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 :
551 175 : 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 2450 : preempt_disable();
557 : if (sched_core_disabled()) {
558 2450 : raw_spin_lock_nested(&rq->__lock, subclass);
559 : /* preempt_count *MUST* be > 1 */
560 2450 : 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 :
600 175 : void raw_spin_rq_unlock(struct rq *rq)
601 : {
602 2450 : raw_spin_unlock(rq_lockp(rq));
603 175 : }
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 (;;) {
635 175 : rq = task_rq(p);
636 175 : raw_spin_rq_lock(rq);
637 350 : if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
638 175 : 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 : */
651 190 : 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 (;;) {
658 190 : raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
659 190 : rq = task_rq(p);
660 190 : 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 : */
678 380 : if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
679 190 : rq_pin_lock(rq, rf);
680 190 : 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 2082 : 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 2082 : 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 2082 : update_rq_clock_pelt(rq, delta);
748 : }
749 :
750 2421 : void update_rq_clock(struct rq *rq)
751 : {
752 : s64 delta;
753 :
754 2421 : lockdep_assert_rq_held(rq);
755 :
756 2421 : 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 2082 : delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
766 2082 : if (delta < 0)
767 : return;
768 2082 : rq->clock += delta;
769 2082 : 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 356 : void resched_curr(struct rq *rq)
1042 : {
1043 356 : struct task_struct *curr = rq->curr;
1044 : int cpu;
1045 :
1046 712 : lockdep_assert_rq_held(rq);
1047 :
1048 356 : if (test_tsk_need_resched(curr))
1049 : return;
1050 :
1051 340 : 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 16 : void resched_cpu(int cpu)
1066 : {
1067 16 : struct rq *rq = cpu_rq(cpu);
1068 : unsigned long flags;
1069 :
1070 32 : raw_spin_rq_lock_irqsave(rq, flags);
1071 16 : if (cpu_online(cpu) || cpu == smp_processor_id())
1072 16 : resched_curr(rq);
1073 32 : raw_spin_rq_unlock_irqrestore(rq, flags);
1074 16 : }
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 1031 : static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2072 : {
2073 1031 : 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 1035 : uclamp_rq_inc(rq, p);
2082 1035 : p->sched_class->enqueue_task(rq, p, flags);
2083 :
2084 1035 : if (sched_core_enabled(rq))
2085 : sched_core_enqueue(rq, p);
2086 1031 : }
2087 :
2088 1029 : static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2089 : {
2090 1033 : if (sched_core_enabled(rq))
2091 : sched_core_dequeue(rq, p, flags);
2092 :
2093 1029 : 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 1033 : uclamp_rq_dec(rq, p);
2102 1033 : p->sched_class->dequeue_task(rq, p, flags);
2103 1029 : }
2104 :
2105 0 : void activate_task(struct rq *rq, struct task_struct *p, int flags)
2106 : {
2107 1031 : 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 1031 : enqueue_task(rq, p, flags);
2113 :
2114 1031 : 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 1029 : p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2120 :
2121 1029 : 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 1031 : void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2202 : {
2203 1031 : if (p->sched_class == rq->curr->sched_class)
2204 1028 : 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 2062 : if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2213 505 : rq_clock_skip_update(rq);
2214 1031 : }
2215 :
2216 : static __always_inline
2217 : int __task_state_match(struct task_struct *p, unsigned int state)
2218 : {
2219 1042 : if (READ_ONCE(p->__state) & state)
2220 : return 1;
2221 :
2222 : #ifdef CONFIG_PREEMPT_RT
2223 : if (READ_ONCE(p->saved_state) & state)
2224 : return -1;
2225 : #endif
2226 : return 0;
2227 : }
2228 :
2229 : static __always_inline
2230 : int task_state_match(struct task_struct *p, unsigned int state)
2231 : {
2232 : #ifdef CONFIG_PREEMPT_RT
2233 : int match;
2234 :
2235 : /*
2236 : * Serialize against current_save_and_set_rtlock_wait_state() and
2237 : * current_restore_rtlock_saved_state().
2238 : */
2239 : raw_spin_lock_irq(&p->pi_lock);
2240 : match = __task_state_match(p, state);
2241 : raw_spin_unlock_irq(&p->pi_lock);
2242 :
2243 : return match;
2244 : #else
2245 0 : return __task_state_match(p, state);
2246 : #endif
2247 : }
2248 :
2249 : /*
2250 : * wait_task_inactive - wait for a thread to unschedule.
2251 : *
2252 : * Wait for the thread to block in any of the states set in @match_state.
2253 : * If it changes, i.e. @p might have woken up, then return zero. When we
2254 : * succeed in waiting for @p to be off its CPU, we return a positive number
2255 : * (its total switch count). If a second call a short while later returns the
2256 : * same number, the caller can be sure that @p has remained unscheduled the
2257 : * whole time.
2258 : *
2259 : * The caller must ensure that the task *will* unschedule sometime soon,
2260 : * else this function might spin for a *long* time. This function can't
2261 : * be called with interrupts off, or it may introduce deadlock with
2262 : * smp_call_function() if an IPI is sent by the same process we are
2263 : * waiting to become inactive.
2264 : */
2265 12 : unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
2266 : {
2267 : int running, queued, match;
2268 : struct rq_flags rf;
2269 : unsigned long ncsw;
2270 : struct rq *rq;
2271 :
2272 : for (;;) {
2273 : /*
2274 : * We do the initial early heuristics without holding
2275 : * any task-queue locks at all. We'll only try to get
2276 : * the runqueue lock when things look like they will
2277 : * work out!
2278 : */
2279 12 : rq = task_rq(p);
2280 :
2281 : /*
2282 : * If the task is actively running on another CPU
2283 : * still, just relax and busy-wait without holding
2284 : * any locks.
2285 : *
2286 : * NOTE! Since we don't hold any locks, it's not
2287 : * even sure that "rq" stays as the right runqueue!
2288 : * But we don't care, since "task_on_cpu()" will
2289 : * return false if the runqueue has changed and p
2290 : * is actually now running somewhere else!
2291 : */
2292 24 : while (task_on_cpu(rq, p)) {
2293 0 : if (!task_state_match(p, match_state))
2294 : return 0;
2295 : cpu_relax();
2296 : }
2297 :
2298 : /*
2299 : * Ok, time to look more closely! We need the rq
2300 : * lock now, to be *sure*. If we're wrong, we'll
2301 : * just go back and repeat.
2302 : */
2303 12 : rq = task_rq_lock(p, &rf);
2304 12 : trace_sched_wait_task(p);
2305 12 : running = task_on_cpu(rq, p);
2306 24 : queued = task_on_rq_queued(p);
2307 12 : ncsw = 0;
2308 12 : if ((match = __task_state_match(p, match_state))) {
2309 : /*
2310 : * When matching on p->saved_state, consider this task
2311 : * still queued so it will wait.
2312 : */
2313 : if (match < 0)
2314 : queued = 1;
2315 12 : ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2316 : }
2317 24 : task_rq_unlock(rq, p, &rf);
2318 :
2319 : /*
2320 : * If it changed from the expected state, bail out now.
2321 : */
2322 12 : if (unlikely(!ncsw))
2323 : break;
2324 :
2325 : /*
2326 : * Was it really running after all now that we
2327 : * checked with the proper locks actually held?
2328 : *
2329 : * Oops. Go back and try again..
2330 : */
2331 12 : if (unlikely(running)) {
2332 : cpu_relax();
2333 0 : continue;
2334 : }
2335 :
2336 : /*
2337 : * It's not enough that it's not actively running,
2338 : * it must be off the runqueue _entirely_, and not
2339 : * preempted!
2340 : *
2341 : * So if it was still runnable (but just not actively
2342 : * running right now), it's preempted, and we should
2343 : * yield - it could be a while.
2344 : */
2345 12 : if (unlikely(queued)) {
2346 0 : ktime_t to = NSEC_PER_SEC / HZ;
2347 :
2348 0 : set_current_state(TASK_UNINTERRUPTIBLE);
2349 0 : schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
2350 0 : continue;
2351 : }
2352 :
2353 : /*
2354 : * Ahh, all good. It wasn't running, and it wasn't
2355 : * runnable, which means that it will never become
2356 : * running in the future either. We're all done!
2357 : */
2358 : break;
2359 : }
2360 :
2361 : return ncsw;
2362 : }
2363 :
2364 : #ifdef CONFIG_SMP
2365 :
2366 : static void
2367 : __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
2368 :
2369 : static int __set_cpus_allowed_ptr(struct task_struct *p,
2370 : struct affinity_context *ctx);
2371 :
2372 : static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2373 : {
2374 : struct affinity_context ac = {
2375 : .new_mask = cpumask_of(rq->cpu),
2376 : .flags = SCA_MIGRATE_DISABLE,
2377 : };
2378 :
2379 : if (likely(!p->migration_disabled))
2380 : return;
2381 :
2382 : if (p->cpus_ptr != &p->cpus_mask)
2383 : return;
2384 :
2385 : /*
2386 : * Violates locking rules! see comment in __do_set_cpus_allowed().
2387 : */
2388 : __do_set_cpus_allowed(p, &ac);
2389 : }
2390 :
2391 : void migrate_disable(void)
2392 : {
2393 : struct task_struct *p = current;
2394 :
2395 : if (p->migration_disabled) {
2396 : p->migration_disabled++;
2397 : return;
2398 : }
2399 :
2400 : preempt_disable();
2401 : this_rq()->nr_pinned++;
2402 : p->migration_disabled = 1;
2403 : preempt_enable();
2404 : }
2405 : EXPORT_SYMBOL_GPL(migrate_disable);
2406 :
2407 : void migrate_enable(void)
2408 : {
2409 : struct task_struct *p = current;
2410 : struct affinity_context ac = {
2411 : .new_mask = &p->cpus_mask,
2412 : .flags = SCA_MIGRATE_ENABLE,
2413 : };
2414 :
2415 : if (p->migration_disabled > 1) {
2416 : p->migration_disabled--;
2417 : return;
2418 : }
2419 :
2420 : if (WARN_ON_ONCE(!p->migration_disabled))
2421 : return;
2422 :
2423 : /*
2424 : * Ensure stop_task runs either before or after this, and that
2425 : * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2426 : */
2427 : preempt_disable();
2428 : if (p->cpus_ptr != &p->cpus_mask)
2429 : __set_cpus_allowed_ptr(p, &ac);
2430 : /*
2431 : * Mustn't clear migration_disabled() until cpus_ptr points back at the
2432 : * regular cpus_mask, otherwise things that race (eg.
2433 : * select_fallback_rq) get confused.
2434 : */
2435 : barrier();
2436 : p->migration_disabled = 0;
2437 : this_rq()->nr_pinned--;
2438 : preempt_enable();
2439 : }
2440 : EXPORT_SYMBOL_GPL(migrate_enable);
2441 :
2442 : static inline bool rq_has_pinned_tasks(struct rq *rq)
2443 : {
2444 : return rq->nr_pinned;
2445 : }
2446 :
2447 : /*
2448 : * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2449 : * __set_cpus_allowed_ptr() and select_fallback_rq().
2450 : */
2451 : static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2452 : {
2453 : /* When not in the task's cpumask, no point in looking further. */
2454 : if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2455 : return false;
2456 :
2457 : /* migrate_disabled() must be allowed to finish. */
2458 : if (is_migration_disabled(p))
2459 : return cpu_online(cpu);
2460 :
2461 : /* Non kernel threads are not allowed during either online or offline. */
2462 : if (!(p->flags & PF_KTHREAD))
2463 : return cpu_active(cpu) && task_cpu_possible(cpu, p);
2464 :
2465 : /* KTHREAD_IS_PER_CPU is always allowed. */
2466 : if (kthread_is_per_cpu(p))
2467 : return cpu_online(cpu);
2468 :
2469 : /* Regular kernel threads don't get to stay during offline. */
2470 : if (cpu_dying(cpu))
2471 : return false;
2472 :
2473 : /* But are allowed during online. */
2474 : return cpu_online(cpu);
2475 : }
2476 :
2477 : /*
2478 : * This is how migration works:
2479 : *
2480 : * 1) we invoke migration_cpu_stop() on the target CPU using
2481 : * stop_one_cpu().
2482 : * 2) stopper starts to run (implicitly forcing the migrated thread
2483 : * off the CPU)
2484 : * 3) it checks whether the migrated task is still in the wrong runqueue.
2485 : * 4) if it's in the wrong runqueue then the migration thread removes
2486 : * it and puts it into the right queue.
2487 : * 5) stopper completes and stop_one_cpu() returns and the migration
2488 : * is done.
2489 : */
2490 :
2491 : /*
2492 : * move_queued_task - move a queued task to new rq.
2493 : *
2494 : * Returns (locked) new rq. Old rq's lock is released.
2495 : */
2496 : static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2497 : struct task_struct *p, int new_cpu)
2498 : {
2499 : lockdep_assert_rq_held(rq);
2500 :
2501 : deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2502 : set_task_cpu(p, new_cpu);
2503 : rq_unlock(rq, rf);
2504 :
2505 : rq = cpu_rq(new_cpu);
2506 :
2507 : rq_lock(rq, rf);
2508 : WARN_ON_ONCE(task_cpu(p) != new_cpu);
2509 : activate_task(rq, p, 0);
2510 : check_preempt_curr(rq, p, 0);
2511 :
2512 : return rq;
2513 : }
2514 :
2515 : struct migration_arg {
2516 : struct task_struct *task;
2517 : int dest_cpu;
2518 : struct set_affinity_pending *pending;
2519 : };
2520 :
2521 : /*
2522 : * @refs: number of wait_for_completion()
2523 : * @stop_pending: is @stop_work in use
2524 : */
2525 : struct set_affinity_pending {
2526 : refcount_t refs;
2527 : unsigned int stop_pending;
2528 : struct completion done;
2529 : struct cpu_stop_work stop_work;
2530 : struct migration_arg arg;
2531 : };
2532 :
2533 : /*
2534 : * Move (not current) task off this CPU, onto the destination CPU. We're doing
2535 : * this because either it can't run here any more (set_cpus_allowed()
2536 : * away from this CPU, or CPU going down), or because we're
2537 : * attempting to rebalance this task on exec (sched_exec).
2538 : *
2539 : * So we race with normal scheduler movements, but that's OK, as long
2540 : * as the task is no longer on this CPU.
2541 : */
2542 : static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2543 : struct task_struct *p, int dest_cpu)
2544 : {
2545 : /* Affinity changed (again). */
2546 : if (!is_cpu_allowed(p, dest_cpu))
2547 : return rq;
2548 :
2549 : rq = move_queued_task(rq, rf, p, dest_cpu);
2550 :
2551 : return rq;
2552 : }
2553 :
2554 : /*
2555 : * migration_cpu_stop - this will be executed by a highprio stopper thread
2556 : * and performs thread migration by bumping thread off CPU then
2557 : * 'pushing' onto another runqueue.
2558 : */
2559 : static int migration_cpu_stop(void *data)
2560 : {
2561 : struct migration_arg *arg = data;
2562 : struct set_affinity_pending *pending = arg->pending;
2563 : struct task_struct *p = arg->task;
2564 : struct rq *rq = this_rq();
2565 : bool complete = false;
2566 : struct rq_flags rf;
2567 :
2568 : /*
2569 : * The original target CPU might have gone down and we might
2570 : * be on another CPU but it doesn't matter.
2571 : */
2572 : local_irq_save(rf.flags);
2573 : /*
2574 : * We need to explicitly wake pending tasks before running
2575 : * __migrate_task() such that we will not miss enforcing cpus_ptr
2576 : * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2577 : */
2578 : flush_smp_call_function_queue();
2579 :
2580 : raw_spin_lock(&p->pi_lock);
2581 : rq_lock(rq, &rf);
2582 :
2583 : /*
2584 : * If we were passed a pending, then ->stop_pending was set, thus
2585 : * p->migration_pending must have remained stable.
2586 : */
2587 : WARN_ON_ONCE(pending && pending != p->migration_pending);
2588 :
2589 : /*
2590 : * If task_rq(p) != rq, it cannot be migrated here, because we're
2591 : * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2592 : * we're holding p->pi_lock.
2593 : */
2594 : if (task_rq(p) == rq) {
2595 : if (is_migration_disabled(p))
2596 : goto out;
2597 :
2598 : if (pending) {
2599 : p->migration_pending = NULL;
2600 : complete = true;
2601 :
2602 : if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2603 : goto out;
2604 : }
2605 :
2606 : if (task_on_rq_queued(p)) {
2607 : update_rq_clock(rq);
2608 : rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2609 : } else {
2610 : p->wake_cpu = arg->dest_cpu;
2611 : }
2612 :
2613 : /*
2614 : * XXX __migrate_task() can fail, at which point we might end
2615 : * up running on a dodgy CPU, AFAICT this can only happen
2616 : * during CPU hotplug, at which point we'll get pushed out
2617 : * anyway, so it's probably not a big deal.
2618 : */
2619 :
2620 : } else if (pending) {
2621 : /*
2622 : * This happens when we get migrated between migrate_enable()'s
2623 : * preempt_enable() and scheduling the stopper task. At that
2624 : * point we're a regular task again and not current anymore.
2625 : *
2626 : * A !PREEMPT kernel has a giant hole here, which makes it far
2627 : * more likely.
2628 : */
2629 :
2630 : /*
2631 : * The task moved before the stopper got to run. We're holding
2632 : * ->pi_lock, so the allowed mask is stable - if it got
2633 : * somewhere allowed, we're done.
2634 : */
2635 : if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2636 : p->migration_pending = NULL;
2637 : complete = true;
2638 : goto out;
2639 : }
2640 :
2641 : /*
2642 : * When migrate_enable() hits a rq mis-match we can't reliably
2643 : * determine is_migration_disabled() and so have to chase after
2644 : * it.
2645 : */
2646 : WARN_ON_ONCE(!pending->stop_pending);
2647 : task_rq_unlock(rq, p, &rf);
2648 : stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2649 : &pending->arg, &pending->stop_work);
2650 : return 0;
2651 : }
2652 : out:
2653 : if (pending)
2654 : pending->stop_pending = false;
2655 : task_rq_unlock(rq, p, &rf);
2656 :
2657 : if (complete)
2658 : complete_all(&pending->done);
2659 :
2660 : return 0;
2661 : }
2662 :
2663 : int push_cpu_stop(void *arg)
2664 : {
2665 : struct rq *lowest_rq = NULL, *rq = this_rq();
2666 : struct task_struct *p = arg;
2667 :
2668 : raw_spin_lock_irq(&p->pi_lock);
2669 : raw_spin_rq_lock(rq);
2670 :
2671 : if (task_rq(p) != rq)
2672 : goto out_unlock;
2673 :
2674 : if (is_migration_disabled(p)) {
2675 : p->migration_flags |= MDF_PUSH;
2676 : goto out_unlock;
2677 : }
2678 :
2679 : p->migration_flags &= ~MDF_PUSH;
2680 :
2681 : if (p->sched_class->find_lock_rq)
2682 : lowest_rq = p->sched_class->find_lock_rq(p, rq);
2683 :
2684 : if (!lowest_rq)
2685 : goto out_unlock;
2686 :
2687 : // XXX validate p is still the highest prio task
2688 : if (task_rq(p) == rq) {
2689 : deactivate_task(rq, p, 0);
2690 : set_task_cpu(p, lowest_rq->cpu);
2691 : activate_task(lowest_rq, p, 0);
2692 : resched_curr(lowest_rq);
2693 : }
2694 :
2695 : double_unlock_balance(rq, lowest_rq);
2696 :
2697 : out_unlock:
2698 : rq->push_busy = false;
2699 : raw_spin_rq_unlock(rq);
2700 : raw_spin_unlock_irq(&p->pi_lock);
2701 :
2702 : put_task_struct(p);
2703 : return 0;
2704 : }
2705 :
2706 : /*
2707 : * sched_class::set_cpus_allowed must do the below, but is not required to
2708 : * actually call this function.
2709 : */
2710 : void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
2711 : {
2712 : if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2713 : p->cpus_ptr = ctx->new_mask;
2714 : return;
2715 : }
2716 :
2717 : cpumask_copy(&p->cpus_mask, ctx->new_mask);
2718 : p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
2719 :
2720 : /*
2721 : * Swap in a new user_cpus_ptr if SCA_USER flag set
2722 : */
2723 : if (ctx->flags & SCA_USER)
2724 : swap(p->user_cpus_ptr, ctx->user_mask);
2725 : }
2726 :
2727 : static void
2728 : __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
2729 : {
2730 : struct rq *rq = task_rq(p);
2731 : bool queued, running;
2732 :
2733 : /*
2734 : * This here violates the locking rules for affinity, since we're only
2735 : * supposed to change these variables while holding both rq->lock and
2736 : * p->pi_lock.
2737 : *
2738 : * HOWEVER, it magically works, because ttwu() is the only code that
2739 : * accesses these variables under p->pi_lock and only does so after
2740 : * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2741 : * before finish_task().
2742 : *
2743 : * XXX do further audits, this smells like something putrid.
2744 : */
2745 : if (ctx->flags & SCA_MIGRATE_DISABLE)
2746 : SCHED_WARN_ON(!p->on_cpu);
2747 : else
2748 : lockdep_assert_held(&p->pi_lock);
2749 :
2750 : queued = task_on_rq_queued(p);
2751 : running = task_current(rq, p);
2752 :
2753 : if (queued) {
2754 : /*
2755 : * Because __kthread_bind() calls this on blocked tasks without
2756 : * holding rq->lock.
2757 : */
2758 : lockdep_assert_rq_held(rq);
2759 : dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2760 : }
2761 : if (running)
2762 : put_prev_task(rq, p);
2763 :
2764 : p->sched_class->set_cpus_allowed(p, ctx);
2765 :
2766 : if (queued)
2767 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2768 : if (running)
2769 : set_next_task(rq, p);
2770 : }
2771 :
2772 : /*
2773 : * Used for kthread_bind() and select_fallback_rq(), in both cases the user
2774 : * affinity (if any) should be destroyed too.
2775 : */
2776 : void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2777 : {
2778 : struct affinity_context ac = {
2779 : .new_mask = new_mask,
2780 : .user_mask = NULL,
2781 : .flags = SCA_USER, /* clear the user requested mask */
2782 : };
2783 : union cpumask_rcuhead {
2784 : cpumask_t cpumask;
2785 : struct rcu_head rcu;
2786 : };
2787 :
2788 : __do_set_cpus_allowed(p, &ac);
2789 :
2790 : /*
2791 : * Because this is called with p->pi_lock held, it is not possible
2792 : * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2793 : * kfree_rcu().
2794 : */
2795 : kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
2796 : }
2797 :
2798 : static cpumask_t *alloc_user_cpus_ptr(int node)
2799 : {
2800 : /*
2801 : * See do_set_cpus_allowed() above for the rcu_head usage.
2802 : */
2803 : int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2804 :
2805 : return kmalloc_node(size, GFP_KERNEL, node);
2806 : }
2807 :
2808 : int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2809 : int node)
2810 : {
2811 : cpumask_t *user_mask;
2812 : unsigned long flags;
2813 :
2814 : /*
2815 : * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2816 : * may differ by now due to racing.
2817 : */
2818 : dst->user_cpus_ptr = NULL;
2819 :
2820 : /*
2821 : * This check is racy and losing the race is a valid situation.
2822 : * It is not worth the extra overhead of taking the pi_lock on
2823 : * every fork/clone.
2824 : */
2825 : if (data_race(!src->user_cpus_ptr))
2826 : return 0;
2827 :
2828 : user_mask = alloc_user_cpus_ptr(node);
2829 : if (!user_mask)
2830 : return -ENOMEM;
2831 :
2832 : /*
2833 : * Use pi_lock to protect content of user_cpus_ptr
2834 : *
2835 : * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2836 : * do_set_cpus_allowed().
2837 : */
2838 : raw_spin_lock_irqsave(&src->pi_lock, flags);
2839 : if (src->user_cpus_ptr) {
2840 : swap(dst->user_cpus_ptr, user_mask);
2841 : cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2842 : }
2843 : raw_spin_unlock_irqrestore(&src->pi_lock, flags);
2844 :
2845 : if (unlikely(user_mask))
2846 : kfree(user_mask);
2847 :
2848 : return 0;
2849 : }
2850 :
2851 : static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2852 : {
2853 : struct cpumask *user_mask = NULL;
2854 :
2855 : swap(p->user_cpus_ptr, user_mask);
2856 :
2857 : return user_mask;
2858 : }
2859 :
2860 : void release_user_cpus_ptr(struct task_struct *p)
2861 : {
2862 : kfree(clear_user_cpus_ptr(p));
2863 : }
2864 :
2865 : /*
2866 : * This function is wildly self concurrent; here be dragons.
2867 : *
2868 : *
2869 : * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2870 : * designated task is enqueued on an allowed CPU. If that task is currently
2871 : * running, we have to kick it out using the CPU stopper.
2872 : *
2873 : * Migrate-Disable comes along and tramples all over our nice sandcastle.
2874 : * Consider:
2875 : *
2876 : * Initial conditions: P0->cpus_mask = [0, 1]
2877 : *
2878 : * P0@CPU0 P1
2879 : *
2880 : * migrate_disable();
2881 : * <preempted>
2882 : * set_cpus_allowed_ptr(P0, [1]);
2883 : *
2884 : * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2885 : * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2886 : * This means we need the following scheme:
2887 : *
2888 : * P0@CPU0 P1
2889 : *
2890 : * migrate_disable();
2891 : * <preempted>
2892 : * set_cpus_allowed_ptr(P0, [1]);
2893 : * <blocks>
2894 : * <resumes>
2895 : * migrate_enable();
2896 : * __set_cpus_allowed_ptr();
2897 : * <wakes local stopper>
2898 : * `--> <woken on migration completion>
2899 : *
2900 : * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2901 : * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2902 : * task p are serialized by p->pi_lock, which we can leverage: the one that
2903 : * should come into effect at the end of the Migrate-Disable region is the last
2904 : * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2905 : * but we still need to properly signal those waiting tasks at the appropriate
2906 : * moment.
2907 : *
2908 : * This is implemented using struct set_affinity_pending. The first
2909 : * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2910 : * setup an instance of that struct and install it on the targeted task_struct.
2911 : * Any and all further callers will reuse that instance. Those then wait for
2912 : * a completion signaled at the tail of the CPU stopper callback (1), triggered
2913 : * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2914 : *
2915 : *
2916 : * (1) In the cases covered above. There is one more where the completion is
2917 : * signaled within affine_move_task() itself: when a subsequent affinity request
2918 : * occurs after the stopper bailed out due to the targeted task still being
2919 : * Migrate-Disable. Consider:
2920 : *
2921 : * Initial conditions: P0->cpus_mask = [0, 1]
2922 : *
2923 : * CPU0 P1 P2
2924 : * <P0>
2925 : * migrate_disable();
2926 : * <preempted>
2927 : * set_cpus_allowed_ptr(P0, [1]);
2928 : * <blocks>
2929 : * <migration/0>
2930 : * migration_cpu_stop()
2931 : * is_migration_disabled()
2932 : * <bails>
2933 : * set_cpus_allowed_ptr(P0, [0, 1]);
2934 : * <signal completion>
2935 : * <awakes>
2936 : *
2937 : * Note that the above is safe vs a concurrent migrate_enable(), as any
2938 : * pending affinity completion is preceded by an uninstallation of
2939 : * p->migration_pending done with p->pi_lock held.
2940 : */
2941 : static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2942 : int dest_cpu, unsigned int flags)
2943 : __releases(rq->lock)
2944 : __releases(p->pi_lock)
2945 : {
2946 : struct set_affinity_pending my_pending = { }, *pending = NULL;
2947 : bool stop_pending, complete = false;
2948 :
2949 : /* Can the task run on the task's current CPU? If so, we're done */
2950 : if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2951 : struct task_struct *push_task = NULL;
2952 :
2953 : if ((flags & SCA_MIGRATE_ENABLE) &&
2954 : (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2955 : rq->push_busy = true;
2956 : push_task = get_task_struct(p);
2957 : }
2958 :
2959 : /*
2960 : * If there are pending waiters, but no pending stop_work,
2961 : * then complete now.
2962 : */
2963 : pending = p->migration_pending;
2964 : if (pending && !pending->stop_pending) {
2965 : p->migration_pending = NULL;
2966 : complete = true;
2967 : }
2968 :
2969 : task_rq_unlock(rq, p, rf);
2970 :
2971 : if (push_task) {
2972 : stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2973 : p, &rq->push_work);
2974 : }
2975 :
2976 : if (complete)
2977 : complete_all(&pending->done);
2978 :
2979 : return 0;
2980 : }
2981 :
2982 : if (!(flags & SCA_MIGRATE_ENABLE)) {
2983 : /* serialized by p->pi_lock */
2984 : if (!p->migration_pending) {
2985 : /* Install the request */
2986 : refcount_set(&my_pending.refs, 1);
2987 : init_completion(&my_pending.done);
2988 : my_pending.arg = (struct migration_arg) {
2989 : .task = p,
2990 : .dest_cpu = dest_cpu,
2991 : .pending = &my_pending,
2992 : };
2993 :
2994 : p->migration_pending = &my_pending;
2995 : } else {
2996 : pending = p->migration_pending;
2997 : refcount_inc(&pending->refs);
2998 : /*
2999 : * Affinity has changed, but we've already installed a
3000 : * pending. migration_cpu_stop() *must* see this, else
3001 : * we risk a completion of the pending despite having a
3002 : * task on a disallowed CPU.
3003 : *
3004 : * Serialized by p->pi_lock, so this is safe.
3005 : */
3006 : pending->arg.dest_cpu = dest_cpu;
3007 : }
3008 : }
3009 : pending = p->migration_pending;
3010 : /*
3011 : * - !MIGRATE_ENABLE:
3012 : * we'll have installed a pending if there wasn't one already.
3013 : *
3014 : * - MIGRATE_ENABLE:
3015 : * we're here because the current CPU isn't matching anymore,
3016 : * the only way that can happen is because of a concurrent
3017 : * set_cpus_allowed_ptr() call, which should then still be
3018 : * pending completion.
3019 : *
3020 : * Either way, we really should have a @pending here.
3021 : */
3022 : if (WARN_ON_ONCE(!pending)) {
3023 : task_rq_unlock(rq, p, rf);
3024 : return -EINVAL;
3025 : }
3026 :
3027 : if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
3028 : /*
3029 : * MIGRATE_ENABLE gets here because 'p == current', but for
3030 : * anything else we cannot do is_migration_disabled(), punt
3031 : * and have the stopper function handle it all race-free.
3032 : */
3033 : stop_pending = pending->stop_pending;
3034 : if (!stop_pending)
3035 : pending->stop_pending = true;
3036 :
3037 : if (flags & SCA_MIGRATE_ENABLE)
3038 : p->migration_flags &= ~MDF_PUSH;
3039 :
3040 : task_rq_unlock(rq, p, rf);
3041 :
3042 : if (!stop_pending) {
3043 : stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
3044 : &pending->arg, &pending->stop_work);
3045 : }
3046 :
3047 : if (flags & SCA_MIGRATE_ENABLE)
3048 : return 0;
3049 : } else {
3050 :
3051 : if (!is_migration_disabled(p)) {
3052 : if (task_on_rq_queued(p))
3053 : rq = move_queued_task(rq, rf, p, dest_cpu);
3054 :
3055 : if (!pending->stop_pending) {
3056 : p->migration_pending = NULL;
3057 : complete = true;
3058 : }
3059 : }
3060 : task_rq_unlock(rq, p, rf);
3061 :
3062 : if (complete)
3063 : complete_all(&pending->done);
3064 : }
3065 :
3066 : wait_for_completion(&pending->done);
3067 :
3068 : if (refcount_dec_and_test(&pending->refs))
3069 : wake_up_var(&pending->refs); /* No UaF, just an address */
3070 :
3071 : /*
3072 : * Block the original owner of &pending until all subsequent callers
3073 : * have seen the completion and decremented the refcount
3074 : */
3075 : wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
3076 :
3077 : /* ARGH */
3078 : WARN_ON_ONCE(my_pending.stop_pending);
3079 :
3080 : return 0;
3081 : }
3082 :
3083 : /*
3084 : * Called with both p->pi_lock and rq->lock held; drops both before returning.
3085 : */
3086 : static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
3087 : struct affinity_context *ctx,
3088 : struct rq *rq,
3089 : struct rq_flags *rf)
3090 : __releases(rq->lock)
3091 : __releases(p->pi_lock)
3092 : {
3093 : const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
3094 : const struct cpumask *cpu_valid_mask = cpu_active_mask;
3095 : bool kthread = p->flags & PF_KTHREAD;
3096 : unsigned int dest_cpu;
3097 : int ret = 0;
3098 :
3099 : update_rq_clock(rq);
3100 :
3101 : if (kthread || is_migration_disabled(p)) {
3102 : /*
3103 : * Kernel threads are allowed on online && !active CPUs,
3104 : * however, during cpu-hot-unplug, even these might get pushed
3105 : * away if not KTHREAD_IS_PER_CPU.
3106 : *
3107 : * Specifically, migration_disabled() tasks must not fail the
3108 : * cpumask_any_and_distribute() pick below, esp. so on
3109 : * SCA_MIGRATE_ENABLE, otherwise we'll not call
3110 : * set_cpus_allowed_common() and actually reset p->cpus_ptr.
3111 : */
3112 : cpu_valid_mask = cpu_online_mask;
3113 : }
3114 :
3115 : if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
3116 : ret = -EINVAL;
3117 : goto out;
3118 : }
3119 :
3120 : /*
3121 : * Must re-check here, to close a race against __kthread_bind(),
3122 : * sched_setaffinity() is not guaranteed to observe the flag.
3123 : */
3124 : if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
3125 : ret = -EINVAL;
3126 : goto out;
3127 : }
3128 :
3129 : if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
3130 : if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) {
3131 : if (ctx->flags & SCA_USER)
3132 : swap(p->user_cpus_ptr, ctx->user_mask);
3133 : goto out;
3134 : }
3135 :
3136 : if (WARN_ON_ONCE(p == current &&
3137 : is_migration_disabled(p) &&
3138 : !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
3139 : ret = -EBUSY;
3140 : goto out;
3141 : }
3142 : }
3143 :
3144 : /*
3145 : * Picking a ~random cpu helps in cases where we are changing affinity
3146 : * for groups of tasks (ie. cpuset), so that load balancing is not
3147 : * immediately required to distribute the tasks within their new mask.
3148 : */
3149 : dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
3150 : if (dest_cpu >= nr_cpu_ids) {
3151 : ret = -EINVAL;
3152 : goto out;
3153 : }
3154 :
3155 : __do_set_cpus_allowed(p, ctx);
3156 :
3157 : return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
3158 :
3159 : out:
3160 : task_rq_unlock(rq, p, rf);
3161 :
3162 : return ret;
3163 : }
3164 :
3165 : /*
3166 : * Change a given task's CPU affinity. Migrate the thread to a
3167 : * proper CPU and schedule it away if the CPU it's executing on
3168 : * is removed from the allowed bitmask.
3169 : *
3170 : * NOTE: the caller must have a valid reference to the task, the
3171 : * task must not exit() & deallocate itself prematurely. The
3172 : * call is not atomic; no spinlocks may be held.
3173 : */
3174 : static int __set_cpus_allowed_ptr(struct task_struct *p,
3175 : struct affinity_context *ctx)
3176 : {
3177 : struct rq_flags rf;
3178 : struct rq *rq;
3179 :
3180 : rq = task_rq_lock(p, &rf);
3181 : /*
3182 : * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3183 : * flags are set.
3184 : */
3185 : if (p->user_cpus_ptr &&
3186 : !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
3187 : cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
3188 : ctx->new_mask = rq->scratch_mask;
3189 :
3190 : return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
3191 : }
3192 :
3193 : int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3194 : {
3195 : struct affinity_context ac = {
3196 : .new_mask = new_mask,
3197 : .flags = 0,
3198 : };
3199 :
3200 : return __set_cpus_allowed_ptr(p, &ac);
3201 : }
3202 : EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
3203 :
3204 : /*
3205 : * Change a given task's CPU affinity to the intersection of its current
3206 : * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3207 : * If user_cpus_ptr is defined, use it as the basis for restricting CPU
3208 : * affinity or use cpu_online_mask instead.
3209 : *
3210 : * If the resulting mask is empty, leave the affinity unchanged and return
3211 : * -EINVAL.
3212 : */
3213 : static int restrict_cpus_allowed_ptr(struct task_struct *p,
3214 : struct cpumask *new_mask,
3215 : const struct cpumask *subset_mask)
3216 : {
3217 : struct affinity_context ac = {
3218 : .new_mask = new_mask,
3219 : .flags = 0,
3220 : };
3221 : struct rq_flags rf;
3222 : struct rq *rq;
3223 : int err;
3224 :
3225 : rq = task_rq_lock(p, &rf);
3226 :
3227 : /*
3228 : * Forcefully restricting the affinity of a deadline task is
3229 : * likely to cause problems, so fail and noisily override the
3230 : * mask entirely.
3231 : */
3232 : if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
3233 : err = -EPERM;
3234 : goto err_unlock;
3235 : }
3236 :
3237 : if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
3238 : err = -EINVAL;
3239 : goto err_unlock;
3240 : }
3241 :
3242 : return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
3243 :
3244 : err_unlock:
3245 : task_rq_unlock(rq, p, &rf);
3246 : return err;
3247 : }
3248 :
3249 : /*
3250 : * Restrict the CPU affinity of task @p so that it is a subset of
3251 : * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
3252 : * old affinity mask. If the resulting mask is empty, we warn and walk
3253 : * up the cpuset hierarchy until we find a suitable mask.
3254 : */
3255 : void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3256 : {
3257 : cpumask_var_t new_mask;
3258 : const struct cpumask *override_mask = task_cpu_possible_mask(p);
3259 :
3260 : alloc_cpumask_var(&new_mask, GFP_KERNEL);
3261 :
3262 : /*
3263 : * __migrate_task() can fail silently in the face of concurrent
3264 : * offlining of the chosen destination CPU, so take the hotplug
3265 : * lock to ensure that the migration succeeds.
3266 : */
3267 : cpus_read_lock();
3268 : if (!cpumask_available(new_mask))
3269 : goto out_set_mask;
3270 :
3271 : if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3272 : goto out_free_mask;
3273 :
3274 : /*
3275 : * We failed to find a valid subset of the affinity mask for the
3276 : * task, so override it based on its cpuset hierarchy.
3277 : */
3278 : cpuset_cpus_allowed(p, new_mask);
3279 : override_mask = new_mask;
3280 :
3281 : out_set_mask:
3282 : if (printk_ratelimit()) {
3283 : printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3284 : task_pid_nr(p), p->comm,
3285 : cpumask_pr_args(override_mask));
3286 : }
3287 :
3288 : WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3289 : out_free_mask:
3290 : cpus_read_unlock();
3291 : free_cpumask_var(new_mask);
3292 : }
3293 :
3294 : static int
3295 : __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
3296 :
3297 : /*
3298 : * Restore the affinity of a task @p which was previously restricted by a
3299 : * call to force_compatible_cpus_allowed_ptr().
3300 : *
3301 : * It is the caller's responsibility to serialise this with any calls to
3302 : * force_compatible_cpus_allowed_ptr(@p).
3303 : */
3304 : void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3305 : {
3306 : struct affinity_context ac = {
3307 : .new_mask = task_user_cpus(p),
3308 : .flags = 0,
3309 : };
3310 : int ret;
3311 :
3312 : /*
3313 : * Try to restore the old affinity mask with __sched_setaffinity().
3314 : * Cpuset masking will be done there too.
3315 : */
3316 : ret = __sched_setaffinity(p, &ac);
3317 : WARN_ON_ONCE(ret);
3318 : }
3319 :
3320 : void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3321 : {
3322 : #ifdef CONFIG_SCHED_DEBUG
3323 : unsigned int state = READ_ONCE(p->__state);
3324 :
3325 : /*
3326 : * We should never call set_task_cpu() on a blocked task,
3327 : * ttwu() will sort out the placement.
3328 : */
3329 : WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3330 :
3331 : /*
3332 : * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3333 : * because schedstat_wait_{start,end} rebase migrating task's wait_start
3334 : * time relying on p->on_rq.
3335 : */
3336 : WARN_ON_ONCE(state == TASK_RUNNING &&
3337 : p->sched_class == &fair_sched_class &&
3338 : (p->on_rq && !task_on_rq_migrating(p)));
3339 :
3340 : #ifdef CONFIG_LOCKDEP
3341 : /*
3342 : * The caller should hold either p->pi_lock or rq->lock, when changing
3343 : * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3344 : *
3345 : * sched_move_task() holds both and thus holding either pins the cgroup,
3346 : * see task_group().
3347 : *
3348 : * Furthermore, all task_rq users should acquire both locks, see
3349 : * task_rq_lock().
3350 : */
3351 : WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3352 : lockdep_is_held(__rq_lockp(task_rq(p)))));
3353 : #endif
3354 : /*
3355 : * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3356 : */
3357 : WARN_ON_ONCE(!cpu_online(new_cpu));
3358 :
3359 : WARN_ON_ONCE(is_migration_disabled(p));
3360 : #endif
3361 :
3362 : trace_sched_migrate_task(p, new_cpu);
3363 :
3364 : if (task_cpu(p) != new_cpu) {
3365 : if (p->sched_class->migrate_task_rq)
3366 : p->sched_class->migrate_task_rq(p, new_cpu);
3367 : p->se.nr_migrations++;
3368 : rseq_migrate(p);
3369 : sched_mm_cid_migrate_from(p);
3370 : perf_event_task_migrate(p);
3371 : }
3372 :
3373 : __set_task_cpu(p, new_cpu);
3374 : }
3375 :
3376 : #ifdef CONFIG_NUMA_BALANCING
3377 : static void __migrate_swap_task(struct task_struct *p, int cpu)
3378 : {
3379 : if (task_on_rq_queued(p)) {
3380 : struct rq *src_rq, *dst_rq;
3381 : struct rq_flags srf, drf;
3382 :
3383 : src_rq = task_rq(p);
3384 : dst_rq = cpu_rq(cpu);
3385 :
3386 : rq_pin_lock(src_rq, &srf);
3387 : rq_pin_lock(dst_rq, &drf);
3388 :
3389 : deactivate_task(src_rq, p, 0);
3390 : set_task_cpu(p, cpu);
3391 : activate_task(dst_rq, p, 0);
3392 : check_preempt_curr(dst_rq, p, 0);
3393 :
3394 : rq_unpin_lock(dst_rq, &drf);
3395 : rq_unpin_lock(src_rq, &srf);
3396 :
3397 : } else {
3398 : /*
3399 : * Task isn't running anymore; make it appear like we migrated
3400 : * it before it went to sleep. This means on wakeup we make the
3401 : * previous CPU our target instead of where it really is.
3402 : */
3403 : p->wake_cpu = cpu;
3404 : }
3405 : }
3406 :
3407 : struct migration_swap_arg {
3408 : struct task_struct *src_task, *dst_task;
3409 : int src_cpu, dst_cpu;
3410 : };
3411 :
3412 : static int migrate_swap_stop(void *data)
3413 : {
3414 : struct migration_swap_arg *arg = data;
3415 : struct rq *src_rq, *dst_rq;
3416 : int ret = -EAGAIN;
3417 :
3418 : if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3419 : return -EAGAIN;
3420 :
3421 : src_rq = cpu_rq(arg->src_cpu);
3422 : dst_rq = cpu_rq(arg->dst_cpu);
3423 :
3424 : double_raw_lock(&arg->src_task->pi_lock,
3425 : &arg->dst_task->pi_lock);
3426 : double_rq_lock(src_rq, dst_rq);
3427 :
3428 : if (task_cpu(arg->dst_task) != arg->dst_cpu)
3429 : goto unlock;
3430 :
3431 : if (task_cpu(arg->src_task) != arg->src_cpu)
3432 : goto unlock;
3433 :
3434 : if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3435 : goto unlock;
3436 :
3437 : if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3438 : goto unlock;
3439 :
3440 : __migrate_swap_task(arg->src_task, arg->dst_cpu);
3441 : __migrate_swap_task(arg->dst_task, arg->src_cpu);
3442 :
3443 : ret = 0;
3444 :
3445 : unlock:
3446 : double_rq_unlock(src_rq, dst_rq);
3447 : raw_spin_unlock(&arg->dst_task->pi_lock);
3448 : raw_spin_unlock(&arg->src_task->pi_lock);
3449 :
3450 : return ret;
3451 : }
3452 :
3453 : /*
3454 : * Cross migrate two tasks
3455 : */
3456 : int migrate_swap(struct task_struct *cur, struct task_struct *p,
3457 : int target_cpu, int curr_cpu)
3458 : {
3459 : struct migration_swap_arg arg;
3460 : int ret = -EINVAL;
3461 :
3462 : arg = (struct migration_swap_arg){
3463 : .src_task = cur,
3464 : .src_cpu = curr_cpu,
3465 : .dst_task = p,
3466 : .dst_cpu = target_cpu,
3467 : };
3468 :
3469 : if (arg.src_cpu == arg.dst_cpu)
3470 : goto out;
3471 :
3472 : /*
3473 : * These three tests are all lockless; this is OK since all of them
3474 : * will be re-checked with proper locks held further down the line.
3475 : */
3476 : if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3477 : goto out;
3478 :
3479 : if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3480 : goto out;
3481 :
3482 : if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3483 : goto out;
3484 :
3485 : trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3486 : ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3487 :
3488 : out:
3489 : return ret;
3490 : }
3491 : #endif /* CONFIG_NUMA_BALANCING */
3492 :
3493 : /***
3494 : * kick_process - kick a running thread to enter/exit the kernel
3495 : * @p: the to-be-kicked thread
3496 : *
3497 : * Cause a process which is running on another CPU to enter
3498 : * kernel-mode, without any delay. (to get signals handled.)
3499 : *
3500 : * NOTE: this function doesn't have to take the runqueue lock,
3501 : * because all it wants to ensure is that the remote task enters
3502 : * the kernel. If the IPI races and the task has been migrated
3503 : * to another CPU then no harm is done and the purpose has been
3504 : * achieved as well.
3505 : */
3506 : void kick_process(struct task_struct *p)
3507 : {
3508 : int cpu;
3509 :
3510 : preempt_disable();
3511 : cpu = task_cpu(p);
3512 : if ((cpu != smp_processor_id()) && task_curr(p))
3513 : smp_send_reschedule(cpu);
3514 : preempt_enable();
3515 : }
3516 : EXPORT_SYMBOL_GPL(kick_process);
3517 :
3518 : /*
3519 : * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3520 : *
3521 : * A few notes on cpu_active vs cpu_online:
3522 : *
3523 : * - cpu_active must be a subset of cpu_online
3524 : *
3525 : * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3526 : * see __set_cpus_allowed_ptr(). At this point the newly online
3527 : * CPU isn't yet part of the sched domains, and balancing will not
3528 : * see it.
3529 : *
3530 : * - on CPU-down we clear cpu_active() to mask the sched domains and
3531 : * avoid the load balancer to place new tasks on the to be removed
3532 : * CPU. Existing tasks will remain running there and will be taken
3533 : * off.
3534 : *
3535 : * This means that fallback selection must not select !active CPUs.
3536 : * And can assume that any active CPU must be online. Conversely
3537 : * select_task_rq() below may allow selection of !active CPUs in order
3538 : * to satisfy the above rules.
3539 : */
3540 : static int select_fallback_rq(int cpu, struct task_struct *p)
3541 : {
3542 : int nid = cpu_to_node(cpu);
3543 : const struct cpumask *nodemask = NULL;
3544 : enum { cpuset, possible, fail } state = cpuset;
3545 : int dest_cpu;
3546 :
3547 : /*
3548 : * If the node that the CPU is on has been offlined, cpu_to_node()
3549 : * will return -1. There is no CPU on the node, and we should
3550 : * select the CPU on the other node.
3551 : */
3552 : if (nid != -1) {
3553 : nodemask = cpumask_of_node(nid);
3554 :
3555 : /* Look for allowed, online CPU in same node. */
3556 : for_each_cpu(dest_cpu, nodemask) {
3557 : if (is_cpu_allowed(p, dest_cpu))
3558 : return dest_cpu;
3559 : }
3560 : }
3561 :
3562 : for (;;) {
3563 : /* Any allowed, online CPU? */
3564 : for_each_cpu(dest_cpu, p->cpus_ptr) {
3565 : if (!is_cpu_allowed(p, dest_cpu))
3566 : continue;
3567 :
3568 : goto out;
3569 : }
3570 :
3571 : /* No more Mr. Nice Guy. */
3572 : switch (state) {
3573 : case cpuset:
3574 : if (cpuset_cpus_allowed_fallback(p)) {
3575 : state = possible;
3576 : break;
3577 : }
3578 : fallthrough;
3579 : case possible:
3580 : /*
3581 : * XXX When called from select_task_rq() we only
3582 : * hold p->pi_lock and again violate locking order.
3583 : *
3584 : * More yuck to audit.
3585 : */
3586 : do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3587 : state = fail;
3588 : break;
3589 : case fail:
3590 : BUG();
3591 : break;
3592 : }
3593 : }
3594 :
3595 : out:
3596 : if (state != cpuset) {
3597 : /*
3598 : * Don't tell them about moving exiting tasks or
3599 : * kernel threads (both mm NULL), since they never
3600 : * leave kernel.
3601 : */
3602 : if (p->mm && printk_ratelimit()) {
3603 : printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3604 : task_pid_nr(p), p->comm, cpu);
3605 : }
3606 : }
3607 :
3608 : return dest_cpu;
3609 : }
3610 :
3611 : /*
3612 : * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3613 : */
3614 : static inline
3615 : int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3616 : {
3617 : lockdep_assert_held(&p->pi_lock);
3618 :
3619 : if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3620 : cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3621 : else
3622 : cpu = cpumask_any(p->cpus_ptr);
3623 :
3624 : /*
3625 : * In order not to call set_task_cpu() on a blocking task we need
3626 : * to rely on ttwu() to place the task on a valid ->cpus_ptr
3627 : * CPU.
3628 : *
3629 : * Since this is common to all placement strategies, this lives here.
3630 : *
3631 : * [ this allows ->select_task() to simply return task_cpu(p) and
3632 : * not worry about this generic constraint ]
3633 : */
3634 : if (unlikely(!is_cpu_allowed(p, cpu)))
3635 : cpu = select_fallback_rq(task_cpu(p), p);
3636 :
3637 : return cpu;
3638 : }
3639 :
3640 : void sched_set_stop_task(int cpu, struct task_struct *stop)
3641 : {
3642 : static struct lock_class_key stop_pi_lock;
3643 : struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3644 : struct task_struct *old_stop = cpu_rq(cpu)->stop;
3645 :
3646 : if (stop) {
3647 : /*
3648 : * Make it appear like a SCHED_FIFO task, its something
3649 : * userspace knows about and won't get confused about.
3650 : *
3651 : * Also, it will make PI more or less work without too
3652 : * much confusion -- but then, stop work should not
3653 : * rely on PI working anyway.
3654 : */
3655 : sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3656 :
3657 : stop->sched_class = &stop_sched_class;
3658 :
3659 : /*
3660 : * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3661 : * adjust the effective priority of a task. As a result,
3662 : * rt_mutex_setprio() can trigger (RT) balancing operations,
3663 : * which can then trigger wakeups of the stop thread to push
3664 : * around the current task.
3665 : *
3666 : * The stop task itself will never be part of the PI-chain, it
3667 : * never blocks, therefore that ->pi_lock recursion is safe.
3668 : * Tell lockdep about this by placing the stop->pi_lock in its
3669 : * own class.
3670 : */
3671 : lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3672 : }
3673 :
3674 : cpu_rq(cpu)->stop = stop;
3675 :
3676 : if (old_stop) {
3677 : /*
3678 : * Reset it back to a normal scheduling class so that
3679 : * it can die in pieces.
3680 : */
3681 : old_stop->sched_class = &rt_sched_class;
3682 : }
3683 : }
3684 :
3685 : #else /* CONFIG_SMP */
3686 :
3687 : static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3688 : struct affinity_context *ctx)
3689 : {
3690 0 : return set_cpus_allowed_ptr(p, ctx->new_mask);
3691 : }
3692 :
3693 : static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3694 :
3695 : static inline bool rq_has_pinned_tasks(struct rq *rq)
3696 : {
3697 : return false;
3698 : }
3699 :
3700 : static inline cpumask_t *alloc_user_cpus_ptr(int node)
3701 : {
3702 : return NULL;
3703 : }
3704 :
3705 : #endif /* !CONFIG_SMP */
3706 :
3707 : static void
3708 : ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3709 : {
3710 : struct rq *rq;
3711 :
3712 : if (!schedstat_enabled())
3713 : return;
3714 :
3715 : rq = this_rq();
3716 :
3717 : #ifdef CONFIG_SMP
3718 : if (cpu == rq->cpu) {
3719 : __schedstat_inc(rq->ttwu_local);
3720 : __schedstat_inc(p->stats.nr_wakeups_local);
3721 : } else {
3722 : struct sched_domain *sd;
3723 :
3724 : __schedstat_inc(p->stats.nr_wakeups_remote);
3725 : rcu_read_lock();
3726 : for_each_domain(rq->cpu, sd) {
3727 : if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3728 : __schedstat_inc(sd->ttwu_wake_remote);
3729 : break;
3730 : }
3731 : }
3732 : rcu_read_unlock();
3733 : }
3734 :
3735 : if (wake_flags & WF_MIGRATED)
3736 : __schedstat_inc(p->stats.nr_wakeups_migrate);
3737 : #endif /* CONFIG_SMP */
3738 :
3739 : __schedstat_inc(rq->ttwu_count);
3740 : __schedstat_inc(p->stats.nr_wakeups);
3741 :
3742 : if (wake_flags & WF_SYNC)
3743 : __schedstat_inc(p->stats.nr_wakeups_sync);
3744 : }
3745 :
3746 : /*
3747 : * Mark the task runnable.
3748 : */
3749 : static inline void ttwu_do_wakeup(struct task_struct *p)
3750 : {
3751 856 : WRITE_ONCE(p->__state, TASK_RUNNING);
3752 856 : trace_sched_wakeup(p);
3753 : }
3754 :
3755 : static void
3756 856 : ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3757 : struct rq_flags *rf)
3758 : {
3759 856 : int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3760 :
3761 856 : lockdep_assert_rq_held(rq);
3762 :
3763 856 : if (p->sched_contributes_to_load)
3764 510 : rq->nr_uninterruptible--;
3765 :
3766 : #ifdef CONFIG_SMP
3767 : if (wake_flags & WF_MIGRATED)
3768 : en_flags |= ENQUEUE_MIGRATED;
3769 : else
3770 : #endif
3771 856 : if (p->in_iowait) {
3772 0 : delayacct_blkio_end(p);
3773 0 : atomic_dec(&task_rq(p)->nr_iowait);
3774 : }
3775 :
3776 856 : activate_task(rq, p, en_flags);
3777 856 : check_preempt_curr(rq, p, wake_flags);
3778 :
3779 856 : ttwu_do_wakeup(p);
3780 :
3781 : #ifdef CONFIG_SMP
3782 : if (p->sched_class->task_woken) {
3783 : /*
3784 : * Our task @p is fully woken up and running; so it's safe to
3785 : * drop the rq->lock, hereafter rq is only used for statistics.
3786 : */
3787 : rq_unpin_lock(rq, rf);
3788 : p->sched_class->task_woken(rq, p);
3789 : rq_repin_lock(rq, rf);
3790 : }
3791 :
3792 : if (rq->idle_stamp) {
3793 : u64 delta = rq_clock(rq) - rq->idle_stamp;
3794 : u64 max = 2*rq->max_idle_balance_cost;
3795 :
3796 : update_avg(&rq->avg_idle, delta);
3797 :
3798 : if (rq->avg_idle > max)
3799 : rq->avg_idle = max;
3800 :
3801 : rq->wake_stamp = jiffies;
3802 : rq->wake_avg_idle = rq->avg_idle / 2;
3803 :
3804 : rq->idle_stamp = 0;
3805 : }
3806 : #endif
3807 856 : }
3808 :
3809 : /*
3810 : * Consider @p being inside a wait loop:
3811 : *
3812 : * for (;;) {
3813 : * set_current_state(TASK_UNINTERRUPTIBLE);
3814 : *
3815 : * if (CONDITION)
3816 : * break;
3817 : *
3818 : * schedule();
3819 : * }
3820 : * __set_current_state(TASK_RUNNING);
3821 : *
3822 : * between set_current_state() and schedule(). In this case @p is still
3823 : * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3824 : * an atomic manner.
3825 : *
3826 : * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3827 : * then schedule() must still happen and p->state can be changed to
3828 : * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3829 : * need to do a full wakeup with enqueue.
3830 : *
3831 : * Returns: %true when the wakeup is done,
3832 : * %false otherwise.
3833 : */
3834 0 : static int ttwu_runnable(struct task_struct *p, int wake_flags)
3835 : {
3836 : struct rq_flags rf;
3837 : struct rq *rq;
3838 0 : int ret = 0;
3839 :
3840 0 : rq = __task_rq_lock(p, &rf);
3841 0 : if (task_on_rq_queued(p)) {
3842 0 : if (!task_on_cpu(rq, p)) {
3843 : /*
3844 : * When on_rq && !on_cpu the task is preempted, see if
3845 : * it should preempt the task that is current now.
3846 : */
3847 0 : update_rq_clock(rq);
3848 0 : check_preempt_curr(rq, p, wake_flags);
3849 : }
3850 0 : ttwu_do_wakeup(p);
3851 0 : ret = 1;
3852 : }
3853 0 : __task_rq_unlock(rq, &rf);
3854 :
3855 0 : return ret;
3856 : }
3857 :
3858 : #ifdef CONFIG_SMP
3859 : void sched_ttwu_pending(void *arg)
3860 : {
3861 : struct llist_node *llist = arg;
3862 : struct rq *rq = this_rq();
3863 : struct task_struct *p, *t;
3864 : struct rq_flags rf;
3865 :
3866 : if (!llist)
3867 : return;
3868 :
3869 : rq_lock_irqsave(rq, &rf);
3870 : update_rq_clock(rq);
3871 :
3872 : llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3873 : if (WARN_ON_ONCE(p->on_cpu))
3874 : smp_cond_load_acquire(&p->on_cpu, !VAL);
3875 :
3876 : if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3877 : set_task_cpu(p, cpu_of(rq));
3878 :
3879 : ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3880 : }
3881 :
3882 : /*
3883 : * Must be after enqueueing at least once task such that
3884 : * idle_cpu() does not observe a false-negative -- if it does,
3885 : * it is possible for select_idle_siblings() to stack a number
3886 : * of tasks on this CPU during that window.
3887 : *
3888 : * It is ok to clear ttwu_pending when another task pending.
3889 : * We will receive IPI after local irq enabled and then enqueue it.
3890 : * Since now nr_running > 0, idle_cpu() will always get correct result.
3891 : */
3892 : WRITE_ONCE(rq->ttwu_pending, 0);
3893 : rq_unlock_irqrestore(rq, &rf);
3894 : }
3895 :
3896 : /*
3897 : * Prepare the scene for sending an IPI for a remote smp_call
3898 : *
3899 : * Returns true if the caller can proceed with sending the IPI.
3900 : * Returns false otherwise.
3901 : */
3902 : bool call_function_single_prep_ipi(int cpu)
3903 : {
3904 : if (set_nr_if_polling(cpu_rq(cpu)->idle)) {
3905 : trace_sched_wake_idle_without_ipi(cpu);
3906 : return false;
3907 : }
3908 :
3909 : return true;
3910 : }
3911 :
3912 : /*
3913 : * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3914 : * necessary. The wakee CPU on receipt of the IPI will queue the task
3915 : * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3916 : * of the wakeup instead of the waker.
3917 : */
3918 : static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3919 : {
3920 : struct rq *rq = cpu_rq(cpu);
3921 :
3922 : p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3923 :
3924 : WRITE_ONCE(rq->ttwu_pending, 1);
3925 : __smp_call_single_queue(cpu, &p->wake_entry.llist);
3926 : }
3927 :
3928 : void wake_up_if_idle(int cpu)
3929 : {
3930 : struct rq *rq = cpu_rq(cpu);
3931 : struct rq_flags rf;
3932 :
3933 : rcu_read_lock();
3934 :
3935 : if (!is_idle_task(rcu_dereference(rq->curr)))
3936 : goto out;
3937 :
3938 : rq_lock_irqsave(rq, &rf);
3939 : if (is_idle_task(rq->curr))
3940 : resched_curr(rq);
3941 : /* Else CPU is not idle, do nothing here: */
3942 : rq_unlock_irqrestore(rq, &rf);
3943 :
3944 : out:
3945 : rcu_read_unlock();
3946 : }
3947 :
3948 : bool cpus_share_cache(int this_cpu, int that_cpu)
3949 : {
3950 : if (this_cpu == that_cpu)
3951 : return true;
3952 :
3953 : return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3954 : }
3955 :
3956 : static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3957 : {
3958 : /*
3959 : * Do not complicate things with the async wake_list while the CPU is
3960 : * in hotplug state.
3961 : */
3962 : if (!cpu_active(cpu))
3963 : return false;
3964 :
3965 : /* Ensure the task will still be allowed to run on the CPU. */
3966 : if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3967 : return false;
3968 :
3969 : /*
3970 : * If the CPU does not share cache, then queue the task on the
3971 : * remote rqs wakelist to avoid accessing remote data.
3972 : */
3973 : if (!cpus_share_cache(smp_processor_id(), cpu))
3974 : return true;
3975 :
3976 : if (cpu == smp_processor_id())
3977 : return false;
3978 :
3979 : /*
3980 : * If the wakee cpu is idle, or the task is descheduling and the
3981 : * only running task on the CPU, then use the wakelist to offload
3982 : * the task activation to the idle (or soon-to-be-idle) CPU as
3983 : * the current CPU is likely busy. nr_running is checked to
3984 : * avoid unnecessary task stacking.
3985 : *
3986 : * Note that we can only get here with (wakee) p->on_rq=0,
3987 : * p->on_cpu can be whatever, we've done the dequeue, so
3988 : * the wakee has been accounted out of ->nr_running.
3989 : */
3990 : if (!cpu_rq(cpu)->nr_running)
3991 : return true;
3992 :
3993 : return false;
3994 : }
3995 :
3996 : static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3997 : {
3998 : if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
3999 : sched_clock_cpu(cpu); /* Sync clocks across CPUs */
4000 : __ttwu_queue_wakelist(p, cpu, wake_flags);
4001 : return true;
4002 : }
4003 :
4004 : return false;
4005 : }
4006 :
4007 : #else /* !CONFIG_SMP */
4008 :
4009 : static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4010 : {
4011 : return false;
4012 : }
4013 :
4014 : #endif /* CONFIG_SMP */
4015 :
4016 : static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
4017 : {
4018 856 : struct rq *rq = cpu_rq(cpu);
4019 : struct rq_flags rf;
4020 :
4021 856 : if (ttwu_queue_wakelist(p, cpu, wake_flags))
4022 : return;
4023 :
4024 856 : rq_lock(rq, &rf);
4025 856 : update_rq_clock(rq);
4026 856 : ttwu_do_activate(rq, p, wake_flags, &rf);
4027 856 : rq_unlock(rq, &rf);
4028 : }
4029 :
4030 : /*
4031 : * Invoked from try_to_wake_up() to check whether the task can be woken up.
4032 : *
4033 : * The caller holds p::pi_lock if p != current or has preemption
4034 : * disabled when p == current.
4035 : *
4036 : * The rules of PREEMPT_RT saved_state:
4037 : *
4038 : * The related locking code always holds p::pi_lock when updating
4039 : * p::saved_state, which means the code is fully serialized in both cases.
4040 : *
4041 : * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
4042 : * bits set. This allows to distinguish all wakeup scenarios.
4043 : */
4044 : static __always_inline
4045 : bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
4046 : {
4047 : int match;
4048 :
4049 : if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
4050 : WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
4051 : state != TASK_RTLOCK_WAIT);
4052 : }
4053 :
4054 1030 : *success = !!(match = __task_state_match(p, state));
4055 :
4056 : #ifdef CONFIG_PREEMPT_RT
4057 : /*
4058 : * Saved state preserves the task state across blocking on
4059 : * an RT lock. If the state matches, set p::saved_state to
4060 : * TASK_RUNNING, but do not wake the task because it waits
4061 : * for a lock wakeup. Also indicate success because from
4062 : * the regular waker's point of view this has succeeded.
4063 : *
4064 : * After acquiring the lock the task will restore p::__state
4065 : * from p::saved_state which ensures that the regular
4066 : * wakeup is not lost. The restore will also set
4067 : * p::saved_state to TASK_RUNNING so any further tests will
4068 : * not result in false positives vs. @success
4069 : */
4070 : if (match < 0)
4071 : p->saved_state = TASK_RUNNING;
4072 : #endif
4073 : return match > 0;
4074 : }
4075 :
4076 : /*
4077 : * Notes on Program-Order guarantees on SMP systems.
4078 : *
4079 : * MIGRATION
4080 : *
4081 : * The basic program-order guarantee on SMP systems is that when a task [t]
4082 : * migrates, all its activity on its old CPU [c0] happens-before any subsequent
4083 : * execution on its new CPU [c1].
4084 : *
4085 : * For migration (of runnable tasks) this is provided by the following means:
4086 : *
4087 : * A) UNLOCK of the rq(c0)->lock scheduling out task t
4088 : * B) migration for t is required to synchronize *both* rq(c0)->lock and
4089 : * rq(c1)->lock (if not at the same time, then in that order).
4090 : * C) LOCK of the rq(c1)->lock scheduling in task
4091 : *
4092 : * Release/acquire chaining guarantees that B happens after A and C after B.
4093 : * Note: the CPU doing B need not be c0 or c1
4094 : *
4095 : * Example:
4096 : *
4097 : * CPU0 CPU1 CPU2
4098 : *
4099 : * LOCK rq(0)->lock
4100 : * sched-out X
4101 : * sched-in Y
4102 : * UNLOCK rq(0)->lock
4103 : *
4104 : * LOCK rq(0)->lock // orders against CPU0
4105 : * dequeue X
4106 : * UNLOCK rq(0)->lock
4107 : *
4108 : * LOCK rq(1)->lock
4109 : * enqueue X
4110 : * UNLOCK rq(1)->lock
4111 : *
4112 : * LOCK rq(1)->lock // orders against CPU2
4113 : * sched-out Z
4114 : * sched-in X
4115 : * UNLOCK rq(1)->lock
4116 : *
4117 : *
4118 : * BLOCKING -- aka. SLEEP + WAKEUP
4119 : *
4120 : * For blocking we (obviously) need to provide the same guarantee as for
4121 : * migration. However the means are completely different as there is no lock
4122 : * chain to provide order. Instead we do:
4123 : *
4124 : * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
4125 : * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
4126 : *
4127 : * Example:
4128 : *
4129 : * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
4130 : *
4131 : * LOCK rq(0)->lock LOCK X->pi_lock
4132 : * dequeue X
4133 : * sched-out X
4134 : * smp_store_release(X->on_cpu, 0);
4135 : *
4136 : * smp_cond_load_acquire(&X->on_cpu, !VAL);
4137 : * X->state = WAKING
4138 : * set_task_cpu(X,2)
4139 : *
4140 : * LOCK rq(2)->lock
4141 : * enqueue X
4142 : * X->state = RUNNING
4143 : * UNLOCK rq(2)->lock
4144 : *
4145 : * LOCK rq(2)->lock // orders against CPU1
4146 : * sched-out Z
4147 : * sched-in X
4148 : * UNLOCK rq(2)->lock
4149 : *
4150 : * UNLOCK X->pi_lock
4151 : * UNLOCK rq(0)->lock
4152 : *
4153 : *
4154 : * However, for wakeups there is a second guarantee we must provide, namely we
4155 : * must ensure that CONDITION=1 done by the caller can not be reordered with
4156 : * accesses to the task state; see try_to_wake_up() and set_current_state().
4157 : */
4158 :
4159 : /**
4160 : * try_to_wake_up - wake up a thread
4161 : * @p: the thread to be awakened
4162 : * @state: the mask of task states that can be woken
4163 : * @wake_flags: wake modifier flags (WF_*)
4164 : *
4165 : * Conceptually does:
4166 : *
4167 : * If (@state & @p->state) @p->state = TASK_RUNNING.
4168 : *
4169 : * If the task was not queued/runnable, also place it back on a runqueue.
4170 : *
4171 : * This function is atomic against schedule() which would dequeue the task.
4172 : *
4173 : * It issues a full memory barrier before accessing @p->state, see the comment
4174 : * with set_current_state().
4175 : *
4176 : * Uses p->pi_lock to serialize against concurrent wake-ups.
4177 : *
4178 : * Relies on p->pi_lock stabilizing:
4179 : * - p->sched_class
4180 : * - p->cpus_ptr
4181 : * - p->sched_task_group
4182 : * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4183 : *
4184 : * Tries really hard to only take one task_rq(p)->lock for performance.
4185 : * Takes rq->lock in:
4186 : * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4187 : * - ttwu_queue() -- new rq, for enqueue of the task;
4188 : * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4189 : *
4190 : * As a consequence we race really badly with just about everything. See the
4191 : * many memory barriers and their comments for details.
4192 : *
4193 : * Return: %true if @p->state changes (an actual wakeup was done),
4194 : * %false otherwise.
4195 : */
4196 : static int
4197 1030 : try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4198 : {
4199 : unsigned long flags;
4200 1030 : int cpu, success = 0;
4201 :
4202 1030 : preempt_disable();
4203 1030 : if (p == current) {
4204 : /*
4205 : * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4206 : * == smp_processor_id()'. Together this means we can special
4207 : * case the whole 'p->on_rq && ttwu_runnable()' case below
4208 : * without taking any locks.
4209 : *
4210 : * In particular:
4211 : * - we rely on Program-Order guarantees for all the ordering,
4212 : * - we're serialized against set_special_state() by virtue of
4213 : * it disabling IRQs (this allows not taking ->pi_lock).
4214 : */
4215 11 : if (!ttwu_state_match(p, state, &success))
4216 : goto out;
4217 :
4218 0 : trace_sched_waking(p);
4219 : ttwu_do_wakeup(p);
4220 : goto out;
4221 : }
4222 :
4223 : /*
4224 : * If we are going to wake up a thread waiting for CONDITION we
4225 : * need to ensure that CONDITION=1 done by the caller can not be
4226 : * reordered with p->state check below. This pairs with smp_store_mb()
4227 : * in set_current_state() that the waiting thread does.
4228 : */
4229 1019 : raw_spin_lock_irqsave(&p->pi_lock, flags);
4230 : smp_mb__after_spinlock();
4231 1019 : if (!ttwu_state_match(p, state, &success))
4232 : goto unlock;
4233 :
4234 856 : trace_sched_waking(p);
4235 :
4236 : /*
4237 : * Ensure we load p->on_rq _after_ p->state, otherwise it would
4238 : * be possible to, falsely, observe p->on_rq == 0 and get stuck
4239 : * in smp_cond_load_acquire() below.
4240 : *
4241 : * sched_ttwu_pending() try_to_wake_up()
4242 : * STORE p->on_rq = 1 LOAD p->state
4243 : * UNLOCK rq->lock
4244 : *
4245 : * __schedule() (switch to task 'p')
4246 : * LOCK rq->lock smp_rmb();
4247 : * smp_mb__after_spinlock();
4248 : * UNLOCK rq->lock
4249 : *
4250 : * [task p]
4251 : * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4252 : *
4253 : * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4254 : * __schedule(). See the comment for smp_mb__after_spinlock().
4255 : *
4256 : * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4257 : */
4258 856 : smp_rmb();
4259 856 : if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4260 : goto unlock;
4261 :
4262 : #ifdef CONFIG_SMP
4263 : /*
4264 : * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4265 : * possible to, falsely, observe p->on_cpu == 0.
4266 : *
4267 : * One must be running (->on_cpu == 1) in order to remove oneself
4268 : * from the runqueue.
4269 : *
4270 : * __schedule() (switch to task 'p') try_to_wake_up()
4271 : * STORE p->on_cpu = 1 LOAD p->on_rq
4272 : * UNLOCK rq->lock
4273 : *
4274 : * __schedule() (put 'p' to sleep)
4275 : * LOCK rq->lock smp_rmb();
4276 : * smp_mb__after_spinlock();
4277 : * STORE p->on_rq = 0 LOAD p->on_cpu
4278 : *
4279 : * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4280 : * __schedule(). See the comment for smp_mb__after_spinlock().
4281 : *
4282 : * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4283 : * schedule()'s deactivate_task() has 'happened' and p will no longer
4284 : * care about it's own p->state. See the comment in __schedule().
4285 : */
4286 : smp_acquire__after_ctrl_dep();
4287 :
4288 : /*
4289 : * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4290 : * == 0), which means we need to do an enqueue, change p->state to
4291 : * TASK_WAKING such that we can unlock p->pi_lock before doing the
4292 : * enqueue, such as ttwu_queue_wakelist().
4293 : */
4294 : WRITE_ONCE(p->__state, TASK_WAKING);
4295 :
4296 : /*
4297 : * If the owning (remote) CPU is still in the middle of schedule() with
4298 : * this task as prev, considering queueing p on the remote CPUs wake_list
4299 : * which potentially sends an IPI instead of spinning on p->on_cpu to
4300 : * let the waker make forward progress. This is safe because IRQs are
4301 : * disabled and the IPI will deliver after on_cpu is cleared.
4302 : *
4303 : * Ensure we load task_cpu(p) after p->on_cpu:
4304 : *
4305 : * set_task_cpu(p, cpu);
4306 : * STORE p->cpu = @cpu
4307 : * __schedule() (switch to task 'p')
4308 : * LOCK rq->lock
4309 : * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4310 : * STORE p->on_cpu = 1 LOAD p->cpu
4311 : *
4312 : * to ensure we observe the correct CPU on which the task is currently
4313 : * scheduling.
4314 : */
4315 : if (smp_load_acquire(&p->on_cpu) &&
4316 : ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4317 : goto unlock;
4318 :
4319 : /*
4320 : * If the owning (remote) CPU is still in the middle of schedule() with
4321 : * this task as prev, wait until it's done referencing the task.
4322 : *
4323 : * Pairs with the smp_store_release() in finish_task().
4324 : *
4325 : * This ensures that tasks getting woken will be fully ordered against
4326 : * their previous state and preserve Program Order.
4327 : */
4328 : smp_cond_load_acquire(&p->on_cpu, !VAL);
4329 :
4330 : cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4331 : if (task_cpu(p) != cpu) {
4332 : if (p->in_iowait) {
4333 : delayacct_blkio_end(p);
4334 : atomic_dec(&task_rq(p)->nr_iowait);
4335 : }
4336 :
4337 : wake_flags |= WF_MIGRATED;
4338 : psi_ttwu_dequeue(p);
4339 : set_task_cpu(p, cpu);
4340 : }
4341 : #else
4342 856 : cpu = task_cpu(p);
4343 : #endif /* CONFIG_SMP */
4344 :
4345 856 : ttwu_queue(p, cpu, wake_flags);
4346 : unlock:
4347 2038 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4348 : out:
4349 : if (success)
4350 : ttwu_stat(p, task_cpu(p), wake_flags);
4351 1030 : preempt_enable();
4352 :
4353 1030 : return success;
4354 : }
4355 :
4356 : static bool __task_needs_rq_lock(struct task_struct *p)
4357 : {
4358 0 : unsigned int state = READ_ONCE(p->__state);
4359 :
4360 : /*
4361 : * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4362 : * the task is blocked. Make sure to check @state since ttwu() can drop
4363 : * locks at the end, see ttwu_queue_wakelist().
4364 : */
4365 0 : if (state == TASK_RUNNING || state == TASK_WAKING)
4366 : return true;
4367 :
4368 : /*
4369 : * Ensure we load p->on_rq after p->__state, otherwise it would be
4370 : * possible to, falsely, observe p->on_rq == 0.
4371 : *
4372 : * See try_to_wake_up() for a longer comment.
4373 : */
4374 0 : smp_rmb();
4375 0 : if (p->on_rq)
4376 : return true;
4377 :
4378 : #ifdef CONFIG_SMP
4379 : /*
4380 : * Ensure the task has finished __schedule() and will not be referenced
4381 : * anymore. Again, see try_to_wake_up() for a longer comment.
4382 : */
4383 : smp_rmb();
4384 : smp_cond_load_acquire(&p->on_cpu, !VAL);
4385 : #endif
4386 :
4387 : return false;
4388 : }
4389 :
4390 : /**
4391 : * task_call_func - Invoke a function on task in fixed state
4392 : * @p: Process for which the function is to be invoked, can be @current.
4393 : * @func: Function to invoke.
4394 : * @arg: Argument to function.
4395 : *
4396 : * Fix the task in it's current state by avoiding wakeups and or rq operations
4397 : * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4398 : * to work out what the state is, if required. Given that @func can be invoked
4399 : * with a runqueue lock held, it had better be quite lightweight.
4400 : *
4401 : * Returns:
4402 : * Whatever @func returns
4403 : */
4404 0 : int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4405 : {
4406 0 : struct rq *rq = NULL;
4407 : struct rq_flags rf;
4408 : int ret;
4409 :
4410 0 : raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4411 :
4412 0 : if (__task_needs_rq_lock(p))
4413 : rq = __task_rq_lock(p, &rf);
4414 :
4415 : /*
4416 : * At this point the task is pinned; either:
4417 : * - blocked and we're holding off wakeups (pi->lock)
4418 : * - woken, and we're holding off enqueue (rq->lock)
4419 : * - queued, and we're holding off schedule (rq->lock)
4420 : * - running, and we're holding off de-schedule (rq->lock)
4421 : *
4422 : * The called function (@func) can use: task_curr(), p->on_rq and
4423 : * p->__state to differentiate between these states.
4424 : */
4425 0 : ret = func(p, arg);
4426 :
4427 0 : if (rq)
4428 0 : rq_unlock(rq, &rf);
4429 :
4430 0 : raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4431 0 : return ret;
4432 : }
4433 :
4434 : /**
4435 : * cpu_curr_snapshot - Return a snapshot of the currently running task
4436 : * @cpu: The CPU on which to snapshot the task.
4437 : *
4438 : * Returns the task_struct pointer of the task "currently" running on
4439 : * the specified CPU. If the same task is running on that CPU throughout,
4440 : * the return value will be a pointer to that task's task_struct structure.
4441 : * If the CPU did any context switches even vaguely concurrently with the
4442 : * execution of this function, the return value will be a pointer to the
4443 : * task_struct structure of a randomly chosen task that was running on
4444 : * that CPU somewhere around the time that this function was executing.
4445 : *
4446 : * If the specified CPU was offline, the return value is whatever it
4447 : * is, perhaps a pointer to the task_struct structure of that CPU's idle
4448 : * task, but there is no guarantee. Callers wishing a useful return
4449 : * value must take some action to ensure that the specified CPU remains
4450 : * online throughout.
4451 : *
4452 : * This function executes full memory barriers before and after fetching
4453 : * the pointer, which permits the caller to confine this function's fetch
4454 : * with respect to the caller's accesses to other shared variables.
4455 : */
4456 0 : struct task_struct *cpu_curr_snapshot(int cpu)
4457 : {
4458 : struct task_struct *t;
4459 :
4460 0 : smp_mb(); /* Pairing determined by caller's synchronization design. */
4461 0 : t = rcu_dereference(cpu_curr(cpu));
4462 0 : smp_mb(); /* Pairing determined by caller's synchronization design. */
4463 0 : return t;
4464 : }
4465 :
4466 : /**
4467 : * wake_up_process - Wake up a specific process
4468 : * @p: The process to be woken up.
4469 : *
4470 : * Attempt to wake up the nominated process and move it to the set of runnable
4471 : * processes.
4472 : *
4473 : * Return: 1 if the process was woken up, 0 if it was already running.
4474 : *
4475 : * This function executes a full memory barrier before accessing the task state.
4476 : */
4477 1029 : int wake_up_process(struct task_struct *p)
4478 : {
4479 1029 : return try_to_wake_up(p, TASK_NORMAL, 0);
4480 : }
4481 : EXPORT_SYMBOL(wake_up_process);
4482 :
4483 1 : int wake_up_state(struct task_struct *p, unsigned int state)
4484 : {
4485 1 : return try_to_wake_up(p, state, 0);
4486 : }
4487 :
4488 : /*
4489 : * Perform scheduler related setup for a newly forked process p.
4490 : * p is forked by current.
4491 : *
4492 : * __sched_fork() is basic setup used by init_idle() too:
4493 : */
4494 176 : static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4495 : {
4496 176 : p->on_rq = 0;
4497 :
4498 176 : p->se.on_rq = 0;
4499 176 : p->se.exec_start = 0;
4500 176 : p->se.sum_exec_runtime = 0;
4501 176 : p->se.prev_sum_exec_runtime = 0;
4502 176 : p->se.nr_migrations = 0;
4503 176 : p->se.vruntime = 0;
4504 352 : INIT_LIST_HEAD(&p->se.group_node);
4505 :
4506 : #ifdef CONFIG_FAIR_GROUP_SCHED
4507 : p->se.cfs_rq = NULL;
4508 : #endif
4509 :
4510 : #ifdef CONFIG_SCHEDSTATS
4511 : /* Even if schedstat is disabled, there should not be garbage */
4512 : memset(&p->stats, 0, sizeof(p->stats));
4513 : #endif
4514 :
4515 176 : RB_CLEAR_NODE(&p->dl.rb_node);
4516 176 : init_dl_task_timer(&p->dl);
4517 176 : init_dl_inactive_task_timer(&p->dl);
4518 176 : __dl_clear_params(p);
4519 :
4520 352 : INIT_LIST_HEAD(&p->rt.run_list);
4521 176 : p->rt.timeout = 0;
4522 176 : p->rt.time_slice = sched_rr_timeslice;
4523 176 : p->rt.on_rq = 0;
4524 176 : p->rt.on_list = 0;
4525 :
4526 : #ifdef CONFIG_PREEMPT_NOTIFIERS
4527 : INIT_HLIST_HEAD(&p->preempt_notifiers);
4528 : #endif
4529 :
4530 : #ifdef CONFIG_COMPACTION
4531 176 : p->capture_control = NULL;
4532 : #endif
4533 176 : init_numa_balancing(clone_flags, p);
4534 : #ifdef CONFIG_SMP
4535 : p->wake_entry.u_flags = CSD_TYPE_TTWU;
4536 : p->migration_pending = NULL;
4537 : #endif
4538 176 : init_sched_mm_cid(p);
4539 176 : }
4540 :
4541 : DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4542 :
4543 : #ifdef CONFIG_NUMA_BALANCING
4544 :
4545 : int sysctl_numa_balancing_mode;
4546 :
4547 : static void __set_numabalancing_state(bool enabled)
4548 : {
4549 : if (enabled)
4550 : static_branch_enable(&sched_numa_balancing);
4551 : else
4552 : static_branch_disable(&sched_numa_balancing);
4553 : }
4554 :
4555 : void set_numabalancing_state(bool enabled)
4556 : {
4557 : if (enabled)
4558 : sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4559 : else
4560 : sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4561 : __set_numabalancing_state(enabled);
4562 : }
4563 :
4564 : #ifdef CONFIG_PROC_SYSCTL
4565 : static void reset_memory_tiering(void)
4566 : {
4567 : struct pglist_data *pgdat;
4568 :
4569 : for_each_online_pgdat(pgdat) {
4570 : pgdat->nbp_threshold = 0;
4571 : pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
4572 : pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
4573 : }
4574 : }
4575 :
4576 : static int sysctl_numa_balancing(struct ctl_table *table, int write,
4577 : void *buffer, size_t *lenp, loff_t *ppos)
4578 : {
4579 : struct ctl_table t;
4580 : int err;
4581 : int state = sysctl_numa_balancing_mode;
4582 :
4583 : if (write && !capable(CAP_SYS_ADMIN))
4584 : return -EPERM;
4585 :
4586 : t = *table;
4587 : t.data = &state;
4588 : err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4589 : if (err < 0)
4590 : return err;
4591 : if (write) {
4592 : if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4593 : (state & NUMA_BALANCING_MEMORY_TIERING))
4594 : reset_memory_tiering();
4595 : sysctl_numa_balancing_mode = state;
4596 : __set_numabalancing_state(state);
4597 : }
4598 : return err;
4599 : }
4600 : #endif
4601 : #endif
4602 :
4603 : #ifdef CONFIG_SCHEDSTATS
4604 :
4605 : DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4606 :
4607 : static void set_schedstats(bool enabled)
4608 : {
4609 : if (enabled)
4610 : static_branch_enable(&sched_schedstats);
4611 : else
4612 : static_branch_disable(&sched_schedstats);
4613 : }
4614 :
4615 : void force_schedstat_enabled(void)
4616 : {
4617 : if (!schedstat_enabled()) {
4618 : pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4619 : static_branch_enable(&sched_schedstats);
4620 : }
4621 : }
4622 :
4623 : static int __init setup_schedstats(char *str)
4624 : {
4625 : int ret = 0;
4626 : if (!str)
4627 : goto out;
4628 :
4629 : if (!strcmp(str, "enable")) {
4630 : set_schedstats(true);
4631 : ret = 1;
4632 : } else if (!strcmp(str, "disable")) {
4633 : set_schedstats(false);
4634 : ret = 1;
4635 : }
4636 : out:
4637 : if (!ret)
4638 : pr_warn("Unable to parse schedstats=\n");
4639 :
4640 : return ret;
4641 : }
4642 : __setup("schedstats=", setup_schedstats);
4643 :
4644 : #ifdef CONFIG_PROC_SYSCTL
4645 : static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4646 : size_t *lenp, loff_t *ppos)
4647 : {
4648 : struct ctl_table t;
4649 : int err;
4650 : int state = static_branch_likely(&sched_schedstats);
4651 :
4652 : if (write && !capable(CAP_SYS_ADMIN))
4653 : return -EPERM;
4654 :
4655 : t = *table;
4656 : t.data = &state;
4657 : err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4658 : if (err < 0)
4659 : return err;
4660 : if (write)
4661 : set_schedstats(state);
4662 : return err;
4663 : }
4664 : #endif /* CONFIG_PROC_SYSCTL */
4665 : #endif /* CONFIG_SCHEDSTATS */
4666 :
4667 : #ifdef CONFIG_SYSCTL
4668 : static struct ctl_table sched_core_sysctls[] = {
4669 : #ifdef CONFIG_SCHEDSTATS
4670 : {
4671 : .procname = "sched_schedstats",
4672 : .data = NULL,
4673 : .maxlen = sizeof(unsigned int),
4674 : .mode = 0644,
4675 : .proc_handler = sysctl_schedstats,
4676 : .extra1 = SYSCTL_ZERO,
4677 : .extra2 = SYSCTL_ONE,
4678 : },
4679 : #endif /* CONFIG_SCHEDSTATS */
4680 : #ifdef CONFIG_UCLAMP_TASK
4681 : {
4682 : .procname = "sched_util_clamp_min",
4683 : .data = &sysctl_sched_uclamp_util_min,
4684 : .maxlen = sizeof(unsigned int),
4685 : .mode = 0644,
4686 : .proc_handler = sysctl_sched_uclamp_handler,
4687 : },
4688 : {
4689 : .procname = "sched_util_clamp_max",
4690 : .data = &sysctl_sched_uclamp_util_max,
4691 : .maxlen = sizeof(unsigned int),
4692 : .mode = 0644,
4693 : .proc_handler = sysctl_sched_uclamp_handler,
4694 : },
4695 : {
4696 : .procname = "sched_util_clamp_min_rt_default",
4697 : .data = &sysctl_sched_uclamp_util_min_rt_default,
4698 : .maxlen = sizeof(unsigned int),
4699 : .mode = 0644,
4700 : .proc_handler = sysctl_sched_uclamp_handler,
4701 : },
4702 : #endif /* CONFIG_UCLAMP_TASK */
4703 : #ifdef CONFIG_NUMA_BALANCING
4704 : {
4705 : .procname = "numa_balancing",
4706 : .data = NULL, /* filled in by handler */
4707 : .maxlen = sizeof(unsigned int),
4708 : .mode = 0644,
4709 : .proc_handler = sysctl_numa_balancing,
4710 : .extra1 = SYSCTL_ZERO,
4711 : .extra2 = SYSCTL_FOUR,
4712 : },
4713 : #endif /* CONFIG_NUMA_BALANCING */
4714 : {}
4715 : };
4716 1 : static int __init sched_core_sysctl_init(void)
4717 : {
4718 1 : register_sysctl_init("kernel", sched_core_sysctls);
4719 1 : return 0;
4720 : }
4721 : late_initcall(sched_core_sysctl_init);
4722 : #endif /* CONFIG_SYSCTL */
4723 :
4724 : /*
4725 : * fork()/clone()-time setup:
4726 : */
4727 175 : int sched_fork(unsigned long clone_flags, struct task_struct *p)
4728 : {
4729 175 : __sched_fork(clone_flags, p);
4730 : /*
4731 : * We mark the process as NEW here. This guarantees that
4732 : * nobody will actually run it, and a signal or other external
4733 : * event cannot wake it up and insert it on the runqueue either.
4734 : */
4735 175 : p->__state = TASK_NEW;
4736 :
4737 : /*
4738 : * Make sure we do not leak PI boosting priority to the child.
4739 : */
4740 175 : p->prio = current->normal_prio;
4741 :
4742 175 : uclamp_fork(p);
4743 :
4744 : /*
4745 : * Revert to default priority/policy on fork if requested.
4746 : */
4747 175 : if (unlikely(p->sched_reset_on_fork)) {
4748 0 : if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4749 0 : p->policy = SCHED_NORMAL;
4750 0 : p->static_prio = NICE_TO_PRIO(0);
4751 0 : p->rt_priority = 0;
4752 0 : } else if (PRIO_TO_NICE(p->static_prio) < 0)
4753 0 : p->static_prio = NICE_TO_PRIO(0);
4754 :
4755 0 : p->prio = p->normal_prio = p->static_prio;
4756 0 : set_load_weight(p, false);
4757 :
4758 : /*
4759 : * We don't need the reset flag anymore after the fork. It has
4760 : * fulfilled its duty:
4761 : */
4762 0 : p->sched_reset_on_fork = 0;
4763 : }
4764 :
4765 350 : if (dl_prio(p->prio))
4766 : return -EAGAIN;
4767 350 : else if (rt_prio(p->prio))
4768 0 : p->sched_class = &rt_sched_class;
4769 : else
4770 175 : p->sched_class = &fair_sched_class;
4771 :
4772 175 : init_entity_runnable_average(&p->se);
4773 :
4774 :
4775 : #ifdef CONFIG_SCHED_INFO
4776 : if (likely(sched_info_on()))
4777 : memset(&p->sched_info, 0, sizeof(p->sched_info));
4778 : #endif
4779 : #if defined(CONFIG_SMP)
4780 : p->on_cpu = 0;
4781 : #endif
4782 175 : init_task_preempt_count(p);
4783 : #ifdef CONFIG_SMP
4784 : plist_node_init(&p->pushable_tasks, MAX_PRIO);
4785 : RB_CLEAR_NODE(&p->pushable_dl_tasks);
4786 : #endif
4787 175 : return 0;
4788 : }
4789 :
4790 175 : void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4791 : {
4792 : unsigned long flags;
4793 :
4794 : /*
4795 : * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4796 : * required yet, but lockdep gets upset if rules are violated.
4797 : */
4798 175 : raw_spin_lock_irqsave(&p->pi_lock, flags);
4799 : #ifdef CONFIG_CGROUP_SCHED
4800 : if (1) {
4801 : struct task_group *tg;
4802 : tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4803 : struct task_group, css);
4804 : tg = autogroup_task_group(p, tg);
4805 : p->sched_task_group = tg;
4806 : }
4807 : #endif
4808 175 : rseq_migrate(p);
4809 : /*
4810 : * We're setting the CPU for the first time, we don't migrate,
4811 : * so use __set_task_cpu().
4812 : */
4813 175 : __set_task_cpu(p, smp_processor_id());
4814 175 : if (p->sched_class->task_fork)
4815 175 : p->sched_class->task_fork(p);
4816 350 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4817 175 : }
4818 :
4819 175 : void sched_post_fork(struct task_struct *p)
4820 : {
4821 175 : uclamp_post_fork(p);
4822 175 : }
4823 :
4824 3 : unsigned long to_ratio(u64 period, u64 runtime)
4825 : {
4826 3 : if (runtime == RUNTIME_INF)
4827 : return BW_UNIT;
4828 :
4829 : /*
4830 : * Doing this here saves a lot of checks in all
4831 : * the calling paths, and returning zero seems
4832 : * safe for them anyway.
4833 : */
4834 3 : if (period == 0)
4835 : return 0;
4836 :
4837 6 : return div64_u64(runtime << BW_SHIFT, period);
4838 : }
4839 :
4840 : /*
4841 : * wake_up_new_task - wake up a newly created task for the first time.
4842 : *
4843 : * This function will do some initial scheduler statistics housekeeping
4844 : * that must be done for every newly created context, then puts the task
4845 : * on the runqueue and wakes it.
4846 : */
4847 175 : void wake_up_new_task(struct task_struct *p)
4848 : {
4849 : struct rq_flags rf;
4850 : struct rq *rq;
4851 :
4852 175 : raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4853 175 : WRITE_ONCE(p->__state, TASK_RUNNING);
4854 : #ifdef CONFIG_SMP
4855 : /*
4856 : * Fork balancing, do it here and not earlier because:
4857 : * - cpus_ptr can change in the fork path
4858 : * - any previously selected CPU might disappear through hotplug
4859 : *
4860 : * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4861 : * as we're not fully set-up yet.
4862 : */
4863 : p->recent_used_cpu = task_cpu(p);
4864 : rseq_migrate(p);
4865 : __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4866 : #endif
4867 175 : rq = __task_rq_lock(p, &rf);
4868 175 : update_rq_clock(rq);
4869 175 : post_init_entity_util_avg(p);
4870 :
4871 175 : activate_task(rq, p, ENQUEUE_NOCLOCK);
4872 175 : trace_sched_wakeup_new(p);
4873 175 : check_preempt_curr(rq, p, WF_FORK);
4874 : #ifdef CONFIG_SMP
4875 : if (p->sched_class->task_woken) {
4876 : /*
4877 : * Nothing relies on rq->lock after this, so it's fine to
4878 : * drop it.
4879 : */
4880 : rq_unpin_lock(rq, &rf);
4881 : p->sched_class->task_woken(rq, p);
4882 : rq_repin_lock(rq, &rf);
4883 : }
4884 : #endif
4885 350 : task_rq_unlock(rq, p, &rf);
4886 175 : }
4887 :
4888 : #ifdef CONFIG_PREEMPT_NOTIFIERS
4889 :
4890 : static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4891 :
4892 : void preempt_notifier_inc(void)
4893 : {
4894 : static_branch_inc(&preempt_notifier_key);
4895 : }
4896 : EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4897 :
4898 : void preempt_notifier_dec(void)
4899 : {
4900 : static_branch_dec(&preempt_notifier_key);
4901 : }
4902 : EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4903 :
4904 : /**
4905 : * preempt_notifier_register - tell me when current is being preempted & rescheduled
4906 : * @notifier: notifier struct to register
4907 : */
4908 : void preempt_notifier_register(struct preempt_notifier *notifier)
4909 : {
4910 : if (!static_branch_unlikely(&preempt_notifier_key))
4911 : WARN(1, "registering preempt_notifier while notifiers disabled\n");
4912 :
4913 : hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4914 : }
4915 : EXPORT_SYMBOL_GPL(preempt_notifier_register);
4916 :
4917 : /**
4918 : * preempt_notifier_unregister - no longer interested in preemption notifications
4919 : * @notifier: notifier struct to unregister
4920 : *
4921 : * This is *not* safe to call from within a preemption notifier.
4922 : */
4923 : void preempt_notifier_unregister(struct preempt_notifier *notifier)
4924 : {
4925 : hlist_del(¬ifier->link);
4926 : }
4927 : EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4928 :
4929 : static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4930 : {
4931 : struct preempt_notifier *notifier;
4932 :
4933 : hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4934 : notifier->ops->sched_in(notifier, raw_smp_processor_id());
4935 : }
4936 :
4937 : static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4938 : {
4939 : if (static_branch_unlikely(&preempt_notifier_key))
4940 : __fire_sched_in_preempt_notifiers(curr);
4941 : }
4942 :
4943 : static void
4944 : __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4945 : struct task_struct *next)
4946 : {
4947 : struct preempt_notifier *notifier;
4948 :
4949 : hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4950 : notifier->ops->sched_out(notifier, next);
4951 : }
4952 :
4953 : static __always_inline void
4954 : fire_sched_out_preempt_notifiers(struct task_struct *curr,
4955 : struct task_struct *next)
4956 : {
4957 : if (static_branch_unlikely(&preempt_notifier_key))
4958 : __fire_sched_out_preempt_notifiers(curr, next);
4959 : }
4960 :
4961 : #else /* !CONFIG_PREEMPT_NOTIFIERS */
4962 :
4963 : static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4964 : {
4965 : }
4966 :
4967 : static inline void
4968 : fire_sched_out_preempt_notifiers(struct task_struct *curr,
4969 : struct task_struct *next)
4970 : {
4971 : }
4972 :
4973 : #endif /* CONFIG_PREEMPT_NOTIFIERS */
4974 :
4975 : static inline void prepare_task(struct task_struct *next)
4976 : {
4977 : #ifdef CONFIG_SMP
4978 : /*
4979 : * Claim the task as running, we do this before switching to it
4980 : * such that any running task will have this set.
4981 : *
4982 : * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4983 : * its ordering comment.
4984 : */
4985 : WRITE_ONCE(next->on_cpu, 1);
4986 : #endif
4987 : }
4988 :
4989 : static inline void finish_task(struct task_struct *prev)
4990 : {
4991 : #ifdef CONFIG_SMP
4992 : /*
4993 : * This must be the very last reference to @prev from this CPU. After
4994 : * p->on_cpu is cleared, the task can be moved to a different CPU. We
4995 : * must ensure this doesn't happen until the switch is completely
4996 : * finished.
4997 : *
4998 : * In particular, the load of prev->state in finish_task_switch() must
4999 : * happen before this.
5000 : *
5001 : * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
5002 : */
5003 : smp_store_release(&prev->on_cpu, 0);
5004 : #endif
5005 : }
5006 :
5007 : #ifdef CONFIG_SMP
5008 :
5009 : static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
5010 : {
5011 : void (*func)(struct rq *rq);
5012 : struct balance_callback *next;
5013 :
5014 : lockdep_assert_rq_held(rq);
5015 :
5016 : while (head) {
5017 : func = (void (*)(struct rq *))head->func;
5018 : next = head->next;
5019 : head->next = NULL;
5020 : head = next;
5021 :
5022 : func(rq);
5023 : }
5024 : }
5025 :
5026 : static void balance_push(struct rq *rq);
5027 :
5028 : /*
5029 : * balance_push_callback is a right abuse of the callback interface and plays
5030 : * by significantly different rules.
5031 : *
5032 : * Where the normal balance_callback's purpose is to be ran in the same context
5033 : * that queued it (only later, when it's safe to drop rq->lock again),
5034 : * balance_push_callback is specifically targeted at __schedule().
5035 : *
5036 : * This abuse is tolerated because it places all the unlikely/odd cases behind
5037 : * a single test, namely: rq->balance_callback == NULL.
5038 : */
5039 : struct balance_callback balance_push_callback = {
5040 : .next = NULL,
5041 : .func = balance_push,
5042 : };
5043 :
5044 : static inline struct balance_callback *
5045 : __splice_balance_callbacks(struct rq *rq, bool split)
5046 : {
5047 : struct balance_callback *head = rq->balance_callback;
5048 :
5049 : if (likely(!head))
5050 : return NULL;
5051 :
5052 : lockdep_assert_rq_held(rq);
5053 : /*
5054 : * Must not take balance_push_callback off the list when
5055 : * splice_balance_callbacks() and balance_callbacks() are not
5056 : * in the same rq->lock section.
5057 : *
5058 : * In that case it would be possible for __schedule() to interleave
5059 : * and observe the list empty.
5060 : */
5061 : if (split && head == &balance_push_callback)
5062 : head = NULL;
5063 : else
5064 : rq->balance_callback = NULL;
5065 :
5066 : return head;
5067 : }
5068 :
5069 : static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5070 : {
5071 : return __splice_balance_callbacks(rq, true);
5072 : }
5073 :
5074 : static void __balance_callbacks(struct rq *rq)
5075 : {
5076 : do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
5077 : }
5078 :
5079 : static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5080 : {
5081 : unsigned long flags;
5082 :
5083 : if (unlikely(head)) {
5084 : raw_spin_rq_lock_irqsave(rq, flags);
5085 : do_balance_callbacks(rq, head);
5086 : raw_spin_rq_unlock_irqrestore(rq, flags);
5087 : }
5088 : }
5089 :
5090 : #else
5091 :
5092 : static inline void __balance_callbacks(struct rq *rq)
5093 : {
5094 : }
5095 :
5096 : static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5097 : {
5098 : return NULL;
5099 : }
5100 :
5101 : static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5102 : {
5103 : }
5104 :
5105 : #endif
5106 :
5107 : static inline void
5108 : prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
5109 : {
5110 : /*
5111 : * Since the runqueue lock will be released by the next
5112 : * task (which is an invalid locking op but in the case
5113 : * of the scheduler it's an obvious special-case), so we
5114 : * do an early lockdep release here:
5115 : */
5116 : rq_unpin_lock(rq, rf);
5117 : spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
5118 : #ifdef CONFIG_DEBUG_SPINLOCK
5119 : /* this is a valid case when another task releases the spinlock */
5120 : rq_lockp(rq)->owner = next;
5121 : #endif
5122 : }
5123 :
5124 : static inline void finish_lock_switch(struct rq *rq)
5125 : {
5126 : /*
5127 : * If we are tracking spinlock dependencies then we have to
5128 : * fix up the runqueue lock - which gets 'carried over' from
5129 : * prev into current:
5130 : */
5131 : spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
5132 1031 : __balance_callbacks(rq);
5133 1031 : raw_spin_rq_unlock_irq(rq);
5134 : }
5135 :
5136 : /*
5137 : * NOP if the arch has not defined these:
5138 : */
5139 :
5140 : #ifndef prepare_arch_switch
5141 : # define prepare_arch_switch(next) do { } while (0)
5142 : #endif
5143 :
5144 : #ifndef finish_arch_post_lock_switch
5145 : # define finish_arch_post_lock_switch() do { } while (0)
5146 : #endif
5147 :
5148 : static inline void kmap_local_sched_out(void)
5149 : {
5150 : #ifdef CONFIG_KMAP_LOCAL
5151 : if (unlikely(current->kmap_ctrl.idx))
5152 : __kmap_local_sched_out();
5153 : #endif
5154 : }
5155 :
5156 : static inline void kmap_local_sched_in(void)
5157 : {
5158 : #ifdef CONFIG_KMAP_LOCAL
5159 : if (unlikely(current->kmap_ctrl.idx))
5160 : __kmap_local_sched_in();
5161 : #endif
5162 : }
5163 :
5164 : /**
5165 : * prepare_task_switch - prepare to switch tasks
5166 : * @rq: the runqueue preparing to switch
5167 : * @prev: the current task that is being switched out
5168 : * @next: the task we are going to switch to.
5169 : *
5170 : * This is called with the rq lock held and interrupts off. It must
5171 : * be paired with a subsequent finish_task_switch after the context
5172 : * switch.
5173 : *
5174 : * prepare_task_switch sets up locking and calls architecture specific
5175 : * hooks.
5176 : */
5177 : static inline void
5178 : prepare_task_switch(struct rq *rq, struct task_struct *prev,
5179 : struct task_struct *next)
5180 : {
5181 : kcov_prepare_switch(prev);
5182 : sched_info_switch(rq, prev, next);
5183 : perf_event_task_sched_out(prev, next);
5184 : rseq_preempt(prev);
5185 : fire_sched_out_preempt_notifiers(prev, next);
5186 : kmap_local_sched_out();
5187 : prepare_task(next);
5188 : prepare_arch_switch(next);
5189 : }
5190 :
5191 : /**
5192 : * finish_task_switch - clean up after a task-switch
5193 : * @prev: the thread we just switched away from.
5194 : *
5195 : * finish_task_switch must be called after the context switch, paired
5196 : * with a prepare_task_switch call before the context switch.
5197 : * finish_task_switch will reconcile locking set up by prepare_task_switch,
5198 : * and do any other architecture-specific cleanup actions.
5199 : *
5200 : * Note that we may have delayed dropping an mm in context_switch(). If
5201 : * so, we finish that here outside of the runqueue lock. (Doing it
5202 : * with the lock held can cause deadlocks; see schedule() for
5203 : * details.)
5204 : *
5205 : * The context switch have flipped the stack from under us and restored the
5206 : * local variables which were saved when this task called schedule() in the
5207 : * past. prev == current is still correct but we need to recalculate this_rq
5208 : * because prev may have moved to another CPU.
5209 : */
5210 1031 : static struct rq *finish_task_switch(struct task_struct *prev)
5211 : __releases(rq->lock)
5212 : {
5213 1031 : struct rq *rq = this_rq();
5214 1031 : struct mm_struct *mm = rq->prev_mm;
5215 : unsigned int prev_state;
5216 :
5217 : /*
5218 : * The previous task will have left us with a preempt_count of 2
5219 : * because it left us after:
5220 : *
5221 : * schedule()
5222 : * preempt_disable(); // 1
5223 : * __schedule()
5224 : * raw_spin_lock_irq(&rq->lock) // 2
5225 : *
5226 : * Also, see FORK_PREEMPT_COUNT.
5227 : */
5228 1031 : if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5229 : "corrupted preempt_count: %s/%d/0x%x\n",
5230 : current->comm, current->pid, preempt_count()))
5231 : preempt_count_set(FORK_PREEMPT_COUNT);
5232 :
5233 1031 : rq->prev_mm = NULL;
5234 :
5235 : /*
5236 : * A task struct has one reference for the use as "current".
5237 : * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5238 : * schedule one last time. The schedule call will never return, and
5239 : * the scheduled task must drop that reference.
5240 : *
5241 : * We must observe prev->state before clearing prev->on_cpu (in
5242 : * finish_task), otherwise a concurrent wakeup can get prev
5243 : * running on another CPU and we could rave with its RUNNING -> DEAD
5244 : * transition, resulting in a double drop.
5245 : */
5246 1031 : prev_state = READ_ONCE(prev->__state);
5247 1031 : vtime_task_switch(prev);
5248 1031 : perf_event_task_sched_in(prev, current);
5249 1031 : finish_task(prev);
5250 : tick_nohz_task_switch();
5251 1031 : finish_lock_switch(rq);
5252 : finish_arch_post_lock_switch();
5253 1031 : kcov_finish_switch(current);
5254 : /*
5255 : * kmap_local_sched_out() is invoked with rq::lock held and
5256 : * interrupts disabled. There is no requirement for that, but the
5257 : * sched out code does not have an interrupt enabled section.
5258 : * Restoring the maps on sched in does not require interrupts being
5259 : * disabled either.
5260 : */
5261 : kmap_local_sched_in();
5262 :
5263 1031 : fire_sched_in_preempt_notifiers(current);
5264 : /*
5265 : * When switching through a kernel thread, the loop in
5266 : * membarrier_{private,global}_expedited() may have observed that
5267 : * kernel thread and not issued an IPI. It is therefore possible to
5268 : * schedule between user->kernel->user threads without passing though
5269 : * switch_mm(). Membarrier requires a barrier after storing to
5270 : * rq->curr, before returning to userspace, so provide them here:
5271 : *
5272 : * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5273 : * provided by mmdrop_lazy_tlb(),
5274 : * - a sync_core for SYNC_CORE.
5275 : */
5276 1031 : if (mm) {
5277 0 : membarrier_mm_sync_core_before_usermode(mm);
5278 : mmdrop_lazy_tlb_sched(mm);
5279 : }
5280 :
5281 1031 : if (unlikely(prev_state == TASK_DEAD)) {
5282 160 : if (prev->sched_class->task_dead)
5283 0 : prev->sched_class->task_dead(prev);
5284 :
5285 : /* Task is done with its stack. */
5286 160 : put_task_stack(prev);
5287 :
5288 160 : put_task_struct_rcu_user(prev);
5289 : }
5290 :
5291 1031 : return rq;
5292 : }
5293 :
5294 : /**
5295 : * schedule_tail - first thing a freshly forked thread must call.
5296 : * @prev: the thread we just switched away from.
5297 : */
5298 175 : asmlinkage __visible void schedule_tail(struct task_struct *prev)
5299 : __releases(rq->lock)
5300 : {
5301 : /*
5302 : * New tasks start with FORK_PREEMPT_COUNT, see there and
5303 : * finish_task_switch() for details.
5304 : *
5305 : * finish_task_switch() will drop rq->lock() and lower preempt_count
5306 : * and the preempt_enable() will end up enabling preemption (on
5307 : * PREEMPT_COUNT kernels).
5308 : */
5309 :
5310 175 : finish_task_switch(prev);
5311 175 : preempt_enable();
5312 :
5313 175 : if (current->set_child_tid)
5314 0 : put_user(task_pid_vnr(current), current->set_child_tid);
5315 :
5316 175 : calculate_sigpending();
5317 175 : }
5318 :
5319 : /*
5320 : * context_switch - switch to the new MM and the new thread's register state.
5321 : */
5322 : static __always_inline struct rq *
5323 : context_switch(struct rq *rq, struct task_struct *prev,
5324 : struct task_struct *next, struct rq_flags *rf)
5325 : {
5326 1031 : prepare_task_switch(rq, prev, next);
5327 :
5328 : /*
5329 : * For paravirt, this is coupled with an exit in switch_to to
5330 : * combine the page table reload and the switch backend into
5331 : * one hypercall.
5332 : */
5333 : arch_start_context_switch(prev);
5334 :
5335 : /*
5336 : * kernel -> kernel lazy + transfer active
5337 : * user -> kernel lazy + mmgrab_lazy_tlb() active
5338 : *
5339 : * kernel -> user switch + mmdrop_lazy_tlb() active
5340 : * user -> user switch
5341 : *
5342 : * switch_mm_cid() needs to be updated if the barriers provided
5343 : * by context_switch() are modified.
5344 : */
5345 1031 : if (!next->mm) { // to kernel
5346 1031 : enter_lazy_tlb(prev->active_mm, next);
5347 :
5348 1031 : next->active_mm = prev->active_mm;
5349 1031 : if (prev->mm) // from user
5350 0 : mmgrab_lazy_tlb(prev->active_mm);
5351 : else
5352 1031 : prev->active_mm = NULL;
5353 : } else { // to user
5354 0 : membarrier_switch_mm(rq, prev->active_mm, next->mm);
5355 : /*
5356 : * sys_membarrier() requires an smp_mb() between setting
5357 : * rq->curr / membarrier_switch_mm() and returning to userspace.
5358 : *
5359 : * The below provides this either through switch_mm(), or in
5360 : * case 'prev->active_mm == next->mm' through
5361 : * finish_task_switch()'s mmdrop().
5362 : */
5363 0 : switch_mm_irqs_off(prev->active_mm, next->mm, next);
5364 0 : lru_gen_use_mm(next->mm);
5365 :
5366 0 : if (!prev->mm) { // from kernel
5367 : /* will mmdrop_lazy_tlb() in finish_task_switch(). */
5368 0 : rq->prev_mm = prev->active_mm;
5369 0 : prev->active_mm = NULL;
5370 : }
5371 : }
5372 :
5373 : /* switch_mm_cid() requires the memory barriers above. */
5374 1031 : switch_mm_cid(rq, prev, next);
5375 :
5376 1031 : rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5377 :
5378 1031 : prepare_lock_switch(rq, next, rf);
5379 :
5380 : /* Here we just switch the register state and the stack. */
5381 1031 : switch_to(prev, next, prev);
5382 856 : barrier();
5383 :
5384 856 : return finish_task_switch(prev);
5385 : }
5386 :
5387 : /*
5388 : * nr_running and nr_context_switches:
5389 : *
5390 : * externally visible scheduler statistics: current number of runnable
5391 : * threads, total number of context switches performed since bootup.
5392 : */
5393 0 : unsigned int nr_running(void)
5394 : {
5395 0 : unsigned int i, sum = 0;
5396 :
5397 0 : for_each_online_cpu(i)
5398 0 : sum += cpu_rq(i)->nr_running;
5399 :
5400 0 : return sum;
5401 : }
5402 :
5403 : /*
5404 : * Check if only the current task is running on the CPU.
5405 : *
5406 : * Caution: this function does not check that the caller has disabled
5407 : * preemption, thus the result might have a time-of-check-to-time-of-use
5408 : * race. The caller is responsible to use it correctly, for example:
5409 : *
5410 : * - from a non-preemptible section (of course)
5411 : *
5412 : * - from a thread that is bound to a single CPU
5413 : *
5414 : * - in a loop with very short iterations (e.g. a polling loop)
5415 : */
5416 0 : bool single_task_running(void)
5417 : {
5418 0 : return raw_rq()->nr_running == 1;
5419 : }
5420 : EXPORT_SYMBOL(single_task_running);
5421 :
5422 0 : unsigned long long nr_context_switches_cpu(int cpu)
5423 : {
5424 0 : return cpu_rq(cpu)->nr_switches;
5425 : }
5426 :
5427 0 : unsigned long long nr_context_switches(void)
5428 : {
5429 : int i;
5430 0 : unsigned long long sum = 0;
5431 :
5432 0 : for_each_possible_cpu(i)
5433 0 : sum += cpu_rq(i)->nr_switches;
5434 :
5435 0 : return sum;
5436 : }
5437 :
5438 : /*
5439 : * Consumers of these two interfaces, like for example the cpuidle menu
5440 : * governor, are using nonsensical data. Preferring shallow idle state selection
5441 : * for a CPU that has IO-wait which might not even end up running the task when
5442 : * it does become runnable.
5443 : */
5444 :
5445 0 : unsigned int nr_iowait_cpu(int cpu)
5446 : {
5447 0 : return atomic_read(&cpu_rq(cpu)->nr_iowait);
5448 : }
5449 :
5450 : /*
5451 : * IO-wait accounting, and how it's mostly bollocks (on SMP).
5452 : *
5453 : * The idea behind IO-wait account is to account the idle time that we could
5454 : * have spend running if it were not for IO. That is, if we were to improve the
5455 : * storage performance, we'd have a proportional reduction in IO-wait time.
5456 : *
5457 : * This all works nicely on UP, where, when a task blocks on IO, we account
5458 : * idle time as IO-wait, because if the storage were faster, it could've been
5459 : * running and we'd not be idle.
5460 : *
5461 : * This has been extended to SMP, by doing the same for each CPU. This however
5462 : * is broken.
5463 : *
5464 : * Imagine for instance the case where two tasks block on one CPU, only the one
5465 : * CPU will have IO-wait accounted, while the other has regular idle. Even
5466 : * though, if the storage were faster, both could've ran at the same time,
5467 : * utilising both CPUs.
5468 : *
5469 : * This means, that when looking globally, the current IO-wait accounting on
5470 : * SMP is a lower bound, by reason of under accounting.
5471 : *
5472 : * Worse, since the numbers are provided per CPU, they are sometimes
5473 : * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5474 : * associated with any one particular CPU, it can wake to another CPU than it
5475 : * blocked on. This means the per CPU IO-wait number is meaningless.
5476 : *
5477 : * Task CPU affinities can make all that even more 'interesting'.
5478 : */
5479 :
5480 0 : unsigned int nr_iowait(void)
5481 : {
5482 0 : unsigned int i, sum = 0;
5483 :
5484 0 : for_each_possible_cpu(i)
5485 0 : sum += nr_iowait_cpu(i);
5486 :
5487 0 : return sum;
5488 : }
5489 :
5490 : #ifdef CONFIG_SMP
5491 :
5492 : /*
5493 : * sched_exec - execve() is a valuable balancing opportunity, because at
5494 : * this point the task has the smallest effective memory and cache footprint.
5495 : */
5496 : void sched_exec(void)
5497 : {
5498 : struct task_struct *p = current;
5499 : unsigned long flags;
5500 : int dest_cpu;
5501 :
5502 : raw_spin_lock_irqsave(&p->pi_lock, flags);
5503 : dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5504 : if (dest_cpu == smp_processor_id())
5505 : goto unlock;
5506 :
5507 : if (likely(cpu_active(dest_cpu))) {
5508 : struct migration_arg arg = { p, dest_cpu };
5509 :
5510 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5511 : stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5512 : return;
5513 : }
5514 : unlock:
5515 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5516 : }
5517 :
5518 : #endif
5519 :
5520 : DEFINE_PER_CPU(struct kernel_stat, kstat);
5521 : DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5522 :
5523 : EXPORT_PER_CPU_SYMBOL(kstat);
5524 : EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5525 :
5526 : /*
5527 : * The function fair_sched_class.update_curr accesses the struct curr
5528 : * and its field curr->exec_start; when called from task_sched_runtime(),
5529 : * we observe a high rate of cache misses in practice.
5530 : * Prefetching this data results in improved performance.
5531 : */
5532 : static inline void prefetch_curr_exec_start(struct task_struct *p)
5533 : {
5534 : #ifdef CONFIG_FAIR_GROUP_SCHED
5535 : struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5536 : #else
5537 0 : struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5538 : #endif
5539 0 : prefetch(curr);
5540 0 : prefetch(&curr->exec_start);
5541 : }
5542 :
5543 : /*
5544 : * Return accounted runtime for the task.
5545 : * In case the task is currently running, return the runtime plus current's
5546 : * pending runtime that have not been accounted yet.
5547 : */
5548 0 : unsigned long long task_sched_runtime(struct task_struct *p)
5549 : {
5550 : struct rq_flags rf;
5551 : struct rq *rq;
5552 : u64 ns;
5553 :
5554 : #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5555 : /*
5556 : * 64-bit doesn't need locks to atomically read a 64-bit value.
5557 : * So we have a optimization chance when the task's delta_exec is 0.
5558 : * Reading ->on_cpu is racy, but this is ok.
5559 : *
5560 : * If we race with it leaving CPU, we'll take a lock. So we're correct.
5561 : * If we race with it entering CPU, unaccounted time is 0. This is
5562 : * indistinguishable from the read occurring a few cycles earlier.
5563 : * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5564 : * been accounted, so we're correct here as well.
5565 : */
5566 : if (!p->on_cpu || !task_on_rq_queued(p))
5567 : return p->se.sum_exec_runtime;
5568 : #endif
5569 :
5570 0 : rq = task_rq_lock(p, &rf);
5571 : /*
5572 : * Must be ->curr _and_ ->on_rq. If dequeued, we would
5573 : * project cycles that may never be accounted to this
5574 : * thread, breaking clock_gettime().
5575 : */
5576 0 : if (task_current(rq, p) && task_on_rq_queued(p)) {
5577 0 : prefetch_curr_exec_start(p);
5578 0 : update_rq_clock(rq);
5579 0 : p->sched_class->update_curr(rq);
5580 : }
5581 0 : ns = p->se.sum_exec_runtime;
5582 0 : task_rq_unlock(rq, p, &rf);
5583 :
5584 0 : return ns;
5585 : }
5586 :
5587 : #ifdef CONFIG_SCHED_DEBUG
5588 : static u64 cpu_resched_latency(struct rq *rq)
5589 : {
5590 : int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5591 : u64 resched_latency, now = rq_clock(rq);
5592 : static bool warned_once;
5593 :
5594 : if (sysctl_resched_latency_warn_once && warned_once)
5595 : return 0;
5596 :
5597 : if (!need_resched() || !latency_warn_ms)
5598 : return 0;
5599 :
5600 : if (system_state == SYSTEM_BOOTING)
5601 : return 0;
5602 :
5603 : if (!rq->last_seen_need_resched_ns) {
5604 : rq->last_seen_need_resched_ns = now;
5605 : rq->ticks_without_resched = 0;
5606 : return 0;
5607 : }
5608 :
5609 : rq->ticks_without_resched++;
5610 : resched_latency = now - rq->last_seen_need_resched_ns;
5611 : if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5612 : return 0;
5613 :
5614 : warned_once = true;
5615 :
5616 : return resched_latency;
5617 : }
5618 :
5619 : static int __init setup_resched_latency_warn_ms(char *str)
5620 : {
5621 : long val;
5622 :
5623 : if ((kstrtol(str, 0, &val))) {
5624 : pr_warn("Unable to set resched_latency_warn_ms\n");
5625 : return 1;
5626 : }
5627 :
5628 : sysctl_resched_latency_warn_ms = val;
5629 : return 1;
5630 : }
5631 : __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5632 : #else
5633 : static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5634 : #endif /* CONFIG_SCHED_DEBUG */
5635 :
5636 : /*
5637 : * This function gets called by the timer code, with HZ frequency.
5638 : * We call it with interrupts disabled.
5639 : */
5640 5 : void scheduler_tick(void)
5641 : {
5642 5 : int cpu = smp_processor_id();
5643 5 : struct rq *rq = cpu_rq(cpu);
5644 5 : struct task_struct *curr = rq->curr;
5645 : struct rq_flags rf;
5646 : unsigned long thermal_pressure;
5647 : u64 resched_latency;
5648 :
5649 5 : if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5650 : arch_scale_freq_tick();
5651 :
5652 : sched_clock_tick();
5653 :
5654 5 : rq_lock(rq, &rf);
5655 :
5656 5 : update_rq_clock(rq);
5657 5 : thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5658 5 : update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5659 5 : curr->sched_class->task_tick(rq, curr, 0);
5660 : if (sched_feat(LATENCY_WARN))
5661 : resched_latency = cpu_resched_latency(rq);
5662 5 : calc_global_load_tick(rq);
5663 5 : sched_core_tick(rq);
5664 5 : task_tick_mm_cid(rq, curr);
5665 :
5666 5 : rq_unlock(rq, &rf);
5667 :
5668 : if (sched_feat(LATENCY_WARN) && resched_latency)
5669 : resched_latency_warn(cpu, resched_latency);
5670 :
5671 : perf_event_task_tick();
5672 :
5673 5 : if (curr->flags & PF_WQ_WORKER)
5674 0 : wq_worker_tick(curr);
5675 :
5676 : #ifdef CONFIG_SMP
5677 : rq->idle_balance = idle_cpu(cpu);
5678 : trigger_load_balance(rq);
5679 : #endif
5680 5 : }
5681 :
5682 : #ifdef CONFIG_NO_HZ_FULL
5683 :
5684 : struct tick_work {
5685 : int cpu;
5686 : atomic_t state;
5687 : struct delayed_work work;
5688 : };
5689 : /* Values for ->state, see diagram below. */
5690 : #define TICK_SCHED_REMOTE_OFFLINE 0
5691 : #define TICK_SCHED_REMOTE_OFFLINING 1
5692 : #define TICK_SCHED_REMOTE_RUNNING 2
5693 :
5694 : /*
5695 : * State diagram for ->state:
5696 : *
5697 : *
5698 : * TICK_SCHED_REMOTE_OFFLINE
5699 : * | ^
5700 : * | |
5701 : * | | sched_tick_remote()
5702 : * | |
5703 : * | |
5704 : * +--TICK_SCHED_REMOTE_OFFLINING
5705 : * | ^
5706 : * | |
5707 : * sched_tick_start() | | sched_tick_stop()
5708 : * | |
5709 : * V |
5710 : * TICK_SCHED_REMOTE_RUNNING
5711 : *
5712 : *
5713 : * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5714 : * and sched_tick_start() are happy to leave the state in RUNNING.
5715 : */
5716 :
5717 : static struct tick_work __percpu *tick_work_cpu;
5718 :
5719 : static void sched_tick_remote(struct work_struct *work)
5720 : {
5721 : struct delayed_work *dwork = to_delayed_work(work);
5722 : struct tick_work *twork = container_of(dwork, struct tick_work, work);
5723 : int cpu = twork->cpu;
5724 : struct rq *rq = cpu_rq(cpu);
5725 : struct task_struct *curr;
5726 : struct rq_flags rf;
5727 : u64 delta;
5728 : int os;
5729 :
5730 : /*
5731 : * Handle the tick only if it appears the remote CPU is running in full
5732 : * dynticks mode. The check is racy by nature, but missing a tick or
5733 : * having one too much is no big deal because the scheduler tick updates
5734 : * statistics and checks timeslices in a time-independent way, regardless
5735 : * of when exactly it is running.
5736 : */
5737 : if (!tick_nohz_tick_stopped_cpu(cpu))
5738 : goto out_requeue;
5739 :
5740 : rq_lock_irq(rq, &rf);
5741 : curr = rq->curr;
5742 : if (cpu_is_offline(cpu))
5743 : goto out_unlock;
5744 :
5745 : update_rq_clock(rq);
5746 :
5747 : if (!is_idle_task(curr)) {
5748 : /*
5749 : * Make sure the next tick runs within a reasonable
5750 : * amount of time.
5751 : */
5752 : delta = rq_clock_task(rq) - curr->se.exec_start;
5753 : WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5754 : }
5755 : curr->sched_class->task_tick(rq, curr, 0);
5756 :
5757 : calc_load_nohz_remote(rq);
5758 : out_unlock:
5759 : rq_unlock_irq(rq, &rf);
5760 : out_requeue:
5761 :
5762 : /*
5763 : * Run the remote tick once per second (1Hz). This arbitrary
5764 : * frequency is large enough to avoid overload but short enough
5765 : * to keep scheduler internal stats reasonably up to date. But
5766 : * first update state to reflect hotplug activity if required.
5767 : */
5768 : os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5769 : WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5770 : if (os == TICK_SCHED_REMOTE_RUNNING)
5771 : queue_delayed_work(system_unbound_wq, dwork, HZ);
5772 : }
5773 :
5774 : static void sched_tick_start(int cpu)
5775 : {
5776 : int os;
5777 : struct tick_work *twork;
5778 :
5779 : if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5780 : return;
5781 :
5782 : WARN_ON_ONCE(!tick_work_cpu);
5783 :
5784 : twork = per_cpu_ptr(tick_work_cpu, cpu);
5785 : os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5786 : WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5787 : if (os == TICK_SCHED_REMOTE_OFFLINE) {
5788 : twork->cpu = cpu;
5789 : INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5790 : queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5791 : }
5792 : }
5793 :
5794 : #ifdef CONFIG_HOTPLUG_CPU
5795 : static void sched_tick_stop(int cpu)
5796 : {
5797 : struct tick_work *twork;
5798 : int os;
5799 :
5800 : if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5801 : return;
5802 :
5803 : WARN_ON_ONCE(!tick_work_cpu);
5804 :
5805 : twork = per_cpu_ptr(tick_work_cpu, cpu);
5806 : /* There cannot be competing actions, but don't rely on stop-machine. */
5807 : os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5808 : WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5809 : /* Don't cancel, as this would mess up the state machine. */
5810 : }
5811 : #endif /* CONFIG_HOTPLUG_CPU */
5812 :
5813 : int __init sched_tick_offload_init(void)
5814 : {
5815 : tick_work_cpu = alloc_percpu(struct tick_work);
5816 : BUG_ON(!tick_work_cpu);
5817 : return 0;
5818 : }
5819 :
5820 : #else /* !CONFIG_NO_HZ_FULL */
5821 : static inline void sched_tick_start(int cpu) { }
5822 : static inline void sched_tick_stop(int cpu) { }
5823 : #endif
5824 :
5825 : #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5826 : defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5827 : /*
5828 : * If the value passed in is equal to the current preempt count
5829 : * then we just disabled preemption. Start timing the latency.
5830 : */
5831 : static inline void preempt_latency_start(int val)
5832 : {
5833 : if (preempt_count() == val) {
5834 : unsigned long ip = get_lock_parent_ip();
5835 : #ifdef CONFIG_DEBUG_PREEMPT
5836 : current->preempt_disable_ip = ip;
5837 : #endif
5838 : trace_preempt_off(CALLER_ADDR0, ip);
5839 : }
5840 : }
5841 :
5842 : void preempt_count_add(int val)
5843 : {
5844 : #ifdef CONFIG_DEBUG_PREEMPT
5845 : /*
5846 : * Underflow?
5847 : */
5848 : if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5849 : return;
5850 : #endif
5851 : __preempt_count_add(val);
5852 : #ifdef CONFIG_DEBUG_PREEMPT
5853 : /*
5854 : * Spinlock count overflowing soon?
5855 : */
5856 : DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5857 : PREEMPT_MASK - 10);
5858 : #endif
5859 : preempt_latency_start(val);
5860 : }
5861 : EXPORT_SYMBOL(preempt_count_add);
5862 : NOKPROBE_SYMBOL(preempt_count_add);
5863 :
5864 : /*
5865 : * If the value passed in equals to the current preempt count
5866 : * then we just enabled preemption. Stop timing the latency.
5867 : */
5868 : static inline void preempt_latency_stop(int val)
5869 : {
5870 : if (preempt_count() == val)
5871 : trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5872 : }
5873 :
5874 : void preempt_count_sub(int val)
5875 : {
5876 : #ifdef CONFIG_DEBUG_PREEMPT
5877 : /*
5878 : * Underflow?
5879 : */
5880 : if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5881 : return;
5882 : /*
5883 : * Is the spinlock portion underflowing?
5884 : */
5885 : if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5886 : !(preempt_count() & PREEMPT_MASK)))
5887 : return;
5888 : #endif
5889 :
5890 : preempt_latency_stop(val);
5891 : __preempt_count_sub(val);
5892 : }
5893 : EXPORT_SYMBOL(preempt_count_sub);
5894 : NOKPROBE_SYMBOL(preempt_count_sub);
5895 :
5896 : #else
5897 : static inline void preempt_latency_start(int val) { }
5898 : static inline void preempt_latency_stop(int val) { }
5899 : #endif
5900 :
5901 : static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5902 : {
5903 : #ifdef CONFIG_DEBUG_PREEMPT
5904 : return p->preempt_disable_ip;
5905 : #else
5906 : return 0;
5907 : #endif
5908 : }
5909 :
5910 : /*
5911 : * Print scheduling while atomic bug:
5912 : */
5913 0 : static noinline void __schedule_bug(struct task_struct *prev)
5914 : {
5915 : /* Save this before calling printk(), since that will clobber it */
5916 0 : unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5917 :
5918 0 : if (oops_in_progress)
5919 : return;
5920 :
5921 0 : printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5922 : prev->comm, prev->pid, preempt_count());
5923 :
5924 0 : debug_show_held_locks(prev);
5925 : print_modules();
5926 0 : if (irqs_disabled())
5927 : print_irqtrace_events(prev);
5928 : if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5929 : && in_atomic_preempt_off()) {
5930 : pr_err("Preemption disabled at:");
5931 : print_ip_sym(KERN_ERR, preempt_disable_ip);
5932 : }
5933 0 : check_panic_on_warn("scheduling while atomic");
5934 :
5935 0 : dump_stack();
5936 0 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5937 : }
5938 :
5939 : /*
5940 : * Various schedule()-time debugging checks and statistics:
5941 : */
5942 : static inline void schedule_debug(struct task_struct *prev, bool preempt)
5943 : {
5944 : #ifdef CONFIG_SCHED_STACK_END_CHECK
5945 : if (task_stack_end_corrupted(prev))
5946 : panic("corrupted stack end detected inside scheduler\n");
5947 :
5948 : if (task_scs_end_corrupted(prev))
5949 : panic("corrupted shadow stack detected inside scheduler\n");
5950 : #endif
5951 :
5952 : #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5953 : if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5954 : printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5955 : prev->comm, prev->pid, prev->non_block_count);
5956 : dump_stack();
5957 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5958 : }
5959 : #endif
5960 :
5961 1032 : if (unlikely(in_atomic_preempt_off())) {
5962 0 : __schedule_bug(prev);
5963 : preempt_count_set(PREEMPT_DISABLED);
5964 : }
5965 : rcu_sleep_check();
5966 : SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5967 :
5968 1032 : profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5969 :
5970 : schedstat_inc(this_rq()->sched_count);
5971 : }
5972 :
5973 : static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5974 : struct rq_flags *rf)
5975 : {
5976 : #ifdef CONFIG_SMP
5977 : const struct sched_class *class;
5978 : /*
5979 : * We must do the balancing pass before put_prev_task(), such
5980 : * that when we release the rq->lock the task is in the same
5981 : * state as before we took rq->lock.
5982 : *
5983 : * We can terminate the balance pass as soon as we know there is
5984 : * a runnable task of @class priority or higher.
5985 : */
5986 : for_class_range(class, prev->sched_class, &idle_sched_class) {
5987 : if (class->balance(rq, prev, rf))
5988 : break;
5989 : }
5990 : #endif
5991 :
5992 0 : put_prev_task(rq, prev);
5993 : }
5994 :
5995 : /*
5996 : * Pick up the highest-prio task:
5997 : */
5998 : static inline struct task_struct *
5999 1032 : __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6000 : {
6001 : const struct sched_class *class;
6002 : struct task_struct *p;
6003 :
6004 : /*
6005 : * Optimization: we know that if all tasks are in the fair class we can
6006 : * call that function directly, but only if the @prev task wasn't of a
6007 : * higher scheduling class, because otherwise those lose the
6008 : * opportunity to pull in more work from other CPUs.
6009 : */
6010 1032 : if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
6011 : rq->nr_running == rq->cfs.h_nr_running)) {
6012 :
6013 1032 : p = pick_next_task_fair(rq, prev, rf);
6014 1032 : if (unlikely(p == RETRY_TASK))
6015 : goto restart;
6016 :
6017 : /* Assume the next prioritized class is idle_sched_class */
6018 1032 : if (!p) {
6019 2 : put_prev_task(rq, prev);
6020 2 : p = pick_next_task_idle(rq);
6021 : }
6022 :
6023 : return p;
6024 : }
6025 :
6026 : restart:
6027 0 : put_prev_task_balance(rq, prev, rf);
6028 :
6029 0 : for_each_class(class) {
6030 0 : p = class->pick_next_task(rq);
6031 0 : if (p)
6032 : return p;
6033 : }
6034 :
6035 0 : BUG(); /* The idle class should always have a runnable task. */
6036 : }
6037 :
6038 : #ifdef CONFIG_SCHED_CORE
6039 : static inline bool is_task_rq_idle(struct task_struct *t)
6040 : {
6041 : return (task_rq(t)->idle == t);
6042 : }
6043 :
6044 : static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
6045 : {
6046 : return is_task_rq_idle(a) || (a->core_cookie == cookie);
6047 : }
6048 :
6049 : static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
6050 : {
6051 : if (is_task_rq_idle(a) || is_task_rq_idle(b))
6052 : return true;
6053 :
6054 : return a->core_cookie == b->core_cookie;
6055 : }
6056 :
6057 : static inline struct task_struct *pick_task(struct rq *rq)
6058 : {
6059 : const struct sched_class *class;
6060 : struct task_struct *p;
6061 :
6062 : for_each_class(class) {
6063 : p = class->pick_task(rq);
6064 : if (p)
6065 : return p;
6066 : }
6067 :
6068 : BUG(); /* The idle class should always have a runnable task. */
6069 : }
6070 :
6071 : extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
6072 :
6073 : static void queue_core_balance(struct rq *rq);
6074 :
6075 : static struct task_struct *
6076 : pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6077 : {
6078 : struct task_struct *next, *p, *max = NULL;
6079 : const struct cpumask *smt_mask;
6080 : bool fi_before = false;
6081 : bool core_clock_updated = (rq == rq->core);
6082 : unsigned long cookie;
6083 : int i, cpu, occ = 0;
6084 : struct rq *rq_i;
6085 : bool need_sync;
6086 :
6087 : if (!sched_core_enabled(rq))
6088 : return __pick_next_task(rq, prev, rf);
6089 :
6090 : cpu = cpu_of(rq);
6091 :
6092 : /* Stopper task is switching into idle, no need core-wide selection. */
6093 : if (cpu_is_offline(cpu)) {
6094 : /*
6095 : * Reset core_pick so that we don't enter the fastpath when
6096 : * coming online. core_pick would already be migrated to
6097 : * another cpu during offline.
6098 : */
6099 : rq->core_pick = NULL;
6100 : return __pick_next_task(rq, prev, rf);
6101 : }
6102 :
6103 : /*
6104 : * If there were no {en,de}queues since we picked (IOW, the task
6105 : * pointers are all still valid), and we haven't scheduled the last
6106 : * pick yet, do so now.
6107 : *
6108 : * rq->core_pick can be NULL if no selection was made for a CPU because
6109 : * it was either offline or went offline during a sibling's core-wide
6110 : * selection. In this case, do a core-wide selection.
6111 : */
6112 : if (rq->core->core_pick_seq == rq->core->core_task_seq &&
6113 : rq->core->core_pick_seq != rq->core_sched_seq &&
6114 : rq->core_pick) {
6115 : WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
6116 :
6117 : next = rq->core_pick;
6118 : if (next != prev) {
6119 : put_prev_task(rq, prev);
6120 : set_next_task(rq, next);
6121 : }
6122 :
6123 : rq->core_pick = NULL;
6124 : goto out;
6125 : }
6126 :
6127 : put_prev_task_balance(rq, prev, rf);
6128 :
6129 : smt_mask = cpu_smt_mask(cpu);
6130 : need_sync = !!rq->core->core_cookie;
6131 :
6132 : /* reset state */
6133 : rq->core->core_cookie = 0UL;
6134 : if (rq->core->core_forceidle_count) {
6135 : if (!core_clock_updated) {
6136 : update_rq_clock(rq->core);
6137 : core_clock_updated = true;
6138 : }
6139 : sched_core_account_forceidle(rq);
6140 : /* reset after accounting force idle */
6141 : rq->core->core_forceidle_start = 0;
6142 : rq->core->core_forceidle_count = 0;
6143 : rq->core->core_forceidle_occupation = 0;
6144 : need_sync = true;
6145 : fi_before = true;
6146 : }
6147 :
6148 : /*
6149 : * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6150 : *
6151 : * @task_seq guards the task state ({en,de}queues)
6152 : * @pick_seq is the @task_seq we did a selection on
6153 : * @sched_seq is the @pick_seq we scheduled
6154 : *
6155 : * However, preemptions can cause multiple picks on the same task set.
6156 : * 'Fix' this by also increasing @task_seq for every pick.
6157 : */
6158 : rq->core->core_task_seq++;
6159 :
6160 : /*
6161 : * Optimize for common case where this CPU has no cookies
6162 : * and there are no cookied tasks running on siblings.
6163 : */
6164 : if (!need_sync) {
6165 : next = pick_task(rq);
6166 : if (!next->core_cookie) {
6167 : rq->core_pick = NULL;
6168 : /*
6169 : * For robustness, update the min_vruntime_fi for
6170 : * unconstrained picks as well.
6171 : */
6172 : WARN_ON_ONCE(fi_before);
6173 : task_vruntime_update(rq, next, false);
6174 : goto out_set_next;
6175 : }
6176 : }
6177 :
6178 : /*
6179 : * For each thread: do the regular task pick and find the max prio task
6180 : * amongst them.
6181 : *
6182 : * Tie-break prio towards the current CPU
6183 : */
6184 : for_each_cpu_wrap(i, smt_mask, cpu) {
6185 : rq_i = cpu_rq(i);
6186 :
6187 : /*
6188 : * Current cpu always has its clock updated on entrance to
6189 : * pick_next_task(). If the current cpu is not the core,
6190 : * the core may also have been updated above.
6191 : */
6192 : if (i != cpu && (rq_i != rq->core || !core_clock_updated))
6193 : update_rq_clock(rq_i);
6194 :
6195 : p = rq_i->core_pick = pick_task(rq_i);
6196 : if (!max || prio_less(max, p, fi_before))
6197 : max = p;
6198 : }
6199 :
6200 : cookie = rq->core->core_cookie = max->core_cookie;
6201 :
6202 : /*
6203 : * For each thread: try and find a runnable task that matches @max or
6204 : * force idle.
6205 : */
6206 : for_each_cpu(i, smt_mask) {
6207 : rq_i = cpu_rq(i);
6208 : p = rq_i->core_pick;
6209 :
6210 : if (!cookie_equals(p, cookie)) {
6211 : p = NULL;
6212 : if (cookie)
6213 : p = sched_core_find(rq_i, cookie);
6214 : if (!p)
6215 : p = idle_sched_class.pick_task(rq_i);
6216 : }
6217 :
6218 : rq_i->core_pick = p;
6219 :
6220 : if (p == rq_i->idle) {
6221 : if (rq_i->nr_running) {
6222 : rq->core->core_forceidle_count++;
6223 : if (!fi_before)
6224 : rq->core->core_forceidle_seq++;
6225 : }
6226 : } else {
6227 : occ++;
6228 : }
6229 : }
6230 :
6231 : if (schedstat_enabled() && rq->core->core_forceidle_count) {
6232 : rq->core->core_forceidle_start = rq_clock(rq->core);
6233 : rq->core->core_forceidle_occupation = occ;
6234 : }
6235 :
6236 : rq->core->core_pick_seq = rq->core->core_task_seq;
6237 : next = rq->core_pick;
6238 : rq->core_sched_seq = rq->core->core_pick_seq;
6239 :
6240 : /* Something should have been selected for current CPU */
6241 : WARN_ON_ONCE(!next);
6242 :
6243 : /*
6244 : * Reschedule siblings
6245 : *
6246 : * NOTE: L1TF -- at this point we're no longer running the old task and
6247 : * sending an IPI (below) ensures the sibling will no longer be running
6248 : * their task. This ensures there is no inter-sibling overlap between
6249 : * non-matching user state.
6250 : */
6251 : for_each_cpu(i, smt_mask) {
6252 : rq_i = cpu_rq(i);
6253 :
6254 : /*
6255 : * An online sibling might have gone offline before a task
6256 : * could be picked for it, or it might be offline but later
6257 : * happen to come online, but its too late and nothing was
6258 : * picked for it. That's Ok - it will pick tasks for itself,
6259 : * so ignore it.
6260 : */
6261 : if (!rq_i->core_pick)
6262 : continue;
6263 :
6264 : /*
6265 : * Update for new !FI->FI transitions, or if continuing to be in !FI:
6266 : * fi_before fi update?
6267 : * 0 0 1
6268 : * 0 1 1
6269 : * 1 0 1
6270 : * 1 1 0
6271 : */
6272 : if (!(fi_before && rq->core->core_forceidle_count))
6273 : task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6274 :
6275 : rq_i->core_pick->core_occupation = occ;
6276 :
6277 : if (i == cpu) {
6278 : rq_i->core_pick = NULL;
6279 : continue;
6280 : }
6281 :
6282 : /* Did we break L1TF mitigation requirements? */
6283 : WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6284 :
6285 : if (rq_i->curr == rq_i->core_pick) {
6286 : rq_i->core_pick = NULL;
6287 : continue;
6288 : }
6289 :
6290 : resched_curr(rq_i);
6291 : }
6292 :
6293 : out_set_next:
6294 : set_next_task(rq, next);
6295 : out:
6296 : if (rq->core->core_forceidle_count && next == rq->idle)
6297 : queue_core_balance(rq);
6298 :
6299 : return next;
6300 : }
6301 :
6302 : static bool try_steal_cookie(int this, int that)
6303 : {
6304 : struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6305 : struct task_struct *p;
6306 : unsigned long cookie;
6307 : bool success = false;
6308 :
6309 : local_irq_disable();
6310 : double_rq_lock(dst, src);
6311 :
6312 : cookie = dst->core->core_cookie;
6313 : if (!cookie)
6314 : goto unlock;
6315 :
6316 : if (dst->curr != dst->idle)
6317 : goto unlock;
6318 :
6319 : p = sched_core_find(src, cookie);
6320 : if (!p)
6321 : goto unlock;
6322 :
6323 : do {
6324 : if (p == src->core_pick || p == src->curr)
6325 : goto next;
6326 :
6327 : if (!is_cpu_allowed(p, this))
6328 : goto next;
6329 :
6330 : if (p->core_occupation > dst->idle->core_occupation)
6331 : goto next;
6332 : /*
6333 : * sched_core_find() and sched_core_next() will ensure that task @p
6334 : * is not throttled now, we also need to check whether the runqueue
6335 : * of the destination CPU is being throttled.
6336 : */
6337 : if (sched_task_is_throttled(p, this))
6338 : goto next;
6339 :
6340 : deactivate_task(src, p, 0);
6341 : set_task_cpu(p, this);
6342 : activate_task(dst, p, 0);
6343 :
6344 : resched_curr(dst);
6345 :
6346 : success = true;
6347 : break;
6348 :
6349 : next:
6350 : p = sched_core_next(p, cookie);
6351 : } while (p);
6352 :
6353 : unlock:
6354 : double_rq_unlock(dst, src);
6355 : local_irq_enable();
6356 :
6357 : return success;
6358 : }
6359 :
6360 : static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6361 : {
6362 : int i;
6363 :
6364 : for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
6365 : if (i == cpu)
6366 : continue;
6367 :
6368 : if (need_resched())
6369 : break;
6370 :
6371 : if (try_steal_cookie(cpu, i))
6372 : return true;
6373 : }
6374 :
6375 : return false;
6376 : }
6377 :
6378 : static void sched_core_balance(struct rq *rq)
6379 : {
6380 : struct sched_domain *sd;
6381 : int cpu = cpu_of(rq);
6382 :
6383 : preempt_disable();
6384 : rcu_read_lock();
6385 : raw_spin_rq_unlock_irq(rq);
6386 : for_each_domain(cpu, sd) {
6387 : if (need_resched())
6388 : break;
6389 :
6390 : if (steal_cookie_task(cpu, sd))
6391 : break;
6392 : }
6393 : raw_spin_rq_lock_irq(rq);
6394 : rcu_read_unlock();
6395 : preempt_enable();
6396 : }
6397 :
6398 : static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
6399 :
6400 : static void queue_core_balance(struct rq *rq)
6401 : {
6402 : if (!sched_core_enabled(rq))
6403 : return;
6404 :
6405 : if (!rq->core->core_cookie)
6406 : return;
6407 :
6408 : if (!rq->nr_running) /* not forced idle */
6409 : return;
6410 :
6411 : queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6412 : }
6413 :
6414 : static void sched_core_cpu_starting(unsigned int cpu)
6415 : {
6416 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6417 : struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6418 : unsigned long flags;
6419 : int t;
6420 :
6421 : sched_core_lock(cpu, &flags);
6422 :
6423 : WARN_ON_ONCE(rq->core != rq);
6424 :
6425 : /* if we're the first, we'll be our own leader */
6426 : if (cpumask_weight(smt_mask) == 1)
6427 : goto unlock;
6428 :
6429 : /* find the leader */
6430 : for_each_cpu(t, smt_mask) {
6431 : if (t == cpu)
6432 : continue;
6433 : rq = cpu_rq(t);
6434 : if (rq->core == rq) {
6435 : core_rq = rq;
6436 : break;
6437 : }
6438 : }
6439 :
6440 : if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6441 : goto unlock;
6442 :
6443 : /* install and validate core_rq */
6444 : for_each_cpu(t, smt_mask) {
6445 : rq = cpu_rq(t);
6446 :
6447 : if (t == cpu)
6448 : rq->core = core_rq;
6449 :
6450 : WARN_ON_ONCE(rq->core != core_rq);
6451 : }
6452 :
6453 : unlock:
6454 : sched_core_unlock(cpu, &flags);
6455 : }
6456 :
6457 : static void sched_core_cpu_deactivate(unsigned int cpu)
6458 : {
6459 : const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6460 : struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6461 : unsigned long flags;
6462 : int t;
6463 :
6464 : sched_core_lock(cpu, &flags);
6465 :
6466 : /* if we're the last man standing, nothing to do */
6467 : if (cpumask_weight(smt_mask) == 1) {
6468 : WARN_ON_ONCE(rq->core != rq);
6469 : goto unlock;
6470 : }
6471 :
6472 : /* if we're not the leader, nothing to do */
6473 : if (rq->core != rq)
6474 : goto unlock;
6475 :
6476 : /* find a new leader */
6477 : for_each_cpu(t, smt_mask) {
6478 : if (t == cpu)
6479 : continue;
6480 : core_rq = cpu_rq(t);
6481 : break;
6482 : }
6483 :
6484 : if (WARN_ON_ONCE(!core_rq)) /* impossible */
6485 : goto unlock;
6486 :
6487 : /* copy the shared state to the new leader */
6488 : core_rq->core_task_seq = rq->core_task_seq;
6489 : core_rq->core_pick_seq = rq->core_pick_seq;
6490 : core_rq->core_cookie = rq->core_cookie;
6491 : core_rq->core_forceidle_count = rq->core_forceidle_count;
6492 : core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6493 : core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6494 :
6495 : /*
6496 : * Accounting edge for forced idle is handled in pick_next_task().
6497 : * Don't need another one here, since the hotplug thread shouldn't
6498 : * have a cookie.
6499 : */
6500 : core_rq->core_forceidle_start = 0;
6501 :
6502 : /* install new leader */
6503 : for_each_cpu(t, smt_mask) {
6504 : rq = cpu_rq(t);
6505 : rq->core = core_rq;
6506 : }
6507 :
6508 : unlock:
6509 : sched_core_unlock(cpu, &flags);
6510 : }
6511 :
6512 : static inline void sched_core_cpu_dying(unsigned int cpu)
6513 : {
6514 : struct rq *rq = cpu_rq(cpu);
6515 :
6516 : if (rq->core != rq)
6517 : rq->core = rq;
6518 : }
6519 :
6520 : #else /* !CONFIG_SCHED_CORE */
6521 :
6522 : static inline void sched_core_cpu_starting(unsigned int cpu) {}
6523 : static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6524 : static inline void sched_core_cpu_dying(unsigned int cpu) {}
6525 :
6526 : static struct task_struct *
6527 : pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6528 : {
6529 1032 : return __pick_next_task(rq, prev, rf);
6530 : }
6531 :
6532 : #endif /* CONFIG_SCHED_CORE */
6533 :
6534 : /*
6535 : * Constants for the sched_mode argument of __schedule().
6536 : *
6537 : * The mode argument allows RT enabled kernels to differentiate a
6538 : * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6539 : * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6540 : * optimize the AND operation out and just check for zero.
6541 : */
6542 : #define SM_NONE 0x0
6543 : #define SM_PREEMPT 0x1
6544 : #define SM_RTLOCK_WAIT 0x2
6545 :
6546 : #ifndef CONFIG_PREEMPT_RT
6547 : # define SM_MASK_PREEMPT (~0U)
6548 : #else
6549 : # define SM_MASK_PREEMPT SM_PREEMPT
6550 : #endif
6551 :
6552 : /*
6553 : * __schedule() is the main scheduler function.
6554 : *
6555 : * The main means of driving the scheduler and thus entering this function are:
6556 : *
6557 : * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6558 : *
6559 : * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6560 : * paths. For example, see arch/x86/entry_64.S.
6561 : *
6562 : * To drive preemption between tasks, the scheduler sets the flag in timer
6563 : * interrupt handler scheduler_tick().
6564 : *
6565 : * 3. Wakeups don't really cause entry into schedule(). They add a
6566 : * task to the run-queue and that's it.
6567 : *
6568 : * Now, if the new task added to the run-queue preempts the current
6569 : * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6570 : * called on the nearest possible occasion:
6571 : *
6572 : * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6573 : *
6574 : * - in syscall or exception context, at the next outmost
6575 : * preempt_enable(). (this might be as soon as the wake_up()'s
6576 : * spin_unlock()!)
6577 : *
6578 : * - in IRQ context, return from interrupt-handler to
6579 : * preemptible context
6580 : *
6581 : * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6582 : * then at the next:
6583 : *
6584 : * - cond_resched() call
6585 : * - explicit schedule() call
6586 : * - return from syscall or exception to user-space
6587 : * - return from interrupt-handler to user-space
6588 : *
6589 : * WARNING: must be called with preemption disabled!
6590 : */
6591 1032 : static void __sched notrace __schedule(unsigned int sched_mode)
6592 : {
6593 : struct task_struct *prev, *next;
6594 : unsigned long *switch_count;
6595 : unsigned long prev_state;
6596 : struct rq_flags rf;
6597 : struct rq *rq;
6598 : int cpu;
6599 :
6600 1032 : cpu = smp_processor_id();
6601 1032 : rq = cpu_rq(cpu);
6602 1032 : prev = rq->curr;
6603 :
6604 2064 : schedule_debug(prev, !!sched_mode);
6605 :
6606 : if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6607 : hrtick_clear(rq);
6608 :
6609 : local_irq_disable();
6610 1032 : rcu_note_context_switch(!!sched_mode);
6611 :
6612 : /*
6613 : * Make sure that signal_pending_state()->signal_pending() below
6614 : * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6615 : * done by the caller to avoid the race with signal_wake_up():
6616 : *
6617 : * __set_current_state(@state) signal_wake_up()
6618 : * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6619 : * wake_up_state(p, state)
6620 : * LOCK rq->lock LOCK p->pi_state
6621 : * smp_mb__after_spinlock() smp_mb__after_spinlock()
6622 : * if (signal_pending_state()) if (p->state & @state)
6623 : *
6624 : * Also, the membarrier system call requires a full memory barrier
6625 : * after coming from user-space, before storing to rq->curr.
6626 : */
6627 1032 : rq_lock(rq, &rf);
6628 : smp_mb__after_spinlock();
6629 :
6630 : /* Promote REQ to ACT */
6631 1032 : rq->clock_update_flags <<= 1;
6632 1032 : update_rq_clock(rq);
6633 :
6634 1032 : switch_count = &prev->nivcsw;
6635 :
6636 : /*
6637 : * We must load prev->state once (task_struct::state is volatile), such
6638 : * that we form a control dependency vs deactivate_task() below.
6639 : */
6640 1032 : prev_state = READ_ONCE(prev->__state);
6641 1032 : if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6642 1029 : if (signal_pending_state(prev_state, prev)) {
6643 0 : WRITE_ONCE(prev->__state, TASK_RUNNING);
6644 : } else {
6645 1029 : prev->sched_contributes_to_load =
6646 : (prev_state & TASK_UNINTERRUPTIBLE) &&
6647 1029 : !(prev_state & TASK_NOLOAD) &&
6648 : !(prev_state & TASK_FROZEN);
6649 :
6650 1029 : if (prev->sched_contributes_to_load)
6651 510 : rq->nr_uninterruptible++;
6652 :
6653 : /*
6654 : * __schedule() ttwu()
6655 : * prev_state = prev->state; if (p->on_rq && ...)
6656 : * if (prev_state) goto out;
6657 : * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6658 : * p->state = TASK_WAKING
6659 : *
6660 : * Where __schedule() and ttwu() have matching control dependencies.
6661 : *
6662 : * After this, schedule() must not care about p->state any more.
6663 : */
6664 1029 : deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6665 :
6666 1029 : if (prev->in_iowait) {
6667 0 : atomic_inc(&rq->nr_iowait);
6668 : delayacct_blkio_start();
6669 : }
6670 : }
6671 1029 : switch_count = &prev->nvcsw;
6672 : }
6673 :
6674 1032 : next = pick_next_task(rq, prev, &rf);
6675 1032 : clear_tsk_need_resched(prev);
6676 : clear_preempt_need_resched();
6677 : #ifdef CONFIG_SCHED_DEBUG
6678 : rq->last_seen_need_resched_ns = 0;
6679 : #endif
6680 :
6681 1032 : if (likely(prev != next)) {
6682 1031 : rq->nr_switches++;
6683 : /*
6684 : * RCU users of rcu_dereference(rq->curr) may not see
6685 : * changes to task_struct made by pick_next_task().
6686 : */
6687 1031 : RCU_INIT_POINTER(rq->curr, next);
6688 : /*
6689 : * The membarrier system call requires each architecture
6690 : * to have a full memory barrier after updating
6691 : * rq->curr, before returning to user-space.
6692 : *
6693 : * Here are the schemes providing that barrier on the
6694 : * various architectures:
6695 : * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6696 : * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6697 : * - finish_lock_switch() for weakly-ordered
6698 : * architectures where spin_unlock is a full barrier,
6699 : * - switch_to() for arm64 (weakly-ordered, spin_unlock
6700 : * is a RELEASE barrier),
6701 : */
6702 1031 : ++*switch_count;
6703 :
6704 1031 : migrate_disable_switch(rq, prev);
6705 1031 : psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6706 :
6707 1031 : trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6708 :
6709 : /* Also unlocks the rq: */
6710 856 : rq = context_switch(rq, prev, next, &rf);
6711 : } else {
6712 1 : rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6713 :
6714 1 : rq_unpin_lock(rq, &rf);
6715 1 : __balance_callbacks(rq);
6716 1 : raw_spin_rq_unlock_irq(rq);
6717 : }
6718 857 : }
6719 :
6720 160 : void __noreturn do_task_dead(void)
6721 : {
6722 : /* Causes final put_task_struct in finish_task_switch(): */
6723 800 : set_special_state(TASK_DEAD);
6724 :
6725 : /* Tell freezer to ignore us: */
6726 160 : current->flags |= PF_NOFREEZE;
6727 :
6728 160 : __schedule(SM_NONE);
6729 0 : BUG();
6730 :
6731 : /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6732 : for (;;)
6733 : cpu_relax();
6734 : }
6735 :
6736 870 : static inline void sched_submit_work(struct task_struct *tsk)
6737 : {
6738 : unsigned int task_flags;
6739 :
6740 870 : if (task_is_running(tsk))
6741 : return;
6742 :
6743 869 : task_flags = tsk->flags;
6744 : /*
6745 : * If a worker goes to sleep, notify and ask workqueue whether it
6746 : * wants to wake up a task to maintain concurrency.
6747 : */
6748 869 : if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6749 24 : if (task_flags & PF_WQ_WORKER)
6750 24 : wq_worker_sleeping(tsk);
6751 : else
6752 0 : io_wq_worker_sleeping(tsk);
6753 : }
6754 :
6755 : /*
6756 : * spinlock and rwlock must not flush block requests. This will
6757 : * deadlock if the callback attempts to acquire a lock which is
6758 : * already acquired.
6759 : */
6760 869 : SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6761 :
6762 : /*
6763 : * If we are going to sleep and we have plugged IO queued,
6764 : * make sure to submit it to avoid deadlocks.
6765 : */
6766 869 : blk_flush_plug(tsk->plug, true);
6767 : }
6768 :
6769 855 : static void sched_update_worker(struct task_struct *tsk)
6770 : {
6771 855 : if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6772 15 : if (tsk->flags & PF_WQ_WORKER)
6773 15 : wq_worker_running(tsk);
6774 : else
6775 0 : io_wq_worker_running(tsk);
6776 : }
6777 855 : }
6778 :
6779 870 : asmlinkage __visible void __sched schedule(void)
6780 : {
6781 870 : struct task_struct *tsk = current;
6782 :
6783 870 : sched_submit_work(tsk);
6784 : do {
6785 870 : preempt_disable();
6786 870 : __schedule(SM_NONE);
6787 855 : sched_preempt_enable_no_resched();
6788 855 : } while (need_resched());
6789 855 : sched_update_worker(tsk);
6790 855 : }
6791 : EXPORT_SYMBOL(schedule);
6792 :
6793 : /*
6794 : * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6795 : * state (have scheduled out non-voluntarily) by making sure that all
6796 : * tasks have either left the run queue or have gone into user space.
6797 : * As idle tasks do not do either, they must not ever be preempted
6798 : * (schedule out non-voluntarily).
6799 : *
6800 : * schedule_idle() is similar to schedule_preempt_disable() except that it
6801 : * never enables preemption because it does not call sched_submit_work().
6802 : */
6803 0 : void __sched schedule_idle(void)
6804 : {
6805 : /*
6806 : * As this skips calling sched_submit_work(), which the idle task does
6807 : * regardless because that function is a nop when the task is in a
6808 : * TASK_RUNNING state, make sure this isn't used someplace that the
6809 : * current task can be in any other state. Note, idle is always in the
6810 : * TASK_RUNNING state.
6811 : */
6812 0 : WARN_ON_ONCE(current->__state);
6813 : do {
6814 0 : __schedule(SM_NONE);
6815 0 : } while (need_resched());
6816 0 : }
6817 :
6818 : #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6819 : asmlinkage __visible void __sched schedule_user(void)
6820 : {
6821 : /*
6822 : * If we come here after a random call to set_need_resched(),
6823 : * or we have been woken up remotely but the IPI has not yet arrived,
6824 : * we haven't yet exited the RCU idle mode. Do it here manually until
6825 : * we find a better solution.
6826 : *
6827 : * NB: There are buggy callers of this function. Ideally we
6828 : * should warn if prev_state != CONTEXT_USER, but that will trigger
6829 : * too frequently to make sense yet.
6830 : */
6831 : enum ctx_state prev_state = exception_enter();
6832 : schedule();
6833 : exception_exit(prev_state);
6834 : }
6835 : #endif
6836 :
6837 : /**
6838 : * schedule_preempt_disabled - called with preemption disabled
6839 : *
6840 : * Returns with preemption disabled. Note: preempt_count must be 1
6841 : */
6842 175 : void __sched schedule_preempt_disabled(void)
6843 : {
6844 175 : sched_preempt_enable_no_resched();
6845 175 : schedule();
6846 174 : preempt_disable();
6847 174 : }
6848 :
6849 : #ifdef CONFIG_PREEMPT_RT
6850 : void __sched notrace schedule_rtlock(void)
6851 : {
6852 : do {
6853 : preempt_disable();
6854 : __schedule(SM_RTLOCK_WAIT);
6855 : sched_preempt_enable_no_resched();
6856 : } while (need_resched());
6857 : }
6858 : NOKPROBE_SYMBOL(schedule_rtlock);
6859 : #endif
6860 :
6861 : static void __sched notrace preempt_schedule_common(void)
6862 : {
6863 : do {
6864 : /*
6865 : * Because the function tracer can trace preempt_count_sub()
6866 : * and it also uses preempt_enable/disable_notrace(), if
6867 : * NEED_RESCHED is set, the preempt_enable_notrace() called
6868 : * by the function tracer will call this function again and
6869 : * cause infinite recursion.
6870 : *
6871 : * Preemption must be disabled here before the function
6872 : * tracer can trace. Break up preempt_disable() into two
6873 : * calls. One to disable preemption without fear of being
6874 : * traced. The other to still record the preemption latency,
6875 : * which can also be traced by the function tracer.
6876 : */
6877 2 : preempt_disable_notrace();
6878 2 : preempt_latency_start(1);
6879 2 : __schedule(SM_PREEMPT);
6880 2 : preempt_latency_stop(1);
6881 2 : preempt_enable_no_resched_notrace();
6882 :
6883 : /*
6884 : * Check again in case we missed a preemption opportunity
6885 : * between schedule and now.
6886 : */
6887 2 : } while (need_resched());
6888 : }
6889 :
6890 : #ifdef CONFIG_PREEMPTION
6891 : /*
6892 : * This is the entry point to schedule() from in-kernel preemption
6893 : * off of preempt_enable.
6894 : */
6895 : asmlinkage __visible void __sched notrace preempt_schedule(void)
6896 : {
6897 : /*
6898 : * If there is a non-zero preempt_count or interrupts are disabled,
6899 : * we do not want to preempt the current task. Just return..
6900 : */
6901 : if (likely(!preemptible()))
6902 : return;
6903 : preempt_schedule_common();
6904 : }
6905 : NOKPROBE_SYMBOL(preempt_schedule);
6906 : EXPORT_SYMBOL(preempt_schedule);
6907 :
6908 : #ifdef CONFIG_PREEMPT_DYNAMIC
6909 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6910 : #ifndef preempt_schedule_dynamic_enabled
6911 : #define preempt_schedule_dynamic_enabled preempt_schedule
6912 : #define preempt_schedule_dynamic_disabled NULL
6913 : #endif
6914 : DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6915 : EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6916 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6917 : static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6918 : void __sched notrace dynamic_preempt_schedule(void)
6919 : {
6920 : if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6921 : return;
6922 : preempt_schedule();
6923 : }
6924 : NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6925 : EXPORT_SYMBOL(dynamic_preempt_schedule);
6926 : #endif
6927 : #endif
6928 :
6929 : /**
6930 : * preempt_schedule_notrace - preempt_schedule called by tracing
6931 : *
6932 : * The tracing infrastructure uses preempt_enable_notrace to prevent
6933 : * recursion and tracing preempt enabling caused by the tracing
6934 : * infrastructure itself. But as tracing can happen in areas coming
6935 : * from userspace or just about to enter userspace, a preempt enable
6936 : * can occur before user_exit() is called. This will cause the scheduler
6937 : * to be called when the system is still in usermode.
6938 : *
6939 : * To prevent this, the preempt_enable_notrace will use this function
6940 : * instead of preempt_schedule() to exit user context if needed before
6941 : * calling the scheduler.
6942 : */
6943 : asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6944 : {
6945 : enum ctx_state prev_ctx;
6946 :
6947 : if (likely(!preemptible()))
6948 : return;
6949 :
6950 : do {
6951 : /*
6952 : * Because the function tracer can trace preempt_count_sub()
6953 : * and it also uses preempt_enable/disable_notrace(), if
6954 : * NEED_RESCHED is set, the preempt_enable_notrace() called
6955 : * by the function tracer will call this function again and
6956 : * cause infinite recursion.
6957 : *
6958 : * Preemption must be disabled here before the function
6959 : * tracer can trace. Break up preempt_disable() into two
6960 : * calls. One to disable preemption without fear of being
6961 : * traced. The other to still record the preemption latency,
6962 : * which can also be traced by the function tracer.
6963 : */
6964 : preempt_disable_notrace();
6965 : preempt_latency_start(1);
6966 : /*
6967 : * Needs preempt disabled in case user_exit() is traced
6968 : * and the tracer calls preempt_enable_notrace() causing
6969 : * an infinite recursion.
6970 : */
6971 : prev_ctx = exception_enter();
6972 : __schedule(SM_PREEMPT);
6973 : exception_exit(prev_ctx);
6974 :
6975 : preempt_latency_stop(1);
6976 : preempt_enable_no_resched_notrace();
6977 : } while (need_resched());
6978 : }
6979 : EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6980 :
6981 : #ifdef CONFIG_PREEMPT_DYNAMIC
6982 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6983 : #ifndef preempt_schedule_notrace_dynamic_enabled
6984 : #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6985 : #define preempt_schedule_notrace_dynamic_disabled NULL
6986 : #endif
6987 : DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6988 : EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6989 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6990 : static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6991 : void __sched notrace dynamic_preempt_schedule_notrace(void)
6992 : {
6993 : if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6994 : return;
6995 : preempt_schedule_notrace();
6996 : }
6997 : NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6998 : EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6999 : #endif
7000 : #endif
7001 :
7002 : #endif /* CONFIG_PREEMPTION */
7003 :
7004 : /*
7005 : * This is the entry point to schedule() from kernel preemption
7006 : * off of irq context.
7007 : * Note, that this is called and return with irqs disabled. This will
7008 : * protect us against recursive calling from irq.
7009 : */
7010 0 : asmlinkage __visible void __sched preempt_schedule_irq(void)
7011 : {
7012 : enum ctx_state prev_state;
7013 :
7014 : /* Catch callers which need to be fixed */
7015 0 : BUG_ON(preempt_count() || !irqs_disabled());
7016 :
7017 : prev_state = exception_enter();
7018 :
7019 : do {
7020 0 : preempt_disable();
7021 : local_irq_enable();
7022 0 : __schedule(SM_PREEMPT);
7023 : local_irq_disable();
7024 0 : sched_preempt_enable_no_resched();
7025 0 : } while (need_resched());
7026 :
7027 : exception_exit(prev_state);
7028 0 : }
7029 :
7030 0 : int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
7031 : void *key)
7032 : {
7033 0 : WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
7034 0 : return try_to_wake_up(curr->private, mode, wake_flags);
7035 : }
7036 : EXPORT_SYMBOL(default_wake_function);
7037 :
7038 : static void __setscheduler_prio(struct task_struct *p, int prio)
7039 : {
7040 0 : if (dl_prio(prio))
7041 0 : p->sched_class = &dl_sched_class;
7042 0 : else if (rt_prio(prio))
7043 0 : p->sched_class = &rt_sched_class;
7044 : else
7045 0 : p->sched_class = &fair_sched_class;
7046 :
7047 0 : p->prio = prio;
7048 : }
7049 :
7050 : #ifdef CONFIG_RT_MUTEXES
7051 :
7052 : static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
7053 : {
7054 0 : if (pi_task)
7055 0 : prio = min(prio, pi_task->prio);
7056 :
7057 : return prio;
7058 : }
7059 :
7060 : static inline int rt_effective_prio(struct task_struct *p, int prio)
7061 : {
7062 0 : struct task_struct *pi_task = rt_mutex_get_top_task(p);
7063 :
7064 0 : return __rt_effective_prio(pi_task, prio);
7065 : }
7066 :
7067 : /*
7068 : * rt_mutex_setprio - set the current priority of a task
7069 : * @p: task to boost
7070 : * @pi_task: donor task
7071 : *
7072 : * This function changes the 'effective' priority of a task. It does
7073 : * not touch ->normal_prio like __setscheduler().
7074 : *
7075 : * Used by the rt_mutex code to implement priority inheritance
7076 : * logic. Call site only calls if the priority of the task changed.
7077 : */
7078 0 : void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
7079 : {
7080 0 : int prio, oldprio, queued, running, queue_flag =
7081 : DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7082 : const struct sched_class *prev_class;
7083 : struct rq_flags rf;
7084 : struct rq *rq;
7085 :
7086 : /* XXX used to be waiter->prio, not waiter->task->prio */
7087 0 : prio = __rt_effective_prio(pi_task, p->normal_prio);
7088 :
7089 : /*
7090 : * If nothing changed; bail early.
7091 : */
7092 0 : if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
7093 : return;
7094 :
7095 0 : rq = __task_rq_lock(p, &rf);
7096 0 : update_rq_clock(rq);
7097 : /*
7098 : * Set under pi_lock && rq->lock, such that the value can be used under
7099 : * either lock.
7100 : *
7101 : * Note that there is loads of tricky to make this pointer cache work
7102 : * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
7103 : * ensure a task is de-boosted (pi_task is set to NULL) before the
7104 : * task is allowed to run again (and can exit). This ensures the pointer
7105 : * points to a blocked task -- which guarantees the task is present.
7106 : */
7107 0 : p->pi_top_task = pi_task;
7108 :
7109 : /*
7110 : * For FIFO/RR we only need to set prio, if that matches we're done.
7111 : */
7112 0 : if (prio == p->prio && !dl_prio(prio))
7113 : goto out_unlock;
7114 :
7115 : /*
7116 : * Idle task boosting is a nono in general. There is one
7117 : * exception, when PREEMPT_RT and NOHZ is active:
7118 : *
7119 : * The idle task calls get_next_timer_interrupt() and holds
7120 : * the timer wheel base->lock on the CPU and another CPU wants
7121 : * to access the timer (probably to cancel it). We can safely
7122 : * ignore the boosting request, as the idle CPU runs this code
7123 : * with interrupts disabled and will complete the lock
7124 : * protected section without being interrupted. So there is no
7125 : * real need to boost.
7126 : */
7127 0 : if (unlikely(p == rq->idle)) {
7128 0 : WARN_ON(p != rq->curr);
7129 0 : WARN_ON(p->pi_blocked_on);
7130 : goto out_unlock;
7131 : }
7132 :
7133 0 : trace_sched_pi_setprio(p, pi_task);
7134 0 : oldprio = p->prio;
7135 :
7136 0 : if (oldprio == prio)
7137 0 : queue_flag &= ~DEQUEUE_MOVE;
7138 :
7139 0 : prev_class = p->sched_class;
7140 0 : queued = task_on_rq_queued(p);
7141 0 : running = task_current(rq, p);
7142 0 : if (queued)
7143 0 : dequeue_task(rq, p, queue_flag);
7144 0 : if (running)
7145 0 : put_prev_task(rq, p);
7146 :
7147 : /*
7148 : * Boosting condition are:
7149 : * 1. -rt task is running and holds mutex A
7150 : * --> -dl task blocks on mutex A
7151 : *
7152 : * 2. -dl task is running and holds mutex A
7153 : * --> -dl task blocks on mutex A and could preempt the
7154 : * running task
7155 : */
7156 0 : if (dl_prio(prio)) {
7157 0 : if (!dl_prio(p->normal_prio) ||
7158 0 : (pi_task && dl_prio(pi_task->prio) &&
7159 0 : dl_entity_preempt(&pi_task->dl, &p->dl))) {
7160 0 : p->dl.pi_se = pi_task->dl.pi_se;
7161 0 : queue_flag |= ENQUEUE_REPLENISH;
7162 : } else {
7163 0 : p->dl.pi_se = &p->dl;
7164 : }
7165 0 : } else if (rt_prio(prio)) {
7166 0 : if (dl_prio(oldprio))
7167 0 : p->dl.pi_se = &p->dl;
7168 0 : if (oldprio < prio)
7169 0 : queue_flag |= ENQUEUE_HEAD;
7170 : } else {
7171 0 : if (dl_prio(oldprio))
7172 0 : p->dl.pi_se = &p->dl;
7173 0 : if (rt_prio(oldprio))
7174 0 : p->rt.timeout = 0;
7175 : }
7176 :
7177 0 : __setscheduler_prio(p, prio);
7178 :
7179 0 : if (queued)
7180 0 : enqueue_task(rq, p, queue_flag);
7181 0 : if (running)
7182 : set_next_task(rq, p);
7183 :
7184 0 : check_class_changed(rq, p, prev_class, oldprio);
7185 : out_unlock:
7186 : /* Avoid rq from going away on us: */
7187 0 : preempt_disable();
7188 :
7189 0 : rq_unpin_lock(rq, &rf);
7190 0 : __balance_callbacks(rq);
7191 0 : raw_spin_rq_unlock(rq);
7192 :
7193 0 : preempt_enable();
7194 : }
7195 : #else
7196 : static inline int rt_effective_prio(struct task_struct *p, int prio)
7197 : {
7198 : return prio;
7199 : }
7200 : #endif
7201 :
7202 9 : void set_user_nice(struct task_struct *p, long nice)
7203 : {
7204 : bool queued, running;
7205 : int old_prio;
7206 : struct rq_flags rf;
7207 : struct rq *rq;
7208 :
7209 18 : if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
7210 4 : return;
7211 : /*
7212 : * We have to be careful, if called from sys_setpriority(),
7213 : * the task might be in the middle of scheduling on another CPU.
7214 : */
7215 5 : rq = task_rq_lock(p, &rf);
7216 5 : update_rq_clock(rq);
7217 :
7218 : /*
7219 : * The RT priorities are set via sched_setscheduler(), but we still
7220 : * allow the 'normal' nice value to be set - but as expected
7221 : * it won't have any effect on scheduling until the task is
7222 : * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7223 : */
7224 15 : if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7225 0 : p->static_prio = NICE_TO_PRIO(nice);
7226 0 : goto out_unlock;
7227 : }
7228 5 : queued = task_on_rq_queued(p);
7229 5 : running = task_current(rq, p);
7230 5 : if (queued)
7231 : dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7232 5 : if (running)
7233 4 : put_prev_task(rq, p);
7234 :
7235 5 : p->static_prio = NICE_TO_PRIO(nice);
7236 5 : set_load_weight(p, true);
7237 5 : old_prio = p->prio;
7238 5 : p->prio = effective_prio(p);
7239 :
7240 5 : if (queued)
7241 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7242 5 : if (running)
7243 : set_next_task(rq, p);
7244 :
7245 : /*
7246 : * If the task increased its priority or is running and
7247 : * lowered its priority, then reschedule its CPU:
7248 : */
7249 5 : p->sched_class->prio_changed(rq, p, old_prio);
7250 :
7251 : out_unlock:
7252 10 : task_rq_unlock(rq, p, &rf);
7253 : }
7254 : EXPORT_SYMBOL(set_user_nice);
7255 :
7256 : /*
7257 : * is_nice_reduction - check if nice value is an actual reduction
7258 : *
7259 : * Similar to can_nice() but does not perform a capability check.
7260 : *
7261 : * @p: task
7262 : * @nice: nice value
7263 : */
7264 : static bool is_nice_reduction(const struct task_struct *p, const int nice)
7265 : {
7266 : /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7267 0 : int nice_rlim = nice_to_rlimit(nice);
7268 :
7269 0 : return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7270 : }
7271 :
7272 : /*
7273 : * can_nice - check if a task can reduce its nice value
7274 : * @p: task
7275 : * @nice: nice value
7276 : */
7277 0 : int can_nice(const struct task_struct *p, const int nice)
7278 : {
7279 0 : return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7280 : }
7281 :
7282 : #ifdef __ARCH_WANT_SYS_NICE
7283 :
7284 : /*
7285 : * sys_nice - change the priority of the current process.
7286 : * @increment: priority increment
7287 : *
7288 : * sys_setpriority is a more generic, but much slower function that
7289 : * does similar things.
7290 : */
7291 0 : SYSCALL_DEFINE1(nice, int, increment)
7292 : {
7293 : long nice, retval;
7294 :
7295 : /*
7296 : * Setpriority might change our priority at the same moment.
7297 : * We don't have to worry. Conceptually one call occurs first
7298 : * and we have a single winner.
7299 : */
7300 0 : increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7301 0 : nice = task_nice(current) + increment;
7302 :
7303 0 : nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7304 0 : if (increment < 0 && !can_nice(current, nice))
7305 : return -EPERM;
7306 :
7307 0 : retval = security_task_setnice(current, nice);
7308 0 : if (retval)
7309 : return retval;
7310 :
7311 0 : set_user_nice(current, nice);
7312 0 : return 0;
7313 : }
7314 :
7315 : #endif
7316 :
7317 : /**
7318 : * task_prio - return the priority value of a given task.
7319 : * @p: the task in question.
7320 : *
7321 : * Return: The priority value as seen by users in /proc.
7322 : *
7323 : * sched policy return value kernel prio user prio/nice
7324 : *
7325 : * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7326 : * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7327 : * deadline -101 -1 0
7328 : */
7329 0 : int task_prio(const struct task_struct *p)
7330 : {
7331 0 : return p->prio - MAX_RT_PRIO;
7332 : }
7333 :
7334 : /**
7335 : * idle_cpu - is a given CPU idle currently?
7336 : * @cpu: the processor in question.
7337 : *
7338 : * Return: 1 if the CPU is currently idle. 0 otherwise.
7339 : */
7340 0 : int idle_cpu(int cpu)
7341 : {
7342 0 : struct rq *rq = cpu_rq(cpu);
7343 :
7344 0 : if (rq->curr != rq->idle)
7345 : return 0;
7346 :
7347 0 : if (rq->nr_running)
7348 : return 0;
7349 :
7350 : #ifdef CONFIG_SMP
7351 : if (rq->ttwu_pending)
7352 : return 0;
7353 : #endif
7354 :
7355 0 : return 1;
7356 : }
7357 :
7358 : /**
7359 : * available_idle_cpu - is a given CPU idle for enqueuing work.
7360 : * @cpu: the CPU in question.
7361 : *
7362 : * Return: 1 if the CPU is currently idle. 0 otherwise.
7363 : */
7364 0 : int available_idle_cpu(int cpu)
7365 : {
7366 0 : if (!idle_cpu(cpu))
7367 : return 0;
7368 :
7369 0 : if (vcpu_is_preempted(cpu))
7370 : return 0;
7371 :
7372 0 : return 1;
7373 : }
7374 :
7375 : /**
7376 : * idle_task - return the idle task for a given CPU.
7377 : * @cpu: the processor in question.
7378 : *
7379 : * Return: The idle task for the CPU @cpu.
7380 : */
7381 0 : struct task_struct *idle_task(int cpu)
7382 : {
7383 0 : return cpu_rq(cpu)->idle;
7384 : }
7385 :
7386 : #ifdef CONFIG_SMP
7387 : /*
7388 : * This function computes an effective utilization for the given CPU, to be
7389 : * used for frequency selection given the linear relation: f = u * f_max.
7390 : *
7391 : * The scheduler tracks the following metrics:
7392 : *
7393 : * cpu_util_{cfs,rt,dl,irq}()
7394 : * cpu_bw_dl()
7395 : *
7396 : * Where the cfs,rt and dl util numbers are tracked with the same metric and
7397 : * synchronized windows and are thus directly comparable.
7398 : *
7399 : * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7400 : * which excludes things like IRQ and steal-time. These latter are then accrued
7401 : * in the irq utilization.
7402 : *
7403 : * The DL bandwidth number otoh is not a measured metric but a value computed
7404 : * based on the task model parameters and gives the minimal utilization
7405 : * required to meet deadlines.
7406 : */
7407 : unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7408 : enum cpu_util_type type,
7409 : struct task_struct *p)
7410 : {
7411 : unsigned long dl_util, util, irq, max;
7412 : struct rq *rq = cpu_rq(cpu);
7413 :
7414 : max = arch_scale_cpu_capacity(cpu);
7415 :
7416 : if (!uclamp_is_used() &&
7417 : type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7418 : return max;
7419 : }
7420 :
7421 : /*
7422 : * Early check to see if IRQ/steal time saturates the CPU, can be
7423 : * because of inaccuracies in how we track these -- see
7424 : * update_irq_load_avg().
7425 : */
7426 : irq = cpu_util_irq(rq);
7427 : if (unlikely(irq >= max))
7428 : return max;
7429 :
7430 : /*
7431 : * Because the time spend on RT/DL tasks is visible as 'lost' time to
7432 : * CFS tasks and we use the same metric to track the effective
7433 : * utilization (PELT windows are synchronized) we can directly add them
7434 : * to obtain the CPU's actual utilization.
7435 : *
7436 : * CFS and RT utilization can be boosted or capped, depending on
7437 : * utilization clamp constraints requested by currently RUNNABLE
7438 : * tasks.
7439 : * When there are no CFS RUNNABLE tasks, clamps are released and
7440 : * frequency will be gracefully reduced with the utilization decay.
7441 : */
7442 : util = util_cfs + cpu_util_rt(rq);
7443 : if (type == FREQUENCY_UTIL)
7444 : util = uclamp_rq_util_with(rq, util, p);
7445 :
7446 : dl_util = cpu_util_dl(rq);
7447 :
7448 : /*
7449 : * For frequency selection we do not make cpu_util_dl() a permanent part
7450 : * of this sum because we want to use cpu_bw_dl() later on, but we need
7451 : * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7452 : * that we select f_max when there is no idle time.
7453 : *
7454 : * NOTE: numerical errors or stop class might cause us to not quite hit
7455 : * saturation when we should -- something for later.
7456 : */
7457 : if (util + dl_util >= max)
7458 : return max;
7459 :
7460 : /*
7461 : * OTOH, for energy computation we need the estimated running time, so
7462 : * include util_dl and ignore dl_bw.
7463 : */
7464 : if (type == ENERGY_UTIL)
7465 : util += dl_util;
7466 :
7467 : /*
7468 : * There is still idle time; further improve the number by using the
7469 : * irq metric. Because IRQ/steal time is hidden from the task clock we
7470 : * need to scale the task numbers:
7471 : *
7472 : * max - irq
7473 : * U' = irq + --------- * U
7474 : * max
7475 : */
7476 : util = scale_irq_capacity(util, irq, max);
7477 : util += irq;
7478 :
7479 : /*
7480 : * Bandwidth required by DEADLINE must always be granted while, for
7481 : * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7482 : * to gracefully reduce the frequency when no tasks show up for longer
7483 : * periods of time.
7484 : *
7485 : * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7486 : * bw_dl as requested freq. However, cpufreq is not yet ready for such
7487 : * an interface. So, we only do the latter for now.
7488 : */
7489 : if (type == FREQUENCY_UTIL)
7490 : util += cpu_bw_dl(rq);
7491 :
7492 : return min(max, util);
7493 : }
7494 :
7495 : unsigned long sched_cpu_util(int cpu)
7496 : {
7497 : return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
7498 : }
7499 : #endif /* CONFIG_SMP */
7500 :
7501 : /**
7502 : * find_process_by_pid - find a process with a matching PID value.
7503 : * @pid: the pid in question.
7504 : *
7505 : * The task of @pid, if found. %NULL otherwise.
7506 : */
7507 : static struct task_struct *find_process_by_pid(pid_t pid)
7508 : {
7509 0 : return pid ? find_task_by_vpid(pid) : current;
7510 : }
7511 :
7512 : /*
7513 : * sched_setparam() passes in -1 for its policy, to let the functions
7514 : * it calls know not to change it.
7515 : */
7516 : #define SETPARAM_POLICY -1
7517 :
7518 0 : static void __setscheduler_params(struct task_struct *p,
7519 : const struct sched_attr *attr)
7520 : {
7521 0 : int policy = attr->sched_policy;
7522 :
7523 0 : if (policy == SETPARAM_POLICY)
7524 0 : policy = p->policy;
7525 :
7526 0 : p->policy = policy;
7527 :
7528 0 : if (dl_policy(policy))
7529 0 : __setparam_dl(p, attr);
7530 0 : else if (fair_policy(policy))
7531 0 : p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7532 :
7533 : /*
7534 : * __sched_setscheduler() ensures attr->sched_priority == 0 when
7535 : * !rt_policy. Always setting this ensures that things like
7536 : * getparam()/getattr() don't report silly values for !rt tasks.
7537 : */
7538 0 : p->rt_priority = attr->sched_priority;
7539 0 : p->normal_prio = normal_prio(p);
7540 0 : set_load_weight(p, true);
7541 0 : }
7542 :
7543 : /*
7544 : * Check the target process has a UID that matches the current process's:
7545 : */
7546 : static bool check_same_owner(struct task_struct *p)
7547 : {
7548 0 : const struct cred *cred = current_cred(), *pcred;
7549 : bool match;
7550 :
7551 : rcu_read_lock();
7552 0 : pcred = __task_cred(p);
7553 0 : match = (uid_eq(cred->euid, pcred->euid) ||
7554 0 : uid_eq(cred->euid, pcred->uid));
7555 : rcu_read_unlock();
7556 : return match;
7557 : }
7558 :
7559 : /*
7560 : * Allow unprivileged RT tasks to decrease priority.
7561 : * Only issue a capable test if needed and only once to avoid an audit
7562 : * event on permitted non-privileged operations:
7563 : */
7564 0 : static int user_check_sched_setscheduler(struct task_struct *p,
7565 : const struct sched_attr *attr,
7566 : int policy, int reset_on_fork)
7567 : {
7568 0 : if (fair_policy(policy)) {
7569 0 : if (attr->sched_nice < task_nice(p) &&
7570 0 : !is_nice_reduction(p, attr->sched_nice))
7571 : goto req_priv;
7572 : }
7573 :
7574 0 : if (rt_policy(policy)) {
7575 0 : unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7576 :
7577 : /* Can't set/change the rt policy: */
7578 0 : if (policy != p->policy && !rlim_rtprio)
7579 : goto req_priv;
7580 :
7581 : /* Can't increase priority: */
7582 0 : if (attr->sched_priority > p->rt_priority &&
7583 0 : attr->sched_priority > rlim_rtprio)
7584 : goto req_priv;
7585 : }
7586 :
7587 : /*
7588 : * Can't set/change SCHED_DEADLINE policy at all for now
7589 : * (safest behavior); in the future we would like to allow
7590 : * unprivileged DL tasks to increase their relative deadline
7591 : * or reduce their runtime (both ways reducing utilization)
7592 : */
7593 0 : if (dl_policy(policy))
7594 : goto req_priv;
7595 :
7596 : /*
7597 : * Treat SCHED_IDLE as nice 20. Only allow a switch to
7598 : * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7599 : */
7600 0 : if (task_has_idle_policy(p) && !idle_policy(policy)) {
7601 0 : if (!is_nice_reduction(p, task_nice(p)))
7602 : goto req_priv;
7603 : }
7604 :
7605 : /* Can't change other user's priorities: */
7606 0 : if (!check_same_owner(p))
7607 : goto req_priv;
7608 :
7609 : /* Normal users shall not reset the sched_reset_on_fork flag: */
7610 0 : if (p->sched_reset_on_fork && !reset_on_fork)
7611 : goto req_priv;
7612 :
7613 : return 0;
7614 :
7615 : req_priv:
7616 0 : if (!capable(CAP_SYS_NICE))
7617 : return -EPERM;
7618 :
7619 : return 0;
7620 : }
7621 :
7622 173 : static int __sched_setscheduler(struct task_struct *p,
7623 : const struct sched_attr *attr,
7624 : bool user, bool pi)
7625 : {
7626 173 : int oldpolicy = -1, policy = attr->sched_policy;
7627 : int retval, oldprio, newprio, queued, running;
7628 : const struct sched_class *prev_class;
7629 : struct balance_callback *head;
7630 : struct rq_flags rf;
7631 : int reset_on_fork;
7632 173 : int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7633 : struct rq *rq;
7634 173 : bool cpuset_locked = false;
7635 :
7636 : /* The pi code expects interrupts enabled */
7637 346 : BUG_ON(pi && in_interrupt());
7638 : recheck:
7639 : /* Double check policy once rq lock held: */
7640 173 : if (policy < 0) {
7641 0 : reset_on_fork = p->sched_reset_on_fork;
7642 0 : policy = oldpolicy = p->policy;
7643 : } else {
7644 173 : reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7645 :
7646 173 : if (!valid_policy(policy))
7647 : return -EINVAL;
7648 : }
7649 :
7650 173 : if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7651 : return -EINVAL;
7652 :
7653 : /*
7654 : * Valid priorities for SCHED_FIFO and SCHED_RR are
7655 : * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7656 : * SCHED_BATCH and SCHED_IDLE is 0.
7657 : */
7658 173 : if (attr->sched_priority > MAX_RT_PRIO-1)
7659 : return -EINVAL;
7660 346 : if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7661 173 : (rt_policy(policy) != (attr->sched_priority != 0)))
7662 : return -EINVAL;
7663 :
7664 173 : if (user) {
7665 0 : retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7666 0 : if (retval)
7667 : return retval;
7668 :
7669 0 : if (attr->sched_flags & SCHED_FLAG_SUGOV)
7670 : return -EINVAL;
7671 :
7672 0 : retval = security_task_setscheduler(p);
7673 0 : if (retval)
7674 : return retval;
7675 : }
7676 :
7677 : /* Update task specific "requested" clamps */
7678 173 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7679 : retval = uclamp_validate(p, attr);
7680 : if (retval)
7681 : return retval;
7682 : }
7683 :
7684 : /*
7685 : * SCHED_DEADLINE bandwidth accounting relies on stable cpusets
7686 : * information.
7687 : */
7688 173 : if (dl_policy(policy) || dl_policy(p->policy)) {
7689 : cpuset_locked = true;
7690 : cpuset_lock();
7691 : }
7692 :
7693 : /*
7694 : * Make sure no PI-waiters arrive (or leave) while we are
7695 : * changing the priority of the task:
7696 : *
7697 : * To be able to change p->policy safely, the appropriate
7698 : * runqueue lock must be held.
7699 : */
7700 173 : rq = task_rq_lock(p, &rf);
7701 173 : update_rq_clock(rq);
7702 :
7703 : /*
7704 : * Changing the policy of the stop threads its a very bad idea:
7705 : */
7706 173 : if (p == rq->stop) {
7707 : retval = -EINVAL;
7708 : goto unlock;
7709 : }
7710 :
7711 : /*
7712 : * If not changing anything there's no need to proceed further,
7713 : * but store a possible modification of reset_on_fork.
7714 : */
7715 173 : if (unlikely(policy == p->policy)) {
7716 346 : if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7717 : goto change;
7718 173 : if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7719 : goto change;
7720 173 : if (dl_policy(policy) && dl_param_changed(p, attr))
7721 : goto change;
7722 173 : if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7723 : goto change;
7724 :
7725 173 : p->sched_reset_on_fork = reset_on_fork;
7726 173 : retval = 0;
7727 173 : goto unlock;
7728 : }
7729 : change:
7730 :
7731 : if (user) {
7732 : #ifdef CONFIG_RT_GROUP_SCHED
7733 : /*
7734 : * Do not allow realtime tasks into groups that have no runtime
7735 : * assigned.
7736 : */
7737 : if (rt_bandwidth_enabled() && rt_policy(policy) &&
7738 : task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7739 : !task_group_is_autogroup(task_group(p))) {
7740 : retval = -EPERM;
7741 : goto unlock;
7742 : }
7743 : #endif
7744 : #ifdef CONFIG_SMP
7745 : if (dl_bandwidth_enabled() && dl_policy(policy) &&
7746 : !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7747 : cpumask_t *span = rq->rd->span;
7748 :
7749 : /*
7750 : * Don't allow tasks with an affinity mask smaller than
7751 : * the entire root_domain to become SCHED_DEADLINE. We
7752 : * will also fail if there's no bandwidth available.
7753 : */
7754 : if (!cpumask_subset(span, p->cpus_ptr) ||
7755 : rq->rd->dl_bw.bw == 0) {
7756 : retval = -EPERM;
7757 : goto unlock;
7758 : }
7759 : }
7760 : #endif
7761 : }
7762 :
7763 : /* Re-check policy now with rq lock held: */
7764 0 : if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7765 0 : policy = oldpolicy = -1;
7766 0 : task_rq_unlock(rq, p, &rf);
7767 : if (cpuset_locked)
7768 : cpuset_unlock();
7769 : goto recheck;
7770 : }
7771 :
7772 : /*
7773 : * If setscheduling to SCHED_DEADLINE (or changing the parameters
7774 : * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7775 : * is available.
7776 : */
7777 0 : if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7778 : retval = -EBUSY;
7779 : goto unlock;
7780 : }
7781 :
7782 0 : p->sched_reset_on_fork = reset_on_fork;
7783 0 : oldprio = p->prio;
7784 :
7785 0 : newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7786 0 : if (pi) {
7787 : /*
7788 : * Take priority boosted tasks into account. If the new
7789 : * effective priority is unchanged, we just store the new
7790 : * normal parameters and do not touch the scheduler class and
7791 : * the runqueue. This will be done when the task deboost
7792 : * itself.
7793 : */
7794 0 : newprio = rt_effective_prio(p, newprio);
7795 0 : if (newprio == oldprio)
7796 0 : queue_flags &= ~DEQUEUE_MOVE;
7797 : }
7798 :
7799 0 : queued = task_on_rq_queued(p);
7800 0 : running = task_current(rq, p);
7801 0 : if (queued)
7802 0 : dequeue_task(rq, p, queue_flags);
7803 0 : if (running)
7804 0 : put_prev_task(rq, p);
7805 :
7806 0 : prev_class = p->sched_class;
7807 :
7808 0 : if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7809 0 : __setscheduler_params(p, attr);
7810 0 : __setscheduler_prio(p, newprio);
7811 : }
7812 0 : __setscheduler_uclamp(p, attr);
7813 :
7814 0 : if (queued) {
7815 : /*
7816 : * We enqueue to tail when the priority of a task is
7817 : * increased (user space view).
7818 : */
7819 0 : if (oldprio < p->prio)
7820 0 : queue_flags |= ENQUEUE_HEAD;
7821 :
7822 0 : enqueue_task(rq, p, queue_flags);
7823 : }
7824 0 : if (running)
7825 : set_next_task(rq, p);
7826 :
7827 0 : check_class_changed(rq, p, prev_class, oldprio);
7828 :
7829 : /* Avoid rq from going away on us: */
7830 0 : preempt_disable();
7831 0 : head = splice_balance_callbacks(rq);
7832 0 : task_rq_unlock(rq, p, &rf);
7833 :
7834 0 : if (pi) {
7835 : if (cpuset_locked)
7836 : cpuset_unlock();
7837 0 : rt_mutex_adjust_pi(p);
7838 : }
7839 :
7840 : /* Run balance callbacks after we've adjusted the PI chain: */
7841 0 : balance_callbacks(rq, head);
7842 0 : preempt_enable();
7843 :
7844 0 : return 0;
7845 :
7846 : unlock:
7847 346 : task_rq_unlock(rq, p, &rf);
7848 : if (cpuset_locked)
7849 : cpuset_unlock();
7850 173 : return retval;
7851 : }
7852 :
7853 173 : static int _sched_setscheduler(struct task_struct *p, int policy,
7854 : const struct sched_param *param, bool check)
7855 : {
7856 519 : struct sched_attr attr = {
7857 : .sched_policy = policy,
7858 173 : .sched_priority = param->sched_priority,
7859 173 : .sched_nice = PRIO_TO_NICE(p->static_prio),
7860 : };
7861 :
7862 : /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7863 173 : if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7864 0 : attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7865 0 : policy &= ~SCHED_RESET_ON_FORK;
7866 0 : attr.sched_policy = policy;
7867 : }
7868 :
7869 173 : return __sched_setscheduler(p, &attr, check, true);
7870 : }
7871 : /**
7872 : * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7873 : * @p: the task in question.
7874 : * @policy: new policy.
7875 : * @param: structure containing the new RT priority.
7876 : *
7877 : * Use sched_set_fifo(), read its comment.
7878 : *
7879 : * Return: 0 on success. An error code otherwise.
7880 : *
7881 : * NOTE that the task may be already dead.
7882 : */
7883 0 : int sched_setscheduler(struct task_struct *p, int policy,
7884 : const struct sched_param *param)
7885 : {
7886 0 : return _sched_setscheduler(p, policy, param, true);
7887 : }
7888 :
7889 0 : int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7890 : {
7891 0 : return __sched_setscheduler(p, attr, true, true);
7892 : }
7893 :
7894 0 : int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7895 : {
7896 0 : return __sched_setscheduler(p, attr, false, true);
7897 : }
7898 : EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7899 :
7900 : /**
7901 : * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7902 : * @p: the task in question.
7903 : * @policy: new policy.
7904 : * @param: structure containing the new RT priority.
7905 : *
7906 : * Just like sched_setscheduler, only don't bother checking if the
7907 : * current context has permission. For example, this is needed in
7908 : * stop_machine(): we create temporary high priority worker threads,
7909 : * but our caller might not have that capability.
7910 : *
7911 : * Return: 0 on success. An error code otherwise.
7912 : */
7913 173 : int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7914 : const struct sched_param *param)
7915 : {
7916 173 : return _sched_setscheduler(p, policy, param, false);
7917 : }
7918 :
7919 : /*
7920 : * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7921 : * incapable of resource management, which is the one thing an OS really should
7922 : * be doing.
7923 : *
7924 : * This is of course the reason it is limited to privileged users only.
7925 : *
7926 : * Worse still; it is fundamentally impossible to compose static priority
7927 : * workloads. You cannot take two correctly working static prio workloads
7928 : * and smash them together and still expect them to work.
7929 : *
7930 : * For this reason 'all' FIFO tasks the kernel creates are basically at:
7931 : *
7932 : * MAX_RT_PRIO / 2
7933 : *
7934 : * The administrator _MUST_ configure the system, the kernel simply doesn't
7935 : * know enough information to make a sensible choice.
7936 : */
7937 0 : void sched_set_fifo(struct task_struct *p)
7938 : {
7939 0 : struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7940 0 : WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7941 0 : }
7942 : EXPORT_SYMBOL_GPL(sched_set_fifo);
7943 :
7944 : /*
7945 : * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7946 : */
7947 0 : void sched_set_fifo_low(struct task_struct *p)
7948 : {
7949 0 : struct sched_param sp = { .sched_priority = 1 };
7950 0 : WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7951 0 : }
7952 : EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7953 :
7954 0 : void sched_set_normal(struct task_struct *p, int nice)
7955 : {
7956 0 : struct sched_attr attr = {
7957 : .sched_policy = SCHED_NORMAL,
7958 : .sched_nice = nice,
7959 : };
7960 0 : WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7961 0 : }
7962 : EXPORT_SYMBOL_GPL(sched_set_normal);
7963 :
7964 : static int
7965 0 : do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7966 : {
7967 : struct sched_param lparam;
7968 : struct task_struct *p;
7969 : int retval;
7970 :
7971 0 : if (!param || pid < 0)
7972 : return -EINVAL;
7973 0 : if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7974 : return -EFAULT;
7975 :
7976 : rcu_read_lock();
7977 0 : retval = -ESRCH;
7978 0 : p = find_process_by_pid(pid);
7979 0 : if (likely(p))
7980 : get_task_struct(p);
7981 : rcu_read_unlock();
7982 :
7983 0 : if (likely(p)) {
7984 0 : retval = sched_setscheduler(p, policy, &lparam);
7985 0 : put_task_struct(p);
7986 : }
7987 :
7988 : return retval;
7989 : }
7990 :
7991 : /*
7992 : * Mimics kernel/events/core.c perf_copy_attr().
7993 : */
7994 0 : static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7995 : {
7996 : u32 size;
7997 : int ret;
7998 :
7999 : /* Zero the full structure, so that a short copy will be nice: */
8000 0 : memset(attr, 0, sizeof(*attr));
8001 :
8002 0 : ret = get_user(size, &uattr->size);
8003 0 : if (ret)
8004 : return ret;
8005 :
8006 : /* ABI compatibility quirk: */
8007 0 : if (!size)
8008 0 : size = SCHED_ATTR_SIZE_VER0;
8009 0 : if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
8010 : goto err_size;
8011 :
8012 0 : ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
8013 0 : if (ret) {
8014 0 : if (ret == -E2BIG)
8015 : goto err_size;
8016 : return ret;
8017 : }
8018 :
8019 0 : if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
8020 : size < SCHED_ATTR_SIZE_VER1)
8021 : return -EINVAL;
8022 :
8023 : /*
8024 : * XXX: Do we want to be lenient like existing syscalls; or do we want
8025 : * to be strict and return an error on out-of-bounds values?
8026 : */
8027 0 : attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
8028 :
8029 0 : return 0;
8030 :
8031 : err_size:
8032 0 : put_user(sizeof(*attr), &uattr->size);
8033 : return -E2BIG;
8034 : }
8035 :
8036 0 : static void get_params(struct task_struct *p, struct sched_attr *attr)
8037 : {
8038 0 : if (task_has_dl_policy(p))
8039 0 : __getparam_dl(p, attr);
8040 0 : else if (task_has_rt_policy(p))
8041 0 : attr->sched_priority = p->rt_priority;
8042 : else
8043 0 : attr->sched_nice = task_nice(p);
8044 0 : }
8045 :
8046 : /**
8047 : * sys_sched_setscheduler - set/change the scheduler policy and RT priority
8048 : * @pid: the pid in question.
8049 : * @policy: new policy.
8050 : * @param: structure containing the new RT priority.
8051 : *
8052 : * Return: 0 on success. An error code otherwise.
8053 : */
8054 0 : SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
8055 : {
8056 0 : if (policy < 0)
8057 : return -EINVAL;
8058 :
8059 0 : return do_sched_setscheduler(pid, policy, param);
8060 : }
8061 :
8062 : /**
8063 : * sys_sched_setparam - set/change the RT priority of a thread
8064 : * @pid: the pid in question.
8065 : * @param: structure containing the new RT priority.
8066 : *
8067 : * Return: 0 on success. An error code otherwise.
8068 : */
8069 0 : SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
8070 : {
8071 0 : return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
8072 : }
8073 :
8074 : /**
8075 : * sys_sched_setattr - same as above, but with extended sched_attr
8076 : * @pid: the pid in question.
8077 : * @uattr: structure containing the extended parameters.
8078 : * @flags: for future extension.
8079 : */
8080 0 : SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
8081 : unsigned int, flags)
8082 : {
8083 : struct sched_attr attr;
8084 : struct task_struct *p;
8085 : int retval;
8086 :
8087 0 : if (!uattr || pid < 0 || flags)
8088 : return -EINVAL;
8089 :
8090 0 : retval = sched_copy_attr(uattr, &attr);
8091 0 : if (retval)
8092 0 : return retval;
8093 :
8094 0 : if ((int)attr.sched_policy < 0)
8095 : return -EINVAL;
8096 0 : if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
8097 0 : attr.sched_policy = SETPARAM_POLICY;
8098 :
8099 : rcu_read_lock();
8100 0 : retval = -ESRCH;
8101 0 : p = find_process_by_pid(pid);
8102 0 : if (likely(p))
8103 : get_task_struct(p);
8104 : rcu_read_unlock();
8105 :
8106 0 : if (likely(p)) {
8107 0 : if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
8108 0 : get_params(p, &attr);
8109 0 : retval = sched_setattr(p, &attr);
8110 0 : put_task_struct(p);
8111 : }
8112 :
8113 0 : return retval;
8114 : }
8115 :
8116 : /**
8117 : * sys_sched_getscheduler - get the policy (scheduling class) of a thread
8118 : * @pid: the pid in question.
8119 : *
8120 : * Return: On success, the policy of the thread. Otherwise, a negative error
8121 : * code.
8122 : */
8123 0 : SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
8124 : {
8125 : struct task_struct *p;
8126 : int retval;
8127 :
8128 0 : if (pid < 0)
8129 : return -EINVAL;
8130 :
8131 0 : retval = -ESRCH;
8132 : rcu_read_lock();
8133 0 : p = find_process_by_pid(pid);
8134 0 : if (p) {
8135 0 : retval = security_task_getscheduler(p);
8136 : if (!retval)
8137 0 : retval = p->policy
8138 0 : | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
8139 : }
8140 : rcu_read_unlock();
8141 0 : return retval;
8142 : }
8143 :
8144 : /**
8145 : * sys_sched_getparam - get the RT priority of a thread
8146 : * @pid: the pid in question.
8147 : * @param: structure containing the RT priority.
8148 : *
8149 : * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8150 : * code.
8151 : */
8152 0 : SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
8153 : {
8154 0 : struct sched_param lp = { .sched_priority = 0 };
8155 : struct task_struct *p;
8156 : int retval;
8157 :
8158 0 : if (!param || pid < 0)
8159 : return -EINVAL;
8160 :
8161 : rcu_read_lock();
8162 0 : p = find_process_by_pid(pid);
8163 0 : retval = -ESRCH;
8164 0 : if (!p)
8165 : goto out_unlock;
8166 :
8167 0 : retval = security_task_getscheduler(p);
8168 : if (retval)
8169 : goto out_unlock;
8170 :
8171 0 : if (task_has_rt_policy(p))
8172 0 : lp.sched_priority = p->rt_priority;
8173 0 : rcu_read_unlock();
8174 :
8175 : /*
8176 : * This one might sleep, we cannot do it with a spinlock held ...
8177 : */
8178 0 : retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
8179 :
8180 0 : return retval;
8181 :
8182 : out_unlock:
8183 : rcu_read_unlock();
8184 0 : return retval;
8185 : }
8186 :
8187 : /*
8188 : * Copy the kernel size attribute structure (which might be larger
8189 : * than what user-space knows about) to user-space.
8190 : *
8191 : * Note that all cases are valid: user-space buffer can be larger or
8192 : * smaller than the kernel-space buffer. The usual case is that both
8193 : * have the same size.
8194 : */
8195 : static int
8196 0 : sched_attr_copy_to_user(struct sched_attr __user *uattr,
8197 : struct sched_attr *kattr,
8198 : unsigned int usize)
8199 : {
8200 0 : unsigned int ksize = sizeof(*kattr);
8201 :
8202 0 : if (!access_ok(uattr, usize))
8203 : return -EFAULT;
8204 :
8205 : /*
8206 : * sched_getattr() ABI forwards and backwards compatibility:
8207 : *
8208 : * If usize == ksize then we just copy everything to user-space and all is good.
8209 : *
8210 : * If usize < ksize then we only copy as much as user-space has space for,
8211 : * this keeps ABI compatibility as well. We skip the rest.
8212 : *
8213 : * If usize > ksize then user-space is using a newer version of the ABI,
8214 : * which part the kernel doesn't know about. Just ignore it - tooling can
8215 : * detect the kernel's knowledge of attributes from the attr->size value
8216 : * which is set to ksize in this case.
8217 : */
8218 0 : kattr->size = min(usize, ksize);
8219 :
8220 0 : if (copy_to_user(uattr, kattr, kattr->size))
8221 : return -EFAULT;
8222 :
8223 0 : return 0;
8224 : }
8225 :
8226 : /**
8227 : * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8228 : * @pid: the pid in question.
8229 : * @uattr: structure containing the extended parameters.
8230 : * @usize: sizeof(attr) for fwd/bwd comp.
8231 : * @flags: for future extension.
8232 : */
8233 0 : SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8234 : unsigned int, usize, unsigned int, flags)
8235 : {
8236 0 : struct sched_attr kattr = { };
8237 : struct task_struct *p;
8238 : int retval;
8239 :
8240 0 : if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8241 0 : usize < SCHED_ATTR_SIZE_VER0 || flags)
8242 : return -EINVAL;
8243 :
8244 : rcu_read_lock();
8245 0 : p = find_process_by_pid(pid);
8246 0 : retval = -ESRCH;
8247 0 : if (!p)
8248 : goto out_unlock;
8249 :
8250 0 : retval = security_task_getscheduler(p);
8251 : if (retval)
8252 : goto out_unlock;
8253 :
8254 0 : kattr.sched_policy = p->policy;
8255 0 : if (p->sched_reset_on_fork)
8256 0 : kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8257 0 : get_params(p, &kattr);
8258 0 : kattr.sched_flags &= SCHED_FLAG_ALL;
8259 :
8260 : #ifdef CONFIG_UCLAMP_TASK
8261 : /*
8262 : * This could race with another potential updater, but this is fine
8263 : * because it'll correctly read the old or the new value. We don't need
8264 : * to guarantee who wins the race as long as it doesn't return garbage.
8265 : */
8266 : kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8267 : kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8268 : #endif
8269 :
8270 : rcu_read_unlock();
8271 :
8272 0 : return sched_attr_copy_to_user(uattr, &kattr, usize);
8273 :
8274 : out_unlock:
8275 : rcu_read_unlock();
8276 0 : return retval;
8277 : }
8278 :
8279 : #ifdef CONFIG_SMP
8280 : int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8281 : {
8282 : int ret = 0;
8283 :
8284 : /*
8285 : * If the task isn't a deadline task or admission control is
8286 : * disabled then we don't care about affinity changes.
8287 : */
8288 : if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8289 : return 0;
8290 :
8291 : /*
8292 : * Since bandwidth control happens on root_domain basis,
8293 : * if admission test is enabled, we only admit -deadline
8294 : * tasks allowed to run on all the CPUs in the task's
8295 : * root_domain.
8296 : */
8297 : rcu_read_lock();
8298 : if (!cpumask_subset(task_rq(p)->rd->span, mask))
8299 : ret = -EBUSY;
8300 : rcu_read_unlock();
8301 : return ret;
8302 : }
8303 : #endif
8304 :
8305 : static int
8306 0 : __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
8307 : {
8308 : int retval;
8309 : cpumask_var_t cpus_allowed, new_mask;
8310 :
8311 0 : if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8312 : return -ENOMEM;
8313 :
8314 0 : if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8315 : retval = -ENOMEM;
8316 : goto out_free_cpus_allowed;
8317 : }
8318 :
8319 0 : cpuset_cpus_allowed(p, cpus_allowed);
8320 0 : cpumask_and(new_mask, ctx->new_mask, cpus_allowed);
8321 :
8322 0 : ctx->new_mask = new_mask;
8323 0 : ctx->flags |= SCA_CHECK;
8324 :
8325 0 : retval = dl_task_check_affinity(p, new_mask);
8326 : if (retval)
8327 : goto out_free_new_mask;
8328 :
8329 0 : retval = __set_cpus_allowed_ptr(p, ctx);
8330 0 : if (retval)
8331 : goto out_free_new_mask;
8332 :
8333 0 : cpuset_cpus_allowed(p, cpus_allowed);
8334 0 : if (!cpumask_subset(new_mask, cpus_allowed)) {
8335 : /*
8336 : * We must have raced with a concurrent cpuset update.
8337 : * Just reset the cpumask to the cpuset's cpus_allowed.
8338 : */
8339 0 : cpumask_copy(new_mask, cpus_allowed);
8340 :
8341 : /*
8342 : * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8343 : * will restore the previous user_cpus_ptr value.
8344 : *
8345 : * In the unlikely event a previous user_cpus_ptr exists,
8346 : * we need to further restrict the mask to what is allowed
8347 : * by that old user_cpus_ptr.
8348 : */
8349 0 : if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
8350 0 : bool empty = !cpumask_and(new_mask, new_mask,
8351 0 : ctx->user_mask);
8352 :
8353 0 : if (WARN_ON_ONCE(empty))
8354 : cpumask_copy(new_mask, cpus_allowed);
8355 : }
8356 0 : __set_cpus_allowed_ptr(p, ctx);
8357 : retval = -EINVAL;
8358 : }
8359 :
8360 : out_free_new_mask:
8361 0 : free_cpumask_var(new_mask);
8362 : out_free_cpus_allowed:
8363 0 : free_cpumask_var(cpus_allowed);
8364 : return retval;
8365 : }
8366 :
8367 0 : long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8368 : {
8369 : struct affinity_context ac;
8370 : struct cpumask *user_mask;
8371 : struct task_struct *p;
8372 : int retval;
8373 :
8374 : rcu_read_lock();
8375 :
8376 0 : p = find_process_by_pid(pid);
8377 0 : if (!p) {
8378 : rcu_read_unlock();
8379 0 : return -ESRCH;
8380 : }
8381 :
8382 : /* Prevent p going away */
8383 0 : get_task_struct(p);
8384 : rcu_read_unlock();
8385 :
8386 0 : if (p->flags & PF_NO_SETAFFINITY) {
8387 : retval = -EINVAL;
8388 : goto out_put_task;
8389 : }
8390 :
8391 0 : if (!check_same_owner(p)) {
8392 : rcu_read_lock();
8393 0 : if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8394 : rcu_read_unlock();
8395 0 : retval = -EPERM;
8396 0 : goto out_put_task;
8397 : }
8398 : rcu_read_unlock();
8399 : }
8400 :
8401 0 : retval = security_task_setscheduler(p);
8402 0 : if (retval)
8403 : goto out_put_task;
8404 :
8405 : /*
8406 : * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8407 : * alloc_user_cpus_ptr() returns NULL.
8408 : */
8409 0 : user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
8410 : if (user_mask) {
8411 : cpumask_copy(user_mask, in_mask);
8412 : } else if (IS_ENABLED(CONFIG_SMP)) {
8413 : retval = -ENOMEM;
8414 : goto out_put_task;
8415 : }
8416 :
8417 0 : ac = (struct affinity_context){
8418 : .new_mask = in_mask,
8419 : .user_mask = user_mask,
8420 : .flags = SCA_USER,
8421 : };
8422 :
8423 0 : retval = __sched_setaffinity(p, &ac);
8424 0 : kfree(ac.user_mask);
8425 :
8426 : out_put_task:
8427 0 : put_task_struct(p);
8428 0 : return retval;
8429 : }
8430 :
8431 0 : static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8432 : struct cpumask *new_mask)
8433 : {
8434 0 : if (len < cpumask_size())
8435 : cpumask_clear(new_mask);
8436 0 : else if (len > cpumask_size())
8437 0 : len = cpumask_size();
8438 :
8439 0 : return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8440 : }
8441 :
8442 : /**
8443 : * sys_sched_setaffinity - set the CPU affinity of a process
8444 : * @pid: pid of the process
8445 : * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8446 : * @user_mask_ptr: user-space pointer to the new CPU mask
8447 : *
8448 : * Return: 0 on success. An error code otherwise.
8449 : */
8450 0 : SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8451 : unsigned long __user *, user_mask_ptr)
8452 : {
8453 : cpumask_var_t new_mask;
8454 : int retval;
8455 :
8456 0 : if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8457 : return -ENOMEM;
8458 :
8459 0 : retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8460 0 : if (retval == 0)
8461 0 : retval = sched_setaffinity(pid, new_mask);
8462 0 : free_cpumask_var(new_mask);
8463 0 : return retval;
8464 : }
8465 :
8466 0 : long sched_getaffinity(pid_t pid, struct cpumask *mask)
8467 : {
8468 : struct task_struct *p;
8469 : unsigned long flags;
8470 : int retval;
8471 :
8472 : rcu_read_lock();
8473 :
8474 0 : retval = -ESRCH;
8475 0 : p = find_process_by_pid(pid);
8476 0 : if (!p)
8477 : goto out_unlock;
8478 :
8479 0 : retval = security_task_getscheduler(p);
8480 : if (retval)
8481 : goto out_unlock;
8482 :
8483 0 : raw_spin_lock_irqsave(&p->pi_lock, flags);
8484 0 : cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8485 0 : raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8486 :
8487 : out_unlock:
8488 : rcu_read_unlock();
8489 :
8490 0 : return retval;
8491 : }
8492 :
8493 : /**
8494 : * sys_sched_getaffinity - get the CPU affinity of a process
8495 : * @pid: pid of the process
8496 : * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8497 : * @user_mask_ptr: user-space pointer to hold the current CPU mask
8498 : *
8499 : * Return: size of CPU mask copied to user_mask_ptr on success. An
8500 : * error code otherwise.
8501 : */
8502 0 : SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8503 : unsigned long __user *, user_mask_ptr)
8504 : {
8505 : int ret;
8506 : cpumask_var_t mask;
8507 :
8508 0 : if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8509 : return -EINVAL;
8510 0 : if (len & (sizeof(unsigned long)-1))
8511 : return -EINVAL;
8512 :
8513 0 : if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
8514 : return -ENOMEM;
8515 :
8516 0 : ret = sched_getaffinity(pid, mask);
8517 0 : if (ret == 0) {
8518 0 : unsigned int retlen = min(len, cpumask_size());
8519 :
8520 0 : if (copy_to_user(user_mask_ptr, cpumask_bits(mask), retlen))
8521 : ret = -EFAULT;
8522 : else
8523 0 : ret = retlen;
8524 : }
8525 0 : free_cpumask_var(mask);
8526 :
8527 0 : return ret;
8528 : }
8529 :
8530 0 : static void do_sched_yield(void)
8531 : {
8532 : struct rq_flags rf;
8533 : struct rq *rq;
8534 :
8535 0 : rq = this_rq_lock_irq(&rf);
8536 :
8537 : schedstat_inc(rq->yld_count);
8538 0 : current->sched_class->yield_task(rq);
8539 :
8540 0 : preempt_disable();
8541 0 : rq_unlock_irq(rq, &rf);
8542 0 : sched_preempt_enable_no_resched();
8543 :
8544 0 : schedule();
8545 0 : }
8546 :
8547 : /**
8548 : * sys_sched_yield - yield the current processor to other threads.
8549 : *
8550 : * This function yields the current CPU to other tasks. If there are no
8551 : * other threads running on this CPU then this function will return.
8552 : *
8553 : * Return: 0.
8554 : */
8555 0 : SYSCALL_DEFINE0(sched_yield)
8556 : {
8557 0 : do_sched_yield();
8558 0 : return 0;
8559 : }
8560 :
8561 : #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8562 721 : int __sched __cond_resched(void)
8563 : {
8564 721 : if (should_resched(0)) {
8565 : preempt_schedule_common();
8566 : return 1;
8567 : }
8568 : /*
8569 : * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8570 : * whether the current CPU is in an RCU read-side critical section,
8571 : * so the tick can report quiescent states even for CPUs looping
8572 : * in kernel context. In contrast, in non-preemptible kernels,
8573 : * RCU readers leave no in-memory hints, which means that CPU-bound
8574 : * processes executing in kernel context might never report an
8575 : * RCU quiescent state. Therefore, the following code causes
8576 : * cond_resched() to report a quiescent state, but only when RCU
8577 : * is in urgent need of one.
8578 : */
8579 : #ifndef CONFIG_PREEMPT_RCU
8580 : rcu_all_qs();
8581 : #endif
8582 719 : return 0;
8583 : }
8584 : EXPORT_SYMBOL(__cond_resched);
8585 : #endif
8586 :
8587 : #ifdef CONFIG_PREEMPT_DYNAMIC
8588 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8589 : #define cond_resched_dynamic_enabled __cond_resched
8590 : #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8591 : DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8592 : EXPORT_STATIC_CALL_TRAMP(cond_resched);
8593 :
8594 : #define might_resched_dynamic_enabled __cond_resched
8595 : #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8596 : DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8597 : EXPORT_STATIC_CALL_TRAMP(might_resched);
8598 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8599 : static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8600 : int __sched dynamic_cond_resched(void)
8601 : {
8602 : klp_sched_try_switch();
8603 : if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8604 : return 0;
8605 : return __cond_resched();
8606 : }
8607 : EXPORT_SYMBOL(dynamic_cond_resched);
8608 :
8609 : static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8610 : int __sched dynamic_might_resched(void)
8611 : {
8612 : if (!static_branch_unlikely(&sk_dynamic_might_resched))
8613 : return 0;
8614 : return __cond_resched();
8615 : }
8616 : EXPORT_SYMBOL(dynamic_might_resched);
8617 : #endif
8618 : #endif
8619 :
8620 : /*
8621 : * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8622 : * call schedule, and on return reacquire the lock.
8623 : *
8624 : * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8625 : * operations here to prevent schedule() from being called twice (once via
8626 : * spin_unlock(), once by hand).
8627 : */
8628 0 : int __cond_resched_lock(spinlock_t *lock)
8629 : {
8630 0 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
8631 0 : int ret = 0;
8632 :
8633 : lockdep_assert_held(lock);
8634 :
8635 0 : if (spin_needbreak(lock) || resched) {
8636 0 : spin_unlock(lock);
8637 0 : if (!_cond_resched())
8638 : cpu_relax();
8639 0 : ret = 1;
8640 : spin_lock(lock);
8641 : }
8642 0 : return ret;
8643 : }
8644 : EXPORT_SYMBOL(__cond_resched_lock);
8645 :
8646 0 : int __cond_resched_rwlock_read(rwlock_t *lock)
8647 : {
8648 0 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
8649 0 : int ret = 0;
8650 :
8651 : lockdep_assert_held_read(lock);
8652 :
8653 0 : if (rwlock_needbreak(lock) || resched) {
8654 0 : read_unlock(lock);
8655 0 : if (!_cond_resched())
8656 : cpu_relax();
8657 0 : ret = 1;
8658 0 : read_lock(lock);
8659 : }
8660 0 : return ret;
8661 : }
8662 : EXPORT_SYMBOL(__cond_resched_rwlock_read);
8663 :
8664 0 : int __cond_resched_rwlock_write(rwlock_t *lock)
8665 : {
8666 0 : int resched = should_resched(PREEMPT_LOCK_OFFSET);
8667 0 : int ret = 0;
8668 :
8669 : lockdep_assert_held_write(lock);
8670 :
8671 0 : if (rwlock_needbreak(lock) || resched) {
8672 0 : write_unlock(lock);
8673 0 : if (!_cond_resched())
8674 : cpu_relax();
8675 0 : ret = 1;
8676 0 : write_lock(lock);
8677 : }
8678 0 : return ret;
8679 : }
8680 : EXPORT_SYMBOL(__cond_resched_rwlock_write);
8681 :
8682 : #ifdef CONFIG_PREEMPT_DYNAMIC
8683 :
8684 : #ifdef CONFIG_GENERIC_ENTRY
8685 : #include <linux/entry-common.h>
8686 : #endif
8687 :
8688 : /*
8689 : * SC:cond_resched
8690 : * SC:might_resched
8691 : * SC:preempt_schedule
8692 : * SC:preempt_schedule_notrace
8693 : * SC:irqentry_exit_cond_resched
8694 : *
8695 : *
8696 : * NONE:
8697 : * cond_resched <- __cond_resched
8698 : * might_resched <- RET0
8699 : * preempt_schedule <- NOP
8700 : * preempt_schedule_notrace <- NOP
8701 : * irqentry_exit_cond_resched <- NOP
8702 : *
8703 : * VOLUNTARY:
8704 : * cond_resched <- __cond_resched
8705 : * might_resched <- __cond_resched
8706 : * preempt_schedule <- NOP
8707 : * preempt_schedule_notrace <- NOP
8708 : * irqentry_exit_cond_resched <- NOP
8709 : *
8710 : * FULL:
8711 : * cond_resched <- RET0
8712 : * might_resched <- RET0
8713 : * preempt_schedule <- preempt_schedule
8714 : * preempt_schedule_notrace <- preempt_schedule_notrace
8715 : * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8716 : */
8717 :
8718 : enum {
8719 : preempt_dynamic_undefined = -1,
8720 : preempt_dynamic_none,
8721 : preempt_dynamic_voluntary,
8722 : preempt_dynamic_full,
8723 : };
8724 :
8725 : int preempt_dynamic_mode = preempt_dynamic_undefined;
8726 :
8727 : int sched_dynamic_mode(const char *str)
8728 : {
8729 : if (!strcmp(str, "none"))
8730 : return preempt_dynamic_none;
8731 :
8732 : if (!strcmp(str, "voluntary"))
8733 : return preempt_dynamic_voluntary;
8734 :
8735 : if (!strcmp(str, "full"))
8736 : return preempt_dynamic_full;
8737 :
8738 : return -EINVAL;
8739 : }
8740 :
8741 : #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8742 : #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8743 : #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8744 : #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8745 : #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8746 : #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8747 : #else
8748 : #error "Unsupported PREEMPT_DYNAMIC mechanism"
8749 : #endif
8750 :
8751 : static DEFINE_MUTEX(sched_dynamic_mutex);
8752 : static bool klp_override;
8753 :
8754 : static void __sched_dynamic_update(int mode)
8755 : {
8756 : /*
8757 : * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8758 : * the ZERO state, which is invalid.
8759 : */
8760 : if (!klp_override)
8761 : preempt_dynamic_enable(cond_resched);
8762 : preempt_dynamic_enable(might_resched);
8763 : preempt_dynamic_enable(preempt_schedule);
8764 : preempt_dynamic_enable(preempt_schedule_notrace);
8765 : preempt_dynamic_enable(irqentry_exit_cond_resched);
8766 :
8767 : switch (mode) {
8768 : case preempt_dynamic_none:
8769 : if (!klp_override)
8770 : preempt_dynamic_enable(cond_resched);
8771 : preempt_dynamic_disable(might_resched);
8772 : preempt_dynamic_disable(preempt_schedule);
8773 : preempt_dynamic_disable(preempt_schedule_notrace);
8774 : preempt_dynamic_disable(irqentry_exit_cond_resched);
8775 : if (mode != preempt_dynamic_mode)
8776 : pr_info("Dynamic Preempt: none\n");
8777 : break;
8778 :
8779 : case preempt_dynamic_voluntary:
8780 : if (!klp_override)
8781 : preempt_dynamic_enable(cond_resched);
8782 : preempt_dynamic_enable(might_resched);
8783 : preempt_dynamic_disable(preempt_schedule);
8784 : preempt_dynamic_disable(preempt_schedule_notrace);
8785 : preempt_dynamic_disable(irqentry_exit_cond_resched);
8786 : if (mode != preempt_dynamic_mode)
8787 : pr_info("Dynamic Preempt: voluntary\n");
8788 : break;
8789 :
8790 : case preempt_dynamic_full:
8791 : if (!klp_override)
8792 : preempt_dynamic_disable(cond_resched);
8793 : preempt_dynamic_disable(might_resched);
8794 : preempt_dynamic_enable(preempt_schedule);
8795 : preempt_dynamic_enable(preempt_schedule_notrace);
8796 : preempt_dynamic_enable(irqentry_exit_cond_resched);
8797 : if (mode != preempt_dynamic_mode)
8798 : pr_info("Dynamic Preempt: full\n");
8799 : break;
8800 : }
8801 :
8802 : preempt_dynamic_mode = mode;
8803 : }
8804 :
8805 : void sched_dynamic_update(int mode)
8806 : {
8807 : mutex_lock(&sched_dynamic_mutex);
8808 : __sched_dynamic_update(mode);
8809 : mutex_unlock(&sched_dynamic_mutex);
8810 : }
8811 :
8812 : #ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
8813 :
8814 : static int klp_cond_resched(void)
8815 : {
8816 : __klp_sched_try_switch();
8817 : return __cond_resched();
8818 : }
8819 :
8820 : void sched_dynamic_klp_enable(void)
8821 : {
8822 : mutex_lock(&sched_dynamic_mutex);
8823 :
8824 : klp_override = true;
8825 : static_call_update(cond_resched, klp_cond_resched);
8826 :
8827 : mutex_unlock(&sched_dynamic_mutex);
8828 : }
8829 :
8830 : void sched_dynamic_klp_disable(void)
8831 : {
8832 : mutex_lock(&sched_dynamic_mutex);
8833 :
8834 : klp_override = false;
8835 : __sched_dynamic_update(preempt_dynamic_mode);
8836 :
8837 : mutex_unlock(&sched_dynamic_mutex);
8838 : }
8839 :
8840 : #endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
8841 :
8842 : static int __init setup_preempt_mode(char *str)
8843 : {
8844 : int mode = sched_dynamic_mode(str);
8845 : if (mode < 0) {
8846 : pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8847 : return 0;
8848 : }
8849 :
8850 : sched_dynamic_update(mode);
8851 : return 1;
8852 : }
8853 : __setup("preempt=", setup_preempt_mode);
8854 :
8855 : static void __init preempt_dynamic_init(void)
8856 : {
8857 : if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8858 : if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8859 : sched_dynamic_update(preempt_dynamic_none);
8860 : } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8861 : sched_dynamic_update(preempt_dynamic_voluntary);
8862 : } else {
8863 : /* Default static call setting, nothing to do */
8864 : WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8865 : preempt_dynamic_mode = preempt_dynamic_full;
8866 : pr_info("Dynamic Preempt: full\n");
8867 : }
8868 : }
8869 : }
8870 :
8871 : #define PREEMPT_MODEL_ACCESSOR(mode) \
8872 : bool preempt_model_##mode(void) \
8873 : { \
8874 : WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8875 : return preempt_dynamic_mode == preempt_dynamic_##mode; \
8876 : } \
8877 : EXPORT_SYMBOL_GPL(preempt_model_##mode)
8878 :
8879 : PREEMPT_MODEL_ACCESSOR(none);
8880 : PREEMPT_MODEL_ACCESSOR(voluntary);
8881 : PREEMPT_MODEL_ACCESSOR(full);
8882 :
8883 : #else /* !CONFIG_PREEMPT_DYNAMIC */
8884 :
8885 : static inline void preempt_dynamic_init(void) { }
8886 :
8887 : #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8888 :
8889 : /**
8890 : * yield - yield the current processor to other threads.
8891 : *
8892 : * Do not ever use this function, there's a 99% chance you're doing it wrong.
8893 : *
8894 : * The scheduler is at all times free to pick the calling task as the most
8895 : * eligible task to run, if removing the yield() call from your code breaks
8896 : * it, it's already broken.
8897 : *
8898 : * Typical broken usage is:
8899 : *
8900 : * while (!event)
8901 : * yield();
8902 : *
8903 : * where one assumes that yield() will let 'the other' process run that will
8904 : * make event true. If the current task is a SCHED_FIFO task that will never
8905 : * happen. Never use yield() as a progress guarantee!!
8906 : *
8907 : * If you want to use yield() to wait for something, use wait_event().
8908 : * If you want to use yield() to be 'nice' for others, use cond_resched().
8909 : * If you still want to use yield(), do not!
8910 : */
8911 0 : void __sched yield(void)
8912 : {
8913 0 : set_current_state(TASK_RUNNING);
8914 0 : do_sched_yield();
8915 0 : }
8916 : EXPORT_SYMBOL(yield);
8917 :
8918 : /**
8919 : * yield_to - yield the current processor to another thread in
8920 : * your thread group, or accelerate that thread toward the
8921 : * processor it's on.
8922 : * @p: target task
8923 : * @preempt: whether task preemption is allowed or not
8924 : *
8925 : * It's the caller's job to ensure that the target task struct
8926 : * can't go away on us before we can do any checks.
8927 : *
8928 : * Return:
8929 : * true (>0) if we indeed boosted the target task.
8930 : * false (0) if we failed to boost the target.
8931 : * -ESRCH if there's no task to yield to.
8932 : */
8933 0 : int __sched yield_to(struct task_struct *p, bool preempt)
8934 : {
8935 0 : struct task_struct *curr = current;
8936 : struct rq *rq, *p_rq;
8937 : unsigned long flags;
8938 0 : int yielded = 0;
8939 :
8940 0 : local_irq_save(flags);
8941 0 : rq = this_rq();
8942 :
8943 : again:
8944 0 : p_rq = task_rq(p);
8945 : /*
8946 : * If we're the only runnable task on the rq and target rq also
8947 : * has only one task, there's absolutely no point in yielding.
8948 : */
8949 0 : if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8950 : yielded = -ESRCH;
8951 : goto out_irq;
8952 : }
8953 :
8954 0 : double_rq_lock(rq, p_rq);
8955 0 : if (task_rq(p) != p_rq) {
8956 : double_rq_unlock(rq, p_rq);
8957 : goto again;
8958 : }
8959 :
8960 0 : if (!curr->sched_class->yield_to_task)
8961 : goto out_unlock;
8962 :
8963 0 : if (curr->sched_class != p->sched_class)
8964 : goto out_unlock;
8965 :
8966 0 : if (task_on_cpu(p_rq, p) || !task_is_running(p))
8967 : goto out_unlock;
8968 :
8969 0 : yielded = curr->sched_class->yield_to_task(rq, p);
8970 : if (yielded) {
8971 : schedstat_inc(rq->yld_count);
8972 : /*
8973 : * Make p's CPU reschedule; pick_next_entity takes care of
8974 : * fairness.
8975 : */
8976 : if (preempt && rq != p_rq)
8977 : resched_curr(p_rq);
8978 : }
8979 :
8980 : out_unlock:
8981 0 : double_rq_unlock(rq, p_rq);
8982 : out_irq:
8983 0 : local_irq_restore(flags);
8984 :
8985 0 : if (yielded > 0)
8986 0 : schedule();
8987 :
8988 0 : return yielded;
8989 : }
8990 : EXPORT_SYMBOL_GPL(yield_to);
8991 :
8992 0 : int io_schedule_prepare(void)
8993 : {
8994 0 : int old_iowait = current->in_iowait;
8995 :
8996 0 : current->in_iowait = 1;
8997 0 : blk_flush_plug(current->plug, true);
8998 0 : return old_iowait;
8999 : }
9000 :
9001 0 : void io_schedule_finish(int token)
9002 : {
9003 0 : current->in_iowait = token;
9004 0 : }
9005 :
9006 : /*
9007 : * This task is about to go to sleep on IO. Increment rq->nr_iowait so
9008 : * that process accounting knows that this is a task in IO wait state.
9009 : */
9010 0 : long __sched io_schedule_timeout(long timeout)
9011 : {
9012 : int token;
9013 : long ret;
9014 :
9015 0 : token = io_schedule_prepare();
9016 0 : ret = schedule_timeout(timeout);
9017 0 : io_schedule_finish(token);
9018 :
9019 0 : return ret;
9020 : }
9021 : EXPORT_SYMBOL(io_schedule_timeout);
9022 :
9023 0 : void __sched io_schedule(void)
9024 : {
9025 : int token;
9026 :
9027 0 : token = io_schedule_prepare();
9028 0 : schedule();
9029 0 : io_schedule_finish(token);
9030 0 : }
9031 : EXPORT_SYMBOL(io_schedule);
9032 :
9033 : /**
9034 : * sys_sched_get_priority_max - return maximum RT priority.
9035 : * @policy: scheduling class.
9036 : *
9037 : * Return: On success, this syscall returns the maximum
9038 : * rt_priority that can be used by a given scheduling class.
9039 : * On failure, a negative error code is returned.
9040 : */
9041 0 : SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
9042 : {
9043 0 : int ret = -EINVAL;
9044 :
9045 : switch (policy) {
9046 : case SCHED_FIFO:
9047 : case SCHED_RR:
9048 0 : ret = MAX_RT_PRIO-1;
9049 : break;
9050 : case SCHED_DEADLINE:
9051 : case SCHED_NORMAL:
9052 : case SCHED_BATCH:
9053 : case SCHED_IDLE:
9054 : ret = 0;
9055 : break;
9056 : }
9057 0 : return ret;
9058 : }
9059 :
9060 : /**
9061 : * sys_sched_get_priority_min - return minimum RT priority.
9062 : * @policy: scheduling class.
9063 : *
9064 : * Return: On success, this syscall returns the minimum
9065 : * rt_priority that can be used by a given scheduling class.
9066 : * On failure, a negative error code is returned.
9067 : */
9068 0 : SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
9069 : {
9070 0 : int ret = -EINVAL;
9071 :
9072 : switch (policy) {
9073 : case SCHED_FIFO:
9074 : case SCHED_RR:
9075 0 : ret = 1;
9076 : break;
9077 : case SCHED_DEADLINE:
9078 : case SCHED_NORMAL:
9079 : case SCHED_BATCH:
9080 : case SCHED_IDLE:
9081 : ret = 0;
9082 : }
9083 0 : return ret;
9084 : }
9085 :
9086 0 : static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
9087 : {
9088 : struct task_struct *p;
9089 : unsigned int time_slice;
9090 : struct rq_flags rf;
9091 : struct rq *rq;
9092 : int retval;
9093 :
9094 0 : if (pid < 0)
9095 : return -EINVAL;
9096 :
9097 0 : retval = -ESRCH;
9098 : rcu_read_lock();
9099 0 : p = find_process_by_pid(pid);
9100 0 : if (!p)
9101 : goto out_unlock;
9102 :
9103 0 : retval = security_task_getscheduler(p);
9104 : if (retval)
9105 : goto out_unlock;
9106 :
9107 0 : rq = task_rq_lock(p, &rf);
9108 0 : time_slice = 0;
9109 0 : if (p->sched_class->get_rr_interval)
9110 0 : time_slice = p->sched_class->get_rr_interval(rq, p);
9111 0 : task_rq_unlock(rq, p, &rf);
9112 :
9113 : rcu_read_unlock();
9114 0 : jiffies_to_timespec64(time_slice, t);
9115 0 : return 0;
9116 :
9117 : out_unlock:
9118 : rcu_read_unlock();
9119 0 : return retval;
9120 : }
9121 :
9122 : /**
9123 : * sys_sched_rr_get_interval - return the default timeslice of a process.
9124 : * @pid: pid of the process.
9125 : * @interval: userspace pointer to the timeslice value.
9126 : *
9127 : * this syscall writes the default timeslice value of a given process
9128 : * into the user-space timespec buffer. A value of '0' means infinity.
9129 : *
9130 : * Return: On success, 0 and the timeslice is in @interval. Otherwise,
9131 : * an error code.
9132 : */
9133 0 : SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
9134 : struct __kernel_timespec __user *, interval)
9135 : {
9136 : struct timespec64 t;
9137 0 : int retval = sched_rr_get_interval(pid, &t);
9138 :
9139 0 : if (retval == 0)
9140 0 : retval = put_timespec64(&t, interval);
9141 :
9142 0 : return retval;
9143 : }
9144 :
9145 : #ifdef CONFIG_COMPAT_32BIT_TIME
9146 : SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
9147 : struct old_timespec32 __user *, interval)
9148 : {
9149 : struct timespec64 t;
9150 : int retval = sched_rr_get_interval(pid, &t);
9151 :
9152 : if (retval == 0)
9153 : retval = put_old_timespec32(&t, interval);
9154 : return retval;
9155 : }
9156 : #endif
9157 :
9158 0 : void sched_show_task(struct task_struct *p)
9159 : {
9160 0 : unsigned long free = 0;
9161 : int ppid;
9162 :
9163 0 : if (!try_get_task_stack(p))
9164 : return;
9165 :
9166 0 : pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
9167 :
9168 0 : if (task_is_running(p))
9169 0 : pr_cont(" running task ");
9170 : #ifdef CONFIG_DEBUG_STACK_USAGE
9171 : free = stack_not_used(p);
9172 : #endif
9173 0 : ppid = 0;
9174 : rcu_read_lock();
9175 0 : if (pid_alive(p))
9176 0 : ppid = task_pid_nr(rcu_dereference(p->real_parent));
9177 : rcu_read_unlock();
9178 0 : pr_cont(" stack:%-5lu pid:%-5d ppid:%-6d flags:0x%08lx\n",
9179 : free, task_pid_nr(p), ppid,
9180 : read_task_thread_flags(p));
9181 :
9182 0 : print_worker_info(KERN_INFO, p);
9183 0 : print_stop_info(KERN_INFO, p);
9184 0 : show_stack(p, NULL, KERN_INFO);
9185 0 : put_task_stack(p);
9186 : }
9187 : EXPORT_SYMBOL_GPL(sched_show_task);
9188 :
9189 : static inline bool
9190 : state_filter_match(unsigned long state_filter, struct task_struct *p)
9191 : {
9192 0 : unsigned int state = READ_ONCE(p->__state);
9193 :
9194 : /* no filter, everything matches */
9195 0 : if (!state_filter)
9196 : return true;
9197 :
9198 : /* filter, but doesn't match */
9199 0 : if (!(state & state_filter))
9200 : return false;
9201 :
9202 : /*
9203 : * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9204 : * TASK_KILLABLE).
9205 : */
9206 0 : if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
9207 : return false;
9208 :
9209 : return true;
9210 : }
9211 :
9212 :
9213 0 : void show_state_filter(unsigned int state_filter)
9214 : {
9215 : struct task_struct *g, *p;
9216 :
9217 : rcu_read_lock();
9218 0 : for_each_process_thread(g, p) {
9219 : /*
9220 : * reset the NMI-timeout, listing all files on a slow
9221 : * console might take a lot of time:
9222 : * Also, reset softlockup watchdogs on all CPUs, because
9223 : * another CPU might be blocked waiting for us to process
9224 : * an IPI.
9225 : */
9226 : touch_nmi_watchdog();
9227 : touch_all_softlockup_watchdogs();
9228 0 : if (state_filter_match(state_filter, p))
9229 0 : sched_show_task(p);
9230 : }
9231 :
9232 : #ifdef CONFIG_SCHED_DEBUG
9233 : if (!state_filter)
9234 : sysrq_sched_debug_show();
9235 : #endif
9236 : rcu_read_unlock();
9237 : /*
9238 : * Only show locks if all tasks are dumped:
9239 : */
9240 : if (!state_filter)
9241 : debug_show_all_locks();
9242 0 : }
9243 :
9244 : /**
9245 : * init_idle - set up an idle thread for a given CPU
9246 : * @idle: task in question
9247 : * @cpu: CPU the idle task belongs to
9248 : *
9249 : * NOTE: this function does not set the idle thread's NEED_RESCHED
9250 : * flag, to make booting more robust.
9251 : */
9252 1 : void __init init_idle(struct task_struct *idle, int cpu)
9253 : {
9254 : #ifdef CONFIG_SMP
9255 : struct affinity_context ac = (struct affinity_context) {
9256 : .new_mask = cpumask_of(cpu),
9257 : .flags = 0,
9258 : };
9259 : #endif
9260 1 : struct rq *rq = cpu_rq(cpu);
9261 : unsigned long flags;
9262 :
9263 1 : __sched_fork(0, idle);
9264 :
9265 1 : raw_spin_lock_irqsave(&idle->pi_lock, flags);
9266 1 : raw_spin_rq_lock(rq);
9267 :
9268 1 : idle->__state = TASK_RUNNING;
9269 1 : idle->se.exec_start = sched_clock();
9270 : /*
9271 : * PF_KTHREAD should already be set at this point; regardless, make it
9272 : * look like a proper per-CPU kthread.
9273 : */
9274 1 : idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
9275 1 : kthread_set_per_cpu(idle, cpu);
9276 :
9277 : #ifdef CONFIG_SMP
9278 : /*
9279 : * It's possible that init_idle() gets called multiple times on a task,
9280 : * in that case do_set_cpus_allowed() will not do the right thing.
9281 : *
9282 : * And since this is boot we can forgo the serialization.
9283 : */
9284 : set_cpus_allowed_common(idle, &ac);
9285 : #endif
9286 : /*
9287 : * We're having a chicken and egg problem, even though we are
9288 : * holding rq->lock, the CPU isn't yet set to this CPU so the
9289 : * lockdep check in task_group() will fail.
9290 : *
9291 : * Similar case to sched_fork(). / Alternatively we could
9292 : * use task_rq_lock() here and obtain the other rq->lock.
9293 : *
9294 : * Silence PROVE_RCU
9295 : */
9296 : rcu_read_lock();
9297 1 : __set_task_cpu(idle, cpu);
9298 : rcu_read_unlock();
9299 :
9300 1 : rq->idle = idle;
9301 1 : rcu_assign_pointer(rq->curr, idle);
9302 1 : idle->on_rq = TASK_ON_RQ_QUEUED;
9303 : #ifdef CONFIG_SMP
9304 : idle->on_cpu = 1;
9305 : #endif
9306 1 : raw_spin_rq_unlock(rq);
9307 2 : raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
9308 :
9309 : /* Set the preempt count _outside_ the spinlocks! */
9310 1 : init_idle_preempt_count(idle, cpu);
9311 :
9312 : /*
9313 : * The idle tasks have their own, simple scheduling class:
9314 : */
9315 1 : idle->sched_class = &idle_sched_class;
9316 1 : ftrace_graph_init_idle_task(idle, cpu);
9317 1 : vtime_init_idle(idle, cpu);
9318 : #ifdef CONFIG_SMP
9319 : sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9320 : #endif
9321 1 : }
9322 :
9323 : #ifdef CONFIG_SMP
9324 :
9325 : int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9326 : const struct cpumask *trial)
9327 : {
9328 : int ret = 1;
9329 :
9330 : if (cpumask_empty(cur))
9331 : return ret;
9332 :
9333 : ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9334 :
9335 : return ret;
9336 : }
9337 :
9338 : int task_can_attach(struct task_struct *p)
9339 : {
9340 : int ret = 0;
9341 :
9342 : /*
9343 : * Kthreads which disallow setaffinity shouldn't be moved
9344 : * to a new cpuset; we don't want to change their CPU
9345 : * affinity and isolating such threads by their set of
9346 : * allowed nodes is unnecessary. Thus, cpusets are not
9347 : * applicable for such threads. This prevents checking for
9348 : * success of set_cpus_allowed_ptr() on all attached tasks
9349 : * before cpus_mask may be changed.
9350 : */
9351 : if (p->flags & PF_NO_SETAFFINITY)
9352 : ret = -EINVAL;
9353 :
9354 : return ret;
9355 : }
9356 :
9357 : bool sched_smp_initialized __read_mostly;
9358 :
9359 : #ifdef CONFIG_NUMA_BALANCING
9360 : /* Migrate current task p to target_cpu */
9361 : int migrate_task_to(struct task_struct *p, int target_cpu)
9362 : {
9363 : struct migration_arg arg = { p, target_cpu };
9364 : int curr_cpu = task_cpu(p);
9365 :
9366 : if (curr_cpu == target_cpu)
9367 : return 0;
9368 :
9369 : if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9370 : return -EINVAL;
9371 :
9372 : /* TODO: This is not properly updating schedstats */
9373 :
9374 : trace_sched_move_numa(p, curr_cpu, target_cpu);
9375 : return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9376 : }
9377 :
9378 : /*
9379 : * Requeue a task on a given node and accurately track the number of NUMA
9380 : * tasks on the runqueues
9381 : */
9382 : void sched_setnuma(struct task_struct *p, int nid)
9383 : {
9384 : bool queued, running;
9385 : struct rq_flags rf;
9386 : struct rq *rq;
9387 :
9388 : rq = task_rq_lock(p, &rf);
9389 : queued = task_on_rq_queued(p);
9390 : running = task_current(rq, p);
9391 :
9392 : if (queued)
9393 : dequeue_task(rq, p, DEQUEUE_SAVE);
9394 : if (running)
9395 : put_prev_task(rq, p);
9396 :
9397 : p->numa_preferred_nid = nid;
9398 :
9399 : if (queued)
9400 : enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9401 : if (running)
9402 : set_next_task(rq, p);
9403 : task_rq_unlock(rq, p, &rf);
9404 : }
9405 : #endif /* CONFIG_NUMA_BALANCING */
9406 :
9407 : #ifdef CONFIG_HOTPLUG_CPU
9408 : /*
9409 : * Ensure that the idle task is using init_mm right before its CPU goes
9410 : * offline.
9411 : */
9412 : void idle_task_exit(void)
9413 : {
9414 : struct mm_struct *mm = current->active_mm;
9415 :
9416 : BUG_ON(cpu_online(smp_processor_id()));
9417 : BUG_ON(current != this_rq()->idle);
9418 :
9419 : if (mm != &init_mm) {
9420 : switch_mm(mm, &init_mm, current);
9421 : finish_arch_post_lock_switch();
9422 : }
9423 :
9424 : /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9425 : }
9426 :
9427 : static int __balance_push_cpu_stop(void *arg)
9428 : {
9429 : struct task_struct *p = arg;
9430 : struct rq *rq = this_rq();
9431 : struct rq_flags rf;
9432 : int cpu;
9433 :
9434 : raw_spin_lock_irq(&p->pi_lock);
9435 : rq_lock(rq, &rf);
9436 :
9437 : update_rq_clock(rq);
9438 :
9439 : if (task_rq(p) == rq && task_on_rq_queued(p)) {
9440 : cpu = select_fallback_rq(rq->cpu, p);
9441 : rq = __migrate_task(rq, &rf, p, cpu);
9442 : }
9443 :
9444 : rq_unlock(rq, &rf);
9445 : raw_spin_unlock_irq(&p->pi_lock);
9446 :
9447 : put_task_struct(p);
9448 :
9449 : return 0;
9450 : }
9451 :
9452 : static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9453 :
9454 : /*
9455 : * Ensure we only run per-cpu kthreads once the CPU goes !active.
9456 : *
9457 : * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9458 : * effective when the hotplug motion is down.
9459 : */
9460 : static void balance_push(struct rq *rq)
9461 : {
9462 : struct task_struct *push_task = rq->curr;
9463 :
9464 : lockdep_assert_rq_held(rq);
9465 :
9466 : /*
9467 : * Ensure the thing is persistent until balance_push_set(.on = false);
9468 : */
9469 : rq->balance_callback = &balance_push_callback;
9470 :
9471 : /*
9472 : * Only active while going offline and when invoked on the outgoing
9473 : * CPU.
9474 : */
9475 : if (!cpu_dying(rq->cpu) || rq != this_rq())
9476 : return;
9477 :
9478 : /*
9479 : * Both the cpu-hotplug and stop task are in this case and are
9480 : * required to complete the hotplug process.
9481 : */
9482 : if (kthread_is_per_cpu(push_task) ||
9483 : is_migration_disabled(push_task)) {
9484 :
9485 : /*
9486 : * If this is the idle task on the outgoing CPU try to wake
9487 : * up the hotplug control thread which might wait for the
9488 : * last task to vanish. The rcuwait_active() check is
9489 : * accurate here because the waiter is pinned on this CPU
9490 : * and can't obviously be running in parallel.
9491 : *
9492 : * On RT kernels this also has to check whether there are
9493 : * pinned and scheduled out tasks on the runqueue. They
9494 : * need to leave the migrate disabled section first.
9495 : */
9496 : if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9497 : rcuwait_active(&rq->hotplug_wait)) {
9498 : raw_spin_rq_unlock(rq);
9499 : rcuwait_wake_up(&rq->hotplug_wait);
9500 : raw_spin_rq_lock(rq);
9501 : }
9502 : return;
9503 : }
9504 :
9505 : get_task_struct(push_task);
9506 : /*
9507 : * Temporarily drop rq->lock such that we can wake-up the stop task.
9508 : * Both preemption and IRQs are still disabled.
9509 : */
9510 : raw_spin_rq_unlock(rq);
9511 : stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9512 : this_cpu_ptr(&push_work));
9513 : /*
9514 : * At this point need_resched() is true and we'll take the loop in
9515 : * schedule(). The next pick is obviously going to be the stop task
9516 : * which kthread_is_per_cpu() and will push this task away.
9517 : */
9518 : raw_spin_rq_lock(rq);
9519 : }
9520 :
9521 : static void balance_push_set(int cpu, bool on)
9522 : {
9523 : struct rq *rq = cpu_rq(cpu);
9524 : struct rq_flags rf;
9525 :
9526 : rq_lock_irqsave(rq, &rf);
9527 : if (on) {
9528 : WARN_ON_ONCE(rq->balance_callback);
9529 : rq->balance_callback = &balance_push_callback;
9530 : } else if (rq->balance_callback == &balance_push_callback) {
9531 : rq->balance_callback = NULL;
9532 : }
9533 : rq_unlock_irqrestore(rq, &rf);
9534 : }
9535 :
9536 : /*
9537 : * Invoked from a CPUs hotplug control thread after the CPU has been marked
9538 : * inactive. All tasks which are not per CPU kernel threads are either
9539 : * pushed off this CPU now via balance_push() or placed on a different CPU
9540 : * during wakeup. Wait until the CPU is quiescent.
9541 : */
9542 : static void balance_hotplug_wait(void)
9543 : {
9544 : struct rq *rq = this_rq();
9545 :
9546 : rcuwait_wait_event(&rq->hotplug_wait,
9547 : rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9548 : TASK_UNINTERRUPTIBLE);
9549 : }
9550 :
9551 : #else
9552 :
9553 : static inline void balance_push(struct rq *rq)
9554 : {
9555 : }
9556 :
9557 : static inline void balance_push_set(int cpu, bool on)
9558 : {
9559 : }
9560 :
9561 : static inline void balance_hotplug_wait(void)
9562 : {
9563 : }
9564 :
9565 : #endif /* CONFIG_HOTPLUG_CPU */
9566 :
9567 : void set_rq_online(struct rq *rq)
9568 : {
9569 : if (!rq->online) {
9570 : const struct sched_class *class;
9571 :
9572 : cpumask_set_cpu(rq->cpu, rq->rd->online);
9573 : rq->online = 1;
9574 :
9575 : for_each_class(class) {
9576 : if (class->rq_online)
9577 : class->rq_online(rq);
9578 : }
9579 : }
9580 : }
9581 :
9582 : void set_rq_offline(struct rq *rq)
9583 : {
9584 : if (rq->online) {
9585 : const struct sched_class *class;
9586 :
9587 : update_rq_clock(rq);
9588 : for_each_class(class) {
9589 : if (class->rq_offline)
9590 : class->rq_offline(rq);
9591 : }
9592 :
9593 : cpumask_clear_cpu(rq->cpu, rq->rd->online);
9594 : rq->online = 0;
9595 : }
9596 : }
9597 :
9598 : /*
9599 : * used to mark begin/end of suspend/resume:
9600 : */
9601 : static int num_cpus_frozen;
9602 :
9603 : /*
9604 : * Update cpusets according to cpu_active mask. If cpusets are
9605 : * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9606 : * around partition_sched_domains().
9607 : *
9608 : * If we come here as part of a suspend/resume, don't touch cpusets because we
9609 : * want to restore it back to its original state upon resume anyway.
9610 : */
9611 : static void cpuset_cpu_active(void)
9612 : {
9613 : if (cpuhp_tasks_frozen) {
9614 : /*
9615 : * num_cpus_frozen tracks how many CPUs are involved in suspend
9616 : * resume sequence. As long as this is not the last online
9617 : * operation in the resume sequence, just build a single sched
9618 : * domain, ignoring cpusets.
9619 : */
9620 : partition_sched_domains(1, NULL, NULL);
9621 : if (--num_cpus_frozen)
9622 : return;
9623 : /*
9624 : * This is the last CPU online operation. So fall through and
9625 : * restore the original sched domains by considering the
9626 : * cpuset configurations.
9627 : */
9628 : cpuset_force_rebuild();
9629 : }
9630 : cpuset_update_active_cpus();
9631 : }
9632 :
9633 : static int cpuset_cpu_inactive(unsigned int cpu)
9634 : {
9635 : if (!cpuhp_tasks_frozen) {
9636 : int ret = dl_bw_check_overflow(cpu);
9637 :
9638 : if (ret)
9639 : return ret;
9640 : cpuset_update_active_cpus();
9641 : } else {
9642 : num_cpus_frozen++;
9643 : partition_sched_domains(1, NULL, NULL);
9644 : }
9645 : return 0;
9646 : }
9647 :
9648 : int sched_cpu_activate(unsigned int cpu)
9649 : {
9650 : struct rq *rq = cpu_rq(cpu);
9651 : struct rq_flags rf;
9652 :
9653 : /*
9654 : * Clear the balance_push callback and prepare to schedule
9655 : * regular tasks.
9656 : */
9657 : balance_push_set(cpu, false);
9658 :
9659 : #ifdef CONFIG_SCHED_SMT
9660 : /*
9661 : * When going up, increment the number of cores with SMT present.
9662 : */
9663 : if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9664 : static_branch_inc_cpuslocked(&sched_smt_present);
9665 : #endif
9666 : set_cpu_active(cpu, true);
9667 :
9668 : if (sched_smp_initialized) {
9669 : sched_update_numa(cpu, true);
9670 : sched_domains_numa_masks_set(cpu);
9671 : cpuset_cpu_active();
9672 : }
9673 :
9674 : /*
9675 : * Put the rq online, if not already. This happens:
9676 : *
9677 : * 1) In the early boot process, because we build the real domains
9678 : * after all CPUs have been brought up.
9679 : *
9680 : * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9681 : * domains.
9682 : */
9683 : rq_lock_irqsave(rq, &rf);
9684 : if (rq->rd) {
9685 : BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9686 : set_rq_online(rq);
9687 : }
9688 : rq_unlock_irqrestore(rq, &rf);
9689 :
9690 : return 0;
9691 : }
9692 :
9693 : int sched_cpu_deactivate(unsigned int cpu)
9694 : {
9695 : struct rq *rq = cpu_rq(cpu);
9696 : struct rq_flags rf;
9697 : int ret;
9698 :
9699 : /*
9700 : * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9701 : * load balancing when not active
9702 : */
9703 : nohz_balance_exit_idle(rq);
9704 :
9705 : set_cpu_active(cpu, false);
9706 :
9707 : /*
9708 : * From this point forward, this CPU will refuse to run any task that
9709 : * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9710 : * push those tasks away until this gets cleared, see
9711 : * sched_cpu_dying().
9712 : */
9713 : balance_push_set(cpu, true);
9714 :
9715 : /*
9716 : * We've cleared cpu_active_mask / set balance_push, wait for all
9717 : * preempt-disabled and RCU users of this state to go away such that
9718 : * all new such users will observe it.
9719 : *
9720 : * Specifically, we rely on ttwu to no longer target this CPU, see
9721 : * ttwu_queue_cond() and is_cpu_allowed().
9722 : *
9723 : * Do sync before park smpboot threads to take care the rcu boost case.
9724 : */
9725 : synchronize_rcu();
9726 :
9727 : rq_lock_irqsave(rq, &rf);
9728 : if (rq->rd) {
9729 : BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9730 : set_rq_offline(rq);
9731 : }
9732 : rq_unlock_irqrestore(rq, &rf);
9733 :
9734 : #ifdef CONFIG_SCHED_SMT
9735 : /*
9736 : * When going down, decrement the number of cores with SMT present.
9737 : */
9738 : if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9739 : static_branch_dec_cpuslocked(&sched_smt_present);
9740 :
9741 : sched_core_cpu_deactivate(cpu);
9742 : #endif
9743 :
9744 : if (!sched_smp_initialized)
9745 : return 0;
9746 :
9747 : sched_update_numa(cpu, false);
9748 : ret = cpuset_cpu_inactive(cpu);
9749 : if (ret) {
9750 : balance_push_set(cpu, false);
9751 : set_cpu_active(cpu, true);
9752 : sched_update_numa(cpu, true);
9753 : return ret;
9754 : }
9755 : sched_domains_numa_masks_clear(cpu);
9756 : return 0;
9757 : }
9758 :
9759 : static void sched_rq_cpu_starting(unsigned int cpu)
9760 : {
9761 : struct rq *rq = cpu_rq(cpu);
9762 :
9763 : rq->calc_load_update = calc_load_update;
9764 : update_max_interval();
9765 : }
9766 :
9767 : int sched_cpu_starting(unsigned int cpu)
9768 : {
9769 : sched_core_cpu_starting(cpu);
9770 : sched_rq_cpu_starting(cpu);
9771 : sched_tick_start(cpu);
9772 : return 0;
9773 : }
9774 :
9775 : #ifdef CONFIG_HOTPLUG_CPU
9776 :
9777 : /*
9778 : * Invoked immediately before the stopper thread is invoked to bring the
9779 : * CPU down completely. At this point all per CPU kthreads except the
9780 : * hotplug thread (current) and the stopper thread (inactive) have been
9781 : * either parked or have been unbound from the outgoing CPU. Ensure that
9782 : * any of those which might be on the way out are gone.
9783 : *
9784 : * If after this point a bound task is being woken on this CPU then the
9785 : * responsible hotplug callback has failed to do it's job.
9786 : * sched_cpu_dying() will catch it with the appropriate fireworks.
9787 : */
9788 : int sched_cpu_wait_empty(unsigned int cpu)
9789 : {
9790 : balance_hotplug_wait();
9791 : return 0;
9792 : }
9793 :
9794 : /*
9795 : * Since this CPU is going 'away' for a while, fold any nr_active delta we
9796 : * might have. Called from the CPU stopper task after ensuring that the
9797 : * stopper is the last running task on the CPU, so nr_active count is
9798 : * stable. We need to take the teardown thread which is calling this into
9799 : * account, so we hand in adjust = 1 to the load calculation.
9800 : *
9801 : * Also see the comment "Global load-average calculations".
9802 : */
9803 : static void calc_load_migrate(struct rq *rq)
9804 : {
9805 : long delta = calc_load_fold_active(rq, 1);
9806 :
9807 : if (delta)
9808 : atomic_long_add(delta, &calc_load_tasks);
9809 : }
9810 :
9811 : static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9812 : {
9813 : struct task_struct *g, *p;
9814 : int cpu = cpu_of(rq);
9815 :
9816 : lockdep_assert_rq_held(rq);
9817 :
9818 : printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9819 : for_each_process_thread(g, p) {
9820 : if (task_cpu(p) != cpu)
9821 : continue;
9822 :
9823 : if (!task_on_rq_queued(p))
9824 : continue;
9825 :
9826 : printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9827 : }
9828 : }
9829 :
9830 : int sched_cpu_dying(unsigned int cpu)
9831 : {
9832 : struct rq *rq = cpu_rq(cpu);
9833 : struct rq_flags rf;
9834 :
9835 : /* Handle pending wakeups and then migrate everything off */
9836 : sched_tick_stop(cpu);
9837 :
9838 : rq_lock_irqsave(rq, &rf);
9839 : if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9840 : WARN(true, "Dying CPU not properly vacated!");
9841 : dump_rq_tasks(rq, KERN_WARNING);
9842 : }
9843 : rq_unlock_irqrestore(rq, &rf);
9844 :
9845 : calc_load_migrate(rq);
9846 : update_max_interval();
9847 : hrtick_clear(rq);
9848 : sched_core_cpu_dying(cpu);
9849 : return 0;
9850 : }
9851 : #endif
9852 :
9853 : void __init sched_init_smp(void)
9854 : {
9855 : sched_init_numa(NUMA_NO_NODE);
9856 :
9857 : /*
9858 : * There's no userspace yet to cause hotplug operations; hence all the
9859 : * CPU masks are stable and all blatant races in the below code cannot
9860 : * happen.
9861 : */
9862 : mutex_lock(&sched_domains_mutex);
9863 : sched_init_domains(cpu_active_mask);
9864 : mutex_unlock(&sched_domains_mutex);
9865 :
9866 : /* Move init over to a non-isolated CPU */
9867 : if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9868 : BUG();
9869 : current->flags &= ~PF_NO_SETAFFINITY;
9870 : sched_init_granularity();
9871 :
9872 : init_sched_rt_class();
9873 : init_sched_dl_class();
9874 :
9875 : sched_smp_initialized = true;
9876 : }
9877 :
9878 : static int __init migration_init(void)
9879 : {
9880 : sched_cpu_starting(smp_processor_id());
9881 : return 0;
9882 : }
9883 : early_initcall(migration_init);
9884 :
9885 : #else
9886 1 : void __init sched_init_smp(void)
9887 : {
9888 1 : sched_init_granularity();
9889 1 : }
9890 : #endif /* CONFIG_SMP */
9891 :
9892 0 : int in_sched_functions(unsigned long addr)
9893 : {
9894 0 : return in_lock_functions(addr) ||
9895 0 : (addr >= (unsigned long)__sched_text_start
9896 0 : && addr < (unsigned long)__sched_text_end);
9897 : }
9898 :
9899 : #ifdef CONFIG_CGROUP_SCHED
9900 : /*
9901 : * Default task group.
9902 : * Every task in system belongs to this group at bootup.
9903 : */
9904 : struct task_group root_task_group;
9905 : LIST_HEAD(task_groups);
9906 :
9907 : /* Cacheline aligned slab cache for task_group */
9908 : static struct kmem_cache *task_group_cache __read_mostly;
9909 : #endif
9910 :
9911 1 : void __init sched_init(void)
9912 : {
9913 1 : unsigned long ptr = 0;
9914 : int i;
9915 :
9916 : /* Make sure the linker didn't screw up */
9917 1 : BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9918 : &fair_sched_class != &rt_sched_class + 1 ||
9919 : &rt_sched_class != &dl_sched_class + 1);
9920 : #ifdef CONFIG_SMP
9921 : BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9922 : #endif
9923 :
9924 1 : wait_bit_init();
9925 :
9926 : #ifdef CONFIG_FAIR_GROUP_SCHED
9927 : ptr += 2 * nr_cpu_ids * sizeof(void **);
9928 : #endif
9929 : #ifdef CONFIG_RT_GROUP_SCHED
9930 : ptr += 2 * nr_cpu_ids * sizeof(void **);
9931 : #endif
9932 : if (ptr) {
9933 : ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9934 :
9935 : #ifdef CONFIG_FAIR_GROUP_SCHED
9936 : root_task_group.se = (struct sched_entity **)ptr;
9937 : ptr += nr_cpu_ids * sizeof(void **);
9938 :
9939 : root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9940 : ptr += nr_cpu_ids * sizeof(void **);
9941 :
9942 : root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9943 : init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9944 : #endif /* CONFIG_FAIR_GROUP_SCHED */
9945 : #ifdef CONFIG_RT_GROUP_SCHED
9946 : root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9947 : ptr += nr_cpu_ids * sizeof(void **);
9948 :
9949 : root_task_group.rt_rq = (struct rt_rq **)ptr;
9950 : ptr += nr_cpu_ids * sizeof(void **);
9951 :
9952 : #endif /* CONFIG_RT_GROUP_SCHED */
9953 : }
9954 :
9955 1 : init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9956 :
9957 : #ifdef CONFIG_SMP
9958 : init_defrootdomain();
9959 : #endif
9960 :
9961 : #ifdef CONFIG_RT_GROUP_SCHED
9962 : init_rt_bandwidth(&root_task_group.rt_bandwidth,
9963 : global_rt_period(), global_rt_runtime());
9964 : #endif /* CONFIG_RT_GROUP_SCHED */
9965 :
9966 : #ifdef CONFIG_CGROUP_SCHED
9967 : task_group_cache = KMEM_CACHE(task_group, 0);
9968 :
9969 : list_add(&root_task_group.list, &task_groups);
9970 : INIT_LIST_HEAD(&root_task_group.children);
9971 : INIT_LIST_HEAD(&root_task_group.siblings);
9972 : autogroup_init(&init_task);
9973 : #endif /* CONFIG_CGROUP_SCHED */
9974 :
9975 2 : for_each_possible_cpu(i) {
9976 : struct rq *rq;
9977 :
9978 1 : rq = cpu_rq(i);
9979 : raw_spin_lock_init(&rq->__lock);
9980 1 : rq->nr_running = 0;
9981 1 : rq->calc_load_active = 0;
9982 1 : rq->calc_load_update = jiffies + LOAD_FREQ;
9983 1 : init_cfs_rq(&rq->cfs);
9984 1 : init_rt_rq(&rq->rt);
9985 1 : init_dl_rq(&rq->dl);
9986 : #ifdef CONFIG_FAIR_GROUP_SCHED
9987 : INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9988 : rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9989 : /*
9990 : * How much CPU bandwidth does root_task_group get?
9991 : *
9992 : * In case of task-groups formed thr' the cgroup filesystem, it
9993 : * gets 100% of the CPU resources in the system. This overall
9994 : * system CPU resource is divided among the tasks of
9995 : * root_task_group and its child task-groups in a fair manner,
9996 : * based on each entity's (task or task-group's) weight
9997 : * (se->load.weight).
9998 : *
9999 : * In other words, if root_task_group has 10 tasks of weight
10000 : * 1024) and two child groups A0 and A1 (of weight 1024 each),
10001 : * then A0's share of the CPU resource is:
10002 : *
10003 : * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
10004 : *
10005 : * We achieve this by letting root_task_group's tasks sit
10006 : * directly in rq->cfs (i.e root_task_group->se[] = NULL).
10007 : */
10008 : init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
10009 : #endif /* CONFIG_FAIR_GROUP_SCHED */
10010 :
10011 1 : rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
10012 : #ifdef CONFIG_RT_GROUP_SCHED
10013 : init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
10014 : #endif
10015 : #ifdef CONFIG_SMP
10016 : rq->sd = NULL;
10017 : rq->rd = NULL;
10018 : rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
10019 : rq->balance_callback = &balance_push_callback;
10020 : rq->active_balance = 0;
10021 : rq->next_balance = jiffies;
10022 : rq->push_cpu = 0;
10023 : rq->cpu = i;
10024 : rq->online = 0;
10025 : rq->idle_stamp = 0;
10026 : rq->avg_idle = 2*sysctl_sched_migration_cost;
10027 : rq->wake_stamp = jiffies;
10028 : rq->wake_avg_idle = rq->avg_idle;
10029 : rq->max_idle_balance_cost = sysctl_sched_migration_cost;
10030 :
10031 : INIT_LIST_HEAD(&rq->cfs_tasks);
10032 :
10033 : rq_attach_root(rq, &def_root_domain);
10034 : #ifdef CONFIG_NO_HZ_COMMON
10035 : rq->last_blocked_load_update_tick = jiffies;
10036 : atomic_set(&rq->nohz_flags, 0);
10037 :
10038 : INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
10039 : #endif
10040 : #ifdef CONFIG_HOTPLUG_CPU
10041 : rcuwait_init(&rq->hotplug_wait);
10042 : #endif
10043 : #endif /* CONFIG_SMP */
10044 1 : hrtick_rq_init(rq);
10045 2 : atomic_set(&rq->nr_iowait, 0);
10046 :
10047 : #ifdef CONFIG_SCHED_CORE
10048 : rq->core = rq;
10049 : rq->core_pick = NULL;
10050 : rq->core_enabled = 0;
10051 : rq->core_tree = RB_ROOT;
10052 : rq->core_forceidle_count = 0;
10053 : rq->core_forceidle_occupation = 0;
10054 : rq->core_forceidle_start = 0;
10055 :
10056 : rq->core_cookie = 0UL;
10057 : #endif
10058 2 : zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
10059 : }
10060 :
10061 1 : set_load_weight(&init_task, false);
10062 :
10063 : /*
10064 : * The boot idle thread does lazy MMU switching as well:
10065 : */
10066 1 : mmgrab_lazy_tlb(&init_mm);
10067 1 : enter_lazy_tlb(&init_mm, current);
10068 :
10069 : /*
10070 : * The idle task doesn't need the kthread struct to function, but it
10071 : * is dressed up as a per-CPU kthread and thus needs to play the part
10072 : * if we want to avoid special-casing it in code that deals with per-CPU
10073 : * kthreads.
10074 : */
10075 1 : WARN_ON(!set_kthread_struct(current));
10076 :
10077 : /*
10078 : * Make us the idle thread. Technically, schedule() should not be
10079 : * called from this thread, however somewhere below it might be,
10080 : * but because we are the idle thread, we just pick up running again
10081 : * when this runqueue becomes "idle".
10082 : */
10083 1 : init_idle(current, smp_processor_id());
10084 :
10085 1 : calc_load_update = jiffies + LOAD_FREQ;
10086 :
10087 : #ifdef CONFIG_SMP
10088 : idle_thread_set_boot_cpu();
10089 : balance_push_set(smp_processor_id(), false);
10090 : #endif
10091 1 : init_sched_fair_class();
10092 :
10093 : psi_init();
10094 :
10095 : init_uclamp();
10096 :
10097 : preempt_dynamic_init();
10098 :
10099 1 : scheduler_running = 1;
10100 1 : }
10101 :
10102 : #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
10103 :
10104 : void __might_sleep(const char *file, int line)
10105 : {
10106 : unsigned int state = get_current_state();
10107 : /*
10108 : * Blocking primitives will set (and therefore destroy) current->state,
10109 : * since we will exit with TASK_RUNNING make sure we enter with it,
10110 : * otherwise we will destroy state.
10111 : */
10112 : WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
10113 : "do not call blocking ops when !TASK_RUNNING; "
10114 : "state=%x set at [<%p>] %pS\n", state,
10115 : (void *)current->task_state_change,
10116 : (void *)current->task_state_change);
10117 :
10118 : __might_resched(file, line, 0);
10119 : }
10120 : EXPORT_SYMBOL(__might_sleep);
10121 :
10122 : static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
10123 : {
10124 : if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
10125 : return;
10126 :
10127 : if (preempt_count() == preempt_offset)
10128 : return;
10129 :
10130 : pr_err("Preemption disabled at:");
10131 : print_ip_sym(KERN_ERR, ip);
10132 : }
10133 :
10134 : static inline bool resched_offsets_ok(unsigned int offsets)
10135 : {
10136 : unsigned int nested = preempt_count();
10137 :
10138 : nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
10139 :
10140 : return nested == offsets;
10141 : }
10142 :
10143 : void __might_resched(const char *file, int line, unsigned int offsets)
10144 : {
10145 : /* Ratelimiting timestamp: */
10146 : static unsigned long prev_jiffy;
10147 :
10148 : unsigned long preempt_disable_ip;
10149 :
10150 : /* WARN_ON_ONCE() by default, no rate limit required: */
10151 : rcu_sleep_check();
10152 :
10153 : if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
10154 : !is_idle_task(current) && !current->non_block_count) ||
10155 : system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10156 : oops_in_progress)
10157 : return;
10158 :
10159 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10160 : return;
10161 : prev_jiffy = jiffies;
10162 :
10163 : /* Save this before calling printk(), since that will clobber it: */
10164 : preempt_disable_ip = get_preempt_disable_ip(current);
10165 :
10166 : pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10167 : file, line);
10168 : pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10169 : in_atomic(), irqs_disabled(), current->non_block_count,
10170 : current->pid, current->comm);
10171 : pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10172 : offsets & MIGHT_RESCHED_PREEMPT_MASK);
10173 :
10174 : if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
10175 : pr_err("RCU nest depth: %d, expected: %u\n",
10176 : rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
10177 : }
10178 :
10179 : if (task_stack_end_corrupted(current))
10180 : pr_emerg("Thread overran stack, or stack corrupted\n");
10181 :
10182 : debug_show_held_locks(current);
10183 : if (irqs_disabled())
10184 : print_irqtrace_events(current);
10185 :
10186 : print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
10187 : preempt_disable_ip);
10188 :
10189 : dump_stack();
10190 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10191 : }
10192 : EXPORT_SYMBOL(__might_resched);
10193 :
10194 : void __cant_sleep(const char *file, int line, int preempt_offset)
10195 : {
10196 : static unsigned long prev_jiffy;
10197 :
10198 : if (irqs_disabled())
10199 : return;
10200 :
10201 : if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10202 : return;
10203 :
10204 : if (preempt_count() > preempt_offset)
10205 : return;
10206 :
10207 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10208 : return;
10209 : prev_jiffy = jiffies;
10210 :
10211 : printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10212 : printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10213 : in_atomic(), irqs_disabled(),
10214 : current->pid, current->comm);
10215 :
10216 : debug_show_held_locks(current);
10217 : dump_stack();
10218 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10219 : }
10220 : EXPORT_SYMBOL_GPL(__cant_sleep);
10221 :
10222 : #ifdef CONFIG_SMP
10223 : void __cant_migrate(const char *file, int line)
10224 : {
10225 : static unsigned long prev_jiffy;
10226 :
10227 : if (irqs_disabled())
10228 : return;
10229 :
10230 : if (is_migration_disabled(current))
10231 : return;
10232 :
10233 : if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10234 : return;
10235 :
10236 : if (preempt_count() > 0)
10237 : return;
10238 :
10239 : if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10240 : return;
10241 : prev_jiffy = jiffies;
10242 :
10243 : pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10244 : pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10245 : in_atomic(), irqs_disabled(), is_migration_disabled(current),
10246 : current->pid, current->comm);
10247 :
10248 : debug_show_held_locks(current);
10249 : dump_stack();
10250 : add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10251 : }
10252 : EXPORT_SYMBOL_GPL(__cant_migrate);
10253 : #endif
10254 : #endif
10255 :
10256 : #ifdef CONFIG_MAGIC_SYSRQ
10257 : void normalize_rt_tasks(void)
10258 : {
10259 : struct task_struct *g, *p;
10260 : struct sched_attr attr = {
10261 : .sched_policy = SCHED_NORMAL,
10262 : };
10263 :
10264 : read_lock(&tasklist_lock);
10265 : for_each_process_thread(g, p) {
10266 : /*
10267 : * Only normalize user tasks:
10268 : */
10269 : if (p->flags & PF_KTHREAD)
10270 : continue;
10271 :
10272 : p->se.exec_start = 0;
10273 : schedstat_set(p->stats.wait_start, 0);
10274 : schedstat_set(p->stats.sleep_start, 0);
10275 : schedstat_set(p->stats.block_start, 0);
10276 :
10277 : if (!dl_task(p) && !rt_task(p)) {
10278 : /*
10279 : * Renice negative nice level userspace
10280 : * tasks back to 0:
10281 : */
10282 : if (task_nice(p) < 0)
10283 : set_user_nice(p, 0);
10284 : continue;
10285 : }
10286 :
10287 : __sched_setscheduler(p, &attr, false, false);
10288 : }
10289 : read_unlock(&tasklist_lock);
10290 : }
10291 :
10292 : #endif /* CONFIG_MAGIC_SYSRQ */
10293 :
10294 : #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
10295 : /*
10296 : * These functions are only useful for the IA64 MCA handling, or kdb.
10297 : *
10298 : * They can only be called when the whole system has been
10299 : * stopped - every CPU needs to be quiescent, and no scheduling
10300 : * activity can take place. Using them for anything else would
10301 : * be a serious bug, and as a result, they aren't even visible
10302 : * under any other configuration.
10303 : */
10304 :
10305 : /**
10306 : * curr_task - return the current task for a given CPU.
10307 : * @cpu: the processor in question.
10308 : *
10309 : * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10310 : *
10311 : * Return: The current task for @cpu.
10312 : */
10313 : struct task_struct *curr_task(int cpu)
10314 : {
10315 : return cpu_curr(cpu);
10316 : }
10317 :
10318 : #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10319 :
10320 : #ifdef CONFIG_IA64
10321 : /**
10322 : * ia64_set_curr_task - set the current task for a given CPU.
10323 : * @cpu: the processor in question.
10324 : * @p: the task pointer to set.
10325 : *
10326 : * Description: This function must only be used when non-maskable interrupts
10327 : * are serviced on a separate stack. It allows the architecture to switch the
10328 : * notion of the current task on a CPU in a non-blocking manner. This function
10329 : * must be called with all CPU's synchronized, and interrupts disabled, the
10330 : * and caller must save the original value of the current task (see
10331 : * curr_task() above) and restore that value before reenabling interrupts and
10332 : * re-starting the system.
10333 : *
10334 : * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10335 : */
10336 : void ia64_set_curr_task(int cpu, struct task_struct *p)
10337 : {
10338 : cpu_curr(cpu) = p;
10339 : }
10340 :
10341 : #endif
10342 :
10343 : #ifdef CONFIG_CGROUP_SCHED
10344 : /* task_group_lock serializes the addition/removal of task groups */
10345 : static DEFINE_SPINLOCK(task_group_lock);
10346 :
10347 : static inline void alloc_uclamp_sched_group(struct task_group *tg,
10348 : struct task_group *parent)
10349 : {
10350 : #ifdef CONFIG_UCLAMP_TASK_GROUP
10351 : enum uclamp_id clamp_id;
10352 :
10353 : for_each_clamp_id(clamp_id) {
10354 : uclamp_se_set(&tg->uclamp_req[clamp_id],
10355 : uclamp_none(clamp_id), false);
10356 : tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10357 : }
10358 : #endif
10359 : }
10360 :
10361 : static void sched_free_group(struct task_group *tg)
10362 : {
10363 : free_fair_sched_group(tg);
10364 : free_rt_sched_group(tg);
10365 : autogroup_free(tg);
10366 : kmem_cache_free(task_group_cache, tg);
10367 : }
10368 :
10369 : static void sched_free_group_rcu(struct rcu_head *rcu)
10370 : {
10371 : sched_free_group(container_of(rcu, struct task_group, rcu));
10372 : }
10373 :
10374 : static void sched_unregister_group(struct task_group *tg)
10375 : {
10376 : unregister_fair_sched_group(tg);
10377 : unregister_rt_sched_group(tg);
10378 : /*
10379 : * We have to wait for yet another RCU grace period to expire, as
10380 : * print_cfs_stats() might run concurrently.
10381 : */
10382 : call_rcu(&tg->rcu, sched_free_group_rcu);
10383 : }
10384 :
10385 : /* allocate runqueue etc for a new task group */
10386 : struct task_group *sched_create_group(struct task_group *parent)
10387 : {
10388 : struct task_group *tg;
10389 :
10390 : tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10391 : if (!tg)
10392 : return ERR_PTR(-ENOMEM);
10393 :
10394 : if (!alloc_fair_sched_group(tg, parent))
10395 : goto err;
10396 :
10397 : if (!alloc_rt_sched_group(tg, parent))
10398 : goto err;
10399 :
10400 : alloc_uclamp_sched_group(tg, parent);
10401 :
10402 : return tg;
10403 :
10404 : err:
10405 : sched_free_group(tg);
10406 : return ERR_PTR(-ENOMEM);
10407 : }
10408 :
10409 : void sched_online_group(struct task_group *tg, struct task_group *parent)
10410 : {
10411 : unsigned long flags;
10412 :
10413 : spin_lock_irqsave(&task_group_lock, flags);
10414 : list_add_rcu(&tg->list, &task_groups);
10415 :
10416 : /* Root should already exist: */
10417 : WARN_ON(!parent);
10418 :
10419 : tg->parent = parent;
10420 : INIT_LIST_HEAD(&tg->children);
10421 : list_add_rcu(&tg->siblings, &parent->children);
10422 : spin_unlock_irqrestore(&task_group_lock, flags);
10423 :
10424 : online_fair_sched_group(tg);
10425 : }
10426 :
10427 : /* rcu callback to free various structures associated with a task group */
10428 : static void sched_unregister_group_rcu(struct rcu_head *rhp)
10429 : {
10430 : /* Now it should be safe to free those cfs_rqs: */
10431 : sched_unregister_group(container_of(rhp, struct task_group, rcu));
10432 : }
10433 :
10434 : void sched_destroy_group(struct task_group *tg)
10435 : {
10436 : /* Wait for possible concurrent references to cfs_rqs complete: */
10437 : call_rcu(&tg->rcu, sched_unregister_group_rcu);
10438 : }
10439 :
10440 : void sched_release_group(struct task_group *tg)
10441 : {
10442 : unsigned long flags;
10443 :
10444 : /*
10445 : * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10446 : * sched_cfs_period_timer()).
10447 : *
10448 : * For this to be effective, we have to wait for all pending users of
10449 : * this task group to leave their RCU critical section to ensure no new
10450 : * user will see our dying task group any more. Specifically ensure
10451 : * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10452 : *
10453 : * We therefore defer calling unregister_fair_sched_group() to
10454 : * sched_unregister_group() which is guarantied to get called only after the
10455 : * current RCU grace period has expired.
10456 : */
10457 : spin_lock_irqsave(&task_group_lock, flags);
10458 : list_del_rcu(&tg->list);
10459 : list_del_rcu(&tg->siblings);
10460 : spin_unlock_irqrestore(&task_group_lock, flags);
10461 : }
10462 :
10463 : static struct task_group *sched_get_task_group(struct task_struct *tsk)
10464 : {
10465 : struct task_group *tg;
10466 :
10467 : /*
10468 : * All callers are synchronized by task_rq_lock(); we do not use RCU
10469 : * which is pointless here. Thus, we pass "true" to task_css_check()
10470 : * to prevent lockdep warnings.
10471 : */
10472 : tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10473 : struct task_group, css);
10474 : tg = autogroup_task_group(tsk, tg);
10475 :
10476 : return tg;
10477 : }
10478 :
10479 : static void sched_change_group(struct task_struct *tsk, struct task_group *group)
10480 : {
10481 : tsk->sched_task_group = group;
10482 :
10483 : #ifdef CONFIG_FAIR_GROUP_SCHED
10484 : if (tsk->sched_class->task_change_group)
10485 : tsk->sched_class->task_change_group(tsk);
10486 : else
10487 : #endif
10488 : set_task_rq(tsk, task_cpu(tsk));
10489 : }
10490 :
10491 : /*
10492 : * Change task's runqueue when it moves between groups.
10493 : *
10494 : * The caller of this function should have put the task in its new group by
10495 : * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10496 : * its new group.
10497 : */
10498 : void sched_move_task(struct task_struct *tsk)
10499 : {
10500 : int queued, running, queue_flags =
10501 : DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10502 : struct task_group *group;
10503 : struct rq_flags rf;
10504 : struct rq *rq;
10505 :
10506 : rq = task_rq_lock(tsk, &rf);
10507 : /*
10508 : * Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
10509 : * group changes.
10510 : */
10511 : group = sched_get_task_group(tsk);
10512 : if (group == tsk->sched_task_group)
10513 : goto unlock;
10514 :
10515 : update_rq_clock(rq);
10516 :
10517 : running = task_current(rq, tsk);
10518 : queued = task_on_rq_queued(tsk);
10519 :
10520 : if (queued)
10521 : dequeue_task(rq, tsk, queue_flags);
10522 : if (running)
10523 : put_prev_task(rq, tsk);
10524 :
10525 : sched_change_group(tsk, group);
10526 :
10527 : if (queued)
10528 : enqueue_task(rq, tsk, queue_flags);
10529 : if (running) {
10530 : set_next_task(rq, tsk);
10531 : /*
10532 : * After changing group, the running task may have joined a
10533 : * throttled one but it's still the running task. Trigger a
10534 : * resched to make sure that task can still run.
10535 : */
10536 : resched_curr(rq);
10537 : }
10538 :
10539 : unlock:
10540 : task_rq_unlock(rq, tsk, &rf);
10541 : }
10542 :
10543 : static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10544 : {
10545 : return css ? container_of(css, struct task_group, css) : NULL;
10546 : }
10547 :
10548 : static struct cgroup_subsys_state *
10549 : cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10550 : {
10551 : struct task_group *parent = css_tg(parent_css);
10552 : struct task_group *tg;
10553 :
10554 : if (!parent) {
10555 : /* This is early initialization for the top cgroup */
10556 : return &root_task_group.css;
10557 : }
10558 :
10559 : tg = sched_create_group(parent);
10560 : if (IS_ERR(tg))
10561 : return ERR_PTR(-ENOMEM);
10562 :
10563 : return &tg->css;
10564 : }
10565 :
10566 : /* Expose task group only after completing cgroup initialization */
10567 : static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10568 : {
10569 : struct task_group *tg = css_tg(css);
10570 : struct task_group *parent = css_tg(css->parent);
10571 :
10572 : if (parent)
10573 : sched_online_group(tg, parent);
10574 :
10575 : #ifdef CONFIG_UCLAMP_TASK_GROUP
10576 : /* Propagate the effective uclamp value for the new group */
10577 : mutex_lock(&uclamp_mutex);
10578 : rcu_read_lock();
10579 : cpu_util_update_eff(css);
10580 : rcu_read_unlock();
10581 : mutex_unlock(&uclamp_mutex);
10582 : #endif
10583 :
10584 : return 0;
10585 : }
10586 :
10587 : static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10588 : {
10589 : struct task_group *tg = css_tg(css);
10590 :
10591 : sched_release_group(tg);
10592 : }
10593 :
10594 : static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10595 : {
10596 : struct task_group *tg = css_tg(css);
10597 :
10598 : /*
10599 : * Relies on the RCU grace period between css_released() and this.
10600 : */
10601 : sched_unregister_group(tg);
10602 : }
10603 :
10604 : #ifdef CONFIG_RT_GROUP_SCHED
10605 : static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10606 : {
10607 : struct task_struct *task;
10608 : struct cgroup_subsys_state *css;
10609 :
10610 : cgroup_taskset_for_each(task, css, tset) {
10611 : if (!sched_rt_can_attach(css_tg(css), task))
10612 : return -EINVAL;
10613 : }
10614 : return 0;
10615 : }
10616 : #endif
10617 :
10618 : static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10619 : {
10620 : struct task_struct *task;
10621 : struct cgroup_subsys_state *css;
10622 :
10623 : cgroup_taskset_for_each(task, css, tset)
10624 : sched_move_task(task);
10625 : }
10626 :
10627 : #ifdef CONFIG_UCLAMP_TASK_GROUP
10628 : static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10629 : {
10630 : struct cgroup_subsys_state *top_css = css;
10631 : struct uclamp_se *uc_parent = NULL;
10632 : struct uclamp_se *uc_se = NULL;
10633 : unsigned int eff[UCLAMP_CNT];
10634 : enum uclamp_id clamp_id;
10635 : unsigned int clamps;
10636 :
10637 : lockdep_assert_held(&uclamp_mutex);
10638 : SCHED_WARN_ON(!rcu_read_lock_held());
10639 :
10640 : css_for_each_descendant_pre(css, top_css) {
10641 : uc_parent = css_tg(css)->parent
10642 : ? css_tg(css)->parent->uclamp : NULL;
10643 :
10644 : for_each_clamp_id(clamp_id) {
10645 : /* Assume effective clamps matches requested clamps */
10646 : eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10647 : /* Cap effective clamps with parent's effective clamps */
10648 : if (uc_parent &&
10649 : eff[clamp_id] > uc_parent[clamp_id].value) {
10650 : eff[clamp_id] = uc_parent[clamp_id].value;
10651 : }
10652 : }
10653 : /* Ensure protection is always capped by limit */
10654 : eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10655 :
10656 : /* Propagate most restrictive effective clamps */
10657 : clamps = 0x0;
10658 : uc_se = css_tg(css)->uclamp;
10659 : for_each_clamp_id(clamp_id) {
10660 : if (eff[clamp_id] == uc_se[clamp_id].value)
10661 : continue;
10662 : uc_se[clamp_id].value = eff[clamp_id];
10663 : uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10664 : clamps |= (0x1 << clamp_id);
10665 : }
10666 : if (!clamps) {
10667 : css = css_rightmost_descendant(css);
10668 : continue;
10669 : }
10670 :
10671 : /* Immediately update descendants RUNNABLE tasks */
10672 : uclamp_update_active_tasks(css);
10673 : }
10674 : }
10675 :
10676 : /*
10677 : * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10678 : * C expression. Since there is no way to convert a macro argument (N) into a
10679 : * character constant, use two levels of macros.
10680 : */
10681 : #define _POW10(exp) ((unsigned int)1e##exp)
10682 : #define POW10(exp) _POW10(exp)
10683 :
10684 : struct uclamp_request {
10685 : #define UCLAMP_PERCENT_SHIFT 2
10686 : #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10687 : s64 percent;
10688 : u64 util;
10689 : int ret;
10690 : };
10691 :
10692 : static inline struct uclamp_request
10693 : capacity_from_percent(char *buf)
10694 : {
10695 : struct uclamp_request req = {
10696 : .percent = UCLAMP_PERCENT_SCALE,
10697 : .util = SCHED_CAPACITY_SCALE,
10698 : .ret = 0,
10699 : };
10700 :
10701 : buf = strim(buf);
10702 : if (strcmp(buf, "max")) {
10703 : req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10704 : &req.percent);
10705 : if (req.ret)
10706 : return req;
10707 : if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10708 : req.ret = -ERANGE;
10709 : return req;
10710 : }
10711 :
10712 : req.util = req.percent << SCHED_CAPACITY_SHIFT;
10713 : req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10714 : }
10715 :
10716 : return req;
10717 : }
10718 :
10719 : static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10720 : size_t nbytes, loff_t off,
10721 : enum uclamp_id clamp_id)
10722 : {
10723 : struct uclamp_request req;
10724 : struct task_group *tg;
10725 :
10726 : req = capacity_from_percent(buf);
10727 : if (req.ret)
10728 : return req.ret;
10729 :
10730 : static_branch_enable(&sched_uclamp_used);
10731 :
10732 : mutex_lock(&uclamp_mutex);
10733 : rcu_read_lock();
10734 :
10735 : tg = css_tg(of_css(of));
10736 : if (tg->uclamp_req[clamp_id].value != req.util)
10737 : uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10738 :
10739 : /*
10740 : * Because of not recoverable conversion rounding we keep track of the
10741 : * exact requested value
10742 : */
10743 : tg->uclamp_pct[clamp_id] = req.percent;
10744 :
10745 : /* Update effective clamps to track the most restrictive value */
10746 : cpu_util_update_eff(of_css(of));
10747 :
10748 : rcu_read_unlock();
10749 : mutex_unlock(&uclamp_mutex);
10750 :
10751 : return nbytes;
10752 : }
10753 :
10754 : static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10755 : char *buf, size_t nbytes,
10756 : loff_t off)
10757 : {
10758 : return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10759 : }
10760 :
10761 : static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10762 : char *buf, size_t nbytes,
10763 : loff_t off)
10764 : {
10765 : return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10766 : }
10767 :
10768 : static inline void cpu_uclamp_print(struct seq_file *sf,
10769 : enum uclamp_id clamp_id)
10770 : {
10771 : struct task_group *tg;
10772 : u64 util_clamp;
10773 : u64 percent;
10774 : u32 rem;
10775 :
10776 : rcu_read_lock();
10777 : tg = css_tg(seq_css(sf));
10778 : util_clamp = tg->uclamp_req[clamp_id].value;
10779 : rcu_read_unlock();
10780 :
10781 : if (util_clamp == SCHED_CAPACITY_SCALE) {
10782 : seq_puts(sf, "max\n");
10783 : return;
10784 : }
10785 :
10786 : percent = tg->uclamp_pct[clamp_id];
10787 : percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10788 : seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10789 : }
10790 :
10791 : static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10792 : {
10793 : cpu_uclamp_print(sf, UCLAMP_MIN);
10794 : return 0;
10795 : }
10796 :
10797 : static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10798 : {
10799 : cpu_uclamp_print(sf, UCLAMP_MAX);
10800 : return 0;
10801 : }
10802 : #endif /* CONFIG_UCLAMP_TASK_GROUP */
10803 :
10804 : #ifdef CONFIG_FAIR_GROUP_SCHED
10805 : static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10806 : struct cftype *cftype, u64 shareval)
10807 : {
10808 : if (shareval > scale_load_down(ULONG_MAX))
10809 : shareval = MAX_SHARES;
10810 : return sched_group_set_shares(css_tg(css), scale_load(shareval));
10811 : }
10812 :
10813 : static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10814 : struct cftype *cft)
10815 : {
10816 : struct task_group *tg = css_tg(css);
10817 :
10818 : return (u64) scale_load_down(tg->shares);
10819 : }
10820 :
10821 : #ifdef CONFIG_CFS_BANDWIDTH
10822 : static DEFINE_MUTEX(cfs_constraints_mutex);
10823 :
10824 : const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10825 : static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10826 : /* More than 203 days if BW_SHIFT equals 20. */
10827 : static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10828 :
10829 : static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10830 :
10831 : static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10832 : u64 burst)
10833 : {
10834 : int i, ret = 0, runtime_enabled, runtime_was_enabled;
10835 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10836 :
10837 : if (tg == &root_task_group)
10838 : return -EINVAL;
10839 :
10840 : /*
10841 : * Ensure we have at some amount of bandwidth every period. This is
10842 : * to prevent reaching a state of large arrears when throttled via
10843 : * entity_tick() resulting in prolonged exit starvation.
10844 : */
10845 : if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10846 : return -EINVAL;
10847 :
10848 : /*
10849 : * Likewise, bound things on the other side by preventing insane quota
10850 : * periods. This also allows us to normalize in computing quota
10851 : * feasibility.
10852 : */
10853 : if (period > max_cfs_quota_period)
10854 : return -EINVAL;
10855 :
10856 : /*
10857 : * Bound quota to defend quota against overflow during bandwidth shift.
10858 : */
10859 : if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10860 : return -EINVAL;
10861 :
10862 : if (quota != RUNTIME_INF && (burst > quota ||
10863 : burst + quota > max_cfs_runtime))
10864 : return -EINVAL;
10865 :
10866 : /*
10867 : * Prevent race between setting of cfs_rq->runtime_enabled and
10868 : * unthrottle_offline_cfs_rqs().
10869 : */
10870 : cpus_read_lock();
10871 : mutex_lock(&cfs_constraints_mutex);
10872 : ret = __cfs_schedulable(tg, period, quota);
10873 : if (ret)
10874 : goto out_unlock;
10875 :
10876 : runtime_enabled = quota != RUNTIME_INF;
10877 : runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10878 : /*
10879 : * If we need to toggle cfs_bandwidth_used, off->on must occur
10880 : * before making related changes, and on->off must occur afterwards
10881 : */
10882 : if (runtime_enabled && !runtime_was_enabled)
10883 : cfs_bandwidth_usage_inc();
10884 : raw_spin_lock_irq(&cfs_b->lock);
10885 : cfs_b->period = ns_to_ktime(period);
10886 : cfs_b->quota = quota;
10887 : cfs_b->burst = burst;
10888 :
10889 : __refill_cfs_bandwidth_runtime(cfs_b);
10890 :
10891 : /* Restart the period timer (if active) to handle new period expiry: */
10892 : if (runtime_enabled)
10893 : start_cfs_bandwidth(cfs_b);
10894 :
10895 : raw_spin_unlock_irq(&cfs_b->lock);
10896 :
10897 : for_each_online_cpu(i) {
10898 : struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10899 : struct rq *rq = cfs_rq->rq;
10900 : struct rq_flags rf;
10901 :
10902 : rq_lock_irq(rq, &rf);
10903 : cfs_rq->runtime_enabled = runtime_enabled;
10904 : cfs_rq->runtime_remaining = 0;
10905 :
10906 : if (cfs_rq->throttled)
10907 : unthrottle_cfs_rq(cfs_rq);
10908 : rq_unlock_irq(rq, &rf);
10909 : }
10910 : if (runtime_was_enabled && !runtime_enabled)
10911 : cfs_bandwidth_usage_dec();
10912 : out_unlock:
10913 : mutex_unlock(&cfs_constraints_mutex);
10914 : cpus_read_unlock();
10915 :
10916 : return ret;
10917 : }
10918 :
10919 : static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10920 : {
10921 : u64 quota, period, burst;
10922 :
10923 : period = ktime_to_ns(tg->cfs_bandwidth.period);
10924 : burst = tg->cfs_bandwidth.burst;
10925 : if (cfs_quota_us < 0)
10926 : quota = RUNTIME_INF;
10927 : else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10928 : quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10929 : else
10930 : return -EINVAL;
10931 :
10932 : return tg_set_cfs_bandwidth(tg, period, quota, burst);
10933 : }
10934 :
10935 : static long tg_get_cfs_quota(struct task_group *tg)
10936 : {
10937 : u64 quota_us;
10938 :
10939 : if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10940 : return -1;
10941 :
10942 : quota_us = tg->cfs_bandwidth.quota;
10943 : do_div(quota_us, NSEC_PER_USEC);
10944 :
10945 : return quota_us;
10946 : }
10947 :
10948 : static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10949 : {
10950 : u64 quota, period, burst;
10951 :
10952 : if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10953 : return -EINVAL;
10954 :
10955 : period = (u64)cfs_period_us * NSEC_PER_USEC;
10956 : quota = tg->cfs_bandwidth.quota;
10957 : burst = tg->cfs_bandwidth.burst;
10958 :
10959 : return tg_set_cfs_bandwidth(tg, period, quota, burst);
10960 : }
10961 :
10962 : static long tg_get_cfs_period(struct task_group *tg)
10963 : {
10964 : u64 cfs_period_us;
10965 :
10966 : cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10967 : do_div(cfs_period_us, NSEC_PER_USEC);
10968 :
10969 : return cfs_period_us;
10970 : }
10971 :
10972 : static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10973 : {
10974 : u64 quota, period, burst;
10975 :
10976 : if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10977 : return -EINVAL;
10978 :
10979 : burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10980 : period = ktime_to_ns(tg->cfs_bandwidth.period);
10981 : quota = tg->cfs_bandwidth.quota;
10982 :
10983 : return tg_set_cfs_bandwidth(tg, period, quota, burst);
10984 : }
10985 :
10986 : static long tg_get_cfs_burst(struct task_group *tg)
10987 : {
10988 : u64 burst_us;
10989 :
10990 : burst_us = tg->cfs_bandwidth.burst;
10991 : do_div(burst_us, NSEC_PER_USEC);
10992 :
10993 : return burst_us;
10994 : }
10995 :
10996 : static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10997 : struct cftype *cft)
10998 : {
10999 : return tg_get_cfs_quota(css_tg(css));
11000 : }
11001 :
11002 : static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
11003 : struct cftype *cftype, s64 cfs_quota_us)
11004 : {
11005 : return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
11006 : }
11007 :
11008 : static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
11009 : struct cftype *cft)
11010 : {
11011 : return tg_get_cfs_period(css_tg(css));
11012 : }
11013 :
11014 : static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
11015 : struct cftype *cftype, u64 cfs_period_us)
11016 : {
11017 : return tg_set_cfs_period(css_tg(css), cfs_period_us);
11018 : }
11019 :
11020 : static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
11021 : struct cftype *cft)
11022 : {
11023 : return tg_get_cfs_burst(css_tg(css));
11024 : }
11025 :
11026 : static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
11027 : struct cftype *cftype, u64 cfs_burst_us)
11028 : {
11029 : return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
11030 : }
11031 :
11032 : struct cfs_schedulable_data {
11033 : struct task_group *tg;
11034 : u64 period, quota;
11035 : };
11036 :
11037 : /*
11038 : * normalize group quota/period to be quota/max_period
11039 : * note: units are usecs
11040 : */
11041 : static u64 normalize_cfs_quota(struct task_group *tg,
11042 : struct cfs_schedulable_data *d)
11043 : {
11044 : u64 quota, period;
11045 :
11046 : if (tg == d->tg) {
11047 : period = d->period;
11048 : quota = d->quota;
11049 : } else {
11050 : period = tg_get_cfs_period(tg);
11051 : quota = tg_get_cfs_quota(tg);
11052 : }
11053 :
11054 : /* note: these should typically be equivalent */
11055 : if (quota == RUNTIME_INF || quota == -1)
11056 : return RUNTIME_INF;
11057 :
11058 : return to_ratio(period, quota);
11059 : }
11060 :
11061 : static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
11062 : {
11063 : struct cfs_schedulable_data *d = data;
11064 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11065 : s64 quota = 0, parent_quota = -1;
11066 :
11067 : if (!tg->parent) {
11068 : quota = RUNTIME_INF;
11069 : } else {
11070 : struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
11071 :
11072 : quota = normalize_cfs_quota(tg, d);
11073 : parent_quota = parent_b->hierarchical_quota;
11074 :
11075 : /*
11076 : * Ensure max(child_quota) <= parent_quota. On cgroup2,
11077 : * always take the min. On cgroup1, only inherit when no
11078 : * limit is set:
11079 : */
11080 : if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
11081 : quota = min(quota, parent_quota);
11082 : } else {
11083 : if (quota == RUNTIME_INF)
11084 : quota = parent_quota;
11085 : else if (parent_quota != RUNTIME_INF && quota > parent_quota)
11086 : return -EINVAL;
11087 : }
11088 : }
11089 : cfs_b->hierarchical_quota = quota;
11090 :
11091 : return 0;
11092 : }
11093 :
11094 : static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
11095 : {
11096 : int ret;
11097 : struct cfs_schedulable_data data = {
11098 : .tg = tg,
11099 : .period = period,
11100 : .quota = quota,
11101 : };
11102 :
11103 : if (quota != RUNTIME_INF) {
11104 : do_div(data.period, NSEC_PER_USEC);
11105 : do_div(data.quota, NSEC_PER_USEC);
11106 : }
11107 :
11108 : rcu_read_lock();
11109 : ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
11110 : rcu_read_unlock();
11111 :
11112 : return ret;
11113 : }
11114 :
11115 : static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
11116 : {
11117 : struct task_group *tg = css_tg(seq_css(sf));
11118 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11119 :
11120 : seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
11121 : seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
11122 : seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
11123 :
11124 : if (schedstat_enabled() && tg != &root_task_group) {
11125 : struct sched_statistics *stats;
11126 : u64 ws = 0;
11127 : int i;
11128 :
11129 : for_each_possible_cpu(i) {
11130 : stats = __schedstats_from_se(tg->se[i]);
11131 : ws += schedstat_val(stats->wait_sum);
11132 : }
11133 :
11134 : seq_printf(sf, "wait_sum %llu\n", ws);
11135 : }
11136 :
11137 : seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
11138 : seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
11139 :
11140 : return 0;
11141 : }
11142 : #endif /* CONFIG_CFS_BANDWIDTH */
11143 : #endif /* CONFIG_FAIR_GROUP_SCHED */
11144 :
11145 : #ifdef CONFIG_RT_GROUP_SCHED
11146 : static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
11147 : struct cftype *cft, s64 val)
11148 : {
11149 : return sched_group_set_rt_runtime(css_tg(css), val);
11150 : }
11151 :
11152 : static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
11153 : struct cftype *cft)
11154 : {
11155 : return sched_group_rt_runtime(css_tg(css));
11156 : }
11157 :
11158 : static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
11159 : struct cftype *cftype, u64 rt_period_us)
11160 : {
11161 : return sched_group_set_rt_period(css_tg(css), rt_period_us);
11162 : }
11163 :
11164 : static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11165 : struct cftype *cft)
11166 : {
11167 : return sched_group_rt_period(css_tg(css));
11168 : }
11169 : #endif /* CONFIG_RT_GROUP_SCHED */
11170 :
11171 : #ifdef CONFIG_FAIR_GROUP_SCHED
11172 : static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11173 : struct cftype *cft)
11174 : {
11175 : return css_tg(css)->idle;
11176 : }
11177 :
11178 : static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11179 : struct cftype *cft, s64 idle)
11180 : {
11181 : return sched_group_set_idle(css_tg(css), idle);
11182 : }
11183 : #endif
11184 :
11185 : static struct cftype cpu_legacy_files[] = {
11186 : #ifdef CONFIG_FAIR_GROUP_SCHED
11187 : {
11188 : .name = "shares",
11189 : .read_u64 = cpu_shares_read_u64,
11190 : .write_u64 = cpu_shares_write_u64,
11191 : },
11192 : {
11193 : .name = "idle",
11194 : .read_s64 = cpu_idle_read_s64,
11195 : .write_s64 = cpu_idle_write_s64,
11196 : },
11197 : #endif
11198 : #ifdef CONFIG_CFS_BANDWIDTH
11199 : {
11200 : .name = "cfs_quota_us",
11201 : .read_s64 = cpu_cfs_quota_read_s64,
11202 : .write_s64 = cpu_cfs_quota_write_s64,
11203 : },
11204 : {
11205 : .name = "cfs_period_us",
11206 : .read_u64 = cpu_cfs_period_read_u64,
11207 : .write_u64 = cpu_cfs_period_write_u64,
11208 : },
11209 : {
11210 : .name = "cfs_burst_us",
11211 : .read_u64 = cpu_cfs_burst_read_u64,
11212 : .write_u64 = cpu_cfs_burst_write_u64,
11213 : },
11214 : {
11215 : .name = "stat",
11216 : .seq_show = cpu_cfs_stat_show,
11217 : },
11218 : #endif
11219 : #ifdef CONFIG_RT_GROUP_SCHED
11220 : {
11221 : .name = "rt_runtime_us",
11222 : .read_s64 = cpu_rt_runtime_read,
11223 : .write_s64 = cpu_rt_runtime_write,
11224 : },
11225 : {
11226 : .name = "rt_period_us",
11227 : .read_u64 = cpu_rt_period_read_uint,
11228 : .write_u64 = cpu_rt_period_write_uint,
11229 : },
11230 : #endif
11231 : #ifdef CONFIG_UCLAMP_TASK_GROUP
11232 : {
11233 : .name = "uclamp.min",
11234 : .flags = CFTYPE_NOT_ON_ROOT,
11235 : .seq_show = cpu_uclamp_min_show,
11236 : .write = cpu_uclamp_min_write,
11237 : },
11238 : {
11239 : .name = "uclamp.max",
11240 : .flags = CFTYPE_NOT_ON_ROOT,
11241 : .seq_show = cpu_uclamp_max_show,
11242 : .write = cpu_uclamp_max_write,
11243 : },
11244 : #endif
11245 : { } /* Terminate */
11246 : };
11247 :
11248 : static int cpu_extra_stat_show(struct seq_file *sf,
11249 : struct cgroup_subsys_state *css)
11250 : {
11251 : #ifdef CONFIG_CFS_BANDWIDTH
11252 : {
11253 : struct task_group *tg = css_tg(css);
11254 : struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11255 : u64 throttled_usec, burst_usec;
11256 :
11257 : throttled_usec = cfs_b->throttled_time;
11258 : do_div(throttled_usec, NSEC_PER_USEC);
11259 : burst_usec = cfs_b->burst_time;
11260 : do_div(burst_usec, NSEC_PER_USEC);
11261 :
11262 : seq_printf(sf, "nr_periods %d\n"
11263 : "nr_throttled %d\n"
11264 : "throttled_usec %llu\n"
11265 : "nr_bursts %d\n"
11266 : "burst_usec %llu\n",
11267 : cfs_b->nr_periods, cfs_b->nr_throttled,
11268 : throttled_usec, cfs_b->nr_burst, burst_usec);
11269 : }
11270 : #endif
11271 : return 0;
11272 : }
11273 :
11274 : #ifdef CONFIG_FAIR_GROUP_SCHED
11275 : static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11276 : struct cftype *cft)
11277 : {
11278 : struct task_group *tg = css_tg(css);
11279 : u64 weight = scale_load_down(tg->shares);
11280 :
11281 : return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11282 : }
11283 :
11284 : static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11285 : struct cftype *cft, u64 weight)
11286 : {
11287 : /*
11288 : * cgroup weight knobs should use the common MIN, DFL and MAX
11289 : * values which are 1, 100 and 10000 respectively. While it loses
11290 : * a bit of range on both ends, it maps pretty well onto the shares
11291 : * value used by scheduler and the round-trip conversions preserve
11292 : * the original value over the entire range.
11293 : */
11294 : if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11295 : return -ERANGE;
11296 :
11297 : weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11298 :
11299 : return sched_group_set_shares(css_tg(css), scale_load(weight));
11300 : }
11301 :
11302 : static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11303 : struct cftype *cft)
11304 : {
11305 : unsigned long weight = scale_load_down(css_tg(css)->shares);
11306 : int last_delta = INT_MAX;
11307 : int prio, delta;
11308 :
11309 : /* find the closest nice value to the current weight */
11310 : for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11311 : delta = abs(sched_prio_to_weight[prio] - weight);
11312 : if (delta >= last_delta)
11313 : break;
11314 : last_delta = delta;
11315 : }
11316 :
11317 : return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11318 : }
11319 :
11320 : static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11321 : struct cftype *cft, s64 nice)
11322 : {
11323 : unsigned long weight;
11324 : int idx;
11325 :
11326 : if (nice < MIN_NICE || nice > MAX_NICE)
11327 : return -ERANGE;
11328 :
11329 : idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11330 : idx = array_index_nospec(idx, 40);
11331 : weight = sched_prio_to_weight[idx];
11332 :
11333 : return sched_group_set_shares(css_tg(css), scale_load(weight));
11334 : }
11335 : #endif
11336 :
11337 : static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11338 : long period, long quota)
11339 : {
11340 : if (quota < 0)
11341 : seq_puts(sf, "max");
11342 : else
11343 : seq_printf(sf, "%ld", quota);
11344 :
11345 : seq_printf(sf, " %ld\n", period);
11346 : }
11347 :
11348 : /* caller should put the current value in *@periodp before calling */
11349 : static int __maybe_unused cpu_period_quota_parse(char *buf,
11350 : u64 *periodp, u64 *quotap)
11351 : {
11352 : char tok[21]; /* U64_MAX */
11353 :
11354 : if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11355 : return -EINVAL;
11356 :
11357 : *periodp *= NSEC_PER_USEC;
11358 :
11359 : if (sscanf(tok, "%llu", quotap))
11360 : *quotap *= NSEC_PER_USEC;
11361 : else if (!strcmp(tok, "max"))
11362 : *quotap = RUNTIME_INF;
11363 : else
11364 : return -EINVAL;
11365 :
11366 : return 0;
11367 : }
11368 :
11369 : #ifdef CONFIG_CFS_BANDWIDTH
11370 : static int cpu_max_show(struct seq_file *sf, void *v)
11371 : {
11372 : struct task_group *tg = css_tg(seq_css(sf));
11373 :
11374 : cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11375 : return 0;
11376 : }
11377 :
11378 : static ssize_t cpu_max_write(struct kernfs_open_file *of,
11379 : char *buf, size_t nbytes, loff_t off)
11380 : {
11381 : struct task_group *tg = css_tg(of_css(of));
11382 : u64 period = tg_get_cfs_period(tg);
11383 : u64 burst = tg_get_cfs_burst(tg);
11384 : u64 quota;
11385 : int ret;
11386 :
11387 : ret = cpu_period_quota_parse(buf, &period, "a);
11388 : if (!ret)
11389 : ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11390 : return ret ?: nbytes;
11391 : }
11392 : #endif
11393 :
11394 : static struct cftype cpu_files[] = {
11395 : #ifdef CONFIG_FAIR_GROUP_SCHED
11396 : {
11397 : .name = "weight",
11398 : .flags = CFTYPE_NOT_ON_ROOT,
11399 : .read_u64 = cpu_weight_read_u64,
11400 : .write_u64 = cpu_weight_write_u64,
11401 : },
11402 : {
11403 : .name = "weight.nice",
11404 : .flags = CFTYPE_NOT_ON_ROOT,
11405 : .read_s64 = cpu_weight_nice_read_s64,
11406 : .write_s64 = cpu_weight_nice_write_s64,
11407 : },
11408 : {
11409 : .name = "idle",
11410 : .flags = CFTYPE_NOT_ON_ROOT,
11411 : .read_s64 = cpu_idle_read_s64,
11412 : .write_s64 = cpu_idle_write_s64,
11413 : },
11414 : #endif
11415 : #ifdef CONFIG_CFS_BANDWIDTH
11416 : {
11417 : .name = "max",
11418 : .flags = CFTYPE_NOT_ON_ROOT,
11419 : .seq_show = cpu_max_show,
11420 : .write = cpu_max_write,
11421 : },
11422 : {
11423 : .name = "max.burst",
11424 : .flags = CFTYPE_NOT_ON_ROOT,
11425 : .read_u64 = cpu_cfs_burst_read_u64,
11426 : .write_u64 = cpu_cfs_burst_write_u64,
11427 : },
11428 : #endif
11429 : #ifdef CONFIG_UCLAMP_TASK_GROUP
11430 : {
11431 : .name = "uclamp.min",
11432 : .flags = CFTYPE_NOT_ON_ROOT,
11433 : .seq_show = cpu_uclamp_min_show,
11434 : .write = cpu_uclamp_min_write,
11435 : },
11436 : {
11437 : .name = "uclamp.max",
11438 : .flags = CFTYPE_NOT_ON_ROOT,
11439 : .seq_show = cpu_uclamp_max_show,
11440 : .write = cpu_uclamp_max_write,
11441 : },
11442 : #endif
11443 : { } /* terminate */
11444 : };
11445 :
11446 : struct cgroup_subsys cpu_cgrp_subsys = {
11447 : .css_alloc = cpu_cgroup_css_alloc,
11448 : .css_online = cpu_cgroup_css_online,
11449 : .css_released = cpu_cgroup_css_released,
11450 : .css_free = cpu_cgroup_css_free,
11451 : .css_extra_stat_show = cpu_extra_stat_show,
11452 : #ifdef CONFIG_RT_GROUP_SCHED
11453 : .can_attach = cpu_cgroup_can_attach,
11454 : #endif
11455 : .attach = cpu_cgroup_attach,
11456 : .legacy_cftypes = cpu_legacy_files,
11457 : .dfl_cftypes = cpu_files,
11458 : .early_init = true,
11459 : .threaded = true,
11460 : };
11461 :
11462 : #endif /* CONFIG_CGROUP_SCHED */
11463 :
11464 0 : void dump_cpu_task(int cpu)
11465 : {
11466 0 : if (cpu == smp_processor_id() && in_hardirq()) {
11467 : struct pt_regs *regs;
11468 :
11469 0 : regs = get_irq_regs();
11470 0 : if (regs) {
11471 0 : show_regs(regs);
11472 0 : return;
11473 : }
11474 : }
11475 :
11476 0 : if (trigger_single_cpu_backtrace(cpu))
11477 : return;
11478 :
11479 0 : pr_info("Task dump for CPU %d:\n", cpu);
11480 0 : sched_show_task(cpu_curr(cpu));
11481 : }
11482 :
11483 : /*
11484 : * Nice levels are multiplicative, with a gentle 10% change for every
11485 : * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11486 : * nice 1, it will get ~10% less CPU time than another CPU-bound task
11487 : * that remained on nice 0.
11488 : *
11489 : * The "10% effect" is relative and cumulative: from _any_ nice level,
11490 : * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11491 : * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11492 : * If a task goes up by ~10% and another task goes down by ~10% then
11493 : * the relative distance between them is ~25%.)
11494 : */
11495 : const int sched_prio_to_weight[40] = {
11496 : /* -20 */ 88761, 71755, 56483, 46273, 36291,
11497 : /* -15 */ 29154, 23254, 18705, 14949, 11916,
11498 : /* -10 */ 9548, 7620, 6100, 4904, 3906,
11499 : /* -5 */ 3121, 2501, 1991, 1586, 1277,
11500 : /* 0 */ 1024, 820, 655, 526, 423,
11501 : /* 5 */ 335, 272, 215, 172, 137,
11502 : /* 10 */ 110, 87, 70, 56, 45,
11503 : /* 15 */ 36, 29, 23, 18, 15,
11504 : };
11505 :
11506 : /*
11507 : * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11508 : *
11509 : * In cases where the weight does not change often, we can use the
11510 : * precalculated inverse to speed up arithmetics by turning divisions
11511 : * into multiplications:
11512 : */
11513 : const u32 sched_prio_to_wmult[40] = {
11514 : /* -20 */ 48388, 59856, 76040, 92818, 118348,
11515 : /* -15 */ 147320, 184698, 229616, 287308, 360437,
11516 : /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11517 : /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11518 : /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11519 : /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11520 : /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11521 : /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11522 : };
11523 :
11524 0 : void call_trace_sched_update_nr_running(struct rq *rq, int count)
11525 : {
11526 0 : trace_sched_update_nr_running_tp(rq, count);
11527 0 : }
11528 :
11529 : #ifdef CONFIG_SCHED_MM_CID
11530 :
11531 : /*
11532 : * @cid_lock: Guarantee forward-progress of cid allocation.
11533 : *
11534 : * Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
11535 : * is only used when contention is detected by the lock-free allocation so
11536 : * forward progress can be guaranteed.
11537 : */
11538 : DEFINE_RAW_SPINLOCK(cid_lock);
11539 :
11540 : /*
11541 : * @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
11542 : *
11543 : * When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
11544 : * detected, it is set to 1 to ensure that all newly coming allocations are
11545 : * serialized by @cid_lock until the allocation which detected contention
11546 : * completes and sets @use_cid_lock back to 0. This guarantees forward progress
11547 : * of a cid allocation.
11548 : */
11549 : int use_cid_lock;
11550 :
11551 : /*
11552 : * mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
11553 : * concurrently with respect to the execution of the source runqueue context
11554 : * switch.
11555 : *
11556 : * There is one basic properties we want to guarantee here:
11557 : *
11558 : * (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
11559 : * used by a task. That would lead to concurrent allocation of the cid and
11560 : * userspace corruption.
11561 : *
11562 : * Provide this guarantee by introducing a Dekker memory ordering to guarantee
11563 : * that a pair of loads observe at least one of a pair of stores, which can be
11564 : * shown as:
11565 : *
11566 : * X = Y = 0
11567 : *
11568 : * w[X]=1 w[Y]=1
11569 : * MB MB
11570 : * r[Y]=y r[X]=x
11571 : *
11572 : * Which guarantees that x==0 && y==0 is impossible. But rather than using
11573 : * values 0 and 1, this algorithm cares about specific state transitions of the
11574 : * runqueue current task (as updated by the scheduler context switch), and the
11575 : * per-mm/cpu cid value.
11576 : *
11577 : * Let's introduce task (Y) which has task->mm == mm and task (N) which has
11578 : * task->mm != mm for the rest of the discussion. There are two scheduler state
11579 : * transitions on context switch we care about:
11580 : *
11581 : * (TSA) Store to rq->curr with transition from (N) to (Y)
11582 : *
11583 : * (TSB) Store to rq->curr with transition from (Y) to (N)
11584 : *
11585 : * On the remote-clear side, there is one transition we care about:
11586 : *
11587 : * (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
11588 : *
11589 : * There is also a transition to UNSET state which can be performed from all
11590 : * sides (scheduler, remote-clear). It is always performed with a cmpxchg which
11591 : * guarantees that only a single thread will succeed:
11592 : *
11593 : * (TMB) cmpxchg to *pcpu_cid to mark UNSET
11594 : *
11595 : * Just to be clear, what we do _not_ want to happen is a transition to UNSET
11596 : * when a thread is actively using the cid (property (1)).
11597 : *
11598 : * Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
11599 : *
11600 : * Scenario A) (TSA)+(TMA) (from next task perspective)
11601 : *
11602 : * CPU0 CPU1
11603 : *
11604 : * Context switch CS-1 Remote-clear
11605 : * - store to rq->curr: (N)->(Y) (TSA) - cmpxchg to *pcpu_id to LAZY (TMA)
11606 : * (implied barrier after cmpxchg)
11607 : * - switch_mm_cid()
11608 : * - memory barrier (see switch_mm_cid()
11609 : * comment explaining how this barrier
11610 : * is combined with other scheduler
11611 : * barriers)
11612 : * - mm_cid_get (next)
11613 : * - READ_ONCE(*pcpu_cid) - rcu_dereference(src_rq->curr)
11614 : *
11615 : * This Dekker ensures that either task (Y) is observed by the
11616 : * rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
11617 : * observed.
11618 : *
11619 : * If task (Y) store is observed by rcu_dereference(), it means that there is
11620 : * still an active task on the cpu. Remote-clear will therefore not transition
11621 : * to UNSET, which fulfills property (1).
11622 : *
11623 : * If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
11624 : * it will move its state to UNSET, which clears the percpu cid perhaps
11625 : * uselessly (which is not an issue for correctness). Because task (Y) is not
11626 : * observed, CPU1 can move ahead to set the state to UNSET. Because moving
11627 : * state to UNSET is done with a cmpxchg expecting that the old state has the
11628 : * LAZY flag set, only one thread will successfully UNSET.
11629 : *
11630 : * If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
11631 : * will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
11632 : * CPU1 will observe task (Y) and do nothing more, which is fine.
11633 : *
11634 : * What we are effectively preventing with this Dekker is a scenario where
11635 : * neither LAZY flag nor store (Y) are observed, which would fail property (1)
11636 : * because this would UNSET a cid which is actively used.
11637 : */
11638 :
11639 : void sched_mm_cid_migrate_from(struct task_struct *t)
11640 : {
11641 : t->migrate_from_cpu = task_cpu(t);
11642 : }
11643 :
11644 : static
11645 : int __sched_mm_cid_migrate_from_fetch_cid(struct rq *src_rq,
11646 : struct task_struct *t,
11647 : struct mm_cid *src_pcpu_cid)
11648 : {
11649 : struct mm_struct *mm = t->mm;
11650 : struct task_struct *src_task;
11651 : int src_cid, last_mm_cid;
11652 :
11653 : if (!mm)
11654 : return -1;
11655 :
11656 : last_mm_cid = t->last_mm_cid;
11657 : /*
11658 : * If the migrated task has no last cid, or if the current
11659 : * task on src rq uses the cid, it means the source cid does not need
11660 : * to be moved to the destination cpu.
11661 : */
11662 : if (last_mm_cid == -1)
11663 : return -1;
11664 : src_cid = READ_ONCE(src_pcpu_cid->cid);
11665 : if (!mm_cid_is_valid(src_cid) || last_mm_cid != src_cid)
11666 : return -1;
11667 :
11668 : /*
11669 : * If we observe an active task using the mm on this rq, it means we
11670 : * are not the last task to be migrated from this cpu for this mm, so
11671 : * there is no need to move src_cid to the destination cpu.
11672 : */
11673 : rcu_read_lock();
11674 : src_task = rcu_dereference(src_rq->curr);
11675 : if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11676 : rcu_read_unlock();
11677 : t->last_mm_cid = -1;
11678 : return -1;
11679 : }
11680 : rcu_read_unlock();
11681 :
11682 : return src_cid;
11683 : }
11684 :
11685 : static
11686 : int __sched_mm_cid_migrate_from_try_steal_cid(struct rq *src_rq,
11687 : struct task_struct *t,
11688 : struct mm_cid *src_pcpu_cid,
11689 : int src_cid)
11690 : {
11691 : struct task_struct *src_task;
11692 : struct mm_struct *mm = t->mm;
11693 : int lazy_cid;
11694 :
11695 : if (src_cid == -1)
11696 : return -1;
11697 :
11698 : /*
11699 : * Attempt to clear the source cpu cid to move it to the destination
11700 : * cpu.
11701 : */
11702 : lazy_cid = mm_cid_set_lazy_put(src_cid);
11703 : if (!try_cmpxchg(&src_pcpu_cid->cid, &src_cid, lazy_cid))
11704 : return -1;
11705 :
11706 : /*
11707 : * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11708 : * rq->curr->mm matches the scheduler barrier in context_switch()
11709 : * between store to rq->curr and load of prev and next task's
11710 : * per-mm/cpu cid.
11711 : *
11712 : * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11713 : * rq->curr->mm_cid_active matches the barrier in
11714 : * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11715 : * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11716 : * load of per-mm/cpu cid.
11717 : */
11718 :
11719 : /*
11720 : * If we observe an active task using the mm on this rq after setting
11721 : * the lazy-put flag, this task will be responsible for transitioning
11722 : * from lazy-put flag set to MM_CID_UNSET.
11723 : */
11724 : rcu_read_lock();
11725 : src_task = rcu_dereference(src_rq->curr);
11726 : if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11727 : rcu_read_unlock();
11728 : /*
11729 : * We observed an active task for this mm, there is therefore
11730 : * no point in moving this cid to the destination cpu.
11731 : */
11732 : t->last_mm_cid = -1;
11733 : return -1;
11734 : }
11735 : rcu_read_unlock();
11736 :
11737 : /*
11738 : * The src_cid is unused, so it can be unset.
11739 : */
11740 : if (!try_cmpxchg(&src_pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11741 : return -1;
11742 : return src_cid;
11743 : }
11744 :
11745 : /*
11746 : * Migration to dst cpu. Called with dst_rq lock held.
11747 : * Interrupts are disabled, which keeps the window of cid ownership without the
11748 : * source rq lock held small.
11749 : */
11750 : void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t)
11751 : {
11752 : struct mm_cid *src_pcpu_cid, *dst_pcpu_cid;
11753 : struct mm_struct *mm = t->mm;
11754 : int src_cid, dst_cid, src_cpu;
11755 : struct rq *src_rq;
11756 :
11757 : lockdep_assert_rq_held(dst_rq);
11758 :
11759 : if (!mm)
11760 : return;
11761 : src_cpu = t->migrate_from_cpu;
11762 : if (src_cpu == -1) {
11763 : t->last_mm_cid = -1;
11764 : return;
11765 : }
11766 : /*
11767 : * Move the src cid if the dst cid is unset. This keeps id
11768 : * allocation closest to 0 in cases where few threads migrate around
11769 : * many cpus.
11770 : *
11771 : * If destination cid is already set, we may have to just clear
11772 : * the src cid to ensure compactness in frequent migrations
11773 : * scenarios.
11774 : *
11775 : * It is not useful to clear the src cid when the number of threads is
11776 : * greater or equal to the number of allowed cpus, because user-space
11777 : * can expect that the number of allowed cids can reach the number of
11778 : * allowed cpus.
11779 : */
11780 : dst_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(dst_rq));
11781 : dst_cid = READ_ONCE(dst_pcpu_cid->cid);
11782 : if (!mm_cid_is_unset(dst_cid) &&
11783 : atomic_read(&mm->mm_users) >= t->nr_cpus_allowed)
11784 : return;
11785 : src_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, src_cpu);
11786 : src_rq = cpu_rq(src_cpu);
11787 : src_cid = __sched_mm_cid_migrate_from_fetch_cid(src_rq, t, src_pcpu_cid);
11788 : if (src_cid == -1)
11789 : return;
11790 : src_cid = __sched_mm_cid_migrate_from_try_steal_cid(src_rq, t, src_pcpu_cid,
11791 : src_cid);
11792 : if (src_cid == -1)
11793 : return;
11794 : if (!mm_cid_is_unset(dst_cid)) {
11795 : __mm_cid_put(mm, src_cid);
11796 : return;
11797 : }
11798 : /* Move src_cid to dst cpu. */
11799 : mm_cid_snapshot_time(dst_rq, mm);
11800 : WRITE_ONCE(dst_pcpu_cid->cid, src_cid);
11801 : }
11802 :
11803 : static void sched_mm_cid_remote_clear(struct mm_struct *mm, struct mm_cid *pcpu_cid,
11804 : int cpu)
11805 : {
11806 : struct rq *rq = cpu_rq(cpu);
11807 : struct task_struct *t;
11808 : unsigned long flags;
11809 : int cid, lazy_cid;
11810 :
11811 : cid = READ_ONCE(pcpu_cid->cid);
11812 : if (!mm_cid_is_valid(cid))
11813 : return;
11814 :
11815 : /*
11816 : * Clear the cpu cid if it is set to keep cid allocation compact. If
11817 : * there happens to be other tasks left on the source cpu using this
11818 : * mm, the next task using this mm will reallocate its cid on context
11819 : * switch.
11820 : */
11821 : lazy_cid = mm_cid_set_lazy_put(cid);
11822 : if (!try_cmpxchg(&pcpu_cid->cid, &cid, lazy_cid))
11823 : return;
11824 :
11825 : /*
11826 : * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11827 : * rq->curr->mm matches the scheduler barrier in context_switch()
11828 : * between store to rq->curr and load of prev and next task's
11829 : * per-mm/cpu cid.
11830 : *
11831 : * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11832 : * rq->curr->mm_cid_active matches the barrier in
11833 : * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11834 : * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11835 : * load of per-mm/cpu cid.
11836 : */
11837 :
11838 : /*
11839 : * If we observe an active task using the mm on this rq after setting
11840 : * the lazy-put flag, that task will be responsible for transitioning
11841 : * from lazy-put flag set to MM_CID_UNSET.
11842 : */
11843 : rcu_read_lock();
11844 : t = rcu_dereference(rq->curr);
11845 : if (READ_ONCE(t->mm_cid_active) && t->mm == mm) {
11846 : rcu_read_unlock();
11847 : return;
11848 : }
11849 : rcu_read_unlock();
11850 :
11851 : /*
11852 : * The cid is unused, so it can be unset.
11853 : * Disable interrupts to keep the window of cid ownership without rq
11854 : * lock small.
11855 : */
11856 : local_irq_save(flags);
11857 : if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11858 : __mm_cid_put(mm, cid);
11859 : local_irq_restore(flags);
11860 : }
11861 :
11862 : static void sched_mm_cid_remote_clear_old(struct mm_struct *mm, int cpu)
11863 : {
11864 : struct rq *rq = cpu_rq(cpu);
11865 : struct mm_cid *pcpu_cid;
11866 : struct task_struct *curr;
11867 : u64 rq_clock;
11868 :
11869 : /*
11870 : * rq->clock load is racy on 32-bit but one spurious clear once in a
11871 : * while is irrelevant.
11872 : */
11873 : rq_clock = READ_ONCE(rq->clock);
11874 : pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11875 :
11876 : /*
11877 : * In order to take care of infrequently scheduled tasks, bump the time
11878 : * snapshot associated with this cid if an active task using the mm is
11879 : * observed on this rq.
11880 : */
11881 : rcu_read_lock();
11882 : curr = rcu_dereference(rq->curr);
11883 : if (READ_ONCE(curr->mm_cid_active) && curr->mm == mm) {
11884 : WRITE_ONCE(pcpu_cid->time, rq_clock);
11885 : rcu_read_unlock();
11886 : return;
11887 : }
11888 : rcu_read_unlock();
11889 :
11890 : if (rq_clock < pcpu_cid->time + SCHED_MM_CID_PERIOD_NS)
11891 : return;
11892 : sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11893 : }
11894 :
11895 : static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
11896 : int weight)
11897 : {
11898 : struct mm_cid *pcpu_cid;
11899 : int cid;
11900 :
11901 : pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11902 : cid = READ_ONCE(pcpu_cid->cid);
11903 : if (!mm_cid_is_valid(cid) || cid < weight)
11904 : return;
11905 : sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11906 : }
11907 :
11908 : static void task_mm_cid_work(struct callback_head *work)
11909 : {
11910 : unsigned long now = jiffies, old_scan, next_scan;
11911 : struct task_struct *t = current;
11912 : struct cpumask *cidmask;
11913 : struct mm_struct *mm;
11914 : int weight, cpu;
11915 :
11916 : SCHED_WARN_ON(t != container_of(work, struct task_struct, cid_work));
11917 :
11918 : work->next = work; /* Prevent double-add */
11919 : if (t->flags & PF_EXITING)
11920 : return;
11921 : mm = t->mm;
11922 : if (!mm)
11923 : return;
11924 : old_scan = READ_ONCE(mm->mm_cid_next_scan);
11925 : next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11926 : if (!old_scan) {
11927 : unsigned long res;
11928 :
11929 : res = cmpxchg(&mm->mm_cid_next_scan, old_scan, next_scan);
11930 : if (res != old_scan)
11931 : old_scan = res;
11932 : else
11933 : old_scan = next_scan;
11934 : }
11935 : if (time_before(now, old_scan))
11936 : return;
11937 : if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
11938 : return;
11939 : cidmask = mm_cidmask(mm);
11940 : /* Clear cids that were not recently used. */
11941 : for_each_possible_cpu(cpu)
11942 : sched_mm_cid_remote_clear_old(mm, cpu);
11943 : weight = cpumask_weight(cidmask);
11944 : /*
11945 : * Clear cids that are greater or equal to the cidmask weight to
11946 : * recompact it.
11947 : */
11948 : for_each_possible_cpu(cpu)
11949 : sched_mm_cid_remote_clear_weight(mm, cpu, weight);
11950 : }
11951 :
11952 : void init_sched_mm_cid(struct task_struct *t)
11953 : {
11954 : struct mm_struct *mm = t->mm;
11955 : int mm_users = 0;
11956 :
11957 : if (mm) {
11958 : mm_users = atomic_read(&mm->mm_users);
11959 : if (mm_users == 1)
11960 : mm->mm_cid_next_scan = jiffies + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11961 : }
11962 : t->cid_work.next = &t->cid_work; /* Protect against double add */
11963 : init_task_work(&t->cid_work, task_mm_cid_work);
11964 : }
11965 :
11966 : void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
11967 : {
11968 : struct callback_head *work = &curr->cid_work;
11969 : unsigned long now = jiffies;
11970 :
11971 : if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
11972 : work->next != work)
11973 : return;
11974 : if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
11975 : return;
11976 : task_work_add(curr, work, TWA_RESUME);
11977 : }
11978 :
11979 : void sched_mm_cid_exit_signals(struct task_struct *t)
11980 : {
11981 : struct mm_struct *mm = t->mm;
11982 : struct rq_flags rf;
11983 : struct rq *rq;
11984 :
11985 : if (!mm)
11986 : return;
11987 :
11988 : preempt_disable();
11989 : rq = this_rq();
11990 : rq_lock_irqsave(rq, &rf);
11991 : preempt_enable_no_resched(); /* holding spinlock */
11992 : WRITE_ONCE(t->mm_cid_active, 0);
11993 : /*
11994 : * Store t->mm_cid_active before loading per-mm/cpu cid.
11995 : * Matches barrier in sched_mm_cid_remote_clear_old().
11996 : */
11997 : smp_mb();
11998 : mm_cid_put(mm);
11999 : t->last_mm_cid = t->mm_cid = -1;
12000 : rq_unlock_irqrestore(rq, &rf);
12001 : }
12002 :
12003 : void sched_mm_cid_before_execve(struct task_struct *t)
12004 : {
12005 : struct mm_struct *mm = t->mm;
12006 : struct rq_flags rf;
12007 : struct rq *rq;
12008 :
12009 : if (!mm)
12010 : return;
12011 :
12012 : preempt_disable();
12013 : rq = this_rq();
12014 : rq_lock_irqsave(rq, &rf);
12015 : preempt_enable_no_resched(); /* holding spinlock */
12016 : WRITE_ONCE(t->mm_cid_active, 0);
12017 : /*
12018 : * Store t->mm_cid_active before loading per-mm/cpu cid.
12019 : * Matches barrier in sched_mm_cid_remote_clear_old().
12020 : */
12021 : smp_mb();
12022 : mm_cid_put(mm);
12023 : t->last_mm_cid = t->mm_cid = -1;
12024 : rq_unlock_irqrestore(rq, &rf);
12025 : }
12026 :
12027 : void sched_mm_cid_after_execve(struct task_struct *t)
12028 : {
12029 : struct mm_struct *mm = t->mm;
12030 : struct rq_flags rf;
12031 : struct rq *rq;
12032 :
12033 : if (!mm)
12034 : return;
12035 :
12036 : preempt_disable();
12037 : rq = this_rq();
12038 : rq_lock_irqsave(rq, &rf);
12039 : preempt_enable_no_resched(); /* holding spinlock */
12040 : WRITE_ONCE(t->mm_cid_active, 1);
12041 : /*
12042 : * Store t->mm_cid_active before loading per-mm/cpu cid.
12043 : * Matches barrier in sched_mm_cid_remote_clear_old().
12044 : */
12045 : smp_mb();
12046 : t->last_mm_cid = t->mm_cid = mm_cid_get(rq, mm);
12047 : rq_unlock_irqrestore(rq, &rf);
12048 : rseq_set_notify_resume(t);
12049 : }
12050 :
12051 : void sched_mm_cid_fork(struct task_struct *t)
12052 : {
12053 : WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
12054 : t->mm_cid_active = 1;
12055 : }
12056 : #endif
|
