAD(&p->se.group_node);
4467 :
4468 : #ifdef CONFIG_FAIR_GROUP_SCHED
4469 : p->se.cfs_rq = NULL;
4470 : #endif
4471 :
4472 : #ifdef CONFIG_SCHEDSTATS
4473 : /* Even if schedstat is disabled, there should not be garbage */
4474 : memset(&p->stats, 0, sizeof(p->stats));
4475 : #endif
4476 :
4477 383 : RB_CLEAR_NODE(&p->dl.rb_node);
4478 383 : init_dl_task_timer(&p->dl);
4479 383 : init_dl_inactive_task_timer(&p->dl);
4480 383 : __dl_clear_params(p);
4481 :
4482 766 : INIT_LIST_HEAD(&p->rt.run_list);
4483 383 : p->rt.timeout = 0;
4484 383 : p->rt.time_slice = sched_rr_timeslice;
4485 383 : p->rt.on_rq = 0;
4486 383 : p->rt.on_list = 0;
4487 :
4488 : #ifdef CONFIG_PREEMPT_NOTIFIERS
4489 : INIT_HLIST_HEAD(&p->preempt_notifiers);
4490 : #endif
4491 :
4492 : #ifdef CONFIG_COMPACTION
4493 383 : p->capture_control = NULL;
4494 : #endif
4495 383 : init_numa_balancing(clone_flags, p);
4496 : #ifdef CONFIG_SMP
4497 : p->wake_entry.u_flags = CSD_TYPE_TTWU;
4498 : p->migration_pending = NULL;
4499 : #endif
4500 383 : init_sched_mm_cid(p);
4501 383 : }
4502 :
4503 : DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4504 :
4505 : #ifdef CONFIG_NUMA_BALANCING
4506 :
4507 : int sysctl_numa_balancing_mode;
4508 :
4509 : static void __set_numabalancing_state(bool enabled)
4510 : {
4511 : if (enabled)
4512 : static_branch_enable(&sched_numa_balancing);
4513 : else
4514 : static_branch_disable(&sched_numa_balancing);
4515 : }
4516 :
4517 : void set_numabalancing_state(bool enabled)
4518 : {
4519 : if (enabled)
4520 : sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4521 : else
4522 : sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4523 : __set_numabalancing_state(enabled);
4524 : }
4525 :
4526 : #ifdef CONFIG_PROC_SYSCTL
4527 : static void reset_memory_tiering(void)
4528 : {
4529 : struct pglist_data *pgdat;
4530 :
4531 : for_each_online_pgdat(pgdat) {
4532 : pgdat->nbp_threshold = 0;
4533 : pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
4534 : pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
4535 : }
4536 : }
4537 :
4538 : static int sysctl_numa_balancing(struct ctl_table *table, int write,
4539 : void *buffer, size_t *lenp, loff_t *ppos)
4540 : {
4541 : struct ctl_table t;
4542 : int err;
4543 : int state = sysctl_numa_balancing_mode;
4544 :
4545 : if (write && !capable(CAP_SYS_ADMIN))
4546 : return -EPERM;
4547 :
4548 : t = *table;
4549 : t.data = &state;
4550 : err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4551 : if (err < 0)
4552 : return err;
4553 : if (write) {
4554 : if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4555 : (state & NUMA_BALANCING_MEMORY_TIERING))
4556 : reset_memory_tiering();
4557 : sysctl_numa_balancing_mode = state;
4558 : __set_numabalancing_state(state);
4559 : }
4560 : return err;
4561 : }
4562 : #endif
4563 : #endif
4564 :
4565 : #ifdef CONFIG_SCHEDSTATS
4566 :
4567 : DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4568 :
4569 : static void set_schedstats(bool enabled)
4570 : {
4571 : if (enabled)
4572 : static_branch_enable(&sched_schedstats);
4573 : else
4574 : static_branch_disable(&sched_schedstats);
4575 : }
4576 :
4577 : void force_schedstat_enabled(void)
4578 : {
4579 : if (!schedstat_enabled()) {
4580 : pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4581 : static_branch_enable(&sched_schedstats);
4582 : }
4583 : }
4584 :
4585 : static int __init setup_schedstats(char *str)
4586 : {
4587 : int ret = 0;
4588 : if (!str)
4589 : goto out;
4590 :
4591 : if (!strcmp(str, "enable")) {
4592 : set_schedstats(true);
4593 : ret = 1;
4594 : } else if (!strcmp(str, "disable")) {
4595 : set_schedstats(false);
4596 : ret = 1;
4597 : }
4598 : out:
4599 : if (!ret)
4600 : pr_warn("Unable to parse schedstats=\n");
4601 :
4602 : return ret;
4603 : }
4604 : __setup("schedstats=", setup_schedstats);
4605 :
4606 : #ifdef CONFIG_PROC_SYSCTL
4607 : static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4608 : size_t *lenp, loff_t *ppos)
4609 : {
4610 : struct ctl_table t;
4611 : int err;
4612 : int state = static_branch_likely(&sched_schedstats);
4613 :
4614 : if (write && !capable(CAP_SYS_ADMIN))
4615 : return -EPERM;
4616 :
4617 : t = *table;
4618 : t.data = &state;
4619 : err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4620 : if (err < 0)
4621 : return err;
4622 : if (write)
4623 : set_schedstats(state);
4624 : return err;
4625 : }
4626 : #endif /* CONFIG_PROC_SYSCTL */
4627 : #endif /* CONFIG_SCHEDSTATS */
4628 :
4629 : #ifdef CONFIG_SYSCTL
4630 : static struct ctl_table sched_core_sysctls[] = {
4631 : #ifdef CONFIG_SCHEDSTATS
4632 : {
4633 : .procname = "sched_schedstats",
4634 : .data = NULL,
4635 : .maxlen = sizeof(unsigned int),
4636 : .mode = 0644,
4637 : .proc_handler = sysctl_schedstats,
4638 : .extra1 = SYSCTL_ZERO,
4639 : .extra2 = SYSCTL_ONE,
4640 : },
4641 : #endif /* CONFIG_SCHEDSTATS */
4642 : #ifdef CONFIG_UCLAMP_TASK
4643 : {
4644 : .procname = "sched_util_clamp_min",
4645 : .data = &sysctl_sched_uclamp_util_min,
4646 : .maxlen = sizeof(unsigned int),
4647 : .mode = 0644,
4648 : .proc_handler = sysctl_sched_uclamp_handler,
4649 : },
4650 : {
4651 : .procname = "sched_util_clamp_max",
4652 : .data = &sysctl_sched_uclamp_util_max,
4653 : .maxlen = sizeof(unsigned int),
4654 : .mode = 0644,
4655 : .proc_handler = sysctl_sched_uclamp_handler,
4656 : },
4657 : {
4658 : .procname = "sched_util_clamp_min_rt_default",
4659 : .data = &sysctl_sched_uclamp_util_min_rt_default,
4660 : .maxlen = sizeof(unsigned int),
4661 : .mode = 0644,
4662 : .proc_handler = sysctl_sched_uclamp_handler,
4663 : },
4664 : #endif /* CONFIG_UCLAMP_TASK */
4665 : #ifdef CONFIG_NUMA_BALANCING
4666 : {
4667 : .procname = "numa_balancing",
4668 : .data = NULL, /* filled in by handler */
4669 : .maxlen = sizeof(unsigned int),
4670 : .mode = 0644,
4671 : .proc_handler = sysctl_numa_balancing,
4672 : .extra1 = SYSCTL_ZERO,
4673 : .extra2 = SYSCTL_FOUR,
4674 : },
4675 : #endif /* CONFIG_NUMA_BALANCING */
4676 : {}
4677 : };
4678 1 : static int __init sched_core_sysctl_init(void)
4679 : {
4680 1 : register_sysctl_init("kernel", sched_core_sysctls);
4681 1 : return 0;
4682 : }
4683 : late_initcall(sched_core_sysctl_init);
4684 : #endif /* CONFIG_SYSCTL */
4685 :
4686 : /*
4687 : * fork()/clone()-time setup:
4688 : */
4689 382 : int sched_fork(unsigned long clone_flags, struct task_struct *p)
4690 : {
4691 382 : __sched_fork(clone_flags, p);
4692 : /*
4693 : * We mark the process as NEW here. This guarantees that
4694 : * nobody will actually run it, and a signal or other external
4695 : * event cannot wake it up and insert it on the runqueue either.
4696 : */
4697 382 : p->__state = TASK_NEW;
4698 :
4699 : /*
4700 : * Make sure we do not leak PI boosting priority to the child.
4701 : */
4702 382 : p->prio = current->normal_prio;
4703 :
4704 382 : uclamp_fork(p);
4705 :
4706 : /*
4707 : * Revert to default priority/policy on fork if requested.
4708 : */
4709 382 : if (unlikely(p->sched_reset_on_fork)) {
4710 0 : if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4711 0 : p->policy = SCHED_NORMAL;
4712 0 : p->static_prio = NICE_TO_PRIO(0);
4713 0 : p->rt_priority = 0;
4714 0 : } else if (PRIO_TO_NICE(p->static_prio) < 0)
4715 0 : p->static_prio = NICE_TO_PRIO(0);
4716 :
4717 0 : p->prio = p->normal_prio = p->static_prio;
4718 0 : set_load_weight(p, false);
4719 :
4720 : /*
4721 : * We don't need the reset flag anymore after the fork. It has
4722 : * fulfilled its duty:
4723 : */
4724 0 : p->sched_reset_on_fork = 0;
4725 : }
4726 :
4727 764 : if (dl_prio(p->prio))
4728 : return -EAGAIN;
4729 764 : else if (rt_prio(p->prio))
4730 0 : p->sched_class = &rt_sched_class;
4731 : else
4732 382 : p->sched_class = &fair_sched_class;
4733 :
4734 382 : init_entity_runnable_average(&p->se);
4735 :
4736 :
4737 : #ifdef CONFIG_SCHED_INFO
4738 : if (likely(sched_info_on()))
4739 : memset(&p->sched_info, 0, sizeof(p->sched_info));
4740 : #endif
4741 : #if defined(CONFIG_SMP)
4742 : p->on_cpu = 0;
4743 : #endif
4744 382 : init_task_preempt_count(p);
4745 : #ifdef CONFIG_SMP
4746 : plist_node_init(&p->pushable_tasks, MAX_PRIO);
4747 : RB_CLEAR_NODE(&p->pushable_dl_tasks);
4748 : #endif
4749 382 : return 0;
4750 : }
4751 :
4752 382 : void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4753 : {
4754 : unsigned long flags;
4755 :
4756 : /*
4757 : * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4758 : * required yet, but lockdep gets upset if rules are violated.
4759 : */
4760 382 : raw_spin_lock_irqsave(&p->pi_lock, flags);
4761 : #ifdef CONFIG_CGROUP_SCHED
4762 : if (1) {
4763 : struct task_group *tg;
4764 : tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4765 : struct task_group, css);
4766 : tg = autogroup_task_group(p, tg);
4767 : p->sched_task_group = tg;
4768 : }
4769 : #endif
4770 382 : rseq_migrate(p);
4771 : /*
4772 : * We're setting the CPU for the first time, we don't migrate,
4773 : * so use __set_task_cpu().
4774 : */
4775 382 : __set_task_cpu(p, smp_processor_id());
4776 382 : if (p->sched_class->task_fork)
4777 382 : p->sched_class->task_fork(p);
4778 764 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4779 382 : }
4780 :
4781 382 : void sched_post_fork(struct task_struct *p)
4782 : {
4783 382 : uclamp_post_fork(p);
4784 382 : }
4785 :
4786 3 : unsigned long to_ratio(u64 period, u64 runtime)
4787 : {
4788 3 : if (runtime == RUNTIME_INF)
4789 : return BW_UNIT;
4790 :
4791 : /*
4792 : * Doing this here saves a lot of checks in all
4793 : * the calling paths, and returning zero seems
4794 : * safe for them anyway.
4795 : */
4796 3 : if (period == 0)
4797 : return 0;
4798 :
4799 6 : return div64_u64(runtime << BW_SHIFT, period);
4800 : }
4801 :
4802 : /*
4803 : * wake_up_new_task - wake up a newly created task for the first time.
4804 : *
4805 : * This function will do some initial scheduler statistics housekeeping
4806 : * that must be done for every newly created context, then puts the task
4807 : * on the runqueue and wakes it.
4808 : */
4809 382 : void wake_up_new_task(struct task_struct *p)
4810 : {
4811 : struct rq_flags rf;
4812 : struct rq *rq;
4813 :
4814 382 : raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4815 382 : WRITE_ONCE(p->__state, TASK_RUNNING);
4816 : #ifdef CONFIG_SMP
4817 : /*
4818 : * Fork balancing, do it here and not earlier because:
4819 : * - cpus_ptr can change in the fork path
4820 : * - any previously selected CPU might disappear through hotplug
4821 : *
4822 : * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4823 : * as we're not fully set-up yet.
4824 : */
4825 : p->recent_used_cpu = task_cpu(p);
4826 : rseq_migrate(p);
4827 : __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4828 : #endif
4829 382 : rq = __task_rq_lock(p, &rf);
4830 382 : update_rq_clock(rq);
4831 382 : post_init_entity_util_avg(p);
4832 :
4833 382 : activate_task(rq, p, ENQUEUE_NOCLOCK);
4834 382 : trace_sched_wakeup_new(p);
4835 382 : check_preempt_curr(rq, p, WF_FORK);
4836 : #ifdef CONFIG_SMP
4837 : if (p->sched_class->task_woken) {
4838 : /*
4839 : * Nothing relies on rq->lock after this, so it's fine to
4840 : * drop it.
4841 : */
4842 : rq_unpin_lock(rq, &rf);
4843 : p->sched_class->task_woken(rq, p);
4844 : rq_repin_lock(rq, &rf);
4845 : }
4846 : #endif
4847 764 : task_rq_unlock(rq, p, &rf);
4848 382 : }
4849 :
4850 : #ifdef CONFIG_PREEMPT_NOTIFIERS
4851 :
4852 : static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4853 :
4854 : void preempt_notifier_inc(void)
4855 : {
4856 : static_branch_inc(&preempt_notifier_key);
4857 : }
4858 : EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4859 :
4860 : void preempt_notifier_dec(void)
4861 : {
4862 : static_branch_dec(&preempt_notifier_key);
4863 : }
4864 : EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4865 :
4866 : /**
4867 : * preempt_notifier_register - tell me when current is being preempted & rescheduled
4868 : * @notifier: notifier struct to register
4869 : */
4870 : void preempt_notifier_register(struct preempt_notifier *notifier)
4871 : {
4872 : if (!static_branch_unlikely(&preempt_notifier_key))
4873 : WARN(1, "registering preempt_notifier while notifiers disabled\n");
4874 :
4875 : hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4876 : }
4877 : EXPORT_SYMBOL_GPL(preempt_notifier_register);
4878 :
4879 : /**
4880 : * preempt_notifier_unregister - no longer interested in preemption notifications
4881 : * @notifier: notifier struct to unregister
4882 : *
4883 : * This is *not* safe to call from within a preemption notifier.
4884 : */
4885 : void preempt_notifier_unregister(struct preempt_notifier *notifier)
4886 : {
4887 : hlist_del(¬ifier->link);
4888 : }
4889 : EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4890 :
4891 : static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4892 : {
4893 : struct preempt_notifier *notifier;
4894 :
4895 : hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4896 : notifier->ops->sched_in(notifier, raw_smp_processor_id());
4897 : }
4898 :
4899 : static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4900 : {
4901 : if (static_branch_unlikely(&preempt_notifier_key))
4902 : __fire_sched_in_preempt_notifiers(curr);
4903 : }
4904 :
4905 : static void
4906 : __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4907 : struct task_struct *next)
4908 : {
4909 : struct preempt_notifier *notifier;
4910 :
4911 : hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4912 : notifier->ops->sched_out(notifier, next);
4913 : }
4914 :
4915 : static __always_inline void
4916 : fire_sched_out_preempt_notifiers(struct task_struct *curr,
4917 : struct task_struct *next)
4918 : {
4919 : if (static_branch_unlikely(&preempt_notifier_key))
4920 : __fire_sched_out_preempt_notifiers(curr, next);
4921 : }
4922 :
4923 : #else /* !CONFIG_PREEMPT_NOTIFIERS */
4924 :
4925 : static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4926 : {
4927 : }
4928 :
4929 : static inline void
4930 : fire_sched_out_preempt_notifiers(struct task_struct *curr,
4931 : struct task_struct *next)
4932 : {
4933 : }
4934 :
4935 : #endif /* CONFIG_PREEMPT_NOTIFIERS */
4936 :
4937 : static inline void prepare_task(struct task_struct *next)
4938 : {
4939 : #ifdef CONFIG_SMP
4940 : /*
4941 : * Claim the task as running, we do this before switching to it
4942 : * such that any running task will have this set.
4943 : *
4944 : * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4945 : * its ordering comment.
4946 : */
4947 : WRITE_ONCE(next->on_cpu, 1);
4948 : #endif
4949 : }
4950 :
4951 : static inline void finish_task(struct task_struct *prev)
4952 : {
4953 : #ifdef CONFIG_SMP
4954 : /*
4955 : * This must be the very last reference to @prev from this CPU. After
4956 : * p->on_cpu is cleared, the task can be moved to a different CPU. We
4957 : * must ensure this doesn't happen until the switch is completely
4958 : * finished.
4959 : *
4960 : * In particular, the load of prev->state in finish_task_switch() must
4961 : * happen before this.
4962 : *
4963 : * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4964 : */
4965 : smp_store_release(&prev->on_cpu, 0);
4966 : #endif
4967 : }
4968 :
4969 : #ifdef CONFIG_SMP
4970 :
4971 : static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
4972 : {
4973 : void (*func)(struct rq *rq);
4974 : struct balance_callback *next;
4975 :
4976 : lockdep_assert_rq_held(rq);
4977 :
4978 : while (head) {
4979 : func = (void (*)(struct rq *))head->func;
4980 : next = head->next;
4981 : head->next = NULL;
4982 : head = next;
4983 :
4984 : func(rq);
4985 : }
4986 : }
4987 :
4988 : static void balance_push(struct rq *rq);
4989 :
4990 : /*
4991 : * balance_push_callback is a right abuse of the callback interface and plays
4992 : * by significantly different rules.
4993 : *
4994 : * Where the normal balance_callback's purpose is to be ran in the same context
4995 : * that queued it (only later, when it's safe to drop rq->lock again),
4996 : * balance_push_callback is specifically targeted at __schedule().
4997 : *
4998 : * This abuse is tolerated because it places all the unlikely/odd cases behind
4999 : * a single test, namely: rq->balance_callback == NULL.
5000 : */
5001 : struct balance_callback balance_push_callback = {
5002 : .next = NULL,
5003 : .func = balance_push,
5004 : };
5005 :
5006 : static inline struct balance_callback *
5007 : __splice_balance_callbacks(struct rq *rq, bool split)
5008 : {
5009 : struct balance_callback *head = rq->balance_callback;
5010 :
5011 : if (likely(!head))
5012 : return NULL;
5013 :
5014 : lockdep_assert_rq_held(rq);
5015 : /*
5016 : * Must not take balance_push_callback off the list when
5017 : * splice_balance_callbacks() and balance_callbacks() are not
5018 : * in the same rq->lock section.
5019 : *
5020 : * In that case it would be possible for __schedule() to interleave
5021 : * and observe the list empty.
5022 : */
5023 : if (split && head == &balance_push_callback)
5024 : head = NULL;
5025 : else
5026 : rq->balance_callback = NULL;
5027 :
5028 : return head;
5029 : }
5030 :
5031 : static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5032 : {
5033 : return __splice_balance_callbacks(rq, true);
5034 : }
5035 :
5036 : static void __balance_callbacks(struct rq *rq)
5037 : {
5038 : do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
5039 : }
5040 :
5041 : static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5042 : {
5043 : unsigned long flags;
5044 :
5045 : if (unlikely(head)) {
5046 : raw_spin_rq_lock_irqsave(rq, flags);
5047 : do_balance_callbacks(rq, head);
5048 : raw_spin_rq_unlock_irqrestore(rq, flags);
5049 : }
5050 : }
5051 :
5052 : #else
5053 :
5054 : static inline void __balance_callbacks(struct rq *rq)
5055 : {
5056 : }
5057 :
5058 : static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5059 : {
5060 : return NULL;
5061 : }
5062 :
5063 : static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5064 : {
5065 : }
5066 :
5067 : #endif
5068 :
5069 : static inline void
5070 : prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
5071 : {
5072 : /*
5073 : * Since the runqueue lock will be released by the next
5074 : * task (which is an invalid locking op but in the case
5075 : * of the scheduler it's an obvious special-case), so we
5076 : * do an early lockdep release here:
5077 : */
5078 : rq_unpin_lock(rq, rf);
5079 : spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
5080 : #ifdef CONFIG_DEBUG_SPINLOCK
5081 : /* this is a valid case when another task releases the spinlock */
5082 : rq_lockp(rq)->owner = next;
5083 : #endif
5084 : }
5085 :
5086 : static inline void finish_lock_switch(struct rq *rq)
5087 : {
5088 : /*
5089 : * If we are tracking spinlock dependencies then we have to
5090 : * fix up the runqueue lock - which gets 'carried over' from
5091 : * prev into current:
5092 : */
5093 : spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
5094 2513 : __balance_callbacks(rq);
5095 2513 : raw_spin_rq_unlock_irq(rq);
5096 : }
5097 :
5098 : /*
5099 : * NOP if the arch has not defined these:
5100 : */
5101 :
5102 : #ifndef prepare_arch_switch
5103 : # define prepare_arch_switch(next) do { } while (0)
5104 : #endif
5105 :
5106 : #ifndef finish_arch_post_lock_switch
5107 : # define finish_arch_post_lock_switch() do { } while (0)
5108 : #endif
5109 :
5110 : static inline void kmap_local_sched_out(void)
5111 : {
5112 : #ifdef CONFIG_KMAP_LOCAL
5113 : if (unlikely(current->kmap_ctrl.idx))
5114 : __kmap_local_sched_out();
5115 : #endif
5116 : }
5117 :
5118 : static inline void kmap_local_sched_in(void)
5119 : {
5120 : #ifdef CONFIG_KMAP_LOCAL
5121 : if (unlikely(current->kmap_ctrl.idx))
5122 : __kmap_local_sched_in();
5123 : #endif
5124 : }
5125 :
5126 : /**
5127 : * prepare_task_switch - prepare to switch tasks
5128 : * @rq: the runqueue preparing to switch
5129 : * @prev: the current task that is being switched out
5130 : * @next: the task we are going to switch to.
5131 : *
5132 : * This is called with the rq lock held and interrupts off. It must
5133 : * be paired with a subsequent finish_task_switch after the context
5134 : * switch.
5135 : *
5136 : * prepare_task_switch sets up locking and calls architecture specific
5137 : * hooks.
5138 : */
5139 : static inline void
5140 : prepare_task_switch(struct rq *rq, struct task_struct *prev,
5141 : struct task_struct *next)
5142 : {
5143 : kcov_prepare_switch(prev);
5144 : sched_info_switch(rq, prev, next);
5145 : perf_event_task_sched_out(prev, next);
5146 : rseq_preempt(prev);
5147 : fire_sched_out_preempt_notifiers(prev, next);
5148 : kmap_local_sched_out();
5149 : prepare_task(next);
5150 : prepare_arch_switch(next);
5151 : }
5152 :
5153 : /**
5154 : * finish_task_switch - clean up after a task-switch
5155 : * @prev: the thread we just switched away from.
5156 : *
5157 : * finish_task_switch must be called after the context switch, paired
5158 : * with a prepare_task_switch call before the context switch.
5159 : * finish_task_switch will reconcile locking set up by prepare_task_switch,
5160 : * and do any other architecture-specific cleanup actions.
5161 : *
5162 : * Note that we may have delayed dropping an mm in context_switch(). If
5163 : * so, we finish that here outside of the runqueue lock. (Doing it
5164 : * with the lock held can cause deadlocks; see schedule() for
5165 : * details.)
5166 : *
5167 : * The context switch have flipped the stack from under us and restored the
5168 : * local variables which were saved when this task called schedule() in the
5169 : * past. prev == current is still correct but we need to recalculate this_rq
5170 : * because prev may have moved to another CPU.
5171 : */
5172 2513 : static struct rq *finish_task_switch(struct task_struct *prev)
5173 : __releases(rq->lock)
5174 : {
5175 2513 : struct rq *rq = this_rq();
5176 2513 : struct mm_struct *mm = rq->prev_mm;
5177 : unsigned int prev_state;
5178 :
5179 : /*
5180 : * The previous task will have left us with a preempt_count of 2
5181 : * because it left us after:
5182 : *
5183 : * schedule()
5184 : * preempt_disable(); // 1
5185 : * __schedule()
5186 : * raw_spin_lock_irq(&rq->lock) // 2
5187 : *
5188 : * Also, see FORK_PREEMPT_COUNT.
5189 : */
5190 2513 : if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5191 : "corrupted preempt_count: %s/%d/0x%x\n",
5192 : current->comm, current->pid, preempt_count()))
5193 : preempt_count_set(FORK_PREEMPT_COUNT);
5194 :
5195 2513 : rq->prev_mm = NULL;
5196 :
5197 : /*
5198 : * A task struct has one reference for the use as "current".
5199 : * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5200 : * schedule one last time. The schedule call will never return, and
5201 : * the scheduled task must drop that reference.
5202 : *
5203 : * We must observe prev->state before clearing prev->on_cpu (in
5204 : * finish_task), otherwise a concurrent wakeup can get prev
5205 : * running on another CPU and we could rave with its RUNNING -> DEAD
5206 : * transition, resulting in a double drop.
5207 : */
5208 2513 : prev_state = READ_ONCE(prev->__state);
5209 2513 : vtime_task_switch(prev);
5210 2513 : perf_event_task_sched_in(prev, current);
5211 2513 : finish_task(prev);
5212 : tick_nohz_task_switch();
5213 2513 : finish_lock_switch(rq);
5214 : finish_arch_post_lock_switch();
5215 2513 : kcov_finish_switch(current);
5216 : /*
5217 : * kmap_local_sched_out() is invoked with rq::lock held and
5218 : * interrupts disabled. There is no requirement for that, but the
5219 : * sched out code does not have an interrupt enabled section.
5220 : * Restoring the maps on sched in does not require interrupts being
5221 : * disabled either.
5222 : */
5223 : kmap_local_sched_in();
5224 :
5225 2513 : fire_sched_in_preempt_notifiers(current);
5226 : /*
5227 : * When switching through a kernel thread, the loop in
5228 : * membarrier_{private,global}_expedited() may have observed that
5229 : * kernel thread and not issued an IPI. It is therefore possible to
5230 : * schedule between user->kernel->user threads without passing though
5231 : * switch_mm(). Membarrier requires a barrier after storing to
5232 : * rq->curr, before returning to userspace, so provide them here:
5233 : *
5234 : * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5235 : * provided by mmdrop_lazy_tlb(),
5236 : * - a sync_core for SYNC_CORE.
5237 : */
5238 2513 : if (mm) {
5239 0 : membarrier_mm_sync_core_before_usermode(mm);
5240 : mmdrop_lazy_tlb_sched(mm);
5241 : }
5242 :
5243 2513 : if (unlikely(prev_state == TASK_DEAD)) {
5244 367 : if (prev->sched_class->task_dead)
5245 0 : prev->sched_class->task_dead(prev);
5246 :
5247 : /* Task is done with its stack. */
5248 367 : put_task_stack(prev);
5249 :
5250 367 : put_task_struct_rcu_user(prev);
5251 : }
5252 :
5253 2513 : return rq;
5254 : }
5255 :
5256 : /**
5257 : * schedule_tail - first thing a freshly forked thread must call.
5258 : * @prev: the thread we just switched away from.
5259 : */
5260 382 : asmlinkage __visible void schedule_tail(struct task_struct *prev)
5261 : __releases(rq->lock)
5262 : {
5263 : /*
5264 : * New tasks start with FORK_PREEMPT_COUNT, see there and
5265 : * finish_task_switch() for details.
5266 : *
5267 : * finish_task_switch() will drop rq->lock() and lower preempt_count
5268 : * and the preempt_enable() will end up enabling preemption (on
5269 : * PREEMPT_COUNT kernels).
5270 : */
5271 :
5272 382 : finish_task_switch(prev);
5273 382 : preempt_enable();
5274 :
5275 382 : if (current->set_child_tid)
5276 0 : put_user(task_pid_vnr(current), current->set_child_tid);
5277 :
5278 382 : calculate_sigpending();
5279 382 : }
5280 :
5281 : /*
5282 : * context_switch - switch to the new MM and the new thread's register state.
5283 : */
5284 : static __always_inline struct rq *
5285 : context_switch(struct rq *rq, struct task_struct *prev,
5286 : struct task_struct *next, struct rq_flags *rf)
5287 : {
5288 2513 : prepare_task_switch(rq, prev, next);
5289 :
5290 : /*
5291 : * For paravirt, this is coupled with an exit in switch_to to
5292 : * combine the page table reload and the switch backend into
5293 : * one hypercall.
5294 : */
5295 : arch_start_context_switch(prev);
5296 :
5297 : /*
5298 : * kernel -> kernel lazy + transfer active
5299 : * user -> kernel lazy + mmgrab_lazy_tlb() active
5300 : *
5301 : * kernel -> user switch + mmdrop_lazy_tlb() active
5302 : * user -> user switch
5303 : *
5304 : * switch_mm_cid() needs to be updated if the barriers provided
5305 : * by context_switch() are modified.
5306 : */
5307 2513 : if (!next->mm) { // to kernel
5308 2513 : enter_lazy_tlb(prev->active_mm, next);
5309 :
5310 2513 : next->active_mm = prev->active_mm;
5311 2513 : if (prev->mm) // from user
5312 0 : mmgrab_lazy_tlb(prev->active_mm);
5313 : else
5314 2513 : prev->active_mm = NULL;
5315 : } else { // to user
5316 0 : membarrier_switch_mm(rq, prev->active_mm, next->mm);
5317 : /*
5318 : * sys_membarrier() requires an smp_mb() between setting
5319 : * rq->curr / membarrier_switch_mm() and returning to userspace.
5320 : *
5321 : * The below provides this either through switch_mm(), or in
5322 : * case 'prev->active_mm == next->mm' through
5323 : * finish_task_switch()'s mmdrop().
5324 : */
5325 0 : switch_mm_irqs_off(prev->active_mm, next->mm, next);
5326 0 : lru_gen_use_mm(next->mm);
5327 :
5328 0 : if (!prev->mm) { // from kernel
5329 : /* will mmdrop_lazy_tlb() in finish_task_switch(). */
5330 0 : rq->prev_mm = prev->active_mm;
5331 0 : prev->active_mm = NULL;
5332 : }
5333 : }
5334 :
5335 : /* switch_mm_cid() requires the memory barriers above. */
5336 2513 : switch_mm_cid(rq, prev, next);
5337 :
5338 2513 : rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5339 :
5340 2513 : prepare_lock_switch(rq, next, rf);
5341 :
5342 : /* Here we just switch the register state and the stack. */
5343 2513 : switch_to(prev, next, prev);
5344 2131 : barrier();
5345 :
5346 2131 : return finish_task_switch(prev);
5347 : }
5348 :
5349 : /*
5350 : * nr_running and nr_context_switches:
5351 : *
5352 : * externally visible scheduler statistics: current number of runnable
5353 : * threads, total number of context switches performed since bootup.
5354 : */
5355 0 : unsigned int nr_running(void)
5356 : {
5357 0 : unsigned int i, sum = 0;
5358 :
5359 0 : for_each_online_cpu(i)
5360 0 : sum += cpu_rq(i)->nr_running;
5361 :
5362 0 : return sum;
5363 : }
5364 :
5365 : /*
5366 : * Check if only the current task is running on the CPU.
5367 : *
5368 : * Caution: this function does not check that the caller has disabled
5369 : * preemption, thus the result might have a time-of-check-to-time-of-use
5370 : * race. The caller is responsible to use it correctly, for example:
5371 : *
5372 : * - from a non-preemptible section (of course)
5373 : *
5374 : * - from a thread that is bound to a single CPU
5375 : *
5376 : * - in a loop with very short iterations (e.g. a polling loop)
5377 : */
5378 0 : bool single_task_running(void)
5379 : {
5380 0 : return raw_rq()->nr_running == 1;
5381 : }
5382 : EXPORT_SYMBOL(single_task_running);
5383 :
5384 0 : unsigned long long nr_context_switches_cpu(int cpu)
5385 : {
5386 0 : return cpu_rq(cpu)->nr_switches;
5387 : }
5388 :
5389 0 : unsigned long long nr_context_switches(void)
5390 : {
5391 : int i;
5392 0 : unsigned long long sum = 0;
5393 :
5394 0 : for_each_possible_cpu(i)
5395 0 : sum += cpu_rq(i)->nr_switches;
5396 :
5397 0 : return sum;
5398 : }
5399 :
5400 : /*
5401 : * Consumers of these two interfaces, like for example the cpuidle menu
5402 : * governor, are using nonsensical data. Preferring shallow idle state selection
5403 : * for a CPU that has IO-wait which might not even end up running the task when
5404 : * it does become runnable.
5405 : */
5406 :
5407 0 : unsigned int nr_iowait_cpu(int cpu)
5408 : {
5409 0 : return atomic_read(&cpu_rq(cpu)->nr_iowait);
5410 : }
5411 :
5412 : /*
5413 : * IO-wait accounting, and how it's mostly bollocks (on SMP).
5414 : *
5415 : * The idea behind IO-wait account is to account the idle time that we could
5416 : * have spend running if it were not for IO. That is, if we were to improve the
5417 : * storage performance, we'd have a proportional reduction in IO-wait time.
5418 : *
5419 : * This all works nicely on UP, where, when a task blocks on IO, we account
5420 : * idle time as IO-wait, because if the storage were faster, it could've been
5421 : * running and we'd not be idle.
5422 : *
5423 : * This has been extended to SMP, by doing the same for each CPU. This however
5424 : * is broken.
5425 : *
5426 : * Imagine for instance the case where two tasks block on one CPU, only the one
5427 : * CPU will have IO-wait accounted, while the other has regular idle. Even
5428 : * though, if the storage were faster, both could've ran at the same time,
5429 : * utilising both CPUs.
5430 : *
5431 : * This means, that when looking globally, the current IO-wait accounting on
5432 : * SMP is a lower bound, by reason of under accounting.
5433 : *
5434 : * Worse, since the numbers are provided per CPU, they are sometimes
5435 : * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5436 : * associated with any one particular CPU, it can wake to another CPU than it
5437 : * blocked on. This means the per CPU IO-wait number is meaningless.
5438 : *
5439 : * Task CPU affinities can make all that even more 'interesting'.
5440 : */
5441 :
5442 0 : unsigned int nr_iowait(void)
5443 : {
5444 0 : unsigned int i, sum = 0;
5445 :
5446 0 : for_each_possible_cpu(i)
5447 0 : sum += nr_iowait_cpu(i);
5448 :
5449 0 : return sum;
5450 : }
5451 :
5452 : #ifdef CONFIG_SMP
5453 :
5454 : /*
5455 : * sched_exec - execve() is a valuable balancing opportunity, because at
5456 : * this point the task has the smallest effective memory and cache footprint.
5457 : */
5458 : void sched_exec(void)
5459 : {
5460 : struct task_struct *p = current;
5461 : unsigned long flags;
5462 : int dest_cpu;
5463 :
5464 : raw_spin_lock_irqsave(&p->pi_lock, flags);
5465 : dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5466 : if (dest_cpu == smp_processor_id())
5467 : goto unlock;
5468 :
5469 : if (likely(cpu_active(dest_cpu))) {
5470 : struct migration_arg arg = { p, dest_cpu };
5471 :
5472 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5473 : stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5474 : return;
5475 : }
5476 : unlock:
5477 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5478 : }
5479 :
5480 : #endif
5481 :
5482 : DEFINE_PER_CPU(struct kernel_stat, kstat);
5483 : DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5484 :
5485 : EXPORT_PER_CPU_SYMBOL(kstat);
5486 : EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5487 :
5488 : /*
5489 : * The function fair_sched_class.update_curr accesses the struct curr
5490 : * and its field curr->exec_start; when called from task_sched_runtime(),
5491 : * we observe a high rate of cache misses in practice.
5492 : * Prefetching this data results in improved performance.
5493 : */
5494 : static inline void prefetch_curr_exec_start(struct task_struct *p)
5495 : {
5496 : #ifdef CONFIG_FAIR_GROUP_SCHED
5497 : struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5498 : #else
5499 0 : struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5500 : #endif
5501 0 : prefetch(curr);
5502 0 : prefetch(&curr->exec_start);
5503 : }
5504 :
5505 : /*
5506 : * Return accounted runtime for the task.
5507 : * In case the task is currently running, return the runtime plus current's
5508 : * pending runtime that have not been accounted yet.
5509 : */
5510 0 : unsigned long long task_sched_runtime(struct task_struct *p)
5511 : {
5512 : struct rq_flags rf;
5513 : struct rq *rq;
5514 : u64 ns;
5515 :
5516 : #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5517 : /*
5518 : * 64-bit doesn't need locks to atomically read a 64-bit value.
5519 : * So we have a optimization chance when the task's delta_exec is 0.
5520 : * Reading ->on_cpu is racy, but this is ok.
5521 : *
5522 : * If we race with it leaving CPU, we'll take a lock. So we're correct.
5523 : * If we race with it entering CPU, unaccounted time is 0. This is
5524 : * indistinguishable from the read occurring a few cycles earlier.
5525 : * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5526 : * been accounted, so we're correct here as well.
5527 : */
5528 : if (!p->on_cpu || !task_on_rq_queued(p))
5529 : return p->se.sum_exec_runtime;
5530 : #endif
5531 :
5532 0 : rq = task_rq_lock(p, &rf);
5533 : /*
5534 : * Must be ->curr _and_ ->on_rq. If dequeued, we would
5535 : * project cycles that may never be accounted to this
5536 : * thread, breaking clock_gettime().
5537 : */
5538 0 : if (task_current(rq, p) && task_on_rq_queued(p)) {
5539 0 : prefetch_curr_exec_start(p);
5540 0 : update_rq_clock(rq);
5541 0 : p->sched_class->update_curr(rq);
5542 : }
5543 0 : ns = p->se.sum_exec_runtime;
5544 0 : task_rq_unlock(rq, p, &rf);
5545 :
5546 0 : return ns;
5547 : }
5548 :
5549 : #ifdef CONFIG_SCHED_DEBUG
5550 : static u64 cpu_resched_latency(struct rq *rq)
5551 : {
5552 : int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5553 : u64 resched_latency, now = rq_clock(rq);
5554 : static bool warned_once;
5555 :
5556 : if (sysctl_resched_latency_warn_once && warned_once)
5557 : return 0;
5558 :
5559 : if (!need_resched() || !latency_warn_ms)
5560 : return 0;
5561 :
5562 : if (system_state == SYSTEM_BOOTING)
5563 : return 0;
5564 :
5565 : if (!rq->last_seen_need_resched_ns) {
5566 : rq->last_seen_need_resched_ns = now;
5567 : rq->ticks_without_resched = 0;
5568 : return 0;
5569 : }
5570 :
5571 : rq->ticks_without_resched++;
5572 : resched_latency = now - rq->last_seen_need_resched_ns;
5573 : if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5574 : return 0;
5575 :
5576 : warned_once = true;
5577 :
5578 : return resched_latency;
5579 : }
5580 :
5581 : static int __init setup_resched_latency_warn_ms(char *str)
5582 : {
5583 : long val;
5584 :
5585 : if ((kstrtol(str, 0, &val))) {
5586 : pr_warn("Unable to set resched_latency_warn_ms\n");
5587 : return 1;
5588 : }
5589 :
5590 : sysctl_resched_latency_warn_ms = val;
5591 : return 1;
5592 : }
5593 : __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5594 : #else
5595 : static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5596 : #endif /* CONFIG_SCHED_DEBUG */
5597 :
5598 : /*
5599 : * This function gets called by the timer code, with HZ frequency.
5600 : * We call it with interrupts disabled.
5601 : */
5602 2943 : void scheduler_tick(void)
5603 : {
5604 2943 : int cpu = smp_processor_id();
5605 2943 : struct rq *rq = cpu_rq(cpu);
5606 2943 : struct task_struct *curr = rq->curr;
5607 : struct rq_flags rf;
5608 : unsigned long thermal_pressure;
5609 : u64 resched_latency;
5610 :
5611 2943 : if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5612 : arch_scale_freq_tick();
5613 :
5614 : sched_clock_tick();
5615 :
5616 2943 : rq_lock(rq, &rf);
5617 :
5618 2943 : update_rq_clock(rq);
5619 2943 : thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5620 2943 : update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5621 2943 : curr->sched_class->task_tick(rq, curr, 0);
5622 : if (sched_feat(LATENCY_WARN))
5623 : resched_latency = cpu_resched_latency(rq);
5624 2943 : calc_global_load_tick(rq);
5625 2943 : sched_core_tick(rq);
5626 2943 : task_tick_mm_cid(rq, curr);
5627 :
5628 2943 : rq_unlock(rq, &rf);
5629 :
5630 : if (sched_feat(LATENCY_WARN) && resched_latency)
5631 : resched_latency_warn(cpu, resched_latency);
5632 :
5633 : perf_event_task_tick();
5634 :
5635 : #ifdef CONFIG_SMP
5636 : rq->idle_balance = idle_cpu(cpu);
5637 : trigger_load_balance(rq);
5638 : #endif
5639 2943 : }
5640 :
5641 : #ifdef CONFIG_NO_HZ_FULL
5642 :
5643 : struct tick_work {
5644 : int cpu;
5645 : atomic_t state;
5646 : struct delayed_work work;
5647 : };
5648 : /* Values for ->state, see diagram below. */
5649 : #define TICK_SCHED_REMOTE_OFFLINE 0
5650 : #define TICK_SCHED_REMOTE_OFFLINING 1
5651 : #define TICK_SCHED_REMOTE_RUNNING 2
5652 :
5653 : /*
5654 : * State diagram for ->state:
5655 : *
5656 : *
5657 : * TICK_SCHED_REMOTE_OFFLINE
5658 : * | ^
5659 : * | |
5660 : * | | sched_tick_remote()
5661 : * | |
5662 : * | |
5663 : * +--TICK_SCHED_REMOTE_OFFLINING
5664 : * | ^
5665 : * | |
5666 : * sched_tick_start() | | sched_tick_stop()
5667 : * | |
5668 : * V |
5669 : * TICK_SCHED_REMOTE_RUNNING
5670 : *
5671 : *
5672 : * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5673 : * and sched_tick_start() are happy to leave the state in RUNNING.
5674 : */
5675 :
5676 : static struct tick_work __percpu *tick_work_cpu;
5677 :
5678 : static void sched_tick_remote(struct work_struct *work)
5679 : {
5680 : struct delayed_work *dwork = to_delayed_work(work);
5681 : struct tick_work *twork = container_of(dwork, struct tick_work, work);
5682 : int cpu = twork->cpu;
5683 : struct rq *rq = cpu_rq(cpu);
5684 : struct task_struct *curr;
5685 : struct rq_flags rf;
5686 : u64 delta;
5687 : int os;
5688 :
5689 : /*
5690 : * Handle the tick only if it appears the remote CPU is running in full
5691 : * dynticks mode. The check is racy by nature, but missing a tick or
5692 : * having one too much is no big deal because the scheduler tick updates
5693 : * statistics and checks timeslices in a time-independent way, regardless
5694 : * of when exactly it is running.
5695 : */
5696 : if (!tick_nohz_tick_stopped_cpu(cpu))
5697 : goto out_requeue;
5698 :
5699 : rq_lock_irq(rq, &rf);
5700 : curr = rq->curr;
5701 : if (cpu_is_offline(cpu))
5702 : goto out_unlock;
5703 :
5704 : update_rq_clock(rq);
5705 :
5706 : if (!is_idle_task(curr)) {
5707 : /*
5708 : * Make sure the next tick runs within a reasonable
5709 : * amount of time.
5710 : */
5711 : delta = rq_clock_task(rq) - curr->se.exec_start;
5712 : WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5713 : }
5714 : curr->sched_class->task_tick(rq, curr, 0);
5715 :
5716 : calc_load_nohz_remote(rq);
5717 : out_unlock:
5718 : rq_unlock_irq(rq, &rf);
5719 : out_requeue:
5720 :
5721 : /*
5722 : * Run the remote tick once per second (1Hz). This arbitrary
5723 : * frequency is large enough to avoid overload but short enough
5724 : * to keep scheduler internal stats reasonably up to date. But
5725 : * first update state to reflect hotplug activity if required.
5726 : */
5727 : os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5728 : WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5729 : if (os == TICK_SCHED_REMOTE_RUNNING)
5730 : queue_delayed_work(system_unbound_wq, dwork, HZ);
5731 : }
5732 :
5733 : static void sched_tick_start(int cpu)
5734 : {
5735 : int os;
5736 : struct tick_work *twork;
5737 :
5738 : if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5739 : return;
5740 :
5741 : WARN_ON_ONCE(!tick_work_cpu);
5742 :
5743 : twork = per_cpu_ptr(tick_work_cpu, cpu);
5744 : os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5745 : WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5746 : if (os == TICK_SCHED_REMOTE_OFFLINE) {
5747 : twork->cpu = cpu;
5748 : INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5749 : queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5750 : }
5751 : }
5752 :
5753 : #ifdef CONFIG_HOTPLUG_CPU
5754 : static void sched_tick_stop(int cpu)
5755 : {
5756 : struct tick_work *twork;
5757 : int os;
5758 :
5759 : if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5760 : return;
5761 :
5762 : WARN_ON_ONCE(!tick_work_cpu);
5763 :
5764 : twork = per_cpu_ptr(tick_work_cpu, cpu);
5765 : /* There cannot be competing actions, but don't rely on stop-machine. */
5766 : os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5767 : WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5768 : /* Don't cancel, as this would mess up the state machine. */
5769 : }
5770 : #endif /* CONFIG_HOTPLUG_CPU */
5771 :
5772 : int __init sched_tick_offload_init(void)
5773 : {
5774 : tick_work_cpu = alloc_percpu(struct tick_work);
5775 : BUG_ON(!tick_work_cpu);
5776 : return 0;
5777 : }
5778 :
5779 : #else /* !CONFIG_NO_HZ_FULL */
5780 : static inline void sched_tick_start(int cpu) { }
5781 : static inline void sched_tick_stop(int cpu) { }
5782 : #endif
5783 :
5784 : #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5785 : defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5786 : /*
5787 : * If the value passed in is equal to the current preempt count
5788 : * then we just disabled preemption. Start timing the latency.
5789 : */
5790 : static inline void preempt_latency_start(int val)
5791 : {
5792 : if (preempt_count() == val) {
5793 : unsigned long ip = get_lock_parent_ip();
5794 : #ifdef CONFIG_DEBUG_PREEMPT
5795 : current->preempt_disable_ip = ip;
5796 : #endif
5797 : trace_preempt_off(CALLER_ADDR0, ip);
5798 : }
5799 : }
5800 :
5801 : void preempt_count_add(int val)
5802 : {
5803 : #ifdef CONFIG_DEBUG_PREEMPT
5804 : /*
5805 : * Underflow?
5806 : */
5807 : if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5808 : return;
5809 : #endif
5810 : __preempt_count_add(val);
5811 : #ifdef CONFIG_DEBUG_PREEMPT
5812 : /*
5813 : * Spinlock count overflowing soon?
5814 : */
5815 : DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5816 : PREEMPT_MASK - 10);
5817 : #endif
5818 : preempt_latency_start(val);
5819 : }
5820 : EXPORT_SYMBOL(preempt_count_add);
5821 : NOKPROBE_SYMBOL(preempt_count_add);
5822 :
5823 : /*
5824 : * If the value passed in equals to the current preempt count
5825 : * then we just enabled preemption. Stop timing the latency.
5826 : */
5827 : static inline void preempt_latency_stop(int val)
5828 : {
5829 : if (preempt_count() == val)
5830 : trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5831 : }
5832 :
5833 : void preempt_count_sub(int val)
5834 : {
5835 : #ifdef CONFIG_DEBUG_PREEMPT
5836 : /*
5837 : * Underflow?
5838 : */
5839 : if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5840 : return;
5841 : /*
5842 : * Is the spinlock portion underflowing?
5843 : */
5844 : if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5845 : !(preempt_count() & PREEMPT_MASK)))
5846 : return;
5847 : #endif
5848 :
5849 : preempt_latency_stop(val);
5850 : __preempt_count_sub(val);
5851 : }
5852 : EXPORT_SYMBOL(preempt_count_sub);
5853 : NOKPROBE_SYMBOL(preempt_count_sub);
5854 :
5855 : #else
5856 : static inline void preempt_latency_start(int val) { }
5857 : static inline void preempt_latency_stop(int val) { }
5858 : #endif
5859 :
5860 : static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5861 : {
5862 : #ifdef CONFIG_DEBUG_PREEMPT
5863 : return p->preempt_disable_ip;
5864 : #else
5865 : return 0;
5866 : #endif
5867 : }
5868 :
5869 : /*
5870 : * Print scheduling while atomic bug:
5871 : */
5872 0 : static noinline void __schedule_bug(struct task_struct *prev)
5873 : {
5874 : /* Save this before calling printk(), since that will clobber it */
5875 0 : unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5876 :
5877 0 : if (oops_in_progress)
5878 : return;
5879 :
5880 0 : printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5881 : prev->comm, prev->pid, preempt_count());
5882 :
5883 0 : debug_show_held_locks(prev);
5884 : print_modules();
5885 0 : if (irqs_disabled())
5886 : print_irqtrace_events(prev);
5887 : if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5888 : && in_atomic_preempt_off()) {
5889 : pr_err("Preemption disabled at:");
5890 : print_ip_sym(KERN_ERR, preempt_disable_ip);
5891 : }
5892 0 : check_panic_on_warn("scheduling while atomic");
5893 :
5894 0 : dump_stack();
5895 0 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5896 : }
5897 :
5898 : /*
5899 : * Various schedule()-time debugging checks and statistics:
5900 : */
5901 : static inline void schedule_debug(struct task_struct *prev, bool preempt)
5902 : {
5903 : #ifdef CONFIG_SCHED_STACK_END_CHECK
5904 : if (task_stack_end_corrupted(prev))
5905 : panic("corrupted stack end detected inside scheduler\n");
5906 :
5907 : if (task_scs_end_corrupted(prev))
5908 : panic("corrupted shadow stack detected inside scheduler\n");
5909 : #endif
5910 :
5911 : #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5912 : if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5913 : printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5914 : prev->comm, prev->pid, prev->non_block_count);
5915 : dump_stack();
5916 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5917 : }
5918 : #endif
5919 :
5920 2513 : if (unlikely(in_atomic_preempt_off())) {
5921 0 : __schedule_bug(prev);
5922 : preempt_count_set(PREEMPT_DISABLED);
5923 : }
5924 : rcu_sleep_check();
5925 : SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5926 :
5927 2513 : profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5928 :
5929 : schedstat_inc(this_rq()->sched_count);
5930 : }
5931 :
5932 : static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5933 : struct rq_flags *rf)
5934 : {
5935 : #ifdef CONFIG_SMP
5936 : const struct sched_class *class;
5937 : /*
5938 : * We must do the balancing pass before put_prev_task(), such
5939 : * that when we release the rq->lock the task is in the same
5940 : * state as before we took rq->lock.
5941 : *
5942 : * We can terminate the balance pass as soon as we know there is
5943 : * a runnable task of @class priority or higher.
5944 : */
5945 : for_class_range(class, prev->sched_class, &idle_sched_class) {
5946 : if (class->balance(rq, prev, rf))
5947 : break;
5948 : }
5949 : #endif
5950 :
5951 0 : put_prev_task(rq, prev);
5952 : }
5953 :
5954 : /*
5955 : * Pick up the highest-prio task:
5956 : */
5957 : static inline struct task_struct *
5958 2513 : __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5959 : {
5960 : const struct sched_class *class;
5961 : struct task_struct *p;
5962 :
5963 : /*
5964 : * Optimization: we know that if all tasks are in the fair class we can
5965 : * call that function directly, but only if the @prev task wasn't of a
5966 : * higher scheduling class, because otherwise those lose the
5967 : * opportunity to pull in more work from other CPUs.
5968 : */
5969 2513 : if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
5970 : rq->nr_running == rq->cfs.h_nr_running)) {
5971 :
5972 2513 : p = pick_next_task_fair(rq, prev, rf);
5973 2513 : if (unlikely(p == RETRY_TASK))
5974 : goto restart;
5975 :
5976 : /* Assume the next prioritized class is idle_sched_class */
5977 2513 : if (!p) {
5978 1 : put_prev_task(rq, prev);
5979 1 : p = pick_next_task_idle(rq);
5980 : }
5981 :
5982 : return p;
5983 : }
5984 :
5985 : restart:
5986 0 : put_prev_task_balance(rq, prev, rf);
5987 :
5988 0 : for_each_class(class) {
5989 0 : p = class->pick_next_task(rq);
5990 0 : if (p)
5991 : return p;
5992 : }
5993 :
5994 0 : BUG(); /* The idle class should always have a runnable task. */
5995 : }
5996 :
5997 : #ifdef CONFIG_SCHED_CORE
5998 : static inline bool is_task_rq_idle(struct task_struct *t)
5999 : {
6000 : return (task_rq(t)->idle == t);
6001 : }
6002 :
6003 : static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
6004 : {
6005 : return is_task_rq_idle(a) || (a->core_cookie == cookie);
6006 : }
6007 :
6008 : static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
6009 : {
6010 : if (is_task_rq_idle(a) || is_task_rq_idle(b))
6011 : return true;
6012 :
6013 : return a->core_cookie == b->core_cookie;
6014 : }
6015 :
6016 : static inline struct task_struct *pick_task(struct rq *rq)
6017 : {
6018 : const struct sched_class *class;
6019 : struct task_struct *p;
6020 :
6021 : for_each_class(class) {
6022 : p = class->pick_task(rq);
6023 : if (p)
6024 : return p;
6025 : }
6026 :
6027 : BUG(); /* The idle class should always have a runnable task. */
6028 : }
6029 :
6030 : extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
6031 :
6032 : static void queue_core_balance(struct rq *rq);
6033 :
6034 : static struct task_struct *
6035 : pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6036 : {
6037 : struct task_struct *next, *p, *max = NULL;
6038 : const struct cpumask *smt_mask;
6039 : bool fi_before = false;
6040 : bool core_clock_updated = (rq == rq->core);
6041 : unsigned long cookie;
6042 : int i, cpu, occ = 0;
6043 : struct rq *rq_i;
6044 : bool need_sync;
6045 :
6046 : if (!sched_core_enabled(rq))
6047 : return __pick_next_task(rq, prev, rf);
6048 :
6049 : cpu = cpu_of(rq);
6050 :
6051 : /* Stopper task is switching into idle, no need core-wide selection. */
6052 : if (cpu_is_offline(cpu)) {
6053 : /*
6054 : * Reset core_pick so that we don't enter the fastpath when
6055 : * coming online. core_pick would already be migrated to
6056 : * another cpu during offline.
6057 : */
6058 : rq->core_pick = NULL;
6059 : return __pick_next_task(rq, prev, rf);
6060 : }
6061 :
6062 : /*
6063 : * If there were no {en,de}queues since we picked (IOW, the task
6064 : * pointers are all still valid), and we haven't scheduled the last
6065 : * pick yet, do so now.
6066 : *
6067 : * rq->core_pick can be NULL if no selection was made for a CPU because
6068 : * it was either offline or went offline during a sibling's core-wide
6069 : * selection. In this case, do a core-wide selection.
6070 : */
6071 : if (rq->core->core_pick_seq == rq->core->core_task_seq &&
6072 : rq->core->core_pick_seq != rq->core_sched_seq &&
6073 : rq->core_pick) {
6074 : WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
6075 :
6076 : next = rq->core_pick;
6077 : if (next != prev) {
6078 : put_prev_task(rq, prev);
6079 : set_next_task(rq, next);
6080 : }
6081 :
6082 : rq->core_pick = NULL;
6083 : goto out;
6084 : }
6085 :
6086 : put_prev_task_balance(rq, prev, rf);
6087 :
6088 : smt_mask = cpu_smt_mask(cpu);
6089 : need_sync = !!rq->core->core_cookie;
6090 :
6091 : /* reset state */
6092 : rq->core->core_cookie = 0UL;
6093 : if (rq->core->core_forceidle_count) {
6094 : if (!core_clock_updated) {
6095 : update_rq_clock(rq->core);
6096 : core_clock_updated = true;
6097 : }
6098 : sched_core_account_forceidle(rq);
6099 : /* reset after accounting force idle */
6100 : rq->core->core_forceidle_start = 0;
6101 : rq->core->core_forceidle_count = 0;
6102 : rq->core->core_forceidle_occupation = 0;
6103 : need_sync = true;
6104 : fi_before = true;
6105 : }
6106 :
6107 : /*
6108 : * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6109 : *
6110 : * @task_seq guards the task state ({en,de}queues)
6111 : * @pick_seq is the @task_seq we did a selection on
6112 : * @sched_seq is the @pick_seq we scheduled
6113 : *
6114 : * However, preemptions can cause multiple picks on the same task set.
6115 : * 'Fix' this by also increasing @task_seq for every pick.
6116 : */
6117 : rq->core->core_task_seq++;
6118 :
6119 : /*
6120 : * Optimize for common case where this CPU has no cookies
6121 : * and there are no cookied tasks running on siblings.
6122 : */
6123 : if (!need_sync) {
6124 : next = pick_task(rq);
6125 : if (!next->core_cookie) {
6126 : rq->core_pick = NULL;
6127 : /*
6128 : * For robustness, update the min_vruntime_fi for
6129 : * unconstrained picks as well.
6130 : */
6131 : WARN_ON_ONCE(fi_before);
6132 : task_vruntime_update(rq, next, false);
6133 : goto out_set_next;
6134 : }
6135 : }
6136 :
6137 : /*
6138 : * For each thread: do the regular task pick and find the max prio task
6139 : * amongst them.
6140 : *
6141 : * Tie-break prio towards the current CPU
6142 : */
6143 : for_each_cpu_wrap(i, smt_mask, cpu) {
6144 : rq_i = cpu_rq(i);
6145 :
6146 : /*
6147 : * Current cpu always has its clock updated on entrance to
6148 : * pick_next_task(). If the current cpu is not the core,
6149 : * the core may also have been updated above.
6150 : */
6151 : if (i != cpu && (rq_i != rq->core || !core_clock_updated))
6152 : update_rq_clock(rq_i);
6153 :
6154 : p = rq_i->core_pick = pick_task(rq_i);
6155 : if (!max || prio_less(max, p, fi_before))
6156 : max = p;
6157 : }
6158 :
6159 : cookie = rq->core->core_cookie = max->core_cookie;
6160 :
6161 : /*
6162 : * For each thread: try and find a runnable task that matches @max or
6163 : * force idle.
6164 : */
6165 : for_each_cpu(i, smt_mask) {
6166 : rq_i = cpu_rq(i);
6167 : p = rq_i->core_pick;
6168 :
6169 : if (!cookie_equals(p, cookie)) {
6170 : p = NULL;
6171 : if (cookie)
6172 : p = sched_core_find(rq_i, cookie);
6173 : if (!p)
6174 : p = idle_sched_class.pick_task(rq_i);
6175 : }
6176 :
6177 : rq_i->core_pick = p;
6178 :
6179 : if (p == rq_i->idle) {
6180 : if (rq_i->nr_running) {
6181 : rq->core->core_forceidle_count++;
6182 : if (!fi_before)
6183 : rq->core->core_forceidle_seq++;
6184 : }
6185 : } else {
6186 : occ++;
6187 : }
6188 : }
6189 :
6190 : if (schedstat_enabled() && rq->core->core_forceidle_count) {
6191 : rq->core->core_forceidle_start = rq_clock(rq->core);
6192 : rq->core->core_forceidle_occupation = occ;
6193 : }
6194 :
6195 : rq->core->core_pick_seq = rq->core->core_task_seq;
6196 : next = rq->core_pick;
6197 : rq->core_sched_seq = rq->core->core_pick_seq;
6198 :
6199 : /* Something should have been selected for current CPU */
6200 : WARN_ON_ONCE(!next);
6201 :
6202 : /*
6203 : * Reschedule siblings
6204 : *
6205 : * NOTE: L1TF -- at this point we're no longer running the old task and
6206 : * sending an IPI (below) ensures the sibling will no longer be running
6207 : * their task. This ensures there is no inter-sibling overlap between
6208 : * non-matching user state.
6209 : */
6210 : for_each_cpu(i, smt_mask) {
6211 : rq_i = cpu_rq(i);
6212 :
6213 : /*
6214 : * An online sibling might have gone offline before a task
6215 : * could be picked for it, or it might be offline but later
6216 : * happen to come online, but its too late and nothing was
6217 : * picked for it. That's Ok - it will pick tasks for itself,
6218 : * so ignore it.
6219 : */
6220 : if (!rq_i->core_pick)
6221 : continue;
6222 :
6223 : /*
6224 : * Update for new !FI->FI transitions, or if continuing to be in !FI:
6225 : * fi_before fi update?
6226 : * 0 0 1
6227 : * 0 1 1
6228 : * 1 0 1
6229 : * 1 1 0
6230 : */
6231 : if (!(fi_before && rq->core->core_forceidle_count))
6232 : task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6233 :
6234 : rq_i->core_pick->core_occupation = occ;
6235 :
6236 : if (i == cpu) {
6237 : rq_i->core_pick = NULL;
6238 : continue;
6239 : }
6240 :
6241 : /* Did we break L1TF mitigation requirements? */
6242 : WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6243 :
6244 : if (rq_i->curr == rq_i->core_pick) {
6245 : rq_i->core_pick = NULL;
6246 : continue;
6247 : }
6248 :
6249 : resched_curr(rq_i);
6250 : }
6251 :
6252 : out_set_next:
6253 : set_next_task(rq, next);
6254 : out:
6255 : if (rq->core->core_forceidle_count && next == rq->idle)
6256 : queue_core_balance(rq);
6257 :
6258 : return next;
6259 : }
6260 :
6261 : static bool try_steal_cookie(int this, int that)
6262 : {
6263 : struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6264 : struct task_struct *p;
6265 : unsigned long cookie;
6266 : bool success = false;
6267 :
6268 : local_irq_disable();
6269 : double_rq_lock(dst, src);
6270 :
6271 : cookie = dst->core->core_cookie;
6272 : if (!cookie)
6273 : goto unlock;
6274 :
6275 : if (dst->curr != dst->idle)
6276 : goto unlock;
6277 :
6278 : p = sched_core_find(src, cookie);
6279 : if (!p)
6280 : goto unlock;
6281 :
6282 : do {
6283 : if (p == src->core_pick || p == src->curr)
6284 : goto next;
6285 :
6286 : if (!is_cpu_allowed(p, this))
6287 : goto next;
6288 :
6289 : if (p->core_occupation > dst->idle->core_occupation)
6290 : goto next;
6291 : /*
6292 : * sched_core_find() and sched_core_next() will ensure that task @p
6293 : * is not throttled now, we also need to check whether the runqueue
6294 : * of the destination CPU is being throttled.
6295 : */
6296 : if (sched_task_is_throttled(p, this))
6297 : goto next;
6298 :
6299 : deactivate_task(src, p, 0);
6300 : set_task_cpu(p, this);
6301 : activate_task(dst, p, 0);
6302 :
6303 : resched_curr(dst);
6304 :
6305 : success = true;
6306 : break;
6307 :
6308 : next:
6309 : p = sched_core_next(p, cookie);
6310 : } while (p);
6311 :
6312 : unlock:
6313 : double_rq_unlock(dst, src);
6314 : local_irq_enable();
6315 :
6316 : return success;
6317 : }
6318 :
6319 : static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6320 : {
6321 : int i;
6322 :
6323 : for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
6324 : if (i == cpu)
6325 : continue;
6326 :
6327 : if (need_resched())
6328 : break;
6329 :
6330 : if (try_steal_cookie(cpu, i))
6331 : return true;
6332 : }
6333 :
6334 : return false;
6335 : }
6336 :
6337 : static void sched_core_balance(struct rq *rq)
6338 : {
6339 : struct sched_domain *sd;
6340 : int cpu = cpu_of(rq);
6341 :
6342 : preempt_disable();
6343 : rcu_read_lock();
6344 : raw_spin_rq_unlock_irq(rq);
6345 : for_each_domain(cpu, sd) {
6346 : if (need_resched())
6347 : break;
6348 :
6349 : if (steal_cookie_task(cpu, sd))
6350 : break;
6351 : }
6352 : raw_spin_rq_lock_irq(rq);
6353 : rcu_read_unlock();
6354 : preempt_enable();
6355 : }
6356 :
6357 : static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
6358 :
6359 : static void queue_core_balance(struct rq *rq)
6360 : {
6361 : if (!sched_core_enabled(rq))
6362 : return;
6363 :
6364 : if (!rq->core->core_cookie)
6365 : return;
6366 :
6367 : if (!rq->nr_running) /* not forced idle */
6368 : return;
6369 :
6370 : queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6371 : }
6372 :
6373 : static void sched_core_cpu_starting(unsigned int cpu)
6374 : {
6375 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6376 : struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6377 : unsigned long flags;
6378 : int t;
6379 :
6380 : sched_core_lock(cpu, &flags);
6381 :
6382 : WARN_ON_ONCE(rq->core != rq);
6383 :
6384 : /* if we're the first, we'll be our own leader */
6385 : if (cpumask_weight(smt_mask) == 1)
6386 : goto unlock;
6387 :
6388 : /* find the leader */
6389 : for_each_cpu(t, smt_mask) {
6390 : if (t == cpu)
6391 : continue;
6392 : rq = cpu_rq(t);
6393 : if (rq->core == rq) {
6394 : core_rq = rq;
6395 : break;
6396 : }
6397 : }
6398 :
6399 : if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6400 : goto unlock;
6401 :
6402 : /* install and validate core_rq */
6403 : for_each_cpu(t, smt_mask) {
6404 : rq = cpu_rq(t);
6405 :
6406 : if (t == cpu)
6407 : rq->core = core_rq;
6408 :
6409 : WARN_ON_ONCE(rq->core != core_rq);
6410 : }
6411 :
6412 : unlock:
6413 : sched_core_unlock(cpu, &flags);
6414 : }
6415 :
6416 : static void sched_core_cpu_deactivate(unsigned int cpu)
6417 : {
6418 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6419 : struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6420 : unsigned long flags;
6421 : int t;
6422 :
6423 : sched_core_lock(cpu, &flags);
6424 :
6425 : /* if we're the last man standing, nothing to do */
6426 : if (cpumask_weight(smt_mask) == 1) {
6427 : WARN_ON_ONCE(rq->core != rq);
6428 : goto unlock;
6429 : }
6430 :
6431 : /* if we're not the leader, nothing to do */
6432 : if (rq->core != rq)
6433 : goto unlock;
6434 :
6435 : /* find a new leader */
6436 : for_each_cpu(t, smt_mask) {
6437 : if (t == cpu)
6438 : continue;
6439 : core_rq = cpu_rq(t);
6440 : break;
6441 : }
6442 :
6443 : if (WARN_ON_ONCE(!core_rq)) /* impossible */
6444 : goto unlock;
6445 :
6446 : /* copy the shared state to the new leader */
6447 : core_rq->core_task_seq = rq->core_task_seq;
6448 : core_rq->core_pick_seq = rq->core_pick_seq;
6449 : core_rq->core_cookie = rq->core_cookie;
6450 : core_rq->core_forceidle_count = rq->core_forceidle_count;
6451 : core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6452 : core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6453 :
6454 : /*
6455 : * Accounting edge for forced idle is handled in pick_next_task().
6456 : * Don't need another one here, since the hotplug thread shouldn't
6457 : * have a cookie.
6458 : */
6459 : core_rq->core_forceidle_start = 0;
6460 :
6461 : /* install new leader */
6462 : for_each_cpu(t, smt_mask) {
6463 : rq = cpu_rq(t);
6464 : rq->core = core_rq;
6465 : }
6466 :
6467 : unlock:
6468 : sched_core_unlock(cpu, &flags);
6469 : }
6470 :
6471 : static inline void sched_core_cpu_dying(unsigned int cpu)
6472 : {
6473 : struct rq *rq = cpu_rq(cpu);
6474 :
6475 : if (rq->core != rq)
6476 : rq->core = rq;
6477 : }
6478 :
6479 : #else /* !CONFIG_SCHED_CORE */
6480 :
6481 : static inline void sched_core_cpu_starting(unsigned int cpu) {}
6482 : static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6483 : static inline void sched_core_cpu_dying(unsigned int cpu) {}
6484 :
6485 : static struct task_struct *
6486 : pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6487 : {
6488 2513 : return __pick_next_task(rq, prev, rf);
6489 : }
6490 :
6491 : #endif /* CONFIG_SCHED_CORE */
6492 :
6493 : /*
6494 : * Constants for the sched_mode argument of __schedule().
6495 : *
6496 : * The mode argument allows RT enabled kernels to differentiate a
6497 : * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6498 : * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6499 : * optimize the AND operation out and just check for zero.
6500 : */
6501 : #define SM_NONE 0x0
6502 : #define SM_PREEMPT 0x1
6503 : #define SM_RTLOCK_WAIT 0x2
6504 :
6505 : #ifndef CONFIG_PREEMPT_RT
6506 : # define SM_MASK_PREEMPT (~0U)
6507 : #else
6508 : # define SM_MASK_PREEMPT SM_PREEMPT
6509 : #endif
6510 :
6511 : /*
6512 : * __schedule() is the main scheduler function.
6513 : *
6514 : * The main means of driving the scheduler and thus entering this function are:
6515 : *
6516 : * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6517 : *
6518 : * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6519 : * paths. For example, see arch/x86/entry_64.S.
6520 : *
6521 : * To drive preemption between tasks, the scheduler sets the flag in timer
6522 : * interrupt handler scheduler_tick().
6523 : *
6524 : * 3. Wakeups don't really cause entry into schedule(). They add a
6525 : * task to the run-queue and that's it.
6526 : *
6527 : * Now, if the new task added to the run-queue preempts the current
6528 : * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6529 : * called on the nearest possible occasion:
6530 : *
6531 : * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6532 : *
6533 : * - in syscall or exception context, at the next outmost
6534 : * preempt_enable(). (this might be as soon as the wake_up()'s
6535 : * spin_unlock()!)
6536 : *
6537 : * - in IRQ context, return from interrupt-handler to
6538 : * preemptible context
6539 : *
6540 : * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6541 : * then at the next:
6542 : *
6543 : * - cond_resched() call
6544 : * - explicit schedule() call
6545 : * - return from syscall or exception to user-space
6546 : * - return from interrupt-handler to user-space
6547 : *
6548 : * WARNING: must be called with preemption disabled!
6549 : */
6550 2513 : static void __sched notrace __schedule(unsigned int sched_mode)
6551 : {
6552 : struct task_struct *prev, *next;
6553 : unsigned long *switch_count;
6554 : unsigned long prev_state;
6555 : struct rq_flags rf;
6556 : struct rq *rq;
6557 : int cpu;
6558 :
6559 2513 : cpu = smp_processor_id();
6560 2513 : rq = cpu_rq(cpu);
6561 2513 : prev = rq->curr;
6562 :
6563 5026 : schedule_debug(prev, !!sched_mode);
6564 :
6565 : if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6566 : hrtick_clear(rq);
6567 :
6568 : local_irq_disable();
6569 2513 : rcu_note_context_switch(!!sched_mode);
6570 :
6571 : /*
6572 : * Make sure that signal_pending_state()->signal_pending() below
6573 : * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6574 : * done by the caller to avoid the race with signal_wake_up():
6575 : *
6576 : * __set_current_state(@state) signal_wake_up()
6577 : * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6578 : * wake_up_state(p, state)
6579 : * LOCK rq->lock LOCK p->pi_state
6580 : * smp_mb__after_spinlock() smp_mb__after_spinlock()
6581 : * if (signal_pending_state()) if (p->state & @state)
6582 : *
6583 : * Also, the membarrier system call requires a full memory barrier
6584 : * after coming from user-space, before storing to rq->curr.
6585 : */
6586 2513 : rq_lock(rq, &rf);
6587 : smp_mb__after_spinlock();
6588 :
6589 : /* Promote REQ to ACT */
6590 2513 : rq->clock_update_flags <<= 1;
6591 2513 : update_rq_clock(rq);
6592 :
6593 2513 : switch_count = &prev->nivcsw;
6594 :
6595 : /*
6596 : * We must load prev->state once (task_struct::state is volatile), such
6597 : * that we form a control dependency vs deactivate_task() below.
6598 : */
6599 2513 : prev_state = READ_ONCE(prev->__state);
6600 2513 : if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6601 2440 : if (signal_pending_state(prev_state, prev)) {
6602 0 : WRITE_ONCE(prev->__state, TASK_RUNNING);
6603 : } else {
6604 2440 : prev->sched_contributes_to_load =
6605 : (prev_state & TASK_UNINTERRUPTIBLE) &&
6606 2440 : !(prev_state & TASK_NOLOAD) &&
6607 : !(prev_state & TASK_FROZEN);
6608 :
6609 2440 : if (prev->sched_contributes_to_load)
6610 1153 : rq->nr_uninterruptible++;
6611 :
6612 : /*
6613 : * __schedule() ttwu()
6614 : * prev_state = prev->state; if (p->on_rq && ...)
6615 : * if (prev_state) goto out;
6616 : * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6617 : * p->state = TASK_WAKING
6618 : *
6619 : * Where __schedule() and ttwu() have matching control dependencies.
6620 : *
6621 : * After this, schedule() must not care about p->state any more.
6622 : */
6623 2440 : deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6624 :
6625 2440 : if (prev->in_iowait) {
6626 0 : atomic_inc(&rq->nr_iowait);
6627 : delayacct_blkio_start();
6628 : }
6629 : }
6630 2440 : switch_count = &prev->nvcsw;
6631 : }
6632 :
6633 2513 : next = pick_next_task(rq, prev, &rf);
6634 2513 : clear_tsk_need_resched(prev);
6635 : clear_preempt_need_resched();
6636 : #ifdef CONFIG_SCHED_DEBUG
6637 : rq->last_seen_need_resched_ns = 0;
6638 : #endif
6639 :
6640 2513 : if (likely(prev != next)) {
6641 2513 : rq->nr_switches++;
6642 : /*
6643 : * RCU users of rcu_dereference(rq->curr) may not see
6644 : * changes to task_struct made by pick_next_task().
6645 : */
6646 2513 : RCU_INIT_POINTER(rq->curr, next);
6647 : /*
6648 : * The membarrier system call requires each architecture
6649 : * to have a full memory barrier after updating
6650 : * rq->curr, before returning to user-space.
6651 : *
6652 : * Here are the schemes providing that barrier on the
6653 : * various architectures:
6654 : * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6655 : * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6656 : * - finish_lock_switch() for weakly-ordered
6657 : * architectures where spin_unlock is a full barrier,
6658 : * - switch_to() for arm64 (weakly-ordered, spin_unlock
6659 : * is a RELEASE barrier),
6660 : */
6661 2513 : ++*switch_count;
6662 :
6663 2513 : migrate_disable_switch(rq, prev);
6664 2513 : psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6665 :
6666 2513 : trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6667 :
6668 : /* Also unlocks the rq: */
6669 2131 : rq = context_switch(rq, prev, next, &rf);
6670 : } else {
6671 0 : rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6672 :
6673 0 : rq_unpin_lock(rq, &rf);
6674 0 : __balance_callbacks(rq);
6675 0 : raw_spin_rq_unlock_irq(rq);
6676 : }
6677 2131 : }
6678 :
6679 367 : void __noreturn do_task_dead(void)
6680 : {
6681 : /* Causes final put_task_struct in finish_task_switch(): */
6682 1835 : set_special_state(TASK_DEAD);
6683 :
6684 : /* Tell freezer to ignore us: */
6685 367 : current->flags |= PF_NOFREEZE;
6686 :
6687 367 : __schedule(SM_NONE);
6688 0 : BUG();
6689 :
6690 : /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6691 : for (;;)
6692 : cpu_relax();
6693 : }
6694 :
6695 2074 : static inline void sched_submit_work(struct task_struct *tsk)
6696 : {
6697 : unsigned int task_flags;
6698 :
6699 2074 : if (task_is_running(tsk))
6700 : return;
6701 :
6702 2073 : task_flags = tsk->flags;
6703 : /*
6704 : * If a worker goes to sleep, notify and ask workqueue whether it
6705 : * wants to wake up a task to maintain concurrency.
6706 : */
6707 2073 : if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6708 94 : if (task_flags & PF_WQ_WORKER)
6709 94 : wq_worker_sleeping(tsk);
6710 : else
6711 0 : io_wq_worker_sleeping(tsk);
6712 : }
6713 :
6714 : /*
6715 : * spinlock and rwlock must not flush block requests. This will
6716 : * deadlock if the callback attempts to acquire a lock which is
6717 : * already acquired.
6718 : */
6719 2073 : SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6720 :
6721 : /*
6722 : * If we are going to sleep and we have plugged IO queued,
6723 : * make sure to submit it to avoid deadlocks.
6724 : */
6725 2073 : blk_flush_plug(tsk->plug, true);
6726 : }
6727 :
6728 2059 : static void sched_update_worker(struct task_struct *tsk)
6729 : {
6730 2059 : if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6731 85 : if (tsk->flags & PF_WQ_WORKER)
6732 85 : wq_worker_running(tsk);
6733 : else
6734 0 : io_wq_worker_running(tsk);
6735 : }
6736 2059 : }
6737 :
6738 2074 : asmlinkage __visible void __sched schedule(void)
6739 : {
6740 2074 : struct task_struct *tsk = current;
6741 :
6742 2074 : sched_submit_work(tsk);
6743 : do {
6744 2074 : preempt_disable();
6745 2074 : __schedule(SM_NONE);
6746 2059 : sched_preempt_enable_no_resched();
6747 2059 : } while (need_resched());
6748 2059 : sched_update_worker(tsk);
6749 2059 : }
6750 : EXPORT_SYMBOL(schedule);
6751 :
6752 : /*
6753 : * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6754 : * state (have scheduled out non-voluntarily) by making sure that all
6755 : * tasks have either left the run queue or have gone into user space.
6756 : * As idle tasks do not do either, they must not ever be preempted
6757 : * (schedule out non-voluntarily).
6758 : *
6759 : * schedule_idle() is similar to schedule_preempt_disable() except that it
6760 : * never enables preemption because it does not call sched_submit_work().
6761 : */
6762 0 : void __sched schedule_idle(void)
6763 : {
6764 : /*
6765 : * As this skips calling sched_submit_work(), which the idle task does
6766 : * regardless because that function is a nop when the task is in a
6767 : * TASK_RUNNING state, make sure this isn't used someplace that the
6768 : * current task can be in any other state. Note, idle is always in the
6769 : * TASK_RUNNING state.
6770 : */
6771 0 : WARN_ON_ONCE(current->__state);
6772 : do {
6773 0 : __schedule(SM_NONE);
6774 0 : } while (need_resched());
6775 0 : }
6776 :
6777 : #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6778 : asmlinkage __visible void __sched schedule_user(void)
6779 : {
6780 : /*
6781 : * If we come here after a random call to set_need_resched(),
6782 : * or we have been woken up remotely but the IPI has not yet arrived,
6783 : * we haven't yet exited the RCU idle mode. Do it here manually until
6784 : * we find a better solution.
6785 : *
6786 : * NB: There are buggy callers of this function. Ideally we
6787 : * should warn if prev_state != CONTEXT_USER, but that will trigger
6788 : * too frequently to make sense yet.
6789 : */
6790 : enum ctx_state prev_state = exception_enter();
6791 : schedule();
6792 : exception_exit(prev_state);
6793 : }
6794 : #endif
6795 :
6796 : /**
6797 : * schedule_preempt_disabled - called with preemption disabled
6798 : *
6799 : * Returns with preemption disabled. Note: preempt_count must be 1
6800 : */
6801 382 : void __sched schedule_preempt_disabled(void)
6802 : {
6803 382 : sched_preempt_enable_no_resched();
6804 382 : schedule();
6805 381 : preempt_disable();
6806 381 : }
6807 :
6808 : #ifdef CONFIG_PREEMPT_RT
6809 : void __sched notrace schedule_rtlock(void)
6810 : {
6811 : do {
6812 : preempt_disable();
6813 : __schedule(SM_RTLOCK_WAIT);
6814 : sched_preempt_enable_no_resched();
6815 : } while (need_resched());
6816 : }
6817 : NOKPROBE_SYMBOL(schedule_rtlock);
6818 : #endif
6819 :
6820 : static void __sched notrace preempt_schedule_common(void)
6821 : {
6822 : do {
6823 : /*
6824 : * Because the function tracer can trace preempt_count_sub()
6825 : * and it also uses preempt_enable/disable_notrace(), if
6826 : * NEED_RESCHED is set, the preempt_enable_notrace() called
6827 : * by the function tracer will call this function again and
6828 : * cause infinite recursion.
6829 : *
6830 : * Preemption must be disabled here before the function
6831 : * tracer can trace. Break up preempt_disable() into two
6832 : * calls. One to disable preemption without fear of being
6833 : * traced. The other to still record the preemption latency,
6834 : * which can also be traced by the function tracer.
6835 : */
6836 72 : preempt_disable_notrace();
6837 72 : preempt_latency_start(1);
6838 72 : __schedule(SM_PREEMPT);
6839 72 : preempt_latency_stop(1);
6840 72 : preempt_enable_no_resched_notrace();
6841 :
6842 : /*
6843 : * Check again in case we missed a preemption opportunity
6844 : * between schedule and now.
6845 : */
6846 72 : } while (need_resched());
6847 : }
6848 :
6849 : #ifdef CONFIG_PREEMPTION
6850 : /*
6851 : * This is the entry point to schedule() from in-kernel preemption
6852 : * off of preempt_enable.
6853 : */
6854 : asmlinkage __visible void __sched notrace preempt_schedule(void)
6855 : {
6856 : /*
6857 : * If there is a non-zero preempt_count or interrupts are disabled,
6858 : * we do not want to preempt the current task. Just return..
6859 : */
6860 : if (likely(!preemptible()))
6861 : return;
6862 : preempt_schedule_common();
6863 : }
6864 : NOKPROBE_SYMBOL(preempt_schedule);
6865 : EXPORT_SYMBOL(preempt_schedule);
6866 :
6867 : #ifdef CONFIG_PREEMPT_DYNAMIC
6868 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6869 : #ifndef preempt_schedule_dynamic_enabled
6870 : #define preempt_schedule_dynamic_enabled preempt_schedule
6871 : #define preempt_schedule_dynamic_disabled NULL
6872 : #endif
6873 : DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6874 : EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6875 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6876 : static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6877 : void __sched notrace dynamic_preempt_schedule(void)
6878 : {
6879 : if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6880 : return;
6881 : preempt_schedule();
6882 : }
6883 : NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6884 : EXPORT_SYMBOL(dynamic_preempt_schedule);
6885 : #endif
6886 : #endif
6887 :
6888 : /**
6889 : * preempt_schedule_notrace - preempt_schedule called by tracing
6890 : *
6891 : * The tracing infrastructure uses preempt_enable_notrace to prevent
6892 : * recursion and tracing preempt enabling caused by the tracing
6893 : * infrastructure itself. But as tracing can happen in areas coming
6894 : * from userspace or just about to enter userspace, a preempt enable
6895 : * can occur before user_exit() is called. This will cause the scheduler
6896 : * to be called when the system is still in usermode.
6897 : *
6898 : * To prevent this, the preempt_enable_notrace will use this function
6899 : * instead of preempt_schedule() to exit user context if needed before
6900 : * calling the scheduler.
6901 : */
6902 : asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6903 : {
6904 : enum ctx_state prev_ctx;
6905 :
6906 : if (likely(!preemptible()))
6907 : return;
6908 :
6909 : do {
6910 : /*
6911 : * Because the function tracer can trace preempt_count_sub()
6912 : * and it also uses preempt_enable/disable_notrace(), if
6913 : * NEED_RESCHED is set, the preempt_enable_notrace() called
6914 : * by the function tracer will call this function again and
6915 : * cause infinite recursion.
6916 : *
6917 : * Preemption must be disabled here before the function
6918 : * tracer can trace. Break up preempt_disable() into two
6919 : * calls. One to disable preemption without fear of being
6920 : * traced. The other to still record the preemption latency,
6921 : * which can also be traced by the function tracer.
6922 : */
6923 : preempt_disable_notrace();
6924 : preempt_latency_start(1);
6925 : /*
6926 : * Needs preempt disabled in case user_exit() is traced
6927 : * and the tracer calls preempt_enable_notrace() causing
6928 : * an infinite recursion.
6929 : */
6930 : prev_ctx = exception_enter();
6931 : __schedule(SM_PREEMPT);
6932 : exception_exit(prev_ctx);
6933 :
6934 : preempt_latency_stop(1);
6935 : preempt_enable_no_resched_notrace();
6936 : } while (need_resched());
6937 : }
6938 : EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6939 :
6940 : #ifdef CONFIG_PREEMPT_DYNAMIC
6941 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6942 : #ifndef preempt_schedule_notrace_dynamic_enabled
6943 : #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6944 : #define preempt_schedule_notrace_dynamic_disabled NULL
6945 : #endif
6946 : DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6947 : EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6948 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6949 : static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6950 : void __sched notrace dynamic_preempt_schedule_notrace(void)
6951 : {
6952 : if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6953 : return;
6954 : preempt_schedule_notrace();
6955 : }
6956 : NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6957 : EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6958 : #endif
6959 : #endif
6960 :
6961 : #endif /* CONFIG_PREEMPTION */
6962 :
6963 : /*
6964 : * This is the entry point to schedule() from kernel preemption
6965 : * off of irq context.
6966 : * Note, that this is called and return with irqs disabled. This will
6967 : * protect us against recursive calling from irq.
6968 : */
6969 0 : asmlinkage __visible void __sched preempt_schedule_irq(void)
6970 : {
6971 : enum ctx_state prev_state;
6972 :
6973 : /* Catch callers which need to be fixed */
6974 0 : BUG_ON(preempt_count() || !irqs_disabled());
6975 :
6976 : prev_state = exception_enter();
6977 :
6978 : do {
6979 0 : preempt_disable();
6980 : local_irq_enable();
6981 0 : __schedule(SM_PREEMPT);
6982 : local_irq_disable();
6983 0 : sched_preempt_enable_no_resched();
6984 0 : } while (need_resched());
6985 :
6986 : exception_exit(prev_state);
6987 0 : }
6988 :
6989 0 : int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6990 : void *key)
6991 : {
6992 0 : WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6993 0 : return try_to_wake_up(curr->private, mode, wake_flags);
6994 : }
6995 : EXPORT_SYMBOL(default_wake_function);
6996 :
6997 : static void __setscheduler_prio(struct task_struct *p, int prio)
6998 : {
6999 0 : if (dl_prio(prio))
7000 0 : p->sched_class = &dl_sched_class;
7001 0 : else if (rt_prio(prio))
7002 0 : p->sched_class = &rt_sched_class;
7003 : else
7004 0 : p->sched_class = &fair_sched_class;
7005 :
7006 0 : p->prio = prio;
7007 : }
7008 :
7009 : #ifdef CONFIG_RT_MUTEXES
7010 :
7011 : static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
7012 : {
7013 0 : if (pi_task)
7014 0 : prio = min(prio, pi_task->prio);
7015 :
7016 : return prio;
7017 : }
7018 :
7019 : static inline int rt_effective_prio(struct task_struct *p, int prio)
7020 : {
7021 0 : struct task_struct *pi_task = rt_mutex_get_top_task(p);
7022 :
7023 0 : return __rt_effective_prio(pi_task, prio);
7024 : }
7025 :
7026 : /*
7027 : * rt_mutex_setprio - set the current priority of a task
7028 : * @p: task to boost
7029 : * @pi_task: donor task
7030 : *
7031 : * This function changes the 'effective' priority of a task. It does
7032 : * not touch ->normal_prio like __setscheduler().
7033 : *
7034 : * Used by the rt_mutex code to implement priority inheritance
7035 : * logic. Call site only calls if the priority of the task changed.
7036 : */
7037 0 : void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
7038 : {
7039 0 : int prio, oldprio, queued, running, queue_flag =
7040 : DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7041 : const struct sched_class *prev_class;
7042 : struct rq_flags rf;
7043 : struct rq *rq;
7044 :
7045 : /* XXX used to be waiter->prio, not waiter->task->prio */
7046 0 : prio = __rt_effective_prio(pi_task, p->normal_prio);
7047 :
7048 : /*
7049 : * If nothing changed; bail early.
7050 : */
7051 0 : if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
7052 : return;
7053 :
7054 0 : rq = __task_rq_lock(p, &rf);
7055 0 : update_rq_clock(rq);
7056 : /*
7057 : * Set under pi_lock && rq->lock, such that the value can be used under
7058 : * either lock.
7059 : *
7060 : * Note that there is loads of tricky to make this pointer cache work
7061 : * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
7062 : * ensure a task is de-boosted (pi_task is set to NULL) before the
7063 : * task is allowed to run again (and can exit). This ensures the pointer
7064 : * points to a blocked task -- which guarantees the task is present.
7065 : */
7066 0 : p->pi_top_task = pi_task;
7067 :
7068 : /*
7069 : * For FIFO/RR we only need to set prio, if that matches we're done.
7070 : */
7071 0 : if (prio == p->prio && !dl_prio(prio))
7072 : goto out_unlock;
7073 :
7074 : /*
7075 : * Idle task boosting is a nono in general. There is one
7076 : * exception, when PREEMPT_RT and NOHZ is active:
7077 : *
7078 : * The idle task calls get_next_timer_interrupt() and holds
7079 : * the timer wheel base->lock on the CPU and another CPU wants
7080 : * to access the timer (probably to cancel it). We can safely
7081 : * ignore the boosting request, as the idle CPU runs this code
7082 : * with interrupts disabled and will complete the lock
7083 : * protected section without being interrupted. So there is no
7084 : * real need to boost.
7085 : */
7086 0 : if (unlikely(p == rq->idle)) {
7087 0 : WARN_ON(p != rq->curr);
7088 0 : WARN_ON(p->pi_blocked_on);
7089 : goto out_unlock;
7090 : }
7091 :
7092 0 : trace_sched_pi_setprio(p, pi_task);
7093 0 : oldprio = p->prio;
7094 :
7095 0 : if (oldprio == prio)
7096 0 : queue_flag &= ~DEQUEUE_MOVE;
7097 :
7098 0 : prev_class = p->sched_class;
7099 0 : queued = task_on_rq_queued(p);
7100 0 : running = task_current(rq, p);
7101 0 : if (queued)
7102 0 : dequeue_task(rq, p, queue_flag);
7103 0 : if (running)
7104 0 : put_prev_task(rq, p);
7105 :
7106 : /*
7107 : * Boosting condition are:
7108 : * 1. -rt task is running and holds mutex A
7109 : * --> -dl task blocks on mutex A
7110 : *
7111 : * 2. -dl task is running and holds mutex A
7112 : * --> -dl task blocks on mutex A and could preempt the
7113 : * running task
7114 : */
7115 0 : if (dl_prio(prio)) {
7116 0 : if (!dl_prio(p->normal_prio) ||
7117 0 : (pi_task && dl_prio(pi_task->prio) &&
7118 0 : dl_entity_preempt(&pi_task->dl, &p->dl))) {
7119 0 : p->dl.pi_se = pi_task->dl.pi_se;
7120 0 : queue_flag |= ENQUEUE_REPLENISH;
7121 : } else {
7122 0 : p->dl.pi_se = &p->dl;
7123 : }
7124 0 : } else if (rt_prio(prio)) {
7125 0 : if (dl_prio(oldprio))
7126 0 : p->dl.pi_se = &p->dl;
7127 0 : if (oldprio < prio)
7128 0 : queue_flag |= ENQUEUE_HEAD;
7129 : } else {
7130 0 : if (dl_prio(oldprio))
7131 0 : p->dl.pi_se = &p->dl;
7132 0 : if (rt_prio(oldprio))
7133 0 : p->rt.timeout = 0;
7134 : }
7135 :
7136 0 : __setscheduler_prio(p, prio);
7137 :
7138 0 : if (queued)
7139 0 : enqueue_task(rq, p, queue_flag);
7140 0 : if (running)
7141 : set_next_task(rq, p);
7142 :
7143 0 : check_class_changed(rq, p, prev_class, oldprio);
7144 : out_unlock:
7145 : /* Avoid rq from going away on us: */
7146 0 : preempt_disable();
7147 :
7148 0 : rq_unpin_lock(rq, &rf);
7149 0 : __balance_callbacks(rq);
7150 0 : raw_spin_rq_unlock(rq);
7151 :
7152 0 : preempt_enable();
7153 : }
7154 : #else
7155 : static inline int rt_effective_prio(struct task_struct *p, int prio)
7156 : {
7157 : return prio;
7158 : }
7159 : #endif
7160 :
7161 9 : void set_user_nice(struct task_struct *p, long nice)
7162 : {
7163 : bool queued, running;
7164 : int old_prio;
7165 : struct rq_flags rf;
7166 : struct rq *rq;
7167 :
7168 18 : if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
7169 4 : return;
7170 : /*
7171 : * We have to be careful, if called from sys_setpriority(),
7172 : * the task might be in the middle of scheduling on another CPU.
7173 : */
7174 5 : rq = task_rq_lock(p, &rf);
7175 5 : update_rq_clock(rq);
7176 :
7177 : /*
7178 : * The RT priorities are set via sched_setscheduler(), but we still
7179 : * allow the 'normal' nice value to be set - but as expected
7180 : * it won't have any effect on scheduling until the task is
7181 : * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7182 : */
7183 15 : if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7184 0 : p->static_prio = NICE_TO_PRIO(nice);
7185 0 : goto out_unlock;
7186 : }
7187 5 : queued = task_on_rq_queued(p);
7188 5 : running = task_current(rq, p);
7189 5 : if (queued)
7190 : dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7191 5 : if (running)
7192 4 : put_prev_task(rq, p);
7193 :
7194 5 : p->static_prio = NICE_TO_PRIO(nice);
7195 5 : set_load_weight(p, true);
7196 5 : old_prio = p->prio;
7197 5 : p->prio = effective_prio(p);
7198 :
7199 5 : if (queued)
7200 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7201 5 : if (running)
7202 : set_next_task(rq, p);
7203 :
7204 : /*
7205 : * If the task increased its priority or is running and
7206 : * lowered its priority, then reschedule its CPU:
7207 : */
7208 5 : p->sched_class->prio_changed(rq, p, old_prio);
7209 :
7210 : out_unlock:
7211 10 : task_rq_unlock(rq, p, &rf);
7212 : }
7213 : EXPORT_SYMBOL(set_user_nice);
7214 :
7215 : /*
7216 : * is_nice_reduction - check if nice value is an actual reduction
7217 : *
7218 : * Similar to can_nice() but does not perform a capability check.
7219 : *
7220 : * @p: task
7221 : * @nice: nice value
7222 : */
7223 : static bool is_nice_reduction(const struct task_struct *p, const int nice)
7224 : {
7225 : /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7226 0 : int nice_rlim = nice_to_rlimit(nice);
7227 :
7228 0 : return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7229 : }
7230 :
7231 : /*
7232 : * can_nice - check if a task can reduce its nice value
7233 : * @p: task
7234 : * @nice: nice value
7235 : */
7236 0 : int can_nice(const struct task_struct *p, const int nice)
7237 : {
7238 0 : return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7239 : }
7240 :
7241 : #ifdef __ARCH_WANT_SYS_NICE
7242 :
7243 : /*
7244 : * sys_nice - change the priority of the current process.
7245 : * @increment: priority increment
7246 : *
7247 : * sys_setpriority is a more generic, but much slower function that
7248 : * does similar things.
7249 : */
7250 0 : SYSCALL_DEFINE1(nice, int, increment)
7251 : {
7252 : long nice, retval;
7253 :
7254 : /*
7255 : * Setpriority might change our priority at the same moment.
7256 : * We don't have to worry. Conceptually one call occurs first
7257 : * and we have a single winner.
7258 : */
7259 0 : increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7260 0 : nice = task_nice(current) + increment;
7261 :
7262 0 : nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7263 0 : if (increment < 0 && !can_nice(current, nice))
7264 : return -EPERM;
7265 :
7266 0 : retval = security_task_setnice(current, nice);
7267 0 : if (retval)
7268 : return retval;
7269 :
7270 0 : set_user_nice(current, nice);
7271 0 : return 0;
7272 : }
7273 :
7274 : #endif
7275 :
7276 : /**
7277 : * task_prio - return the priority value of a given task.
7278 : * @p: the task in question.
7279 : *
7280 : * Return: The priority value as seen by users in /proc.
7281 : *
7282 : * sched policy return value kernel prio user prio/nice
7283 : *
7284 : * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7285 : * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7286 : * deadline -101 -1 0
7287 : */
7288 0 : int task_prio(const struct task_struct *p)
7289 : {
7290 0 : return p->prio - MAX_RT_PRIO;
7291 : }
7292 :
7293 : /**
7294 : * idle_cpu - is a given CPU idle currently?
7295 : * @cpu: the processor in question.
7296 : *
7297 : * Return: 1 if the CPU is currently idle. 0 otherwise.
7298 : */
7299 0 : int idle_cpu(int cpu)
7300 : {
7301 0 : struct rq *rq = cpu_rq(cpu);
7302 :
7303 0 : if (rq->curr != rq->idle)
7304 : return 0;
7305 :
7306 0 : if (rq->nr_running)
7307 : return 0;
7308 :
7309 : #ifdef CONFIG_SMP
7310 : if (rq->ttwu_pending)
7311 : return 0;
7312 : #endif
7313 :
7314 0 : return 1;
7315 : }
7316 :
7317 : /**
7318 : * available_idle_cpu - is a given CPU idle for enqueuing work.
7319 : * @cpu: the CPU in question.
7320 : *
7321 : * Return: 1 if the CPU is currently idle. 0 otherwise.
7322 : */
7323 0 : int available_idle_cpu(int cpu)
7324 : {
7325 0 : if (!idle_cpu(cpu))
7326 : return 0;
7327 :
7328 0 : if (vcpu_is_preempted(cpu))
7329 : return 0;
7330 :
7331 0 : return 1;
7332 : }
7333 :
7334 : /**
7335 : * idle_task - return the idle task for a given CPU.
7336 : * @cpu: the processor in question.
7337 : *
7338 : * Return: The idle task for the CPU @cpu.
7339 : */
7340 0 : struct task_struct *idle_task(int cpu)
7341 : {
7342 0 : return cpu_rq(cpu)->idle;
7343 : }
7344 :
7345 : #ifdef CONFIG_SMP
7346 : /*
7347 : * This function computes an effective utilization for the given CPU, to be
7348 : * used for frequency selection given the linear relation: f = u * f_max.
7349 : *
7350 : * The scheduler tracks the following metrics:
7351 : *
7352 : * cpu_util_{cfs,rt,dl,irq}()
7353 : * cpu_bw_dl()
7354 : *
7355 : * Where the cfs,rt and dl util numbers are tracked with the same metric and
7356 : * synchronized windows and are thus directly comparable.
7357 : *
7358 : * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7359 : * which excludes things like IRQ and steal-time. These latter are then accrued
7360 : * in the irq utilization.
7361 : *
7362 : * The DL bandwidth number otoh is not a measured metric but a value computed
7363 : * based on the task model parameters and gives the minimal utilization
7364 : * required to meet deadlines.
7365 : */
7366 : unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7367 : enum cpu_util_type type,
7368 : struct task_struct *p)
7369 : {
7370 : unsigned long dl_util, util, irq, max;
7371 : struct rq *rq = cpu_rq(cpu);
7372 :
7373 : max = arch_scale_cpu_capacity(cpu);
7374 :
7375 : if (!uclamp_is_used() &&
7376 : type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7377 : return max;
7378 : }
7379 :
7380 : /*
7381 : * Early check to see if IRQ/steal time saturates the CPU, can be
7382 : * because of inaccuracies in how we track these -- see
7383 : * update_irq_load_avg().
7384 : */
7385 : irq = cpu_util_irq(rq);
7386 : if (unlikely(irq >= max))
7387 : return max;
7388 :
7389 : /*
7390 : * Because the time spend on RT/DL tasks is visible as 'lost' time to
7391 : * CFS tasks and we use the same metric to track the effective
7392 : * utilization (PELT windows are synchronized) we can directly add them
7393 : * to obtain the CPU's actual utilization.
7394 : *
7395 : * CFS and RT utilization can be boosted or capped, depending on
7396 : * utilization clamp constraints requested by currently RUNNABLE
7397 : * tasks.
7398 : * When there are no CFS RUNNABLE tasks, clamps are released and
7399 : * frequency will be gracefully reduced with the utilization decay.
7400 : */
7401 : util = util_cfs + cpu_util_rt(rq);
7402 : if (type == FREQUENCY_UTIL)
7403 : util = uclamp_rq_util_with(rq, util, p);
7404 :
7405 : dl_util = cpu_util_dl(rq);
7406 :
7407 : /*
7408 : * For frequency selection we do not make cpu_util_dl() a permanent part
7409 : * of this sum because we want to use cpu_bw_dl() later on, but we need
7410 : * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7411 : * that we select f_max when there is no idle time.
7412 : *
7413 : * NOTE: numerical errors or stop class might cause us to not quite hit
7414 : * saturation when we should -- something for later.
7415 : */
7416 : if (util + dl_util >= max)
7417 : return max;
7418 :
7419 : /*
7420 : * OTOH, for energy computation we need the estimated running time, so
7421 : * include util_dl and ignore dl_bw.
7422 : */
7423 : if (type == ENERGY_UTIL)
7424 : util += dl_util;
7425 :
7426 : /*
7427 : * There is still idle time; further improve the number by using the
7428 : * irq metric. Because IRQ/steal time is hidden from the task clock we
7429 : * need to scale the task numbers:
7430 : *
7431 : * max - irq
7432 : * U' = irq + --------- * U
7433 : * max
7434 : */
7435 : util = scale_irq_capacity(util, irq, max);
7436 : util += irq;
7437 :
7438 : /*
7439 : * Bandwidth required by DEADLINE must always be granted while, for
7440 : * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7441 : * to gracefully reduce the frequency when no tasks show up for longer
7442 : * periods of time.
7443 : *
7444 : * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7445 : * bw_dl as requested freq. However, cpufreq is not yet ready for such
7446 : * an interface. So, we only do the latter for now.
7447 : */
7448 : if (type == FREQUENCY_UTIL)
7449 : util += cpu_bw_dl(rq);
7450 :
7451 : return min(max, util);
7452 : }
7453 :
7454 : unsigned long sched_cpu_util(int cpu)
7455 : {
7456 : return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
7457 : }
7458 : #endif /* CONFIG_SMP */
7459 :
7460 : /**
7461 : * find_process_by_pid - find a process with a matching PID value.
7462 : * @pid: the pid in question.
7463 : *
7464 : * The task of @pid, if found. %NULL otherwise.
7465 : */
7466 : static struct task_struct *find_process_by_pid(pid_t pid)
7467 : {
7468 0 : return pid ? find_task_by_vpid(pid) : current;
7469 : }
7470 :
7471 : /*
7472 : * sched_setparam() passes in -1 for its policy, to let the functions
7473 : * it calls know not to change it.
7474 : */
7475 : #define SETPARAM_POLICY -1
7476 :
7477 0 : static void __setscheduler_params(struct task_struct *p,
7478 : const struct sched_attr *attr)
7479 : {
7480 0 : int policy = attr->sched_policy;
7481 :
7482 0 : if (policy == SETPARAM_POLICY)
7483 0 : policy = p->policy;
7484 :
7485 0 : p->policy = policy;
7486 :
7487 0 : if (dl_policy(policy))
7488 0 : __setparam_dl(p, attr);
7489 0 : else if (fair_policy(policy))
7490 0 : p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7491 :
7492 : /*
7493 : * __sched_setscheduler() ensures attr->sched_priority == 0 when
7494 : * !rt_policy. Always setting this ensures that things like
7495 : * getparam()/getattr() don't report silly values for !rt tasks.
7496 : */
7497 0 : p->rt_priority = attr->sched_priority;
7498 0 : p->normal_prio = normal_prio(p);
7499 0 : set_load_weight(p, true);
7500 0 : }
7501 :
7502 : /*
7503 : * Check the target process has a UID that matches the current process's:
7504 : */
7505 : static bool check_same_owner(struct task_struct *p)
7506 : {
7507 0 : const struct cred *cred = current_cred(), *pcred;
7508 : bool match;
7509 :
7510 : rcu_read_lock();
7511 0 : pcred = __task_cred(p);
7512 0 : match = (uid_eq(cred->euid, pcred->euid) ||
7513 0 : uid_eq(cred->euid, pcred->uid));
7514 : rcu_read_unlock();
7515 : return match;
7516 : }
7517 :
7518 : /*
7519 : * Allow unprivileged RT tasks to decrease priority.
7520 : * Only issue a capable test if needed and only once to avoid an audit
7521 : * event on permitted non-privileged operations:
7522 : */
7523 0 : static int user_check_sched_setscheduler(struct task_struct *p,
7524 : const struct sched_attr *attr,
7525 : int policy, int reset_on_fork)
7526 : {
7527 0 : if (fair_policy(policy)) {
7528 0 : if (attr->sched_nice < task_nice(p) &&
7529 0 : !is_nice_reduction(p, attr->sched_nice))
7530 : goto req_priv;
7531 : }
7532 :
7533 0 : if (rt_policy(policy)) {
7534 0 : unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7535 :
7536 : /* Can't set/change the rt policy: */
7537 0 : if (policy != p->policy && !rlim_rtprio)
7538 : goto req_priv;
7539 :
7540 : /* Can't increase priority: */
7541 0 : if (attr->sched_priority > p->rt_priority &&
7542 0 : attr->sched_priority > rlim_rtprio)
7543 : goto req_priv;
7544 : }
7545 :
7546 : /*
7547 : * Can't set/change SCHED_DEADLINE policy at all for now
7548 : * (safest behavior); in the future we would like to allow
7549 : * unprivileged DL tasks to increase their relative deadline
7550 : * or reduce their runtime (both ways reducing utilization)
7551 : */
7552 0 : if (dl_policy(policy))
7553 : goto req_priv;
7554 :
7555 : /*
7556 : * Treat SCHED_IDLE as nice 20. Only allow a switch to
7557 : * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7558 : */
7559 0 : if (task_has_idle_policy(p) && !idle_policy(policy)) {
7560 0 : if (!is_nice_reduction(p, task_nice(p)))
7561 : goto req_priv;
7562 : }
7563 :
7564 : /* Can't change other user's priorities: */
7565 0 : if (!check_same_owner(p))
7566 : goto req_priv;
7567 :
7568 : /* Normal users shall not reset the sched_reset_on_fork flag: */
7569 0 : if (p->sched_reset_on_fork && !reset_on_fork)
7570 : goto req_priv;
7571 :
7572 : return 0;
7573 :
7574 : req_priv:
7575 0 : if (!capable(CAP_SYS_NICE))
7576 : return -EPERM;
7577 :
7578 : return 0;
7579 : }
7580 :
7581 380 : static int __sched_setscheduler(struct task_struct *p,
7582 : const struct sched_attr *attr,
7583 : bool user, bool pi)
7584 : {
7585 380 : int oldpolicy = -1, policy = attr->sched_policy;
7586 : int retval, oldprio, newprio, queued, running;
7587 : const struct sched_class *prev_class;
7588 : struct balance_callback *head;
7589 : struct rq_flags rf;
7590 : int reset_on_fork;
7591 380 : int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7592 : struct rq *rq;
7593 :
7594 : /* The pi code expects interrupts enabled */
7595 760 : BUG_ON(pi && in_interrupt());
7596 : recheck:
7597 : /* Double check policy once rq lock held: */
7598 380 : if (policy < 0) {
7599 0 : reset_on_fork = p->sched_reset_on_fork;
7600 0 : policy = oldpolicy = p->policy;
7601 : } else {
7602 380 : reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7603 :
7604 380 : if (!valid_policy(policy))
7605 : return -EINVAL;
7606 : }
7607 :
7608 380 : if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7609 : return -EINVAL;
7610 :
7611 : /*
7612 : * Valid priorities for SCHED_FIFO and SCHED_RR are
7613 : * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7614 : * SCHED_BATCH and SCHED_IDLE is 0.
7615 : */
7616 380 : if (attr->sched_priority > MAX_RT_PRIO-1)
7617 : return -EINVAL;
7618 760 : if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7619 380 : (rt_policy(policy) != (attr->sched_priority != 0)))
7620 : return -EINVAL;
7621 :
7622 380 : if (user) {
7623 0 : retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7624 0 : if (retval)
7625 : return retval;
7626 :
7627 0 : if (attr->sched_flags & SCHED_FLAG_SUGOV)
7628 : return -EINVAL;
7629 :
7630 0 : retval = security_task_setscheduler(p);
7631 0 : if (retval)
7632 : return retval;
7633 : }
7634 :
7635 : /* Update task specific "requested" clamps */
7636 380 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7637 : retval = uclamp_validate(p, attr);
7638 : if (retval)
7639 : return retval;
7640 : }
7641 :
7642 : if (pi)
7643 : cpuset_read_lock();
7644 :
7645 : /*
7646 : * Make sure no PI-waiters arrive (or leave) while we are
7647 : * changing the priority of the task:
7648 : *
7649 : * To be able to change p->policy safely, the appropriate
7650 : * runqueue lock must be held.
7651 : */
7652 380 : rq = task_rq_lock(p, &rf);
7653 380 : update_rq_clock(rq);
7654 :
7655 : /*
7656 : * Changing the policy of the stop threads its a very bad idea:
7657 : */
7658 380 : if (p == rq->stop) {
7659 : retval = -EINVAL;
7660 : goto unlock;
7661 : }
7662 :
7663 : /*
7664 : * If not changing anything there's no need to proceed further,
7665 : * but store a possible modification of reset_on_fork.
7666 : */
7667 380 : if (unlikely(policy == p->policy)) {
7668 760 : if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7669 : goto change;
7670 380 : if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7671 : goto change;
7672 380 : if (dl_policy(policy) && dl_param_changed(p, attr))
7673 : goto change;
7674 380 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7675 : goto change;
7676 :
7677 380 : p->sched_reset_on_fork = reset_on_fork;
7678 380 : retval = 0;
7679 380 : goto unlock;
7680 : }
7681 : change:
7682 :
7683 : if (user) {
7684 : #ifdef CONFIG_RT_GROUP_SCHED
7685 : /*
7686 : * Do not allow realtime tasks into groups that have no runtime
7687 : * assigned.
7688 : */
7689 : if (rt_bandwidth_enabled() && rt_policy(policy) &&
7690 : task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7691 : !task_group_is_autogroup(task_group(p))) {
7692 : retval = -EPERM;
7693 : goto unlock;
7694 : }
7695 : #endif
7696 : #ifdef CONFIG_SMP
7697 : if (dl_bandwidth_enabled() && dl_policy(policy) &&
7698 : !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7699 : cpumask_t *span = rq->rd->span;
7700 :
7701 : /*
7702 : * Don't allow tasks with an affinity mask smaller than
7703 : * the entire root_domain to become SCHED_DEADLINE. We
7704 : * will also fail if there's no bandwidth available.
7705 : */
7706 : if (!cpumask_subset(span, p->cpus_ptr) ||
7707 : rq->rd->dl_bw.bw == 0) {
7708 : retval = -EPERM;
7709 : goto unlock;
7710 : }
7711 : }
7712 : #endif
7713 : }
7714 :
7715 : /* Re-check policy now with rq lock held: */
7716 0 : if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7717 0 : policy = oldpolicy = -1;
7718 0 : task_rq_unlock(rq, p, &rf);
7719 : if (pi)
7720 : cpuset_read_unlock();
7721 : goto recheck;
7722 : }
7723 :
7724 : /*
7725 : * If setscheduling to SCHED_DEADLINE (or changing the parameters
7726 : * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7727 : * is available.
7728 : */
7729 0 : if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7730 : retval = -EBUSY;
7731 : goto unlock;
7732 : }
7733 :
7734 0 : p->sched_reset_on_fork = reset_on_fork;
7735 0 : oldprio = p->prio;
7736 :
7737 0 : newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7738 0 : if (pi) {
7739 : /*
7740 : * Take priority boosted tasks into account. If the new
7741 : * effective priority is unchanged, we just store the new
7742 : * normal parameters and do not touch the scheduler class and
7743 : * the runqueue. This will be done when the task deboost
7744 : * itself.
7745 : */
7746 0 : newprio = rt_effective_prio(p, newprio);
7747 0 : if (newprio == oldprio)
7748 0 : queue_flags &= ~DEQUEUE_MOVE;
7749 : }
7750 :
7751 0 : queued = task_on_rq_queued(p);
7752 0 : running = task_current(rq, p);
7753 0 : if (queued)
7754 0 : dequeue_task(rq, p, queue_flags);
7755 0 : if (running)
7756 0 : put_prev_task(rq, p);
7757 :
7758 0 : prev_class = p->sched_class;
7759 :
7760 0 : if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7761 0 : __setscheduler_params(p, attr);
7762 0 : __setscheduler_prio(p, newprio);
7763 : }
7764 0 : __setscheduler_uclamp(p, attr);
7765 :
7766 0 : if (queued) {
7767 : /*
7768 : * We enqueue to tail when the priority of a task is
7769 : * increased (user space view).
7770 : */
7771 0 : if (oldprio < p->prio)
7772 0 : queue_flags |= ENQUEUE_HEAD;
7773 :
7774 0 : enqueue_task(rq, p, queue_flags);
7775 : }
7776 0 : if (running)
7777 : set_next_task(rq, p);
7778 :
7779 0 : check_class_changed(rq, p, prev_class, oldprio);
7780 :
7781 : /* Avoid rq from going away on us: */
7782 0 : preempt_disable();
7783 0 : head = splice_balance_callbacks(rq);
7784 0 : task_rq_unlock(rq, p, &rf);
7785 :
7786 0 : if (pi) {
7787 : cpuset_read_unlock();
7788 0 : rt_mutex_adjust_pi(p);
7789 : }
7790 :
7791 : /* Run balance callbacks after we've adjusted the PI chain: */
7792 0 : balance_callbacks(rq, head);
7793 0 : preempt_enable();
7794 :
7795 0 : return 0;
7796 :
7797 : unlock:
7798 760 : task_rq_unlock(rq, p, &rf);
7799 : if (pi)
7800 : cpuset_read_unlock();
7801 380 : return retval;
7802 : }
7803 :
7804 380 : static int _sched_setscheduler(struct task_struct *p, int policy,
7805 : const struct sched_param *param, bool check)
7806 : {
7807 1140 : struct sched_attr attr = {
7808 : .sched_policy = policy,
7809 380 : .sched_priority = param->sched_priority,
7810 380 : .sched_nice = PRIO_TO_NICE(p->static_prio),
7811 : };
7812 :
7813 : /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7814 380 : if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7815 0 : attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7816 0 : policy &= ~SCHED_RESET_ON_FORK;
7817 0 : attr.sched_policy = policy;
7818 : }
7819 :
7820 380 : return __sched_setscheduler(p, &attr, check, true);
7821 : }
7822 : /**
7823 : * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7824 : * @p: the task in question.
7825 : * @policy: new policy.
7826 : * @param: structure containing the new RT priority.
7827 : *
7828 : * Use sched_set_fifo(), read its comment.
7829 : *
7830 : * Return: 0 on success. An error code otherwise.
7831 : *
7832 : * NOTE that the task may be already dead.
7833 : */
7834 0 : int sched_setscheduler(struct task_struct *p, int policy,
7835 : const struct sched_param *param)
7836 : {
7837 0 : return _sched_setscheduler(p, policy, param, true);
7838 : }
7839 :
7840 0 : int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7841 : {
7842 0 : return __sched_setscheduler(p, attr, true, true);
7843 : }
7844 :
7845 0 : int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7846 : {
7847 0 : return __sched_setscheduler(p, attr, false, true);
7848 : }
7849 : EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7850 :
7851 : /**
7852 : * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7853 : * @p: the task in question.
7854 : * @policy: new policy.
7855 : * @param: structure containing the new RT priority.
7856 : *
7857 : * Just like sched_setscheduler, only don't bother checking if the
7858 : * current context has permission. For example, this is needed in
7859 : * stop_machine(): we create temporary high priority worker threads,
7860 : * but our caller might not have that capability.
7861 : *
7862 : * Return: 0 on success. An error code otherwise.
7863 : */
7864 380 : int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7865 : const struct sched_param *param)
7866 : {
7867 380 : return _sched_setscheduler(p, policy, param, false);
7868 : }
7869 :
7870 : /*
7871 : * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7872 : * incapable of resource management, which is the one thing an OS really should
7873 : * be doing.
7874 : *
7875 : * This is of course the reason it is limited to privileged users only.
7876 : *
7877 : * Worse still; it is fundamentally impossible to compose static priority
7878 : * workloads. You cannot take two correctly working static prio workloads
7879 : * and smash them together and still expect them to work.
7880 : *
7881 : * For this reason 'all' FIFO tasks the kernel creates are basically at:
7882 : *
7883 : * MAX_RT_PRIO / 2
7884 : *
7885 : * The administrator _MUST_ configure the system, the kernel simply doesn't
7886 : * know enough information to make a sensible choice.
7887 : */
7888 0 : void sched_set_fifo(struct task_struct *p)
7889 : {
7890 0 : struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7891 0 : WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7892 0 : }
7893 : EXPORT_SYMBOL_GPL(sched_set_fifo);
7894 :
7895 : /*
7896 : * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7897 : */
7898 0 : void sched_set_fifo_low(struct task_struct *p)
7899 : {
7900 0 : struct sched_param sp = { .sched_priority = 1 };
7901 0 : WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7902 0 : }
7903 : EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7904 :
7905 0 : void sched_set_normal(struct task_struct *p, int nice)
7906 : {
7907 0 : struct sched_attr attr = {
7908 : .sched_policy = SCHED_NORMAL,
7909 : .sched_nice = nice,
7910 : };
7911 0 : WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7912 0 : }
7913 : EXPORT_SYMBOL_GPL(sched_set_normal);
7914 :
7915 : static int
7916 0 : do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7917 : {
7918 : struct sched_param lparam;
7919 : struct task_struct *p;
7920 : int retval;
7921 :
7922 0 : if (!param || pid < 0)
7923 : return -EINVAL;
7924 0 : if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7925 : return -EFAULT;
7926 :
7927 : rcu_read_lock();
7928 0 : retval = -ESRCH;
7929 0 : p = find_process_by_pid(pid);
7930 0 : if (likely(p))
7931 : get_task_struct(p);
7932 : rcu_read_unlock();
7933 :
7934 0 : if (likely(p)) {
7935 0 : retval = sched_setscheduler(p, policy, &lparam);
7936 0 : put_task_struct(p);
7937 : }
7938 :
7939 : return retval;
7940 : }
7941 :
7942 : /*
7943 : * Mimics kernel/events/core.c perf_copy_attr().
7944 : */
7945 0 : static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7946 : {
7947 : u32 size;
7948 : int ret;
7949 :
7950 : /* Zero the full structure, so that a short copy will be nice: */
7951 0 : memset(attr, 0, sizeof(*attr));
7952 :
7953 0 : ret = get_user(size, &uattr->size);
7954 0 : if (ret)
7955 : return ret;
7956 :
7957 : /* ABI compatibility quirk: */
7958 0 : if (!size)
7959 0 : size = SCHED_ATTR_SIZE_VER0;
7960 0 : if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7961 : goto err_size;
7962 :
7963 0 : ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7964 0 : if (ret) {
7965 0 : if (ret == -E2BIG)
7966 : goto err_size;
7967 : return ret;
7968 : }
7969 :
7970 0 : if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7971 : size < SCHED_ATTR_SIZE_VER1)
7972 : return -EINVAL;
7973 :
7974 : /*
7975 : * XXX: Do we want to be lenient like existing syscalls; or do we want
7976 : * to be strict and return an error on out-of-bounds values?
7977 : */
7978 0 : attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7979 :
7980 0 : return 0;
7981 :
7982 : err_size:
7983 0 : put_user(sizeof(*attr), &uattr->size);
7984 : return -E2BIG;
7985 : }
7986 :
7987 0 : static void get_params(struct task_struct *p, struct sched_attr *attr)
7988 : {
7989 0 : if (task_has_dl_policy(p))
7990 0 : __getparam_dl(p, attr);
7991 0 : else if (task_has_rt_policy(p))
7992 0 : attr->sched_priority = p->rt_priority;
7993 : else
7994 0 : attr->sched_nice = task_nice(p);
7995 0 : }
7996 :
7997 : /**
7998 : * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7999 : * @pid: the pid in question.
8000 : * @policy: new policy.
8001 : * @param: structure containing the new RT priority.
8002 : *
8003 : * Return: 0 on success. An error code otherwise.
8004 : */
8005 0 : SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
8006 : {
8007 0 : if (policy < 0)
8008 : return -EINVAL;
8009 :
8010 0 : return do_sched_setscheduler(pid, policy, param);
8011 : }
8012 :
8013 : /**
8014 : * sys_sched_setparam - set/change the RT priority of a thread
8015 : * @pid: the pid in question.
8016 : * @param: structure containing the new RT priority.
8017 : *
8018 : * Return: 0 on success. An error code otherwise.
8019 : */
8020 0 : SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
8021 : {
8022 0 : return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
8023 : }
8024 :
8025 : /**
8026 : * sys_sched_setattr - same as above, but with extended sched_attr
8027 : * @pid: the pid in question.
8028 : * @uattr: structure containing the extended parameters.
8029 : * @flags: for future extension.
8030 : */
8031 0 : SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
8032 : unsigned int, flags)
8033 : {
8034 : struct sched_attr attr;
8035 : struct task_struct *p;
8036 : int retval;
8037 :
8038 0 : if (!uattr || pid < 0 || flags)
8039 : return -EINVAL;
8040 :
8041 0 : retval = sched_copy_attr(uattr, &attr);
8042 0 : if (retval)
8043 0 : return retval;
8044 :
8045 0 : if ((int)attr.sched_policy < 0)
8046 : return -EINVAL;
8047 0 : if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
8048 0 : attr.sched_policy = SETPARAM_POLICY;
8049 :
8050 : rcu_read_lock();
8051 0 : retval = -ESRCH;
8052 0 : p = find_process_by_pid(pid);
8053 0 : if (likely(p))
8054 : get_task_struct(p);
8055 : rcu_read_unlock();
8056 :
8057 0 : if (likely(p)) {
8058 0 : if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
8059 0 : get_params(p, &attr);
8060 0 : retval = sched_setattr(p, &attr);
8061 0 : put_task_struct(p);
8062 : }
8063 :
8064 0 : return retval;
8065 : }
8066 :
8067 : /**
8068 : * sys_sched_getscheduler - get the policy (scheduling class) of a thread
8069 : * @pid: the pid in question.
8070 : *
8071 : * Return: On success, the policy of the thread. Otherwise, a negative error
8072 : * code.
8073 : */
8074 0 : SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
8075 : {
8076 : struct task_struct *p;
8077 : int retval;
8078 :
8079 0 : if (pid < 0)
8080 : return -EINVAL;
8081 :
8082 0 : retval = -ESRCH;
8083 : rcu_read_lock();
8084 0 : p = find_process_by_pid(pid);
8085 0 : if (p) {
8086 0 : retval = security_task_getscheduler(p);
8087 : if (!retval)
8088 0 : retval = p->policy
8089 0 : | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
8090 : }
8091 : rcu_read_unlock();
8092 0 : return retval;
8093 : }
8094 :
8095 : /**
8096 : * sys_sched_getparam - get the RT priority of a thread
8097 : * @pid: the pid in question.
8098 : * @param: structure containing the RT priority.
8099 : *
8100 : * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8101 : * code.
8102 : */
8103 0 : SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
8104 : {
8105 0 : struct sched_param lp = { .sched_priority = 0 };
8106 : struct task_struct *p;
8107 : int retval;
8108 :
8109 0 : if (!param || pid < 0)
8110 : return -EINVAL;
8111 :
8112 : rcu_read_lock();
8113 0 : p = find_process_by_pid(pid);
8114 0 : retval = -ESRCH;
8115 0 : if (!p)
8116 : goto out_unlock;
8117 :
8118 0 : retval = security_task_getscheduler(p);
8119 : if (retval)
8120 : goto out_unlock;
8121 :
8122 0 : if (task_has_rt_policy(p))
8123 0 : lp.sched_priority = p->rt_priority;
8124 0 : rcu_read_unlock();
8125 :
8126 : /*
8127 : * This one might sleep, we cannot do it with a spinlock held ...
8128 : */
8129 0 : retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
8130 :
8131 0 : return retval;
8132 :
8133 : out_unlock:
8134 : rcu_read_unlock();
8135 0 : return retval;
8136 : }
8137 :
8138 : /*
8139 : * Copy the kernel size attribute structure (which might be larger
8140 : * than what user-space knows about) to user-space.
8141 : *
8142 : * Note that all cases are valid: user-space buffer can be larger or
8143 : * smaller than the kernel-space buffer. The usual case is that both
8144 : * have the same size.
8145 : */
8146 : static int
8147 0 : sched_attr_copy_to_user(struct sched_attr __user *uattr,
8148 : struct sched_attr *kattr,
8149 : unsigned int usize)
8150 : {
8151 0 : unsigned int ksize = sizeof(*kattr);
8152 :
8153 0 : if (!access_ok(uattr, usize))
8154 : return -EFAULT;
8155 :
8156 : /*
8157 : * sched_getattr() ABI forwards and backwards compatibility:
8158 : *
8159 : * If usize == ksize then we just copy everything to user-space and all is good.
8160 : *
8161 : * If usize < ksize then we only copy as much as user-space has space for,
8162 : * this keeps ABI compatibility as well. We skip the rest.
8163 : *
8164 : * If usize > ksize then user-space is using a newer version of the ABI,
8165 : * which part the kernel doesn't know about. Just ignore it - tooling can
8166 : * detect the kernel's knowledge of attributes from the attr->size value
8167 : * which is set to ksize in this case.
8168 : */
8169 0 : kattr->size = min(usize, ksize);
8170 :
8171 0 : if (copy_to_user(uattr, kattr, kattr->size))
8172 : return -EFAULT;
8173 :
8174 0 : return 0;
8175 : }
8176 :
8177 : /**
8178 : * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8179 : * @pid: the pid in question.
8180 : * @uattr: structure containing the extended parameters.
8181 : * @usize: sizeof(attr) for fwd/bwd comp.
8182 : * @flags: for future extension.
8183 : */
8184 0 : SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8185 : unsigned int, usize, unsigned int, flags)
8186 : {
8187 0 : struct sched_attr kattr = { };
8188 : struct task_struct *p;
8189 : int retval;
8190 :
8191 0 : if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8192 0 : usize < SCHED_ATTR_SIZE_VER0 || flags)
8193 : return -EINVAL;
8194 :
8195 : rcu_read_lock();
8196 0 : p = find_process_by_pid(pid);
8197 0 : retval = -ESRCH;
8198 0 : if (!p)
8199 : goto out_unlock;
8200 :
8201 0 : retval = security_task_getscheduler(p);
8202 : if (retval)
8203 : goto out_unlock;
8204 :
8205 0 : kattr.sched_policy = p->policy;
8206 0 : if (p->sched_reset_on_fork)
8207 0 : kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8208 0 : get_params(p, &kattr);
8209 0 : kattr.sched_flags &= SCHED_FLAG_ALL;
8210 :
8211 : #ifdef CONFIG_UCLAMP_TASK
8212 : /*
8213 : * This could race with another potential updater, but this is fine
8214 : * because it'll correctly read the old or the new value. We don't need
8215 : * to guarantee who wins the race as long as it doesn't return garbage.
8216 : */
8217 : kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8218 : kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8219 : #endif
8220 :
8221 : rcu_read_unlock();
8222 :
8223 0 : return sched_attr_copy_to_user(uattr, &kattr, usize);
8224 :
8225 : out_unlock:
8226 : rcu_read_unlock();
8227 0 : return retval;
8228 : }
8229 :
8230 : #ifdef CONFIG_SMP
8231 : int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8232 : {
8233 : int ret = 0;
8234 :
8235 : /*
8236 : * If the task isn't a deadline task or admission control is
8237 : * disabled then we don't care about affinity changes.
8238 : */
8239 : if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8240 : return 0;
8241 :
8242 : /*
8243 : * Since bandwidth control happens on root_domain basis,
8244 : * if admission test is enabled, we only admit -deadline
8245 : * tasks allowed to run on all the CPUs in the task's
8246 : * root_domain.
8247 : */
8248 : rcu_read_lock();
8249 : if (!cpumask_subset(task_rq(p)->rd->span, mask))
8250 : ret = -EBUSY;
8251 : rcu_read_unlock();
8252 : return ret;
8253 : }
8254 : #endif
8255 :
8256 : static int
8257 0 : __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
8258 : {
8259 : int retval;
8260 : cpumask_var_t cpus_allowed, new_mask;
8261 :
8262 0 : if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8263 : return -ENOMEM;
8264 :
8265 0 : if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8266 : retval = -ENOMEM;
8267 : goto out_free_cpus_allowed;
8268 : }
8269 :
8270 0 : cpuset_cpus_allowed(p, cpus_allowed);
8271 0 : cpumask_and(new_mask, ctx->new_mask, cpus_allowed);
8272 :
8273 0 : ctx->new_mask = new_mask;
8274 0 : ctx->flags |= SCA_CHECK;
8275 :
8276 0 : retval = dl_task_check_affinity(p, new_mask);
8277 : if (retval)
8278 : goto out_free_new_mask;
8279 :
8280 0 : retval = __set_cpus_allowed_ptr(p, ctx);
8281 0 : if (retval)
8282 : goto out_free_new_mask;
8283 :
8284 0 : cpuset_cpus_allowed(p, cpus_allowed);
8285 0 : if (!cpumask_subset(new_mask, cpus_allowed)) {
8286 : /*
8287 : * We must have raced with a concurrent cpuset update.
8288 : * Just reset the cpumask to the cpuset's cpus_allowed.
8289 : */
8290 0 : cpumask_copy(new_mask, cpus_allowed);
8291 :
8292 : /*
8293 : * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8294 : * will restore the previous user_cpus_ptr value.
8295 : *
8296 : * In the unlikely event a previous user_cpus_ptr exists,
8297 : * we need to further restrict the mask to what is allowed
8298 : * by that old user_cpus_ptr.
8299 : */
8300 0 : if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
8301 0 : bool empty = !cpumask_and(new_mask, new_mask,
8302 0 : ctx->user_mask);
8303 :
8304 0 : if (WARN_ON_ONCE(empty))
8305 : cpumask_copy(new_mask, cpus_allowed);
8306 : }
8307 0 : __set_cpus_allowed_ptr(p, ctx);
8308 : retval = -EINVAL;
8309 : }
8310 :
8311 : out_free_new_mask:
8312 0 : free_cpumask_var(new_mask);
8313 : out_free_cpus_allowed:
8314 0 : free_cpumask_var(cpus_allowed);
8315 : return retval;
8316 : }
8317 :
8318 0 : long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8319 : {
8320 : struct affinity_context ac;
8321 : struct cpumask *user_mask;
8322 : struct task_struct *p;
8323 : int retval;
8324 :
8325 : rcu_read_lock();
8326 :
8327 0 : p = find_process_by_pid(pid);
8328 0 : if (!p) {
8329 : rcu_read_unlock();
8330 0 : return -ESRCH;
8331 : }
8332 :
8333 : /* Prevent p going away */
8334 0 : get_task_struct(p);
8335 : rcu_read_unlock();
8336 :
8337 0 : if (p->flags & PF_NO_SETAFFINITY) {
8338 : retval = -EINVAL;
8339 : goto out_put_task;
8340 : }
8341 :
8342 0 : if (!check_same_owner(p)) {
8343 : rcu_read_lock();
8344 0 : if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8345 : rcu_read_unlock();
8346 0 : retval = -EPERM;
8347 0 : goto out_put_task;
8348 : }
8349 : rcu_read_unlock();
8350 : }
8351 :
8352 0 : retval = security_task_setscheduler(p);
8353 0 : if (retval)
8354 : goto out_put_task;
8355 :
8356 : /*
8357 : * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8358 : * alloc_user_cpus_ptr() returns NULL.
8359 : */
8360 0 : user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
8361 : if (user_mask) {
8362 : cpumask_copy(user_mask, in_mask);
8363 : } else if (IS_ENABLED(CONFIG_SMP)) {
8364 : retval = -ENOMEM;
8365 : goto out_put_task;
8366 : }
8367 :
8368 0 : ac = (struct affinity_context){
8369 : .new_mask = in_mask,
8370 : .user_mask = user_mask,
8371 : .flags = SCA_USER,
8372 : };
8373 :
8374 0 : retval = __sched_setaffinity(p, &ac);
8375 0 : kfree(ac.user_mask);
8376 :
8377 : out_put_task:
8378 0 : put_task_struct(p);
8379 0 : return retval;
8380 : }
8381 :
8382 0 : static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8383 : struct cpumask *new_mask)
8384 : {
8385 0 : if (len < cpumask_size())
8386 : cpumask_clear(new_mask);
8387 0 : else if (len > cpumask_size())
8388 0 : len = cpumask_size();
8389 :
8390 0 : return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8391 : }
8392 :
8393 : /**
8394 : * sys_sched_setaffinity - set the CPU affinity of a process
8395 : * @pid: pid of the process
8396 : * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8397 : * @user_mask_ptr: user-space pointer to the new CPU mask
8398 : *
8399 : * Return: 0 on success. An error code otherwise.
8400 : */
8401 0 : SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8402 : unsigned long __user *, user_mask_ptr)
8403 : {
8404 : cpumask_var_t new_mask;
8405 : int retval;
8406 :
8407 0 : if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8408 : return -ENOMEM;
8409 :
8410 0 : retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8411 0 : if (retval == 0)
8412 0 : retval = sched_setaffinity(pid, new_mask);
8413 0 : free_cpumask_var(new_mask);
8414 0 : return retval;
8415 : }
8416 :
8417 0 : long sched_getaffinity(pid_t pid, struct cpumask *mask)
8418 : {
8419 : struct task_struct *p;
8420 : unsigned long flags;
8421 : int retval;
8422 :
8423 : rcu_read_lock();
8424 :
8425 0 : retval = -ESRCH;
8426 0 : p = find_process_by_pid(pid);
8427 0 : if (!p)
8428 : goto out_unlock;
8429 :
8430 0 : retval = security_task_getscheduler(p);
8431 : if (retval)
8432 : goto out_unlock;
8433 :
8434 0 : raw_spin_lock_irqsave(&p->pi_lock, flags);
8435 0 : cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8436 0 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8437 :
8438 : out_unlock:
8439 : rcu_read_unlock();
8440 :
8441 0 : return retval;
8442 : }
8443 :
8444 : /**
8445 : * sys_sched_getaffinity - get the CPU affinity of a process
8446 : * @pid: pid of the process
8447 : * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8448 : * @user_mask_ptr: user-space pointer to hold the current CPU mask
8449 : *
8450 : * Return: size of CPU mask copied to user_mask_ptr on success. An
8451 : * error code otherwise.
8452 : */
8453 0 : SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8454 : unsigned long __user *, user_mask_ptr)
8455 : {
8456 : int ret;
8457 : cpumask_var_t mask;
8458 :
8459 0 : if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8460 : return -EINVAL;
8461 0 : if (len & (sizeof(unsigned long)-1))
8462 : return -EINVAL;
8463 :
8464 0 : if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
8465 : return -ENOMEM;
8466 :
8467 0 : ret = sched_getaffinity(pid, mask);
8468 0 : if (ret == 0) {
8469 0 : unsigned int retlen = min(len, cpumask_size());
8470 :
8471 0 : if (copy_to_user(user_mask_ptr, cpumask_bits(mask), retlen))
8472 : ret = -EFAULT;
8473 : else
8474 0 : ret = retlen;
8475 : }
8476 0 : free_cpumask_var(mask);
8477 :
8478 0 : return ret;
8479 : }
8480 :
8481 0 : static void do_sched_yield(void)
8482 : {
8483 : struct rq_flags rf;
8484 : struct rq *rq;
8485 :
8486 0 : rq = this_rq_lock_irq(&rf);
8487 :
8488 : schedstat_inc(rq->yld_count);
8489 0 : current->sched_class->yield_task(rq);
8490 :
8491 0 : preempt_disable();
8492 0 : rq_unlock_irq(rq, &rf);
8493 0 : sched_preempt_enable_no_resched();
8494 :
8495 0 : schedule();
8496 0 : }
8497 :
8498 : /**
8499 : * sys_sched_yield - yield the current processor to other threads.
8500 : *
8501 : * This function yields the current CPU to other tasks. If there are no
8502 : * other threads running on this CPU then this function will return.
8503 : *
8504 : * Return: 0.
8505 : */
8506 0 : SYSCALL_DEFINE0(sched_yield)
8507 : {
8508 0 : do_sched_yield();
8509 0 : return 0;
8510 : }
8511 :
8512 : #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8513 47766 : int __sched __cond_resched(void)
8514 : {
8515 47766 : if (should_resched(0)) {
8516 : preempt_schedule_common();
8517 : return 1;
8518 : }
8519 : /*
8520 : * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8521 : * whether the current CPU is in an RCU read-side critical section,
8522 : * so the tick can report quiescent states even for CPUs looping
8523 : * in kernel context. In contrast, in non-preemptible kernels,
8524 : * RCU readers leave no in-memory hints, which means that CPU-bound
8525 : * processes executing in kernel context might never report an
8526 : * RCU quiescent state. Therefore, the following code causes
8527 : * cond_resched() to report a quiescent state, but only when RCU
8528 : * is in urgent need of one.
8529 : */
8530 : #ifndef CONFIG_PREEMPT_RCU
8531 : rcu_all_qs();
8532 : #endif
8533 47694 : return 0;
8534 : }
8535 : EXPORT_SYMBOL(__cond_resched);
8536 : #endif
8537 :
8538 : #ifdef CONFIG_PREEMPT_DYNAMIC
8539 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8540 : #define cond_resched_dynamic_enabled __cond_resched
8541 : #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8542 : DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8543 : EXPORT_STATIC_CALL_TRAMP(cond_resched);
8544 :
8545 : #define might_resched_dynamic_enabled __cond_resched
8546 : #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8547 : DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8548 : EXPORT_STATIC_CALL_TRAMP(might_resched);
8549 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8550 : static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8551 : int __sched dynamic_cond_resched(void)
8552 : {
8553 : klp_sched_try_switch();
8554 : if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8555 : return 0;
8556 : return __cond_resched();
8557 : }
8558 : EXPORT_SYMBOL(dynamic_cond_resched);
8559 :
8560 : static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8561 : int __sched dynamic_might_resched(void)
8562 : {
8563 : if (!static_branch_unlikely(&sk_dynamic_might_resched))
8564 : return 0;
8565 : return __cond_resched();
8566 : }
8567 : EXPORT_SYMBOL(dynamic_might_resched);
8568 : #endif
8569 : #endif
8570 :
8571 : /*
8572 : * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8573 : * call schedule, and on return reacquire the lock.
8574 : *
8575 : * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8576 : * operations here to prevent schedule() from being called twice (once via
8577 : * spin_unlock(), once by hand).
8578 : */
8579 5 : int __cond_resched_lock(spinlock_t *lock)
8580 : {
8581 5 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
8582 5 : int ret = 0;
8583 :
8584 : lockdep_assert_held(lock);
8585 :
8586 5 : if (spin_needbreak(lock) || resched) {
8587 2 : spin_unlock(lock);
8588 2 : if (!_cond_resched())
8589 : cpu_relax();
8590 2 : ret = 1;
8591 : spin_lock(lock);
8592 : }
8593 5 : return ret;
8594 : }
8595 : EXPORT_SYMBOL(__cond_resched_lock);
8596 :
8597 0 : int __cond_resched_rwlock_read(rwlock_t *lock)
8598 : {
8599 0 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
8600 0 : int ret = 0;
8601 :
8602 : lockdep_assert_held_read(lock);
8603 :
8604 0 : if (rwlock_needbreak(lock) || resched) {
8605 0 : read_unlock(lock);
8606 0 : if (!_cond_resched())
8607 : cpu_relax();
8608 0 : ret = 1;
8609 0 : read_lock(lock);
8610 : }
8611 0 : return ret;
8612 : }
8613 : EXPORT_SYMBOL(__cond_resched_rwlock_read);
8614 :
8615 0 : int __cond_resched_rwlock_write(rwlock_t *lock)
8616 : {
8617 0 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
8618 0 : int ret = 0;
8619 :
8620 : lockdep_assert_held_write(lock);
8621 :
8622 0 : if (rwlock_needbreak(lock) || resched) {
8623 0 : write_unlock(lock);
8624 0 : if (!_cond_resched())
8625 : cpu_relax();
8626 0 : ret = 1;
8627 0 : write_lock(lock);
8628 : }
8629 0 : return ret;
8630 : }
8631 : EXPORT_SYMBOL(__cond_resched_rwlock_write);
8632 :
8633 : #ifdef CONFIG_PREEMPT_DYNAMIC
8634 :
8635 : #ifdef CONFIG_GENERIC_ENTRY
8636 : #include <linux/entry-common.h>
8637 : #endif
8638 :
8639 : /*
8640 : * SC:cond_resched
8641 : * SC:might_resched
8642 : * SC:preempt_schedule
8643 : * SC:preempt_schedule_notrace
8644 : * SC:irqentry_exit_cond_resched
8645 : *
8646 : *
8647 : * NONE:
8648 : * cond_resched <- __cond_resched
8649 : * might_resched <- RET0
8650 : * preempt_schedule <- NOP
8651 : * preempt_schedule_notrace <- NOP
8652 : * irqentry_exit_cond_resched <- NOP
8653 : *
8654 : * VOLUNTARY:
8655 : * cond_resched <- __cond_resched
8656 : * might_resched <- __cond_resched
8657 : * preempt_schedule <- NOP
8658 : * preempt_schedule_notrace <- NOP
8659 : * irqentry_exit_cond_resched <- NOP
8660 : *
8661 : * FULL:
8662 : * cond_resched <- RET0
8663 : * might_resched <- RET0
8664 : * preempt_schedule <- preempt_schedule
8665 : * preempt_schedule_notrace <- preempt_schedule_notrace
8666 : * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8667 : */
8668 :
8669 : enum {
8670 : preempt_dynamic_undefined = -1,
8671 : preempt_dynamic_none,
8672 : preempt_dynamic_voluntary,
8673 : preempt_dynamic_full,
8674 : };
8675 :
8676 : int preempt_dynamic_mode = preempt_dynamic_undefined;
8677 :
8678 : int sched_dynamic_mode(const char *str)
8679 : {
8680 : if (!strcmp(str, "none"))
8681 : return preempt_dynamic_none;
8682 :
8683 : if (!strcmp(str, "voluntary"))
8684 : return preempt_dynamic_voluntary;
8685 :
8686 : if (!strcmp(str, "full"))
8687 : return preempt_dynamic_full;
8688 :
8689 : return -EINVAL;
8690 : }
8691 :
8692 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8693 : #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8694 : #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8695 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8696 : #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8697 : #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8698 : #else
8699 : #error "Unsupported PREEMPT_DYNAMIC mechanism"
8700 : #endif
8701 :
8702 : static DEFINE_MUTEX(sched_dynamic_mutex);
8703 : static bool klp_override;
8704 :
8705 : static void __sched_dynamic_update(int mode)
8706 : {
8707 : /*
8708 : * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8709 : * the ZERO state, which is invalid.
8710 : */
8711 : if (!klp_override)
8712 : preempt_dynamic_enable(cond_resched);
8713 : preempt_dynamic_enable(might_resched);
8714 : preempt_dynamic_enable(preempt_schedule);
8715 : preempt_dynamic_enable(preempt_schedule_notrace);
8716 : preempt_dynamic_enable(irqentry_exit_cond_resched);
8717 :
8718 : switch (mode) {
8719 : case preempt_dynamic_none:
8720 : if (!klp_override)
8721 : preempt_dynamic_enable(cond_resched);
8722 : preempt_dynamic_disable(might_resched);
8723 : preempt_dynamic_disable(preempt_schedule);
8724 : preempt_dynamic_disable(preempt_schedule_notrace);
8725 : preempt_dynamic_disable(irqentry_exit_cond_resched);
8726 : if (mode != preempt_dynamic_mode)
8727 : pr_info("Dynamic Preempt: none\n");
8728 : break;
8729 :
8730 : case preempt_dynamic_voluntary:
8731 : if (!klp_override)
8732 : preempt_dynamic_enable(cond_resched);
8733 : preempt_dynamic_enable(might_resched);
8734 : preempt_dynamic_disable(preempt_schedule);
8735 : preempt_dynamic_disable(preempt_schedule_notrace);
8736 : preempt_dynamic_disable(irqentry_exit_cond_resched);
8737 : if (mode != preempt_dynamic_mode)
8738 : pr_info("Dynamic Preempt: voluntary\n");
8739 : break;
8740 :
8741 : case preempt_dynamic_full:
8742 : if (!klp_override)
8743 : preempt_dynamic_disable(cond_resched);
8744 : preempt_dynamic_disable(might_resched);
8745 : preempt_dynamic_enable(preempt_schedule);
8746 : preempt_dynamic_enable(preempt_schedule_notrace);
8747 : preempt_dynamic_enable(irqentry_exit_cond_resched);
8748 : if (mode != preempt_dynamic_mode)
8749 : pr_info("Dynamic Preempt: full\n");
8750 : break;
8751 : }
8752 :
8753 : preempt_dynamic_mode = mode;
8754 : }
8755 :
8756 : void sched_dynamic_update(int mode)
8757 : {
8758 : mutex_lock(&sched_dynamic_mutex);
8759 : __sched_dynamic_update(mode);
8760 : mutex_unlock(&sched_dynamic_mutex);
8761 : }
8762 :
8763 : #ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
8764 :
8765 : static int klp_cond_resched(void)
8766 : {
8767 : __klp_sched_try_switch();
8768 : return __cond_resched();
8769 : }
8770 :
8771 : void sched_dynamic_klp_enable(void)
8772 : {
8773 : mutex_lock(&sched_dynamic_mutex);
8774 :
8775 : klp_override = true;
8776 : static_call_update(cond_resched, klp_cond_resched);
8777 :
8778 : mutex_unlock(&sched_dynamic_mutex);
8779 : }
8780 :
8781 : void sched_dynamic_klp_disable(void)
8782 : {
8783 : mutex_lock(&sched_dynamic_mutex);
8784 :
8785 : klp_override = false;
8786 : __sched_dynamic_update(preempt_dynamic_mode);
8787 :
8788 : mutex_unlock(&sched_dynamic_mutex);
8789 : }
8790 :
8791 : #endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
8792 :
8793 : static int __init setup_preempt_mode(char *str)
8794 : {
8795 : int mode = sched_dynamic_mode(str);
8796 : if (mode < 0) {
8797 : pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8798 : return 0;
8799 : }
8800 :
8801 : sched_dynamic_update(mode);
8802 : return 1;
8803 : }
8804 : __setup("preempt=", setup_preempt_mode);
8805 :
8806 : static void __init preempt_dynamic_init(void)
8807 : {
8808 : if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8809 : if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8810 : sched_dynamic_update(preempt_dynamic_none);
8811 : } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8812 : sched_dynamic_update(preempt_dynamic_voluntary);
8813 : } else {
8814 : /* Default static call setting, nothing to do */
8815 : WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8816 : preempt_dynamic_mode = preempt_dynamic_full;
8817 : pr_info("Dynamic Preempt: full\n");
8818 : }
8819 : }
8820 : }
8821 :
8822 : #define PREEMPT_MODEL_ACCESSOR(mode) \
8823 : bool preempt_model_##mode(void) \
8824 : { \
8825 : WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8826 : return preempt_dynamic_mode == preempt_dynamic_##mode; \
8827 : } \
8828 : EXPORT_SYMBOL_GPL(preempt_model_##mode)
8829 :
8830 : PREEMPT_MODEL_ACCESSOR(none);
8831 : PREEMPT_MODEL_ACCESSOR(voluntary);
8832 : PREEMPT_MODEL_ACCESSOR(full);
8833 :
8834 : #else /* !CONFIG_PREEMPT_DYNAMIC */
8835 :
8836 : static inline void preempt_dynamic_init(void) { }
8837 :
8838 : #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8839 :
8840 : /**
8841 : * yield - yield the current processor to other threads.
8842 : *
8843 : * Do not ever use this function, there's a 99% chance you're doing it wrong.
8844 : *
8845 : * The scheduler is at all times free to pick the calling task as the most
8846 : * eligible task to run, if removing the yield() call from your code breaks
8847 : * it, it's already broken.
8848 : *
8849 : * Typical broken usage is:
8850 : *
8851 : * while (!event)
8852 : * yield();
8853 : *
8854 : * where one assumes that yield() will let 'the other' process run that will
8855 : * make event true. If the current task is a SCHED_FIFO task that will never
8856 : * happen. Never use yield() as a progress guarantee!!
8857 : *
8858 : * If you want to use yield() to wait for something, use wait_event().
8859 : * If you want to use yield() to be 'nice' for others, use cond_resched().
8860 : * If you still want to use yield(), do not!
8861 : */
8862 0 : void __sched yield(void)
8863 : {
8864 0 : set_current_state(TASK_RUNNING);
8865 0 : do_sched_yield();
8866 0 : }
8867 : EXPORT_SYMBOL(yield);
8868 :
8869 : /**
8870 : * yield_to - yield the current processor to another thread in
8871 : * your thread group, or accelerate that thread toward the
8872 : * processor it's on.
8873 : * @p: target task
8874 : * @preempt: whether task preemption is allowed or not
8875 : *
8876 : * It's the caller's job to ensure that the target task struct
8877 : * can't go away on us before we can do any checks.
8878 : *
8879 : * Return:
8880 : * true (>0) if we indeed boosted the target task.
8881 : * false (0) if we failed to boost the target.
8882 : * -ESRCH if there's no task to yield to.
8883 : */
8884 0 : int __sched yield_to(struct task_struct *p, bool preempt)
8885 : {
8886 0 : struct task_struct *curr = current;
8887 : struct rq *rq, *p_rq;
8888 : unsigned long flags;
8889 0 : int yielded = 0;
8890 :
8891 0 : local_irq_save(flags);
8892 0 : rq = this_rq();
8893 :
8894 : again:
8895 0 : p_rq = task_rq(p);
8896 : /*
8897 : * If we're the only runnable task on the rq and target rq also
8898 : * has only one task, there's absolutely no point in yielding.
8899 : */
8900 0 : if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8901 : yielded = -ESRCH;
8902 : goto out_irq;
8903 : }
8904 :
8905 0 : double_rq_lock(rq, p_rq);
8906 0 : if (task_rq(p) != p_rq) {
8907 : double_rq_unlock(rq, p_rq);
8908 : goto again;
8909 : }
8910 :
8911 0 : if (!curr->sched_class->yield_to_task)
8912 : goto out_unlock;
8913 :
8914 0 : if (curr->sched_class != p->sched_class)
8915 : goto out_unlock;
8916 :
8917 0 : if (task_on_cpu(p_rq, p) || !task_is_running(p))
8918 : goto out_unlock;
8919 :
8920 0 : yielded = curr->sched_class->yield_to_task(rq, p);
8921 : if (yielded) {
8922 : schedstat_inc(rq->yld_count);
8923 : /*
8924 : * Make p's CPU reschedule; pick_next_entity takes care of
8925 : * fairness.
8926 : */
8927 : if (preempt && rq != p_rq)
8928 : resched_curr(p_rq);
8929 : }
8930 :
8931 : out_unlock:
8932 0 : double_rq_unlock(rq, p_rq);
8933 : out_irq:
8934 0 : local_irq_restore(flags);
8935 :
8936 0 : if (yielded > 0)
8937 0 : schedule();
8938 :
8939 0 : return yielded;
8940 : }
8941 : EXPORT_SYMBOL_GPL(yield_to);
8942 :
8943 0 : int io_schedule_prepare(void)
8944 : {
8945 0 : int old_iowait = current->in_iowait;
8946 :
8947 0 : current->in_iowait = 1;
8948 0 : blk_flush_plug(current->plug, true);
8949 0 : return old_iowait;
8950 : }
8951 :
8952 0 : void io_schedule_finish(int token)
8953 : {
8954 0 : current->in_iowait = token;
8955 0 : }
8956 :
8957 : /*
8958 : * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8959 : * that process accounting knows that this is a task in IO wait state.
8960 : */
8961 0 : long __sched io_schedule_timeout(long timeout)
8962 : {
8963 : int token;
8964 : long ret;
8965 :
8966 0 : token = io_schedule_prepare();
8967 0 : ret = schedule_timeout(timeout);
8968 0 : io_schedule_finish(token);
8969 :
8970 0 : return ret;
8971 : }
8972 : EXPORT_SYMBOL(io_schedule_timeout);
8973 :
8974 0 : void __sched io_schedule(void)
8975 : {
8976 : int token;
8977 :
8978 0 : token = io_schedule_prepare();
8979 0 : schedule();
8980 0 : io_schedule_finish(token);
8981 0 : }
8982 : EXPORT_SYMBOL(io_schedule);
8983 :
8984 : /**
8985 : * sys_sched_get_priority_max - return maximum RT priority.
8986 : * @policy: scheduling class.
8987 : *
8988 : * Return: On success, this syscall returns the maximum
8989 : * rt_priority that can be used by a given scheduling class.
8990 : * On failure, a negative error code is returned.
8991 : */
8992 0 : SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8993 : {
8994 0 : int ret = -EINVAL;
8995 :
8996 : switch (policy) {
8997 : case SCHED_FIFO:
8998 : case SCHED_RR:
8999 0 : ret = MAX_RT_PRIO-1;
9000 : break;
9001 : case SCHED_DEADLINE:
9002 : case SCHED_NORMAL:
9003 : case SCHED_BATCH:
9004 : case SCHED_IDLE:
9005 : ret = 0;
9006 : break;
9007 : }
9008 0 : return ret;
9009 : }
9010 :
9011 : /**
9012 : * sys_sched_get_priority_min - return minimum RT priority.
9013 : * @policy: scheduling class.
9014 : *
9015 : * Return: On success, this syscall returns the minimum
9016 : * rt_priority that can be used by a given scheduling class.
9017 : * On failure, a negative error code is returned.
9018 : */
9019 0 : SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
9020 : {
9021 0 : int ret = -EINVAL;
9022 :
9023 : switch (policy) {
9024 : case SCHED_FIFO:
9025 : case SCHED_RR:
9026 0 : ret = 1;
9027 : break;
9028 : case SCHED_DEADLINE:
9029 : case SCHED_NORMAL:
9030 : case SCHED_BATCH:
9031 : case SCHED_IDLE:
9032 : ret = 0;
9033 : }
9034 0 : return ret;
9035 : }
9036 :
9037 0 : static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
9038 : {
9039 : struct task_struct *p;
9040 : unsigned int time_slice;
9041 : struct rq_flags rf;
9042 : struct rq *rq;
9043 : int retval;
9044 :
9045 0 : if (pid < 0)
9046 : return -EINVAL;
9047 :
9048 0 : retval = -ESRCH;
9049 : rcu_read_lock();
9050 0 : p = find_process_by_pid(pid);
9051 0 : if (!p)
9052 : goto out_unlock;
9053 :
9054 0 : retval = security_task_getscheduler(p);
9055 : if (retval)
9056 : goto out_unlock;
9057 :
9058 0 : rq = task_rq_lock(p, &rf);
9059 0 : time_slice = 0;
9060 0 : if (p->sched_class->get_rr_interval)
9061 0 : time_slice = p->sched_class->get_rr_interval(rq, p);
9062 0 : task_rq_unlock(rq, p, &rf);
9063 :
9064 : rcu_read_unlock();
9065 0 : jiffies_to_timespec64(time_slice, t);
9066 0 : return 0;
9067 :
9068 : out_unlock:
9069 : rcu_read_unlock();
9070 0 : return retval;
9071 : }
9072 :
9073 : /**
9074 : * sys_sched_rr_get_interval - return the default timeslice of a process.
9075 : * @pid: pid of the process.
9076 : * @interval: userspace pointer to the timeslice value.
9077 : *
9078 : * this syscall writes the default timeslice value of a given process
9079 : * into the user-space timespec buffer. A value of '0' means infinity.
9080 : *
9081 : * Return: On success, 0 and the timeslice is in @interval. Otherwise,
9082 : * an error code.
9083 : */
9084 0 : SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
9085 : struct __kernel_timespec __user *, interval)
9086 : {
9087 : struct timespec64 t;
9088 0 : int retval = sched_rr_get_interval(pid, &t);
9089 :
9090 0 : if (retval == 0)
9091 0 : retval = put_timespec64(&t, interval);
9092 :
9093 0 : return retval;
9094 : }
9095 :
9096 : #ifdef CONFIG_COMPAT_32BIT_TIME
9097 : SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
9098 : struct old_timespec32 __user *, interval)
9099 : {
9100 : struct timespec64 t;
9101 : int retval = sched_rr_get_interval(pid, &t);
9102 :
9103 : if (retval == 0)
9104 : retval = put_old_timespec32(&t, interval);
9105 : return retval;
9106 : }
9107 : #endif
9108 :
9109 0 : void sched_show_task(struct task_struct *p)
9110 : {
9111 0 : unsigned long free = 0;
9112 : int ppid;
9113 :
9114 0 : if (!try_get_task_stack(p))
9115 : return;
9116 :
9117 0 : pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
9118 :
9119 0 : if (task_is_running(p))
9120 0 : pr_cont(" running task ");
9121 : #ifdef CONFIG_DEBUG_STACK_USAGE
9122 : free = stack_not_used(p);
9123 : #endif
9124 0 : ppid = 0;
9125 : rcu_read_lock();
9126 0 : if (pid_alive(p))
9127 0 : ppid = task_pid_nr(rcu_dereference(p->real_parent));
9128 : rcu_read_unlock();
9129 0 : pr_cont(" stack:%-5lu pid:%-5d ppid:%-6d flags:0x%08lx\n",
9130 : free, task_pid_nr(p), ppid,
9131 : read_task_thread_flags(p));
9132 :
9133 0 : print_worker_info(KERN_INFO, p);
9134 0 : print_stop_info(KERN_INFO, p);
9135 0 : show_stack(p, NULL, KERN_INFO);
9136 0 : put_task_stack(p);
9137 : }
9138 : EXPORT_SYMBOL_GPL(sched_show_task);
9139 :
9140 : static inline bool
9141 : state_filter_match(unsigned long state_filter, struct task_struct *p)
9142 : {
9143 0 : unsigned int state = READ_ONCE(p->__state);
9144 :
9145 : /* no filter, everything matches */
9146 0 : if (!state_filter)
9147 : return true;
9148 :
9149 : /* filter, but doesn't match */
9150 0 : if (!(state & state_filter))
9151 : return false;
9152 :
9153 : /*
9154 : * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9155 : * TASK_KILLABLE).
9156 : */
9157 0 : if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
9158 : return false;
9159 :
9160 : return true;
9161 : }
9162 :
9163 :
9164 0 : void show_state_filter(unsigned int state_filter)
9165 : {
9166 : struct task_struct *g, *p;
9167 :
9168 : rcu_read_lock();
9169 0 : for_each_process_thread(g, p) {
9170 : /*
9171 : * reset the NMI-timeout, listing all files on a slow
9172 : * console might take a lot of time:
9173 : * Also, reset softlockup watchdogs on all CPUs, because
9174 : * another CPU might be blocked waiting for us to process
9175 : * an IPI.
9176 : */
9177 : touch_nmi_watchdog();
9178 : touch_all_softlockup_watchdogs();
9179 0 : if (state_filter_match(state_filter, p))
9180 0 : sched_show_task(p);
9181 : }
9182 :
9183 : #ifdef CONFIG_SCHED_DEBUG
9184 : if (!state_filter)
9185 : sysrq_sched_debug_show();
9186 : #endif
9187 : rcu_read_unlock();
9188 : /*
9189 : * Only show locks if all tasks are dumped:
9190 : */
9191 : if (!state_filter)
9192 : debug_show_all_locks();
9193 0 : }
9194 :
9195 : /**
9196 : * init_idle - set up an idle thread for a given CPU
9197 : * @idle: task in question
9198 : * @cpu: CPU the idle task belongs to
9199 : *
9200 : * NOTE: this function does not set the idle thread's NEED_RESCHED
9201 : * flag, to make booting more robust.
9202 : */
9203 1 : void __init init_idle(struct task_struct *idle, int cpu)
9204 : {
9205 : #ifdef CONFIG_SMP
9206 : struct affinity_context ac = (struct affinity_context) {
9207 : .new_mask = cpumask_of(cpu),
9208 : .flags = 0,
9209 : };
9210 : #endif
9211 1 : struct rq *rq = cpu_rq(cpu);
9212 : unsigned long flags;
9213 :
9214 1 : __sched_fork(0, idle);
9215 :
9216 1 : raw_spin_lock_irqsave(&idle->pi_lock, flags);
9217 1 : raw_spin_rq_lock(rq);
9218 :
9219 1 : idle->__state = TASK_RUNNING;
9220 1 : idle->se.exec_start = sched_clock();
9221 : /*
9222 : * PF_KTHREAD should already be set at this point; regardless, make it
9223 : * look like a proper per-CPU kthread.
9224 : */
9225 1 : idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
9226 1 : kthread_set_per_cpu(idle, cpu);
9227 :
9228 : #ifdef CONFIG_SMP
9229 : /*
9230 : * It's possible that init_idle() gets called multiple times on a task,
9231 : * in that case do_set_cpus_allowed() will not do the right thing.
9232 : *
9233 : * And since this is boot we can forgo the serialization.
9234 : */
9235 : set_cpus_allowed_common(idle, &ac);
9236 : #endif
9237 : /*
9238 : * We're having a chicken and egg problem, even though we are
9239 : * holding rq->lock, the CPU isn't yet set to this CPU so the
9240 : * lockdep check in task_group() will fail.
9241 : *
9242 : * Similar case to sched_fork(). / Alternatively we could
9243 : * use task_rq_lock() here and obtain the other rq->lock.
9244 : *
9245 : * Silence PROVE_RCU
9246 : */
9247 : rcu_read_lock();
9248 1 : __set_task_cpu(idle, cpu);
9249 : rcu_read_unlock();
9250 :
9251 1 : rq->idle = idle;
9252 1 : rcu_assign_pointer(rq->curr, idle);
9253 1 : idle->on_rq = TASK_ON_RQ_QUEUED;
9254 : #ifdef CONFIG_SMP
9255 : idle->on_cpu = 1;
9256 : #endif
9257 1 : raw_spin_rq_unlock(rq);
9258 2 : raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
9259 :
9260 : /* Set the preempt count _outside_ the spinlocks! */
9261 1 : init_idle_preempt_count(idle, cpu);
9262 :
9263 : /*
9264 : * The idle tasks have their own, simple scheduling class:
9265 : */
9266 1 : idle->sched_class = &idle_sched_class;
9267 1 : ftrace_graph_init_idle_task(idle, cpu);
9268 1 : vtime_init_idle(idle, cpu);
9269 : #ifdef CONFIG_SMP
9270 : sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9271 : #endif
9272 1 : }
9273 :
9274 : #ifdef CONFIG_SMP
9275 :
9276 : int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9277 : const struct cpumask *trial)
9278 : {
9279 : int ret = 1;
9280 :
9281 : if (cpumask_empty(cur))
9282 : return ret;
9283 :
9284 : ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9285 :
9286 : return ret;
9287 : }
9288 :
9289 : int task_can_attach(struct task_struct *p,
9290 : const struct cpumask *cs_effective_cpus)
9291 : {
9292 : int ret = 0;
9293 :
9294 : /*
9295 : * Kthreads which disallow setaffinity shouldn't be moved
9296 : * to a new cpuset; we don't want to change their CPU
9297 : * affinity and isolating such threads by their set of
9298 : * allowed nodes is unnecessary. Thus, cpusets are not
9299 : * applicable for such threads. This prevents checking for
9300 : * success of set_cpus_allowed_ptr() on all attached tasks
9301 : * before cpus_mask may be changed.
9302 : */
9303 : if (p->flags & PF_NO_SETAFFINITY) {
9304 : ret = -EINVAL;
9305 : goto out;
9306 : }
9307 :
9308 : if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
9309 : cs_effective_cpus)) {
9310 : int cpu = cpumask_any_and(cpu_active_mask, cs_effective_cpus);
9311 :
9312 : if (unlikely(cpu >= nr_cpu_ids))
9313 : return -EINVAL;
9314 : ret = dl_cpu_busy(cpu, p);
9315 : }
9316 :
9317 : out:
9318 : return ret;
9319 : }
9320 :
9321 : bool sched_smp_initialized __read_mostly;
9322 :
9323 : #ifdef CONFIG_NUMA_BALANCING
9324 : /* Migrate current task p to target_cpu */
9325 : int migrate_task_to(struct task_struct *p, int target_cpu)
9326 : {
9327 : struct migration_arg arg = { p, target_cpu };
9328 : int curr_cpu = task_cpu(p);
9329 :
9330 : if (curr_cpu == target_cpu)
9331 : return 0;
9332 :
9333 : if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9334 : return -EINVAL;
9335 :
9336 : /* TODO: This is not properly updating schedstats */
9337 :
9338 : trace_sched_move_numa(p, curr_cpu, target_cpu);
9339 : return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9340 : }
9341 :
9342 : /*
9343 : * Requeue a task on a given node and accurately track the number of NUMA
9344 : * tasks on the runqueues
9345 : */
9346 : void sched_setnuma(struct task_struct *p, int nid)
9347 : {
9348 : bool queued, running;
9349 : struct rq_flags rf;
9350 : struct rq *rq;
9351 :
9352 : rq = task_rq_lock(p, &rf);
9353 : queued = task_on_rq_queued(p);
9354 : running = task_current(rq, p);
9355 :
9356 : if (queued)
9357 : dequeue_task(rq, p, DEQUEUE_SAVE);
9358 : if (running)
9359 : put_prev_task(rq, p);
9360 :
9361 : p->numa_preferred_nid = nid;
9362 :
9363 : if (queued)
9364 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9365 : if (running)
9366 : set_next_task(rq, p);
9367 : task_rq_unlock(rq, p, &rf);
9368 : }
9369 : #endif /* CONFIG_NUMA_BALANCING */
9370 :
9371 : #ifdef CONFIG_HOTPLUG_CPU
9372 : /*
9373 : * Ensure that the idle task is using init_mm right before its CPU goes
9374 : * offline.
9375 : */
9376 : void idle_task_exit(void)
9377 : {
9378 : struct mm_struct *mm = current->active_mm;
9379 :
9380 : BUG_ON(cpu_online(smp_processor_id()));
9381 : BUG_ON(current != this_rq()->idle);
9382 :
9383 : if (mm != &init_mm) {
9384 : switch_mm(mm, &init_mm, current);
9385 : finish_arch_post_lock_switch();
9386 : }
9387 :
9388 : /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9389 : }
9390 :
9391 : static int __balance_push_cpu_stop(void *arg)
9392 : {
9393 : struct task_struct *p = arg;
9394 : struct rq *rq = this_rq();
9395 : struct rq_flags rf;
9396 : int cpu;
9397 :
9398 : raw_spin_lock_irq(&p->pi_lock);
9399 : rq_lock(rq, &rf);
9400 :
9401 : update_rq_clock(rq);
9402 :
9403 : if (task_rq(p) == rq && task_on_rq_queued(p)) {
9404 : cpu = select_fallback_rq(rq->cpu, p);
9405 : rq = __migrate_task(rq, &rf, p, cpu);
9406 : }
9407 :
9408 : rq_unlock(rq, &rf);
9409 : raw_spin_unlock_irq(&p->pi_lock);
9410 :
9411 : put_task_struct(p);
9412 :
9413 : return 0;
9414 : }
9415 :
9416 : static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9417 :
9418 : /*
9419 : * Ensure we only run per-cpu kthreads once the CPU goes !active.
9420 : *
9421 : * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9422 : * effective when the hotplug motion is down.
9423 : */
9424 : static void balance_push(struct rq *rq)
9425 : {
9426 : struct task_struct *push_task = rq->curr;
9427 :
9428 : lockdep_assert_rq_held(rq);
9429 :
9430 : /*
9431 : * Ensure the thing is persistent until balance_push_set(.on = false);
9432 : */
9433 : rq->balance_callback = &balance_push_callback;
9434 :
9435 : /*
9436 : * Only active while going offline and when invoked on the outgoing
9437 : * CPU.
9438 : */
9439 : if (!cpu_dying(rq->cpu) || rq != this_rq())
9440 : return;
9441 :
9442 : /*
9443 : * Both the cpu-hotplug and stop task are in this case and are
9444 : * required to complete the hotplug process.
9445 : */
9446 : if (kthread_is_per_cpu(push_task) ||
9447 : is_migration_disabled(push_task)) {
9448 :
9449 : /*
9450 : * If this is the idle task on the outgoing CPU try to wake
9451 : * up the hotplug control thread which might wait for the
9452 : * last task to vanish. The rcuwait_active() check is
9453 : * accurate here because the waiter is pinned on this CPU
9454 : * and can't obviously be running in parallel.
9455 : *
9456 : * On RT kernels this also has to check whether there are
9457 : * pinned and scheduled out tasks on the runqueue. They
9458 : * need to leave the migrate disabled section first.
9459 : */
9460 : if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9461 : rcuwait_active(&rq->hotplug_wait)) {
9462 : raw_spin_rq_unlock(rq);
9463 : rcuwait_wake_up(&rq->hotplug_wait);
9464 : raw_spin_rq_lock(rq);
9465 : }
9466 : return;
9467 : }
9468 :
9469 : get_task_struct(push_task);
9470 : /*
9471 : * Temporarily drop rq->lock such that we can wake-up the stop task.
9472 : * Both preemption and IRQs are still disabled.
9473 : */
9474 : raw_spin_rq_unlock(rq);
9475 : stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9476 : this_cpu_ptr(&push_work));
9477 : /*
9478 : * At this point need_resched() is true and we'll take the loop in
9479 : * schedule(). The next pick is obviously going to be the stop task
9480 : * which kthread_is_per_cpu() and will push this task away.
9481 : */
9482 : raw_spin_rq_lock(rq);
9483 : }
9484 :
9485 : static void balance_push_set(int cpu, bool on)
9486 : {
9487 : struct rq *rq = cpu_rq(cpu);
9488 : struct rq_flags rf;
9489 :
9490 : rq_lock_irqsave(rq, &rf);
9491 : if (on) {
9492 : WARN_ON_ONCE(rq->balance_callback);
9493 : rq->balance_callback = &balance_push_callback;
9494 : } else if (rq->balance_callback == &balance_push_callback) {
9495 : rq->balance_callback = NULL;
9496 : }
9497 : rq_unlock_irqrestore(rq, &rf);
9498 : }
9499 :
9500 : /*
9501 : * Invoked from a CPUs hotplug control thread after the CPU has been marked
9502 : * inactive. All tasks which are not per CPU kernel threads are either
9503 : * pushed off this CPU now via balance_push() or placed on a different CPU
9504 : * during wakeup. Wait until the CPU is quiescent.
9505 : */
9506 : static void balance_hotplug_wait(void)
9507 : {
9508 : struct rq *rq = this_rq();
9509 :
9510 : rcuwait_wait_event(&rq->hotplug_wait,
9511 : rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9512 : TASK_UNINTERRUPTIBLE);
9513 : }
9514 :
9515 : #else
9516 :
9517 : static inline void balance_push(struct rq *rq)
9518 : {
9519 : }
9520 :
9521 : static inline void balance_push_set(int cpu, bool on)
9522 : {
9523 : }
9524 :
9525 : static inline void balance_hotplug_wait(void)
9526 : {
9527 : }
9528 :
9529 : #endif /* CONFIG_HOTPLUG_CPU */
9530 :
9531 : void set_rq_online(struct rq *rq)
9532 : {
9533 : if (!rq->online) {
9534 : const struct sched_class *class;
9535 :
9536 : cpumask_set_cpu(rq->cpu, rq->rd->online);
9537 : rq->online = 1;
9538 :
9539 : for_each_class(class) {
9540 : if (class->rq_online)
9541 : class->rq_online(rq);
9542 : }
9543 : }
9544 : }
9545 :
9546 : void set_rq_offline(struct rq *rq)
9547 : {
9548 : if (rq->online) {
9549 : const struct sched_class *class;
9550 :
9551 : for_each_class(class) {
9552 : if (class->rq_offline)
9553 : class->rq_offline(rq);
9554 : }
9555 :
9556 : cpumask_clear_cpu(rq->cpu, rq->rd->online);
9557 : rq->online = 0;
9558 : }
9559 : }
9560 :
9561 : /*
9562 : * used to mark begin/end of suspend/resume:
9563 : */
9564 : static int num_cpus_frozen;
9565 :
9566 : /*
9567 : * Update cpusets according to cpu_active mask. If cpusets are
9568 : * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9569 : * around partition_sched_domains().
9570 : *
9571 : * If we come here as part of a suspend/resume, don't touch cpusets because we
9572 : * want to restore it back to its original state upon resume anyway.
9573 : */
9574 : static void cpuset_cpu_active(void)
9575 : {
9576 : if (cpuhp_tasks_frozen) {
9577 : /*
9578 : * num_cpus_frozen tracks how many CPUs are involved in suspend
9579 : * resume sequence. As long as this is not the last online
9580 : * operation in the resume sequence, just build a single sched
9581 : * domain, ignoring cpusets.
9582 : */
9583 : partition_sched_domains(1, NULL, NULL);
9584 : if (--num_cpus_frozen)
9585 : return;
9586 : /*
9587 : * This is the last CPU online operation. So fall through and
9588 : * restore the original sched domains by considering the
9589 : * cpuset configurations.
9590 : */
9591 : cpuset_force_rebuild();
9592 : }
9593 : cpuset_update_active_cpus();
9594 : }
9595 :
9596 : static int cpuset_cpu_inactive(unsigned int cpu)
9597 : {
9598 : if (!cpuhp_tasks_frozen) {
9599 : int ret = dl_cpu_busy(cpu, NULL);
9600 :
9601 : if (ret)
9602 : return ret;
9603 : cpuset_update_active_cpus();
9604 : } else {
9605 : num_cpus_frozen++;
9606 : partition_sched_domains(1, NULL, NULL);
9607 : }
9608 : return 0;
9609 : }
9610 :
9611 : int sched_cpu_activate(unsigned int cpu)
9612 : {
9613 : struct rq *rq = cpu_rq(cpu);
9614 : struct rq_flags rf;
9615 :
9616 : /*
9617 : * Clear the balance_push callback and prepare to schedule
9618 : * regular tasks.
9619 : */
9620 : balance_push_set(cpu, false);
9621 :
9622 : #ifdef CONFIG_SCHED_SMT
9623 : /*
9624 : * When going up, increment the number of cores with SMT present.
9625 : */
9626 : if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9627 : static_branch_inc_cpuslocked(&sched_smt_present);
9628 : #endif
9629 : set_cpu_active(cpu, true);
9630 :
9631 : if (sched_smp_initialized) {
9632 : sched_update_numa(cpu, true);
9633 : sched_domains_numa_masks_set(cpu);
9634 : cpuset_cpu_active();
9635 : }
9636 :
9637 : /*
9638 : * Put the rq online, if not already. This happens:
9639 : *
9640 : * 1) In the early boot process, because we build the real domains
9641 : * after all CPUs have been brought up.
9642 : *
9643 : * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9644 : * domains.
9645 : */
9646 : rq_lock_irqsave(rq, &rf);
9647 : if (rq->rd) {
9648 : BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9649 : set_rq_online(rq);
9650 : }
9651 : rq_unlock_irqrestore(rq, &rf);
9652 :
9653 : return 0;
9654 : }
9655 :
9656 : int sched_cpu_deactivate(unsigned int cpu)
9657 : {
9658 : struct rq *rq = cpu_rq(cpu);
9659 : struct rq_flags rf;
9660 : int ret;
9661 :
9662 : /*
9663 : * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9664 : * load balancing when not active
9665 : */
9666 : nohz_balance_exit_idle(rq);
9667 :
9668 : set_cpu_active(cpu, false);
9669 :
9670 : /*
9671 : * From this point forward, this CPU will refuse to run any task that
9672 : * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9673 : * push those tasks away until this gets cleared, see
9674 : * sched_cpu_dying().
9675 : */
9676 : balance_push_set(cpu, true);
9677 :
9678 : /*
9679 : * We've cleared cpu_active_mask / set balance_push, wait for all
9680 : * preempt-disabled and RCU users of this state to go away such that
9681 : * all new such users will observe it.
9682 : *
9683 : * Specifically, we rely on ttwu to no longer target this CPU, see
9684 : * ttwu_queue_cond() and is_cpu_allowed().
9685 : *
9686 : * Do sync before park smpboot threads to take care the rcu boost case.
9687 : */
9688 : synchronize_rcu();
9689 :
9690 : rq_lock_irqsave(rq, &rf);
9691 : if (rq->rd) {
9692 : update_rq_clock(rq);
9693 : BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9694 : set_rq_offline(rq);
9695 : }
9696 : rq_unlock_irqrestore(rq, &rf);
9697 :
9698 : #ifdef CONFIG_SCHED_SMT
9699 : /*
9700 : * When going down, decrement the number of cores with SMT present.
9701 : */
9702 : if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9703 : static_branch_dec_cpuslocked(&sched_smt_present);
9704 :
9705 : sched_core_cpu_deactivate(cpu);
9706 : #endif
9707 :
9708 : if (!sched_smp_initialized)
9709 : return 0;
9710 :
9711 : sched_update_numa(cpu, false);
9712 : ret = cpuset_cpu_inactive(cpu);
9713 : if (ret) {
9714 : balance_push_set(cpu, false);
9715 : set_cpu_active(cpu, true);
9716 : sched_update_numa(cpu, true);
9717 : return ret;
9718 : }
9719 : sched_domains_numa_masks_clear(cpu);
9720 : return 0;
9721 : }
9722 :
9723 : static void sched_rq_cpu_starting(unsigned int cpu)
9724 : {
9725 : struct rq *rq = cpu_rq(cpu);
9726 :
9727 : rq->calc_load_update = calc_load_update;
9728 : update_max_interval();
9729 : }
9730 :
9731 : int sched_cpu_starting(unsigned int cpu)
9732 : {
9733 : sched_core_cpu_starting(cpu);
9734 : sched_rq_cpu_starting(cpu);
9735 : sched_tick_start(cpu);
9736 : return 0;
9737 : }
9738 :
9739 : #ifdef CONFIG_HOTPLUG_CPU
9740 :
9741 : /*
9742 : * Invoked immediately before the stopper thread is invoked to bring the
9743 : * CPU down completely. At this point all per CPU kthreads except the
9744 : * hotplug thread (current) and the stopper thread (inactive) have been
9745 : * either parked or have been unbound from the outgoing CPU. Ensure that
9746 : * any of those which might be on the way out are gone.
9747 : *
9748 : * If after this point a bound task is being woken on this CPU then the
9749 : * responsible hotplug callback has failed to do it's job.
9750 : * sched_cpu_dying() will catch it with the appropriate fireworks.
9751 : */
9752 : int sched_cpu_wait_empty(unsigned int cpu)
9753 : {
9754 : balance_hotplug_wait();
9755 : return 0;
9756 : }
9757 :
9758 : /*
9759 : * Since this CPU is going 'away' for a while, fold any nr_active delta we
9760 : * might have. Called from the CPU stopper task after ensuring that the
9761 : * stopper is the last running task on the CPU, so nr_active count is
9762 : * stable. We need to take the teardown thread which is calling this into
9763 : * account, so we hand in adjust = 1 to the load calculation.
9764 : *
9765 : * Also see the comment "Global load-average calculations".
9766 : */
9767 : static void calc_load_migrate(struct rq *rq)
9768 : {
9769 : long delta = calc_load_fold_active(rq, 1);
9770 :
9771 : if (delta)
9772 : atomic_long_add(delta, &calc_load_tasks);
9773 : }
9774 :
9775 : static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9776 : {
9777 : struct task_struct *g, *p;
9778 : int cpu = cpu_of(rq);
9779 :
9780 : lockdep_assert_rq_held(rq);
9781 :
9782 : printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9783 : for_each_process_thread(g, p) {
9784 : if (task_cpu(p) != cpu)
9785 : continue;
9786 :
9787 : if (!task_on_rq_queued(p))
9788 : continue;
9789 :
9790 : printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9791 : }
9792 : }
9793 :
9794 : int sched_cpu_dying(unsigned int cpu)
9795 : {
9796 : struct rq *rq = cpu_rq(cpu);
9797 : struct rq_flags rf;
9798 :
9799 : /* Handle pending wakeups and then migrate everything off */
9800 : sched_tick_stop(cpu);
9801 :
9802 : rq_lock_irqsave(rq, &rf);
9803 : if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9804 : WARN(true, "Dying CPU not properly vacated!");
9805 : dump_rq_tasks(rq, KERN_WARNING);
9806 : }
9807 : rq_unlock_irqrestore(rq, &rf);
9808 :
9809 : calc_load_migrate(rq);
9810 : update_max_interval();
9811 : hrtick_clear(rq);
9812 : sched_core_cpu_dying(cpu);
9813 : return 0;
9814 : }
9815 : #endif
9816 :
9817 : void __init sched_init_smp(void)
9818 : {
9819 : sched_init_numa(NUMA_NO_NODE);
9820 :
9821 : /*
9822 : * There's no userspace yet to cause hotplug operations; hence all the
9823 : * CPU masks are stable and all blatant races in the below code cannot
9824 : * happen.
9825 : */
9826 : mutex_lock(&sched_domains_mutex);
9827 : sched_init_domains(cpu_active_mask);
9828 : mutex_unlock(&sched_domains_mutex);
9829 :
9830 : /* Move init over to a non-isolated CPU */
9831 : if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9832 : BUG();
9833 : current->flags &= ~PF_NO_SETAFFINITY;
9834 : sched_init_granularity();
9835 :
9836 : init_sched_rt_class();
9837 : init_sched_dl_class();
9838 :
9839 : sched_smp_initialized = true;
9840 : }
9841 :
9842 : static int __init migration_init(void)
9843 : {
9844 : sched_cpu_starting(smp_processor_id());
9845 : return 0;
9846 : }
9847 : early_initcall(migration_init);
9848 :
9849 : #else
9850 1 : void __init sched_init_smp(void)
9851 : {
9852 1 : sched_init_granularity();
9853 1 : }
9854 : #endif /* CONFIG_SMP */
9855 :
9856 0 : int in_sched_functions(unsigned long addr)
9857 : {
9858 0 : return in_lock_functions(addr) ||
9859 0 : (addr >= (unsigned long)__sched_text_start
9860 0 : && addr < (unsigned long)__sched_text_end);
9861 : }
9862 :
9863 : #ifdef CONFIG_CGROUP_SCHED
9864 : /*
9865 : * Default task group.
9866 : * Every task in system belongs to this group at bootup.
9867 : */
9868 : struct task_group root_task_group;
9869 : LIST_HEAD(task_groups);
9870 :
9871 : /* Cacheline aligned slab cache for task_group */
9872 : static struct kmem_cache *task_group_cache __read_mostly;
9873 : #endif
9874 :
9875 1 : void __init sched_init(void)
9876 : {
9877 1 : unsigned long ptr = 0;
9878 : int i;
9879 :
9880 : /* Make sure the linker didn't screw up */
9881 1 : BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9882 : &fair_sched_class != &rt_sched_class + 1 ||
9883 : &rt_sched_class != &dl_sched_class + 1);
9884 : #ifdef CONFIG_SMP
9885 : BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9886 : #endif
9887 :
9888 1 : wait_bit_init();
9889 :
9890 : #ifdef CONFIG_FAIR_GROUP_SCHED
9891 : ptr += 2 * nr_cpu_ids * sizeof(void **);
9892 : #endif
9893 : #ifdef CONFIG_RT_GROUP_SCHED
9894 : ptr += 2 * nr_cpu_ids * sizeof(void **);
9895 : #endif
9896 : if (ptr) {
9897 : ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9898 :
9899 : #ifdef CONFIG_FAIR_GROUP_SCHED
9900 : root_task_group.se = (struct sched_entity **)ptr;
9901 : ptr += nr_cpu_ids * sizeof(void **);
9902 :
9903 : root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9904 : ptr += nr_cpu_ids * sizeof(void **);
9905 :
9906 : root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9907 : init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9908 : #endif /* CONFIG_FAIR_GROUP_SCHED */
9909 : #ifdef CONFIG_RT_GROUP_SCHED
9910 : root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9911 : ptr += nr_cpu_ids * sizeof(void **);
9912 :
9913 : root_task_group.rt_rq = (struct rt_rq **)ptr;
9914 : ptr += nr_cpu_ids * sizeof(void **);
9915 :
9916 : #endif /* CONFIG_RT_GROUP_SCHED */
9917 : }
9918 :
9919 1 : init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9920 :
9921 : #ifdef CONFIG_SMP
9922 : init_defrootdomain();
9923 : #endif
9924 :
9925 : #ifdef CONFIG_RT_GROUP_SCHED
9926 : init_rt_bandwidth(&root_task_group.rt_bandwidth,
9927 : global_rt_period(), global_rt_runtime());
9928 : #endif /* CONFIG_RT_GROUP_SCHED */
9929 :
9930 : #ifdef CONFIG_CGROUP_SCHED
9931 : task_group_cache = KMEM_CACHE(task_group, 0);
9932 :
9933 : list_add(&root_task_group.list, &task_groups);
9934 : INIT_LIST_HEAD(&root_task_group.children);
9935 : INIT_LIST_HEAD(&root_task_group.siblings);
9936 : autogroup_init(&init_task);
9937 : #endif /* CONFIG_CGROUP_SCHED */
9938 :
9939 2 : for_each_possible_cpu(i) {
9940 : struct rq *rq;
9941 :
9942 1 : rq = cpu_rq(i);
9943 : raw_spin_lock_init(&rq->__lock);
9944 1 : rq->nr_running = 0;
9945 1 : rq->calc_load_active = 0;
9946 1 : rq->calc_load_update = jiffies + LOAD_FREQ;
9947 1 : init_cfs_rq(&rq->cfs);
9948 1 : init_rt_rq(&rq->rt);
9949 1 : init_dl_rq(&rq->dl);
9950 : #ifdef CONFIG_FAIR_GROUP_SCHED
9951 : INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9952 : rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9953 : /*
9954 : * How much CPU bandwidth does root_task_group get?
9955 : *
9956 : * In case of task-groups formed thr' the cgroup filesystem, it
9957 : * gets 100% of the CPU resources in the system. This overall
9958 : * system CPU resource is divided among the tasks of
9959 : * root_task_group and its child task-groups in a fair manner,
9960 : * based on each entity's (task or task-group's) weight
9961 : * (se->load.weight).
9962 : *
9963 : * In other words, if root_task_group has 10 tasks of weight
9964 : * 1024) and two child groups A0 and A1 (of weight 1024 each),
9965 : * then A0's share of the CPU resource is:
9966 : *
9967 : * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9968 : *
9969 : * We achieve this by letting root_task_group's tasks sit
9970 : * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9971 : */
9972 : init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9973 : #endif /* CONFIG_FAIR_GROUP_SCHED */
9974 :
9975 1 : rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9976 : #ifdef CONFIG_RT_GROUP_SCHED
9977 : init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9978 : #endif
9979 : #ifdef CONFIG_SMP
9980 : rq->sd = NULL;
9981 : rq->rd = NULL;
9982 : rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9983 : rq->balance_callback = &balance_push_callback;
9984 : rq->active_balance = 0;
9985 : rq->next_balance = jiffies;
9986 : rq->push_cpu = 0;
9987 : rq->cpu = i;
9988 : rq->online = 0;
9989 : rq->idle_stamp = 0;
9990 : rq->avg_idle = 2*sysctl_sched_migration_cost;
9991 : rq->wake_stamp = jiffies;
9992 : rq->wake_avg_idle = rq->avg_idle;
9993 : rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9994 :
9995 : INIT_LIST_HEAD(&rq->cfs_tasks);
9996 :
9997 : rq_attach_root(rq, &def_root_domain);
9998 : #ifdef CONFIG_NO_HZ_COMMON
9999 : rq->last_blocked_load_update_tick = jiffies;
10000 : atomic_set(&rq->nohz_flags, 0);
10001 :
10002 : INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
10003 : #endif
10004 : #ifdef CONFIG_HOTPLUG_CPU
10005 : rcuwait_init(&rq->hotplug_wait);
10006 : #endif
10007 : #endif /* CONFIG_SMP */
10008 1 : hrtick_rq_init(rq);
10009 2 : atomic_set(&rq->nr_iowait, 0);
10010 :
10011 : #ifdef CONFIG_SCHED_CORE
10012 : rq->core = rq;
10013 : rq->core_pick = NULL;
10014 : rq->core_enabled = 0;
10015 : rq->core_tree = RB_ROOT;
10016 : rq->core_forceidle_count = 0;
10017 : rq->core_forceidle_occupation = 0;
10018 : rq->core_forceidle_start = 0;
10019 :
10020 : rq->core_cookie = 0UL;
10021 : #endif
10022 2 : zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
10023 : }
10024 :
10025 1 : set_load_weight(&init_task, false);
10026 :
10027 : /*
10028 : * The boot idle thread does lazy MMU switching as well:
10029 : */
10030 1 : mmgrab_lazy_tlb(&init_mm);
10031 1 : enter_lazy_tlb(&init_mm, current);
10032 :
10033 : /*
10034 : * The idle task doesn't need the kthread struct to function, but it
10035 : * is dressed up as a per-CPU kthread and thus needs to play the part
10036 : * if we want to avoid special-casing it in code that deals with per-CPU
10037 : * kthreads.
10038 : */
10039 1 : WARN_ON(!set_kthread_struct(current));
10040 :
10041 : /*
10042 : * Make us the idle thread. Technically, schedule() should not be
10043 : * called from this thread, however somewhere below it might be,
10044 : * but because we are the idle thread, we just pick up running again
10045 : * when this runqueue becomes "idle".
10046 : */
10047 1 : init_idle(current, smp_processor_id());
10048 :
10049 1 : calc_load_update = jiffies + LOAD_FREQ;
10050 :
10051 : #ifdef CONFIG_SMP
10052 : idle_thread_set_boot_cpu();
10053 : balance_push_set(smp_processor_id(), false);
10054 : #endif
10055 1 : init_sched_fair_class();
10056 :
10057 : psi_init();
10058 :
10059 : init_uclamp();
10060 :
10061 : preempt_dynamic_init();
10062 :
10063 1 : scheduler_running = 1;
10064 1 : }
10065 :
10066 : #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
10067 :
10068 : void __might_sleep(const char *file, int line)
10069 : {
10070 : unsigned int state = get_current_state();
10071 : /*
10072 : * Blocking primitives will set (and therefore destroy) current->state,
10073 : * since we will exit with TASK_RUNNING make sure we enter with it,
10074 : * otherwise we will destroy state.
10075 : */
10076 : WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
10077 : "do not call blocking ops when !TASK_RUNNING; "
10078 : "state=%x set at [<%p>] %pS\n", state,
10079 : (void *)current->task_state_change,
10080 : (void *)current->task_state_change);
10081 :
10082 : __might_resched(file, line, 0);
10083 : }
10084 : EXPORT_SYMBOL(__might_sleep);
10085 :
10086 : static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
10087 : {
10088 : if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
10089 : return;
10090 :
10091 : if (preempt_count() == preempt_offset)
10092 : return;
10093 :
10094 : pr_err("Preemption disabled at:");
10095 : print_ip_sym(KERN_ERR, ip);
10096 : }
10097 :
10098 : static inline bool resched_offsets_ok(unsigned int offsets)
10099 : {
10100 : unsigned int nested = preempt_count();
10101 :
10102 : nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
10103 :
10104 : return nested == offsets;
10105 : }
10106 :
10107 : void __might_resched(const char *file, int line, unsigned int offsets)
10108 : {
10109 : /* Ratelimiting timestamp: */
10110 : static unsigned long prev_jiffy;
10111 :
10112 : unsigned long preempt_disable_ip;
10113 :
10114 : /* WARN_ON_ONCE() by default, no rate limit required: */
10115 : rcu_sleep_check();
10116 :
10117 : if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
10118 : !is_idle_task(current) && !current->non_block_count) ||
10119 : system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10120 : oops_in_progress)
10121 : return;
10122 :
10123 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10124 : return;
10125 : prev_jiffy = jiffies;
10126 :
10127 : /* Save this before calling printk(), since that will clobber it: */
10128 : preempt_disable_ip = get_preempt_disable_ip(current);
10129 :
10130 : pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10131 : file, line);
10132 : pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10133 : in_atomic(), irqs_disabled(), current->non_block_count,
10134 : current->pid, current->comm);
10135 : pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10136 : offsets & MIGHT_RESCHED_PREEMPT_MASK);
10137 :
10138 : if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
10139 : pr_err("RCU nest depth: %d, expected: %u\n",
10140 : rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
10141 : }
10142 :
10143 : if (task_stack_end_corrupted(current))
10144 : pr_emerg("Thread overran stack, or stack corrupted\n");
10145 :
10146 : debug_show_held_locks(current);
10147 : if (irqs_disabled())
10148 : print_irqtrace_events(current);
10149 :
10150 : print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
10151 : preempt_disable_ip);
10152 :
10153 : dump_stack();
10154 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10155 : }
10156 : EXPORT_SYMBOL(__might_resched);
10157 :
10158 : void __cant_sleep(const char *file, int line, int preempt_offset)
10159 : {
10160 : static unsigned long prev_jiffy;
10161 :
10162 : if (irqs_disabled())
10163 : return;
10164 :
10165 : if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10166 : return;
10167 :
10168 : if (preempt_count() > preempt_offset)
10169 : return;
10170 :
10171 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10172 : return;
10173 : prev_jiffy = jiffies;
10174 :
10175 : printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10176 : printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10177 : in_atomic(), irqs_disabled(),
10178 : current->pid, current->comm);
10179 :
10180 : debug_show_held_locks(current);
10181 : dump_stack();
10182 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10183 : }
10184 : EXPORT_SYMBOL_GPL(__cant_sleep);
10185 :
10186 : #ifdef CONFIG_SMP
10187 : void __cant_migrate(const char *file, int line)
10188 : {
10189 : static unsigned long prev_jiffy;
10190 :
10191 : if (irqs_disabled())
10192 : return;
10193 :
10194 : if (is_migration_disabled(current))
10195 : return;
10196 :
10197 : if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10198 : return;
10199 :
10200 : if (preempt_count() > 0)
10201 : return;
10202 :
10203 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10204 : return;
10205 : prev_jiffy = jiffies;
10206 :
10207 : pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10208 : pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10209 : in_atomic(), irqs_disabled(), is_migration_disabled(current),
10210 : current->pid, current->comm);
10211 :
10212 : debug_show_held_locks(current);
10213 : dump_stack();
10214 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10215 : }
10216 : EXPORT_SYMBOL_GPL(__cant_migrate);
10217 : #endif
10218 : #endif
10219 :
10220 : #ifdef CONFIG_MAGIC_SYSRQ
10221 : void normalize_rt_tasks(void)
10222 : {
10223 : struct task_struct *g, *p;
10224 : struct sched_attr attr = {
10225 : .sched_policy = SCHED_NORMAL,
10226 : };
10227 :
10228 : read_lock(&tasklist_lock);
10229 : for_each_process_thread(g, p) {
10230 : /*
10231 : * Only normalize user tasks:
10232 : */
10233 : if (p->flags & PF_KTHREAD)
10234 : continue;
10235 :
10236 : p->se.exec_start = 0;
10237 : schedstat_set(p->stats.wait_start, 0);
10238 : schedstat_set(p->stats.sleep_start, 0);
10239 : schedstat_set(p->stats.block_start, 0);
10240 :
10241 : if (!dl_task(p) && !rt_task(p)) {
10242 : /*
10243 : * Renice negative nice level userspace
10244 : * tasks back to 0:
10245 : */
10246 : if (task_nice(p) < 0)
10247 : set_user_nice(p, 0);
10248 : continue;
10249 : }
10250 :
10251 : __sched_setscheduler(p, &attr, false, false);
10252 : }
10253 : read_unlock(&tasklist_lock);
10254 : }
10255 :
10256 : #endif /* CONFIG_MAGIC_SYSRQ */
10257 :
10258 : #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
10259 : /*
10260 : * These functions are only useful for the IA64 MCA handling, or kdb.
10261 : *
10262 : * They can only be called when the whole system has been
10263 : * stopped - every CPU needs to be quiescent, and no scheduling
10264 : * activity can take place. Using them for anything else would
10265 : * be a serious bug, and as a result, they aren't even visible
10266 : * under any other configuration.
10267 : */
10268 :
10269 : /**
10270 : * curr_task - return the current task for a given CPU.
10271 : * @cpu: the processor in question.
10272 : *
10273 : * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10274 : *
10275 : * Return: The current task for @cpu.
10276 : */
10277 : struct task_struct *curr_task(int cpu)
10278 : {
10279 : return cpu_curr(cpu);
10280 : }
10281 :
10282 : #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10283 :
10284 : #ifdef CONFIG_IA64
10285 : /**
10286 : * ia64_set_curr_task - set the current task for a given CPU.
10287 : * @cpu: the processor in question.
10288 : * @p: the task pointer to set.
10289 : *
10290 : * Description: This function must only be used when non-maskable interrupts
10291 : * are serviced on a separate stack. It allows the architecture to switch the
10292 : * notion of the current task on a CPU in a non-blocking manner. This function
10293 : * must be called with all CPU's synchronized, and interrupts disabled, the
10294 : * and caller must save the original value of the current task (see
10295 : * curr_task() above) and restore that value before reenabling interrupts and
10296 : * re-starting the system.
10297 : *
10298 : * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10299 : */
10300 : void ia64_set_curr_task(int cpu, struct task_struct *p)
10301 : {
10302 : cpu_curr(cpu) = p;
10303 : }
10304 :
10305 : #endif
10306 :
10307 : #ifdef CONFIG_CGROUP_SCHED
10308 : /* task_group_lock serializes the addition/removal of task groups */
10309 : static DEFINE_SPINLOCK(task_group_lock);
10310 :
10311 : static inline void alloc_uclamp_sched_group(struct task_group *tg,
10312 : struct task_group *parent)
10313 : {
10314 : #ifdef CONFIG_UCLAMP_TASK_GROUP
10315 : enum uclamp_id clamp_id;
10316 :
10317 : for_each_clamp_id(clamp_id) {
10318 : uclamp_se_set(&tg->uclamp_req[clamp_id],
10319 : uclamp_none(clamp_id), false);
10320 : tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10321 : }
10322 : #endif
10323 : }
10324 :
10325 : static void sched_free_group(struct task_group *tg)
10326 : {
10327 : free_fair_sched_group(tg);
10328 : free_rt_sched_group(tg);
10329 : autogroup_free(tg);
10330 : kmem_cache_free(task_group_cache, tg);
10331 : }
10332 :
10333 : static void sched_free_group_rcu(struct rcu_head *rcu)
10334 : {
10335 : sched_free_group(container_of(rcu, struct task_group, rcu));
10336 : }
10337 :
10338 : static void sched_unregister_group(struct task_group *tg)
10339 : {
10340 : unregister_fair_sched_group(tg);
10341 : unregister_rt_sched_group(tg);
10342 : /*
10343 : * We have to wait for yet another RCU grace period to expire, as
10344 : * print_cfs_stats() might run concurrently.
10345 : */
10346 : call_rcu(&tg->rcu, sched_free_group_rcu);
10347 : }
10348 :
10349 : /* allocate runqueue etc for a new task group */
10350 : struct task_group *sched_create_group(struct task_group *parent)
10351 : {
10352 : struct task_group *tg;
10353 :
10354 : tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10355 : if (!tg)
10356 : return ERR_PTR(-ENOMEM);
10357 :
10358 : if (!alloc_fair_sched_group(tg, parent))
10359 : goto err;
10360 :
10361 : if (!alloc_rt_sched_group(tg, parent))
10362 : goto err;
10363 :
10364 : alloc_uclamp_sched_group(tg, parent);
10365 :
10366 : return tg;
10367 :
10368 : err:
10369 : sched_free_group(tg);
10370 : return ERR_PTR(-ENOMEM);
10371 : }
10372 :
10373 : void sched_online_group(struct task_group *tg, struct task_group *parent)
10374 : {
10375 : unsigned long flags;
10376 :
10377 : spin_lock_irqsave(&task_group_lock, flags);
10378 : list_add_rcu(&tg->list, &task_groups);
10379 :
10380 : /* Root should already exist: */
10381 : WARN_ON(!parent);
10382 :
10383 : tg->parent = parent;
10384 : INIT_LIST_HEAD(&tg->children);
10385 : list_add_rcu(&tg->siblings, &parent->children);
10386 : spin_unlock_irqrestore(&task_group_lock, flags);
10387 :
10388 : online_fair_sched_group(tg);
10389 : }
10390 :
10391 : /* rcu callback to free various structures associated with a task group */
10392 : static void sched_unregister_group_rcu(struct rcu_head *rhp)
10393 : {
10394 : /* Now it should be safe to free those cfs_rqs: */
10395 : sched_unregister_group(container_of(rhp, struct task_group, rcu));
10396 : }
10397 :
10398 : void sched_destroy_group(struct task_group *tg)
10399 : {
10400 : /* Wait for possible concurrent references to cfs_rqs complete: */
10401 : call_rcu(&tg->rcu, sched_unregister_group_rcu);
10402 : }
10403 :
10404 : void sched_release_group(struct task_group *tg)
10405 : {
10406 : unsigned long flags;
10407 :
10408 : /*
10409 : * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10410 : * sched_cfs_period_timer()).
10411 : *
10412 : * For this to be effective, we have to wait for all pending users of
10413 : * this task group to leave their RCU critical section to ensure no new
10414 : * user will see our dying task group any more. Specifically ensure
10415 : * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10416 : *
10417 : * We therefore defer calling unregister_fair_sched_group() to
10418 : * sched_unregister_group() which is guarantied to get called only after the
10419 : * current RCU grace period has expired.
10420 : */
10421 : spin_lock_irqsave(&task_group_lock, flags);
10422 : list_del_rcu(&tg->list);
10423 : list_del_rcu(&tg->siblings);
10424 : spin_unlock_irqrestore(&task_group_lock, flags);
10425 : }
10426 :
10427 : static struct task_group *sched_get_task_group(struct task_struct *tsk)
10428 : {
10429 : struct task_group *tg;
10430 :
10431 : /*
10432 : * All callers are synchronized by task_rq_lock(); we do not use RCU
10433 : * which is pointless here. Thus, we pass "true" to task_css_check()
10434 : * to prevent lockdep warnings.
10435 : */
10436 : tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10437 : struct task_group, css);
10438 : tg = autogroup_task_group(tsk, tg);
10439 :
10440 : return tg;
10441 : }
10442 :
10443 : static void sched_change_group(struct task_struct *tsk, struct task_group *group)
10444 : {
10445 : tsk->sched_task_group = group;
10446 :
10447 : #ifdef CONFIG_FAIR_GROUP_SCHED
10448 : if (tsk->sched_class->task_change_group)
10449 : tsk->sched_class->task_change_group(tsk);
10450 : else
10451 : #endif
10452 : set_task_rq(tsk, task_cpu(tsk));
10453 : }
10454 :
10455 : /*
10456 : * Change task's runqueue when it moves between groups.
10457 : *
10458 : * The caller of this function should have put the task in its new group by
10459 : * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10460 : * its new group.
10461 : */
10462 : void sched_move_task(struct task_struct *tsk)
10463 : {
10464 : int queued, running, queue_flags =
10465 : DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10466 : struct task_group *group;
10467 : struct rq_flags rf;
10468 : struct rq *rq;
10469 :
10470 : rq = task_rq_lock(tsk, &rf);
10471 : /*
10472 : * Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
10473 : * group changes.
10474 : */
10475 : group = sched_get_task_group(tsk);
10476 : if (group == tsk->sched_task_group)
10477 : goto unlock;
10478 :
10479 : update_rq_clock(rq);
10480 :
10481 : running = task_current(rq, tsk);
10482 : queued = task_on_rq_queued(tsk);
10483 :
10484 : if (queued)
10485 : dequeue_task(rq, tsk, queue_flags);
10486 : if (running)
10487 : put_prev_task(rq, tsk);
10488 :
10489 : sched_change_group(tsk, group);
10490 :
10491 : if (queued)
10492 : enqueue_task(rq, tsk, queue_flags);
10493 : if (running) {
10494 : set_next_task(rq, tsk);
10495 : /*
10496 : * After changing group, the running task may have joined a
10497 : * throttled one but it's still the running task. Trigger a
10498 : * resched to make sure that task can still run.
10499 : */
10500 : resched_curr(rq);
10501 : }
10502 :
10503 : unlock:
10504 : task_rq_unlock(rq, tsk, &rf);
10505 : }
10506 :
10507 : static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10508 : {
10509 : return css ? container_of(css, struct task_group, css) : NULL;
10510 : }
10511 :
10512 : static struct cgroup_subsys_state *
10513 : cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10514 : {
10515 : struct task_group *parent = css_tg(parent_css);
10516 : struct task_group *tg;
10517 :
10518 : if (!parent) {
10519 : /* This is early initialization for the top cgroup */
10520 : return &root_task_group.css;
10521 : }
10522 :
10523 : tg = sched_create_group(parent);
10524 : if (IS_ERR(tg))
10525 : return ERR_PTR(-ENOMEM);
10526 :
10527 : return &tg->css;
10528 : }
10529 :
10530 : /* Expose task group only after completing cgroup initialization */
10531 : static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10532 : {
10533 : struct task_group *tg = css_tg(css);
10534 : struct task_group *parent = css_tg(css->parent);
10535 :
10536 : if (parent)
10537 : sched_online_group(tg, parent);
10538 :
10539 : #ifdef CONFIG_UCLAMP_TASK_GROUP
10540 : /* Propagate the effective uclamp value for the new group */
10541 : mutex_lock(&uclamp_mutex);
10542 : rcu_read_lock();
10543 : cpu_util_update_eff(css);
10544 : rcu_read_unlock();
10545 : mutex_unlock(&uclamp_mutex);
10546 : #endif
10547 :
10548 : return 0;
10549 : }
10550 :
10551 : static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10552 : {
10553 : struct task_group *tg = css_tg(css);
10554 :
10555 : sched_release_group(tg);
10556 : }
10557 :
10558 : static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10559 : {
10560 : struct task_group *tg = css_tg(css);
10561 :
10562 : /*
10563 : * Relies on the RCU grace period between css_released() and this.
10564 : */
10565 : sched_unregister_group(tg);
10566 : }
10567 :
10568 : #ifdef CONFIG_RT_GROUP_SCHED
10569 : static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10570 : {
10571 : struct task_struct *task;
10572 : struct cgroup_subsys_state *css;
10573 :
10574 : cgroup_taskset_for_each(task, css, tset) {
10575 : if (!sched_rt_can_attach(css_tg(css), task))
10576 : return -EINVAL;
10577 : }
10578 : return 0;
10579 : }
10580 : #endif
10581 :
10582 : static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10583 : {
10584 : struct task_struct *task;
10585 : struct cgroup_subsys_state *css;
10586 :
10587 : cgroup_taskset_for_each(task, css, tset)
10588 : sched_move_task(task);
10589 : }
10590 :
10591 : #ifdef CONFIG_UCLAMP_TASK_GROUP
10592 : static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10593 : {
10594 : struct cgroup_subsys_state *top_css = css;
10595 : struct uclamp_se *uc_parent = NULL;
10596 : struct uclamp_se *uc_se = NULL;
10597 : unsigned int eff[UCLAMP_CNT];
10598 : enum uclamp_id clamp_id;
10599 : unsigned int clamps;
10600 :
10601 : lockdep_assert_held(&uclamp_mutex);
10602 : SCHED_WARN_ON(!rcu_read_lock_held());
10603 :
10604 : css_for_each_descendant_pre(css, top_css) {
10605 : uc_parent = css_tg(css)->parent
10606 : ? css_tg(css)->parent->uclamp : NULL;
10607 :
10608 : for_each_clamp_id(clamp_id) {
10609 : /* Assume effective clamps matches requested clamps */
10610 : eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10611 : /* Cap effective clamps with parent's effective clamps */
10612 : if (uc_parent &&
10613 : eff[clamp_id] > uc_parent[clamp_id].value) {
10614 : eff[clamp_id] = uc_parent[clamp_id].value;
10615 : }
10616 : }
10617 : /* Ensure protection is always capped by limit */
10618 : eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10619 :
10620 : /* Propagate most restrictive effective clamps */
10621 : clamps = 0x0;
10622 : uc_se = css_tg(css)->uclamp;
10623 : for_each_clamp_id(clamp_id) {
10624 : if (eff[clamp_id] == uc_se[clamp_id].value)
10625 : continue;
10626 : uc_se[clamp_id].value = eff[clamp_id];
10627 : uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10628 : clamps |= (0x1 << clamp_id);
10629 : }
10630 : if (!clamps) {
10631 : css = css_rightmost_descendant(css);
10632 : continue;
10633 : }
10634 :
10635 : /* Immediately update descendants RUNNABLE tasks */
10636 : uclamp_update_active_tasks(css);
10637 : }
10638 : }
10639 :
10640 : /*
10641 : * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10642 : * C expression. Since there is no way to convert a macro argument (N) into a
10643 : * character constant, use two levels of macros.
10644 : */
10645 : #define _POW10(exp) ((unsigned int)1e##exp)
10646 : #define POW10(exp) _POW10(exp)
10647 :
10648 : struct uclamp_request {
10649 : #define UCLAMP_PERCENT_SHIFT 2
10650 : #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10651 : s64 percent;
10652 : u64 util;
10653 : int ret;
10654 : };
10655 :
10656 : static inline struct uclamp_request
10657 : capacity_from_percent(char *buf)
10658 : {
10659 : struct uclamp_request req = {
10660 : .percent = UCLAMP_PERCENT_SCALE,
10661 : .util = SCHED_CAPACITY_SCALE,
10662 : .ret = 0,
10663 : };
10664 :
10665 : buf = strim(buf);
10666 : if (strcmp(buf, "max")) {
10667 : req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10668 : &req.percent);
10669 : if (req.ret)
10670 : return req;
10671 : if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10672 : req.ret = -ERANGE;
10673 : return req;
10674 : }
10675 :
10676 : req.util = req.percent << SCHED_CAPACITY_SHIFT;
10677 : req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10678 : }
10679 :
10680 : return req;
10681 : }
10682 :
10683 : static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10684 : size_t nbytes, loff_t off,
10685 : enum uclamp_id clamp_id)
10686 : {
10687 : struct uclamp_request req;
10688 : struct task_group *tg;
10689 :
10690 : req = capacity_from_percent(buf);
10691 : if (req.ret)
10692 : return req.ret;
10693 :
10694 : static_branch_enable(&sched_uclamp_used);
10695 :
10696 : mutex_lock(&uclamp_mutex);
10697 : rcu_read_lock();
10698 :
10699 : tg = css_tg(of_css(of));
10700 : if (tg->uclamp_req[clamp_id].value != req.util)
10701 : uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10702 :
10703 : /*
10704 : * Because of not recoverable conversion rounding we keep track of the
10705 : * exact requested value
10706 : */
10707 : tg->uclamp_pct[clamp_id] = req.percent;
10708 :
10709 : /* Update effective clamps to track the most restrictive value */
10710 : cpu_util_update_eff(of_css(of));
10711 :
10712 : rcu_read_unlock();
10713 : mutex_unlock(&uclamp_mutex);
10714 :
10715 : return nbytes;
10716 : }
10717 :
10718 : static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10719 : char *buf, size_t nbytes,
10720 : loff_t off)
10721 : {
10722 : return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10723 : }
10724 :
10725 : static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10726 : char *buf, size_t nbytes,
10727 : loff_t off)
10728 : {
10729 : return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10730 : }
10731 :
10732 : static inline void cpu_uclamp_print(struct seq_file *sf,
10733 : enum uclamp_id clamp_id)
10734 : {
10735 : struct task_group *tg;
10736 : u64 util_clamp;
10737 : u64 percent;
10738 : u32 rem;
10739 :
10740 : rcu_read_lock();
10741 : tg = css_tg(seq_css(sf));
10742 : util_clamp = tg->uclamp_req[clamp_id].value;
10743 : rcu_read_unlock();
10744 :
10745 : if (util_clamp == SCHED_CAPACITY_SCALE) {
10746 : seq_puts(sf, "max\n");
10747 : return;
10748 : }
10749 :
10750 : percent = tg->uclamp_pct[clamp_id];
10751 : percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10752 : seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10753 : }
10754 :
10755 : static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10756 : {
10757 : cpu_uclamp_print(sf, UCLAMP_MIN);
10758 : return 0;
10759 : }
10760 :
10761 : static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10762 : {
10763 : cpu_uclamp_print(sf, UCLAMP_MAX);
10764 : return 0;
10765 : }
10766 : #endif /* CONFIG_UCLAMP_TASK_GROUP */
10767 :
10768 : #ifdef CONFIG_FAIR_GROUP_SCHED
10769 : static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10770 : struct cftype *cftype, u64 shareval)
10771 : {
10772 : if (shareval > scale_load_down(ULONG_MAX))
10773 : shareval = MAX_SHARES;
10774 : return sched_group_set_shares(css_tg(css), scale_load(shareval));
10775 : }
10776 :
10777 : static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10778 : struct cftype *cft)
10779 : {
10780 : struct task_group *tg = css_tg(css);
10781 :
10782 : return (u64) scale_load_down(tg->shares);
10783 : }
10784 :
10785 : #ifdef CONFIG_CFS_BANDWIDTH
10786 : static DEFINE_MUTEX(cfs_constraints_mutex);
10787 :
10788 : const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10789 : static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10790 : /* More than 203 days if BW_SHIFT equals 20. */
10791 : static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10792 :
10793 : static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10794 :
10795 : static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10796 : u64 burst)
10797 : {
10798 : int i, ret = 0, runtime_enabled, runtime_was_enabled;
10799 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10800 :
10801 : if (tg == &root_task_group)
10802 : return -EINVAL;
10803 :
10804 : /*
10805 : * Ensure we have at some amount of bandwidth every period. This is
10806 : * to prevent reaching a state of large arrears when throttled via
10807 : * entity_tick() resulting in prolonged exit starvation.
10808 : */
10809 : if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10810 : return -EINVAL;
10811 :
10812 : /*
10813 : * Likewise, bound things on the other side by preventing insane quota
10814 : * periods. This also allows us to normalize in computing quota
10815 : * feasibility.
10816 : */
10817 : if (period > max_cfs_quota_period)
10818 : return -EINVAL;
10819 :
10820 : /*
10821 : * Bound quota to defend quota against overflow during bandwidth shift.
10822 : */
10823 : if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10824 : return -EINVAL;
10825 :
10826 : if (quota != RUNTIME_INF && (burst > quota ||
10827 : burst + quota > max_cfs_runtime))
10828 : return -EINVAL;
10829 :
10830 : /*
10831 : * Prevent race between setting of cfs_rq->runtime_enabled and
10832 : * unthrottle_offline_cfs_rqs().
10833 : */
10834 : cpus_read_lock();
10835 : mutex_lock(&cfs_constraints_mutex);
10836 : ret = __cfs_schedulable(tg, period, quota);
10837 : if (ret)
10838 : goto out_unlock;
10839 :
10840 : runtime_enabled = quota != RUNTIME_INF;
10841 : runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10842 : /*
10843 : * If we need to toggle cfs_bandwidth_used, off->on must occur
10844 : * before making related changes, and on->off must occur afterwards
10845 : */
10846 : if (runtime_enabled && !runtime_was_enabled)
10847 : cfs_bandwidth_usage_inc();
10848 : raw_spin_lock_irq(&cfs_b->lock);
10849 : cfs_b->period = ns_to_ktime(period);
10850 : cfs_b->quota = quota;
10851 : cfs_b->burst = burst;
10852 :
10853 : __refill_cfs_bandwidth_runtime(cfs_b);
10854 :
10855 : /* Restart the period timer (if active) to handle new period expiry: */
10856 : if (runtime_enabled)
10857 : start_cfs_bandwidth(cfs_b);
10858 :
10859 : raw_spin_unlock_irq(&cfs_b->lock);
10860 :
10861 : for_each_online_cpu(i) {
10862 : struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10863 : struct rq *rq = cfs_rq->rq;
10864 : struct rq_flags rf;
10865 :
10866 : rq_lock_irq(rq, &rf);
10867 : cfs_rq->runtime_enabled = runtime_enabled;
10868 : cfs_rq->runtime_remaining = 0;
10869 :
10870 : if (cfs_rq->throttled)
10871 : unthrottle_cfs_rq(cfs_rq);
10872 : rq_unlock_irq(rq, &rf);
10873 : }
10874 : if (runtime_was_enabled && !runtime_enabled)
10875 : cfs_bandwidth_usage_dec();
10876 : out_unlock:
10877 : mutex_unlock(&cfs_constraints_mutex);
10878 : cpus_read_unlock();
10879 :
10880 : return ret;
10881 : }
10882 :
10883 : static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10884 : {
10885 : u64 quota, period, burst;
10886 :
10887 : period = ktime_to_ns(tg->cfs_bandwidth.period);
10888 : burst = tg->cfs_bandwidth.burst;
10889 : if (cfs_quota_us < 0)
10890 : quota = RUNTIME_INF;
10891 : else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10892 : quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10893 : else
10894 : return -EINVAL;
10895 :
10896 : return tg_set_cfs_bandwidth(tg, period, quota, burst);
10897 : }
10898 :
10899 : static long tg_get_cfs_quota(struct task_group *tg)
10900 : {
10901 : u64 quota_us;
10902 :
10903 : if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10904 : return -1;
10905 :
10906 : quota_us = tg->cfs_bandwidth.quota;
10907 : do_div(quota_us, NSEC_PER_USEC);
10908 :
10909 : return quota_us;
10910 : }
10911 :
10912 : static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10913 : {
10914 : u64 quota, period, burst;
10915 :
10916 : if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10917 : return -EINVAL;
10918 :
10919 : period = (u64)cfs_period_us * NSEC_PER_USEC;
10920 : quota = tg->cfs_bandwidth.quota;
10921 : burst = tg->cfs_bandwidth.burst;
10922 :
10923 : return tg_set_cfs_bandwidth(tg, period, quota, burst);
10924 : }
10925 :
10926 : static long tg_get_cfs_period(struct task_group *tg)
10927 : {
10928 : u64 cfs_period_us;
10929 :
10930 : cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10931 : do_div(cfs_period_us, NSEC_PER_USEC);
10932 :
10933 : return cfs_period_us;
10934 : }
10935 :
10936 : static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10937 : {
10938 : u64 quota, period, burst;
10939 :
10940 : if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10941 : return -EINVAL;
10942 :
10943 : burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10944 : period = ktime_to_ns(tg->cfs_bandwidth.period);
10945 : quota = tg->cfs_bandwidth.quota;
10946 :
10947 : return tg_set_cfs_bandwidth(tg, period, quota, burst);
10948 : }
10949 :
10950 : static long tg_get_cfs_burst(struct task_group *tg)
10951 : {
10952 : u64 burst_us;
10953 :
10954 : burst_us = tg->cfs_bandwidth.burst;
10955 : do_div(burst_us, NSEC_PER_USEC);
10956 :
10957 : return burst_us;
10958 : }
10959 :
10960 : static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10961 : struct cftype *cft)
10962 : {
10963 : return tg_get_cfs_quota(css_tg(css));
10964 : }
10965 :
10966 : static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10967 : struct cftype *cftype, s64 cfs_quota_us)
10968 : {
10969 : return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10970 : }
10971 :
10972 : static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10973 : struct cftype *cft)
10974 : {
10975 : return tg_get_cfs_period(css_tg(css));
10976 : }
10977 :
10978 : static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10979 : struct cftype *cftype, u64 cfs_period_us)
10980 : {
10981 : return tg_set_cfs_period(css_tg(css), cfs_period_us);
10982 : }
10983 :
10984 : static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10985 : struct cftype *cft)
10986 : {
10987 : return tg_get_cfs_burst(css_tg(css));
10988 : }
10989 :
10990 : static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10991 : struct cftype *cftype, u64 cfs_burst_us)
10992 : {
10993 : return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10994 : }
10995 :
10996 : struct cfs_schedulable_data {
10997 : struct task_group *tg;
10998 : u64 period, quota;
10999 : };
11000 :
11001 : /*
11002 : * normalize group quota/period to be quota/max_period
11003 : * note: units are usecs
11004 : */
11005 : static u64 normalize_cfs_quota(struct task_group *tg,
11006 : struct cfs_schedulable_data *d)
11007 : {
11008 : u64 quota, period;
11009 :
11010 : if (tg == d->tg) {
11011 : period = d->period;
11012 : quota = d->quota;
11013 : } else {
11014 : period = tg_get_cfs_period(tg);
11015 : quota = tg_get_cfs_quota(tg);
11016 : }
11017 :
11018 : /* note: these should typically be equivalent */
11019 : if (quota == RUNTIME_INF || quota == -1)
11020 : return RUNTIME_INF;
11021 :
11022 : return to_ratio(period, quota);
11023 : }
11024 :
11025 : static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
11026 : {
11027 : struct cfs_schedulable_data *d = data;
11028 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11029 : s64 quota = 0, parent_quota = -1;
11030 :
11031 : if (!tg->parent) {
11032 : quota = RUNTIME_INF;
11033 : } else {
11034 : struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
11035 :
11036 : quota = normalize_cfs_quota(tg, d);
11037 : parent_quota = parent_b->hierarchical_quota;
11038 :
11039 : /*
11040 : * Ensure max(child_quota) <= parent_quota. On cgroup2,
11041 : * always take the min. On cgroup1, only inherit when no
11042 : * limit is set:
11043 : */
11044 : if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
11045 : quota = min(quota, parent_quota);
11046 : } else {
11047 : if (quota == RUNTIME_INF)
11048 : quota = parent_quota;
11049 : else if (parent_quota != RUNTIME_INF && quota > parent_quota)
11050 : return -EINVAL;
11051 : }
11052 : }
11053 : cfs_b->hierarchical_quota = quota;
11054 :
11055 : return 0;
11056 : }
11057 :
11058 : static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
11059 : {
11060 : int ret;
11061 : struct cfs_schedulable_data data = {
11062 : .tg = tg,
11063 : .period = period,
11064 : .quota = quota,
11065 : };
11066 :
11067 : if (quota != RUNTIME_INF) {
11068 : do_div(data.period, NSEC_PER_USEC);
11069 : do_div(data.quota, NSEC_PER_USEC);
11070 : }
11071 :
11072 : rcu_read_lock();
11073 : ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
11074 : rcu_read_unlock();
11075 :
11076 : return ret;
11077 : }
11078 :
11079 : static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
11080 : {
11081 : struct task_group *tg = css_tg(seq_css(sf));
11082 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11083 :
11084 : seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
11085 : seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
11086 : seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
11087 :
11088 : if (schedstat_enabled() && tg != &root_task_group) {
11089 : struct sched_statistics *stats;
11090 : u64 ws = 0;
11091 : int i;
11092 :
11093 : for_each_possible_cpu(i) {
11094 : stats = __schedstats_from_se(tg->se[i]);
11095 : ws += schedstat_val(stats->wait_sum);
11096 : }
11097 :
11098 : seq_printf(sf, "wait_sum %llu\n", ws);
11099 : }
11100 :
11101 : seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
11102 : seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
11103 :
11104 : return 0;
11105 : }
11106 : #endif /* CONFIG_CFS_BANDWIDTH */
11107 : #endif /* CONFIG_FAIR_GROUP_SCHED */
11108 :
11109 : #ifdef CONFIG_RT_GROUP_SCHED
11110 : static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
11111 : struct cftype *cft, s64 val)
11112 : {
11113 : return sched_group_set_rt_runtime(css_tg(css), val);
11114 : }
11115 :
11116 : static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
11117 : struct cftype *cft)
11118 : {
11119 : return sched_group_rt_runtime(css_tg(css));
11120 : }
11121 :
11122 : static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
11123 : struct cftype *cftype, u64 rt_period_us)
11124 : {
11125 : return sched_group_set_rt_period(css_tg(css), rt_period_us);
11126 : }
11127 :
11128 : static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11129 : struct cftype *cft)
11130 : {
11131 : return sched_group_rt_period(css_tg(css));
11132 : }
11133 : #endif /* CONFIG_RT_GROUP_SCHED */
11134 :
11135 : #ifdef CONFIG_FAIR_GROUP_SCHED
11136 : static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11137 : struct cftype *cft)
11138 : {
11139 : return css_tg(css)->idle;
11140 : }
11141 :
11142 : static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11143 : struct cftype *cft, s64 idle)
11144 : {
11145 : return sched_group_set_idle(css_tg(css), idle);
11146 : }
11147 : #endif
11148 :
11149 : static struct cftype cpu_legacy_files[] = {
11150 : #ifdef CONFIG_FAIR_GROUP_SCHED
11151 : {
11152 : .name = "shares",
11153 : .read_u64 = cpu_shares_read_u64,
11154 : .write_u64 = cpu_shares_write_u64,
11155 : },
11156 : {
11157 : .name = "idle",
11158 : .read_s64 = cpu_idle_read_s64,
11159 : .write_s64 = cpu_idle_write_s64,
11160 : },
11161 : #endif
11162 : #ifdef CONFIG_CFS_BANDWIDTH
11163 : {
11164 : .name = "cfs_quota_us",
11165 : .read_s64 = cpu_cfs_quota_read_s64,
11166 : .write_s64 = cpu_cfs_quota_write_s64,
11167 : },
11168 : {
11169 : .name = "cfs_period_us",
11170 : .read_u64 = cpu_cfs_period_read_u64,
11171 : .write_u64 = cpu_cfs_period_write_u64,
11172 : },
11173 : {
11174 : .name = "cfs_burst_us",
11175 : .read_u64 = cpu_cfs_burst_read_u64,
11176 : .write_u64 = cpu_cfs_burst_write_u64,
11177 : },
11178 : {
11179 : .name = "stat",
11180 : .seq_show = cpu_cfs_stat_show,
11181 : },
11182 : #endif
11183 : #ifdef CONFIG_RT_GROUP_SCHED
11184 : {
11185 : .name = "rt_runtime_us",
11186 : .read_s64 = cpu_rt_runtime_read,
11187 : .write_s64 = cpu_rt_runtime_write,
11188 : },
11189 : {
11190 : .name = "rt_period_us",
11191 : .read_u64 = cpu_rt_period_read_uint,
11192 : .write_u64 = cpu_rt_period_write_uint,
11193 : },
11194 : #endif
11195 : #ifdef CONFIG_UCLAMP_TASK_GROUP
11196 : {
11197 : .name = "uclamp.min",
11198 : .flags = CFTYPE_NOT_ON_ROOT,
11199 : .seq_show = cpu_uclamp_min_show,
11200 : .write = cpu_uclamp_min_write,
11201 : },
11202 : {
11203 : .name = "uclamp.max",
11204 : .flags = CFTYPE_NOT_ON_ROOT,
11205 : .seq_show = cpu_uclamp_max_show,
11206 : .write = cpu_uclamp_max_write,
11207 : },
11208 : #endif
11209 : { } /* Terminate */
11210 : };
11211 :
11212 : static int cpu_extra_stat_show(struct seq_file *sf,
11213 : struct cgroup_subsys_state *css)
11214 : {
11215 : #ifdef CONFIG_CFS_BANDWIDTH
11216 : {
11217 : struct task_group *tg = css_tg(css);
11218 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11219 : u64 throttled_usec, burst_usec;
11220 :
11221 : throttled_usec = cfs_b->throttled_time;
11222 : do_div(throttled_usec, NSEC_PER_USEC);
11223 : burst_usec = cfs_b->burst_time;
11224 : do_div(burst_usec, NSEC_PER_USEC);
11225 :
11226 : seq_printf(sf, "nr_periods %d\n"
11227 : "nr_throttled %d\n"
11228 : "throttled_usec %llu\n"
11229 : "nr_bursts %d\n"
11230 : "burst_usec %llu\n",
11231 : cfs_b->nr_periods, cfs_b->nr_throttled,
11232 : throttled_usec, cfs_b->nr_burst, burst_usec);
11233 : }
11234 : #endif
11235 : return 0;
11236 : }
11237 :
11238 : #ifdef CONFIG_FAIR_GROUP_SCHED
11239 : static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11240 : struct cftype *cft)
11241 : {
11242 : struct task_group *tg = css_tg(css);
11243 : u64 weight = scale_load_down(tg->shares);
11244 :
11245 : return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11246 : }
11247 :
11248 : static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11249 : struct cftype *cft, u64 weight)
11250 : {
11251 : /*
11252 : * cgroup weight knobs should use the common MIN, DFL and MAX
11253 : * values which are 1, 100 and 10000 respectively. While it loses
11254 : * a bit of range on both ends, it maps pretty well onto the shares
11255 : * value used by scheduler and the round-trip conversions preserve
11256 : * the original value over the entire range.
11257 : */
11258 : if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11259 : return -ERANGE;
11260 :
11261 : weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11262 :
11263 : return sched_group_set_shares(css_tg(css), scale_load(weight));
11264 : }
11265 :
11266 : static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11267 : struct cftype *cft)
11268 : {
11269 : unsigned long weight = scale_load_down(css_tg(css)->shares);
11270 : int last_delta = INT_MAX;
11271 : int prio, delta;
11272 :
11273 : /* find the closest nice value to the current weight */
11274 : for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11275 : delta = abs(sched_prio_to_weight[prio] - weight);
11276 : if (delta >= last_delta)
11277 : break;
11278 : last_delta = delta;
11279 : }
11280 :
11281 : return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11282 : }
11283 :
11284 : static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11285 : struct cftype *cft, s64 nice)
11286 : {
11287 : unsigned long weight;
11288 : int idx;
11289 :
11290 : if (nice < MIN_NICE || nice > MAX_NICE)
11291 : return -ERANGE;
11292 :
11293 : idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11294 : idx = array_index_nospec(idx, 40);
11295 : weight = sched_prio_to_weight[idx];
11296 :
11297 : return sched_group_set_shares(css_tg(css), scale_load(weight));
11298 : }
11299 : #endif
11300 :
11301 : static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11302 : long period, long quota)
11303 : {
11304 : if (quota < 0)
11305 : seq_puts(sf, "max");
11306 : else
11307 : seq_printf(sf, "%ld", quota);
11308 :
11309 : seq_printf(sf, " %ld\n", period);
11310 : }
11311 :
11312 : /* caller should put the current value in *@periodp before calling */
11313 : static int __maybe_unused cpu_period_quota_parse(char *buf,
11314 : u64 *periodp, u64 *quotap)
11315 : {
11316 : char tok[21]; /* U64_MAX */
11317 :
11318 : if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11319 : return -EINVAL;
11320 :
11321 : *periodp *= NSEC_PER_USEC;
11322 :
11323 : if (sscanf(tok, "%llu", quotap))
11324 : *quotap *= NSEC_PER_USEC;
11325 : else if (!strcmp(tok, "max"))
11326 : *quotap = RUNTIME_INF;
11327 : else
11328 : return -EINVAL;
11329 :
11330 : return 0;
11331 : }
11332 :
11333 : #ifdef CONFIG_CFS_BANDWIDTH
11334 : static int cpu_max_show(struct seq_file *sf, void *v)
11335 : {
11336 : struct task_group *tg = css_tg(seq_css(sf));
11337 :
11338 : cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11339 : return 0;
11340 : }
11341 :
11342 : static ssize_t cpu_max_write(struct kernfs_open_file *of,
11343 : char *buf, size_t nbytes, loff_t off)
11344 : {
11345 : struct task_group *tg = css_tg(of_css(of));
11346 : u64 period = tg_get_cfs_period(tg);
11347 : u64 burst = tg_get_cfs_burst(tg);
11348 : u64 quota;
11349 : int ret;
11350 :
11351 : ret = cpu_period_quota_parse(buf, &period, "a);
11352 : if (!ret)
11353 : ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11354 : return ret ?: nbytes;
11355 : }
11356 : #endif
11357 :
11358 : static struct cftype cpu_files[] = {
11359 : #ifdef CONFIG_FAIR_GROUP_SCHED
11360 : {
11361 : .name = "weight",
11362 : .flags = CFTYPE_NOT_ON_ROOT,
11363 : .read_u64 = cpu_weight_read_u64,
11364 : .write_u64 = cpu_weight_write_u64,
11365 : },
11366 : {
11367 : .name = "weight.nice",
11368 : .flags = CFTYPE_NOT_ON_ROOT,
11369 : .read_s64 = cpu_weight_nice_read_s64,
11370 : .write_s64 = cpu_weight_nice_write_s64,
11371 : },
11372 : {
11373 : .name = "idle",
11374 : .flags = CFTYPE_NOT_ON_ROOT,
11375 : .read_s64 = cpu_idle_read_s64,
11376 : .write_s64 = cpu_idle_write_s64,
11377 : },
11378 : #endif
11379 : #ifdef CONFIG_CFS_BANDWIDTH
11380 : {
11381 : .name = "max",
11382 : .flags = CFTYPE_NOT_ON_ROOT,
11383 : .seq_show = cpu_max_show,
11384 : .write = cpu_max_write,
11385 : },
11386 : {
11387 : .name = "max.burst",
11388 : .flags = CFTYPE_NOT_ON_ROOT,
11389 : .read_u64 = cpu_cfs_burst_read_u64,
11390 : .write_u64 = cpu_cfs_burst_write_u64,
11391 : },
11392 : #endif
11393 : #ifdef CONFIG_UCLAMP_TASK_GROUP
11394 : {
11395 : .name = "uclamp.min",
11396 : .flags = CFTYPE_NOT_ON_ROOT,
11397 : .seq_show = cpu_uclamp_min_show,
11398 : .write = cpu_uclamp_min_write,
11399 : },
11400 : {
11401 : .name = "uclamp.max",
11402 : .flags = CFTYPE_NOT_ON_ROOT,
11403 : .seq_show = cpu_uclamp_max_show,
11404 : .write = cpu_uclamp_max_write,
11405 : },
11406 : #endif
11407 : { } /* terminate */
11408 : };
11409 :
11410 : struct cgroup_subsys cpu_cgrp_subsys = {
11411 : .css_alloc = cpu_cgroup_css_alloc,
11412 : .css_online = cpu_cgroup_css_online,
11413 : .css_released = cpu_cgroup_css_released,
11414 : .css_free = cpu_cgroup_css_free,
11415 : .css_extra_stat_show = cpu_extra_stat_show,
11416 : #ifdef CONFIG_RT_GROUP_SCHED
11417 : .can_attach = cpu_cgroup_can_attach,
11418 : #endif
11419 : .attach = cpu_cgroup_attach,
11420 : .legacy_cftypes = cpu_legacy_files,
11421 : .dfl_cftypes = cpu_files,
11422 : .early_init = true,
11423 : .threaded = true,
11424 : };
11425 :
11426 : #endif /* CONFIG_CGROUP_SCHED */
11427 :
11428 0 : void dump_cpu_task(int cpu)
11429 : {
11430 0 : if (cpu == smp_processor_id() && in_hardirq()) {
11431 : struct pt_regs *regs;
11432 :
11433 0 : regs = get_irq_regs();
11434 0 : if (regs) {
11435 0 : show_regs(regs);
11436 0 : return;
11437 : }
11438 : }
11439 :
11440 0 : if (trigger_single_cpu_backtrace(cpu))
11441 : return;
11442 :
11443 0 : pr_info("Task dump for CPU %d:\n", cpu);
11444 0 : sched_show_task(cpu_curr(cpu));
11445 : }
11446 :
11447 : /*
11448 : * Nice levels are multiplicative, with a gentle 10% change for every
11449 : * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11450 : * nice 1, it will get ~10% less CPU time than another CPU-bound task
11451 : * that remained on nice 0.
11452 : *
11453 : * The "10% effect" is relative and cumulative: from _any_ nice level,
11454 : * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11455 : * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11456 : * If a task goes up by ~10% and another task goes down by ~10% then
11457 : * the relative distance between them is ~25%.)
11458 : */
11459 : const int sched_prio_to_weight[40] = {
11460 : /* -20 */ 88761, 71755, 56483, 46273, 36291,
11461 : /* -15 */ 29154, 23254, 18705, 14949, 11916,
11462 : /* -10 */ 9548, 7620, 6100, 4904, 3906,
11463 : /* -5 */ 3121, 2501, 1991, 1586, 1277,
11464 : /* 0 */ 1024, 820, 655, 526, 423,
11465 : /* 5 */ 335, 272, 215, 172, 137,
11466 : /* 10 */ 110, 87, 70, 56, 45,
11467 : /* 15 */ 36, 29, 23, 18, 15,
11468 : };
11469 :
11470 : /*
11471 : * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11472 : *
11473 : * In cases where the weight does not change often, we can use the
11474 : * precalculated inverse to speed up arithmetics by turning divisions
11475 : * into multiplications:
11476 : */
11477 : const u32 sched_prio_to_wmult[40] = {
11478 : /* -20 */ 48388, 59856, 76040, 92818, 118348,
11479 : /* -15 */ 147320, 184698, 229616, 287308, 360437,
11480 : /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11481 : /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11482 : /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11483 : /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11484 : /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11485 : /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11486 : };
11487 :
11488 0 : void call_trace_sched_update_nr_running(struct rq *rq, int count)
11489 : {
11490 0 : trace_sched_update_nr_running_tp(rq, count);
11491 0 : }
11492 :
11493 : #ifdef CONFIG_SCHED_MM_CID
11494 :
11495 : /*
11496 : * @cid_lock: Guarantee forward-progress of cid allocation.
11497 : *
11498 : * Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
11499 : * is only used when contention is detected by the lock-free allocation so
11500 : * forward progress can be guaranteed.
11501 : */
11502 : DEFINE_RAW_SPINLOCK(cid_lock);
11503 :
11504 : /*
11505 : * @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
11506 : *
11507 : * When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
11508 : * detected, it is set to 1 to ensure that all newly coming allocations are
11509 : * serialized by @cid_lock until the allocation which detected contention
11510 : * completes and sets @use_cid_lock back to 0. This guarantees forward progress
11511 : * of a cid allocation.
11512 : */
11513 : int use_cid_lock;
11514 :
11515 : /*
11516 : * mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
11517 : * concurrently with respect to the execution of the source runqueue context
11518 : * switch.
11519 : *
11520 : * There is one basic properties we want to guarantee here:
11521 : *
11522 : * (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
11523 : * used by a task. That would lead to concurrent allocation of the cid and
11524 : * userspace corruption.
11525 : *
11526 : * Provide this guarantee by introducing a Dekker memory ordering to guarantee
11527 : * that a pair of loads observe at least one of a pair of stores, which can be
11528 : * shown as:
11529 : *
11530 : * X = Y = 0
11531 : *
11532 : * w[X]=1 w[Y]=1
11533 : * MB MB
11534 : * r[Y]=y r[X]=x
11535 : *
11536 : * Which guarantees that x==0 && y==0 is impossible. But rather than using
11537 : * values 0 and 1, this algorithm cares about specific state transitions of the
11538 : * runqueue current task (as updated by the scheduler context switch), and the
11539 : * per-mm/cpu cid value.
11540 : *
11541 : * Let's introduce task (Y) which has task->mm == mm and task (N) which has
11542 : * task->mm != mm for the rest of the discussion. There are two scheduler state
11543 : * transitions on context switch we care about:
11544 : *
11545 : * (TSA) Store to rq->curr with transition from (N) to (Y)
11546 : *
11547 : * (TSB) Store to rq->curr with transition from (Y) to (N)
11548 : *
11549 : * On the remote-clear side, there is one transition we care about:
11550 : *
11551 : * (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
11552 : *
11553 : * There is also a transition to UNSET state which can be performed from all
11554 : * sides (scheduler, remote-clear). It is always performed with a cmpxchg which
11555 : * guarantees that only a single thread will succeed:
11556 : *
11557 : * (TMB) cmpxchg to *pcpu_cid to mark UNSET
11558 : *
11559 : * Just to be clear, what we do _not_ want to happen is a transition to UNSET
11560 : * when a thread is actively using the cid (property (1)).
11561 : *
11562 : * Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
11563 : *
11564 : * Scenario A) (TSA)+(TMA) (from next task perspective)
11565 : *
11566 : * CPU0 CPU1
11567 : *
11568 : * Context switch CS-1 Remote-clear
11569 : * - store to rq->curr: (N)->(Y) (TSA) - cmpxchg to *pcpu_id to LAZY (TMA)
11570 : * (implied barrier after cmpxchg)
11571 : * - switch_mm_cid()
11572 : * - memory barrier (see switch_mm_cid()
11573 : * comment explaining how this barrier
11574 : * is combined with other scheduler
11575 : * barriers)
11576 : * - mm_cid_get (next)
11577 : * - READ_ONCE(*pcpu_cid) - rcu_dereference(src_rq->curr)
11578 : *
11579 : * This Dekker ensures that either task (Y) is observed by the
11580 : * rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
11581 : * observed.
11582 : *
11583 : * If task (Y) store is observed by rcu_dereference(), it means that there is
11584 : * still an active task on the cpu. Remote-clear will therefore not transition
11585 : * to UNSET, which fulfills property (1).
11586 : *
11587 : * If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
11588 : * it will move its state to UNSET, which clears the percpu cid perhaps
11589 : * uselessly (which is not an issue for correctness). Because task (Y) is not
11590 : * observed, CPU1 can move ahead to set the state to UNSET. Because moving
11591 : * state to UNSET is done with a cmpxchg expecting that the old state has the
11592 : * LAZY flag set, only one thread will successfully UNSET.
11593 : *
11594 : * If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
11595 : * will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
11596 : * CPU1 will observe task (Y) and do nothing more, which is fine.
11597 : *
11598 : * What we are effectively preventing with this Dekker is a scenario where
11599 : * neither LAZY flag nor store (Y) are observed, which would fail property (1)
11600 : * because this would UNSET a cid which is actively used.
11601 : */
11602 :
11603 : void sched_mm_cid_migrate_from(struct task_struct *t)
11604 : {
11605 : t->migrate_from_cpu = task_cpu(t);
11606 : }
11607 :
11608 : static
11609 : int __sched_mm_cid_migrate_from_fetch_cid(struct rq *src_rq,
11610 : struct task_struct *t,
11611 : struct mm_cid *src_pcpu_cid)
11612 : {
11613 : struct mm_struct *mm = t->mm;
11614 : struct task_struct *src_task;
11615 : int src_cid, last_mm_cid;
11616 :
11617 : if (!mm)
11618 : return -1;
11619 :
11620 : last_mm_cid = t->last_mm_cid;
11621 : /*
11622 : * If the migrated task has no last cid, or if the current
11623 : * task on src rq uses the cid, it means the source cid does not need
11624 : * to be moved to the destination cpu.
11625 : */
11626 : if (last_mm_cid == -1)
11627 : return -1;
11628 : src_cid = READ_ONCE(src_pcpu_cid->cid);
11629 : if (!mm_cid_is_valid(src_cid) || last_mm_cid != src_cid)
11630 : return -1;
11631 :
11632 : /*
11633 : * If we observe an active task using the mm on this rq, it means we
11634 : * are not the last task to be migrated from this cpu for this mm, so
11635 : * there is no need to move src_cid to the destination cpu.
11636 : */
11637 : rcu_read_lock();
11638 : src_task = rcu_dereference(src_rq->curr);
11639 : if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11640 : rcu_read_unlock();
11641 : t->last_mm_cid = -1;
11642 : return -1;
11643 : }
11644 : rcu_read_unlock();
11645 :
11646 : return src_cid;
11647 : }
11648 :
11649 : static
11650 : int __sched_mm_cid_migrate_from_try_steal_cid(struct rq *src_rq,
11651 : struct task_struct *t,
11652 : struct mm_cid *src_pcpu_cid,
11653 : int src_cid)
11654 : {
11655 : struct task_struct *src_task;
11656 : struct mm_struct *mm = t->mm;
11657 : int lazy_cid;
11658 :
11659 : if (src_cid == -1)
11660 : return -1;
11661 :
11662 : /*
11663 : * Attempt to clear the source cpu cid to move it to the destination
11664 : * cpu.
11665 : */
11666 : lazy_cid = mm_cid_set_lazy_put(src_cid);
11667 : if (!try_cmpxchg(&src_pcpu_cid->cid, &src_cid, lazy_cid))
11668 : return -1;
11669 :
11670 : /*
11671 : * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11672 : * rq->curr->mm matches the scheduler barrier in context_switch()
11673 : * between store to rq->curr and load of prev and next task's
11674 : * per-mm/cpu cid.
11675 : *
11676 : * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11677 : * rq->curr->mm_cid_active matches the barrier in
11678 : * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11679 : * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11680 : * load of per-mm/cpu cid.
11681 : */
11682 :
11683 : /*
11684 : * If we observe an active task using the mm on this rq after setting
11685 : * the lazy-put flag, this task will be responsible for transitioning
11686 : * from lazy-put flag set to MM_CID_UNSET.
11687 : */
11688 : rcu_read_lock();
11689 : src_task = rcu_dereference(src_rq->curr);
11690 : if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11691 : rcu_read_unlock();
11692 : /*
11693 : * We observed an active task for this mm, there is therefore
11694 : * no point in moving this cid to the destination cpu.
11695 : */
11696 : t->last_mm_cid = -1;
11697 : return -1;
11698 : }
11699 : rcu_read_unlock();
11700 :
11701 : /*
11702 : * The src_cid is unused, so it can be unset.
11703 : */
11704 : if (!try_cmpxchg(&src_pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11705 : return -1;
11706 : return src_cid;
11707 : }
11708 :
11709 : /*
11710 : * Migration to dst cpu. Called with dst_rq lock held.
11711 : * Interrupts are disabled, which keeps the window of cid ownership without the
11712 : * source rq lock held small.
11713 : */
11714 : void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t)
11715 : {
11716 : struct mm_cid *src_pcpu_cid, *dst_pcpu_cid;
11717 : struct mm_struct *mm = t->mm;
11718 : int src_cid, dst_cid, src_cpu;
11719 : struct rq *src_rq;
11720 :
11721 : lockdep_assert_rq_held(dst_rq);
11722 :
11723 : if (!mm)
11724 : return;
11725 : src_cpu = t->migrate_from_cpu;
11726 : if (src_cpu == -1) {
11727 : t->last_mm_cid = -1;
11728 : return;
11729 : }
11730 : /*
11731 : * Move the src cid if the dst cid is unset. This keeps id
11732 : * allocation closest to 0 in cases where few threads migrate around
11733 : * many cpus.
11734 : *
11735 : * If destination cid is already set, we may have to just clear
11736 : * the src cid to ensure compactness in frequent migrations
11737 : * scenarios.
11738 : *
11739 : * It is not useful to clear the src cid when the number of threads is
11740 : * greater or equal to the number of allowed cpus, because user-space
11741 : * can expect that the number of allowed cids can reach the number of
11742 : * allowed cpus.
11743 : */
11744 : dst_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(dst_rq));
11745 : dst_cid = READ_ONCE(dst_pcpu_cid->cid);
11746 : if (!mm_cid_is_unset(dst_cid) &&
11747 : atomic_read(&mm->mm_users) >= t->nr_cpus_allowed)
11748 : return;
11749 : src_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, src_cpu);
11750 : src_rq = cpu_rq(src_cpu);
11751 : src_cid = __sched_mm_cid_migrate_from_fetch_cid(src_rq, t, src_pcpu_cid);
11752 : if (src_cid == -1)
11753 : return;
11754 : src_cid = __sched_mm_cid_migrate_from_try_steal_cid(src_rq, t, src_pcpu_cid,
11755 : src_cid);
11756 : if (src_cid == -1)
11757 : return;
11758 : if (!mm_cid_is_unset(dst_cid)) {
11759 : __mm_cid_put(mm, src_cid);
11760 : return;
11761 : }
11762 : /* Move src_cid to dst cpu. */
11763 : mm_cid_snapshot_time(dst_rq, mm);
11764 : WRITE_ONCE(dst_pcpu_cid->cid, src_cid);
11765 : }
11766 :
11767 : static void sched_mm_cid_remote_clear(struct mm_struct *mm, struct mm_cid *pcpu_cid,
11768 : int cpu)
11769 : {
11770 : struct rq *rq = cpu_rq(cpu);
11771 : struct task_struct *t;
11772 : unsigned long flags;
11773 : int cid, lazy_cid;
11774 :
11775 : cid = READ_ONCE(pcpu_cid->cid);
11776 : if (!mm_cid_is_valid(cid))
11777 : return;
11778 :
11779 : /*
11780 : * Clear the cpu cid if it is set to keep cid allocation compact. If
11781 : * there happens to be other tasks left on the source cpu using this
11782 : * mm, the next task using this mm will reallocate its cid on context
11783 : * switch.
11784 : */
11785 : lazy_cid = mm_cid_set_lazy_put(cid);
11786 : if (!try_cmpxchg(&pcpu_cid->cid, &cid, lazy_cid))
11787 : return;
11788 :
11789 : /*
11790 : * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11791 : * rq->curr->mm matches the scheduler barrier in context_switch()
11792 : * between store to rq->curr and load of prev and next task's
11793 : * per-mm/cpu cid.
11794 : *
11795 : * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11796 : * rq->curr->mm_cid_active matches the barrier in
11797 : * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11798 : * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11799 : * load of per-mm/cpu cid.
11800 : */
11801 :
11802 : /*
11803 : * If we observe an active task using the mm on this rq after setting
11804 : * the lazy-put flag, that task will be responsible for transitioning
11805 : * from lazy-put flag set to MM_CID_UNSET.
11806 : */
11807 : rcu_read_lock();
11808 : t = rcu_dereference(rq->curr);
11809 : if (READ_ONCE(t->mm_cid_active) && t->mm == mm) {
11810 : rcu_read_unlock();
11811 : return;
11812 : }
11813 : rcu_read_unlock();
11814 :
11815 : /*
11816 : * The cid is unused, so it can be unset.
11817 : * Disable interrupts to keep the window of cid ownership without rq
11818 : * lock small.
11819 : */
11820 : local_irq_save(flags);
11821 : if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11822 : __mm_cid_put(mm, cid);
11823 : local_irq_restore(flags);
11824 : }
11825 :
11826 : static void sched_mm_cid_remote_clear_old(struct mm_struct *mm, int cpu)
11827 : {
11828 : struct rq *rq = cpu_rq(cpu);
11829 : struct mm_cid *pcpu_cid;
11830 : struct task_struct *curr;
11831 : u64 rq_clock;
11832 :
11833 : /*
11834 : * rq->clock load is racy on 32-bit but one spurious clear once in a
11835 : * while is irrelevant.
11836 : */
11837 : rq_clock = READ_ONCE(rq->clock);
11838 : pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11839 :
11840 : /*
11841 : * In order to take care of infrequently scheduled tasks, bump the time
11842 : * snapshot associated with this cid if an active task using the mm is
11843 : * observed on this rq.
11844 : */
11845 : rcu_read_lock();
11846 : curr = rcu_dereference(rq->curr);
11847 : if (READ_ONCE(curr->mm_cid_active) && curr->mm == mm) {
11848 : WRITE_ONCE(pcpu_cid->time, rq_clock);
11849 : rcu_read_unlock();
11850 : return;
11851 : }
11852 : rcu_read_unlock();
11853 :
11854 : if (rq_clock < pcpu_cid->time + SCHED_MM_CID_PERIOD_NS)
11855 : return;
11856 : sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11857 : }
11858 :
11859 : static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
11860 : int weight)
11861 : {
11862 : struct mm_cid *pcpu_cid;
11863 : int cid;
11864 :
11865 : pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11866 : cid = READ_ONCE(pcpu_cid->cid);
11867 : if (!mm_cid_is_valid(cid) || cid < weight)
11868 : return;
11869 : sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11870 : }
11871 :
11872 : static void task_mm_cid_work(struct callback_head *work)
11873 : {
11874 : unsigned long now = jiffies, old_scan, next_scan;
11875 : struct task_struct *t = current;
11876 : struct cpumask *cidmask;
11877 : struct mm_struct *mm;
11878 : int weight, cpu;
11879 :
11880 : SCHED_WARN_ON(t != container_of(work, struct task_struct, cid_work));
11881 :
11882 : work->next = work; /* Prevent double-add */
11883 : if (t->flags & PF_EXITING)
11884 : return;
11885 : mm = t->mm;
11886 : if (!mm)
11887 : return;
11888 : old_scan = READ_ONCE(mm->mm_cid_next_scan);
11889 : next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11890 : if (!old_scan) {
11891 : unsigned long res;
11892 :
11893 : res = cmpxchg(&mm->mm_cid_next_scan, old_scan, next_scan);
11894 : if (res != old_scan)
11895 : old_scan = res;
11896 : else
11897 : old_scan = next_scan;
11898 : }
11899 : if (time_before(now, old_scan))
11900 : return;
11901 : if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
11902 : return;
11903 : cidmask = mm_cidmask(mm);
11904 : /* Clear cids that were not recently used. */
11905 : for_each_possible_cpu(cpu)
11906 : sched_mm_cid_remote_clear_old(mm, cpu);
11907 : weight = cpumask_weight(cidmask);
11908 : /*
11909 : * Clear cids that are greater or equal to the cidmask weight to
11910 : * recompact it.
11911 : */
11912 : for_each_possible_cpu(cpu)
11913 : sched_mm_cid_remote_clear_weight(mm, cpu, weight);
11914 : }
11915 :
11916 : void init_sched_mm_cid(struct task_struct *t)
11917 : {
11918 : struct mm_struct *mm = t->mm;
11919 : int mm_users = 0;
11920 :
11921 : if (mm) {
11922 : mm_users = atomic_read(&mm->mm_users);
11923 : if (mm_users == 1)
11924 : mm->mm_cid_next_scan = jiffies + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11925 : }
11926 : t->cid_work.next = &t->cid_work; /* Protect against double add */
11927 : init_task_work(&t->cid_work, task_mm_cid_work);
11928 : }
11929 :
11930 : void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
11931 : {
11932 : struct callback_head *work = &curr->cid_work;
11933 : unsigned long now = jiffies;
11934 :
11935 : if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
11936 : work->next != work)
11937 : return;
11938 : if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
11939 : return;
11940 : task_work_add(curr, work, TWA_RESUME);
11941 : }
11942 :
11943 : void sched_mm_cid_exit_signals(struct task_struct *t)
11944 : {
11945 : struct mm_struct *mm = t->mm;
11946 : struct rq_flags rf;
11947 : struct rq *rq;
11948 :
11949 : if (!mm)
11950 : return;
11951 :
11952 : preempt_disable();
11953 : rq = this_rq();
11954 : rq_lock_irqsave(rq, &rf);
11955 : preempt_enable_no_resched(); /* holding spinlock */
11956 : WRITE_ONCE(t->mm_cid_active, 0);
11957 : /*
11958 : * Store t->mm_cid_active before loading per-mm/cpu cid.
11959 : * Matches barrier in sched_mm_cid_remote_clear_old().
11960 : */
11961 : smp_mb();
11962 : mm_cid_put(mm);
11963 : t->last_mm_cid = t->mm_cid = -1;
11964 : rq_unlock_irqrestore(rq, &rf);
11965 : }
11966 :
11967 : void sched_mm_cid_before_execve(struct task_struct *t)
11968 : {
11969 : struct mm_struct *mm = t->mm;
11970 : struct rq_flags rf;
11971 : struct rq *rq;
11972 :
11973 : if (!mm)
11974 : return;
11975 :
11976 : preempt_disable();
11977 : rq = this_rq();
11978 : rq_lock_irqsave(rq, &rf);
11979 : preempt_enable_no_resched(); /* holding spinlock */
11980 : WRITE_ONCE(t->mm_cid_active, 0);
11981 : /*
11982 : * Store t->mm_cid_active before loading per-mm/cpu cid.
11983 : * Matches barrier in sched_mm_cid_remote_clear_old().
11984 : */
11985 : smp_mb();
11986 : mm_cid_put(mm);
11987 : t->last_mm_cid = t->mm_cid = -1;
11988 : rq_unlock_irqrestore(rq, &rf);
11989 : }
11990 :
11991 : void sched_mm_cid_after_execve(struct task_struct *t)
11992 : {
11993 : struct mm_struct *mm = t->mm;
11994 : struct rq_flags rf;
11995 : struct rq *rq;
11996 :
11997 : if (!mm)
11998 : return;
11999 :
12000 : preempt_disable();
12001 : rq = this_rq();
12002 : rq_lock_irqsave(rq, &rf);
12003 : preempt_enable_no_resched(); /* holding spinlock */
12004 : WRITE_ONCE(t->mm_cid_active, 1);
12005 : /*
12006 : * Store t->mm_cid_active before loading per-mm/cpu cid.
12007 : * Matches barrier in sched_mm_cid_remote_clear_old().
12008 : */
12009 : smp_mb();
12010 : t->last_mm_cid = t->mm_cid = mm_cid_get(rq, mm);
12011 : rq_unlock_irqrestore(rq, &rf);
12012 : rseq_set_notify_resume(t);
12013 : }
12014 :
12015 : void sched_mm_cid_fork(struct task_struct *t)
12016 : {
12017 : WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
12018 : t->mm_cid_active = 1;
12019 : }
12020 : #endif
|
