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/rseq.h>
61 #include <linux/sched/wake_q.h>
62 #include <linux/scs.h>
63 #include <linux/slab.h>
64 #include <linux/syscalls.h>
65 #include <linux/vtime.h>
66 #include <linux/wait_api.h>
67 #include <linux/workqueue_api.h>
68
69 #ifdef CONFIG_PREEMPT_DYNAMIC
70 # ifdef CONFIG_GENERIC_ENTRY
71 # include <linux/entry-common.h>
72 # endif
73 #endif
74
75 #include <uapi/linux/sched/types.h>
76
77 #include <asm/irq_regs.h>
78 #include <asm/switch_to.h>
79 #include <asm/tlb.h>
80
81 #define CREATE_TRACE_POINTS
82 #include <linux/sched/rseq_api.h>
83 #include <trace/events/sched.h>
84 #include <trace/events/ipi.h>
85 #undef CREATE_TRACE_POINTS
86
87 #include "sched.h"
88 #include "stats.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_hw_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 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_compute_energy_tp);
118
119 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
120
121 #ifdef CONFIG_SCHED_DEBUG
122 /*
123 * Debugging: various feature bits
124 *
125 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
126 * sysctl_sched_features, defined in sched.h, to allow constants propagation
127 * at compile time and compiler optimization based on features default.
128 */
129 #define SCHED_FEAT(name, enabled) \
130 (1UL << __SCHED_FEAT_##name) * enabled |
131 const_debug unsigned int sysctl_sched_features =
132 #include "features.h"
133 0;
134 #undef SCHED_FEAT
135
136 /*
137 * Print a warning if need_resched is set for the given duration (if
138 * LATENCY_WARN is enabled).
139 *
140 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
141 * per boot.
142 */
143 __read_mostly int sysctl_resched_latency_warn_ms = 100;
144 __read_mostly int sysctl_resched_latency_warn_once = 1;
145 #endif /* CONFIG_SCHED_DEBUG */
146
147 /*
148 * Number of tasks to iterate in a single balance run.
149 * Limited because this is done with IRQs disabled.
150 */
151 const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
152
153 __read_mostly int scheduler_running;
154
155 #ifdef CONFIG_SCHED_CORE
156
157 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
158
159 /* kernel prio, less is more */
__task_prio(const struct task_struct * p)160 static inline int __task_prio(const struct task_struct *p)
161 {
162 if (p->sched_class == &stop_sched_class) /* trumps deadline */
163 return -2;
164
165 if (rt_prio(p->prio)) /* includes deadline */
166 return p->prio; /* [-1, 99] */
167
168 if (p->sched_class == &idle_sched_class)
169 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
170
171 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
172 }
173
174 /*
175 * l(a,b)
176 * le(a,b) := !l(b,a)
177 * g(a,b) := l(b,a)
178 * ge(a,b) := !l(a,b)
179 */
180
181 /* real prio, less is less */
prio_less(const struct task_struct * a,const struct task_struct * b,bool in_fi)182 static inline bool prio_less(const struct task_struct *a,
183 const struct task_struct *b, bool in_fi)
184 {
185
186 int pa = __task_prio(a), pb = __task_prio(b);
187
188 if (-pa < -pb)
189 return true;
190
191 if (-pb < -pa)
192 return false;
193
194 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
195 return !dl_time_before(a->dl.deadline, b->dl.deadline);
196
197 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
198 return cfs_prio_less(a, b, in_fi);
199
200 return false;
201 }
202
__sched_core_less(const struct task_struct * a,const struct task_struct * b)203 static inline bool __sched_core_less(const struct task_struct *a,
204 const struct task_struct *b)
205 {
206 if (a->core_cookie < b->core_cookie)
207 return true;
208
209 if (a->core_cookie > b->core_cookie)
210 return false;
211
212 /* flip prio, so high prio is leftmost */
213 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
214 return true;
215
216 return false;
217 }
218
219 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
220
rb_sched_core_less(struct rb_node * a,const struct rb_node * b)221 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
222 {
223 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
224 }
225
rb_sched_core_cmp(const void * key,const struct rb_node * node)226 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
227 {
228 const struct task_struct *p = __node_2_sc(node);
229 unsigned long cookie = (unsigned long)key;
230
231 if (cookie < p->core_cookie)
232 return -1;
233
234 if (cookie > p->core_cookie)
235 return 1;
236
237 return 0;
238 }
239
sched_core_enqueue(struct rq * rq,struct task_struct * p)240 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
241 {
242 rq->core->core_task_seq++;
243
244 if (!p->core_cookie)
245 return;
246
247 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
248 }
249
sched_core_dequeue(struct rq * rq,struct task_struct * p,int flags)250 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
251 {
252 rq->core->core_task_seq++;
253
254 if (sched_core_enqueued(p)) {
255 rb_erase(&p->core_node, &rq->core_tree);
256 RB_CLEAR_NODE(&p->core_node);
257 }
258
259 /*
260 * Migrating the last task off the cpu, with the cpu in forced idle
261 * state. Reschedule to create an accounting edge for forced idle,
262 * and re-examine whether the core is still in forced idle state.
263 */
264 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
265 rq->core->core_forceidle_count && rq->curr == rq->idle)
266 resched_curr(rq);
267 }
268
sched_task_is_throttled(struct task_struct * p,int cpu)269 static int sched_task_is_throttled(struct task_struct *p, int cpu)
270 {
271 if (p->sched_class->task_is_throttled)
272 return p->sched_class->task_is_throttled(p, cpu);
273
274 return 0;
275 }
276
sched_core_next(struct task_struct * p,unsigned long cookie)277 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
278 {
279 struct rb_node *node = &p->core_node;
280 int cpu = task_cpu(p);
281
282 do {
283 node = rb_next(node);
284 if (!node)
285 return NULL;
286
287 p = __node_2_sc(node);
288 if (p->core_cookie != cookie)
289 return NULL;
290
291 } while (sched_task_is_throttled(p, cpu));
292
293 return p;
294 }
295
296 /*
297 * Find left-most (aka, highest priority) and unthrottled task matching @cookie.
298 * If no suitable task is found, NULL will be returned.
299 */
sched_core_find(struct rq * rq,unsigned long cookie)300 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
301 {
302 struct task_struct *p;
303 struct rb_node *node;
304
305 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
306 if (!node)
307 return NULL;
308
309 p = __node_2_sc(node);
310 if (!sched_task_is_throttled(p, rq->cpu))
311 return p;
312
313 return sched_core_next(p, cookie);
314 }
315
316 /*
317 * Magic required such that:
318 *
319 * raw_spin_rq_lock(rq);
320 * ...
321 * raw_spin_rq_unlock(rq);
322 *
323 * ends up locking and unlocking the _same_ lock, and all CPUs
324 * always agree on what rq has what lock.
325 *
326 * XXX entirely possible to selectively enable cores, don't bother for now.
327 */
328
329 static DEFINE_MUTEX(sched_core_mutex);
330 static atomic_t sched_core_count;
331 static struct cpumask sched_core_mask;
332
sched_core_lock(int cpu,unsigned long * flags)333 static void sched_core_lock(int cpu, unsigned long *flags)
334 {
335 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
336 int t, i = 0;
337
338 local_irq_save(*flags);
339 for_each_cpu(t, smt_mask)
340 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
341 }
342
sched_core_unlock(int cpu,unsigned long * flags)343 static void sched_core_unlock(int cpu, unsigned long *flags)
344 {
345 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
346 int t;
347
348 for_each_cpu(t, smt_mask)
349 raw_spin_unlock(&cpu_rq(t)->__lock);
350 local_irq_restore(*flags);
351 }
352
__sched_core_flip(bool enabled)353 static void __sched_core_flip(bool enabled)
354 {
355 unsigned long flags;
356 int cpu, t;
357
358 cpus_read_lock();
359
360 /*
361 * Toggle the online cores, one by one.
362 */
363 cpumask_copy(&sched_core_mask, cpu_online_mask);
364 for_each_cpu(cpu, &sched_core_mask) {
365 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
366
367 sched_core_lock(cpu, &flags);
368
369 for_each_cpu(t, smt_mask)
370 cpu_rq(t)->core_enabled = enabled;
371
372 cpu_rq(cpu)->core->core_forceidle_start = 0;
373
374 sched_core_unlock(cpu, &flags);
375
376 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
377 }
378
379 /*
380 * Toggle the offline CPUs.
381 */
382 for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
383 cpu_rq(cpu)->core_enabled = enabled;
384
385 cpus_read_unlock();
386 }
387
sched_core_assert_empty(void)388 static void sched_core_assert_empty(void)
389 {
390 int cpu;
391
392 for_each_possible_cpu(cpu)
393 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
394 }
395
__sched_core_enable(void)396 static void __sched_core_enable(void)
397 {
398 static_branch_enable(&__sched_core_enabled);
399 /*
400 * Ensure all previous instances of raw_spin_rq_*lock() have finished
401 * and future ones will observe !sched_core_disabled().
402 */
403 synchronize_rcu();
404 __sched_core_flip(true);
405 sched_core_assert_empty();
406 }
407
__sched_core_disable(void)408 static void __sched_core_disable(void)
409 {
410 sched_core_assert_empty();
411 __sched_core_flip(false);
412 static_branch_disable(&__sched_core_enabled);
413 }
414
sched_core_get(void)415 void sched_core_get(void)
416 {
417 if (atomic_inc_not_zero(&sched_core_count))
418 return;
419
420 mutex_lock(&sched_core_mutex);
421 if (!atomic_read(&sched_core_count))
422 __sched_core_enable();
423
424 smp_mb__before_atomic();
425 atomic_inc(&sched_core_count);
426 mutex_unlock(&sched_core_mutex);
427 }
428
__sched_core_put(struct work_struct * work)429 static void __sched_core_put(struct work_struct *work)
430 {
431 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
432 __sched_core_disable();
433 mutex_unlock(&sched_core_mutex);
434 }
435 }
436
sched_core_put(void)437 void sched_core_put(void)
438 {
439 static DECLARE_WORK(_work, __sched_core_put);
440
441 /*
442 * "There can be only one"
443 *
444 * Either this is the last one, or we don't actually need to do any
445 * 'work'. If it is the last *again*, we rely on
446 * WORK_STRUCT_PENDING_BIT.
447 */
448 if (!atomic_add_unless(&sched_core_count, -1, 1))
449 schedule_work(&_work);
450 }
451
452 #else /* !CONFIG_SCHED_CORE */
453
sched_core_enqueue(struct rq * rq,struct task_struct * p)454 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
455 static inline void
sched_core_dequeue(struct rq * rq,struct task_struct * p,int flags)456 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
457
458 #endif /* CONFIG_SCHED_CORE */
459
460 /*
461 * Serialization rules:
462 *
463 * Lock order:
464 *
465 * p->pi_lock
466 * rq->lock
467 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
468 *
469 * rq1->lock
470 * rq2->lock where: rq1 < rq2
471 *
472 * Regular state:
473 *
474 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
475 * local CPU's rq->lock, it optionally removes the task from the runqueue and
476 * always looks at the local rq data structures to find the most eligible task
477 * to run next.
478 *
479 * Task enqueue is also under rq->lock, possibly taken from another CPU.
480 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
481 * the local CPU to avoid bouncing the runqueue state around [ see
482 * ttwu_queue_wakelist() ]
483 *
484 * Task wakeup, specifically wakeups that involve migration, are horribly
485 * complicated to avoid having to take two rq->locks.
486 *
487 * Special state:
488 *
489 * System-calls and anything external will use task_rq_lock() which acquires
490 * both p->pi_lock and rq->lock. As a consequence the state they change is
491 * stable while holding either lock:
492 *
493 * - sched_setaffinity()/
494 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
495 * - set_user_nice(): p->se.load, p->*prio
496 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
497 * p->se.load, p->rt_priority,
498 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
499 * - sched_setnuma(): p->numa_preferred_nid
500 * - sched_move_task(): p->sched_task_group
501 * - uclamp_update_active() p->uclamp*
502 *
503 * p->state <- TASK_*:
504 *
505 * is changed locklessly using set_current_state(), __set_current_state() or
506 * set_special_state(), see their respective comments, or by
507 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
508 * concurrent self.
509 *
510 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
511 *
512 * is set by activate_task() and cleared by deactivate_task(), under
513 * rq->lock. Non-zero indicates the task is runnable, the special
514 * ON_RQ_MIGRATING state is used for migration without holding both
515 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
516 *
517 * p->on_cpu <- { 0, 1 }:
518 *
519 * is set by prepare_task() and cleared by finish_task() such that it will be
520 * set before p is scheduled-in and cleared after p is scheduled-out, both
521 * under rq->lock. Non-zero indicates the task is running on its CPU.
522 *
523 * [ The astute reader will observe that it is possible for two tasks on one
524 * CPU to have ->on_cpu = 1 at the same time. ]
525 *
526 * task_cpu(p): is changed by set_task_cpu(), the rules are:
527 *
528 * - Don't call set_task_cpu() on a blocked task:
529 *
530 * We don't care what CPU we're not running on, this simplifies hotplug,
531 * the CPU assignment of blocked tasks isn't required to be valid.
532 *
533 * - for try_to_wake_up(), called under p->pi_lock:
534 *
535 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
536 *
537 * - for migration called under rq->lock:
538 * [ see task_on_rq_migrating() in task_rq_lock() ]
539 *
540 * o move_queued_task()
541 * o detach_task()
542 *
543 * - for migration called under double_rq_lock():
544 *
545 * o __migrate_swap_task()
546 * o push_rt_task() / pull_rt_task()
547 * o push_dl_task() / pull_dl_task()
548 * o dl_task_offline_migration()
549 *
550 */
551
raw_spin_rq_lock_nested(struct rq * rq,int subclass)552 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
553 {
554 raw_spinlock_t *lock;
555
556 /* Matches synchronize_rcu() in __sched_core_enable() */
557 preempt_disable();
558 if (sched_core_disabled()) {
559 raw_spin_lock_nested(&rq->__lock, subclass);
560 /* preempt_count *MUST* be > 1 */
561 preempt_enable_no_resched();
562 return;
563 }
564
565 for (;;) {
566 lock = __rq_lockp(rq);
567 raw_spin_lock_nested(lock, subclass);
568 if (likely(lock == __rq_lockp(rq))) {
569 /* preempt_count *MUST* be > 1 */
570 preempt_enable_no_resched();
571 return;
572 }
573 raw_spin_unlock(lock);
574 }
575 }
576
raw_spin_rq_trylock(struct rq * rq)577 bool raw_spin_rq_trylock(struct rq *rq)
578 {
579 raw_spinlock_t *lock;
580 bool ret;
581
582 /* Matches synchronize_rcu() in __sched_core_enable() */
583 preempt_disable();
584 if (sched_core_disabled()) {
585 ret = raw_spin_trylock(&rq->__lock);
586 preempt_enable();
587 return ret;
588 }
589
590 for (;;) {
591 lock = __rq_lockp(rq);
592 ret = raw_spin_trylock(lock);
593 if (!ret || (likely(lock == __rq_lockp(rq)))) {
594 preempt_enable();
595 return ret;
596 }
597 raw_spin_unlock(lock);
598 }
599 }
600
raw_spin_rq_unlock(struct rq * rq)601 void raw_spin_rq_unlock(struct rq *rq)
602 {
603 raw_spin_unlock(rq_lockp(rq));
604 }
605
606 #ifdef CONFIG_SMP
607 /*
608 * double_rq_lock - safely lock two runqueues
609 */
double_rq_lock(struct rq * rq1,struct rq * rq2)610 void double_rq_lock(struct rq *rq1, struct rq *rq2)
611 {
612 lockdep_assert_irqs_disabled();
613
614 if (rq_order_less(rq2, rq1))
615 swap(rq1, rq2);
616
617 raw_spin_rq_lock(rq1);
618 if (__rq_lockp(rq1) != __rq_lockp(rq2))
619 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
620
621 double_rq_clock_clear_update(rq1, rq2);
622 }
623 #endif
624
625 /*
626 * __task_rq_lock - lock the rq @p resides on.
627 */
__task_rq_lock(struct task_struct * p,struct rq_flags * rf)628 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
629 __acquires(rq->lock)
630 {
631 struct rq *rq;
632
633 lockdep_assert_held(&p->pi_lock);
634
635 for (;;) {
636 rq = task_rq(p);
637 raw_spin_rq_lock(rq);
638 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
639 rq_pin_lock(rq, rf);
640 return rq;
641 }
642 raw_spin_rq_unlock(rq);
643
644 while (unlikely(task_on_rq_migrating(p)))
645 cpu_relax();
646 }
647 }
648
649 /*
650 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
651 */
task_rq_lock(struct task_struct * p,struct rq_flags * rf)652 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
653 __acquires(p->pi_lock)
654 __acquires(rq->lock)
655 {
656 struct rq *rq;
657
658 for (;;) {
659 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
660 rq = task_rq(p);
661 raw_spin_rq_lock(rq);
662 /*
663 * move_queued_task() task_rq_lock()
664 *
665 * ACQUIRE (rq->lock)
666 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
667 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
668 * [S] ->cpu = new_cpu [L] task_rq()
669 * [L] ->on_rq
670 * RELEASE (rq->lock)
671 *
672 * If we observe the old CPU in task_rq_lock(), the acquire of
673 * the old rq->lock will fully serialize against the stores.
674 *
675 * If we observe the new CPU in task_rq_lock(), the address
676 * dependency headed by '[L] rq = task_rq()' and the acquire
677 * will pair with the WMB to ensure we then also see migrating.
678 */
679 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
680 rq_pin_lock(rq, rf);
681 return rq;
682 }
683 raw_spin_rq_unlock(rq);
684 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
685
686 while (unlikely(task_on_rq_migrating(p)))
687 cpu_relax();
688 }
689 }
690
691 /*
692 * RQ-clock updating methods:
693 */
694
update_rq_clock_task(struct rq * rq,s64 delta)695 static void update_rq_clock_task(struct rq *rq, s64 delta)
696 {
697 /*
698 * In theory, the compile should just see 0 here, and optimize out the call
699 * to sched_rt_avg_update. But I don't trust it...
700 */
701 s64 __maybe_unused steal = 0, irq_delta = 0;
702
703 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
704 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
705
706 /*
707 * Since irq_time is only updated on {soft,}irq_exit, we might run into
708 * this case when a previous update_rq_clock() happened inside a
709 * {soft,}irq region.
710 *
711 * When this happens, we stop ->clock_task and only update the
712 * prev_irq_time stamp to account for the part that fit, so that a next
713 * update will consume the rest. This ensures ->clock_task is
714 * monotonic.
715 *
716 * It does however cause some slight miss-attribution of {soft,}irq
717 * time, a more accurate solution would be to update the irq_time using
718 * the current rq->clock timestamp, except that would require using
719 * atomic ops.
720 */
721 if (irq_delta > delta)
722 irq_delta = delta;
723
724 rq->prev_irq_time += irq_delta;
725 delta -= irq_delta;
726 psi_account_irqtime(rq->curr, irq_delta);
727 delayacct_irq(rq->curr, irq_delta);
728 #endif
729 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
730 if (static_key_false((¶virt_steal_rq_enabled))) {
731 steal = paravirt_steal_clock(cpu_of(rq));
732 steal -= rq->prev_steal_time_rq;
733
734 if (unlikely(steal > delta))
735 steal = delta;
736
737 rq->prev_steal_time_rq += steal;
738 delta -= steal;
739 }
740 #endif
741
742 rq->clock_task += delta;
743
744 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
745 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
746 update_irq_load_avg(rq, irq_delta + steal);
747 #endif
748 update_rq_clock_pelt(rq, delta);
749 }
750
update_rq_clock(struct rq * rq)751 void update_rq_clock(struct rq *rq)
752 {
753 s64 delta;
754
755 lockdep_assert_rq_held(rq);
756
757 if (rq->clock_update_flags & RQCF_ACT_SKIP)
758 return;
759
760 #ifdef CONFIG_SCHED_DEBUG
761 if (sched_feat(WARN_DOUBLE_CLOCK))
762 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
763 rq->clock_update_flags |= RQCF_UPDATED;
764 #endif
765
766 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
767 if (delta < 0)
768 return;
769 rq->clock += delta;
770 update_rq_clock_task(rq, delta);
771 }
772
773 #ifdef CONFIG_SCHED_HRTICK
774 /*
775 * Use HR-timers to deliver accurate preemption points.
776 */
777
hrtick_clear(struct rq * rq)778 static void hrtick_clear(struct rq *rq)
779 {
780 if (hrtimer_active(&rq->hrtick_timer))
781 hrtimer_cancel(&rq->hrtick_timer);
782 }
783
784 /*
785 * High-resolution timer tick.
786 * Runs from hardirq context with interrupts disabled.
787 */
hrtick(struct hrtimer * timer)788 static enum hrtimer_restart hrtick(struct hrtimer *timer)
789 {
790 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
791 struct rq_flags rf;
792
793 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
794
795 rq_lock(rq, &rf);
796 update_rq_clock(rq);
797 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
798 rq_unlock(rq, &rf);
799
800 return HRTIMER_NORESTART;
801 }
802
803 #ifdef CONFIG_SMP
804
__hrtick_restart(struct rq * rq)805 static void __hrtick_restart(struct rq *rq)
806 {
807 struct hrtimer *timer = &rq->hrtick_timer;
808 ktime_t time = rq->hrtick_time;
809
810 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
811 }
812
813 /*
814 * called from hardirq (IPI) context
815 */
__hrtick_start(void * arg)816 static void __hrtick_start(void *arg)
817 {
818 struct rq *rq = arg;
819 struct rq_flags rf;
820
821 rq_lock(rq, &rf);
822 __hrtick_restart(rq);
823 rq_unlock(rq, &rf);
824 }
825
826 /*
827 * Called to set the hrtick timer state.
828 *
829 * called with rq->lock held and irqs disabled
830 */
hrtick_start(struct rq * rq,u64 delay)831 void hrtick_start(struct rq *rq, u64 delay)
832 {
833 struct hrtimer *timer = &rq->hrtick_timer;
834 s64 delta;
835
836 /*
837 * Don't schedule slices shorter than 10000ns, that just
838 * doesn't make sense and can cause timer DoS.
839 */
840 delta = max_t(s64, delay, 10000LL);
841 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
842
843 if (rq == this_rq())
844 __hrtick_restart(rq);
845 else
846 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
847 }
848
849 #else
850 /*
851 * Called to set the hrtick timer state.
852 *
853 * called with rq->lock held and irqs disabled
854 */
hrtick_start(struct rq * rq,u64 delay)855 void hrtick_start(struct rq *rq, u64 delay)
856 {
857 /*
858 * Don't schedule slices shorter than 10000ns, that just
859 * doesn't make sense. Rely on vruntime for fairness.
860 */
861 delay = max_t(u64, delay, 10000LL);
862 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
863 HRTIMER_MODE_REL_PINNED_HARD);
864 }
865
866 #endif /* CONFIG_SMP */
867
hrtick_rq_init(struct rq * rq)868 static void hrtick_rq_init(struct rq *rq)
869 {
870 #ifdef CONFIG_SMP
871 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
872 #endif
873 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
874 rq->hrtick_timer.function = hrtick;
875 }
876 #else /* CONFIG_SCHED_HRTICK */
hrtick_clear(struct rq * rq)877 static inline void hrtick_clear(struct rq *rq)
878 {
879 }
880
hrtick_rq_init(struct rq * rq)881 static inline void hrtick_rq_init(struct rq *rq)
882 {
883 }
884 #endif /* CONFIG_SCHED_HRTICK */
885
886 /*
887 * cmpxchg based fetch_or, macro so it works for different integer types
888 */
889 #define fetch_or(ptr, mask) \
890 ({ \
891 typeof(ptr) _ptr = (ptr); \
892 typeof(mask) _mask = (mask); \
893 typeof(*_ptr) _val = *_ptr; \
894 \
895 do { \
896 } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
897 _val; \
898 })
899
900 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
901 /*
902 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
903 * this avoids any races wrt polling state changes and thereby avoids
904 * spurious IPIs.
905 */
set_nr_and_not_polling(struct task_struct * p)906 static inline bool set_nr_and_not_polling(struct task_struct *p)
907 {
908 struct thread_info *ti = task_thread_info(p);
909 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
910 }
911
912 /*
913 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
914 *
915 * If this returns true, then the idle task promises to call
916 * sched_ttwu_pending() and reschedule soon.
917 */
set_nr_if_polling(struct task_struct * p)918 static bool set_nr_if_polling(struct task_struct *p)
919 {
920 struct thread_info *ti = task_thread_info(p);
921 typeof(ti->flags) val = READ_ONCE(ti->flags);
922
923 do {
924 if (!(val & _TIF_POLLING_NRFLAG))
925 return false;
926 if (val & _TIF_NEED_RESCHED)
927 return true;
928 } while (!try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED));
929
930 return true;
931 }
932
933 #else
set_nr_and_not_polling(struct task_struct * p)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
set_nr_if_polling(struct task_struct * p)941 static inline bool set_nr_if_polling(struct task_struct *p)
942 {
943 return false;
944 }
945 #endif
946 #endif
947
__wake_q_add(struct wake_q_head * head,struct task_struct * task)948 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
949 {
950 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 smp_mb__before_atomic();
961 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 *head->lastp = node;
968 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 */
wake_q_add(struct wake_q_head * head,struct task_struct * task)984 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
985 {
986 if (__wake_q_add(head, task))
987 get_task_struct(task);
988 }
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 */
wake_q_add_safe(struct wake_q_head * head,struct task_struct * task)1007 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
1008 {
1009 if (!__wake_q_add(head, task))
1010 put_task_struct(task);
1011 }
1012
wake_up_q(struct wake_q_head * head)1013 void wake_up_q(struct wake_q_head *head)
1014 {
1015 struct wake_q_node *node = head->first;
1016
1017 while (node != WAKE_Q_TAIL) {
1018 struct task_struct *task;
1019
1020 task = container_of(node, struct task_struct, wake_q);
1021 /* Task can safely be re-inserted now: */
1022 node = node->next;
1023 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 wake_up_process(task);
1030 put_task_struct(task);
1031 }
1032 }
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 */
resched_curr(struct rq * rq)1041 void resched_curr(struct rq *rq)
1042 {
1043 struct task_struct *curr = rq->curr;
1044 int cpu;
1045
1046 lockdep_assert_rq_held(rq);
1047
1048 if (test_tsk_need_resched(curr))
1049 return;
1050
1051 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
resched_cpu(int cpu)1065 void resched_cpu(int cpu)
1066 {
1067 struct rq *rq = cpu_rq(cpu);
1068 unsigned long flags;
1069
1070 raw_spin_rq_lock_irqsave(rq, flags);
1071 if (cpu_online(cpu) || cpu == smp_processor_id())
1072 resched_curr(rq);
1073 raw_spin_rq_unlock_irqrestore(rq, flags);
1074 }
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 */
get_nohz_timer_target(void)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 guard(rcu)();
1101
1102 for_each_domain(cpu, sd) {
1103 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1104 if (cpu == i)
1105 continue;
1106
1107 if (!idle_cpu(i))
1108 return i;
1109 }
1110 }
1111
1112 if (default_cpu == -1)
1113 default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1114
1115 return default_cpu;
1116 }
1117
1118 /*
1119 * When add_timer_on() enqueues a timer into the timer wheel of an
1120 * idle CPU then this timer might expire before the next timer event
1121 * which is scheduled to wake up that CPU. In case of a completely
1122 * idle system the next event might even be infinite time into the
1123 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1124 * leaves the inner idle loop so the newly added timer is taken into
1125 * account when the CPU goes back to idle and evaluates the timer
1126 * wheel for the next timer event.
1127 */
wake_up_idle_cpu(int cpu)1128 static void wake_up_idle_cpu(int cpu)
1129 {
1130 struct rq *rq = cpu_rq(cpu);
1131
1132 if (cpu == smp_processor_id())
1133 return;
1134
1135 /*
1136 * Set TIF_NEED_RESCHED and send an IPI if in the non-polling
1137 * part of the idle loop. This forces an exit from the idle loop
1138 * and a round trip to schedule(). Now this could be optimized
1139 * because a simple new idle loop iteration is enough to
1140 * re-evaluate the next tick. Provided some re-ordering of tick
1141 * nohz functions that would need to follow TIF_NR_POLLING
1142 * clearing:
1143 *
1144 * - On most archs, a simple fetch_or on ti::flags with a
1145 * "0" value would be enough to know if an IPI needs to be sent.
1146 *
1147 * - x86 needs to perform a last need_resched() check between
1148 * monitor and mwait which doesn't take timers into account.
1149 * There a dedicated TIF_TIMER flag would be required to
1150 * fetch_or here and be checked along with TIF_NEED_RESCHED
1151 * before mwait().
1152 *
1153 * However, remote timer enqueue is not such a frequent event
1154 * and testing of the above solutions didn't appear to report
1155 * much benefits.
1156 */
1157 if (set_nr_and_not_polling(rq->idle))
1158 smp_send_reschedule(cpu);
1159 else
1160 trace_sched_wake_idle_without_ipi(cpu);
1161 }
1162
wake_up_full_nohz_cpu(int cpu)1163 static bool wake_up_full_nohz_cpu(int cpu)
1164 {
1165 /*
1166 * We just need the target to call irq_exit() and re-evaluate
1167 * the next tick. The nohz full kick at least implies that.
1168 * If needed we can still optimize that later with an
1169 * empty IRQ.
1170 */
1171 if (cpu_is_offline(cpu))
1172 return true; /* Don't try to wake offline CPUs. */
1173 if (tick_nohz_full_cpu(cpu)) {
1174 if (cpu != smp_processor_id() ||
1175 tick_nohz_tick_stopped())
1176 tick_nohz_full_kick_cpu(cpu);
1177 return true;
1178 }
1179
1180 return false;
1181 }
1182
1183 /*
1184 * Wake up the specified CPU. If the CPU is going offline, it is the
1185 * caller's responsibility to deal with the lost wakeup, for example,
1186 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1187 */
wake_up_nohz_cpu(int cpu)1188 void wake_up_nohz_cpu(int cpu)
1189 {
1190 if (!wake_up_full_nohz_cpu(cpu))
1191 wake_up_idle_cpu(cpu);
1192 }
1193
nohz_csd_func(void * info)1194 static void nohz_csd_func(void *info)
1195 {
1196 struct rq *rq = info;
1197 int cpu = cpu_of(rq);
1198 unsigned int flags;
1199
1200 /*
1201 * Release the rq::nohz_csd.
1202 */
1203 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1204 WARN_ON(!(flags & NOHZ_KICK_MASK));
1205
1206 rq->idle_balance = idle_cpu(cpu);
1207 if (rq->idle_balance && !need_resched()) {
1208 rq->nohz_idle_balance = flags;
1209 raise_softirq_irqoff(SCHED_SOFTIRQ);
1210 }
1211 }
1212
1213 #endif /* CONFIG_NO_HZ_COMMON */
1214
1215 #ifdef CONFIG_NO_HZ_FULL
__need_bw_check(struct rq * rq,struct task_struct * p)1216 static inline bool __need_bw_check(struct rq *rq, struct task_struct *p)
1217 {
1218 if (rq->nr_running != 1)
1219 return false;
1220
1221 if (p->sched_class != &fair_sched_class)
1222 return false;
1223
1224 if (!task_on_rq_queued(p))
1225 return false;
1226
1227 return true;
1228 }
1229
sched_can_stop_tick(struct rq * rq)1230 bool sched_can_stop_tick(struct rq *rq)
1231 {
1232 int fifo_nr_running;
1233
1234 /* Deadline tasks, even if single, need the tick */
1235 if (rq->dl.dl_nr_running)
1236 return false;
1237
1238 /*
1239 * If there are more than one RR tasks, we need the tick to affect the
1240 * actual RR behaviour.
1241 */
1242 if (rq->rt.rr_nr_running) {
1243 if (rq->rt.rr_nr_running == 1)
1244 return true;
1245 else
1246 return false;
1247 }
1248
1249 /*
1250 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1251 * forced preemption between FIFO tasks.
1252 */
1253 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1254 if (fifo_nr_running)
1255 return true;
1256
1257 /*
1258 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1259 * if there's more than one we need the tick for involuntary
1260 * preemption.
1261 */
1262 if (rq->nr_running > 1)
1263 return false;
1264
1265 /*
1266 * If there is one task and it has CFS runtime bandwidth constraints
1267 * and it's on the cpu now we don't want to stop the tick.
1268 * This check prevents clearing the bit if a newly enqueued task here is
1269 * dequeued by migrating while the constrained task continues to run.
1270 * E.g. going from 2->1 without going through pick_next_task().
1271 */
1272 if (sched_feat(HZ_BW) && __need_bw_check(rq, rq->curr)) {
1273 if (cfs_task_bw_constrained(rq->curr))
1274 return false;
1275 }
1276
1277 return true;
1278 }
1279 #endif /* CONFIG_NO_HZ_FULL */
1280 #endif /* CONFIG_SMP */
1281
1282 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1283 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1284 /*
1285 * Iterate task_group tree rooted at *from, calling @down when first entering a
1286 * node and @up when leaving it for the final time.
1287 *
1288 * Caller must hold rcu_lock or sufficient equivalent.
1289 */
walk_tg_tree_from(struct task_group * from,tg_visitor down,tg_visitor up,void * data)1290 int walk_tg_tree_from(struct task_group *from,
1291 tg_visitor down, tg_visitor up, void *data)
1292 {
1293 struct task_group *parent, *child;
1294 int ret;
1295
1296 parent = from;
1297
1298 down:
1299 ret = (*down)(parent, data);
1300 if (ret)
1301 goto out;
1302 list_for_each_entry_rcu(child, &parent->children, siblings) {
1303 parent = child;
1304 goto down;
1305
1306 up:
1307 continue;
1308 }
1309 ret = (*up)(parent, data);
1310 if (ret || parent == from)
1311 goto out;
1312
1313 child = parent;
1314 parent = parent->parent;
1315 if (parent)
1316 goto up;
1317 out:
1318 return ret;
1319 }
1320
tg_nop(struct task_group * tg,void * data)1321 int tg_nop(struct task_group *tg, void *data)
1322 {
1323 return 0;
1324 }
1325 #endif
1326
set_load_weight(struct task_struct * p,bool update_load)1327 static void set_load_weight(struct task_struct *p, bool update_load)
1328 {
1329 int prio = p->static_prio - MAX_RT_PRIO;
1330 struct load_weight *load = &p->se.load;
1331
1332 /*
1333 * SCHED_IDLE tasks get minimal weight:
1334 */
1335 if (task_has_idle_policy(p)) {
1336 load->weight = scale_load(WEIGHT_IDLEPRIO);
1337 load->inv_weight = WMULT_IDLEPRIO;
1338 return;
1339 }
1340
1341 /*
1342 * SCHED_OTHER tasks have to update their load when changing their
1343 * weight
1344 */
1345 if (update_load && p->sched_class == &fair_sched_class) {
1346 reweight_task(p, prio);
1347 } else {
1348 load->weight = scale_load(sched_prio_to_weight[prio]);
1349 load->inv_weight = sched_prio_to_wmult[prio];
1350 }
1351 }
1352
1353 #ifdef CONFIG_UCLAMP_TASK
1354 /*
1355 * Serializes updates of utilization clamp values
1356 *
1357 * The (slow-path) user-space triggers utilization clamp value updates which
1358 * can require updates on (fast-path) scheduler's data structures used to
1359 * support enqueue/dequeue operations.
1360 * While the per-CPU rq lock protects fast-path update operations, user-space
1361 * requests are serialized using a mutex to reduce the risk of conflicting
1362 * updates or API abuses.
1363 */
1364 static DEFINE_MUTEX(uclamp_mutex);
1365
1366 /* Max allowed minimum utilization */
1367 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1368
1369 /* Max allowed maximum utilization */
1370 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1371
1372 /*
1373 * By default RT tasks run at the maximum performance point/capacity of the
1374 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1375 * SCHED_CAPACITY_SCALE.
1376 *
1377 * This knob allows admins to change the default behavior when uclamp is being
1378 * used. In battery powered devices, particularly, running at the maximum
1379 * capacity and frequency will increase energy consumption and shorten the
1380 * battery life.
1381 *
1382 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1383 *
1384 * This knob will not override the system default sched_util_clamp_min defined
1385 * above.
1386 */
1387 static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1388
1389 /* All clamps are required to be less or equal than these values */
1390 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1391
1392 /*
1393 * This static key is used to reduce the uclamp overhead in the fast path. It
1394 * primarily disables the call to uclamp_rq_{inc, dec}() in
1395 * enqueue/dequeue_task().
1396 *
1397 * This allows users to continue to enable uclamp in their kernel config with
1398 * minimum uclamp overhead in the fast path.
1399 *
1400 * As soon as userspace modifies any of the uclamp knobs, the static key is
1401 * enabled, since we have an actual users that make use of uclamp
1402 * functionality.
1403 *
1404 * The knobs that would enable this static key are:
1405 *
1406 * * A task modifying its uclamp value with sched_setattr().
1407 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1408 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1409 */
1410 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1411
1412 /* Integer rounded range for each bucket */
1413 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1414
1415 #define for_each_clamp_id(clamp_id) \
1416 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1417
uclamp_bucket_id(unsigned int clamp_value)1418 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1419 {
1420 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1421 }
1422
uclamp_none(enum uclamp_id clamp_id)1423 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1424 {
1425 if (clamp_id == UCLAMP_MIN)
1426 return 0;
1427 return SCHED_CAPACITY_SCALE;
1428 }
1429
uclamp_se_set(struct uclamp_se * uc_se,unsigned int value,bool user_defined)1430 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1431 unsigned int value, bool user_defined)
1432 {
1433 uc_se->value = value;
1434 uc_se->bucket_id = uclamp_bucket_id(value);
1435 uc_se->user_defined = user_defined;
1436 }
1437
1438 static inline unsigned int
uclamp_idle_value(struct rq * rq,enum uclamp_id clamp_id,unsigned int clamp_value)1439 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1440 unsigned int clamp_value)
1441 {
1442 /*
1443 * Avoid blocked utilization pushing up the frequency when we go
1444 * idle (which drops the max-clamp) by retaining the last known
1445 * max-clamp.
1446 */
1447 if (clamp_id == UCLAMP_MAX) {
1448 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1449 return clamp_value;
1450 }
1451
1452 return uclamp_none(UCLAMP_MIN);
1453 }
1454
uclamp_idle_reset(struct rq * rq,enum uclamp_id clamp_id,unsigned int clamp_value)1455 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1456 unsigned int clamp_value)
1457 {
1458 /* Reset max-clamp retention only on idle exit */
1459 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1460 return;
1461
1462 uclamp_rq_set(rq, clamp_id, clamp_value);
1463 }
1464
1465 static inline
uclamp_rq_max_value(struct rq * rq,enum uclamp_id clamp_id,unsigned int clamp_value)1466 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1467 unsigned int clamp_value)
1468 {
1469 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1470 int bucket_id = UCLAMP_BUCKETS - 1;
1471
1472 /*
1473 * Since both min and max clamps are max aggregated, find the
1474 * top most bucket with tasks in.
1475 */
1476 for ( ; bucket_id >= 0; bucket_id--) {
1477 if (!bucket[bucket_id].tasks)
1478 continue;
1479 return bucket[bucket_id].value;
1480 }
1481
1482 /* No tasks -- default clamp values */
1483 return uclamp_idle_value(rq, clamp_id, clamp_value);
1484 }
1485
__uclamp_update_util_min_rt_default(struct task_struct * p)1486 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1487 {
1488 unsigned int default_util_min;
1489 struct uclamp_se *uc_se;
1490
1491 lockdep_assert_held(&p->pi_lock);
1492
1493 uc_se = &p->uclamp_req[UCLAMP_MIN];
1494
1495 /* Only sync if user didn't override the default */
1496 if (uc_se->user_defined)
1497 return;
1498
1499 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1500 uclamp_se_set(uc_se, default_util_min, false);
1501 }
1502
uclamp_update_util_min_rt_default(struct task_struct * p)1503 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1504 {
1505 if (!rt_task(p))
1506 return;
1507
1508 /* Protect updates to p->uclamp_* */
1509 guard(task_rq_lock)(p);
1510 __uclamp_update_util_min_rt_default(p);
1511 }
1512
1513 static inline struct uclamp_se
uclamp_tg_restrict(struct task_struct * p,enum uclamp_id clamp_id)1514 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1515 {
1516 /* Copy by value as we could modify it */
1517 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1518 #ifdef CONFIG_UCLAMP_TASK_GROUP
1519 unsigned int tg_min, tg_max, value;
1520
1521 /*
1522 * Tasks in autogroups or root task group will be
1523 * restricted by system defaults.
1524 */
1525 if (task_group_is_autogroup(task_group(p)))
1526 return uc_req;
1527 if (task_group(p) == &root_task_group)
1528 return uc_req;
1529
1530 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1531 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1532 value = uc_req.value;
1533 value = clamp(value, tg_min, tg_max);
1534 uclamp_se_set(&uc_req, value, false);
1535 #endif
1536
1537 return uc_req;
1538 }
1539
1540 /*
1541 * The effective clamp bucket index of a task depends on, by increasing
1542 * priority:
1543 * - the task specific clamp value, when explicitly requested from userspace
1544 * - the task group effective clamp value, for tasks not either in the root
1545 * group or in an autogroup
1546 * - the system default clamp value, defined by the sysadmin
1547 */
1548 static inline struct uclamp_se
uclamp_eff_get(struct task_struct * p,enum uclamp_id clamp_id)1549 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1550 {
1551 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1552 struct uclamp_se uc_max = uclamp_default[clamp_id];
1553
1554 /* System default restrictions always apply */
1555 if (unlikely(uc_req.value > uc_max.value))
1556 return uc_max;
1557
1558 return uc_req;
1559 }
1560
uclamp_eff_value(struct task_struct * p,enum uclamp_id clamp_id)1561 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1562 {
1563 struct uclamp_se uc_eff;
1564
1565 /* Task currently refcounted: use back-annotated (effective) value */
1566 if (p->uclamp[clamp_id].active)
1567 return (unsigned long)p->uclamp[clamp_id].value;
1568
1569 uc_eff = uclamp_eff_get(p, clamp_id);
1570
1571 return (unsigned long)uc_eff.value;
1572 }
1573
1574 /*
1575 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1576 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1577 * updates the rq's clamp value if required.
1578 *
1579 * Tasks can have a task-specific value requested from user-space, track
1580 * within each bucket the maximum value for tasks refcounted in it.
1581 * This "local max aggregation" allows to track the exact "requested" value
1582 * for each bucket when all its RUNNABLE tasks require the same clamp.
1583 */
uclamp_rq_inc_id(struct rq * rq,struct task_struct * p,enum uclamp_id clamp_id)1584 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1585 enum uclamp_id clamp_id)
1586 {
1587 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1588 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1589 struct uclamp_bucket *bucket;
1590
1591 lockdep_assert_rq_held(rq);
1592
1593 /* Update task effective clamp */
1594 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1595
1596 bucket = &uc_rq->bucket[uc_se->bucket_id];
1597 bucket->tasks++;
1598 uc_se->active = true;
1599
1600 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1601
1602 /*
1603 * Local max aggregation: rq buckets always track the max
1604 * "requested" clamp value of its RUNNABLE tasks.
1605 */
1606 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1607 bucket->value = uc_se->value;
1608
1609 if (uc_se->value > uclamp_rq_get(rq, clamp_id))
1610 uclamp_rq_set(rq, clamp_id, uc_se->value);
1611 }
1612
1613 /*
1614 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1615 * is released. If this is the last task reference counting the rq's max
1616 * active clamp value, then the rq's clamp value is updated.
1617 *
1618 * Both refcounted tasks and rq's cached clamp values are expected to be
1619 * always valid. If it's detected they are not, as defensive programming,
1620 * enforce the expected state and warn.
1621 */
uclamp_rq_dec_id(struct rq * rq,struct task_struct * p,enum uclamp_id clamp_id)1622 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1623 enum uclamp_id clamp_id)
1624 {
1625 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1626 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1627 struct uclamp_bucket *bucket;
1628 unsigned int bkt_clamp;
1629 unsigned int rq_clamp;
1630
1631 lockdep_assert_rq_held(rq);
1632
1633 /*
1634 * If sched_uclamp_used was enabled after task @p was enqueued,
1635 * we could end up with unbalanced call to uclamp_rq_dec_id().
1636 *
1637 * In this case the uc_se->active flag should be false since no uclamp
1638 * accounting was performed at enqueue time and we can just return
1639 * here.
1640 *
1641 * Need to be careful of the following enqueue/dequeue ordering
1642 * problem too
1643 *
1644 * enqueue(taskA)
1645 * // sched_uclamp_used gets enabled
1646 * enqueue(taskB)
1647 * dequeue(taskA)
1648 * // Must not decrement bucket->tasks here
1649 * dequeue(taskB)
1650 *
1651 * where we could end up with stale data in uc_se and
1652 * bucket[uc_se->bucket_id].
1653 *
1654 * The following check here eliminates the possibility of such race.
1655 */
1656 if (unlikely(!uc_se->active))
1657 return;
1658
1659 bucket = &uc_rq->bucket[uc_se->bucket_id];
1660
1661 SCHED_WARN_ON(!bucket->tasks);
1662 if (likely(bucket->tasks))
1663 bucket->tasks--;
1664
1665 uc_se->active = false;
1666
1667 /*
1668 * Keep "local max aggregation" simple and accept to (possibly)
1669 * overboost some RUNNABLE tasks in the same bucket.
1670 * The rq clamp bucket value is reset to its base value whenever
1671 * there are no more RUNNABLE tasks refcounting it.
1672 */
1673 if (likely(bucket->tasks))
1674 return;
1675
1676 rq_clamp = uclamp_rq_get(rq, clamp_id);
1677 /*
1678 * Defensive programming: this should never happen. If it happens,
1679 * e.g. due to future modification, warn and fixup the expected value.
1680 */
1681 SCHED_WARN_ON(bucket->value > rq_clamp);
1682 if (bucket->value >= rq_clamp) {
1683 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1684 uclamp_rq_set(rq, clamp_id, bkt_clamp);
1685 }
1686 }
1687
uclamp_rq_inc(struct rq * rq,struct task_struct * p)1688 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1689 {
1690 enum uclamp_id clamp_id;
1691
1692 /*
1693 * Avoid any overhead until uclamp is actually used by the userspace.
1694 *
1695 * The condition is constructed such that a NOP is generated when
1696 * sched_uclamp_used is disabled.
1697 */
1698 if (!static_branch_unlikely(&sched_uclamp_used))
1699 return;
1700
1701 if (unlikely(!p->sched_class->uclamp_enabled))
1702 return;
1703
1704 for_each_clamp_id(clamp_id)
1705 uclamp_rq_inc_id(rq, p, clamp_id);
1706
1707 /* Reset clamp idle holding when there is one RUNNABLE task */
1708 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1709 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1710 }
1711
uclamp_rq_dec(struct rq * rq,struct task_struct * p)1712 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1713 {
1714 enum uclamp_id clamp_id;
1715
1716 /*
1717 * Avoid any overhead until uclamp is actually used by the userspace.
1718 *
1719 * The condition is constructed such that a NOP is generated when
1720 * sched_uclamp_used is disabled.
1721 */
1722 if (!static_branch_unlikely(&sched_uclamp_used))
1723 return;
1724
1725 if (unlikely(!p->sched_class->uclamp_enabled))
1726 return;
1727
1728 for_each_clamp_id(clamp_id)
1729 uclamp_rq_dec_id(rq, p, clamp_id);
1730 }
1731
uclamp_rq_reinc_id(struct rq * rq,struct task_struct * p,enum uclamp_id clamp_id)1732 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1733 enum uclamp_id clamp_id)
1734 {
1735 if (!p->uclamp[clamp_id].active)
1736 return;
1737
1738 uclamp_rq_dec_id(rq, p, clamp_id);
1739 uclamp_rq_inc_id(rq, p, clamp_id);
1740
1741 /*
1742 * Make sure to clear the idle flag if we've transiently reached 0
1743 * active tasks on rq.
1744 */
1745 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1746 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1747 }
1748
1749 static inline void
uclamp_update_active(struct task_struct * p)1750 uclamp_update_active(struct task_struct *p)
1751 {
1752 enum uclamp_id clamp_id;
1753 struct rq_flags rf;
1754 struct rq *rq;
1755
1756 /*
1757 * Lock the task and the rq where the task is (or was) queued.
1758 *
1759 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1760 * price to pay to safely serialize util_{min,max} updates with
1761 * enqueues, dequeues and migration operations.
1762 * This is the same locking schema used by __set_cpus_allowed_ptr().
1763 */
1764 rq = task_rq_lock(p, &rf);
1765
1766 /*
1767 * Setting the clamp bucket is serialized by task_rq_lock().
1768 * If the task is not yet RUNNABLE and its task_struct is not
1769 * affecting a valid clamp bucket, the next time it's enqueued,
1770 * it will already see the updated clamp bucket value.
1771 */
1772 for_each_clamp_id(clamp_id)
1773 uclamp_rq_reinc_id(rq, p, clamp_id);
1774
1775 task_rq_unlock(rq, p, &rf);
1776 }
1777
1778 #ifdef CONFIG_UCLAMP_TASK_GROUP
1779 static inline void
uclamp_update_active_tasks(struct cgroup_subsys_state * css)1780 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1781 {
1782 struct css_task_iter it;
1783 struct task_struct *p;
1784
1785 css_task_iter_start(css, 0, &it);
1786 while ((p = css_task_iter_next(&it)))
1787 uclamp_update_active(p);
1788 css_task_iter_end(&it);
1789 }
1790
1791 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1792 #endif
1793
1794 #ifdef CONFIG_SYSCTL
1795 #ifdef CONFIG_UCLAMP_TASK_GROUP
uclamp_update_root_tg(void)1796 static void uclamp_update_root_tg(void)
1797 {
1798 struct task_group *tg = &root_task_group;
1799
1800 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1801 sysctl_sched_uclamp_util_min, false);
1802 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1803 sysctl_sched_uclamp_util_max, false);
1804
1805 guard(rcu)();
1806 cpu_util_update_eff(&root_task_group.css);
1807 }
1808 #else
uclamp_update_root_tg(void)1809 static void uclamp_update_root_tg(void) { }
1810 #endif
1811
uclamp_sync_util_min_rt_default(void)1812 static void uclamp_sync_util_min_rt_default(void)
1813 {
1814 struct task_struct *g, *p;
1815
1816 /*
1817 * copy_process() sysctl_uclamp
1818 * uclamp_min_rt = X;
1819 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1820 * // link thread smp_mb__after_spinlock()
1821 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1822 * sched_post_fork() for_each_process_thread()
1823 * __uclamp_sync_rt() __uclamp_sync_rt()
1824 *
1825 * Ensures that either sched_post_fork() will observe the new
1826 * uclamp_min_rt or for_each_process_thread() will observe the new
1827 * task.
1828 */
1829 read_lock(&tasklist_lock);
1830 smp_mb__after_spinlock();
1831 read_unlock(&tasklist_lock);
1832
1833 guard(rcu)();
1834 for_each_process_thread(g, p)
1835 uclamp_update_util_min_rt_default(p);
1836 }
1837
sysctl_sched_uclamp_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)1838 static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1839 void *buffer, size_t *lenp, loff_t *ppos)
1840 {
1841 bool update_root_tg = false;
1842 int old_min, old_max, old_min_rt;
1843 int result;
1844
1845 guard(mutex)(&uclamp_mutex);
1846
1847 old_min = sysctl_sched_uclamp_util_min;
1848 old_max = sysctl_sched_uclamp_util_max;
1849 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1850
1851 result = proc_dointvec(table, write, buffer, lenp, ppos);
1852 if (result)
1853 goto undo;
1854 if (!write)
1855 return 0;
1856
1857 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1858 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1859 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1860
1861 result = -EINVAL;
1862 goto undo;
1863 }
1864
1865 if (old_min != sysctl_sched_uclamp_util_min) {
1866 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1867 sysctl_sched_uclamp_util_min, false);
1868 update_root_tg = true;
1869 }
1870 if (old_max != sysctl_sched_uclamp_util_max) {
1871 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1872 sysctl_sched_uclamp_util_max, false);
1873 update_root_tg = true;
1874 }
1875
1876 if (update_root_tg) {
1877 static_branch_enable(&sched_uclamp_used);
1878 uclamp_update_root_tg();
1879 }
1880
1881 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1882 static_branch_enable(&sched_uclamp_used);
1883 uclamp_sync_util_min_rt_default();
1884 }
1885
1886 /*
1887 * We update all RUNNABLE tasks only when task groups are in use.
1888 * Otherwise, keep it simple and do just a lazy update at each next
1889 * task enqueue time.
1890 */
1891 return 0;
1892
1893 undo:
1894 sysctl_sched_uclamp_util_min = old_min;
1895 sysctl_sched_uclamp_util_max = old_max;
1896 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1897 return result;
1898 }
1899 #endif
1900
uclamp_validate(struct task_struct * p,const struct sched_attr * attr)1901 static int uclamp_validate(struct task_struct *p,
1902 const struct sched_attr *attr)
1903 {
1904 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1905 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1906
1907 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1908 util_min = attr->sched_util_min;
1909
1910 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1911 return -EINVAL;
1912 }
1913
1914 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1915 util_max = attr->sched_util_max;
1916
1917 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1918 return -EINVAL;
1919 }
1920
1921 if (util_min != -1 && util_max != -1 && util_min > util_max)
1922 return -EINVAL;
1923
1924 /*
1925 * We have valid uclamp attributes; make sure uclamp is enabled.
1926 *
1927 * We need to do that here, because enabling static branches is a
1928 * blocking operation which obviously cannot be done while holding
1929 * scheduler locks.
1930 */
1931 static_branch_enable(&sched_uclamp_used);
1932
1933 return 0;
1934 }
1935
uclamp_reset(const struct sched_attr * attr,enum uclamp_id clamp_id,struct uclamp_se * uc_se)1936 static bool uclamp_reset(const struct sched_attr *attr,
1937 enum uclamp_id clamp_id,
1938 struct uclamp_se *uc_se)
1939 {
1940 /* Reset on sched class change for a non user-defined clamp value. */
1941 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1942 !uc_se->user_defined)
1943 return true;
1944
1945 /* Reset on sched_util_{min,max} == -1. */
1946 if (clamp_id == UCLAMP_MIN &&
1947 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1948 attr->sched_util_min == -1) {
1949 return true;
1950 }
1951
1952 if (clamp_id == UCLAMP_MAX &&
1953 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1954 attr->sched_util_max == -1) {
1955 return true;
1956 }
1957
1958 return false;
1959 }
1960
__setscheduler_uclamp(struct task_struct * p,const struct sched_attr * attr)1961 static void __setscheduler_uclamp(struct task_struct *p,
1962 const struct sched_attr *attr)
1963 {
1964 enum uclamp_id clamp_id;
1965
1966 for_each_clamp_id(clamp_id) {
1967 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1968 unsigned int value;
1969
1970 if (!uclamp_reset(attr, clamp_id, uc_se))
1971 continue;
1972
1973 /*
1974 * RT by default have a 100% boost value that could be modified
1975 * at runtime.
1976 */
1977 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1978 value = sysctl_sched_uclamp_util_min_rt_default;
1979 else
1980 value = uclamp_none(clamp_id);
1981
1982 uclamp_se_set(uc_se, value, false);
1983
1984 }
1985
1986 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1987 return;
1988
1989 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1990 attr->sched_util_min != -1) {
1991 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1992 attr->sched_util_min, true);
1993 }
1994
1995 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1996 attr->sched_util_max != -1) {
1997 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1998 attr->sched_util_max, true);
1999 }
2000 }
2001
uclamp_fork(struct task_struct * p)2002 static void uclamp_fork(struct task_struct *p)
2003 {
2004 enum uclamp_id clamp_id;
2005
2006 /*
2007 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
2008 * as the task is still at its early fork stages.
2009 */
2010 for_each_clamp_id(clamp_id)
2011 p->uclamp[clamp_id].active = false;
2012
2013 if (likely(!p->sched_reset_on_fork))
2014 return;
2015
2016 for_each_clamp_id(clamp_id) {
2017 uclamp_se_set(&p->uclamp_req[clamp_id],
2018 uclamp_none(clamp_id), false);
2019 }
2020 }
2021
uclamp_post_fork(struct task_struct * p)2022 static void uclamp_post_fork(struct task_struct *p)
2023 {
2024 uclamp_update_util_min_rt_default(p);
2025 }
2026
init_uclamp_rq(struct rq * rq)2027 static void __init init_uclamp_rq(struct rq *rq)
2028 {
2029 enum uclamp_id clamp_id;
2030 struct uclamp_rq *uc_rq = rq->uclamp;
2031
2032 for_each_clamp_id(clamp_id) {
2033 uc_rq[clamp_id] = (struct uclamp_rq) {
2034 .value = uclamp_none(clamp_id)
2035 };
2036 }
2037
2038 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
2039 }
2040
init_uclamp(void)2041 static void __init init_uclamp(void)
2042 {
2043 struct uclamp_se uc_max = {};
2044 enum uclamp_id clamp_id;
2045 int cpu;
2046
2047 for_each_possible_cpu(cpu)
2048 init_uclamp_rq(cpu_rq(cpu));
2049
2050 for_each_clamp_id(clamp_id) {
2051 uclamp_se_set(&init_task.uclamp_req[clamp_id],
2052 uclamp_none(clamp_id), false);
2053 }
2054
2055 /* System defaults allow max clamp values for both indexes */
2056 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2057 for_each_clamp_id(clamp_id) {
2058 uclamp_default[clamp_id] = uc_max;
2059 #ifdef CONFIG_UCLAMP_TASK_GROUP
2060 root_task_group.uclamp_req[clamp_id] = uc_max;
2061 root_task_group.uclamp[clamp_id] = uc_max;
2062 #endif
2063 }
2064 }
2065
2066 #else /* !CONFIG_UCLAMP_TASK */
uclamp_rq_inc(struct rq * rq,struct task_struct * p)2067 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
uclamp_rq_dec(struct rq * rq,struct task_struct * p)2068 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
uclamp_validate(struct task_struct * p,const struct sched_attr * attr)2069 static inline int uclamp_validate(struct task_struct *p,
2070 const struct sched_attr *attr)
2071 {
2072 return -EOPNOTSUPP;
2073 }
__setscheduler_uclamp(struct task_struct * p,const struct sched_attr * attr)2074 static void __setscheduler_uclamp(struct task_struct *p,
2075 const struct sched_attr *attr) { }
uclamp_fork(struct task_struct * p)2076 static inline void uclamp_fork(struct task_struct *p) { }
uclamp_post_fork(struct task_struct * p)2077 static inline void uclamp_post_fork(struct task_struct *p) { }
init_uclamp(void)2078 static inline void init_uclamp(void) { }
2079 #endif /* CONFIG_UCLAMP_TASK */
2080
sched_task_on_rq(struct task_struct * p)2081 bool sched_task_on_rq(struct task_struct *p)
2082 {
2083 return task_on_rq_queued(p);
2084 }
2085
get_wchan(struct task_struct * p)2086 unsigned long get_wchan(struct task_struct *p)
2087 {
2088 unsigned long ip = 0;
2089 unsigned int state;
2090
2091 if (!p || p == current)
2092 return 0;
2093
2094 /* Only get wchan if task is blocked and we can keep it that way. */
2095 raw_spin_lock_irq(&p->pi_lock);
2096 state = READ_ONCE(p->__state);
2097 smp_rmb(); /* see try_to_wake_up() */
2098 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2099 ip = __get_wchan(p);
2100 raw_spin_unlock_irq(&p->pi_lock);
2101
2102 return ip;
2103 }
2104
enqueue_task(struct rq * rq,struct task_struct * p,int flags)2105 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2106 {
2107 if (!(flags & ENQUEUE_NOCLOCK))
2108 update_rq_clock(rq);
2109
2110 if (!(flags & ENQUEUE_RESTORE)) {
2111 sched_info_enqueue(rq, p);
2112 psi_enqueue(p, (flags & ENQUEUE_WAKEUP) && !(flags & ENQUEUE_MIGRATED));
2113 }
2114
2115 uclamp_rq_inc(rq, p);
2116 p->sched_class->enqueue_task(rq, p, flags);
2117
2118 if (sched_core_enabled(rq))
2119 sched_core_enqueue(rq, p);
2120 }
2121
dequeue_task(struct rq * rq,struct task_struct * p,int flags)2122 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2123 {
2124 if (sched_core_enabled(rq))
2125 sched_core_dequeue(rq, p, flags);
2126
2127 if (!(flags & DEQUEUE_NOCLOCK))
2128 update_rq_clock(rq);
2129
2130 if (!(flags & DEQUEUE_SAVE)) {
2131 sched_info_dequeue(rq, p);
2132 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2133 }
2134
2135 uclamp_rq_dec(rq, p);
2136 p->sched_class->dequeue_task(rq, p, flags);
2137 }
2138
activate_task(struct rq * rq,struct task_struct * p,int flags)2139 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2140 {
2141 if (task_on_rq_migrating(p))
2142 flags |= ENQUEUE_MIGRATED;
2143 if (flags & ENQUEUE_MIGRATED)
2144 sched_mm_cid_migrate_to(rq, p);
2145
2146 enqueue_task(rq, p, flags);
2147
2148 WRITE_ONCE(p->on_rq, TASK_ON_RQ_QUEUED);
2149 ASSERT_EXCLUSIVE_WRITER(p->on_rq);
2150 }
2151
deactivate_task(struct rq * rq,struct task_struct * p,int flags)2152 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2153 {
2154 WRITE_ONCE(p->on_rq, (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING);
2155 ASSERT_EXCLUSIVE_WRITER(p->on_rq);
2156
2157 dequeue_task(rq, p, flags);
2158 }
2159
__normal_prio(int policy,int rt_prio,int nice)2160 static inline int __normal_prio(int policy, int rt_prio, int nice)
2161 {
2162 int prio;
2163
2164 if (dl_policy(policy))
2165 prio = MAX_DL_PRIO - 1;
2166 else if (rt_policy(policy))
2167 prio = MAX_RT_PRIO - 1 - rt_prio;
2168 else
2169 prio = NICE_TO_PRIO(nice);
2170
2171 return prio;
2172 }
2173
2174 /*
2175 * Calculate the expected normal priority: i.e. priority
2176 * without taking RT-inheritance into account. Might be
2177 * boosted by interactivity modifiers. Changes upon fork,
2178 * setprio syscalls, and whenever the interactivity
2179 * estimator recalculates.
2180 */
normal_prio(struct task_struct * p)2181 static inline int normal_prio(struct task_struct *p)
2182 {
2183 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2184 }
2185
2186 /*
2187 * Calculate the current priority, i.e. the priority
2188 * taken into account by the scheduler. This value might
2189 * be boosted by RT tasks, or might be boosted by
2190 * interactivity modifiers. Will be RT if the task got
2191 * RT-boosted. If not then it returns p->normal_prio.
2192 */
effective_prio(struct task_struct * p)2193 static int effective_prio(struct task_struct *p)
2194 {
2195 p->normal_prio = normal_prio(p);
2196 /*
2197 * If we are RT tasks or we were boosted to RT priority,
2198 * keep the priority unchanged. Otherwise, update priority
2199 * to the normal priority:
2200 */
2201 if (!rt_prio(p->prio))
2202 return p->normal_prio;
2203 return p->prio;
2204 }
2205
2206 /**
2207 * task_curr - is this task currently executing on a CPU?
2208 * @p: the task in question.
2209 *
2210 * Return: 1 if the task is currently executing. 0 otherwise.
2211 */
task_curr(const struct task_struct * p)2212 inline int task_curr(const struct task_struct *p)
2213 {
2214 return cpu_curr(task_cpu(p)) == p;
2215 }
2216
2217 /*
2218 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2219 * use the balance_callback list if you want balancing.
2220 *
2221 * this means any call to check_class_changed() must be followed by a call to
2222 * balance_callback().
2223 */
check_class_changed(struct rq * rq,struct task_struct * p,const struct sched_class * prev_class,int oldprio)2224 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2225 const struct sched_class *prev_class,
2226 int oldprio)
2227 {
2228 if (prev_class != p->sched_class) {
2229 if (prev_class->switched_from)
2230 prev_class->switched_from(rq, p);
2231
2232 p->sched_class->switched_to(rq, p);
2233 } else if (oldprio != p->prio || dl_task(p))
2234 p->sched_class->prio_changed(rq, p, oldprio);
2235 }
2236
wakeup_preempt(struct rq * rq,struct task_struct * p,int flags)2237 void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags)
2238 {
2239 if (p->sched_class == rq->curr->sched_class)
2240 rq->curr->sched_class->wakeup_preempt(rq, p, flags);
2241 else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2242 resched_curr(rq);
2243
2244 /*
2245 * A queue event has occurred, and we're going to schedule. In
2246 * this case, we can save a useless back to back clock update.
2247 */
2248 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2249 rq_clock_skip_update(rq);
2250 }
2251
2252 static __always_inline
__task_state_match(struct task_struct * p,unsigned int state)2253 int __task_state_match(struct task_struct *p, unsigned int state)
2254 {
2255 if (READ_ONCE(p->__state) & state)
2256 return 1;
2257
2258 if (READ_ONCE(p->saved_state) & state)
2259 return -1;
2260
2261 return 0;
2262 }
2263
2264 static __always_inline
task_state_match(struct task_struct * p,unsigned int state)2265 int task_state_match(struct task_struct *p, unsigned int state)
2266 {
2267 /*
2268 * Serialize against current_save_and_set_rtlock_wait_state(),
2269 * current_restore_rtlock_saved_state(), and __refrigerator().
2270 */
2271 guard(raw_spinlock_irq)(&p->pi_lock);
2272 return __task_state_match(p, state);
2273 }
2274
2275 /*
2276 * wait_task_inactive - wait for a thread to unschedule.
2277 *
2278 * Wait for the thread to block in any of the states set in @match_state.
2279 * If it changes, i.e. @p might have woken up, then return zero. When we
2280 * succeed in waiting for @p to be off its CPU, we return a positive number
2281 * (its total switch count). If a second call a short while later returns the
2282 * same number, the caller can be sure that @p has remained unscheduled the
2283 * whole time.
2284 *
2285 * The caller must ensure that the task *will* unschedule sometime soon,
2286 * else this function might spin for a *long* time. This function can't
2287 * be called with interrupts off, or it may introduce deadlock with
2288 * smp_call_function() if an IPI is sent by the same process we are
2289 * waiting to become inactive.
2290 */
wait_task_inactive(struct task_struct * p,unsigned int match_state)2291 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
2292 {
2293 int running, queued, match;
2294 struct rq_flags rf;
2295 unsigned long ncsw;
2296 struct rq *rq;
2297
2298 for (;;) {
2299 /*
2300 * We do the initial early heuristics without holding
2301 * any task-queue locks at all. We'll only try to get
2302 * the runqueue lock when things look like they will
2303 * work out!
2304 */
2305 rq = task_rq(p);
2306
2307 /*
2308 * If the task is actively running on another CPU
2309 * still, just relax and busy-wait without holding
2310 * any locks.
2311 *
2312 * NOTE! Since we don't hold any locks, it's not
2313 * even sure that "rq" stays as the right runqueue!
2314 * But we don't care, since "task_on_cpu()" will
2315 * return false if the runqueue has changed and p
2316 * is actually now running somewhere else!
2317 */
2318 while (task_on_cpu(rq, p)) {
2319 if (!task_state_match(p, match_state))
2320 return 0;
2321 cpu_relax();
2322 }
2323
2324 /*
2325 * Ok, time to look more closely! We need the rq
2326 * lock now, to be *sure*. If we're wrong, we'll
2327 * just go back and repeat.
2328 */
2329 rq = task_rq_lock(p, &rf);
2330 trace_sched_wait_task(p);
2331 running = task_on_cpu(rq, p);
2332 queued = task_on_rq_queued(p);
2333 ncsw = 0;
2334 if ((match = __task_state_match(p, match_state))) {
2335 /*
2336 * When matching on p->saved_state, consider this task
2337 * still queued so it will wait.
2338 */
2339 if (match < 0)
2340 queued = 1;
2341 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2342 }
2343 task_rq_unlock(rq, p, &rf);
2344
2345 /*
2346 * If it changed from the expected state, bail out now.
2347 */
2348 if (unlikely(!ncsw))
2349 break;
2350
2351 /*
2352 * Was it really running after all now that we
2353 * checked with the proper locks actually held?
2354 *
2355 * Oops. Go back and try again..
2356 */
2357 if (unlikely(running)) {
2358 cpu_relax();
2359 continue;
2360 }
2361
2362 /*
2363 * It's not enough that it's not actively running,
2364 * it must be off the runqueue _entirely_, and not
2365 * preempted!
2366 *
2367 * So if it was still runnable (but just not actively
2368 * running right now), it's preempted, and we should
2369 * yield - it could be a while.
2370 */
2371 if (unlikely(queued)) {
2372 ktime_t to = NSEC_PER_SEC / HZ;
2373
2374 set_current_state(TASK_UNINTERRUPTIBLE);
2375 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
2376 continue;
2377 }
2378
2379 /*
2380 * Ahh, all good. It wasn't running, and it wasn't
2381 * runnable, which means that it will never become
2382 * running in the future either. We're all done!
2383 */
2384 break;
2385 }
2386
2387 return ncsw;
2388 }
2389
2390 #ifdef CONFIG_SMP
2391
2392 static void
2393 __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
2394
2395 static int __set_cpus_allowed_ptr(struct task_struct *p,
2396 struct affinity_context *ctx);
2397
migrate_disable_switch(struct rq * rq,struct task_struct * p)2398 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2399 {
2400 struct affinity_context ac = {
2401 .new_mask = cpumask_of(rq->cpu),
2402 .flags = SCA_MIGRATE_DISABLE,
2403 };
2404
2405 if (likely(!p->migration_disabled))
2406 return;
2407
2408 if (p->cpus_ptr != &p->cpus_mask)
2409 return;
2410
2411 /*
2412 * Violates locking rules! see comment in __do_set_cpus_allowed().
2413 */
2414 __do_set_cpus_allowed(p, &ac);
2415 }
2416
migrate_disable(void)2417 void migrate_disable(void)
2418 {
2419 struct task_struct *p = current;
2420
2421 if (p->migration_disabled) {
2422 p->migration_disabled++;
2423 return;
2424 }
2425
2426 guard(preempt)();
2427 this_rq()->nr_pinned++;
2428 p->migration_disabled = 1;
2429 }
2430 EXPORT_SYMBOL_GPL(migrate_disable);
2431
migrate_enable(void)2432 void migrate_enable(void)
2433 {
2434 struct task_struct *p = current;
2435 struct affinity_context ac = {
2436 .new_mask = &p->cpus_mask,
2437 .flags = SCA_MIGRATE_ENABLE,
2438 };
2439
2440 if (p->migration_disabled > 1) {
2441 p->migration_disabled--;
2442 return;
2443 }
2444
2445 if (WARN_ON_ONCE(!p->migration_disabled))
2446 return;
2447
2448 /*
2449 * Ensure stop_task runs either before or after this, and that
2450 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2451 */
2452 guard(preempt)();
2453 if (p->cpus_ptr != &p->cpus_mask)
2454 __set_cpus_allowed_ptr(p, &ac);
2455 /*
2456 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2457 * regular cpus_mask, otherwise things that race (eg.
2458 * select_fallback_rq) get confused.
2459 */
2460 barrier();
2461 p->migration_disabled = 0;
2462 this_rq()->nr_pinned--;
2463 }
2464 EXPORT_SYMBOL_GPL(migrate_enable);
2465
rq_has_pinned_tasks(struct rq * rq)2466 static inline bool rq_has_pinned_tasks(struct rq *rq)
2467 {
2468 return rq->nr_pinned;
2469 }
2470
2471 /*
2472 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2473 * __set_cpus_allowed_ptr() and select_fallback_rq().
2474 */
is_cpu_allowed(struct task_struct * p,int cpu)2475 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2476 {
2477 /* When not in the task's cpumask, no point in looking further. */
2478 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2479 return false;
2480
2481 /* migrate_disabled() must be allowed to finish. */
2482 if (is_migration_disabled(p))
2483 return cpu_online(cpu);
2484
2485 /* Non kernel threads are not allowed during either online or offline. */
2486 if (!(p->flags & PF_KTHREAD))
2487 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2488
2489 /* KTHREAD_IS_PER_CPU is always allowed. */
2490 if (kthread_is_per_cpu(p))
2491 return cpu_online(cpu);
2492
2493 /* Regular kernel threads don't get to stay during offline. */
2494 if (cpu_dying(cpu))
2495 return false;
2496
2497 /* But are allowed during online. */
2498 return cpu_online(cpu);
2499 }
2500
2501 /*
2502 * This is how migration works:
2503 *
2504 * 1) we invoke migration_cpu_stop() on the target CPU using
2505 * stop_one_cpu().
2506 * 2) stopper starts to run (implicitly forcing the migrated thread
2507 * off the CPU)
2508 * 3) it checks whether the migrated task is still in the wrong runqueue.
2509 * 4) if it's in the wrong runqueue then the migration thread removes
2510 * it and puts it into the right queue.
2511 * 5) stopper completes and stop_one_cpu() returns and the migration
2512 * is done.
2513 */
2514
2515 /*
2516 * move_queued_task - move a queued task to new rq.
2517 *
2518 * Returns (locked) new rq. Old rq's lock is released.
2519 */
move_queued_task(struct rq * rq,struct rq_flags * rf,struct task_struct * p,int new_cpu)2520 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2521 struct task_struct *p, int new_cpu)
2522 {
2523 lockdep_assert_rq_held(rq);
2524
2525 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2526 set_task_cpu(p, new_cpu);
2527 rq_unlock(rq, rf);
2528
2529 rq = cpu_rq(new_cpu);
2530
2531 rq_lock(rq, rf);
2532 WARN_ON_ONCE(task_cpu(p) != new_cpu);
2533 activate_task(rq, p, 0);
2534 wakeup_preempt(rq, p, 0);
2535
2536 return rq;
2537 }
2538
2539 struct migration_arg {
2540 struct task_struct *task;
2541 int dest_cpu;
2542 struct set_affinity_pending *pending;
2543 };
2544
2545 /*
2546 * @refs: number of wait_for_completion()
2547 * @stop_pending: is @stop_work in use
2548 */
2549 struct set_affinity_pending {
2550 refcount_t refs;
2551 unsigned int stop_pending;
2552 struct completion done;
2553 struct cpu_stop_work stop_work;
2554 struct migration_arg arg;
2555 };
2556
2557 /*
2558 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2559 * this because either it can't run here any more (set_cpus_allowed()
2560 * away from this CPU, or CPU going down), or because we're
2561 * attempting to rebalance this task on exec (sched_exec).
2562 *
2563 * So we race with normal scheduler movements, but that's OK, as long
2564 * as the task is no longer on this CPU.
2565 */
__migrate_task(struct rq * rq,struct rq_flags * rf,struct task_struct * p,int dest_cpu)2566 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2567 struct task_struct *p, int dest_cpu)
2568 {
2569 /* Affinity changed (again). */
2570 if (!is_cpu_allowed(p, dest_cpu))
2571 return rq;
2572
2573 rq = move_queued_task(rq, rf, p, dest_cpu);
2574
2575 return rq;
2576 }
2577
2578 /*
2579 * migration_cpu_stop - this will be executed by a highprio stopper thread
2580 * and performs thread migration by bumping thread off CPU then
2581 * 'pushing' onto another runqueue.
2582 */
migration_cpu_stop(void * data)2583 static int migration_cpu_stop(void *data)
2584 {
2585 struct migration_arg *arg = data;
2586 struct set_affinity_pending *pending = arg->pending;
2587 struct task_struct *p = arg->task;
2588 struct rq *rq = this_rq();
2589 bool complete = false;
2590 struct rq_flags rf;
2591
2592 /*
2593 * The original target CPU might have gone down and we might
2594 * be on another CPU but it doesn't matter.
2595 */
2596 local_irq_save(rf.flags);
2597 /*
2598 * We need to explicitly wake pending tasks before running
2599 * __migrate_task() such that we will not miss enforcing cpus_ptr
2600 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2601 */
2602 flush_smp_call_function_queue();
2603
2604 raw_spin_lock(&p->pi_lock);
2605 rq_lock(rq, &rf);
2606
2607 /*
2608 * If we were passed a pending, then ->stop_pending was set, thus
2609 * p->migration_pending must have remained stable.
2610 */
2611 WARN_ON_ONCE(pending && pending != p->migration_pending);
2612
2613 /*
2614 * If task_rq(p) != rq, it cannot be migrated here, because we're
2615 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2616 * we're holding p->pi_lock.
2617 */
2618 if (task_rq(p) == rq) {
2619 if (is_migration_disabled(p))
2620 goto out;
2621
2622 if (pending) {
2623 p->migration_pending = NULL;
2624 complete = true;
2625
2626 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2627 goto out;
2628 }
2629
2630 if (task_on_rq_queued(p)) {
2631 update_rq_clock(rq);
2632 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2633 } else {
2634 p->wake_cpu = arg->dest_cpu;
2635 }
2636
2637 /*
2638 * XXX __migrate_task() can fail, at which point we might end
2639 * up running on a dodgy CPU, AFAICT this can only happen
2640 * during CPU hotplug, at which point we'll get pushed out
2641 * anyway, so it's probably not a big deal.
2642 */
2643
2644 } else if (pending) {
2645 /*
2646 * This happens when we get migrated between migrate_enable()'s
2647 * preempt_enable() and scheduling the stopper task. At that
2648 * point we're a regular task again and not current anymore.
2649 *
2650 * A !PREEMPT kernel has a giant hole here, which makes it far
2651 * more likely.
2652 */
2653
2654 /*
2655 * The task moved before the stopper got to run. We're holding
2656 * ->pi_lock, so the allowed mask is stable - if it got
2657 * somewhere allowed, we're done.
2658 */
2659 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2660 p->migration_pending = NULL;
2661 complete = true;
2662 goto out;
2663 }
2664
2665 /*
2666 * When migrate_enable() hits a rq mis-match we can't reliably
2667 * determine is_migration_disabled() and so have to chase after
2668 * it.
2669 */
2670 WARN_ON_ONCE(!pending->stop_pending);
2671 preempt_disable();
2672 task_rq_unlock(rq, p, &rf);
2673 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2674 &pending->arg, &pending->stop_work);
2675 preempt_enable();
2676 return 0;
2677 }
2678 out:
2679 if (pending)
2680 pending->stop_pending = false;
2681 task_rq_unlock(rq, p, &rf);
2682
2683 if (complete)
2684 complete_all(&pending->done);
2685
2686 return 0;
2687 }
2688
push_cpu_stop(void * arg)2689 int push_cpu_stop(void *arg)
2690 {
2691 struct rq *lowest_rq = NULL, *rq = this_rq();
2692 struct task_struct *p = arg;
2693
2694 raw_spin_lock_irq(&p->pi_lock);
2695 raw_spin_rq_lock(rq);
2696
2697 if (task_rq(p) != rq)
2698 goto out_unlock;
2699
2700 if (is_migration_disabled(p)) {
2701 p->migration_flags |= MDF_PUSH;
2702 goto out_unlock;
2703 }
2704
2705 p->migration_flags &= ~MDF_PUSH;
2706
2707 if (p->sched_class->find_lock_rq)
2708 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2709
2710 if (!lowest_rq)
2711 goto out_unlock;
2712
2713 // XXX validate p is still the highest prio task
2714 if (task_rq(p) == rq) {
2715 deactivate_task(rq, p, 0);
2716 set_task_cpu(p, lowest_rq->cpu);
2717 activate_task(lowest_rq, p, 0);
2718 resched_curr(lowest_rq);
2719 }
2720
2721 double_unlock_balance(rq, lowest_rq);
2722
2723 out_unlock:
2724 rq->push_busy = false;
2725 raw_spin_rq_unlock(rq);
2726 raw_spin_unlock_irq(&p->pi_lock);
2727
2728 put_task_struct(p);
2729 return 0;
2730 }
2731
2732 /*
2733 * sched_class::set_cpus_allowed must do the below, but is not required to
2734 * actually call this function.
2735 */
set_cpus_allowed_common(struct task_struct * p,struct affinity_context * ctx)2736 void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
2737 {
2738 if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2739 p->cpus_ptr = ctx->new_mask;
2740 return;
2741 }
2742
2743 cpumask_copy(&p->cpus_mask, ctx->new_mask);
2744 p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
2745
2746 /*
2747 * Swap in a new user_cpus_ptr if SCA_USER flag set
2748 */
2749 if (ctx->flags & SCA_USER)
2750 swap(p->user_cpus_ptr, ctx->user_mask);
2751 }
2752
2753 static void
__do_set_cpus_allowed(struct task_struct * p,struct affinity_context * ctx)2754 __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
2755 {
2756 struct rq *rq = task_rq(p);
2757 bool queued, running;
2758
2759 /*
2760 * This here violates the locking rules for affinity, since we're only
2761 * supposed to change these variables while holding both rq->lock and
2762 * p->pi_lock.
2763 *
2764 * HOWEVER, it magically works, because ttwu() is the only code that
2765 * accesses these variables under p->pi_lock and only does so after
2766 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2767 * before finish_task().
2768 *
2769 * XXX do further audits, this smells like something putrid.
2770 */
2771 if (ctx->flags & SCA_MIGRATE_DISABLE)
2772 SCHED_WARN_ON(!p->on_cpu);
2773 else
2774 lockdep_assert_held(&p->pi_lock);
2775
2776 queued = task_on_rq_queued(p);
2777 running = task_current(rq, p);
2778
2779 if (queued) {
2780 /*
2781 * Because __kthread_bind() calls this on blocked tasks without
2782 * holding rq->lock.
2783 */
2784 lockdep_assert_rq_held(rq);
2785 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2786 }
2787 if (running)
2788 put_prev_task(rq, p);
2789
2790 p->sched_class->set_cpus_allowed(p, ctx);
2791
2792 if (queued)
2793 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2794 if (running)
2795 set_next_task(rq, p);
2796 }
2797
2798 /*
2799 * Used for kthread_bind() and select_fallback_rq(), in both cases the user
2800 * affinity (if any) should be destroyed too.
2801 */
do_set_cpus_allowed(struct task_struct * p,const struct cpumask * new_mask)2802 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2803 {
2804 struct affinity_context ac = {
2805 .new_mask = new_mask,
2806 .user_mask = NULL,
2807 .flags = SCA_USER, /* clear the user requested mask */
2808 };
2809 union cpumask_rcuhead {
2810 cpumask_t cpumask;
2811 struct rcu_head rcu;
2812 };
2813
2814 __do_set_cpus_allowed(p, &ac);
2815
2816 /*
2817 * Because this is called with p->pi_lock held, it is not possible
2818 * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2819 * kfree_rcu().
2820 */
2821 kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
2822 }
2823
alloc_user_cpus_ptr(int node)2824 static cpumask_t *alloc_user_cpus_ptr(int node)
2825 {
2826 /*
2827 * See do_set_cpus_allowed() above for the rcu_head usage.
2828 */
2829 int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2830
2831 return kmalloc_node(size, GFP_KERNEL, node);
2832 }
2833
dup_user_cpus_ptr(struct task_struct * dst,struct task_struct * src,int node)2834 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2835 int node)
2836 {
2837 cpumask_t *user_mask;
2838 unsigned long flags;
2839
2840 /*
2841 * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2842 * may differ by now due to racing.
2843 */
2844 dst->user_cpus_ptr = NULL;
2845
2846 /*
2847 * This check is racy and losing the race is a valid situation.
2848 * It is not worth the extra overhead of taking the pi_lock on
2849 * every fork/clone.
2850 */
2851 if (data_race(!src->user_cpus_ptr))
2852 return 0;
2853
2854 user_mask = alloc_user_cpus_ptr(node);
2855 if (!user_mask)
2856 return -ENOMEM;
2857
2858 /*
2859 * Use pi_lock to protect content of user_cpus_ptr
2860 *
2861 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2862 * do_set_cpus_allowed().
2863 */
2864 raw_spin_lock_irqsave(&src->pi_lock, flags);
2865 if (src->user_cpus_ptr) {
2866 swap(dst->user_cpus_ptr, user_mask);
2867 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2868 }
2869 raw_spin_unlock_irqrestore(&src->pi_lock, flags);
2870
2871 if (unlikely(user_mask))
2872 kfree(user_mask);
2873
2874 return 0;
2875 }
2876
clear_user_cpus_ptr(struct task_struct * p)2877 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2878 {
2879 struct cpumask *user_mask = NULL;
2880
2881 swap(p->user_cpus_ptr, user_mask);
2882
2883 return user_mask;
2884 }
2885
release_user_cpus_ptr(struct task_struct * p)2886 void release_user_cpus_ptr(struct task_struct *p)
2887 {
2888 kfree(clear_user_cpus_ptr(p));
2889 }
2890
2891 /*
2892 * This function is wildly self concurrent; here be dragons.
2893 *
2894 *
2895 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2896 * designated task is enqueued on an allowed CPU. If that task is currently
2897 * running, we have to kick it out using the CPU stopper.
2898 *
2899 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2900 * Consider:
2901 *
2902 * Initial conditions: P0->cpus_mask = [0, 1]
2903 *
2904 * P0@CPU0 P1
2905 *
2906 * migrate_disable();
2907 * <preempted>
2908 * set_cpus_allowed_ptr(P0, [1]);
2909 *
2910 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2911 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2912 * This means we need the following scheme:
2913 *
2914 * P0@CPU0 P1
2915 *
2916 * migrate_disable();
2917 * <preempted>
2918 * set_cpus_allowed_ptr(P0, [1]);
2919 * <blocks>
2920 * <resumes>
2921 * migrate_enable();
2922 * __set_cpus_allowed_ptr();
2923 * <wakes local stopper>
2924 * `--> <woken on migration completion>
2925 *
2926 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2927 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2928 * task p are serialized by p->pi_lock, which we can leverage: the one that
2929 * should come into effect at the end of the Migrate-Disable region is the last
2930 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2931 * but we still need to properly signal those waiting tasks at the appropriate
2932 * moment.
2933 *
2934 * This is implemented using struct set_affinity_pending. The first
2935 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2936 * setup an instance of that struct and install it on the targeted task_struct.
2937 * Any and all further callers will reuse that instance. Those then wait for
2938 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2939 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2940 *
2941 *
2942 * (1) In the cases covered above. There is one more where the completion is
2943 * signaled within affine_move_task() itself: when a subsequent affinity request
2944 * occurs after the stopper bailed out due to the targeted task still being
2945 * Migrate-Disable. Consider:
2946 *
2947 * Initial conditions: P0->cpus_mask = [0, 1]
2948 *
2949 * CPU0 P1 P2
2950 * <P0>
2951 * migrate_disable();
2952 * <preempted>
2953 * set_cpus_allowed_ptr(P0, [1]);
2954 * <blocks>
2955 * <migration/0>
2956 * migration_cpu_stop()
2957 * is_migration_disabled()
2958 * <bails>
2959 * set_cpus_allowed_ptr(P0, [0, 1]);
2960 * <signal completion>
2961 * <awakes>
2962 *
2963 * Note that the above is safe vs a concurrent migrate_enable(), as any
2964 * pending affinity completion is preceded by an uninstallation of
2965 * p->migration_pending done with p->pi_lock held.
2966 */
affine_move_task(struct rq * rq,struct task_struct * p,struct rq_flags * rf,int dest_cpu,unsigned int flags)2967 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2968 int dest_cpu, unsigned int flags)
2969 __releases(rq->lock)
2970 __releases(p->pi_lock)
2971 {
2972 struct set_affinity_pending my_pending = { }, *pending = NULL;
2973 bool stop_pending, complete = false;
2974
2975 /* Can the task run on the task's current CPU? If so, we're done */
2976 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2977 struct task_struct *push_task = NULL;
2978
2979 if ((flags & SCA_MIGRATE_ENABLE) &&
2980 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2981 rq->push_busy = true;
2982 push_task = get_task_struct(p);
2983 }
2984
2985 /*
2986 * If there are pending waiters, but no pending stop_work,
2987 * then complete now.
2988 */
2989 pending = p->migration_pending;
2990 if (pending && !pending->stop_pending) {
2991 p->migration_pending = NULL;
2992 complete = true;
2993 }
2994
2995 preempt_disable();
2996 task_rq_unlock(rq, p, rf);
2997 if (push_task) {
2998 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2999 p, &rq->push_work);
3000 }
3001 preempt_enable();
3002
3003 if (complete)
3004 complete_all(&pending->done);
3005
3006 return 0;
3007 }
3008
3009 if (!(flags & SCA_MIGRATE_ENABLE)) {
3010 /* serialized by p->pi_lock */
3011 if (!p->migration_pending) {
3012 /* Install the request */
3013 refcount_set(&my_pending.refs, 1);
3014 init_completion(&my_pending.done);
3015 my_pending.arg = (struct migration_arg) {
3016 .task = p,
3017 .dest_cpu = dest_cpu,
3018 .pending = &my_pending,
3019 };
3020
3021 p->migration_pending = &my_pending;
3022 } else {
3023 pending = p->migration_pending;
3024 refcount_inc(&pending->refs);
3025 /*
3026 * Affinity has changed, but we've already installed a
3027 * pending. migration_cpu_stop() *must* see this, else
3028 * we risk a completion of the pending despite having a
3029 * task on a disallowed CPU.
3030 *
3031 * Serialized by p->pi_lock, so this is safe.
3032 */
3033 pending->arg.dest_cpu = dest_cpu;
3034 }
3035 }
3036 pending = p->migration_pending;
3037 /*
3038 * - !MIGRATE_ENABLE:
3039 * we'll have installed a pending if there wasn't one already.
3040 *
3041 * - MIGRATE_ENABLE:
3042 * we're here because the current CPU isn't matching anymore,
3043 * the only way that can happen is because of a concurrent
3044 * set_cpus_allowed_ptr() call, which should then still be
3045 * pending completion.
3046 *
3047 * Either way, we really should have a @pending here.
3048 */
3049 if (WARN_ON_ONCE(!pending)) {
3050 task_rq_unlock(rq, p, rf);
3051 return -EINVAL;
3052 }
3053
3054 if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
3055 /*
3056 * MIGRATE_ENABLE gets here because 'p == current', but for
3057 * anything else we cannot do is_migration_disabled(), punt
3058 * and have the stopper function handle it all race-free.
3059 */
3060 stop_pending = pending->stop_pending;
3061 if (!stop_pending)
3062 pending->stop_pending = true;
3063
3064 if (flags & SCA_MIGRATE_ENABLE)
3065 p->migration_flags &= ~MDF_PUSH;
3066
3067 preempt_disable();
3068 task_rq_unlock(rq, p, rf);
3069 if (!stop_pending) {
3070 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
3071 &pending->arg, &pending->stop_work);
3072 }
3073 preempt_enable();
3074
3075 if (flags & SCA_MIGRATE_ENABLE)
3076 return 0;
3077 } else {
3078
3079 if (!is_migration_disabled(p)) {
3080 if (task_on_rq_queued(p))
3081 rq = move_queued_task(rq, rf, p, dest_cpu);
3082
3083 if (!pending->stop_pending) {
3084 p->migration_pending = NULL;
3085 complete = true;
3086 }
3087 }
3088 task_rq_unlock(rq, p, rf);
3089
3090 if (complete)
3091 complete_all(&pending->done);
3092 }
3093
3094 wait_for_completion(&pending->done);
3095
3096 if (refcount_dec_and_test(&pending->refs))
3097 wake_up_var(&pending->refs); /* No UaF, just an address */
3098
3099 /*
3100 * Block the original owner of &pending until all subsequent callers
3101 * have seen the completion and decremented the refcount
3102 */
3103 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
3104
3105 /* ARGH */
3106 WARN_ON_ONCE(my_pending.stop_pending);
3107
3108 return 0;
3109 }
3110
3111 /*
3112 * Called with both p->pi_lock and rq->lock held; drops both before returning.
3113 */
__set_cpus_allowed_ptr_locked(struct task_struct * p,struct affinity_context * ctx,struct rq * rq,struct rq_flags * rf)3114 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
3115 struct affinity_context *ctx,
3116 struct rq *rq,
3117 struct rq_flags *rf)
3118 __releases(rq->lock)
3119 __releases(p->pi_lock)
3120 {
3121 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
3122 const struct cpumask *cpu_valid_mask = cpu_active_mask;
3123 bool kthread = p->flags & PF_KTHREAD;
3124 unsigned int dest_cpu;
3125 int ret = 0;
3126
3127 update_rq_clock(rq);
3128
3129 if (kthread || is_migration_disabled(p)) {
3130 /*
3131 * Kernel threads are allowed on online && !active CPUs,
3132 * however, during cpu-hot-unplug, even these might get pushed
3133 * away if not KTHREAD_IS_PER_CPU.
3134 *
3135 * Specifically, migration_disabled() tasks must not fail the
3136 * cpumask_any_and_distribute() pick below, esp. so on
3137 * SCA_MIGRATE_ENABLE, otherwise we'll not call
3138 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
3139 */
3140 cpu_valid_mask = cpu_online_mask;
3141 }
3142
3143 if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
3144 ret = -EINVAL;
3145 goto out;
3146 }
3147
3148 /*
3149 * Must re-check here, to close a race against __kthread_bind(),
3150 * sched_setaffinity() is not guaranteed to observe the flag.
3151 */
3152 if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
3153 ret = -EINVAL;
3154 goto out;
3155 }
3156
3157 if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
3158 if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) {
3159 if (ctx->flags & SCA_USER)
3160 swap(p->user_cpus_ptr, ctx->user_mask);
3161 goto out;
3162 }
3163
3164 if (WARN_ON_ONCE(p == current &&
3165 is_migration_disabled(p) &&
3166 !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
3167 ret = -EBUSY;
3168 goto out;
3169 }
3170 }
3171
3172 /*
3173 * Picking a ~random cpu helps in cases where we are changing affinity
3174 * for groups of tasks (ie. cpuset), so that load balancing is not
3175 * immediately required to distribute the tasks within their new mask.
3176 */
3177 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
3178 if (dest_cpu >= nr_cpu_ids) {
3179 ret = -EINVAL;
3180 goto out;
3181 }
3182
3183 __do_set_cpus_allowed(p, ctx);
3184
3185 return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
3186
3187 out:
3188 task_rq_unlock(rq, p, rf);
3189
3190 return ret;
3191 }
3192
3193 /*
3194 * Change a given task's CPU affinity. Migrate the thread to a
3195 * proper CPU and schedule it away if the CPU it's executing on
3196 * is removed from the allowed bitmask.
3197 *
3198 * NOTE: the caller must have a valid reference to the task, the
3199 * task must not exit() & deallocate itself prematurely. The
3200 * call is not atomic; no spinlocks may be held.
3201 */
__set_cpus_allowed_ptr(struct task_struct * p,struct affinity_context * ctx)3202 static int __set_cpus_allowed_ptr(struct task_struct *p,
3203 struct affinity_context *ctx)
3204 {
3205 struct rq_flags rf;
3206 struct rq *rq;
3207
3208 rq = task_rq_lock(p, &rf);
3209 /*
3210 * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3211 * flags are set.
3212 */
3213 if (p->user_cpus_ptr &&
3214 !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
3215 cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
3216 ctx->new_mask = rq->scratch_mask;
3217
3218 return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
3219 }
3220
set_cpus_allowed_ptr(struct task_struct * p,const struct cpumask * new_mask)3221 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3222 {
3223 struct affinity_context ac = {
3224 .new_mask = new_mask,
3225 .flags = 0,
3226 };
3227
3228 return __set_cpus_allowed_ptr(p, &ac);
3229 }
3230 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
3231
3232 /*
3233 * Change a given task's CPU affinity to the intersection of its current
3234 * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3235 * If user_cpus_ptr is defined, use it as the basis for restricting CPU
3236 * affinity or use cpu_online_mask instead.
3237 *
3238 * If the resulting mask is empty, leave the affinity unchanged and return
3239 * -EINVAL.
3240 */
restrict_cpus_allowed_ptr(struct task_struct * p,struct cpumask * new_mask,const struct cpumask * subset_mask)3241 static int restrict_cpus_allowed_ptr(struct task_struct *p,
3242 struct cpumask *new_mask,
3243 const struct cpumask *subset_mask)
3244 {
3245 struct affinity_context ac = {
3246 .new_mask = new_mask,
3247 .flags = 0,
3248 };
3249 struct rq_flags rf;
3250 struct rq *rq;
3251 int err;
3252
3253 rq = task_rq_lock(p, &rf);
3254
3255 /*
3256 * Forcefully restricting the affinity of a deadline task is
3257 * likely to cause problems, so fail and noisily override the
3258 * mask entirely.
3259 */
3260 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
3261 err = -EPERM;
3262 goto err_unlock;
3263 }
3264
3265 if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
3266 err = -EINVAL;
3267 goto err_unlock;
3268 }
3269
3270 return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
3271
3272 err_unlock:
3273 task_rq_unlock(rq, p, &rf);
3274 return err;
3275 }
3276
3277 /*
3278 * Restrict the CPU affinity of task @p so that it is a subset of
3279 * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
3280 * old affinity mask. If the resulting mask is empty, we warn and walk
3281 * up the cpuset hierarchy until we find a suitable mask.
3282 */
force_compatible_cpus_allowed_ptr(struct task_struct * p)3283 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3284 {
3285 cpumask_var_t new_mask;
3286 const struct cpumask *override_mask = task_cpu_possible_mask(p);
3287
3288 alloc_cpumask_var(&new_mask, GFP_KERNEL);
3289
3290 /*
3291 * __migrate_task() can fail silently in the face of concurrent
3292 * offlining of the chosen destination CPU, so take the hotplug
3293 * lock to ensure that the migration succeeds.
3294 */
3295 cpus_read_lock();
3296 if (!cpumask_available(new_mask))
3297 goto out_set_mask;
3298
3299 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3300 goto out_free_mask;
3301
3302 /*
3303 * We failed to find a valid subset of the affinity mask for the
3304 * task, so override it based on its cpuset hierarchy.
3305 */
3306 cpuset_cpus_allowed(p, new_mask);
3307 override_mask = new_mask;
3308
3309 out_set_mask:
3310 if (printk_ratelimit()) {
3311 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3312 task_pid_nr(p), p->comm,
3313 cpumask_pr_args(override_mask));
3314 }
3315
3316 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3317 out_free_mask:
3318 cpus_read_unlock();
3319 free_cpumask_var(new_mask);
3320 }
3321
3322 static int
3323 __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
3324
3325 /*
3326 * Restore the affinity of a task @p which was previously restricted by a
3327 * call to force_compatible_cpus_allowed_ptr().
3328 *
3329 * It is the caller's responsibility to serialise this with any calls to
3330 * force_compatible_cpus_allowed_ptr(@p).
3331 */
relax_compatible_cpus_allowed_ptr(struct task_struct * p)3332 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3333 {
3334 struct affinity_context ac = {
3335 .new_mask = task_user_cpus(p),
3336 .flags = 0,
3337 };
3338 int ret;
3339
3340 /*
3341 * Try to restore the old affinity mask with __sched_setaffinity().
3342 * Cpuset masking will be done there too.
3343 */
3344 ret = __sched_setaffinity(p, &ac);
3345 WARN_ON_ONCE(ret);
3346 }
3347
set_task_cpu(struct task_struct * p,unsigned int new_cpu)3348 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3349 {
3350 #ifdef CONFIG_SCHED_DEBUG
3351 unsigned int state = READ_ONCE(p->__state);
3352
3353 /*
3354 * We should never call set_task_cpu() on a blocked task,
3355 * ttwu() will sort out the placement.
3356 */
3357 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3358
3359 /*
3360 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3361 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3362 * time relying on p->on_rq.
3363 */
3364 WARN_ON_ONCE(state == TASK_RUNNING &&
3365 p->sched_class == &fair_sched_class &&
3366 (p->on_rq && !task_on_rq_migrating(p)));
3367
3368 #ifdef CONFIG_LOCKDEP
3369 /*
3370 * The caller should hold either p->pi_lock or rq->lock, when changing
3371 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3372 *
3373 * sched_move_task() holds both and thus holding either pins the cgroup,
3374 * see task_group().
3375 *
3376 * Furthermore, all task_rq users should acquire both locks, see
3377 * task_rq_lock().
3378 */
3379 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3380 lockdep_is_held(__rq_lockp(task_rq(p)))));
3381 #endif
3382 /*
3383 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3384 */
3385 WARN_ON_ONCE(!cpu_online(new_cpu));
3386
3387 WARN_ON_ONCE(is_migration_disabled(p));
3388 #endif
3389
3390 trace_sched_migrate_task(p, new_cpu);
3391
3392 if (task_cpu(p) != new_cpu) {
3393 if (p->sched_class->migrate_task_rq)
3394 p->sched_class->migrate_task_rq(p, new_cpu);
3395 p->se.nr_migrations++;
3396 rseq_migrate(p);
3397 sched_mm_cid_migrate_from(p);
3398 perf_event_task_migrate(p);
3399 }
3400
3401 __set_task_cpu(p, new_cpu);
3402 }
3403
3404 #ifdef CONFIG_NUMA_BALANCING
__migrate_swap_task(struct task_struct * p,int cpu)3405 static void __migrate_swap_task(struct task_struct *p, int cpu)
3406 {
3407 if (task_on_rq_queued(p)) {
3408 struct rq *src_rq, *dst_rq;
3409 struct rq_flags srf, drf;
3410
3411 src_rq = task_rq(p);
3412 dst_rq = cpu_rq(cpu);
3413
3414 rq_pin_lock(src_rq, &srf);
3415 rq_pin_lock(dst_rq, &drf);
3416
3417 deactivate_task(src_rq, p, 0);
3418 set_task_cpu(p, cpu);
3419 activate_task(dst_rq, p, 0);
3420 wakeup_preempt(dst_rq, p, 0);
3421
3422 rq_unpin_lock(dst_rq, &drf);
3423 rq_unpin_lock(src_rq, &srf);
3424
3425 } else {
3426 /*
3427 * Task isn't running anymore; make it appear like we migrated
3428 * it before it went to sleep. This means on wakeup we make the
3429 * previous CPU our target instead of where it really is.
3430 */
3431 p->wake_cpu = cpu;
3432 }
3433 }
3434
3435 struct migration_swap_arg {
3436 struct task_struct *src_task, *dst_task;
3437 int src_cpu, dst_cpu;
3438 };
3439
migrate_swap_stop(void * data)3440 static int migrate_swap_stop(void *data)
3441 {
3442 struct migration_swap_arg *arg = data;
3443 struct rq *src_rq, *dst_rq;
3444
3445 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3446 return -EAGAIN;
3447
3448 src_rq = cpu_rq(arg->src_cpu);
3449 dst_rq = cpu_rq(arg->dst_cpu);
3450
3451 guard(double_raw_spinlock)(&arg->src_task->pi_lock, &arg->dst_task->pi_lock);
3452 guard(double_rq_lock)(src_rq, dst_rq);
3453
3454 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3455 return -EAGAIN;
3456
3457 if (task_cpu(arg->src_task) != arg->src_cpu)
3458 return -EAGAIN;
3459
3460 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3461 return -EAGAIN;
3462
3463 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3464 return -EAGAIN;
3465
3466 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3467 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3468
3469 return 0;
3470 }
3471
3472 /*
3473 * Cross migrate two tasks
3474 */
migrate_swap(struct task_struct * cur,struct task_struct * p,int target_cpu,int curr_cpu)3475 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3476 int target_cpu, int curr_cpu)
3477 {
3478 struct migration_swap_arg arg;
3479 int ret = -EINVAL;
3480
3481 arg = (struct migration_swap_arg){
3482 .src_task = cur,
3483 .src_cpu = curr_cpu,
3484 .dst_task = p,
3485 .dst_cpu = target_cpu,
3486 };
3487
3488 if (arg.src_cpu == arg.dst_cpu)
3489 goto out;
3490
3491 /*
3492 * These three tests are all lockless; this is OK since all of them
3493 * will be re-checked with proper locks held further down the line.
3494 */
3495 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3496 goto out;
3497
3498 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3499 goto out;
3500
3501 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3502 goto out;
3503
3504 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3505 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3506
3507 out:
3508 return ret;
3509 }
3510 #endif /* CONFIG_NUMA_BALANCING */
3511
3512 /***
3513 * kick_process - kick a running thread to enter/exit the kernel
3514 * @p: the to-be-kicked thread
3515 *
3516 * Cause a process which is running on another CPU to enter
3517 * kernel-mode, without any delay. (to get signals handled.)
3518 *
3519 * NOTE: this function doesn't have to take the runqueue lock,
3520 * because all it wants to ensure is that the remote task enters
3521 * the kernel. If the IPI races and the task has been migrated
3522 * to another CPU then no harm is done and the purpose has been
3523 * achieved as well.
3524 */
kick_process(struct task_struct * p)3525 void kick_process(struct task_struct *p)
3526 {
3527 guard(preempt)();
3528 int cpu = task_cpu(p);
3529
3530 if ((cpu != smp_processor_id()) && task_curr(p))
3531 smp_send_reschedule(cpu);
3532 }
3533 EXPORT_SYMBOL_GPL(kick_process);
3534
3535 /*
3536 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3537 *
3538 * A few notes on cpu_active vs cpu_online:
3539 *
3540 * - cpu_active must be a subset of cpu_online
3541 *
3542 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3543 * see __set_cpus_allowed_ptr(). At this point the newly online
3544 * CPU isn't yet part of the sched domains, and balancing will not
3545 * see it.
3546 *
3547 * - on CPU-down we clear cpu_active() to mask the sched domains and
3548 * avoid the load balancer to place new tasks on the to be removed
3549 * CPU. Existing tasks will remain running there and will be taken
3550 * off.
3551 *
3552 * This means that fallback selection must not select !active CPUs.
3553 * And can assume that any active CPU must be online. Conversely
3554 * select_task_rq() below may allow selection of !active CPUs in order
3555 * to satisfy the above rules.
3556 */
select_fallback_rq(int cpu,struct task_struct * p)3557 static int select_fallback_rq(int cpu, struct task_struct *p)
3558 {
3559 int nid = cpu_to_node(cpu);
3560 const struct cpumask *nodemask = NULL;
3561 enum { cpuset, possible, fail } state = cpuset;
3562 int dest_cpu;
3563
3564 /*
3565 * If the node that the CPU is on has been offlined, cpu_to_node()
3566 * will return -1. There is no CPU on the node, and we should
3567 * select the CPU on the other node.
3568 */
3569 if (nid != -1) {
3570 nodemask = cpumask_of_node(nid);
3571
3572 /* Look for allowed, online CPU in same node. */
3573 for_each_cpu(dest_cpu, nodemask) {
3574 if (is_cpu_allowed(p, dest_cpu))
3575 return dest_cpu;
3576 }
3577 }
3578
3579 for (;;) {
3580 /* Any allowed, online CPU? */
3581 for_each_cpu(dest_cpu, p->cpus_ptr) {
3582 if (!is_cpu_allowed(p, dest_cpu))
3583 continue;
3584
3585 goto out;
3586 }
3587
3588 /* No more Mr. Nice Guy. */
3589 switch (state) {
3590 case cpuset:
3591 if (cpuset_cpus_allowed_fallback(p)) {
3592 state = possible;
3593 break;
3594 }
3595 fallthrough;
3596 case possible:
3597 /*
3598 * XXX When called from select_task_rq() we only
3599 * hold p->pi_lock and again violate locking order.
3600 *
3601 * More yuck to audit.
3602 */
3603 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3604 state = fail;
3605 break;
3606 case fail:
3607 BUG();
3608 break;
3609 }
3610 }
3611
3612 out:
3613 if (state != cpuset) {
3614 /*
3615 * Don't tell them about moving exiting tasks or
3616 * kernel threads (both mm NULL), since they never
3617 * leave kernel.
3618 */
3619 if (p->mm && printk_ratelimit()) {
3620 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3621 task_pid_nr(p), p->comm, cpu);
3622 }
3623 }
3624
3625 return dest_cpu;
3626 }
3627
3628 /*
3629 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3630 */
3631 static inline
select_task_rq(struct task_struct * p,int cpu,int wake_flags)3632 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3633 {
3634 lockdep_assert_held(&p->pi_lock);
3635
3636 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3637 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3638 else
3639 cpu = cpumask_any(p->cpus_ptr);
3640
3641 /*
3642 * In order not to call set_task_cpu() on a blocking task we need
3643 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3644 * CPU.
3645 *
3646 * Since this is common to all placement strategies, this lives here.
3647 *
3648 * [ this allows ->select_task() to simply return task_cpu(p) and
3649 * not worry about this generic constraint ]
3650 */
3651 if (unlikely(!is_cpu_allowed(p, cpu)))
3652 cpu = select_fallback_rq(task_cpu(p), p);
3653
3654 return cpu;
3655 }
3656
sched_set_stop_task(int cpu,struct task_struct * stop)3657 void sched_set_stop_task(int cpu, struct task_struct *stop)
3658 {
3659 static struct lock_class_key stop_pi_lock;
3660 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3661 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3662
3663 if (stop) {
3664 /*
3665 * Make it appear like a SCHED_FIFO task, its something
3666 * userspace knows about and won't get confused about.
3667 *
3668 * Also, it will make PI more or less work without too
3669 * much confusion -- but then, stop work should not
3670 * rely on PI working anyway.
3671 */
3672 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3673
3674 stop->sched_class = &stop_sched_class;
3675
3676 /*
3677 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3678 * adjust the effective priority of a task. As a result,
3679 * rt_mutex_setprio() can trigger (RT) balancing operations,
3680 * which can then trigger wakeups of the stop thread to push
3681 * around the current task.
3682 *
3683 * The stop task itself will never be part of the PI-chain, it
3684 * never blocks, therefore that ->pi_lock recursion is safe.
3685 * Tell lockdep about this by placing the stop->pi_lock in its
3686 * own class.
3687 */
3688 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3689 }
3690
3691 cpu_rq(cpu)->stop = stop;
3692
3693 if (old_stop) {
3694 /*
3695 * Reset it back to a normal scheduling class so that
3696 * it can die in pieces.
3697 */
3698 old_stop->sched_class = &rt_sched_class;
3699 }
3700 }
3701
3702 #else /* CONFIG_SMP */
3703
__set_cpus_allowed_ptr(struct task_struct * p,struct affinity_context * ctx)3704 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3705 struct affinity_context *ctx)
3706 {
3707 return set_cpus_allowed_ptr(p, ctx->new_mask);
3708 }
3709
migrate_disable_switch(struct rq * rq,struct task_struct * p)3710 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3711
rq_has_pinned_tasks(struct rq * rq)3712 static inline bool rq_has_pinned_tasks(struct rq *rq)
3713 {
3714 return false;
3715 }
3716
alloc_user_cpus_ptr(int node)3717 static inline cpumask_t *alloc_user_cpus_ptr(int node)
3718 {
3719 return NULL;
3720 }
3721
3722 #endif /* !CONFIG_SMP */
3723
3724 static void
ttwu_stat(struct task_struct * p,int cpu,int wake_flags)3725 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3726 {
3727 struct rq *rq;
3728
3729 if (!schedstat_enabled())
3730 return;
3731
3732 rq = this_rq();
3733
3734 #ifdef CONFIG_SMP
3735 if (cpu == rq->cpu) {
3736 __schedstat_inc(rq->ttwu_local);
3737 __schedstat_inc(p->stats.nr_wakeups_local);
3738 } else {
3739 struct sched_domain *sd;
3740
3741 __schedstat_inc(p->stats.nr_wakeups_remote);
3742
3743 guard(rcu)();
3744 for_each_domain(rq->cpu, sd) {
3745 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3746 __schedstat_inc(sd->ttwu_wake_remote);
3747 break;
3748 }
3749 }
3750 }
3751
3752 if (wake_flags & WF_MIGRATED)
3753 __schedstat_inc(p->stats.nr_wakeups_migrate);
3754 #endif /* CONFIG_SMP */
3755
3756 __schedstat_inc(rq->ttwu_count);
3757 __schedstat_inc(p->stats.nr_wakeups);
3758
3759 if (wake_flags & WF_SYNC)
3760 __schedstat_inc(p->stats.nr_wakeups_sync);
3761 }
3762
3763 /*
3764 * Mark the task runnable.
3765 */
ttwu_do_wakeup(struct task_struct * p)3766 static inline void ttwu_do_wakeup(struct task_struct *p)
3767 {
3768 WRITE_ONCE(p->__state, TASK_RUNNING);
3769 trace_sched_wakeup(p);
3770 }
3771
3772 static void
ttwu_do_activate(struct rq * rq,struct task_struct * p,int wake_flags,struct rq_flags * rf)3773 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3774 struct rq_flags *rf)
3775 {
3776 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3777
3778 lockdep_assert_rq_held(rq);
3779
3780 if (p->sched_contributes_to_load)
3781 rq->nr_uninterruptible--;
3782
3783 #ifdef CONFIG_SMP
3784 if (wake_flags & WF_MIGRATED)
3785 en_flags |= ENQUEUE_MIGRATED;
3786 else
3787 #endif
3788 if (p->in_iowait) {
3789 delayacct_blkio_end(p);
3790 atomic_dec(&task_rq(p)->nr_iowait);
3791 }
3792
3793 activate_task(rq, p, en_flags);
3794 wakeup_preempt(rq, p, wake_flags);
3795
3796 ttwu_do_wakeup(p);
3797
3798 #ifdef CONFIG_SMP
3799 if (p->sched_class->task_woken) {
3800 /*
3801 * Our task @p is fully woken up and running; so it's safe to
3802 * drop the rq->lock, hereafter rq is only used for statistics.
3803 */
3804 rq_unpin_lock(rq, rf);
3805 p->sched_class->task_woken(rq, p);
3806 rq_repin_lock(rq, rf);
3807 }
3808
3809 if (rq->idle_stamp) {
3810 u64 delta = rq_clock(rq) - rq->idle_stamp;
3811 u64 max = 2*rq->max_idle_balance_cost;
3812
3813 update_avg(&rq->avg_idle, delta);
3814
3815 if (rq->avg_idle > max)
3816 rq->avg_idle = max;
3817
3818 rq->idle_stamp = 0;
3819 }
3820 #endif
3821
3822 p->dl_server = NULL;
3823 }
3824
3825 /*
3826 * Consider @p being inside a wait loop:
3827 *
3828 * for (;;) {
3829 * set_current_state(TASK_UNINTERRUPTIBLE);
3830 *
3831 * if (CONDITION)
3832 * break;
3833 *
3834 * schedule();
3835 * }
3836 * __set_current_state(TASK_RUNNING);
3837 *
3838 * between set_current_state() and schedule(). In this case @p is still
3839 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3840 * an atomic manner.
3841 *
3842 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3843 * then schedule() must still happen and p->state can be changed to
3844 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3845 * need to do a full wakeup with enqueue.
3846 *
3847 * Returns: %true when the wakeup is done,
3848 * %false otherwise.
3849 */
ttwu_runnable(struct task_struct * p,int wake_flags)3850 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3851 {
3852 struct rq_flags rf;
3853 struct rq *rq;
3854 int ret = 0;
3855
3856 rq = __task_rq_lock(p, &rf);
3857 if (task_on_rq_queued(p)) {
3858 if (!task_on_cpu(rq, p)) {
3859 /*
3860 * When on_rq && !on_cpu the task is preempted, see if
3861 * it should preempt the task that is current now.
3862 */
3863 update_rq_clock(rq);
3864 wakeup_preempt(rq, p, wake_flags);
3865 }
3866 ttwu_do_wakeup(p);
3867 ret = 1;
3868 }
3869 __task_rq_unlock(rq, &rf);
3870
3871 return ret;
3872 }
3873
3874 #ifdef CONFIG_SMP
sched_ttwu_pending(void * arg)3875 void sched_ttwu_pending(void *arg)
3876 {
3877 struct llist_node *llist = arg;
3878 struct rq *rq = this_rq();
3879 struct task_struct *p, *t;
3880 struct rq_flags rf;
3881
3882 if (!llist)
3883 return;
3884
3885 rq_lock_irqsave(rq, &rf);
3886 update_rq_clock(rq);
3887
3888 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3889 if (WARN_ON_ONCE(p->on_cpu))
3890 smp_cond_load_acquire(&p->on_cpu, !VAL);
3891
3892 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3893 set_task_cpu(p, cpu_of(rq));
3894
3895 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3896 }
3897
3898 /*
3899 * Must be after enqueueing at least once task such that
3900 * idle_cpu() does not observe a false-negative -- if it does,
3901 * it is possible for select_idle_siblings() to stack a number
3902 * of tasks on this CPU during that window.
3903 *
3904 * It is ok to clear ttwu_pending when another task pending.
3905 * We will receive IPI after local irq enabled and then enqueue it.
3906 * Since now nr_running > 0, idle_cpu() will always get correct result.
3907 */
3908 WRITE_ONCE(rq->ttwu_pending, 0);
3909 rq_unlock_irqrestore(rq, &rf);
3910 }
3911
3912 /*
3913 * Prepare the scene for sending an IPI for a remote smp_call
3914 *
3915 * Returns true if the caller can proceed with sending the IPI.
3916 * Returns false otherwise.
3917 */
call_function_single_prep_ipi(int cpu)3918 bool call_function_single_prep_ipi(int cpu)
3919 {
3920 if (set_nr_if_polling(cpu_rq(cpu)->idle)) {
3921 trace_sched_wake_idle_without_ipi(cpu);
3922 return false;
3923 }
3924
3925 return true;
3926 }
3927
3928 /*
3929 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3930 * necessary. The wakee CPU on receipt of the IPI will queue the task
3931 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3932 * of the wakeup instead of the waker.
3933 */
__ttwu_queue_wakelist(struct task_struct * p,int cpu,int wake_flags)3934 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3935 {
3936 struct rq *rq = cpu_rq(cpu);
3937
3938 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3939
3940 WRITE_ONCE(rq->ttwu_pending, 1);
3941 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3942 }
3943
wake_up_if_idle(int cpu)3944 void wake_up_if_idle(int cpu)
3945 {
3946 struct rq *rq = cpu_rq(cpu);
3947
3948 guard(rcu)();
3949 if (is_idle_task(rcu_dereference(rq->curr))) {
3950 guard(rq_lock_irqsave)(rq);
3951 if (is_idle_task(rq->curr))
3952 resched_curr(rq);
3953 }
3954 }
3955
cpus_equal_capacity(int this_cpu,int that_cpu)3956 bool cpus_equal_capacity(int this_cpu, int that_cpu)
3957 {
3958 if (!sched_asym_cpucap_active())
3959 return true;
3960
3961 if (this_cpu == that_cpu)
3962 return true;
3963
3964 return arch_scale_cpu_capacity(this_cpu) == arch_scale_cpu_capacity(that_cpu);
3965 }
3966
cpus_share_cache(int this_cpu,int that_cpu)3967 bool cpus_share_cache(int this_cpu, int that_cpu)
3968 {
3969 if (this_cpu == that_cpu)
3970 return true;
3971
3972 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3973 }
3974
3975 /*
3976 * Whether CPUs are share cache resources, which means LLC on non-cluster
3977 * machines and LLC tag or L2 on machines with clusters.
3978 */
cpus_share_resources(int this_cpu,int that_cpu)3979 bool cpus_share_resources(int this_cpu, int that_cpu)
3980 {
3981 if (this_cpu == that_cpu)
3982 return true;
3983
3984 return per_cpu(sd_share_id, this_cpu) == per_cpu(sd_share_id, that_cpu);
3985 }
3986
ttwu_queue_cond(struct task_struct * p,int cpu)3987 static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3988 {
3989 /*
3990 * Do not complicate things with the async wake_list while the CPU is
3991 * in hotplug state.
3992 */
3993 if (!cpu_active(cpu))
3994 return false;
3995
3996 /* Ensure the task will still be allowed to run on the CPU. */
3997 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3998 return false;
3999
4000 /*
4001 * If the CPU does not share cache, then queue the task on the
4002 * remote rqs wakelist to avoid accessing remote data.
4003 */
4004 if (!cpus_share_cache(smp_processor_id(), cpu))
4005 return true;
4006
4007 if (cpu == smp_processor_id())
4008 return false;
4009
4010 /*
4011 * If the wakee cpu is idle, or the task is descheduling and the
4012 * only running task on the CPU, then use the wakelist to offload
4013 * the task activation to the idle (or soon-to-be-idle) CPU as
4014 * the current CPU is likely busy. nr_running is checked to
4015 * avoid unnecessary task stacking.
4016 *
4017 * Note that we can only get here with (wakee) p->on_rq=0,
4018 * p->on_cpu can be whatever, we've done the dequeue, so
4019 * the wakee has been accounted out of ->nr_running.
4020 */
4021 if (!cpu_rq(cpu)->nr_running)
4022 return true;
4023
4024 return false;
4025 }
4026
ttwu_queue_wakelist(struct task_struct * p,int cpu,int wake_flags)4027 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4028 {
4029 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
4030 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
4031 __ttwu_queue_wakelist(p, cpu, wake_flags);
4032 return true;
4033 }
4034
4035 return false;
4036 }
4037
4038 #else /* !CONFIG_SMP */
4039
ttwu_queue_wakelist(struct task_struct * p,int cpu,int wake_flags)4040 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4041 {
4042 return false;
4043 }
4044
4045 #endif /* CONFIG_SMP */
4046
ttwu_queue(struct task_struct * p,int cpu,int wake_flags)4047 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
4048 {
4049 struct rq *rq = cpu_rq(cpu);
4050 struct rq_flags rf;
4051
4052 if (ttwu_queue_wakelist(p, cpu, wake_flags))
4053 return;
4054
4055 rq_lock(rq, &rf);
4056 update_rq_clock(rq);
4057 ttwu_do_activate(rq, p, wake_flags, &rf);
4058 rq_unlock(rq, &rf);
4059 }
4060
4061 /*
4062 * Invoked from try_to_wake_up() to check whether the task can be woken up.
4063 *
4064 * The caller holds p::pi_lock if p != current or has preemption
4065 * disabled when p == current.
4066 *
4067 * The rules of saved_state:
4068 *
4069 * The related locking code always holds p::pi_lock when updating
4070 * p::saved_state, which means the code is fully serialized in both cases.
4071 *
4072 * For PREEMPT_RT, the lock wait and lock wakeups happen via TASK_RTLOCK_WAIT.
4073 * No other bits set. This allows to distinguish all wakeup scenarios.
4074 *
4075 * For FREEZER, the wakeup happens via TASK_FROZEN. No other bits set. This
4076 * allows us to prevent early wakeup of tasks before they can be run on
4077 * asymmetric ISA architectures (eg ARMv9).
4078 */
4079 static __always_inline
ttwu_state_match(struct task_struct * p,unsigned int state,int * success)4080 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
4081 {
4082 int match;
4083
4084 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
4085 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
4086 state != TASK_RTLOCK_WAIT);
4087 }
4088
4089 *success = !!(match = __task_state_match(p, state));
4090
4091 /*
4092 * Saved state preserves the task state across blocking on
4093 * an RT lock or TASK_FREEZABLE tasks. If the state matches,
4094 * set p::saved_state to TASK_RUNNING, but do not wake the task
4095 * because it waits for a lock wakeup or __thaw_task(). Also
4096 * indicate success because from the regular waker's point of
4097 * view this has succeeded.
4098 *
4099 * After acquiring the lock the task will restore p::__state
4100 * from p::saved_state which ensures that the regular
4101 * wakeup is not lost. The restore will also set
4102 * p::saved_state to TASK_RUNNING so any further tests will
4103 * not result in false positives vs. @success
4104 */
4105 if (match < 0)
4106 p->saved_state = TASK_RUNNING;
4107
4108 return match > 0;
4109 }
4110
4111 /*
4112 * Notes on Program-Order guarantees on SMP systems.
4113 *
4114 * MIGRATION
4115 *
4116 * The basic program-order guarantee on SMP systems is that when a task [t]
4117 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
4118 * execution on its new CPU [c1].
4119 *
4120 * For migration (of runnable tasks) this is provided by the following means:
4121 *
4122 * A) UNLOCK of the rq(c0)->lock scheduling out task t
4123 * B) migration for t is required to synchronize *both* rq(c0)->lock and
4124 * rq(c1)->lock (if not at the same time, then in that order).
4125 * C) LOCK of the rq(c1)->lock scheduling in task
4126 *
4127 * Release/acquire chaining guarantees that B happens after A and C after B.
4128 * Note: the CPU doing B need not be c0 or c1
4129 *
4130 * Example:
4131 *
4132 * CPU0 CPU1 CPU2
4133 *
4134 * LOCK rq(0)->lock
4135 * sched-out X
4136 * sched-in Y
4137 * UNLOCK rq(0)->lock
4138 *
4139 * LOCK rq(0)->lock // orders against CPU0
4140 * dequeue X
4141 * UNLOCK rq(0)->lock
4142 *
4143 * LOCK rq(1)->lock
4144 * enqueue X
4145 * UNLOCK rq(1)->lock
4146 *
4147 * LOCK rq(1)->lock // orders against CPU2
4148 * sched-out Z
4149 * sched-in X
4150 * UNLOCK rq(1)->lock
4151 *
4152 *
4153 * BLOCKING -- aka. SLEEP + WAKEUP
4154 *
4155 * For blocking we (obviously) need to provide the same guarantee as for
4156 * migration. However the means are completely different as there is no lock
4157 * chain to provide order. Instead we do:
4158 *
4159 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
4160 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
4161 *
4162 * Example:
4163 *
4164 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
4165 *
4166 * LOCK rq(0)->lock LOCK X->pi_lock
4167 * dequeue X
4168 * sched-out X
4169 * smp_store_release(X->on_cpu, 0);
4170 *
4171 * smp_cond_load_acquire(&X->on_cpu, !VAL);
4172 * X->state = WAKING
4173 * set_task_cpu(X,2)
4174 *
4175 * LOCK rq(2)->lock
4176 * enqueue X
4177 * X->state = RUNNING
4178 * UNLOCK rq(2)->lock
4179 *
4180 * LOCK rq(2)->lock // orders against CPU1
4181 * sched-out Z
4182 * sched-in X
4183 * UNLOCK rq(2)->lock
4184 *
4185 * UNLOCK X->pi_lock
4186 * UNLOCK rq(0)->lock
4187 *
4188 *
4189 * However, for wakeups there is a second guarantee we must provide, namely we
4190 * must ensure that CONDITION=1 done by the caller can not be reordered with
4191 * accesses to the task state; see try_to_wake_up() and set_current_state().
4192 */
4193
4194 /**
4195 * try_to_wake_up - wake up a thread
4196 * @p: the thread to be awakened
4197 * @state: the mask of task states that can be woken
4198 * @wake_flags: wake modifier flags (WF_*)
4199 *
4200 * Conceptually does:
4201 *
4202 * If (@state & @p->state) @p->state = TASK_RUNNING.
4203 *
4204 * If the task was not queued/runnable, also place it back on a runqueue.
4205 *
4206 * This function is atomic against schedule() which would dequeue the task.
4207 *
4208 * It issues a full memory barrier before accessing @p->state, see the comment
4209 * with set_current_state().
4210 *
4211 * Uses p->pi_lock to serialize against concurrent wake-ups.
4212 *
4213 * Relies on p->pi_lock stabilizing:
4214 * - p->sched_class
4215 * - p->cpus_ptr
4216 * - p->sched_task_group
4217 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4218 *
4219 * Tries really hard to only take one task_rq(p)->lock for performance.
4220 * Takes rq->lock in:
4221 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4222 * - ttwu_queue() -- new rq, for enqueue of the task;
4223 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4224 *
4225 * As a consequence we race really badly with just about everything. See the
4226 * many memory barriers and their comments for details.
4227 *
4228 * Return: %true if @p->state changes (an actual wakeup was done),
4229 * %false otherwise.
4230 */
try_to_wake_up(struct task_struct * p,unsigned int state,int wake_flags)4231 int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4232 {
4233 guard(preempt)();
4234 int cpu, success = 0;
4235
4236 if (p == current) {
4237 /*
4238 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4239 * == smp_processor_id()'. Together this means we can special
4240 * case the whole 'p->on_rq && ttwu_runnable()' case below
4241 * without taking any locks.
4242 *
4243 * In particular:
4244 * - we rely on Program-Order guarantees for all the ordering,
4245 * - we're serialized against set_special_state() by virtue of
4246 * it disabling IRQs (this allows not taking ->pi_lock).
4247 */
4248 if (!ttwu_state_match(p, state, &success))
4249 goto out;
4250
4251 trace_sched_waking(p);
4252 ttwu_do_wakeup(p);
4253 goto out;
4254 }
4255
4256 /*
4257 * If we are going to wake up a thread waiting for CONDITION we
4258 * need to ensure that CONDITION=1 done by the caller can not be
4259 * reordered with p->state check below. This pairs with smp_store_mb()
4260 * in set_current_state() that the waiting thread does.
4261 */
4262 scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
4263 smp_mb__after_spinlock();
4264 if (!ttwu_state_match(p, state, &success))
4265 break;
4266
4267 trace_sched_waking(p);
4268
4269 /*
4270 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4271 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4272 * in smp_cond_load_acquire() below.
4273 *
4274 * sched_ttwu_pending() try_to_wake_up()
4275 * STORE p->on_rq = 1 LOAD p->state
4276 * UNLOCK rq->lock
4277 *
4278 * __schedule() (switch to task 'p')
4279 * LOCK rq->lock smp_rmb();
4280 * smp_mb__after_spinlock();
4281 * UNLOCK rq->lock
4282 *
4283 * [task p]
4284 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4285 *
4286 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4287 * __schedule(). See the comment for smp_mb__after_spinlock().
4288 *
4289 * A similar smp_rmb() lives in __task_needs_rq_lock().
4290 */
4291 smp_rmb();
4292 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4293 break;
4294
4295 #ifdef CONFIG_SMP
4296 /*
4297 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4298 * possible to, falsely, observe p->on_cpu == 0.
4299 *
4300 * One must be running (->on_cpu == 1) in order to remove oneself
4301 * from the runqueue.
4302 *
4303 * __schedule() (switch to task 'p') try_to_wake_up()
4304 * STORE p->on_cpu = 1 LOAD p->on_rq
4305 * UNLOCK rq->lock
4306 *
4307 * __schedule() (put 'p' to sleep)
4308 * LOCK rq->lock smp_rmb();
4309 * smp_mb__after_spinlock();
4310 * STORE p->on_rq = 0 LOAD p->on_cpu
4311 *
4312 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4313 * __schedule(). See the comment for smp_mb__after_spinlock().
4314 *
4315 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4316 * schedule()'s deactivate_task() has 'happened' and p will no longer
4317 * care about it's own p->state. See the comment in __schedule().
4318 */
4319 smp_acquire__after_ctrl_dep();
4320
4321 /*
4322 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4323 * == 0), which means we need to do an enqueue, change p->state to
4324 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4325 * enqueue, such as ttwu_queue_wakelist().
4326 */
4327 WRITE_ONCE(p->__state, TASK_WAKING);
4328
4329 /*
4330 * If the owning (remote) CPU is still in the middle of schedule() with
4331 * this task as prev, considering queueing p on the remote CPUs wake_list
4332 * which potentially sends an IPI instead of spinning on p->on_cpu to
4333 * let the waker make forward progress. This is safe because IRQs are
4334 * disabled and the IPI will deliver after on_cpu is cleared.
4335 *
4336 * Ensure we load task_cpu(p) after p->on_cpu:
4337 *
4338 * set_task_cpu(p, cpu);
4339 * STORE p->cpu = @cpu
4340 * __schedule() (switch to task 'p')
4341 * LOCK rq->lock
4342 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4343 * STORE p->on_cpu = 1 LOAD p->cpu
4344 *
4345 * to ensure we observe the correct CPU on which the task is currently
4346 * scheduling.
4347 */
4348 if (smp_load_acquire(&p->on_cpu) &&
4349 ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4350 break;
4351
4352 /*
4353 * If the owning (remote) CPU is still in the middle of schedule() with
4354 * this task as prev, wait until it's done referencing the task.
4355 *
4356 * Pairs with the smp_store_release() in finish_task().
4357 *
4358 * This ensures that tasks getting woken will be fully ordered against
4359 * their previous state and preserve Program Order.
4360 */
4361 smp_cond_load_acquire(&p->on_cpu, !VAL);
4362
4363 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4364 if (task_cpu(p) != cpu) {
4365 if (p->in_iowait) {
4366 delayacct_blkio_end(p);
4367 atomic_dec(&task_rq(p)->nr_iowait);
4368 }
4369
4370 wake_flags |= WF_MIGRATED;
4371 psi_ttwu_dequeue(p);
4372 set_task_cpu(p, cpu);
4373 }
4374 #else
4375 cpu = task_cpu(p);
4376 #endif /* CONFIG_SMP */
4377
4378 ttwu_queue(p, cpu, wake_flags);
4379 }
4380 out:
4381 if (success)
4382 ttwu_stat(p, task_cpu(p), wake_flags);
4383
4384 return success;
4385 }
4386
__task_needs_rq_lock(struct task_struct * p)4387 static bool __task_needs_rq_lock(struct task_struct *p)
4388 {
4389 unsigned int state = READ_ONCE(p->__state);
4390
4391 /*
4392 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4393 * the task is blocked. Make sure to check @state since ttwu() can drop
4394 * locks at the end, see ttwu_queue_wakelist().
4395 */
4396 if (state == TASK_RUNNING || state == TASK_WAKING)
4397 return true;
4398
4399 /*
4400 * Ensure we load p->on_rq after p->__state, otherwise it would be
4401 * possible to, falsely, observe p->on_rq == 0.
4402 *
4403 * See try_to_wake_up() for a longer comment.
4404 */
4405 smp_rmb();
4406 if (p->on_rq)
4407 return true;
4408
4409 #ifdef CONFIG_SMP
4410 /*
4411 * Ensure the task has finished __schedule() and will not be referenced
4412 * anymore. Again, see try_to_wake_up() for a longer comment.
4413 */
4414 smp_rmb();
4415 smp_cond_load_acquire(&p->on_cpu, !VAL);
4416 #endif
4417
4418 return false;
4419 }
4420
4421 /**
4422 * task_call_func - Invoke a function on task in fixed state
4423 * @p: Process for which the function is to be invoked, can be @current.
4424 * @func: Function to invoke.
4425 * @arg: Argument to function.
4426 *
4427 * Fix the task in it's current state by avoiding wakeups and or rq operations
4428 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4429 * to work out what the state is, if required. Given that @func can be invoked
4430 * with a runqueue lock held, it had better be quite lightweight.
4431 *
4432 * Returns:
4433 * Whatever @func returns
4434 */
task_call_func(struct task_struct * p,task_call_f func,void * arg)4435 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4436 {
4437 struct rq *rq = NULL;
4438 struct rq_flags rf;
4439 int ret;
4440
4441 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4442
4443 if (__task_needs_rq_lock(p))
4444 rq = __task_rq_lock(p, &rf);
4445
4446 /*
4447 * At this point the task is pinned; either:
4448 * - blocked and we're holding off wakeups (pi->lock)
4449 * - woken, and we're holding off enqueue (rq->lock)
4450 * - queued, and we're holding off schedule (rq->lock)
4451 * - running, and we're holding off de-schedule (rq->lock)
4452 *
4453 * The called function (@func) can use: task_curr(), p->on_rq and
4454 * p->__state to differentiate between these states.
4455 */
4456 ret = func(p, arg);
4457
4458 if (rq)
4459 rq_unlock(rq, &rf);
4460
4461 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4462 return ret;
4463 }
4464
4465 /**
4466 * cpu_curr_snapshot - Return a snapshot of the currently running task
4467 * @cpu: The CPU on which to snapshot the task.
4468 *
4469 * Returns the task_struct pointer of the task "currently" running on
4470 * the specified CPU. If the same task is running on that CPU throughout,
4471 * the return value will be a pointer to that task's task_struct structure.
4472 * If the CPU did any context switches even vaguely concurrently with the
4473 * execution of this function, the return value will be a pointer to the
4474 * task_struct structure of a randomly chosen task that was running on
4475 * that CPU somewhere around the time that this function was executing.
4476 *
4477 * If the specified CPU was offline, the return value is whatever it
4478 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4479 * task, but there is no guarantee. Callers wishing a useful return
4480 * value must take some action to ensure that the specified CPU remains
4481 * online throughout.
4482 *
4483 * This function executes full memory barriers before and after fetching
4484 * the pointer, which permits the caller to confine this function's fetch
4485 * with respect to the caller's accesses to other shared variables.
4486 */
cpu_curr_snapshot(int cpu)4487 struct task_struct *cpu_curr_snapshot(int cpu)
4488 {
4489 struct task_struct *t;
4490
4491 smp_mb(); /* Pairing determined by caller's synchronization design. */
4492 t = rcu_dereference(cpu_curr(cpu));
4493 smp_mb(); /* Pairing determined by caller's synchronization design. */
4494 return t;
4495 }
4496
4497 /**
4498 * wake_up_process - Wake up a specific process
4499 * @p: The process to be woken up.
4500 *
4501 * Attempt to wake up the nominated process and move it to the set of runnable
4502 * processes.
4503 *
4504 * Return: 1 if the process was woken up, 0 if it was already running.
4505 *
4506 * This function executes a full memory barrier before accessing the task state.
4507 */
wake_up_process(struct task_struct * p)4508 int wake_up_process(struct task_struct *p)
4509 {
4510 return try_to_wake_up(p, TASK_NORMAL, 0);
4511 }
4512 EXPORT_SYMBOL(wake_up_process);
4513
wake_up_state(struct task_struct * p,unsigned int state)4514 int wake_up_state(struct task_struct *p, unsigned int state)
4515 {
4516 return try_to_wake_up(p, state, 0);
4517 }
4518
4519 /*
4520 * Perform scheduler related setup for a newly forked process p.
4521 * p is forked by current.
4522 *
4523 * __sched_fork() is basic setup used by init_idle() too:
4524 */
__sched_fork(unsigned long clone_flags,struct task_struct * p)4525 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4526 {
4527 p->on_rq = 0;
4528
4529 p->se.on_rq = 0;
4530 p->se.exec_start = 0;
4531 p->se.sum_exec_runtime = 0;
4532 p->se.prev_sum_exec_runtime = 0;
4533 p->se.nr_migrations = 0;
4534 p->se.vruntime = 0;
4535 p->se.vlag = 0;
4536 p->se.slice = sysctl_sched_base_slice;
4537 INIT_LIST_HEAD(&p->se.group_node);
4538
4539 #ifdef CONFIG_FAIR_GROUP_SCHED
4540 p->se.cfs_rq = NULL;
4541 #endif
4542
4543 #ifdef CONFIG_SCHEDSTATS
4544 /* Even if schedstat is disabled, there should not be garbage */
4545 memset(&p->stats, 0, sizeof(p->stats));
4546 #endif
4547
4548 init_dl_entity(&p->dl);
4549
4550 INIT_LIST_HEAD(&p->rt.run_list);
4551 p->rt.timeout = 0;
4552 p->rt.time_slice = sched_rr_timeslice;
4553 p->rt.on_rq = 0;
4554 p->rt.on_list = 0;
4555
4556 #ifdef CONFIG_PREEMPT_NOTIFIERS
4557 INIT_HLIST_HEAD(&p->preempt_notifiers);
4558 #endif
4559
4560 #ifdef CONFIG_COMPACTION
4561 p->capture_control = NULL;
4562 #endif
4563 init_numa_balancing(clone_flags, p);
4564 #ifdef CONFIG_SMP
4565 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4566 p->migration_pending = NULL;
4567 #endif
4568 init_sched_mm_cid(p);
4569 }
4570
4571 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4572
4573 #ifdef CONFIG_NUMA_BALANCING
4574
4575 int sysctl_numa_balancing_mode;
4576
__set_numabalancing_state(bool enabled)4577 static void __set_numabalancing_state(bool enabled)
4578 {
4579 if (enabled)
4580 static_branch_enable(&sched_numa_balancing);
4581 else
4582 static_branch_disable(&sched_numa_balancing);
4583 }
4584
set_numabalancing_state(bool enabled)4585 void set_numabalancing_state(bool enabled)
4586 {
4587 if (enabled)
4588 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4589 else
4590 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4591 __set_numabalancing_state(enabled);
4592 }
4593
4594 #ifdef CONFIG_PROC_SYSCTL
reset_memory_tiering(void)4595 static void reset_memory_tiering(void)
4596 {
4597 struct pglist_data *pgdat;
4598
4599 for_each_online_pgdat(pgdat) {
4600 pgdat->nbp_threshold = 0;
4601 pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
4602 pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
4603 }
4604 }
4605
sysctl_numa_balancing(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)4606 static int sysctl_numa_balancing(struct ctl_table *table, int write,
4607 void *buffer, size_t *lenp, loff_t *ppos)
4608 {
4609 struct ctl_table t;
4610 int err;
4611 int state = sysctl_numa_balancing_mode;
4612
4613 if (write && !capable(CAP_SYS_ADMIN))
4614 return -EPERM;
4615
4616 t = *table;
4617 t.data = &state;
4618 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4619 if (err < 0)
4620 return err;
4621 if (write) {
4622 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4623 (state & NUMA_BALANCING_MEMORY_TIERING))
4624 reset_memory_tiering();
4625 sysctl_numa_balancing_mode = state;
4626 __set_numabalancing_state(state);
4627 }
4628 return err;
4629 }
4630 #endif
4631 #endif
4632
4633 #ifdef CONFIG_SCHEDSTATS
4634
4635 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4636
set_schedstats(bool enabled)4637 static void set_schedstats(bool enabled)
4638 {
4639 if (enabled)
4640 static_branch_enable(&sched_schedstats);
4641 else
4642 static_branch_disable(&sched_schedstats);
4643 }
4644
force_schedstat_enabled(void)4645 void force_schedstat_enabled(void)
4646 {
4647 if (!schedstat_enabled()) {
4648 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4649 static_branch_enable(&sched_schedstats);
4650 }
4651 }
4652
setup_schedstats(char * str)4653 static int __init setup_schedstats(char *str)
4654 {
4655 int ret = 0;
4656 if (!str)
4657 goto out;
4658
4659 if (!strcmp(str, "enable")) {
4660 set_schedstats(true);
4661 ret = 1;
4662 } else if (!strcmp(str, "disable")) {
4663 set_schedstats(false);
4664 ret = 1;
4665 }
4666 out:
4667 if (!ret)
4668 pr_warn("Unable to parse schedstats=\n");
4669
4670 return ret;
4671 }
4672 __setup("schedstats=", setup_schedstats);
4673
4674 #ifdef CONFIG_PROC_SYSCTL
sysctl_schedstats(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)4675 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4676 size_t *lenp, loff_t *ppos)
4677 {
4678 struct ctl_table t;
4679 int err;
4680 int state = static_branch_likely(&sched_schedstats);
4681
4682 if (write && !capable(CAP_SYS_ADMIN))
4683 return -EPERM;
4684
4685 t = *table;
4686 t.data = &state;
4687 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4688 if (err < 0)
4689 return err;
4690 if (write)
4691 set_schedstats(state);
4692 return err;
4693 }
4694 #endif /* CONFIG_PROC_SYSCTL */
4695 #endif /* CONFIG_SCHEDSTATS */
4696
4697 #ifdef CONFIG_SYSCTL
4698 static struct ctl_table sched_core_sysctls[] = {
4699 #ifdef CONFIG_SCHEDSTATS
4700 {
4701 .procname = "sched_schedstats",
4702 .data = NULL,
4703 .maxlen = sizeof(unsigned int),
4704 .mode = 0644,
4705 .proc_handler = sysctl_schedstats,
4706 .extra1 = SYSCTL_ZERO,
4707 .extra2 = SYSCTL_ONE,
4708 },
4709 #endif /* CONFIG_SCHEDSTATS */
4710 #ifdef CONFIG_UCLAMP_TASK
4711 {
4712 .procname = "sched_util_clamp_min",
4713 .data = &sysctl_sched_uclamp_util_min,
4714 .maxlen = sizeof(unsigned int),
4715 .mode = 0644,
4716 .proc_handler = sysctl_sched_uclamp_handler,
4717 },
4718 {
4719 .procname = "sched_util_clamp_max",
4720 .data = &sysctl_sched_uclamp_util_max,
4721 .maxlen = sizeof(unsigned int),
4722 .mode = 0644,
4723 .proc_handler = sysctl_sched_uclamp_handler,
4724 },
4725 {
4726 .procname = "sched_util_clamp_min_rt_default",
4727 .data = &sysctl_sched_uclamp_util_min_rt_default,
4728 .maxlen = sizeof(unsigned int),
4729 .mode = 0644,
4730 .proc_handler = sysctl_sched_uclamp_handler,
4731 },
4732 #endif /* CONFIG_UCLAMP_TASK */
4733 #ifdef CONFIG_NUMA_BALANCING
4734 {
4735 .procname = "numa_balancing",
4736 .data = NULL, /* filled in by handler */
4737 .maxlen = sizeof(unsigned int),
4738 .mode = 0644,
4739 .proc_handler = sysctl_numa_balancing,
4740 .extra1 = SYSCTL_ZERO,
4741 .extra2 = SYSCTL_FOUR,
4742 },
4743 #endif /* CONFIG_NUMA_BALANCING */
4744 };
sched_core_sysctl_init(void)4745 static int __init sched_core_sysctl_init(void)
4746 {
4747 register_sysctl_init("kernel", sched_core_sysctls);
4748 return 0;
4749 }
4750 late_initcall(sched_core_sysctl_init);
4751 #endif /* CONFIG_SYSCTL */
4752
4753 /*
4754 * fork()/clone()-time setup:
4755 */
sched_fork(unsigned long clone_flags,struct task_struct * p)4756 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4757 {
4758 __sched_fork(clone_flags, p);
4759 /*
4760 * We mark the process as NEW here. This guarantees that
4761 * nobody will actually run it, and a signal or other external
4762 * event cannot wake it up and insert it on the runqueue either.
4763 */
4764 p->__state = TASK_NEW;
4765
4766 /*
4767 * Make sure we do not leak PI boosting priority to the child.
4768 */
4769 p->prio = current->normal_prio;
4770
4771 uclamp_fork(p);
4772
4773 /*
4774 * Revert to default priority/policy on fork if requested.
4775 */
4776 if (unlikely(p->sched_reset_on_fork)) {
4777 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4778 p->policy = SCHED_NORMAL;
4779 p->static_prio = NICE_TO_PRIO(0);
4780 p->rt_priority = 0;
4781 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4782 p->static_prio = NICE_TO_PRIO(0);
4783
4784 p->prio = p->normal_prio = p->static_prio;
4785 set_load_weight(p, false);
4786
4787 /*
4788 * We don't need the reset flag anymore after the fork. It has
4789 * fulfilled its duty:
4790 */
4791 p->sched_reset_on_fork = 0;
4792 }
4793
4794 if (dl_prio(p->prio))
4795 return -EAGAIN;
4796 else if (rt_prio(p->prio))
4797 p->sched_class = &rt_sched_class;
4798 else
4799 p->sched_class = &fair_sched_class;
4800
4801 init_entity_runnable_average(&p->se);
4802
4803
4804 #ifdef CONFIG_SCHED_INFO
4805 if (likely(sched_info_on()))
4806 memset(&p->sched_info, 0, sizeof(p->sched_info));
4807 #endif
4808 #if defined(CONFIG_SMP)
4809 p->on_cpu = 0;
4810 #endif
4811 init_task_preempt_count(p);
4812 #ifdef CONFIG_SMP
4813 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4814 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4815 #endif
4816 return 0;
4817 }
4818
sched_cgroup_fork(struct task_struct * p,struct kernel_clone_args * kargs)4819 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4820 {
4821 unsigned long flags;
4822
4823 /*
4824 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4825 * required yet, but lockdep gets upset if rules are violated.
4826 */
4827 raw_spin_lock_irqsave(&p->pi_lock, flags);
4828 #ifdef CONFIG_CGROUP_SCHED
4829 if (1) {
4830 struct task_group *tg;
4831 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4832 struct task_group, css);
4833 tg = autogroup_task_group(p, tg);
4834 p->sched_task_group = tg;
4835 }
4836 #endif
4837 rseq_migrate(p);
4838 /*
4839 * We're setting the CPU for the first time, we don't migrate,
4840 * so use __set_task_cpu().
4841 */
4842 __set_task_cpu(p, smp_processor_id());
4843 if (p->sched_class->task_fork)
4844 p->sched_class->task_fork(p);
4845 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4846 }
4847
sched_post_fork(struct task_struct * p)4848 void sched_post_fork(struct task_struct *p)
4849 {
4850 uclamp_post_fork(p);
4851 }
4852
to_ratio(u64 period,u64 runtime)4853 unsigned long to_ratio(u64 period, u64 runtime)
4854 {
4855 if (runtime == RUNTIME_INF)
4856 return BW_UNIT;
4857
4858 /*
4859 * Doing this here saves a lot of checks in all
4860 * the calling paths, and returning zero seems
4861 * safe for them anyway.
4862 */
4863 if (period == 0)
4864 return 0;
4865
4866 return div64_u64(runtime << BW_SHIFT, period);
4867 }
4868
4869 /*
4870 * wake_up_new_task - wake up a newly created task for the first time.
4871 *
4872 * This function will do some initial scheduler statistics housekeeping
4873 * that must be done for every newly created context, then puts the task
4874 * on the runqueue and wakes it.
4875 */
wake_up_new_task(struct task_struct * p)4876 void wake_up_new_task(struct task_struct *p)
4877 {
4878 struct rq_flags rf;
4879 struct rq *rq;
4880
4881 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4882 WRITE_ONCE(p->__state, TASK_RUNNING);
4883 #ifdef CONFIG_SMP
4884 /*
4885 * Fork balancing, do it here and not earlier because:
4886 * - cpus_ptr can change in the fork path
4887 * - any previously selected CPU might disappear through hotplug
4888 *
4889 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4890 * as we're not fully set-up yet.
4891 */
4892 p->recent_used_cpu = task_cpu(p);
4893 rseq_migrate(p);
4894 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4895 #endif
4896 rq = __task_rq_lock(p, &rf);
4897 update_rq_clock(rq);
4898 post_init_entity_util_avg(p);
4899
4900 activate_task(rq, p, ENQUEUE_NOCLOCK);
4901 trace_sched_wakeup_new(p);
4902 wakeup_preempt(rq, p, WF_FORK);
4903 #ifdef CONFIG_SMP
4904 if (p->sched_class->task_woken) {
4905 /*
4906 * Nothing relies on rq->lock after this, so it's fine to
4907 * drop it.
4908 */
4909 rq_unpin_lock(rq, &rf);
4910 p->sched_class->task_woken(rq, p);
4911 rq_repin_lock(rq, &rf);
4912 }
4913 #endif
4914 task_rq_unlock(rq, p, &rf);
4915 }
4916
4917 #ifdef CONFIG_PREEMPT_NOTIFIERS
4918
4919 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4920
preempt_notifier_inc(void)4921 void preempt_notifier_inc(void)
4922 {
4923 static_branch_inc(&preempt_notifier_key);
4924 }
4925 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4926
preempt_notifier_dec(void)4927 void preempt_notifier_dec(void)
4928 {
4929 static_branch_dec(&preempt_notifier_key);
4930 }
4931 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4932
4933 /**
4934 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4935 * @notifier: notifier struct to register
4936 */
preempt_notifier_register(struct preempt_notifier * notifier)4937 void preempt_notifier_register(struct preempt_notifier *notifier)
4938 {
4939 if (!static_branch_unlikely(&preempt_notifier_key))
4940 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4941
4942 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4943 }
4944 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4945
4946 /**
4947 * preempt_notifier_unregister - no longer interested in preemption notifications
4948 * @notifier: notifier struct to unregister
4949 *
4950 * This is *not* safe to call from within a preemption notifier.
4951 */
preempt_notifier_unregister(struct preempt_notifier * notifier)4952 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4953 {
4954 hlist_del(¬ifier->link);
4955 }
4956 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4957
__fire_sched_in_preempt_notifiers(struct task_struct * curr)4958 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4959 {
4960 struct preempt_notifier *notifier;
4961
4962 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4963 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4964 }
4965
fire_sched_in_preempt_notifiers(struct task_struct * curr)4966 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4967 {
4968 if (static_branch_unlikely(&preempt_notifier_key))
4969 __fire_sched_in_preempt_notifiers(curr);
4970 }
4971
4972 static void
__fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)4973 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4974 struct task_struct *next)
4975 {
4976 struct preempt_notifier *notifier;
4977
4978 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4979 notifier->ops->sched_out(notifier, next);
4980 }
4981
4982 static __always_inline void
fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)4983 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4984 struct task_struct *next)
4985 {
4986 if (static_branch_unlikely(&preempt_notifier_key))
4987 __fire_sched_out_preempt_notifiers(curr, next);
4988 }
4989
4990 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4991
fire_sched_in_preempt_notifiers(struct task_struct * curr)4992 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4993 {
4994 }
4995
4996 static inline void
fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)4997 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4998 struct task_struct *next)
4999 {
5000 }
5001
5002 #endif /* CONFIG_PREEMPT_NOTIFIERS */
5003
prepare_task(struct task_struct * next)5004 static inline void prepare_task(struct task_struct *next)
5005 {
5006 #ifdef CONFIG_SMP
5007 /*
5008 * Claim the task as running, we do this before switching to it
5009 * such that any running task will have this set.
5010 *
5011 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
5012 * its ordering comment.
5013 */
5014 WRITE_ONCE(next->on_cpu, 1);
5015 #endif
5016 }
5017
finish_task(struct task_struct * prev)5018 static inline void finish_task(struct task_struct *prev)
5019 {
5020 #ifdef CONFIG_SMP
5021 /*
5022 * This must be the very last reference to @prev from this CPU. After
5023 * p->on_cpu is cleared, the task can be moved to a different CPU. We
5024 * must ensure this doesn't happen until the switch is completely
5025 * finished.
5026 *
5027 * In particular, the load of prev->state in finish_task_switch() must
5028 * happen before this.
5029 *
5030 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
5031 */
5032 smp_store_release(&prev->on_cpu, 0);
5033 #endif
5034 }
5035
5036 #ifdef CONFIG_SMP
5037
do_balance_callbacks(struct rq * rq,struct balance_callback * head)5038 static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
5039 {
5040 void (*func)(struct rq *rq);
5041 struct balance_callback *next;
5042
5043 lockdep_assert_rq_held(rq);
5044
5045 while (head) {
5046 func = (void (*)(struct rq *))head->func;
5047 next = head->next;
5048 head->next = NULL;
5049 head = next;
5050
5051 func(rq);
5052 }
5053 }
5054
5055 static void balance_push(struct rq *rq);
5056
5057 /*
5058 * balance_push_callback is a right abuse of the callback interface and plays
5059 * by significantly different rules.
5060 *
5061 * Where the normal balance_callback's purpose is to be ran in the same context
5062 * that queued it (only later, when it's safe to drop rq->lock again),
5063 * balance_push_callback is specifically targeted at __schedule().
5064 *
5065 * This abuse is tolerated because it places all the unlikely/odd cases behind
5066 * a single test, namely: rq->balance_callback == NULL.
5067 */
5068 struct balance_callback balance_push_callback = {
5069 .next = NULL,
5070 .func = balance_push,
5071 };
5072
5073 static inline struct balance_callback *
__splice_balance_callbacks(struct rq * rq,bool split)5074 __splice_balance_callbacks(struct rq *rq, bool split)
5075 {
5076 struct balance_callback *head = rq->balance_callback;
5077
5078 if (likely(!head))
5079 return NULL;
5080
5081 lockdep_assert_rq_held(rq);
5082 /*
5083 * Must not take balance_push_callback off the list when
5084 * splice_balance_callbacks() and balance_callbacks() are not
5085 * in the same rq->lock section.
5086 *
5087 * In that case it would be possible for __schedule() to interleave
5088 * and observe the list empty.
5089 */
5090 if (split && head == &balance_push_callback)
5091 head = NULL;
5092 else
5093 rq->balance_callback = NULL;
5094
5095 return head;
5096 }
5097
splice_balance_callbacks(struct rq * rq)5098 static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5099 {
5100 return __splice_balance_callbacks(rq, true);
5101 }
5102
__balance_callbacks(struct rq * rq)5103 static void __balance_callbacks(struct rq *rq)
5104 {
5105 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
5106 }
5107
balance_callbacks(struct rq * rq,struct balance_callback * head)5108 static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5109 {
5110 unsigned long flags;
5111
5112 if (unlikely(head)) {
5113 raw_spin_rq_lock_irqsave(rq, flags);
5114 do_balance_callbacks(rq, head);
5115 raw_spin_rq_unlock_irqrestore(rq, flags);
5116 }
5117 }
5118
5119 #else
5120
__balance_callbacks(struct rq * rq)5121 static inline void __balance_callbacks(struct rq *rq)
5122 {
5123 }
5124
splice_balance_callbacks(struct rq * rq)5125 static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5126 {
5127 return NULL;
5128 }
5129
balance_callbacks(struct rq * rq,struct balance_callback * head)5130 static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5131 {
5132 }
5133
5134 #endif
5135
5136 static inline void
prepare_lock_switch(struct rq * rq,struct task_struct * next,struct rq_flags * rf)5137 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
5138 {
5139 /*
5140 * Since the runqueue lock will be released by the next
5141 * task (which is an invalid locking op but in the case
5142 * of the scheduler it's an obvious special-case), so we
5143 * do an early lockdep release here:
5144 */
5145 rq_unpin_lock(rq, rf);
5146 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
5147 #ifdef CONFIG_DEBUG_SPINLOCK
5148 /* this is a valid case when another task releases the spinlock */
5149 rq_lockp(rq)->owner = next;
5150 #endif
5151 }
5152
finish_lock_switch(struct rq * rq)5153 static inline void finish_lock_switch(struct rq *rq)
5154 {
5155 /*
5156 * If we are tracking spinlock dependencies then we have to
5157 * fix up the runqueue lock - which gets 'carried over' from
5158 * prev into current:
5159 */
5160 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
5161 __balance_callbacks(rq);
5162 raw_spin_rq_unlock_irq(rq);
5163 }
5164
5165 /*
5166 * NOP if the arch has not defined these:
5167 */
5168
5169 #ifndef prepare_arch_switch
5170 # define prepare_arch_switch(next) do { } while (0)
5171 #endif
5172
5173 #ifndef finish_arch_post_lock_switch
5174 # define finish_arch_post_lock_switch() do { } while (0)
5175 #endif
5176
kmap_local_sched_out(void)5177 static inline void kmap_local_sched_out(void)
5178 {
5179 #ifdef CONFIG_KMAP_LOCAL
5180 if (unlikely(current->kmap_ctrl.idx))
5181 __kmap_local_sched_out();
5182 #endif
5183 }
5184
kmap_local_sched_in(void)5185 static inline void kmap_local_sched_in(void)
5186 {
5187 #ifdef CONFIG_KMAP_LOCAL
5188 if (unlikely(current->kmap_ctrl.idx))
5189 __kmap_local_sched_in();
5190 #endif
5191 }
5192
5193 /**
5194 * prepare_task_switch - prepare to switch tasks
5195 * @rq: the runqueue preparing to switch
5196 * @prev: the current task that is being switched out
5197 * @next: the task we are going to switch to.
5198 *
5199 * This is called with the rq lock held and interrupts off. It must
5200 * be paired with a subsequent finish_task_switch after the context
5201 * switch.
5202 *
5203 * prepare_task_switch sets up locking and calls architecture specific
5204 * hooks.
5205 */
5206 static inline void
prepare_task_switch(struct rq * rq,struct task_struct * prev,struct task_struct * next)5207 prepare_task_switch(struct rq *rq, struct task_struct *prev,
5208 struct task_struct *next)
5209 {
5210 kcov_prepare_switch(prev);
5211 sched_info_switch(rq, prev, next);
5212 perf_event_task_sched_out(prev, next);
5213 rseq_preempt(prev);
5214 fire_sched_out_preempt_notifiers(prev, next);
5215 kmap_local_sched_out();
5216 prepare_task(next);
5217 prepare_arch_switch(next);
5218 }
5219
5220 /**
5221 * finish_task_switch - clean up after a task-switch
5222 * @prev: the thread we just switched away from.
5223 *
5224 * finish_task_switch must be called after the context switch, paired
5225 * with a prepare_task_switch call before the context switch.
5226 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5227 * and do any other architecture-specific cleanup actions.
5228 *
5229 * Note that we may have delayed dropping an mm in context_switch(). If
5230 * so, we finish that here outside of the runqueue lock. (Doing it
5231 * with the lock held can cause deadlocks; see schedule() for
5232 * details.)
5233 *
5234 * The context switch have flipped the stack from under us and restored the
5235 * local variables which were saved when this task called schedule() in the
5236 * past. prev == current is still correct but we need to recalculate this_rq
5237 * because prev may have moved to another CPU.
5238 */
finish_task_switch(struct task_struct * prev)5239 static struct rq *finish_task_switch(struct task_struct *prev)
5240 __releases(rq->lock)
5241 {
5242 struct rq *rq = this_rq();
5243 struct mm_struct *mm = rq->prev_mm;
5244 unsigned int prev_state;
5245
5246 /*
5247 * The previous task will have left us with a preempt_count of 2
5248 * because it left us after:
5249 *
5250 * schedule()
5251 * preempt_disable(); // 1
5252 * __schedule()
5253 * raw_spin_lock_irq(&rq->lock) // 2
5254 *
5255 * Also, see FORK_PREEMPT_COUNT.
5256 */
5257 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5258 "corrupted preempt_count: %s/%d/0x%x\n",
5259 current->comm, current->pid, preempt_count()))
5260 preempt_count_set(FORK_PREEMPT_COUNT);
5261
5262 rq->prev_mm = NULL;
5263
5264 /*
5265 * A task struct has one reference for the use as "current".
5266 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5267 * schedule one last time. The schedule call will never return, and
5268 * the scheduled task must drop that reference.
5269 *
5270 * We must observe prev->state before clearing prev->on_cpu (in
5271 * finish_task), otherwise a concurrent wakeup can get prev
5272 * running on another CPU and we could rave with its RUNNING -> DEAD
5273 * transition, resulting in a double drop.
5274 */
5275 prev_state = READ_ONCE(prev->__state);
5276 vtime_task_switch(prev);
5277 perf_event_task_sched_in(prev, current);
5278 finish_task(prev);
5279 tick_nohz_task_switch();
5280 finish_lock_switch(rq);
5281 finish_arch_post_lock_switch();
5282 kcov_finish_switch(current);
5283 /*
5284 * kmap_local_sched_out() is invoked with rq::lock held and
5285 * interrupts disabled. There is no requirement for that, but the
5286 * sched out code does not have an interrupt enabled section.
5287 * Restoring the maps on sched in does not require interrupts being
5288 * disabled either.
5289 */
5290 kmap_local_sched_in();
5291
5292 fire_sched_in_preempt_notifiers(current);
5293 /*
5294 * When switching through a kernel thread, the loop in
5295 * membarrier_{private,global}_expedited() may have observed that
5296 * kernel thread and not issued an IPI. It is therefore possible to
5297 * schedule between user->kernel->user threads without passing though
5298 * switch_mm(). Membarrier requires a barrier after storing to
5299 * rq->curr, before returning to userspace, so provide them here:
5300 *
5301 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5302 * provided by mmdrop_lazy_tlb(),
5303 * - a sync_core for SYNC_CORE.
5304 */
5305 if (mm) {
5306 membarrier_mm_sync_core_before_usermode(mm);
5307 mmdrop_lazy_tlb_sched(mm);
5308 }
5309
5310 if (unlikely(prev_state == TASK_DEAD)) {
5311 if (prev->sched_class->task_dead)
5312 prev->sched_class->task_dead(prev);
5313
5314 /* Task is done with its stack. */
5315 put_task_stack(prev);
5316
5317 put_task_struct_rcu_user(prev);
5318 }
5319
5320 return rq;
5321 }
5322
5323 /**
5324 * schedule_tail - first thing a freshly forked thread must call.
5325 * @prev: the thread we just switched away from.
5326 */
schedule_tail(struct task_struct * prev)5327 asmlinkage __visible void schedule_tail(struct task_struct *prev)
5328 __releases(rq->lock)
5329 {
5330 /*
5331 * New tasks start with FORK_PREEMPT_COUNT, see there and
5332 * finish_task_switch() for details.
5333 *
5334 * finish_task_switch() will drop rq->lock() and lower preempt_count
5335 * and the preempt_enable() will end up enabling preemption (on
5336 * PREEMPT_COUNT kernels).
5337 */
5338
5339 finish_task_switch(prev);
5340 preempt_enable();
5341
5342 if (current->set_child_tid)
5343 put_user(task_pid_vnr(current), current->set_child_tid);
5344
5345 calculate_sigpending();
5346 }
5347
5348 /*
5349 * context_switch - switch to the new MM and the new thread's register state.
5350 */
5351 static __always_inline struct rq *
context_switch(struct rq * rq,struct task_struct * prev,struct task_struct * next,struct rq_flags * rf)5352 context_switch(struct rq *rq, struct task_struct *prev,
5353 struct task_struct *next, struct rq_flags *rf)
5354 {
5355 prepare_task_switch(rq, prev, next);
5356
5357 /*
5358 * For paravirt, this is coupled with an exit in switch_to to
5359 * combine the page table reload and the switch backend into
5360 * one hypercall.
5361 */
5362 arch_start_context_switch(prev);
5363
5364 /*
5365 * kernel -> kernel lazy + transfer active
5366 * user -> kernel lazy + mmgrab_lazy_tlb() active
5367 *
5368 * kernel -> user switch + mmdrop_lazy_tlb() active
5369 * user -> user switch
5370 *
5371 * switch_mm_cid() needs to be updated if the barriers provided
5372 * by context_switch() are modified.
5373 */
5374 if (!next->mm) { // to kernel
5375 enter_lazy_tlb(prev->active_mm, next);
5376
5377 next->active_mm = prev->active_mm;
5378 if (prev->mm) // from user
5379 mmgrab_lazy_tlb(prev->active_mm);
5380 else
5381 prev->active_mm = NULL;
5382 } else { // to user
5383 membarrier_switch_mm(rq, prev->active_mm, next->mm);
5384 /*
5385 * sys_membarrier() requires an smp_mb() between setting
5386 * rq->curr / membarrier_switch_mm() and returning to userspace.
5387 *
5388 * The below provides this either through switch_mm(), or in
5389 * case 'prev->active_mm == next->mm' through
5390 * finish_task_switch()'s mmdrop().
5391 */
5392 switch_mm_irqs_off(prev->active_mm, next->mm, next);
5393 lru_gen_use_mm(next->mm);
5394
5395 if (!prev->mm) { // from kernel
5396 /* will mmdrop_lazy_tlb() in finish_task_switch(). */
5397 rq->prev_mm = prev->active_mm;
5398 prev->active_mm = NULL;
5399 }
5400 }
5401
5402 /* switch_mm_cid() requires the memory barriers above. */
5403 switch_mm_cid(rq, prev, next);
5404
5405 prepare_lock_switch(rq, next, rf);
5406
5407 /* Here we just switch the register state and the stack. */
5408 switch_to(prev, next, prev);
5409 barrier();
5410
5411 return finish_task_switch(prev);
5412 }
5413
5414 /*
5415 * nr_running and nr_context_switches:
5416 *
5417 * externally visible scheduler statistics: current number of runnable
5418 * threads, total number of context switches performed since bootup.
5419 */
nr_running(void)5420 unsigned int nr_running(void)
5421 {
5422 unsigned int i, sum = 0;
5423
5424 for_each_online_cpu(i)
5425 sum += cpu_rq(i)->nr_running;
5426
5427 return sum;
5428 }
5429
5430 /*
5431 * Check if only the current task is running on the CPU.
5432 *
5433 * Caution: this function does not check that the caller has disabled
5434 * preemption, thus the result might have a time-of-check-to-time-of-use
5435 * race. The caller is responsible to use it correctly, for example:
5436 *
5437 * - from a non-preemptible section (of course)
5438 *
5439 * - from a thread that is bound to a single CPU
5440 *
5441 * - in a loop with very short iterations (e.g. a polling loop)
5442 */
single_task_running(void)5443 bool single_task_running(void)
5444 {
5445 return raw_rq()->nr_running == 1;
5446 }
5447 EXPORT_SYMBOL(single_task_running);
5448
nr_context_switches_cpu(int cpu)5449 unsigned long long nr_context_switches_cpu(int cpu)
5450 {
5451 return cpu_rq(cpu)->nr_switches;
5452 }
5453
nr_context_switches(void)5454 unsigned long long nr_context_switches(void)
5455 {
5456 int i;
5457 unsigned long long sum = 0;
5458
5459 for_each_possible_cpu(i)
5460 sum += cpu_rq(i)->nr_switches;
5461
5462 return sum;
5463 }
5464
5465 /*
5466 * Consumers of these two interfaces, like for example the cpuidle menu
5467 * governor, are using nonsensical data. Preferring shallow idle state selection
5468 * for a CPU that has IO-wait which might not even end up running the task when
5469 * it does become runnable.
5470 */
5471
nr_iowait_cpu(int cpu)5472 unsigned int nr_iowait_cpu(int cpu)
5473 {
5474 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5475 }
5476
5477 /*
5478 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5479 *
5480 * The idea behind IO-wait account is to account the idle time that we could
5481 * have spend running if it were not for IO. That is, if we were to improve the
5482 * storage performance, we'd have a proportional reduction in IO-wait time.
5483 *
5484 * This all works nicely on UP, where, when a task blocks on IO, we account
5485 * idle time as IO-wait, because if the storage were faster, it could've been
5486 * running and we'd not be idle.
5487 *
5488 * This has been extended to SMP, by doing the same for each CPU. This however
5489 * is broken.
5490 *
5491 * Imagine for instance the case where two tasks block on one CPU, only the one
5492 * CPU will have IO-wait accounted, while the other has regular idle. Even
5493 * though, if the storage were faster, both could've ran at the same time,
5494 * utilising both CPUs.
5495 *
5496 * This means, that when looking globally, the current IO-wait accounting on
5497 * SMP is a lower bound, by reason of under accounting.
5498 *
5499 * Worse, since the numbers are provided per CPU, they are sometimes
5500 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5501 * associated with any one particular CPU, it can wake to another CPU than it
5502 * blocked on. This means the per CPU IO-wait number is meaningless.
5503 *
5504 * Task CPU affinities can make all that even more 'interesting'.
5505 */
5506
nr_iowait(void)5507 unsigned int nr_iowait(void)
5508 {
5509 unsigned int i, sum = 0;
5510
5511 for_each_possible_cpu(i)
5512 sum += nr_iowait_cpu(i);
5513
5514 return sum;
5515 }
5516
5517 #ifdef CONFIG_SMP
5518
5519 /*
5520 * sched_exec - execve() is a valuable balancing opportunity, because at
5521 * this point the task has the smallest effective memory and cache footprint.
5522 */
sched_exec(void)5523 void sched_exec(void)
5524 {
5525 struct task_struct *p = current;
5526 struct migration_arg arg;
5527 int dest_cpu;
5528
5529 scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
5530 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5531 if (dest_cpu == smp_processor_id())
5532 return;
5533
5534 if (unlikely(!cpu_active(dest_cpu)))
5535 return;
5536
5537 arg = (struct migration_arg){ p, dest_cpu };
5538 }
5539 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5540 }
5541
5542 #endif
5543
5544 DEFINE_PER_CPU(struct kernel_stat, kstat);
5545 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5546
5547 EXPORT_PER_CPU_SYMBOL(kstat);
5548 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5549
5550 /*
5551 * The function fair_sched_class.update_curr accesses the struct curr
5552 * and its field curr->exec_start; when called from task_sched_runtime(),
5553 * we observe a high rate of cache misses in practice.
5554 * Prefetching this data results in improved performance.
5555 */
prefetch_curr_exec_start(struct task_struct * p)5556 static inline void prefetch_curr_exec_start(struct task_struct *p)
5557 {
5558 #ifdef CONFIG_FAIR_GROUP_SCHED
5559 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5560 #else
5561 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5562 #endif
5563 prefetch(curr);
5564 prefetch(&curr->exec_start);
5565 }
5566
5567 /*
5568 * Return accounted runtime for the task.
5569 * In case the task is currently running, return the runtime plus current's
5570 * pending runtime that have not been accounted yet.
5571 */
task_sched_runtime(struct task_struct * p)5572 unsigned long long task_sched_runtime(struct task_struct *p)
5573 {
5574 struct rq_flags rf;
5575 struct rq *rq;
5576 u64 ns;
5577
5578 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5579 /*
5580 * 64-bit doesn't need locks to atomically read a 64-bit value.
5581 * So we have a optimization chance when the task's delta_exec is 0.
5582 * Reading ->on_cpu is racy, but this is ok.
5583 *
5584 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5585 * If we race with it entering CPU, unaccounted time is 0. This is
5586 * indistinguishable from the read occurring a few cycles earlier.
5587 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5588 * been accounted, so we're correct here as well.
5589 */
5590 if (!p->on_cpu || !task_on_rq_queued(p))
5591 return p->se.sum_exec_runtime;
5592 #endif
5593
5594 rq = task_rq_lock(p, &rf);
5595 /*
5596 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5597 * project cycles that may never be accounted to this
5598 * thread, breaking clock_gettime().
5599 */
5600 if (task_current(rq, p) && task_on_rq_queued(p)) {
5601 prefetch_curr_exec_start(p);
5602 update_rq_clock(rq);
5603 p->sched_class->update_curr(rq);
5604 }
5605 ns = p->se.sum_exec_runtime;
5606 task_rq_unlock(rq, p, &rf);
5607
5608 return ns;
5609 }
5610
5611 #ifdef CONFIG_SCHED_DEBUG
cpu_resched_latency(struct rq * rq)5612 static u64 cpu_resched_latency(struct rq *rq)
5613 {
5614 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5615 u64 resched_latency, now = rq_clock(rq);
5616 static bool warned_once;
5617
5618 if (sysctl_resched_latency_warn_once && warned_once)
5619 return 0;
5620
5621 if (!need_resched() || !latency_warn_ms)
5622 return 0;
5623
5624 if (system_state == SYSTEM_BOOTING)
5625 return 0;
5626
5627 if (!rq->last_seen_need_resched_ns) {
5628 rq->last_seen_need_resched_ns = now;
5629 rq->ticks_without_resched = 0;
5630 return 0;
5631 }
5632
5633 rq->ticks_without_resched++;
5634 resched_latency = now - rq->last_seen_need_resched_ns;
5635 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5636 return 0;
5637
5638 warned_once = true;
5639
5640 return resched_latency;
5641 }
5642
setup_resched_latency_warn_ms(char * str)5643 static int __init setup_resched_latency_warn_ms(char *str)
5644 {
5645 long val;
5646
5647 if ((kstrtol(str, 0, &val))) {
5648 pr_warn("Unable to set resched_latency_warn_ms\n");
5649 return 1;
5650 }
5651
5652 sysctl_resched_latency_warn_ms = val;
5653 return 1;
5654 }
5655 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5656 #else
cpu_resched_latency(struct rq * rq)5657 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5658 #endif /* CONFIG_SCHED_DEBUG */
5659
5660 /*
5661 * This function gets called by the timer code, with HZ frequency.
5662 * We call it with interrupts disabled.
5663 */
sched_tick(void)5664 void sched_tick(void)
5665 {
5666 int cpu = smp_processor_id();
5667 struct rq *rq = cpu_rq(cpu);
5668 struct task_struct *curr = rq->curr;
5669 struct rq_flags rf;
5670 unsigned long hw_pressure;
5671 u64 resched_latency;
5672
5673 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5674 arch_scale_freq_tick();
5675
5676 sched_clock_tick();
5677
5678 rq_lock(rq, &rf);
5679
5680 update_rq_clock(rq);
5681 hw_pressure = arch_scale_hw_pressure(cpu_of(rq));
5682 update_hw_load_avg(rq_clock_task(rq), rq, hw_pressure);
5683 curr->sched_class->task_tick(rq, curr, 0);
5684 if (sched_feat(LATENCY_WARN))
5685 resched_latency = cpu_resched_latency(rq);
5686 calc_global_load_tick(rq);
5687 sched_core_tick(rq);
5688 task_tick_mm_cid(rq, curr);
5689
5690 rq_unlock(rq, &rf);
5691
5692 if (sched_feat(LATENCY_WARN) && resched_latency)
5693 resched_latency_warn(cpu, resched_latency);
5694
5695 perf_event_task_tick();
5696
5697 if (curr->flags & PF_WQ_WORKER)
5698 wq_worker_tick(curr);
5699
5700 #ifdef CONFIG_SMP
5701 rq->idle_balance = idle_cpu(cpu);
5702 sched_balance_trigger(rq);
5703 #endif
5704 }
5705
5706 #ifdef CONFIG_NO_HZ_FULL
5707
5708 struct tick_work {
5709 int cpu;
5710 atomic_t state;
5711 struct delayed_work work;
5712 };
5713 /* Values for ->state, see diagram below. */
5714 #define TICK_SCHED_REMOTE_OFFLINE 0
5715 #define TICK_SCHED_REMOTE_OFFLINING 1
5716 #define TICK_SCHED_REMOTE_RUNNING 2
5717
5718 /*
5719 * State diagram for ->state:
5720 *
5721 *
5722 * TICK_SCHED_REMOTE_OFFLINE
5723 * | ^
5724 * | |
5725 * | | sched_tick_remote()
5726 * | |
5727 * | |
5728 * +--TICK_SCHED_REMOTE_OFFLINING
5729 * | ^
5730 * | |
5731 * sched_tick_start() | | sched_tick_stop()
5732 * | |
5733 * V |
5734 * TICK_SCHED_REMOTE_RUNNING
5735 *
5736 *
5737 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5738 * and sched_tick_start() are happy to leave the state in RUNNING.
5739 */
5740
5741 static struct tick_work __percpu *tick_work_cpu;
5742
sched_tick_remote(struct work_struct * work)5743 static void sched_tick_remote(struct work_struct *work)
5744 {
5745 struct delayed_work *dwork = to_delayed_work(work);
5746 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5747 int cpu = twork->cpu;
5748 struct rq *rq = cpu_rq(cpu);
5749 int os;
5750
5751 /*
5752 * Handle the tick only if it appears the remote CPU is running in full
5753 * dynticks mode. The check is racy by nature, but missing a tick or
5754 * having one too much is no big deal because the scheduler tick updates
5755 * statistics and checks timeslices in a time-independent way, regardless
5756 * of when exactly it is running.
5757 */
5758 if (tick_nohz_tick_stopped_cpu(cpu)) {
5759 guard(rq_lock_irq)(rq);
5760 struct task_struct *curr = rq->curr;
5761
5762 if (cpu_online(cpu)) {
5763 update_rq_clock(rq);
5764
5765 if (!is_idle_task(curr)) {
5766 /*
5767 * Make sure the next tick runs within a
5768 * reasonable amount of time.
5769 */
5770 u64 delta = rq_clock_task(rq) - curr->se.exec_start;
5771 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5772 }
5773 curr->sched_class->task_tick(rq, curr, 0);
5774
5775 calc_load_nohz_remote(rq);
5776 }
5777 }
5778
5779 /*
5780 * Run the remote tick once per second (1Hz). This arbitrary
5781 * frequency is large enough to avoid overload but short enough
5782 * to keep scheduler internal stats reasonably up to date. But
5783 * first update state to reflect hotplug activity if required.
5784 */
5785 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5786 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5787 if (os == TICK_SCHED_REMOTE_RUNNING)
5788 queue_delayed_work(system_unbound_wq, dwork, HZ);
5789 }
5790
sched_tick_start(int cpu)5791 static void sched_tick_start(int cpu)
5792 {
5793 int os;
5794 struct tick_work *twork;
5795
5796 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5797 return;
5798
5799 WARN_ON_ONCE(!tick_work_cpu);
5800
5801 twork = per_cpu_ptr(tick_work_cpu, cpu);
5802 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5803 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5804 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5805 twork->cpu = cpu;
5806 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5807 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5808 }
5809 }
5810
5811 #ifdef CONFIG_HOTPLUG_CPU
sched_tick_stop(int cpu)5812 static void sched_tick_stop(int cpu)
5813 {
5814 struct tick_work *twork;
5815 int os;
5816
5817 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5818 return;
5819
5820 WARN_ON_ONCE(!tick_work_cpu);
5821
5822 twork = per_cpu_ptr(tick_work_cpu, cpu);
5823 /* There cannot be competing actions, but don't rely on stop-machine. */
5824 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5825 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5826 /* Don't cancel, as this would mess up the state machine. */
5827 }
5828 #endif /* CONFIG_HOTPLUG_CPU */
5829
sched_tick_offload_init(void)5830 int __init sched_tick_offload_init(void)
5831 {
5832 tick_work_cpu = alloc_percpu(struct tick_work);
5833 BUG_ON(!tick_work_cpu);
5834 return 0;
5835 }
5836
5837 #else /* !CONFIG_NO_HZ_FULL */
sched_tick_start(int cpu)5838 static inline void sched_tick_start(int cpu) { }
sched_tick_stop(int cpu)5839 static inline void sched_tick_stop(int cpu) { }
5840 #endif
5841
5842 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5843 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5844 /*
5845 * If the value passed in is equal to the current preempt count
5846 * then we just disabled preemption. Start timing the latency.
5847 */
preempt_latency_start(int val)5848 static inline void preempt_latency_start(int val)
5849 {
5850 if (preempt_count() == val) {
5851 unsigned long ip = get_lock_parent_ip();
5852 #ifdef CONFIG_DEBUG_PREEMPT
5853 current->preempt_disable_ip = ip;
5854 #endif
5855 trace_preempt_off(CALLER_ADDR0, ip);
5856 }
5857 }
5858
preempt_count_add(int val)5859 void preempt_count_add(int val)
5860 {
5861 #ifdef CONFIG_DEBUG_PREEMPT
5862 /*
5863 * Underflow?
5864 */
5865 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5866 return;
5867 #endif
5868 __preempt_count_add(val);
5869 #ifdef CONFIG_DEBUG_PREEMPT
5870 /*
5871 * Spinlock count overflowing soon?
5872 */
5873 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5874 PREEMPT_MASK - 10);
5875 #endif
5876 preempt_latency_start(val);
5877 }
5878 EXPORT_SYMBOL(preempt_count_add);
5879 NOKPROBE_SYMBOL(preempt_count_add);
5880
5881 /*
5882 * If the value passed in equals to the current preempt count
5883 * then we just enabled preemption. Stop timing the latency.
5884 */
preempt_latency_stop(int val)5885 static inline void preempt_latency_stop(int val)
5886 {
5887 if (preempt_count() == val)
5888 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5889 }
5890
preempt_count_sub(int val)5891 void preempt_count_sub(int val)
5892 {
5893 #ifdef CONFIG_DEBUG_PREEMPT
5894 /*
5895 * Underflow?
5896 */
5897 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5898 return;
5899 /*
5900 * Is the spinlock portion underflowing?
5901 */
5902 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5903 !(preempt_count() & PREEMPT_MASK)))
5904 return;
5905 #endif
5906
5907 preempt_latency_stop(val);
5908 __preempt_count_sub(val);
5909 }
5910 EXPORT_SYMBOL(preempt_count_sub);
5911 NOKPROBE_SYMBOL(preempt_count_sub);
5912
5913 #else
preempt_latency_start(int val)5914 static inline void preempt_latency_start(int val) { }
preempt_latency_stop(int val)5915 static inline void preempt_latency_stop(int val) { }
5916 #endif
5917
get_preempt_disable_ip(struct task_struct * p)5918 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5919 {
5920 #ifdef CONFIG_DEBUG_PREEMPT
5921 return p->preempt_disable_ip;
5922 #else
5923 return 0;
5924 #endif
5925 }
5926
5927 /*
5928 * Print scheduling while atomic bug:
5929 */
__schedule_bug(struct task_struct * prev)5930 static noinline void __schedule_bug(struct task_struct *prev)
5931 {
5932 /* Save this before calling printk(), since that will clobber it */
5933 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5934
5935 if (oops_in_progress)
5936 return;
5937
5938 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5939 prev->comm, prev->pid, preempt_count());
5940
5941 debug_show_held_locks(prev);
5942 print_modules();
5943 if (irqs_disabled())
5944 print_irqtrace_events(prev);
5945 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
5946 pr_err("Preemption disabled at:");
5947 print_ip_sym(KERN_ERR, preempt_disable_ip);
5948 }
5949 check_panic_on_warn("scheduling while atomic");
5950
5951 dump_stack();
5952 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5953 }
5954
5955 /*
5956 * Various schedule()-time debugging checks and statistics:
5957 */
schedule_debug(struct task_struct * prev,bool preempt)5958 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5959 {
5960 #ifdef CONFIG_SCHED_STACK_END_CHECK
5961 if (task_stack_end_corrupted(prev))
5962 panic("corrupted stack end detected inside scheduler\n");
5963
5964 if (task_scs_end_corrupted(prev))
5965 panic("corrupted shadow stack detected inside scheduler\n");
5966 #endif
5967
5968 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5969 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5970 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5971 prev->comm, prev->pid, prev->non_block_count);
5972 dump_stack();
5973 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5974 }
5975 #endif
5976
5977 if (unlikely(in_atomic_preempt_off())) {
5978 __schedule_bug(prev);
5979 preempt_count_set(PREEMPT_DISABLED);
5980 }
5981 rcu_sleep_check();
5982 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5983
5984 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5985
5986 schedstat_inc(this_rq()->sched_count);
5987 }
5988
put_prev_task_balance(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)5989 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5990 struct rq_flags *rf)
5991 {
5992 #ifdef CONFIG_SMP
5993 const struct sched_class *class;
5994 /*
5995 * We must do the balancing pass before put_prev_task(), such
5996 * that when we release the rq->lock the task is in the same
5997 * state as before we took rq->lock.
5998 *
5999 * We can terminate the balance pass as soon as we know there is
6000 * a runnable task of @class priority or higher.
6001 */
6002 for_class_range(class, prev->sched_class, &idle_sched_class) {
6003 if (class->balance(rq, prev, rf))
6004 break;
6005 }
6006 #endif
6007
6008 put_prev_task(rq, prev);
6009 }
6010
6011 /*
6012 * Pick up the highest-prio task:
6013 */
6014 static inline struct task_struct *
__pick_next_task(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)6015 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6016 {
6017 const struct sched_class *class;
6018 struct task_struct *p;
6019
6020 /*
6021 * Optimization: we know that if all tasks are in the fair class we can
6022 * call that function directly, but only if the @prev task wasn't of a
6023 * higher scheduling class, because otherwise those lose the
6024 * opportunity to pull in more work from other CPUs.
6025 */
6026 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
6027 rq->nr_running == rq->cfs.h_nr_running)) {
6028
6029 p = pick_next_task_fair(rq, prev, rf);
6030 if (unlikely(p == RETRY_TASK))
6031 goto restart;
6032
6033 /* Assume the next prioritized class is idle_sched_class */
6034 if (!p) {
6035 put_prev_task(rq, prev);
6036 p = pick_next_task_idle(rq);
6037 }
6038
6039 /*
6040 * This is the fast path; it cannot be a DL server pick;
6041 * therefore even if @p == @prev, ->dl_server must be NULL.
6042 */
6043 if (p->dl_server)
6044 p->dl_server = NULL;
6045
6046 return p;
6047 }
6048
6049 restart:
6050 put_prev_task_balance(rq, prev, rf);
6051
6052 /*
6053 * We've updated @prev and no longer need the server link, clear it.
6054 * Must be done before ->pick_next_task() because that can (re)set
6055 * ->dl_server.
6056 */
6057 if (prev->dl_server)
6058 prev->dl_server = NULL;
6059
6060 for_each_class(class) {
6061 p = class->pick_next_task(rq);
6062 if (p)
6063 return p;
6064 }
6065
6066 BUG(); /* The idle class should always have a runnable task. */
6067 }
6068
6069 #ifdef CONFIG_SCHED_CORE
is_task_rq_idle(struct task_struct * t)6070 static inline bool is_task_rq_idle(struct task_struct *t)
6071 {
6072 return (task_rq(t)->idle == t);
6073 }
6074
cookie_equals(struct task_struct * a,unsigned long cookie)6075 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
6076 {
6077 return is_task_rq_idle(a) || (a->core_cookie == cookie);
6078 }
6079
cookie_match(struct task_struct * a,struct task_struct * b)6080 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
6081 {
6082 if (is_task_rq_idle(a) || is_task_rq_idle(b))
6083 return true;
6084
6085 return a->core_cookie == b->core_cookie;
6086 }
6087
pick_task(struct rq * rq)6088 static inline struct task_struct *pick_task(struct rq *rq)
6089 {
6090 const struct sched_class *class;
6091 struct task_struct *p;
6092
6093 for_each_class(class) {
6094 p = class->pick_task(rq);
6095 if (p)
6096 return p;
6097 }
6098
6099 BUG(); /* The idle class should always have a runnable task. */
6100 }
6101
6102 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
6103
6104 static void queue_core_balance(struct rq *rq);
6105
6106 static struct task_struct *
pick_next_task(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)6107 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6108 {
6109 struct task_struct *next, *p, *max = NULL;
6110 const struct cpumask *smt_mask;
6111 bool fi_before = false;
6112 bool core_clock_updated = (rq == rq->core);
6113 unsigned long cookie;
6114 int i, cpu, occ = 0;
6115 struct rq *rq_i;
6116 bool need_sync;
6117
6118 if (!sched_core_enabled(rq))
6119 return __pick_next_task(rq, prev, rf);
6120
6121 cpu = cpu_of(rq);
6122
6123 /* Stopper task is switching into idle, no need core-wide selection. */
6124 if (cpu_is_offline(cpu)) {
6125 /*
6126 * Reset core_pick so that we don't enter the fastpath when
6127 * coming online. core_pick would already be migrated to
6128 * another cpu during offline.
6129 */
6130 rq->core_pick = NULL;
6131 return __pick_next_task(rq, prev, rf);
6132 }
6133
6134 /*
6135 * If there were no {en,de}queues since we picked (IOW, the task
6136 * pointers are all still valid), and we haven't scheduled the last
6137 * pick yet, do so now.
6138 *
6139 * rq->core_pick can be NULL if no selection was made for a CPU because
6140 * it was either offline or went offline during a sibling's core-wide
6141 * selection. In this case, do a core-wide selection.
6142 */
6143 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
6144 rq->core->core_pick_seq != rq->core_sched_seq &&
6145 rq->core_pick) {
6146 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
6147
6148 next = rq->core_pick;
6149 if (next != prev) {
6150 put_prev_task(rq, prev);
6151 set_next_task(rq, next);
6152 }
6153
6154 rq->core_pick = NULL;
6155 goto out;
6156 }
6157
6158 put_prev_task_balance(rq, prev, rf);
6159
6160 smt_mask = cpu_smt_mask(cpu);
6161 need_sync = !!rq->core->core_cookie;
6162
6163 /* reset state */
6164 rq->core->core_cookie = 0UL;
6165 if (rq->core->core_forceidle_count) {
6166 if (!core_clock_updated) {
6167 update_rq_clock(rq->core);
6168 core_clock_updated = true;
6169 }
6170 sched_core_account_forceidle(rq);
6171 /* reset after accounting force idle */
6172 rq->core->core_forceidle_start = 0;
6173 rq->core->core_forceidle_count = 0;
6174 rq->core->core_forceidle_occupation = 0;
6175 need_sync = true;
6176 fi_before = true;
6177 }
6178
6179 /*
6180 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6181 *
6182 * @task_seq guards the task state ({en,de}queues)
6183 * @pick_seq is the @task_seq we did a selection on
6184 * @sched_seq is the @pick_seq we scheduled
6185 *
6186 * However, preemptions can cause multiple picks on the same task set.
6187 * 'Fix' this by also increasing @task_seq for every pick.
6188 */
6189 rq->core->core_task_seq++;
6190
6191 /*
6192 * Optimize for common case where this CPU has no cookies
6193 * and there are no cookied tasks running on siblings.
6194 */
6195 if (!need_sync) {
6196 next = pick_task(rq);
6197 if (!next->core_cookie) {
6198 rq->core_pick = NULL;
6199 /*
6200 * For robustness, update the min_vruntime_fi for
6201 * unconstrained picks as well.
6202 */
6203 WARN_ON_ONCE(fi_before);
6204 task_vruntime_update(rq, next, false);
6205 goto out_set_next;
6206 }
6207 }
6208
6209 /*
6210 * For each thread: do the regular task pick and find the max prio task
6211 * amongst them.
6212 *
6213 * Tie-break prio towards the current CPU
6214 */
6215 for_each_cpu_wrap(i, smt_mask, cpu) {
6216 rq_i = cpu_rq(i);
6217
6218 /*
6219 * Current cpu always has its clock updated on entrance to
6220 * pick_next_task(). If the current cpu is not the core,
6221 * the core may also have been updated above.
6222 */
6223 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
6224 update_rq_clock(rq_i);
6225
6226 p = rq_i->core_pick = pick_task(rq_i);
6227 if (!max || prio_less(max, p, fi_before))
6228 max = p;
6229 }
6230
6231 cookie = rq->core->core_cookie = max->core_cookie;
6232
6233 /*
6234 * For each thread: try and find a runnable task that matches @max or
6235 * force idle.
6236 */
6237 for_each_cpu(i, smt_mask) {
6238 rq_i = cpu_rq(i);
6239 p = rq_i->core_pick;
6240
6241 if (!cookie_equals(p, cookie)) {
6242 p = NULL;
6243 if (cookie)
6244 p = sched_core_find(rq_i, cookie);
6245 if (!p)
6246 p = idle_sched_class.pick_task(rq_i);
6247 }
6248
6249 rq_i->core_pick = p;
6250
6251 if (p == rq_i->idle) {
6252 if (rq_i->nr_running) {
6253 rq->core->core_forceidle_count++;
6254 if (!fi_before)
6255 rq->core->core_forceidle_seq++;
6256 }
6257 } else {
6258 occ++;
6259 }
6260 }
6261
6262 if (schedstat_enabled() && rq->core->core_forceidle_count) {
6263 rq->core->core_forceidle_start = rq_clock(rq->core);
6264 rq->core->core_forceidle_occupation = occ;
6265 }
6266
6267 rq->core->core_pick_seq = rq->core->core_task_seq;
6268 next = rq->core_pick;
6269 rq->core_sched_seq = rq->core->core_pick_seq;
6270
6271 /* Something should have been selected for current CPU */
6272 WARN_ON_ONCE(!next);
6273
6274 /*
6275 * Reschedule siblings
6276 *
6277 * NOTE: L1TF -- at this point we're no longer running the old task and
6278 * sending an IPI (below) ensures the sibling will no longer be running
6279 * their task. This ensures there is no inter-sibling overlap between
6280 * non-matching user state.
6281 */
6282 for_each_cpu(i, smt_mask) {
6283 rq_i = cpu_rq(i);
6284
6285 /*
6286 * An online sibling might have gone offline before a task
6287 * could be picked for it, or it might be offline but later
6288 * happen to come online, but its too late and nothing was
6289 * picked for it. That's Ok - it will pick tasks for itself,
6290 * so ignore it.
6291 */
6292 if (!rq_i->core_pick)
6293 continue;
6294
6295 /*
6296 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6297 * fi_before fi update?
6298 * 0 0 1
6299 * 0 1 1
6300 * 1 0 1
6301 * 1 1 0
6302 */
6303 if (!(fi_before && rq->core->core_forceidle_count))
6304 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6305
6306 rq_i->core_pick->core_occupation = occ;
6307
6308 if (i == cpu) {
6309 rq_i->core_pick = NULL;
6310 continue;
6311 }
6312
6313 /* Did we break L1TF mitigation requirements? */
6314 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6315
6316 if (rq_i->curr == rq_i->core_pick) {
6317 rq_i->core_pick = NULL;
6318 continue;
6319 }
6320
6321 resched_curr(rq_i);
6322 }
6323
6324 out_set_next:
6325 set_next_task(rq, next);
6326 out:
6327 if (rq->core->core_forceidle_count && next == rq->idle)
6328 queue_core_balance(rq);
6329
6330 return next;
6331 }
6332
try_steal_cookie(int this,int that)6333 static bool try_steal_cookie(int this, int that)
6334 {
6335 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6336 struct task_struct *p;
6337 unsigned long cookie;
6338 bool success = false;
6339
6340 guard(irq)();
6341 guard(double_rq_lock)(dst, src);
6342
6343 cookie = dst->core->core_cookie;
6344 if (!cookie)
6345 return false;
6346
6347 if (dst->curr != dst->idle)
6348 return false;
6349
6350 p = sched_core_find(src, cookie);
6351 if (!p)
6352 return false;
6353
6354 do {
6355 if (p == src->core_pick || p == src->curr)
6356 goto next;
6357
6358 if (!is_cpu_allowed(p, this))
6359 goto next;
6360
6361 if (p->core_occupation > dst->idle->core_occupation)
6362 goto next;
6363 /*
6364 * sched_core_find() and sched_core_next() will ensure
6365 * that task @p is not throttled now, we also need to
6366 * check whether the runqueue of the destination CPU is
6367 * being throttled.
6368 */
6369 if (sched_task_is_throttled(p, this))
6370 goto next;
6371
6372 deactivate_task(src, p, 0);
6373 set_task_cpu(p, this);
6374 activate_task(dst, p, 0);
6375
6376 resched_curr(dst);
6377
6378 success = true;
6379 break;
6380
6381 next:
6382 p = sched_core_next(p, cookie);
6383 } while (p);
6384
6385 return success;
6386 }
6387
steal_cookie_task(int cpu,struct sched_domain * sd)6388 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6389 {
6390 int i;
6391
6392 for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
6393 if (i == cpu)
6394 continue;
6395
6396 if (need_resched())
6397 break;
6398
6399 if (try_steal_cookie(cpu, i))
6400 return true;
6401 }
6402
6403 return false;
6404 }
6405
sched_core_balance(struct rq * rq)6406 static void sched_core_balance(struct rq *rq)
6407 {
6408 struct sched_domain *sd;
6409 int cpu = cpu_of(rq);
6410
6411 guard(preempt)();
6412 guard(rcu)();
6413
6414 raw_spin_rq_unlock_irq(rq);
6415 for_each_domain(cpu, sd) {
6416 if (need_resched())
6417 break;
6418
6419 if (steal_cookie_task(cpu, sd))
6420 break;
6421 }
6422 raw_spin_rq_lock_irq(rq);
6423 }
6424
6425 static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
6426
queue_core_balance(struct rq * rq)6427 static void queue_core_balance(struct rq *rq)
6428 {
6429 if (!sched_core_enabled(rq))
6430 return;
6431
6432 if (!rq->core->core_cookie)
6433 return;
6434
6435 if (!rq->nr_running) /* not forced idle */
6436 return;
6437
6438 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6439 }
6440
6441 DEFINE_LOCK_GUARD_1(core_lock, int,
6442 sched_core_lock(*_T->lock, &_T->flags),
6443 sched_core_unlock(*_T->lock, &_T->flags),
6444 unsigned long flags)
6445
sched_core_cpu_starting(unsigned int cpu)6446 static void sched_core_cpu_starting(unsigned int cpu)
6447 {
6448 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6449 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6450 int t;
6451
6452 guard(core_lock)(&cpu);
6453
6454 WARN_ON_ONCE(rq->core != rq);
6455
6456 /* if we're the first, we'll be our own leader */
6457 if (cpumask_weight(smt_mask) == 1)
6458 return;
6459
6460 /* find the leader */
6461 for_each_cpu(t, smt_mask) {
6462 if (t == cpu)
6463 continue;
6464 rq = cpu_rq(t);
6465 if (rq->core == rq) {
6466 core_rq = rq;
6467 break;
6468 }
6469 }
6470
6471 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6472 return;
6473
6474 /* install and validate core_rq */
6475 for_each_cpu(t, smt_mask) {
6476 rq = cpu_rq(t);
6477
6478 if (t == cpu)
6479 rq->core = core_rq;
6480
6481 WARN_ON_ONCE(rq->core != core_rq);
6482 }
6483 }
6484
sched_core_cpu_deactivate(unsigned int cpu)6485 static void sched_core_cpu_deactivate(unsigned int cpu)
6486 {
6487 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6488 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6489 int t;
6490
6491 guard(core_lock)(&cpu);
6492
6493 /* if we're the last man standing, nothing to do */
6494 if (cpumask_weight(smt_mask) == 1) {
6495 WARN_ON_ONCE(rq->core != rq);
6496 return;
6497 }
6498
6499 /* if we're not the leader, nothing to do */
6500 if (rq->core != rq)
6501 return;
6502
6503 /* find a new leader */
6504 for_each_cpu(t, smt_mask) {
6505 if (t == cpu)
6506 continue;
6507 core_rq = cpu_rq(t);
6508 break;
6509 }
6510
6511 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6512 return;
6513
6514 /* copy the shared state to the new leader */
6515 core_rq->core_task_seq = rq->core_task_seq;
6516 core_rq->core_pick_seq = rq->core_pick_seq;
6517 core_rq->core_cookie = rq->core_cookie;
6518 core_rq->core_forceidle_count = rq->core_forceidle_count;
6519 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6520 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6521
6522 /*
6523 * Accounting edge for forced idle is handled in pick_next_task().
6524 * Don't need another one here, since the hotplug thread shouldn't
6525 * have a cookie.
6526 */
6527 core_rq->core_forceidle_start = 0;
6528
6529 /* install new leader */
6530 for_each_cpu(t, smt_mask) {
6531 rq = cpu_rq(t);
6532 rq->core = core_rq;
6533 }
6534 }
6535
sched_core_cpu_dying(unsigned int cpu)6536 static inline void sched_core_cpu_dying(unsigned int cpu)
6537 {
6538 struct rq *rq = cpu_rq(cpu);
6539
6540 if (rq->core != rq)
6541 rq->core = rq;
6542 }
6543
6544 #else /* !CONFIG_SCHED_CORE */
6545
sched_core_cpu_starting(unsigned int cpu)6546 static inline void sched_core_cpu_starting(unsigned int cpu) {}
sched_core_cpu_deactivate(unsigned int cpu)6547 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
sched_core_cpu_dying(unsigned int cpu)6548 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6549
6550 static struct task_struct *
pick_next_task(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)6551 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6552 {
6553 return __pick_next_task(rq, prev, rf);
6554 }
6555
6556 #endif /* CONFIG_SCHED_CORE */
6557
6558 /*
6559 * Constants for the sched_mode argument of __schedule().
6560 *
6561 * The mode argument allows RT enabled kernels to differentiate a
6562 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6563 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6564 * optimize the AND operation out and just check for zero.
6565 */
6566 #define SM_NONE 0x0
6567 #define SM_PREEMPT 0x1
6568 #define SM_RTLOCK_WAIT 0x2
6569
6570 #ifndef CONFIG_PREEMPT_RT
6571 # define SM_MASK_PREEMPT (~0U)
6572 #else
6573 # define SM_MASK_PREEMPT SM_PREEMPT
6574 #endif
6575
6576 /*
6577 * __schedule() is the main scheduler function.
6578 *
6579 * The main means of driving the scheduler and thus entering this function are:
6580 *
6581 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6582 *
6583 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6584 * paths. For example, see arch/x86/entry_64.S.
6585 *
6586 * To drive preemption between tasks, the scheduler sets the flag in timer
6587 * interrupt handler sched_tick().
6588 *
6589 * 3. Wakeups don't really cause entry into schedule(). They add a
6590 * task to the run-queue and that's it.
6591 *
6592 * Now, if the new task added to the run-queue preempts the current
6593 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6594 * called on the nearest possible occasion:
6595 *
6596 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6597 *
6598 * - in syscall or exception context, at the next outmost
6599 * preempt_enable(). (this might be as soon as the wake_up()'s
6600 * spin_unlock()!)
6601 *
6602 * - in IRQ context, return from interrupt-handler to
6603 * preemptible context
6604 *
6605 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6606 * then at the next:
6607 *
6608 * - cond_resched() call
6609 * - explicit schedule() call
6610 * - return from syscall or exception to user-space
6611 * - return from interrupt-handler to user-space
6612 *
6613 * WARNING: must be called with preemption disabled!
6614 */
__schedule(unsigned int sched_mode)6615 static void __sched notrace __schedule(unsigned int sched_mode)
6616 {
6617 struct task_struct *prev, *next;
6618 unsigned long *switch_count;
6619 unsigned long prev_state;
6620 struct rq_flags rf;
6621 struct rq *rq;
6622 int cpu;
6623
6624 cpu = smp_processor_id();
6625 rq = cpu_rq(cpu);
6626 prev = rq->curr;
6627
6628 schedule_debug(prev, !!sched_mode);
6629
6630 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6631 hrtick_clear(rq);
6632
6633 local_irq_disable();
6634 rcu_note_context_switch(!!sched_mode);
6635
6636 /*
6637 * Make sure that signal_pending_state()->signal_pending() below
6638 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6639 * done by the caller to avoid the race with signal_wake_up():
6640 *
6641 * __set_current_state(@state) signal_wake_up()
6642 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6643 * wake_up_state(p, state)
6644 * LOCK rq->lock LOCK p->pi_state
6645 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6646 * if (signal_pending_state()) if (p->state & @state)
6647 *
6648 * Also, the membarrier system call requires a full memory barrier
6649 * after coming from user-space, before storing to rq->curr; this
6650 * barrier matches a full barrier in the proximity of the membarrier
6651 * system call exit.
6652 */
6653 rq_lock(rq, &rf);
6654 smp_mb__after_spinlock();
6655
6656 /* Promote REQ to ACT */
6657 rq->clock_update_flags <<= 1;
6658 update_rq_clock(rq);
6659 rq->clock_update_flags = RQCF_UPDATED;
6660
6661 switch_count = &prev->nivcsw;
6662
6663 /*
6664 * We must load prev->state once (task_struct::state is volatile), such
6665 * that we form a control dependency vs deactivate_task() below.
6666 */
6667 prev_state = READ_ONCE(prev->__state);
6668 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6669 if (signal_pending_state(prev_state, prev)) {
6670 WRITE_ONCE(prev->__state, TASK_RUNNING);
6671 } else {
6672 prev->sched_contributes_to_load =
6673 (prev_state & TASK_UNINTERRUPTIBLE) &&
6674 !(prev_state & TASK_NOLOAD) &&
6675 !(prev_state & TASK_FROZEN);
6676
6677 if (prev->sched_contributes_to_load)
6678 rq->nr_uninterruptible++;
6679
6680 /*
6681 * __schedule() ttwu()
6682 * prev_state = prev->state; if (p->on_rq && ...)
6683 * if (prev_state) goto out;
6684 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6685 * p->state = TASK_WAKING
6686 *
6687 * Where __schedule() and ttwu() have matching control dependencies.
6688 *
6689 * After this, schedule() must not care about p->state any more.
6690 */
6691 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6692
6693 if (prev->in_iowait) {
6694 atomic_inc(&rq->nr_iowait);
6695 delayacct_blkio_start();
6696 }
6697 }
6698 switch_count = &prev->nvcsw;
6699 }
6700
6701 next = pick_next_task(rq, prev, &rf);
6702 clear_tsk_need_resched(prev);
6703 clear_preempt_need_resched();
6704 #ifdef CONFIG_SCHED_DEBUG
6705 rq->last_seen_need_resched_ns = 0;
6706 #endif
6707
6708 if (likely(prev != next)) {
6709 rq->nr_switches++;
6710 /*
6711 * RCU users of rcu_dereference(rq->curr) may not see
6712 * changes to task_struct made by pick_next_task().
6713 */
6714 RCU_INIT_POINTER(rq->curr, next);
6715 /*
6716 * The membarrier system call requires each architecture
6717 * to have a full memory barrier after updating
6718 * rq->curr, before returning to user-space.
6719 *
6720 * Here are the schemes providing that barrier on the
6721 * various architectures:
6722 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC,
6723 * RISC-V. switch_mm() relies on membarrier_arch_switch_mm()
6724 * on PowerPC and on RISC-V.
6725 * - finish_lock_switch() for weakly-ordered
6726 * architectures where spin_unlock is a full barrier,
6727 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6728 * is a RELEASE barrier),
6729 *
6730 * The barrier matches a full barrier in the proximity of
6731 * the membarrier system call entry.
6732 *
6733 * On RISC-V, this barrier pairing is also needed for the
6734 * SYNC_CORE command when switching between processes, cf.
6735 * the inline comments in membarrier_arch_switch_mm().
6736 */
6737 ++*switch_count;
6738
6739 migrate_disable_switch(rq, prev);
6740 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6741
6742 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6743
6744 /* Also unlocks the rq: */
6745 rq = context_switch(rq, prev, next, &rf);
6746 } else {
6747 rq_unpin_lock(rq, &rf);
6748 __balance_callbacks(rq);
6749 raw_spin_rq_unlock_irq(rq);
6750 }
6751 }
6752
do_task_dead(void)6753 void __noreturn do_task_dead(void)
6754 {
6755 /* Causes final put_task_struct in finish_task_switch(): */
6756 set_special_state(TASK_DEAD);
6757
6758 /* Tell freezer to ignore us: */
6759 current->flags |= PF_NOFREEZE;
6760
6761 __schedule(SM_NONE);
6762 BUG();
6763
6764 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6765 for (;;)
6766 cpu_relax();
6767 }
6768
sched_submit_work(struct task_struct * tsk)6769 static inline void sched_submit_work(struct task_struct *tsk)
6770 {
6771 static DEFINE_WAIT_OVERRIDE_MAP(sched_map, LD_WAIT_CONFIG);
6772 unsigned int task_flags;
6773
6774 /*
6775 * Establish LD_WAIT_CONFIG context to ensure none of the code called
6776 * will use a blocking primitive -- which would lead to recursion.
6777 */
6778 lock_map_acquire_try(&sched_map);
6779
6780 task_flags = tsk->flags;
6781 /*
6782 * If a worker goes to sleep, notify and ask workqueue whether it
6783 * wants to wake up a task to maintain concurrency.
6784 */
6785 if (task_flags & PF_WQ_WORKER)
6786 wq_worker_sleeping(tsk);
6787 else if (task_flags & PF_IO_WORKER)
6788 io_wq_worker_sleeping(tsk);
6789
6790 /*
6791 * spinlock and rwlock must not flush block requests. This will
6792 * deadlock if the callback attempts to acquire a lock which is
6793 * already acquired.
6794 */
6795 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6796
6797 /*
6798 * If we are going to sleep and we have plugged IO queued,
6799 * make sure to submit it to avoid deadlocks.
6800 */
6801 blk_flush_plug(tsk->plug, true);
6802
6803 lock_map_release(&sched_map);
6804 }
6805
sched_update_worker(struct task_struct * tsk)6806 static void sched_update_worker(struct task_struct *tsk)
6807 {
6808 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER | PF_BLOCK_TS)) {
6809 if (tsk->flags & PF_BLOCK_TS)
6810 blk_plug_invalidate_ts(tsk);
6811 if (tsk->flags & PF_WQ_WORKER)
6812 wq_worker_running(tsk);
6813 else if (tsk->flags & PF_IO_WORKER)
6814 io_wq_worker_running(tsk);
6815 }
6816 }
6817
__schedule_loop(unsigned int sched_mode)6818 static __always_inline void __schedule_loop(unsigned int sched_mode)
6819 {
6820 do {
6821 preempt_disable();
6822 __schedule(sched_mode);
6823 sched_preempt_enable_no_resched();
6824 } while (need_resched());
6825 }
6826
schedule(void)6827 asmlinkage __visible void __sched schedule(void)
6828 {
6829 struct task_struct *tsk = current;
6830
6831 #ifdef CONFIG_RT_MUTEXES
6832 lockdep_assert(!tsk->sched_rt_mutex);
6833 #endif
6834
6835 if (!task_is_running(tsk))
6836 sched_submit_work(tsk);
6837 __schedule_loop(SM_NONE);
6838 sched_update_worker(tsk);
6839 }
6840 EXPORT_SYMBOL(schedule);
6841
6842 /*
6843 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6844 * state (have scheduled out non-voluntarily) by making sure that all
6845 * tasks have either left the run queue or have gone into user space.
6846 * As idle tasks do not do either, they must not ever be preempted
6847 * (schedule out non-voluntarily).
6848 *
6849 * schedule_idle() is similar to schedule_preempt_disable() except that it
6850 * never enables preemption because it does not call sched_submit_work().
6851 */
schedule_idle(void)6852 void __sched schedule_idle(void)
6853 {
6854 /*
6855 * As this skips calling sched_submit_work(), which the idle task does
6856 * regardless because that function is a nop when the task is in a
6857 * TASK_RUNNING state, make sure this isn't used someplace that the
6858 * current task can be in any other state. Note, idle is always in the
6859 * TASK_RUNNING state.
6860 */
6861 WARN_ON_ONCE(current->__state);
6862 do {
6863 __schedule(SM_NONE);
6864 } while (need_resched());
6865 }
6866
6867 #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
schedule_user(void)6868 asmlinkage __visible void __sched schedule_user(void)
6869 {
6870 /*
6871 * If we come here after a random call to set_need_resched(),
6872 * or we have been woken up remotely but the IPI has not yet arrived,
6873 * we haven't yet exited the RCU idle mode. Do it here manually until
6874 * we find a better solution.
6875 *
6876 * NB: There are buggy callers of this function. Ideally we
6877 * should warn if prev_state != CONTEXT_USER, but that will trigger
6878 * too frequently to make sense yet.
6879 */
6880 enum ctx_state prev_state = exception_enter();
6881 schedule();
6882 exception_exit(prev_state);
6883 }
6884 #endif
6885
6886 /**
6887 * schedule_preempt_disabled - called with preemption disabled
6888 *
6889 * Returns with preemption disabled. Note: preempt_count must be 1
6890 */
schedule_preempt_disabled(void)6891 void __sched schedule_preempt_disabled(void)
6892 {
6893 sched_preempt_enable_no_resched();
6894 schedule();
6895 preempt_disable();
6896 }
6897
6898 #ifdef CONFIG_PREEMPT_RT
schedule_rtlock(void)6899 void __sched notrace schedule_rtlock(void)
6900 {
6901 __schedule_loop(SM_RTLOCK_WAIT);
6902 }
6903 NOKPROBE_SYMBOL(schedule_rtlock);
6904 #endif
6905
preempt_schedule_common(void)6906 static void __sched notrace preempt_schedule_common(void)
6907 {
6908 do {
6909 /*
6910 * Because the function tracer can trace preempt_count_sub()
6911 * and it also uses preempt_enable/disable_notrace(), if
6912 * NEED_RESCHED is set, the preempt_enable_notrace() called
6913 * by the function tracer will call this function again and
6914 * cause infinite recursion.
6915 *
6916 * Preemption must be disabled here before the function
6917 * tracer can trace. Break up preempt_disable() into two
6918 * calls. One to disable preemption without fear of being
6919 * traced. The other to still record the preemption latency,
6920 * which can also be traced by the function tracer.
6921 */
6922 preempt_disable_notrace();
6923 preempt_latency_start(1);
6924 __schedule(SM_PREEMPT);
6925 preempt_latency_stop(1);
6926 preempt_enable_no_resched_notrace();
6927
6928 /*
6929 * Check again in case we missed a preemption opportunity
6930 * between schedule and now.
6931 */
6932 } while (need_resched());
6933 }
6934
6935 #ifdef CONFIG_PREEMPTION
6936 /*
6937 * This is the entry point to schedule() from in-kernel preemption
6938 * off of preempt_enable.
6939 */
preempt_schedule(void)6940 asmlinkage __visible void __sched notrace preempt_schedule(void)
6941 {
6942 /*
6943 * If there is a non-zero preempt_count or interrupts are disabled,
6944 * we do not want to preempt the current task. Just return..
6945 */
6946 if (likely(!preemptible()))
6947 return;
6948 preempt_schedule_common();
6949 }
6950 NOKPROBE_SYMBOL(preempt_schedule);
6951 EXPORT_SYMBOL(preempt_schedule);
6952
6953 #ifdef CONFIG_PREEMPT_DYNAMIC
6954 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6955 #ifndef preempt_schedule_dynamic_enabled
6956 #define preempt_schedule_dynamic_enabled preempt_schedule
6957 #define preempt_schedule_dynamic_disabled NULL
6958 #endif
6959 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6960 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6961 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6962 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
dynamic_preempt_schedule(void)6963 void __sched notrace dynamic_preempt_schedule(void)
6964 {
6965 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6966 return;
6967 preempt_schedule();
6968 }
6969 NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6970 EXPORT_SYMBOL(dynamic_preempt_schedule);
6971 #endif
6972 #endif
6973
6974 /**
6975 * preempt_schedule_notrace - preempt_schedule called by tracing
6976 *
6977 * The tracing infrastructure uses preempt_enable_notrace to prevent
6978 * recursion and tracing preempt enabling caused by the tracing
6979 * infrastructure itself. But as tracing can happen in areas coming
6980 * from userspace or just about to enter userspace, a preempt enable
6981 * can occur before user_exit() is called. This will cause the scheduler
6982 * to be called when the system is still in usermode.
6983 *
6984 * To prevent this, the preempt_enable_notrace will use this function
6985 * instead of preempt_schedule() to exit user context if needed before
6986 * calling the scheduler.
6987 */
preempt_schedule_notrace(void)6988 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6989 {
6990 enum ctx_state prev_ctx;
6991
6992 if (likely(!preemptible()))
6993 return;
6994
6995 do {
6996 /*
6997 * Because the function tracer can trace preempt_count_sub()
6998 * and it also uses preempt_enable/disable_notrace(), if
6999 * NEED_RESCHED is set, the preempt_enable_notrace() called
7000 * by the function tracer will call this function again and
7001 * cause infinite recursion.
7002 *
7003 * Preemption must be disabled here before the function
7004 * tracer can trace. Break up preempt_disable() into two
7005 * calls. One to disable preemption without fear of being
7006 * traced. The other to still record the preemption latency,
7007 * which can also be traced by the function tracer.
7008 */
7009 preempt_disable_notrace();
7010 preempt_latency_start(1);
7011 /*
7012 * Needs preempt disabled in case user_exit() is traced
7013 * and the tracer calls preempt_enable_notrace() causing
7014 * an infinite recursion.
7015 */
7016 prev_ctx = exception_enter();
7017 __schedule(SM_PREEMPT);
7018 exception_exit(prev_ctx);
7019
7020 preempt_latency_stop(1);
7021 preempt_enable_no_resched_notrace();
7022 } while (need_resched());
7023 }
7024 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
7025
7026 #ifdef CONFIG_PREEMPT_DYNAMIC
7027 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
7028 #ifndef preempt_schedule_notrace_dynamic_enabled
7029 #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
7030 #define preempt_schedule_notrace_dynamic_disabled NULL
7031 #endif
7032 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
7033 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
7034 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
7035 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
dynamic_preempt_schedule_notrace(void)7036 void __sched notrace dynamic_preempt_schedule_notrace(void)
7037 {
7038 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
7039 return;
7040 preempt_schedule_notrace();
7041 }
7042 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
7043 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
7044 #endif
7045 #endif
7046
7047 #endif /* CONFIG_PREEMPTION */
7048
7049 /*
7050 * This is the entry point to schedule() from kernel preemption
7051 * off of irq context.
7052 * Note, that this is called and return with irqs disabled. This will
7053 * protect us against recursive calling from irq.
7054 */
preempt_schedule_irq(void)7055 asmlinkage __visible void __sched preempt_schedule_irq(void)
7056 {
7057 enum ctx_state prev_state;
7058
7059 /* Catch callers which need to be fixed */
7060 BUG_ON(preempt_count() || !irqs_disabled());
7061
7062 prev_state = exception_enter();
7063
7064 do {
7065 preempt_disable();
7066 local_irq_enable();
7067 __schedule(SM_PREEMPT);
7068 local_irq_disable();
7069 sched_preempt_enable_no_resched();
7070 } while (need_resched());
7071
7072 exception_exit(prev_state);
7073 }
7074
default_wake_function(wait_queue_entry_t * curr,unsigned mode,int wake_flags,void * key)7075 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
7076 void *key)
7077 {
7078 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~(WF_SYNC|WF_CURRENT_CPU));
7079 return try_to_wake_up(curr->private, mode, wake_flags);
7080 }
7081 EXPORT_SYMBOL(default_wake_function);
7082
__setscheduler_prio(struct task_struct * p,int prio)7083 static void __setscheduler_prio(struct task_struct *p, int prio)
7084 {
7085 if (dl_prio(prio))
7086 p->sched_class = &dl_sched_class;
7087 else if (rt_prio(prio))
7088 p->sched_class = &rt_sched_class;
7089 else
7090 p->sched_class = &fair_sched_class;
7091
7092 p->prio = prio;
7093 }
7094
7095 #ifdef CONFIG_RT_MUTEXES
7096
7097 /*
7098 * Would be more useful with typeof()/auto_type but they don't mix with
7099 * bit-fields. Since it's a local thing, use int. Keep the generic sounding
7100 * name such that if someone were to implement this function we get to compare
7101 * notes.
7102 */
7103 #define fetch_and_set(x, v) ({ int _x = (x); (x) = (v); _x; })
7104
rt_mutex_pre_schedule(void)7105 void rt_mutex_pre_schedule(void)
7106 {
7107 lockdep_assert(!fetch_and_set(current->sched_rt_mutex, 1));
7108 sched_submit_work(current);
7109 }
7110
rt_mutex_schedule(void)7111 void rt_mutex_schedule(void)
7112 {
7113 lockdep_assert(current->sched_rt_mutex);
7114 __schedule_loop(SM_NONE);
7115 }
7116
rt_mutex_post_schedule(void)7117 void rt_mutex_post_schedule(void)
7118 {
7119 sched_update_worker(current);
7120 lockdep_assert(fetch_and_set(current->sched_rt_mutex, 0));
7121 }
7122
__rt_effective_prio(struct task_struct * pi_task,int prio)7123 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
7124 {
7125 if (pi_task)
7126 prio = min(prio, pi_task->prio);
7127
7128 return prio;
7129 }
7130
rt_effective_prio(struct task_struct * p,int prio)7131 static inline int rt_effective_prio(struct task_struct *p, int prio)
7132 {
7133 struct task_struct *pi_task = rt_mutex_get_top_task(p);
7134
7135 return __rt_effective_prio(pi_task, prio);
7136 }
7137
7138 /*
7139 * rt_mutex_setprio - set the current priority of a task
7140 * @p: task to boost
7141 * @pi_task: donor task
7142 *
7143 * This function changes the 'effective' priority of a task. It does
7144 * not touch ->normal_prio like __setscheduler().
7145 *
7146 * Used by the rt_mutex code to implement priority inheritance
7147 * logic. Call site only calls if the priority of the task changed.
7148 */
rt_mutex_setprio(struct task_struct * p,struct task_struct * pi_task)7149 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
7150 {
7151 int prio, oldprio, queued, running, queue_flag =
7152 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7153 const struct sched_class *prev_class;
7154 struct rq_flags rf;
7155 struct rq *rq;
7156
7157 /* XXX used to be waiter->prio, not waiter->task->prio */
7158 prio = __rt_effective_prio(pi_task, p->normal_prio);
7159
7160 /*
7161 * If nothing changed; bail early.
7162 */
7163 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
7164 return;
7165
7166 rq = __task_rq_lock(p, &rf);
7167 update_rq_clock(rq);
7168 /*
7169 * Set under pi_lock && rq->lock, such that the value can be used under
7170 * either lock.
7171 *
7172 * Note that there is loads of tricky to make this pointer cache work
7173 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
7174 * ensure a task is de-boosted (pi_task is set to NULL) before the
7175 * task is allowed to run again (and can exit). This ensures the pointer
7176 * points to a blocked task -- which guarantees the task is present.
7177 */
7178 p->pi_top_task = pi_task;
7179
7180 /*
7181 * For FIFO/RR we only need to set prio, if that matches we're done.
7182 */
7183 if (prio == p->prio && !dl_prio(prio))
7184 goto out_unlock;
7185
7186 /*
7187 * Idle task boosting is a nono in general. There is one
7188 * exception, when PREEMPT_RT and NOHZ is active:
7189 *
7190 * The idle task calls get_next_timer_interrupt() and holds
7191 * the timer wheel base->lock on the CPU and another CPU wants
7192 * to access the timer (probably to cancel it). We can safely
7193 * ignore the boosting request, as the idle CPU runs this code
7194 * with interrupts disabled and will complete the lock
7195 * protected section without being interrupted. So there is no
7196 * real need to boost.
7197 */
7198 if (unlikely(p == rq->idle)) {
7199 WARN_ON(p != rq->curr);
7200 WARN_ON(p->pi_blocked_on);
7201 goto out_unlock;
7202 }
7203
7204 trace_sched_pi_setprio(p, pi_task);
7205 oldprio = p->prio;
7206
7207 if (oldprio == prio)
7208 queue_flag &= ~DEQUEUE_MOVE;
7209
7210 prev_class = p->sched_class;
7211 queued = task_on_rq_queued(p);
7212 running = task_current(rq, p);
7213 if (queued)
7214 dequeue_task(rq, p, queue_flag);
7215 if (running)
7216 put_prev_task(rq, p);
7217
7218 /*
7219 * Boosting condition are:
7220 * 1. -rt task is running and holds mutex A
7221 * --> -dl task blocks on mutex A
7222 *
7223 * 2. -dl task is running and holds mutex A
7224 * --> -dl task blocks on mutex A and could preempt the
7225 * running task
7226 */
7227 if (dl_prio(prio)) {
7228 if (!dl_prio(p->normal_prio) ||
7229 (pi_task && dl_prio(pi_task->prio) &&
7230 dl_entity_preempt(&pi_task->dl, &p->dl))) {
7231 p->dl.pi_se = pi_task->dl.pi_se;
7232 queue_flag |= ENQUEUE_REPLENISH;
7233 } else {
7234 p->dl.pi_se = &p->dl;
7235 }
7236 } else if (rt_prio(prio)) {
7237 if (dl_prio(oldprio))
7238 p->dl.pi_se = &p->dl;
7239 if (oldprio < prio)
7240 queue_flag |= ENQUEUE_HEAD;
7241 } else {
7242 if (dl_prio(oldprio))
7243 p->dl.pi_se = &p->dl;
7244 if (rt_prio(oldprio))
7245 p->rt.timeout = 0;
7246 }
7247
7248 __setscheduler_prio(p, prio);
7249
7250 if (queued)
7251 enqueue_task(rq, p, queue_flag);
7252 if (running)
7253 set_next_task(rq, p);
7254
7255 check_class_changed(rq, p, prev_class, oldprio);
7256 out_unlock:
7257 /* Avoid rq from going away on us: */
7258 preempt_disable();
7259
7260 rq_unpin_lock(rq, &rf);
7261 __balance_callbacks(rq);
7262 raw_spin_rq_unlock(rq);
7263
7264 preempt_enable();
7265 }
7266 #else
rt_effective_prio(struct task_struct * p,int prio)7267 static inline int rt_effective_prio(struct task_struct *p, int prio)
7268 {
7269 return prio;
7270 }
7271 #endif
7272
set_user_nice(struct task_struct * p,long nice)7273 void set_user_nice(struct task_struct *p, long nice)
7274 {
7275 bool queued, running;
7276 struct rq *rq;
7277 int old_prio;
7278
7279 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
7280 return;
7281 /*
7282 * We have to be careful, if called from sys_setpriority(),
7283 * the task might be in the middle of scheduling on another CPU.
7284 */
7285 CLASS(task_rq_lock, rq_guard)(p);
7286 rq = rq_guard.rq;
7287
7288 update_rq_clock(rq);
7289
7290 /*
7291 * The RT priorities are set via sched_setscheduler(), but we still
7292 * allow the 'normal' nice value to be set - but as expected
7293 * it won't have any effect on scheduling until the task is
7294 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7295 */
7296 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7297 p->static_prio = NICE_TO_PRIO(nice);
7298 return;
7299 }
7300
7301 queued = task_on_rq_queued(p);
7302 running = task_current(rq, p);
7303 if (queued)
7304 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7305 if (running)
7306 put_prev_task(rq, p);
7307
7308 p->static_prio = NICE_TO_PRIO(nice);
7309 set_load_weight(p, true);
7310 old_prio = p->prio;
7311 p->prio = effective_prio(p);
7312
7313 if (queued)
7314 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7315 if (running)
7316 set_next_task(rq, p);
7317
7318 /*
7319 * If the task increased its priority or is running and
7320 * lowered its priority, then reschedule its CPU:
7321 */
7322 p->sched_class->prio_changed(rq, p, old_prio);
7323 }
7324 EXPORT_SYMBOL(set_user_nice);
7325
7326 /*
7327 * is_nice_reduction - check if nice value is an actual reduction
7328 *
7329 * Similar to can_nice() but does not perform a capability check.
7330 *
7331 * @p: task
7332 * @nice: nice value
7333 */
is_nice_reduction(const struct task_struct * p,const int nice)7334 static bool is_nice_reduction(const struct task_struct *p, const int nice)
7335 {
7336 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7337 int nice_rlim = nice_to_rlimit(nice);
7338
7339 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7340 }
7341
7342 /*
7343 * can_nice - check if a task can reduce its nice value
7344 * @p: task
7345 * @nice: nice value
7346 */
can_nice(const struct task_struct * p,const int nice)7347 int can_nice(const struct task_struct *p, const int nice)
7348 {
7349 return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7350 }
7351
7352 #ifdef __ARCH_WANT_SYS_NICE
7353
7354 /*
7355 * sys_nice - change the priority of the current process.
7356 * @increment: priority increment
7357 *
7358 * sys_setpriority is a more generic, but much slower function that
7359 * does similar things.
7360 */
SYSCALL_DEFINE1(nice,int,increment)7361 SYSCALL_DEFINE1(nice, int, increment)
7362 {
7363 long nice, retval;
7364
7365 /*
7366 * Setpriority might change our priority at the same moment.
7367 * We don't have to worry. Conceptually one call occurs first
7368 * and we have a single winner.
7369 */
7370 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7371 nice = task_nice(current) + increment;
7372
7373 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7374 if (increment < 0 && !can_nice(current, nice))
7375 return -EPERM;
7376
7377 retval = security_task_setnice(current, nice);
7378 if (retval)
7379 return retval;
7380
7381 set_user_nice(current, nice);
7382 return 0;
7383 }
7384
7385 #endif
7386
7387 /**
7388 * task_prio - return the priority value of a given task.
7389 * @p: the task in question.
7390 *
7391 * Return: The priority value as seen by users in /proc.
7392 *
7393 * sched policy return value kernel prio user prio/nice
7394 *
7395 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7396 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7397 * deadline -101 -1 0
7398 */
task_prio(const struct task_struct * p)7399 int task_prio(const struct task_struct *p)
7400 {
7401 return p->prio - MAX_RT_PRIO;
7402 }
7403
7404 /**
7405 * idle_cpu - is a given CPU idle currently?
7406 * @cpu: the processor in question.
7407 *
7408 * Return: 1 if the CPU is currently idle. 0 otherwise.
7409 */
idle_cpu(int cpu)7410 int idle_cpu(int cpu)
7411 {
7412 struct rq *rq = cpu_rq(cpu);
7413
7414 if (rq->curr != rq->idle)
7415 return 0;
7416
7417 if (rq->nr_running)
7418 return 0;
7419
7420 #ifdef CONFIG_SMP
7421 if (rq->ttwu_pending)
7422 return 0;
7423 #endif
7424
7425 return 1;
7426 }
7427
7428 /**
7429 * available_idle_cpu - is a given CPU idle for enqueuing work.
7430 * @cpu: the CPU in question.
7431 *
7432 * Return: 1 if the CPU is currently idle. 0 otherwise.
7433 */
available_idle_cpu(int cpu)7434 int available_idle_cpu(int cpu)
7435 {
7436 if (!idle_cpu(cpu))
7437 return 0;
7438
7439 if (vcpu_is_preempted(cpu))
7440 return 0;
7441
7442 return 1;
7443 }
7444
7445 /**
7446 * idle_task - return the idle task for a given CPU.
7447 * @cpu: the processor in question.
7448 *
7449 * Return: The idle task for the CPU @cpu.
7450 */
idle_task(int cpu)7451 struct task_struct *idle_task(int cpu)
7452 {
7453 return cpu_rq(cpu)->idle;
7454 }
7455
7456 #ifdef CONFIG_SCHED_CORE
sched_core_idle_cpu(int cpu)7457 int sched_core_idle_cpu(int cpu)
7458 {
7459 struct rq *rq = cpu_rq(cpu);
7460
7461 if (sched_core_enabled(rq) && rq->curr == rq->idle)
7462 return 1;
7463
7464 return idle_cpu(cpu);
7465 }
7466
7467 #endif
7468
7469 #ifdef CONFIG_SMP
7470 /*
7471 * This function computes an effective utilization for the given CPU, to be
7472 * used for frequency selection given the linear relation: f = u * f_max.
7473 *
7474 * The scheduler tracks the following metrics:
7475 *
7476 * cpu_util_{cfs,rt,dl,irq}()
7477 * cpu_bw_dl()
7478 *
7479 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7480 * synchronized windows and are thus directly comparable.
7481 *
7482 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7483 * which excludes things like IRQ and steal-time. These latter are then accrued
7484 * in the irq utilization.
7485 *
7486 * The DL bandwidth number otoh is not a measured metric but a value computed
7487 * based on the task model parameters and gives the minimal utilization
7488 * required to meet deadlines.
7489 */
effective_cpu_util(int cpu,unsigned long util_cfs,unsigned long * min,unsigned long * max)7490 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7491 unsigned long *min,
7492 unsigned long *max)
7493 {
7494 unsigned long util, irq, scale;
7495 struct rq *rq = cpu_rq(cpu);
7496
7497 scale = arch_scale_cpu_capacity(cpu);
7498
7499 /*
7500 * Early check to see if IRQ/steal time saturates the CPU, can be
7501 * because of inaccuracies in how we track these -- see
7502 * update_irq_load_avg().
7503 */
7504 irq = cpu_util_irq(rq);
7505 if (unlikely(irq >= scale)) {
7506 if (min)
7507 *min = scale;
7508 if (max)
7509 *max = scale;
7510 return scale;
7511 }
7512
7513 if (min) {
7514 /*
7515 * The minimum utilization returns the highest level between:
7516 * - the computed DL bandwidth needed with the IRQ pressure which
7517 * steals time to the deadline task.
7518 * - The minimum performance requirement for CFS and/or RT.
7519 */
7520 *min = max(irq + cpu_bw_dl(rq), uclamp_rq_get(rq, UCLAMP_MIN));
7521
7522 /*
7523 * When an RT task is runnable and uclamp is not used, we must
7524 * ensure that the task will run at maximum compute capacity.
7525 */
7526 if (!uclamp_is_used() && rt_rq_is_runnable(&rq->rt))
7527 *min = max(*min, scale);
7528 }
7529
7530 /*
7531 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7532 * CFS tasks and we use the same metric to track the effective
7533 * utilization (PELT windows are synchronized) we can directly add them
7534 * to obtain the CPU's actual utilization.
7535 */
7536 util = util_cfs + cpu_util_rt(rq);
7537 util += cpu_util_dl(rq);
7538
7539 /*
7540 * The maximum hint is a soft bandwidth requirement, which can be lower
7541 * than the actual utilization because of uclamp_max requirements.
7542 */
7543 if (max)
7544 *max = min(scale, uclamp_rq_get(rq, UCLAMP_MAX));
7545
7546 if (util >= scale)
7547 return scale;
7548
7549 /*
7550 * There is still idle time; further improve the number by using the
7551 * irq metric. Because IRQ/steal time is hidden from the task clock we
7552 * need to scale the task numbers:
7553 *
7554 * max - irq
7555 * U' = irq + --------- * U
7556 * max
7557 */
7558 util = scale_irq_capacity(util, irq, scale);
7559 util += irq;
7560
7561 return min(scale, util);
7562 }
7563
sched_cpu_util(int cpu)7564 unsigned long sched_cpu_util(int cpu)
7565 {
7566 return effective_cpu_util(cpu, cpu_util_cfs(cpu), NULL, NULL);
7567 }
7568 #endif /* CONFIG_SMP */
7569
7570 /**
7571 * find_process_by_pid - find a process with a matching PID value.
7572 * @pid: the pid in question.
7573 *
7574 * The task of @pid, if found. %NULL otherwise.
7575 */
find_process_by_pid(pid_t pid)7576 static struct task_struct *find_process_by_pid(pid_t pid)
7577 {
7578 return pid ? find_task_by_vpid(pid) : current;
7579 }
7580
find_get_task(pid_t pid)7581 static struct task_struct *find_get_task(pid_t pid)
7582 {
7583 struct task_struct *p;
7584 guard(rcu)();
7585
7586 p = find_process_by_pid(pid);
7587 if (likely(p))
7588 get_task_struct(p);
7589
7590 return p;
7591 }
7592
DEFINE_CLASS(find_get_task,struct task_struct *,if (_T)put_task_struct (_T),find_get_task (pid),pid_t pid)7593 DEFINE_CLASS(find_get_task, struct task_struct *, if (_T) put_task_struct(_T),
7594 find_get_task(pid), pid_t pid)
7595
7596 /*
7597 * sched_setparam() passes in -1 for its policy, to let the functions
7598 * it calls know not to change it.
7599 */
7600 #define SETPARAM_POLICY -1
7601
7602 static void __setscheduler_params(struct task_struct *p,
7603 const struct sched_attr *attr)
7604 {
7605 int policy = attr->sched_policy;
7606
7607 if (policy == SETPARAM_POLICY)
7608 policy = p->policy;
7609
7610 p->policy = policy;
7611
7612 if (dl_policy(policy))
7613 __setparam_dl(p, attr);
7614 else if (fair_policy(policy))
7615 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7616
7617 /*
7618 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7619 * !rt_policy. Always setting this ensures that things like
7620 * getparam()/getattr() don't report silly values for !rt tasks.
7621 */
7622 p->rt_priority = attr->sched_priority;
7623 p->normal_prio = normal_prio(p);
7624 set_load_weight(p, true);
7625 }
7626
7627 /*
7628 * Check the target process has a UID that matches the current process's:
7629 */
check_same_owner(struct task_struct * p)7630 static bool check_same_owner(struct task_struct *p)
7631 {
7632 const struct cred *cred = current_cred(), *pcred;
7633 guard(rcu)();
7634
7635 pcred = __task_cred(p);
7636 return (uid_eq(cred->euid, pcred->euid) ||
7637 uid_eq(cred->euid, pcred->uid));
7638 }
7639
7640 /*
7641 * Allow unprivileged RT tasks to decrease priority.
7642 * Only issue a capable test if needed and only once to avoid an audit
7643 * event on permitted non-privileged operations:
7644 */
user_check_sched_setscheduler(struct task_struct * p,const struct sched_attr * attr,int policy,int reset_on_fork)7645 static int user_check_sched_setscheduler(struct task_struct *p,
7646 const struct sched_attr *attr,
7647 int policy, int reset_on_fork)
7648 {
7649 if (fair_policy(policy)) {
7650 if (attr->sched_nice < task_nice(p) &&
7651 !is_nice_reduction(p, attr->sched_nice))
7652 goto req_priv;
7653 }
7654
7655 if (rt_policy(policy)) {
7656 unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7657
7658 /* Can't set/change the rt policy: */
7659 if (policy != p->policy && !rlim_rtprio)
7660 goto req_priv;
7661
7662 /* Can't increase priority: */
7663 if (attr->sched_priority > p->rt_priority &&
7664 attr->sched_priority > rlim_rtprio)
7665 goto req_priv;
7666 }
7667
7668 /*
7669 * Can't set/change SCHED_DEADLINE policy at all for now
7670 * (safest behavior); in the future we would like to allow
7671 * unprivileged DL tasks to increase their relative deadline
7672 * or reduce their runtime (both ways reducing utilization)
7673 */
7674 if (dl_policy(policy))
7675 goto req_priv;
7676
7677 /*
7678 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7679 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7680 */
7681 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7682 if (!is_nice_reduction(p, task_nice(p)))
7683 goto req_priv;
7684 }
7685
7686 /* Can't change other user's priorities: */
7687 if (!check_same_owner(p))
7688 goto req_priv;
7689
7690 /* Normal users shall not reset the sched_reset_on_fork flag: */
7691 if (p->sched_reset_on_fork && !reset_on_fork)
7692 goto req_priv;
7693
7694 return 0;
7695
7696 req_priv:
7697 if (!capable(CAP_SYS_NICE))
7698 return -EPERM;
7699
7700 return 0;
7701 }
7702
__sched_setscheduler(struct task_struct * p,const struct sched_attr * attr,bool user,bool pi)7703 static int __sched_setscheduler(struct task_struct *p,
7704 const struct sched_attr *attr,
7705 bool user, bool pi)
7706 {
7707 int oldpolicy = -1, policy = attr->sched_policy;
7708 int retval, oldprio, newprio, queued, running;
7709 const struct sched_class *prev_class;
7710 struct balance_callback *head;
7711 struct rq_flags rf;
7712 int reset_on_fork;
7713 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7714 struct rq *rq;
7715 bool cpuset_locked = false;
7716
7717 /* The pi code expects interrupts enabled */
7718 BUG_ON(pi && in_interrupt());
7719 recheck:
7720 /* Double check policy once rq lock held: */
7721 if (policy < 0) {
7722 reset_on_fork = p->sched_reset_on_fork;
7723 policy = oldpolicy = p->policy;
7724 } else {
7725 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7726
7727 if (!valid_policy(policy))
7728 return -EINVAL;
7729 }
7730
7731 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7732 return -EINVAL;
7733
7734 /*
7735 * Valid priorities for SCHED_FIFO and SCHED_RR are
7736 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7737 * SCHED_BATCH and SCHED_IDLE is 0.
7738 */
7739 if (attr->sched_priority > MAX_RT_PRIO-1)
7740 return -EINVAL;
7741 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7742 (rt_policy(policy) != (attr->sched_priority != 0)))
7743 return -EINVAL;
7744
7745 if (user) {
7746 retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7747 if (retval)
7748 return retval;
7749
7750 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7751 return -EINVAL;
7752
7753 retval = security_task_setscheduler(p);
7754 if (retval)
7755 return retval;
7756 }
7757
7758 /* Update task specific "requested" clamps */
7759 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7760 retval = uclamp_validate(p, attr);
7761 if (retval)
7762 return retval;
7763 }
7764
7765 /*
7766 * SCHED_DEADLINE bandwidth accounting relies on stable cpusets
7767 * information.
7768 */
7769 if (dl_policy(policy) || dl_policy(p->policy)) {
7770 cpuset_locked = true;
7771 cpuset_lock();
7772 }
7773
7774 /*
7775 * Make sure no PI-waiters arrive (or leave) while we are
7776 * changing the priority of the task:
7777 *
7778 * To be able to change p->policy safely, the appropriate
7779 * runqueue lock must be held.
7780 */
7781 rq = task_rq_lock(p, &rf);
7782 update_rq_clock(rq);
7783
7784 /*
7785 * Changing the policy of the stop threads its a very bad idea:
7786 */
7787 if (p == rq->stop) {
7788 retval = -EINVAL;
7789 goto unlock;
7790 }
7791
7792 /*
7793 * If not changing anything there's no need to proceed further,
7794 * but store a possible modification of reset_on_fork.
7795 */
7796 if (unlikely(policy == p->policy)) {
7797 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7798 goto change;
7799 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7800 goto change;
7801 if (dl_policy(policy) && dl_param_changed(p, attr))
7802 goto change;
7803 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7804 goto change;
7805
7806 p->sched_reset_on_fork = reset_on_fork;
7807 retval = 0;
7808 goto unlock;
7809 }
7810 change:
7811
7812 if (user) {
7813 #ifdef CONFIG_RT_GROUP_SCHED
7814 /*
7815 * Do not allow realtime tasks into groups that have no runtime
7816 * assigned.
7817 */
7818 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7819 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7820 !task_group_is_autogroup(task_group(p))) {
7821 retval = -EPERM;
7822 goto unlock;
7823 }
7824 #endif
7825 #ifdef CONFIG_SMP
7826 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7827 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7828 cpumask_t *span = rq->rd->span;
7829
7830 /*
7831 * Don't allow tasks with an affinity mask smaller than
7832 * the entire root_domain to become SCHED_DEADLINE. We
7833 * will also fail if there's no bandwidth available.
7834 */
7835 if (!cpumask_subset(span, p->cpus_ptr) ||
7836 rq->rd->dl_bw.bw == 0) {
7837 retval = -EPERM;
7838 goto unlock;
7839 }
7840 }
7841 #endif
7842 }
7843
7844 /* Re-check policy now with rq lock held: */
7845 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7846 policy = oldpolicy = -1;
7847 task_rq_unlock(rq, p, &rf);
7848 if (cpuset_locked)
7849 cpuset_unlock();
7850 goto recheck;
7851 }
7852
7853 /*
7854 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7855 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7856 * is available.
7857 */
7858 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7859 retval = -EBUSY;
7860 goto unlock;
7861 }
7862
7863 p->sched_reset_on_fork = reset_on_fork;
7864 oldprio = p->prio;
7865
7866 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7867 if (pi) {
7868 /*
7869 * Take priority boosted tasks into account. If the new
7870 * effective priority is unchanged, we just store the new
7871 * normal parameters and do not touch the scheduler class and
7872 * the runqueue. This will be done when the task deboost
7873 * itself.
7874 */
7875 newprio = rt_effective_prio(p, newprio);
7876 if (newprio == oldprio)
7877 queue_flags &= ~DEQUEUE_MOVE;
7878 }
7879
7880 queued = task_on_rq_queued(p);
7881 running = task_current(rq, p);
7882 if (queued)
7883 dequeue_task(rq, p, queue_flags);
7884 if (running)
7885 put_prev_task(rq, p);
7886
7887 prev_class = p->sched_class;
7888
7889 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7890 __setscheduler_params(p, attr);
7891 __setscheduler_prio(p, newprio);
7892 }
7893 __setscheduler_uclamp(p, attr);
7894
7895 if (queued) {
7896 /*
7897 * We enqueue to tail when the priority of a task is
7898 * increased (user space view).
7899 */
7900 if (oldprio < p->prio)
7901 queue_flags |= ENQUEUE_HEAD;
7902
7903 enqueue_task(rq, p, queue_flags);
7904 }
7905 if (running)
7906 set_next_task(rq, p);
7907
7908 check_class_changed(rq, p, prev_class, oldprio);
7909
7910 /* Avoid rq from going away on us: */
7911 preempt_disable();
7912 head = splice_balance_callbacks(rq);
7913 task_rq_unlock(rq, p, &rf);
7914
7915 if (pi) {
7916 if (cpuset_locked)
7917 cpuset_unlock();
7918 rt_mutex_adjust_pi(p);
7919 }
7920
7921 /* Run balance callbacks after we've adjusted the PI chain: */
7922 balance_callbacks(rq, head);
7923 preempt_enable();
7924
7925 return 0;
7926
7927 unlock:
7928 task_rq_unlock(rq, p, &rf);
7929 if (cpuset_locked)
7930 cpuset_unlock();
7931 return retval;
7932 }
7933
_sched_setscheduler(struct task_struct * p,int policy,const struct sched_param * param,bool check)7934 static int _sched_setscheduler(struct task_struct *p, int policy,
7935 const struct sched_param *param, bool check)
7936 {
7937 struct sched_attr attr = {
7938 .sched_policy = policy,
7939 .sched_priority = param->sched_priority,
7940 .sched_nice = PRIO_TO_NICE(p->static_prio),
7941 };
7942
7943 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7944 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7945 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7946 policy &= ~SCHED_RESET_ON_FORK;
7947 attr.sched_policy = policy;
7948 }
7949
7950 return __sched_setscheduler(p, &attr, check, true);
7951 }
7952 /**
7953 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7954 * @p: the task in question.
7955 * @policy: new policy.
7956 * @param: structure containing the new RT priority.
7957 *
7958 * Use sched_set_fifo(), read its comment.
7959 *
7960 * Return: 0 on success. An error code otherwise.
7961 *
7962 * NOTE that the task may be already dead.
7963 */
sched_setscheduler(struct task_struct * p,int policy,const struct sched_param * param)7964 int sched_setscheduler(struct task_struct *p, int policy,
7965 const struct sched_param *param)
7966 {
7967 return _sched_setscheduler(p, policy, param, true);
7968 }
7969
sched_setattr(struct task_struct * p,const struct sched_attr * attr)7970 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7971 {
7972 return __sched_setscheduler(p, attr, true, true);
7973 }
7974
sched_setattr_nocheck(struct task_struct * p,const struct sched_attr * attr)7975 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7976 {
7977 return __sched_setscheduler(p, attr, false, true);
7978 }
7979 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7980
7981 /**
7982 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7983 * @p: the task in question.
7984 * @policy: new policy.
7985 * @param: structure containing the new RT priority.
7986 *
7987 * Just like sched_setscheduler, only don't bother checking if the
7988 * current context has permission. For example, this is needed in
7989 * stop_machine(): we create temporary high priority worker threads,
7990 * but our caller might not have that capability.
7991 *
7992 * Return: 0 on success. An error code otherwise.
7993 */
sched_setscheduler_nocheck(struct task_struct * p,int policy,const struct sched_param * param)7994 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7995 const struct sched_param *param)
7996 {
7997 return _sched_setscheduler(p, policy, param, false);
7998 }
7999
8000 /*
8001 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
8002 * incapable of resource management, which is the one thing an OS really should
8003 * be doing.
8004 *
8005 * This is of course the reason it is limited to privileged users only.
8006 *
8007 * Worse still; it is fundamentally impossible to compose static priority
8008 * workloads. You cannot take two correctly working static prio workloads
8009 * and smash them together and still expect them to work.
8010 *
8011 * For this reason 'all' FIFO tasks the kernel creates are basically at:
8012 *
8013 * MAX_RT_PRIO / 2
8014 *
8015 * The administrator _MUST_ configure the system, the kernel simply doesn't
8016 * know enough information to make a sensible choice.
8017 */
sched_set_fifo(struct task_struct * p)8018 void sched_set_fifo(struct task_struct *p)
8019 {
8020 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
8021 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
8022 }
8023 EXPORT_SYMBOL_GPL(sched_set_fifo);
8024
8025 /*
8026 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
8027 */
sched_set_fifo_low(struct task_struct * p)8028 void sched_set_fifo_low(struct task_struct *p)
8029 {
8030 struct sched_param sp = { .sched_priority = 1 };
8031 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
8032 }
8033 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
8034
sched_set_normal(struct task_struct * p,int nice)8035 void sched_set_normal(struct task_struct *p, int nice)
8036 {
8037 struct sched_attr attr = {
8038 .sched_policy = SCHED_NORMAL,
8039 .sched_nice = nice,
8040 };
8041 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
8042 }
8043 EXPORT_SYMBOL_GPL(sched_set_normal);
8044
8045 static int
do_sched_setscheduler(pid_t pid,int policy,struct sched_param __user * param)8046 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
8047 {
8048 struct sched_param lparam;
8049
8050 if (!param || pid < 0)
8051 return -EINVAL;
8052 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
8053 return -EFAULT;
8054
8055 CLASS(find_get_task, p)(pid);
8056 if (!p)
8057 return -ESRCH;
8058
8059 return sched_setscheduler(p, policy, &lparam);
8060 }
8061
8062 /*
8063 * Mimics kernel/events/core.c perf_copy_attr().
8064 */
sched_copy_attr(struct sched_attr __user * uattr,struct sched_attr * attr)8065 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
8066 {
8067 u32 size;
8068 int ret;
8069
8070 /* Zero the full structure, so that a short copy will be nice: */
8071 memset(attr, 0, sizeof(*attr));
8072
8073 ret = get_user(size, &uattr->size);
8074 if (ret)
8075 return ret;
8076
8077 /* ABI compatibility quirk: */
8078 if (!size)
8079 size = SCHED_ATTR_SIZE_VER0;
8080 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
8081 goto err_size;
8082
8083 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
8084 if (ret) {
8085 if (ret == -E2BIG)
8086 goto err_size;
8087 return ret;
8088 }
8089
8090 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
8091 size < SCHED_ATTR_SIZE_VER1)
8092 return -EINVAL;
8093
8094 /*
8095 * XXX: Do we want to be lenient like existing syscalls; or do we want
8096 * to be strict and return an error on out-of-bounds values?
8097 */
8098 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
8099
8100 return 0;
8101
8102 err_size:
8103 put_user(sizeof(*attr), &uattr->size);
8104 return -E2BIG;
8105 }
8106
get_params(struct task_struct * p,struct sched_attr * attr)8107 static void get_params(struct task_struct *p, struct sched_attr *attr)
8108 {
8109 if (task_has_dl_policy(p))
8110 __getparam_dl(p, attr);
8111 else if (task_has_rt_policy(p))
8112 attr->sched_priority = p->rt_priority;
8113 else
8114 attr->sched_nice = task_nice(p);
8115 }
8116
8117 /**
8118 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
8119 * @pid: the pid in question.
8120 * @policy: new policy.
8121 * @param: structure containing the new RT priority.
8122 *
8123 * Return: 0 on success. An error code otherwise.
8124 */
SYSCALL_DEFINE3(sched_setscheduler,pid_t,pid,int,policy,struct sched_param __user *,param)8125 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
8126 {
8127 if (policy < 0)
8128 return -EINVAL;
8129
8130 return do_sched_setscheduler(pid, policy, param);
8131 }
8132
8133 /**
8134 * sys_sched_setparam - set/change the RT priority of a thread
8135 * @pid: the pid in question.
8136 * @param: structure containing the new RT priority.
8137 *
8138 * Return: 0 on success. An error code otherwise.
8139 */
SYSCALL_DEFINE2(sched_setparam,pid_t,pid,struct sched_param __user *,param)8140 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
8141 {
8142 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
8143 }
8144
8145 /**
8146 * sys_sched_setattr - same as above, but with extended sched_attr
8147 * @pid: the pid in question.
8148 * @uattr: structure containing the extended parameters.
8149 * @flags: for future extension.
8150 */
SYSCALL_DEFINE3(sched_setattr,pid_t,pid,struct sched_attr __user *,uattr,unsigned int,flags)8151 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
8152 unsigned int, flags)
8153 {
8154 struct sched_attr attr;
8155 int retval;
8156
8157 if (!uattr || pid < 0 || flags)
8158 return -EINVAL;
8159
8160 retval = sched_copy_attr(uattr, &attr);
8161 if (retval)
8162 return retval;
8163
8164 if ((int)attr.sched_policy < 0)
8165 return -EINVAL;
8166 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
8167 attr.sched_policy = SETPARAM_POLICY;
8168
8169 CLASS(find_get_task, p)(pid);
8170 if (!p)
8171 return -ESRCH;
8172
8173 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
8174 get_params(p, &attr);
8175
8176 return sched_setattr(p, &attr);
8177 }
8178
8179 /**
8180 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
8181 * @pid: the pid in question.
8182 *
8183 * Return: On success, the policy of the thread. Otherwise, a negative error
8184 * code.
8185 */
SYSCALL_DEFINE1(sched_getscheduler,pid_t,pid)8186 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
8187 {
8188 struct task_struct *p;
8189 int retval;
8190
8191 if (pid < 0)
8192 return -EINVAL;
8193
8194 guard(rcu)();
8195 p = find_process_by_pid(pid);
8196 if (!p)
8197 return -ESRCH;
8198
8199 retval = security_task_getscheduler(p);
8200 if (!retval) {
8201 retval = p->policy;
8202 if (p->sched_reset_on_fork)
8203 retval |= SCHED_RESET_ON_FORK;
8204 }
8205 return retval;
8206 }
8207
8208 /**
8209 * sys_sched_getparam - get the RT priority of a thread
8210 * @pid: the pid in question.
8211 * @param: structure containing the RT priority.
8212 *
8213 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8214 * code.
8215 */
SYSCALL_DEFINE2(sched_getparam,pid_t,pid,struct sched_param __user *,param)8216 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
8217 {
8218 struct sched_param lp = { .sched_priority = 0 };
8219 struct task_struct *p;
8220 int retval;
8221
8222 if (!param || pid < 0)
8223 return -EINVAL;
8224
8225 scoped_guard (rcu) {
8226 p = find_process_by_pid(pid);
8227 if (!p)
8228 return -ESRCH;
8229
8230 retval = security_task_getscheduler(p);
8231 if (retval)
8232 return retval;
8233
8234 if (task_has_rt_policy(p))
8235 lp.sched_priority = p->rt_priority;
8236 }
8237
8238 /*
8239 * This one might sleep, we cannot do it with a spinlock held ...
8240 */
8241 return copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
8242 }
8243
8244 /*
8245 * Copy the kernel size attribute structure (which might be larger
8246 * than what user-space knows about) to user-space.
8247 *
8248 * Note that all cases are valid: user-space buffer can be larger or
8249 * smaller than the kernel-space buffer. The usual case is that both
8250 * have the same size.
8251 */
8252 static int
sched_attr_copy_to_user(struct sched_attr __user * uattr,struct sched_attr * kattr,unsigned int usize)8253 sched_attr_copy_to_user(struct sched_attr __user *uattr,
8254 struct sched_attr *kattr,
8255 unsigned int usize)
8256 {
8257 unsigned int ksize = sizeof(*kattr);
8258
8259 if (!access_ok(uattr, usize))
8260 return -EFAULT;
8261
8262 /*
8263 * sched_getattr() ABI forwards and backwards compatibility:
8264 *
8265 * If usize == ksize then we just copy everything to user-space and all is good.
8266 *
8267 * If usize < ksize then we only copy as much as user-space has space for,
8268 * this keeps ABI compatibility as well. We skip the rest.
8269 *
8270 * If usize > ksize then user-space is using a newer version of the ABI,
8271 * which part the kernel doesn't know about. Just ignore it - tooling can
8272 * detect the kernel's knowledge of attributes from the attr->size value
8273 * which is set to ksize in this case.
8274 */
8275 kattr->size = min(usize, ksize);
8276
8277 if (copy_to_user(uattr, kattr, kattr->size))
8278 return -EFAULT;
8279
8280 return 0;
8281 }
8282
8283 /**
8284 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8285 * @pid: the pid in question.
8286 * @uattr: structure containing the extended parameters.
8287 * @usize: sizeof(attr) for fwd/bwd comp.
8288 * @flags: for future extension.
8289 */
SYSCALL_DEFINE4(sched_getattr,pid_t,pid,struct sched_attr __user *,uattr,unsigned int,usize,unsigned int,flags)8290 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8291 unsigned int, usize, unsigned int, flags)
8292 {
8293 struct sched_attr kattr = { };
8294 struct task_struct *p;
8295 int retval;
8296
8297 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8298 usize < SCHED_ATTR_SIZE_VER0 || flags)
8299 return -EINVAL;
8300
8301 scoped_guard (rcu) {
8302 p = find_process_by_pid(pid);
8303 if (!p)
8304 return -ESRCH;
8305
8306 retval = security_task_getscheduler(p);
8307 if (retval)
8308 return retval;
8309
8310 kattr.sched_policy = p->policy;
8311 if (p->sched_reset_on_fork)
8312 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8313 get_params(p, &kattr);
8314 kattr.sched_flags &= SCHED_FLAG_ALL;
8315
8316 #ifdef CONFIG_UCLAMP_TASK
8317 /*
8318 * This could race with another potential updater, but this is fine
8319 * because it'll correctly read the old or the new value. We don't need
8320 * to guarantee who wins the race as long as it doesn't return garbage.
8321 */
8322 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8323 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8324 #endif
8325 }
8326
8327 return sched_attr_copy_to_user(uattr, &kattr, usize);
8328 }
8329
8330 #ifdef CONFIG_SMP
dl_task_check_affinity(struct task_struct * p,const struct cpumask * mask)8331 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8332 {
8333 /*
8334 * If the task isn't a deadline task or admission control is
8335 * disabled then we don't care about affinity changes.
8336 */
8337 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8338 return 0;
8339
8340 /*
8341 * Since bandwidth control happens on root_domain basis,
8342 * if admission test is enabled, we only admit -deadline
8343 * tasks allowed to run on all the CPUs in the task's
8344 * root_domain.
8345 */
8346 guard(rcu)();
8347 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8348 return -EBUSY;
8349
8350 return 0;
8351 }
8352 #endif
8353
8354 static int
__sched_setaffinity(struct task_struct * p,struct affinity_context * ctx)8355 __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
8356 {
8357 int retval;
8358 cpumask_var_t cpus_allowed, new_mask;
8359
8360 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8361 return -ENOMEM;
8362
8363 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8364 retval = -ENOMEM;
8365 goto out_free_cpus_allowed;
8366 }
8367
8368 cpuset_cpus_allowed(p, cpus_allowed);
8369 cpumask_and(new_mask, ctx->new_mask, cpus_allowed);
8370
8371 ctx->new_mask = new_mask;
8372 ctx->flags |= SCA_CHECK;
8373
8374 retval = dl_task_check_affinity(p, new_mask);
8375 if (retval)
8376 goto out_free_new_mask;
8377
8378 retval = __set_cpus_allowed_ptr(p, ctx);
8379 if (retval)
8380 goto out_free_new_mask;
8381
8382 cpuset_cpus_allowed(p, cpus_allowed);
8383 if (!cpumask_subset(new_mask, cpus_allowed)) {
8384 /*
8385 * We must have raced with a concurrent cpuset update.
8386 * Just reset the cpumask to the cpuset's cpus_allowed.
8387 */
8388 cpumask_copy(new_mask, cpus_allowed);
8389
8390 /*
8391 * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8392 * will restore the previous user_cpus_ptr value.
8393 *
8394 * In the unlikely event a previous user_cpus_ptr exists,
8395 * we need to further restrict the mask to what is allowed
8396 * by that old user_cpus_ptr.
8397 */
8398 if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
8399 bool empty = !cpumask_and(new_mask, new_mask,
8400 ctx->user_mask);
8401
8402 if (WARN_ON_ONCE(empty))
8403 cpumask_copy(new_mask, cpus_allowed);
8404 }
8405 __set_cpus_allowed_ptr(p, ctx);
8406 retval = -EINVAL;
8407 }
8408
8409 out_free_new_mask:
8410 free_cpumask_var(new_mask);
8411 out_free_cpus_allowed:
8412 free_cpumask_var(cpus_allowed);
8413 return retval;
8414 }
8415
sched_setaffinity(pid_t pid,const struct cpumask * in_mask)8416 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8417 {
8418 struct affinity_context ac;
8419 struct cpumask *user_mask;
8420 int retval;
8421
8422 CLASS(find_get_task, p)(pid);
8423 if (!p)
8424 return -ESRCH;
8425
8426 if (p->flags & PF_NO_SETAFFINITY)
8427 return -EINVAL;
8428
8429 if (!check_same_owner(p)) {
8430 guard(rcu)();
8431 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE))
8432 return -EPERM;
8433 }
8434
8435 retval = security_task_setscheduler(p);
8436 if (retval)
8437 return retval;
8438
8439 /*
8440 * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8441 * alloc_user_cpus_ptr() returns NULL.
8442 */
8443 user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
8444 if (user_mask) {
8445 cpumask_copy(user_mask, in_mask);
8446 } else if (IS_ENABLED(CONFIG_SMP)) {
8447 return -ENOMEM;
8448 }
8449
8450 ac = (struct affinity_context){
8451 .new_mask = in_mask,
8452 .user_mask = user_mask,
8453 .flags = SCA_USER,
8454 };
8455
8456 retval = __sched_setaffinity(p, &ac);
8457 kfree(ac.user_mask);
8458
8459 return retval;
8460 }
8461
get_user_cpu_mask(unsigned long __user * user_mask_ptr,unsigned len,struct cpumask * new_mask)8462 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8463 struct cpumask *new_mask)
8464 {
8465 if (len < cpumask_size())
8466 cpumask_clear(new_mask);
8467 else if (len > cpumask_size())
8468 len = cpumask_size();
8469
8470 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8471 }
8472
8473 /**
8474 * sys_sched_setaffinity - set the CPU affinity of a process
8475 * @pid: pid of the process
8476 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8477 * @user_mask_ptr: user-space pointer to the new CPU mask
8478 *
8479 * Return: 0 on success. An error code otherwise.
8480 */
SYSCALL_DEFINE3(sched_setaffinity,pid_t,pid,unsigned int,len,unsigned long __user *,user_mask_ptr)8481 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8482 unsigned long __user *, user_mask_ptr)
8483 {
8484 cpumask_var_t new_mask;
8485 int retval;
8486
8487 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8488 return -ENOMEM;
8489
8490 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8491 if (retval == 0)
8492 retval = sched_setaffinity(pid, new_mask);
8493 free_cpumask_var(new_mask);
8494 return retval;
8495 }
8496
sched_getaffinity(pid_t pid,struct cpumask * mask)8497 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8498 {
8499 struct task_struct *p;
8500 int retval;
8501
8502 guard(rcu)();
8503 p = find_process_by_pid(pid);
8504 if (!p)
8505 return -ESRCH;
8506
8507 retval = security_task_getscheduler(p);
8508 if (retval)
8509 return retval;
8510
8511 guard(raw_spinlock_irqsave)(&p->pi_lock);
8512 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8513
8514 return 0;
8515 }
8516
8517 /**
8518 * sys_sched_getaffinity - get the CPU affinity of a process
8519 * @pid: pid of the process
8520 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8521 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8522 *
8523 * Return: size of CPU mask copied to user_mask_ptr on success. An
8524 * error code otherwise.
8525 */
SYSCALL_DEFINE3(sched_getaffinity,pid_t,pid,unsigned int,len,unsigned long __user *,user_mask_ptr)8526 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8527 unsigned long __user *, user_mask_ptr)
8528 {
8529 int ret;
8530 cpumask_var_t mask;
8531
8532 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8533 return -EINVAL;
8534 if (len & (sizeof(unsigned long)-1))
8535 return -EINVAL;
8536
8537 if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
8538 return -ENOMEM;
8539
8540 ret = sched_getaffinity(pid, mask);
8541 if (ret == 0) {
8542 unsigned int retlen = min(len, cpumask_size());
8543
8544 if (copy_to_user(user_mask_ptr, cpumask_bits(mask), retlen))
8545 ret = -EFAULT;
8546 else
8547 ret = retlen;
8548 }
8549 free_cpumask_var(mask);
8550
8551 return ret;
8552 }
8553
do_sched_yield(void)8554 static void do_sched_yield(void)
8555 {
8556 struct rq_flags rf;
8557 struct rq *rq;
8558
8559 rq = this_rq_lock_irq(&rf);
8560
8561 schedstat_inc(rq->yld_count);
8562 current->sched_class->yield_task(rq);
8563
8564 preempt_disable();
8565 rq_unlock_irq(rq, &rf);
8566 sched_preempt_enable_no_resched();
8567
8568 schedule();
8569 }
8570
8571 /**
8572 * sys_sched_yield - yield the current processor to other threads.
8573 *
8574 * This function yields the current CPU to other tasks. If there are no
8575 * other threads running on this CPU then this function will return.
8576 *
8577 * Return: 0.
8578 */
SYSCALL_DEFINE0(sched_yield)8579 SYSCALL_DEFINE0(sched_yield)
8580 {
8581 do_sched_yield();
8582 return 0;
8583 }
8584
8585 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
__cond_resched(void)8586 int __sched __cond_resched(void)
8587 {
8588 if (should_resched(0)) {
8589 preempt_schedule_common();
8590 return 1;
8591 }
8592 /*
8593 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8594 * whether the current CPU is in an RCU read-side critical section,
8595 * so the tick can report quiescent states even for CPUs looping
8596 * in kernel context. In contrast, in non-preemptible kernels,
8597 * RCU readers leave no in-memory hints, which means that CPU-bound
8598 * processes executing in kernel context might never report an
8599 * RCU quiescent state. Therefore, the following code causes
8600 * cond_resched() to report a quiescent state, but only when RCU
8601 * is in urgent need of one.
8602 */
8603 #ifndef CONFIG_PREEMPT_RCU
8604 rcu_all_qs();
8605 #endif
8606 return 0;
8607 }
8608 EXPORT_SYMBOL(__cond_resched);
8609 #endif
8610
8611 #ifdef CONFIG_PREEMPT_DYNAMIC
8612 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8613 #define cond_resched_dynamic_enabled __cond_resched
8614 #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8615 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8616 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8617
8618 #define might_resched_dynamic_enabled __cond_resched
8619 #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8620 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8621 EXPORT_STATIC_CALL_TRAMP(might_resched);
8622 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8623 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
dynamic_cond_resched(void)8624 int __sched dynamic_cond_resched(void)
8625 {
8626 klp_sched_try_switch();
8627 if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8628 return 0;
8629 return __cond_resched();
8630 }
8631 EXPORT_SYMBOL(dynamic_cond_resched);
8632
8633 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
dynamic_might_resched(void)8634 int __sched dynamic_might_resched(void)
8635 {
8636 if (!static_branch_unlikely(&sk_dynamic_might_resched))
8637 return 0;
8638 return __cond_resched();
8639 }
8640 EXPORT_SYMBOL(dynamic_might_resched);
8641 #endif
8642 #endif
8643
8644 /*
8645 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8646 * call schedule, and on return reacquire the lock.
8647 *
8648 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8649 * operations here to prevent schedule() from being called twice (once via
8650 * spin_unlock(), once by hand).
8651 */
__cond_resched_lock(spinlock_t * lock)8652 int __cond_resched_lock(spinlock_t *lock)
8653 {
8654 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8655 int ret = 0;
8656
8657 lockdep_assert_held(lock);
8658
8659 if (spin_needbreak(lock) || resched) {
8660 spin_unlock(lock);
8661 if (!_cond_resched())
8662 cpu_relax();
8663 ret = 1;
8664 spin_lock(lock);
8665 }
8666 return ret;
8667 }
8668 EXPORT_SYMBOL(__cond_resched_lock);
8669
__cond_resched_rwlock_read(rwlock_t * lock)8670 int __cond_resched_rwlock_read(rwlock_t *lock)
8671 {
8672 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8673 int ret = 0;
8674
8675 lockdep_assert_held_read(lock);
8676
8677 if (rwlock_needbreak(lock) || resched) {
8678 read_unlock(lock);
8679 if (!_cond_resched())
8680 cpu_relax();
8681 ret = 1;
8682 read_lock(lock);
8683 }
8684 return ret;
8685 }
8686 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8687
__cond_resched_rwlock_write(rwlock_t * lock)8688 int __cond_resched_rwlock_write(rwlock_t *lock)
8689 {
8690 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8691 int ret = 0;
8692
8693 lockdep_assert_held_write(lock);
8694
8695 if (rwlock_needbreak(lock) || resched) {
8696 write_unlock(lock);
8697 if (!_cond_resched())
8698 cpu_relax();
8699 ret = 1;
8700 write_lock(lock);
8701 }
8702 return ret;
8703 }
8704 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8705
8706 #ifdef CONFIG_PREEMPT_DYNAMIC
8707
8708 #ifdef CONFIG_GENERIC_ENTRY
8709 #include <linux/entry-common.h>
8710 #endif
8711
8712 /*
8713 * SC:cond_resched
8714 * SC:might_resched
8715 * SC:preempt_schedule
8716 * SC:preempt_schedule_notrace
8717 * SC:irqentry_exit_cond_resched
8718 *
8719 *
8720 * NONE:
8721 * cond_resched <- __cond_resched
8722 * might_resched <- RET0
8723 * preempt_schedule <- NOP
8724 * preempt_schedule_notrace <- NOP
8725 * irqentry_exit_cond_resched <- NOP
8726 *
8727 * VOLUNTARY:
8728 * cond_resched <- __cond_resched
8729 * might_resched <- __cond_resched
8730 * preempt_schedule <- NOP
8731 * preempt_schedule_notrace <- NOP
8732 * irqentry_exit_cond_resched <- NOP
8733 *
8734 * FULL:
8735 * cond_resched <- RET0
8736 * might_resched <- RET0
8737 * preempt_schedule <- preempt_schedule
8738 * preempt_schedule_notrace <- preempt_schedule_notrace
8739 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8740 */
8741
8742 enum {
8743 preempt_dynamic_undefined = -1,
8744 preempt_dynamic_none,
8745 preempt_dynamic_voluntary,
8746 preempt_dynamic_full,
8747 };
8748
8749 int preempt_dynamic_mode = preempt_dynamic_undefined;
8750
sched_dynamic_mode(const char * str)8751 int sched_dynamic_mode(const char *str)
8752 {
8753 if (!strcmp(str, "none"))
8754 return preempt_dynamic_none;
8755
8756 if (!strcmp(str, "voluntary"))
8757 return preempt_dynamic_voluntary;
8758
8759 if (!strcmp(str, "full"))
8760 return preempt_dynamic_full;
8761
8762 return -EINVAL;
8763 }
8764
8765 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8766 #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8767 #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8768 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8769 #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8770 #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8771 #else
8772 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8773 #endif
8774
8775 static DEFINE_MUTEX(sched_dynamic_mutex);
8776 static bool klp_override;
8777
__sched_dynamic_update(int mode)8778 static void __sched_dynamic_update(int mode)
8779 {
8780 /*
8781 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8782 * the ZERO state, which is invalid.
8783 */
8784 if (!klp_override)
8785 preempt_dynamic_enable(cond_resched);
8786 preempt_dynamic_enable(might_resched);
8787 preempt_dynamic_enable(preempt_schedule);
8788 preempt_dynamic_enable(preempt_schedule_notrace);
8789 preempt_dynamic_enable(irqentry_exit_cond_resched);
8790
8791 switch (mode) {
8792 case preempt_dynamic_none:
8793 if (!klp_override)
8794 preempt_dynamic_enable(cond_resched);
8795 preempt_dynamic_disable(might_resched);
8796 preempt_dynamic_disable(preempt_schedule);
8797 preempt_dynamic_disable(preempt_schedule_notrace);
8798 preempt_dynamic_disable(irqentry_exit_cond_resched);
8799 if (mode != preempt_dynamic_mode)
8800 pr_info("Dynamic Preempt: none\n");
8801 break;
8802
8803 case preempt_dynamic_voluntary:
8804 if (!klp_override)
8805 preempt_dynamic_enable(cond_resched);
8806 preempt_dynamic_enable(might_resched);
8807 preempt_dynamic_disable(preempt_schedule);
8808 preempt_dynamic_disable(preempt_schedule_notrace);
8809 preempt_dynamic_disable(irqentry_exit_cond_resched);
8810 if (mode != preempt_dynamic_mode)
8811 pr_info("Dynamic Preempt: voluntary\n");
8812 break;
8813
8814 case preempt_dynamic_full:
8815 if (!klp_override)
8816 preempt_dynamic_disable(cond_resched);
8817 preempt_dynamic_disable(might_resched);
8818 preempt_dynamic_enable(preempt_schedule);
8819 preempt_dynamic_enable(preempt_schedule_notrace);
8820 preempt_dynamic_enable(irqentry_exit_cond_resched);
8821 if (mode != preempt_dynamic_mode)
8822 pr_info("Dynamic Preempt: full\n");
8823 break;
8824 }
8825
8826 preempt_dynamic_mode = mode;
8827 }
8828
sched_dynamic_update(int mode)8829 void sched_dynamic_update(int mode)
8830 {
8831 mutex_lock(&sched_dynamic_mutex);
8832 __sched_dynamic_update(mode);
8833 mutex_unlock(&sched_dynamic_mutex);
8834 }
8835
8836 #ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
8837
klp_cond_resched(void)8838 static int klp_cond_resched(void)
8839 {
8840 __klp_sched_try_switch();
8841 return __cond_resched();
8842 }
8843
sched_dynamic_klp_enable(void)8844 void sched_dynamic_klp_enable(void)
8845 {
8846 mutex_lock(&sched_dynamic_mutex);
8847
8848 klp_override = true;
8849 static_call_update(cond_resched, klp_cond_resched);
8850
8851 mutex_unlock(&sched_dynamic_mutex);
8852 }
8853
sched_dynamic_klp_disable(void)8854 void sched_dynamic_klp_disable(void)
8855 {
8856 mutex_lock(&sched_dynamic_mutex);
8857
8858 klp_override = false;
8859 __sched_dynamic_update(preempt_dynamic_mode);
8860
8861 mutex_unlock(&sched_dynamic_mutex);
8862 }
8863
8864 #endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
8865
setup_preempt_mode(char * str)8866 static int __init setup_preempt_mode(char *str)
8867 {
8868 int mode = sched_dynamic_mode(str);
8869 if (mode < 0) {
8870 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8871 return 0;
8872 }
8873
8874 sched_dynamic_update(mode);
8875 return 1;
8876 }
8877 __setup("preempt=", setup_preempt_mode);
8878
preempt_dynamic_init(void)8879 static void __init preempt_dynamic_init(void)
8880 {
8881 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8882 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8883 sched_dynamic_update(preempt_dynamic_none);
8884 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8885 sched_dynamic_update(preempt_dynamic_voluntary);
8886 } else {
8887 /* Default static call setting, nothing to do */
8888 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8889 preempt_dynamic_mode = preempt_dynamic_full;
8890 pr_info("Dynamic Preempt: full\n");
8891 }
8892 }
8893 }
8894
8895 #define PREEMPT_MODEL_ACCESSOR(mode) \
8896 bool preempt_model_##mode(void) \
8897 { \
8898 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8899 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8900 } \
8901 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8902
8903 PREEMPT_MODEL_ACCESSOR(none);
8904 PREEMPT_MODEL_ACCESSOR(voluntary);
8905 PREEMPT_MODEL_ACCESSOR(full);
8906
8907 #else /* !CONFIG_PREEMPT_DYNAMIC */
8908
preempt_dynamic_init(void)8909 static inline void preempt_dynamic_init(void) { }
8910
8911 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8912
8913 /**
8914 * yield - yield the current processor to other threads.
8915 *
8916 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8917 *
8918 * The scheduler is at all times free to pick the calling task as the most
8919 * eligible task to run, if removing the yield() call from your code breaks
8920 * it, it's already broken.
8921 *
8922 * Typical broken usage is:
8923 *
8924 * while (!event)
8925 * yield();
8926 *
8927 * where one assumes that yield() will let 'the other' process run that will
8928 * make event true. If the current task is a SCHED_FIFO task that will never
8929 * happen. Never use yield() as a progress guarantee!!
8930 *
8931 * If you want to use yield() to wait for something, use wait_event().
8932 * If you want to use yield() to be 'nice' for others, use cond_resched().
8933 * If you still want to use yield(), do not!
8934 */
yield(void)8935 void __sched yield(void)
8936 {
8937 set_current_state(TASK_RUNNING);
8938 do_sched_yield();
8939 }
8940 EXPORT_SYMBOL(yield);
8941
8942 /**
8943 * yield_to - yield the current processor to another thread in
8944 * your thread group, or accelerate that thread toward the
8945 * processor it's on.
8946 * @p: target task
8947 * @preempt: whether task preemption is allowed or not
8948 *
8949 * It's the caller's job to ensure that the target task struct
8950 * can't go away on us before we can do any checks.
8951 *
8952 * Return:
8953 * true (>0) if we indeed boosted the target task.
8954 * false (0) if we failed to boost the target.
8955 * -ESRCH if there's no task to yield to.
8956 */
yield_to(struct task_struct * p,bool preempt)8957 int __sched yield_to(struct task_struct *p, bool preempt)
8958 {
8959 struct task_struct *curr = current;
8960 struct rq *rq, *p_rq;
8961 int yielded = 0;
8962
8963 scoped_guard (irqsave) {
8964 rq = this_rq();
8965
8966 again:
8967 p_rq = task_rq(p);
8968 /*
8969 * If we're the only runnable task on the rq and target rq also
8970 * has only one task, there's absolutely no point in yielding.
8971 */
8972 if (rq->nr_running == 1 && p_rq->nr_running == 1)
8973 return -ESRCH;
8974
8975 guard(double_rq_lock)(rq, p_rq);
8976 if (task_rq(p) != p_rq)
8977 goto again;
8978
8979 if (!curr->sched_class->yield_to_task)
8980 return 0;
8981
8982 if (curr->sched_class != p->sched_class)
8983 return 0;
8984
8985 if (task_on_cpu(p_rq, p) || !task_is_running(p))
8986 return 0;
8987
8988 yielded = curr->sched_class->yield_to_task(rq, p);
8989 if (yielded) {
8990 schedstat_inc(rq->yld_count);
8991 /*
8992 * Make p's CPU reschedule; pick_next_entity
8993 * takes care of fairness.
8994 */
8995 if (preempt && rq != p_rq)
8996 resched_curr(p_rq);
8997 }
8998 }
8999
9000 if (yielded)
9001 schedule();
9002
9003 return yielded;
9004 }
9005 EXPORT_SYMBOL_GPL(yield_to);
9006
io_schedule_prepare(void)9007 int io_schedule_prepare(void)
9008 {
9009 int old_iowait = current->in_iowait;
9010
9011 current->in_iowait = 1;
9012 blk_flush_plug(current->plug, true);
9013 return old_iowait;
9014 }
9015
io_schedule_finish(int token)9016 void io_schedule_finish(int token)
9017 {
9018 current->in_iowait = token;
9019 }
9020
9021 /*
9022 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
9023 * that process accounting knows that this is a task in IO wait state.
9024 */
io_schedule_timeout(long timeout)9025 long __sched io_schedule_timeout(long timeout)
9026 {
9027 int token;
9028 long ret;
9029
9030 token = io_schedule_prepare();
9031 ret = schedule_timeout(timeout);
9032 io_schedule_finish(token);
9033
9034 return ret;
9035 }
9036 EXPORT_SYMBOL(io_schedule_timeout);
9037
io_schedule(void)9038 void __sched io_schedule(void)
9039 {
9040 int token;
9041
9042 token = io_schedule_prepare();
9043 schedule();
9044 io_schedule_finish(token);
9045 }
9046 EXPORT_SYMBOL(io_schedule);
9047
9048 /**
9049 * sys_sched_get_priority_max - return maximum RT priority.
9050 * @policy: scheduling class.
9051 *
9052 * Return: On success, this syscall returns the maximum
9053 * rt_priority that can be used by a given scheduling class.
9054 * On failure, a negative error code is returned.
9055 */
SYSCALL_DEFINE1(sched_get_priority_max,int,policy)9056 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
9057 {
9058 int ret = -EINVAL;
9059
9060 switch (policy) {
9061 case SCHED_FIFO:
9062 case SCHED_RR:
9063 ret = MAX_RT_PRIO-1;
9064 break;
9065 case SCHED_DEADLINE:
9066 case SCHED_NORMAL:
9067 case SCHED_BATCH:
9068 case SCHED_IDLE:
9069 ret = 0;
9070 break;
9071 }
9072 return ret;
9073 }
9074
9075 /**
9076 * sys_sched_get_priority_min - return minimum RT priority.
9077 * @policy: scheduling class.
9078 *
9079 * Return: On success, this syscall returns the minimum
9080 * rt_priority that can be used by a given scheduling class.
9081 * On failure, a negative error code is returned.
9082 */
SYSCALL_DEFINE1(sched_get_priority_min,int,policy)9083 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
9084 {
9085 int ret = -EINVAL;
9086
9087 switch (policy) {
9088 case SCHED_FIFO:
9089 case SCHED_RR:
9090 ret = 1;
9091 break;
9092 case SCHED_DEADLINE:
9093 case SCHED_NORMAL:
9094 case SCHED_BATCH:
9095 case SCHED_IDLE:
9096 ret = 0;
9097 }
9098 return ret;
9099 }
9100
sched_rr_get_interval(pid_t pid,struct timespec64 * t)9101 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
9102 {
9103 unsigned int time_slice = 0;
9104 int retval;
9105
9106 if (pid < 0)
9107 return -EINVAL;
9108
9109 scoped_guard (rcu) {
9110 struct task_struct *p = find_process_by_pid(pid);
9111 if (!p)
9112 return -ESRCH;
9113
9114 retval = security_task_getscheduler(p);
9115 if (retval)
9116 return retval;
9117
9118 scoped_guard (task_rq_lock, p) {
9119 struct rq *rq = scope.rq;
9120 if (p->sched_class->get_rr_interval)
9121 time_slice = p->sched_class->get_rr_interval(rq, p);
9122 }
9123 }
9124
9125 jiffies_to_timespec64(time_slice, t);
9126 return 0;
9127 }
9128
9129 /**
9130 * sys_sched_rr_get_interval - return the default timeslice of a process.
9131 * @pid: pid of the process.
9132 * @interval: userspace pointer to the timeslice value.
9133 *
9134 * this syscall writes the default timeslice value of a given process
9135 * into the user-space timespec buffer. A value of '0' means infinity.
9136 *
9137 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
9138 * an error code.
9139 */
SYSCALL_DEFINE2(sched_rr_get_interval,pid_t,pid,struct __kernel_timespec __user *,interval)9140 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
9141 struct __kernel_timespec __user *, interval)
9142 {
9143 struct timespec64 t;
9144 int retval = sched_rr_get_interval(pid, &t);
9145
9146 if (retval == 0)
9147 retval = put_timespec64(&t, interval);
9148
9149 return retval;
9150 }
9151
9152 #ifdef CONFIG_COMPAT_32BIT_TIME
SYSCALL_DEFINE2(sched_rr_get_interval_time32,pid_t,pid,struct old_timespec32 __user *,interval)9153 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
9154 struct old_timespec32 __user *, interval)
9155 {
9156 struct timespec64 t;
9157 int retval = sched_rr_get_interval(pid, &t);
9158
9159 if (retval == 0)
9160 retval = put_old_timespec32(&t, interval);
9161 return retval;
9162 }
9163 #endif
9164
sched_show_task(struct task_struct * p)9165 void sched_show_task(struct task_struct *p)
9166 {
9167 unsigned long free = 0;
9168 int ppid;
9169
9170 if (!try_get_task_stack(p))
9171 return;
9172
9173 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
9174
9175 if (task_is_running(p))
9176 pr_cont(" running task ");
9177 #ifdef CONFIG_DEBUG_STACK_USAGE
9178 free = stack_not_used(p);
9179 #endif
9180 ppid = 0;
9181 rcu_read_lock();
9182 if (pid_alive(p))
9183 ppid = task_pid_nr(rcu_dereference(p->real_parent));
9184 rcu_read_unlock();
9185 pr_cont(" stack:%-5lu pid:%-5d tgid:%-5d ppid:%-6d flags:0x%08lx\n",
9186 free, task_pid_nr(p), task_tgid_nr(p),
9187 ppid, read_task_thread_flags(p));
9188
9189 print_worker_info(KERN_INFO, p);
9190 print_stop_info(KERN_INFO, p);
9191 show_stack(p, NULL, KERN_INFO);
9192 put_task_stack(p);
9193 }
9194 EXPORT_SYMBOL_GPL(sched_show_task);
9195
9196 static inline bool
state_filter_match(unsigned long state_filter,struct task_struct * p)9197 state_filter_match(unsigned long state_filter, struct task_struct *p)
9198 {
9199 unsigned int state = READ_ONCE(p->__state);
9200
9201 /* no filter, everything matches */
9202 if (!state_filter)
9203 return true;
9204
9205 /* filter, but doesn't match */
9206 if (!(state & state_filter))
9207 return false;
9208
9209 /*
9210 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9211 * TASK_KILLABLE).
9212 */
9213 if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
9214 return false;
9215
9216 return true;
9217 }
9218
9219
show_state_filter(unsigned int state_filter)9220 void show_state_filter(unsigned int state_filter)
9221 {
9222 struct task_struct *g, *p;
9223
9224 rcu_read_lock();
9225 for_each_process_thread(g, p) {
9226 /*
9227 * reset the NMI-timeout, listing all files on a slow
9228 * console might take a lot of time:
9229 * Also, reset softlockup watchdogs on all CPUs, because
9230 * another CPU might be blocked waiting for us to process
9231 * an IPI.
9232 */
9233 touch_nmi_watchdog();
9234 touch_all_softlockup_watchdogs();
9235 if (state_filter_match(state_filter, p))
9236 sched_show_task(p);
9237 }
9238
9239 #ifdef CONFIG_SCHED_DEBUG
9240 if (!state_filter)
9241 sysrq_sched_debug_show();
9242 #endif
9243 rcu_read_unlock();
9244 /*
9245 * Only show locks if all tasks are dumped:
9246 */
9247 if (!state_filter)
9248 debug_show_all_locks();
9249 }
9250
9251 /**
9252 * init_idle - set up an idle thread for a given CPU
9253 * @idle: task in question
9254 * @cpu: CPU the idle task belongs to
9255 *
9256 * NOTE: this function does not set the idle thread's NEED_RESCHED
9257 * flag, to make booting more robust.
9258 */
init_idle(struct task_struct * idle,int cpu)9259 void __init init_idle(struct task_struct *idle, int cpu)
9260 {
9261 #ifdef CONFIG_SMP
9262 struct affinity_context ac = (struct affinity_context) {
9263 .new_mask = cpumask_of(cpu),
9264 .flags = 0,
9265 };
9266 #endif
9267 struct rq *rq = cpu_rq(cpu);
9268 unsigned long flags;
9269
9270 __sched_fork(0, idle);
9271
9272 raw_spin_lock_irqsave(&idle->pi_lock, flags);
9273 raw_spin_rq_lock(rq);
9274
9275 idle->__state = TASK_RUNNING;
9276 idle->se.exec_start = sched_clock();
9277 /*
9278 * PF_KTHREAD should already be set at this point; regardless, make it
9279 * look like a proper per-CPU kthread.
9280 */
9281 idle->flags |= PF_KTHREAD | PF_NO_SETAFFINITY;
9282 kthread_set_per_cpu(idle, cpu);
9283
9284 #ifdef CONFIG_SMP
9285 /*
9286 * It's possible that init_idle() gets called multiple times on a task,
9287 * in that case do_set_cpus_allowed() will not do the right thing.
9288 *
9289 * And since this is boot we can forgo the serialization.
9290 */
9291 set_cpus_allowed_common(idle, &ac);
9292 #endif
9293 /*
9294 * We're having a chicken and egg problem, even though we are
9295 * holding rq->lock, the CPU isn't yet set to this CPU so the
9296 * lockdep check in task_group() will fail.
9297 *
9298 * Similar case to sched_fork(). / Alternatively we could
9299 * use task_rq_lock() here and obtain the other rq->lock.
9300 *
9301 * Silence PROVE_RCU
9302 */
9303 rcu_read_lock();
9304 __set_task_cpu(idle, cpu);
9305 rcu_read_unlock();
9306
9307 rq->idle = idle;
9308 rcu_assign_pointer(rq->curr, idle);
9309 idle->on_rq = TASK_ON_RQ_QUEUED;
9310 #ifdef CONFIG_SMP
9311 idle->on_cpu = 1;
9312 #endif
9313 raw_spin_rq_unlock(rq);
9314 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
9315
9316 /* Set the preempt count _outside_ the spinlocks! */
9317 init_idle_preempt_count(idle, cpu);
9318
9319 /*
9320 * The idle tasks have their own, simple scheduling class:
9321 */
9322 idle->sched_class = &idle_sched_class;
9323 ftrace_graph_init_idle_task(idle, cpu);
9324 vtime_init_idle(idle, cpu);
9325 #ifdef CONFIG_SMP
9326 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9327 #endif
9328 }
9329
9330 #ifdef CONFIG_SMP
9331
cpuset_cpumask_can_shrink(const struct cpumask * cur,const struct cpumask * trial)9332 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9333 const struct cpumask *trial)
9334 {
9335 int ret = 1;
9336
9337 if (cpumask_empty(cur))
9338 return ret;
9339
9340 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9341
9342 return ret;
9343 }
9344
task_can_attach(struct task_struct * p)9345 int task_can_attach(struct task_struct *p)
9346 {
9347 int ret = 0;
9348
9349 /*
9350 * Kthreads which disallow setaffinity shouldn't be moved
9351 * to a new cpuset; we don't want to change their CPU
9352 * affinity and isolating such threads by their set of
9353 * allowed nodes is unnecessary. Thus, cpusets are not
9354 * applicable for such threads. This prevents checking for
9355 * success of set_cpus_allowed_ptr() on all attached tasks
9356 * before cpus_mask may be changed.
9357 */
9358 if (p->flags & PF_NO_SETAFFINITY)
9359 ret = -EINVAL;
9360
9361 return ret;
9362 }
9363
9364 bool sched_smp_initialized __read_mostly;
9365
9366 #ifdef CONFIG_NUMA_BALANCING
9367 /* Migrate current task p to target_cpu */
migrate_task_to(struct task_struct * p,int target_cpu)9368 int migrate_task_to(struct task_struct *p, int target_cpu)
9369 {
9370 struct migration_arg arg = { p, target_cpu };
9371 int curr_cpu = task_cpu(p);
9372
9373 if (curr_cpu == target_cpu)
9374 return 0;
9375
9376 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9377 return -EINVAL;
9378
9379 /* TODO: This is not properly updating schedstats */
9380
9381 trace_sched_move_numa(p, curr_cpu, target_cpu);
9382 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9383 }
9384
9385 /*
9386 * Requeue a task on a given node and accurately track the number of NUMA
9387 * tasks on the runqueues
9388 */
sched_setnuma(struct task_struct * p,int nid)9389 void sched_setnuma(struct task_struct *p, int nid)
9390 {
9391 bool queued, running;
9392 struct rq_flags rf;
9393 struct rq *rq;
9394
9395 rq = task_rq_lock(p, &rf);
9396 queued = task_on_rq_queued(p);
9397 running = task_current(rq, p);
9398
9399 if (queued)
9400 dequeue_task(rq, p, DEQUEUE_SAVE);
9401 if (running)
9402 put_prev_task(rq, p);
9403
9404 p->numa_preferred_nid = nid;
9405
9406 if (queued)
9407 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9408 if (running)
9409 set_next_task(rq, p);
9410 task_rq_unlock(rq, p, &rf);
9411 }
9412 #endif /* CONFIG_NUMA_BALANCING */
9413
9414 #ifdef CONFIG_HOTPLUG_CPU
9415 /*
9416 * Ensure that the idle task is using init_mm right before its CPU goes
9417 * offline.
9418 */
idle_task_exit(void)9419 void idle_task_exit(void)
9420 {
9421 struct mm_struct *mm = current->active_mm;
9422
9423 BUG_ON(cpu_online(smp_processor_id()));
9424 BUG_ON(current != this_rq()->idle);
9425
9426 if (mm != &init_mm) {
9427 switch_mm(mm, &init_mm, current);
9428 finish_arch_post_lock_switch();
9429 }
9430
9431 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9432 }
9433
__balance_push_cpu_stop(void * arg)9434 static int __balance_push_cpu_stop(void *arg)
9435 {
9436 struct task_struct *p = arg;
9437 struct rq *rq = this_rq();
9438 struct rq_flags rf;
9439 int cpu;
9440
9441 raw_spin_lock_irq(&p->pi_lock);
9442 rq_lock(rq, &rf);
9443
9444 update_rq_clock(rq);
9445
9446 if (task_rq(p) == rq && task_on_rq_queued(p)) {
9447 cpu = select_fallback_rq(rq->cpu, p);
9448 rq = __migrate_task(rq, &rf, p, cpu);
9449 }
9450
9451 rq_unlock(rq, &rf);
9452 raw_spin_unlock_irq(&p->pi_lock);
9453
9454 put_task_struct(p);
9455
9456 return 0;
9457 }
9458
9459 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9460
9461 /*
9462 * Ensure we only run per-cpu kthreads once the CPU goes !active.
9463 *
9464 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9465 * effective when the hotplug motion is down.
9466 */
balance_push(struct rq * rq)9467 static void balance_push(struct rq *rq)
9468 {
9469 struct task_struct *push_task = rq->curr;
9470
9471 lockdep_assert_rq_held(rq);
9472
9473 /*
9474 * Ensure the thing is persistent until balance_push_set(.on = false);
9475 */
9476 rq->balance_callback = &balance_push_callback;
9477
9478 /*
9479 * Only active while going offline and when invoked on the outgoing
9480 * CPU.
9481 */
9482 if (!cpu_dying(rq->cpu) || rq != this_rq())
9483 return;
9484
9485 /*
9486 * Both the cpu-hotplug and stop task are in this case and are
9487 * required to complete the hotplug process.
9488 */
9489 if (kthread_is_per_cpu(push_task) ||
9490 is_migration_disabled(push_task)) {
9491
9492 /*
9493 * If this is the idle task on the outgoing CPU try to wake
9494 * up the hotplug control thread which might wait for the
9495 * last task to vanish. The rcuwait_active() check is
9496 * accurate here because the waiter is pinned on this CPU
9497 * and can't obviously be running in parallel.
9498 *
9499 * On RT kernels this also has to check whether there are
9500 * pinned and scheduled out tasks on the runqueue. They
9501 * need to leave the migrate disabled section first.
9502 */
9503 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9504 rcuwait_active(&rq->hotplug_wait)) {
9505 raw_spin_rq_unlock(rq);
9506 rcuwait_wake_up(&rq->hotplug_wait);
9507 raw_spin_rq_lock(rq);
9508 }
9509 return;
9510 }
9511
9512 get_task_struct(push_task);
9513 /*
9514 * Temporarily drop rq->lock such that we can wake-up the stop task.
9515 * Both preemption and IRQs are still disabled.
9516 */
9517 preempt_disable();
9518 raw_spin_rq_unlock(rq);
9519 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9520 this_cpu_ptr(&push_work));
9521 preempt_enable();
9522 /*
9523 * At this point need_resched() is true and we'll take the loop in
9524 * schedule(). The next pick is obviously going to be the stop task
9525 * which kthread_is_per_cpu() and will push this task away.
9526 */
9527 raw_spin_rq_lock(rq);
9528 }
9529
balance_push_set(int cpu,bool on)9530 static void balance_push_set(int cpu, bool on)
9531 {
9532 struct rq *rq = cpu_rq(cpu);
9533 struct rq_flags rf;
9534
9535 rq_lock_irqsave(rq, &rf);
9536 if (on) {
9537 WARN_ON_ONCE(rq->balance_callback);
9538 rq->balance_callback = &balance_push_callback;
9539 } else if (rq->balance_callback == &balance_push_callback) {
9540 rq->balance_callback = NULL;
9541 }
9542 rq_unlock_irqrestore(rq, &rf);
9543 }
9544
9545 /*
9546 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9547 * inactive. All tasks which are not per CPU kernel threads are either
9548 * pushed off this CPU now via balance_push() or placed on a different CPU
9549 * during wakeup. Wait until the CPU is quiescent.
9550 */
balance_hotplug_wait(void)9551 static void balance_hotplug_wait(void)
9552 {
9553 struct rq *rq = this_rq();
9554
9555 rcuwait_wait_event(&rq->hotplug_wait,
9556 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9557 TASK_UNINTERRUPTIBLE);
9558 }
9559
9560 #else
9561
balance_push(struct rq * rq)9562 static inline void balance_push(struct rq *rq)
9563 {
9564 }
9565
balance_push_set(int cpu,bool on)9566 static inline void balance_push_set(int cpu, bool on)
9567 {
9568 }
9569
balance_hotplug_wait(void)9570 static inline void balance_hotplug_wait(void)
9571 {
9572 }
9573
9574 #endif /* CONFIG_HOTPLUG_CPU */
9575
set_rq_online(struct rq * rq)9576 void set_rq_online(struct rq *rq)
9577 {
9578 if (!rq->online) {
9579 const struct sched_class *class;
9580
9581 cpumask_set_cpu(rq->cpu, rq->rd->online);
9582 rq->online = 1;
9583
9584 for_each_class(class) {
9585 if (class->rq_online)
9586 class->rq_online(rq);
9587 }
9588 }
9589 }
9590
set_rq_offline(struct rq * rq)9591 void set_rq_offline(struct rq *rq)
9592 {
9593 if (rq->online) {
9594 const struct sched_class *class;
9595
9596 update_rq_clock(rq);
9597 for_each_class(class) {
9598 if (class->rq_offline)
9599 class->rq_offline(rq);
9600 }
9601
9602 cpumask_clear_cpu(rq->cpu, rq->rd->online);
9603 rq->online = 0;
9604 }
9605 }
9606
9607 /*
9608 * used to mark begin/end of suspend/resume:
9609 */
9610 static int num_cpus_frozen;
9611
9612 /*
9613 * Update cpusets according to cpu_active mask. If cpusets are
9614 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9615 * around partition_sched_domains().
9616 *
9617 * If we come here as part of a suspend/resume, don't touch cpusets because we
9618 * want to restore it back to its original state upon resume anyway.
9619 */
cpuset_cpu_active(void)9620 static void cpuset_cpu_active(void)
9621 {
9622 if (cpuhp_tasks_frozen) {
9623 /*
9624 * num_cpus_frozen tracks how many CPUs are involved in suspend
9625 * resume sequence. As long as this is not the last online
9626 * operation in the resume sequence, just build a single sched
9627 * domain, ignoring cpusets.
9628 */
9629 partition_sched_domains(1, NULL, NULL);
9630 if (--num_cpus_frozen)
9631 return;
9632 /*
9633 * This is the last CPU online operation. So fall through and
9634 * restore the original sched domains by considering the
9635 * cpuset configurations.
9636 */
9637 cpuset_force_rebuild();
9638 }
9639 cpuset_update_active_cpus();
9640 }
9641
cpuset_cpu_inactive(unsigned int cpu)9642 static int cpuset_cpu_inactive(unsigned int cpu)
9643 {
9644 if (!cpuhp_tasks_frozen) {
9645 int ret = dl_bw_check_overflow(cpu);
9646
9647 if (ret)
9648 return ret;
9649 cpuset_update_active_cpus();
9650 } else {
9651 num_cpus_frozen++;
9652 partition_sched_domains(1, NULL, NULL);
9653 }
9654 return 0;
9655 }
9656
sched_cpu_activate(unsigned int cpu)9657 int sched_cpu_activate(unsigned int cpu)
9658 {
9659 struct rq *rq = cpu_rq(cpu);
9660 struct rq_flags rf;
9661
9662 /*
9663 * Clear the balance_push callback and prepare to schedule
9664 * regular tasks.
9665 */
9666 balance_push_set(cpu, false);
9667
9668 #ifdef CONFIG_SCHED_SMT
9669 /*
9670 * When going up, increment the number of cores with SMT present.
9671 */
9672 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9673 static_branch_inc_cpuslocked(&sched_smt_present);
9674 #endif
9675 set_cpu_active(cpu, true);
9676
9677 if (sched_smp_initialized) {
9678 sched_update_numa(cpu, true);
9679 sched_domains_numa_masks_set(cpu);
9680 cpuset_cpu_active();
9681 }
9682
9683 /*
9684 * Put the rq online, if not already. This happens:
9685 *
9686 * 1) In the early boot process, because we build the real domains
9687 * after all CPUs have been brought up.
9688 *
9689 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9690 * domains.
9691 */
9692 rq_lock_irqsave(rq, &rf);
9693 if (rq->rd) {
9694 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9695 set_rq_online(rq);
9696 }
9697 rq_unlock_irqrestore(rq, &rf);
9698
9699 return 0;
9700 }
9701
sched_cpu_deactivate(unsigned int cpu)9702 int sched_cpu_deactivate(unsigned int cpu)
9703 {
9704 struct rq *rq = cpu_rq(cpu);
9705 struct rq_flags rf;
9706 int ret;
9707
9708 /*
9709 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9710 * load balancing when not active
9711 */
9712 nohz_balance_exit_idle(rq);
9713
9714 set_cpu_active(cpu, false);
9715
9716 /*
9717 * From this point forward, this CPU will refuse to run any task that
9718 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9719 * push those tasks away until this gets cleared, see
9720 * sched_cpu_dying().
9721 */
9722 balance_push_set(cpu, true);
9723
9724 /*
9725 * We've cleared cpu_active_mask / set balance_push, wait for all
9726 * preempt-disabled and RCU users of this state to go away such that
9727 * all new such users will observe it.
9728 *
9729 * Specifically, we rely on ttwu to no longer target this CPU, see
9730 * ttwu_queue_cond() and is_cpu_allowed().
9731 *
9732 * Do sync before park smpboot threads to take care the rcu boost case.
9733 */
9734 synchronize_rcu();
9735
9736 rq_lock_irqsave(rq, &rf);
9737 if (rq->rd) {
9738 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9739 set_rq_offline(rq);
9740 }
9741 rq_unlock_irqrestore(rq, &rf);
9742
9743 #ifdef CONFIG_SCHED_SMT
9744 /*
9745 * When going down, decrement the number of cores with SMT present.
9746 */
9747 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9748 static_branch_dec_cpuslocked(&sched_smt_present);
9749
9750 sched_core_cpu_deactivate(cpu);
9751 #endif
9752
9753 if (!sched_smp_initialized)
9754 return 0;
9755
9756 sched_update_numa(cpu, false);
9757 ret = cpuset_cpu_inactive(cpu);
9758 if (ret) {
9759 balance_push_set(cpu, false);
9760 set_cpu_active(cpu, true);
9761 sched_update_numa(cpu, true);
9762 return ret;
9763 }
9764 sched_domains_numa_masks_clear(cpu);
9765 return 0;
9766 }
9767
sched_rq_cpu_starting(unsigned int cpu)9768 static void sched_rq_cpu_starting(unsigned int cpu)
9769 {
9770 struct rq *rq = cpu_rq(cpu);
9771
9772 rq->calc_load_update = calc_load_update;
9773 update_max_interval();
9774 }
9775
sched_cpu_starting(unsigned int cpu)9776 int sched_cpu_starting(unsigned int cpu)
9777 {
9778 sched_core_cpu_starting(cpu);
9779 sched_rq_cpu_starting(cpu);
9780 sched_tick_start(cpu);
9781 return 0;
9782 }
9783
9784 #ifdef CONFIG_HOTPLUG_CPU
9785
9786 /*
9787 * Invoked immediately before the stopper thread is invoked to bring the
9788 * CPU down completely. At this point all per CPU kthreads except the
9789 * hotplug thread (current) and the stopper thread (inactive) have been
9790 * either parked or have been unbound from the outgoing CPU. Ensure that
9791 * any of those which might be on the way out are gone.
9792 *
9793 * If after this point a bound task is being woken on this CPU then the
9794 * responsible hotplug callback has failed to do it's job.
9795 * sched_cpu_dying() will catch it with the appropriate fireworks.
9796 */
sched_cpu_wait_empty(unsigned int cpu)9797 int sched_cpu_wait_empty(unsigned int cpu)
9798 {
9799 balance_hotplug_wait();
9800 return 0;
9801 }
9802
9803 /*
9804 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9805 * might have. Called from the CPU stopper task after ensuring that the
9806 * stopper is the last running task on the CPU, so nr_active count is
9807 * stable. We need to take the teardown thread which is calling this into
9808 * account, so we hand in adjust = 1 to the load calculation.
9809 *
9810 * Also see the comment "Global load-average calculations".
9811 */
calc_load_migrate(struct rq * rq)9812 static void calc_load_migrate(struct rq *rq)
9813 {
9814 long delta = calc_load_fold_active(rq, 1);
9815
9816 if (delta)
9817 atomic_long_add(delta, &calc_load_tasks);
9818 }
9819
dump_rq_tasks(struct rq * rq,const char * loglvl)9820 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9821 {
9822 struct task_struct *g, *p;
9823 int cpu = cpu_of(rq);
9824
9825 lockdep_assert_rq_held(rq);
9826
9827 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9828 for_each_process_thread(g, p) {
9829 if (task_cpu(p) != cpu)
9830 continue;
9831
9832 if (!task_on_rq_queued(p))
9833 continue;
9834
9835 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9836 }
9837 }
9838
sched_cpu_dying(unsigned int cpu)9839 int sched_cpu_dying(unsigned int cpu)
9840 {
9841 struct rq *rq = cpu_rq(cpu);
9842 struct rq_flags rf;
9843
9844 /* Handle pending wakeups and then migrate everything off */
9845 sched_tick_stop(cpu);
9846
9847 rq_lock_irqsave(rq, &rf);
9848 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9849 WARN(true, "Dying CPU not properly vacated!");
9850 dump_rq_tasks(rq, KERN_WARNING);
9851 }
9852 rq_unlock_irqrestore(rq, &rf);
9853
9854 calc_load_migrate(rq);
9855 update_max_interval();
9856 hrtick_clear(rq);
9857 sched_core_cpu_dying(cpu);
9858 return 0;
9859 }
9860 #endif
9861
sched_init_smp(void)9862 void __init sched_init_smp(void)
9863 {
9864 sched_init_numa(NUMA_NO_NODE);
9865
9866 /*
9867 * There's no userspace yet to cause hotplug operations; hence all the
9868 * CPU masks are stable and all blatant races in the below code cannot
9869 * happen.
9870 */
9871 mutex_lock(&sched_domains_mutex);
9872 sched_init_domains(cpu_active_mask);
9873 mutex_unlock(&sched_domains_mutex);
9874
9875 /* Move init over to a non-isolated CPU */
9876 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9877 BUG();
9878 current->flags &= ~PF_NO_SETAFFINITY;
9879 sched_init_granularity();
9880
9881 init_sched_rt_class();
9882 init_sched_dl_class();
9883
9884 sched_smp_initialized = true;
9885 }
9886
migration_init(void)9887 static int __init migration_init(void)
9888 {
9889 sched_cpu_starting(smp_processor_id());
9890 return 0;
9891 }
9892 early_initcall(migration_init);
9893
9894 #else
sched_init_smp(void)9895 void __init sched_init_smp(void)
9896 {
9897 sched_init_granularity();
9898 }
9899 #endif /* CONFIG_SMP */
9900
in_sched_functions(unsigned long addr)9901 int in_sched_functions(unsigned long addr)
9902 {
9903 return in_lock_functions(addr) ||
9904 (addr >= (unsigned long)__sched_text_start
9905 && addr < (unsigned long)__sched_text_end);
9906 }
9907
9908 #ifdef CONFIG_CGROUP_SCHED
9909 /*
9910 * Default task group.
9911 * Every task in system belongs to this group at bootup.
9912 */
9913 struct task_group root_task_group;
9914 LIST_HEAD(task_groups);
9915
9916 /* Cacheline aligned slab cache for task_group */
9917 static struct kmem_cache *task_group_cache __ro_after_init;
9918 #endif
9919
sched_init(void)9920 void __init sched_init(void)
9921 {
9922 unsigned long ptr = 0;
9923 int i;
9924
9925 /* Make sure the linker didn't screw up */
9926 BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9927 &fair_sched_class != &rt_sched_class + 1 ||
9928 &rt_sched_class != &dl_sched_class + 1);
9929 #ifdef CONFIG_SMP
9930 BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9931 #endif
9932
9933 wait_bit_init();
9934
9935 #ifdef CONFIG_FAIR_GROUP_SCHED
9936 ptr += 2 * nr_cpu_ids * sizeof(void **);
9937 #endif
9938 #ifdef CONFIG_RT_GROUP_SCHED
9939 ptr += 2 * nr_cpu_ids * sizeof(void **);
9940 #endif
9941 if (ptr) {
9942 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9943
9944 #ifdef CONFIG_FAIR_GROUP_SCHED
9945 root_task_group.se = (struct sched_entity **)ptr;
9946 ptr += nr_cpu_ids * sizeof(void **);
9947
9948 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9949 ptr += nr_cpu_ids * sizeof(void **);
9950
9951 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9952 init_cfs_bandwidth(&root_task_group.cfs_bandwidth, NULL);
9953 #endif /* CONFIG_FAIR_GROUP_SCHED */
9954 #ifdef CONFIG_RT_GROUP_SCHED
9955 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9956 ptr += nr_cpu_ids * sizeof(void **);
9957
9958 root_task_group.rt_rq = (struct rt_rq **)ptr;
9959 ptr += nr_cpu_ids * sizeof(void **);
9960
9961 #endif /* CONFIG_RT_GROUP_SCHED */
9962 }
9963
9964 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9965
9966 #ifdef CONFIG_SMP
9967 init_defrootdomain();
9968 #endif
9969
9970 #ifdef CONFIG_RT_GROUP_SCHED
9971 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9972 global_rt_period(), global_rt_runtime());
9973 #endif /* CONFIG_RT_GROUP_SCHED */
9974
9975 #ifdef CONFIG_CGROUP_SCHED
9976 task_group_cache = KMEM_CACHE(task_group, 0);
9977
9978 list_add(&root_task_group.list, &task_groups);
9979 INIT_LIST_HEAD(&root_task_group.children);
9980 INIT_LIST_HEAD(&root_task_group.siblings);
9981 autogroup_init(&init_task);
9982 #endif /* CONFIG_CGROUP_SCHED */
9983
9984 for_each_possible_cpu(i) {
9985 struct rq *rq;
9986
9987 rq = cpu_rq(i);
9988 raw_spin_lock_init(&rq->__lock);
9989 rq->nr_running = 0;
9990 rq->calc_load_active = 0;
9991 rq->calc_load_update = jiffies + LOAD_FREQ;
9992 init_cfs_rq(&rq->cfs);
9993 init_rt_rq(&rq->rt);
9994 init_dl_rq(&rq->dl);
9995 #ifdef CONFIG_FAIR_GROUP_SCHED
9996 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9997 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9998 /*
9999 * How much CPU bandwidth does root_task_group get?
10000 *
10001 * In case of task-groups formed thr' the cgroup filesystem, it
10002 * gets 100% of the CPU resources in the system. This overall
10003 * system CPU resource is divided among the tasks of
10004 * root_task_group and its child task-groups in a fair manner,
10005 * based on each entity's (task or task-group's) weight
10006 * (se->load.weight).
10007 *
10008 * In other words, if root_task_group has 10 tasks of weight
10009 * 1024) and two child groups A0 and A1 (of weight 1024 each),
10010 * then A0's share of the CPU resource is:
10011 *
10012 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
10013 *
10014 * We achieve this by letting root_task_group's tasks sit
10015 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
10016 */
10017 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
10018 #endif /* CONFIG_FAIR_GROUP_SCHED */
10019
10020 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
10021 #ifdef CONFIG_RT_GROUP_SCHED
10022 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
10023 #endif
10024 #ifdef CONFIG_SMP
10025 rq->sd = NULL;
10026 rq->rd = NULL;
10027 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
10028 rq->balance_callback = &balance_push_callback;
10029 rq->active_balance = 0;
10030 rq->next_balance = jiffies;
10031 rq->push_cpu = 0;
10032 rq->cpu = i;
10033 rq->online = 0;
10034 rq->idle_stamp = 0;
10035 rq->avg_idle = 2*sysctl_sched_migration_cost;
10036 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
10037
10038 INIT_LIST_HEAD(&rq->cfs_tasks);
10039
10040 rq_attach_root(rq, &def_root_domain);
10041 #ifdef CONFIG_NO_HZ_COMMON
10042 rq->last_blocked_load_update_tick = jiffies;
10043 atomic_set(&rq->nohz_flags, 0);
10044
10045 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
10046 #endif
10047 #ifdef CONFIG_HOTPLUG_CPU
10048 rcuwait_init(&rq->hotplug_wait);
10049 #endif
10050 #endif /* CONFIG_SMP */
10051 hrtick_rq_init(rq);
10052 atomic_set(&rq->nr_iowait, 0);
10053
10054 #ifdef CONFIG_SCHED_CORE
10055 rq->core = rq;
10056 rq->core_pick = NULL;
10057 rq->core_enabled = 0;
10058 rq->core_tree = RB_ROOT;
10059 rq->core_forceidle_count = 0;
10060 rq->core_forceidle_occupation = 0;
10061 rq->core_forceidle_start = 0;
10062
10063 rq->core_cookie = 0UL;
10064 #endif
10065 zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
10066 }
10067
10068 set_load_weight(&init_task, false);
10069
10070 /*
10071 * The boot idle thread does lazy MMU switching as well:
10072 */
10073 mmgrab_lazy_tlb(&init_mm);
10074 enter_lazy_tlb(&init_mm, current);
10075
10076 /*
10077 * The idle task doesn't need the kthread struct to function, but it
10078 * is dressed up as a per-CPU kthread and thus needs to play the part
10079 * if we want to avoid special-casing it in code that deals with per-CPU
10080 * kthreads.
10081 */
10082 WARN_ON(!set_kthread_struct(current));
10083
10084 /*
10085 * Make us the idle thread. Technically, schedule() should not be
10086 * called from this thread, however somewhere below it might be,
10087 * but because we are the idle thread, we just pick up running again
10088 * when this runqueue becomes "idle".
10089 */
10090 init_idle(current, smp_processor_id());
10091
10092 calc_load_update = jiffies + LOAD_FREQ;
10093
10094 #ifdef CONFIG_SMP
10095 idle_thread_set_boot_cpu();
10096 balance_push_set(smp_processor_id(), false);
10097 #endif
10098 init_sched_fair_class();
10099
10100 psi_init();
10101
10102 init_uclamp();
10103
10104 preempt_dynamic_init();
10105
10106 scheduler_running = 1;
10107 }
10108
10109 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
10110
__might_sleep(const char * file,int line)10111 void __might_sleep(const char *file, int line)
10112 {
10113 unsigned int state = get_current_state();
10114 /*
10115 * Blocking primitives will set (and therefore destroy) current->state,
10116 * since we will exit with TASK_RUNNING make sure we enter with it,
10117 * otherwise we will destroy state.
10118 */
10119 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
10120 "do not call blocking ops when !TASK_RUNNING; "
10121 "state=%x set at [<%p>] %pS\n", state,
10122 (void *)current->task_state_change,
10123 (void *)current->task_state_change);
10124
10125 __might_resched(file, line, 0);
10126 }
10127 EXPORT_SYMBOL(__might_sleep);
10128
print_preempt_disable_ip(int preempt_offset,unsigned long ip)10129 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
10130 {
10131 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
10132 return;
10133
10134 if (preempt_count() == preempt_offset)
10135 return;
10136
10137 pr_err("Preemption disabled at:");
10138 print_ip_sym(KERN_ERR, ip);
10139 }
10140
resched_offsets_ok(unsigned int offsets)10141 static inline bool resched_offsets_ok(unsigned int offsets)
10142 {
10143 unsigned int nested = preempt_count();
10144
10145 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
10146
10147 return nested == offsets;
10148 }
10149
__might_resched(const char * file,int line,unsigned int offsets)10150 void __might_resched(const char *file, int line, unsigned int offsets)
10151 {
10152 /* Ratelimiting timestamp: */
10153 static unsigned long prev_jiffy;
10154
10155 unsigned long preempt_disable_ip;
10156
10157 /* WARN_ON_ONCE() by default, no rate limit required: */
10158 rcu_sleep_check();
10159
10160 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
10161 !is_idle_task(current) && !current->non_block_count) ||
10162 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10163 oops_in_progress)
10164 return;
10165
10166 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10167 return;
10168 prev_jiffy = jiffies;
10169
10170 /* Save this before calling printk(), since that will clobber it: */
10171 preempt_disable_ip = get_preempt_disable_ip(current);
10172
10173 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10174 file, line);
10175 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10176 in_atomic(), irqs_disabled(), current->non_block_count,
10177 current->pid, current->comm);
10178 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10179 offsets & MIGHT_RESCHED_PREEMPT_MASK);
10180
10181 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
10182 pr_err("RCU nest depth: %d, expected: %u\n",
10183 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
10184 }
10185
10186 if (task_stack_end_corrupted(current))
10187 pr_emerg("Thread overran stack, or stack corrupted\n");
10188
10189 debug_show_held_locks(current);
10190 if (irqs_disabled())
10191 print_irqtrace_events(current);
10192
10193 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
10194 preempt_disable_ip);
10195
10196 dump_stack();
10197 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10198 }
10199 EXPORT_SYMBOL(__might_resched);
10200
__cant_sleep(const char * file,int line,int preempt_offset)10201 void __cant_sleep(const char *file, int line, int preempt_offset)
10202 {
10203 static unsigned long prev_jiffy;
10204
10205 if (irqs_disabled())
10206 return;
10207
10208 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10209 return;
10210
10211 if (preempt_count() > preempt_offset)
10212 return;
10213
10214 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10215 return;
10216 prev_jiffy = jiffies;
10217
10218 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10219 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10220 in_atomic(), irqs_disabled(),
10221 current->pid, current->comm);
10222
10223 debug_show_held_locks(current);
10224 dump_stack();
10225 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10226 }
10227 EXPORT_SYMBOL_GPL(__cant_sleep);
10228
10229 #ifdef CONFIG_SMP
__cant_migrate(const char * file,int line)10230 void __cant_migrate(const char *file, int line)
10231 {
10232 static unsigned long prev_jiffy;
10233
10234 if (irqs_disabled())
10235 return;
10236
10237 if (is_migration_disabled(current))
10238 return;
10239
10240 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10241 return;
10242
10243 if (preempt_count() > 0)
10244 return;
10245
10246 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10247 return;
10248 prev_jiffy = jiffies;
10249
10250 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10251 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10252 in_atomic(), irqs_disabled(), is_migration_disabled(current),
10253 current->pid, current->comm);
10254
10255 debug_show_held_locks(current);
10256 dump_stack();
10257 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10258 }
10259 EXPORT_SYMBOL_GPL(__cant_migrate);
10260 #endif
10261 #endif
10262
10263 #ifdef CONFIG_MAGIC_SYSRQ
normalize_rt_tasks(void)10264 void normalize_rt_tasks(void)
10265 {
10266 struct task_struct *g, *p;
10267 struct sched_attr attr = {
10268 .sched_policy = SCHED_NORMAL,
10269 };
10270
10271 read_lock(&tasklist_lock);
10272 for_each_process_thread(g, p) {
10273 /*
10274 * Only normalize user tasks:
10275 */
10276 if (p->flags & PF_KTHREAD)
10277 continue;
10278
10279 p->se.exec_start = 0;
10280 schedstat_set(p->stats.wait_start, 0);
10281 schedstat_set(p->stats.sleep_start, 0);
10282 schedstat_set(p->stats.block_start, 0);
10283
10284 if (!dl_task(p) && !rt_task(p)) {
10285 /*
10286 * Renice negative nice level userspace
10287 * tasks back to 0:
10288 */
10289 if (task_nice(p) < 0)
10290 set_user_nice(p, 0);
10291 continue;
10292 }
10293
10294 __sched_setscheduler(p, &attr, false, false);
10295 }
10296 read_unlock(&tasklist_lock);
10297 }
10298
10299 #endif /* CONFIG_MAGIC_SYSRQ */
10300
10301 #if defined(CONFIG_KGDB_KDB)
10302 /*
10303 * These functions are only useful for kdb.
10304 *
10305 * They can only be called when the whole system has been
10306 * stopped - every CPU needs to be quiescent, and no scheduling
10307 * activity can take place. Using them for anything else would
10308 * be a serious bug, and as a result, they aren't even visible
10309 * under any other configuration.
10310 */
10311
10312 /**
10313 * curr_task - return the current task for a given CPU.
10314 * @cpu: the processor in question.
10315 *
10316 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10317 *
10318 * Return: The current task for @cpu.
10319 */
curr_task(int cpu)10320 struct task_struct *curr_task(int cpu)
10321 {
10322 return cpu_curr(cpu);
10323 }
10324
10325 #endif /* defined(CONFIG_KGDB_KDB) */
10326
10327 #ifdef CONFIG_CGROUP_SCHED
10328 /* task_group_lock serializes the addition/removal of task groups */
10329 static DEFINE_SPINLOCK(task_group_lock);
10330
alloc_uclamp_sched_group(struct task_group * tg,struct task_group * parent)10331 static inline void alloc_uclamp_sched_group(struct task_group *tg,
10332 struct task_group *parent)
10333 {
10334 #ifdef CONFIG_UCLAMP_TASK_GROUP
10335 enum uclamp_id clamp_id;
10336
10337 for_each_clamp_id(clamp_id) {
10338 uclamp_se_set(&tg->uclamp_req[clamp_id],
10339 uclamp_none(clamp_id), false);
10340 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10341 }
10342 #endif
10343 }
10344
sched_free_group(struct task_group * tg)10345 static void sched_free_group(struct task_group *tg)
10346 {
10347 free_fair_sched_group(tg);
10348 free_rt_sched_group(tg);
10349 autogroup_free(tg);
10350 kmem_cache_free(task_group_cache, tg);
10351 }
10352
sched_free_group_rcu(struct rcu_head * rcu)10353 static void sched_free_group_rcu(struct rcu_head *rcu)
10354 {
10355 sched_free_group(container_of(rcu, struct task_group, rcu));
10356 }
10357
sched_unregister_group(struct task_group * tg)10358 static void sched_unregister_group(struct task_group *tg)
10359 {
10360 unregister_fair_sched_group(tg);
10361 unregister_rt_sched_group(tg);
10362 /*
10363 * We have to wait for yet another RCU grace period to expire, as
10364 * print_cfs_stats() might run concurrently.
10365 */
10366 call_rcu(&tg->rcu, sched_free_group_rcu);
10367 }
10368
10369 /* allocate runqueue etc for a new task group */
sched_create_group(struct task_group * parent)10370 struct task_group *sched_create_group(struct task_group *parent)
10371 {
10372 struct task_group *tg;
10373
10374 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10375 if (!tg)
10376 return ERR_PTR(-ENOMEM);
10377
10378 if (!alloc_fair_sched_group(tg, parent))
10379 goto err;
10380
10381 if (!alloc_rt_sched_group(tg, parent))
10382 goto err;
10383
10384 alloc_uclamp_sched_group(tg, parent);
10385
10386 return tg;
10387
10388 err:
10389 sched_free_group(tg);
10390 return ERR_PTR(-ENOMEM);
10391 }
10392
sched_online_group(struct task_group * tg,struct task_group * parent)10393 void sched_online_group(struct task_group *tg, struct task_group *parent)
10394 {
10395 unsigned long flags;
10396
10397 spin_lock_irqsave(&task_group_lock, flags);
10398 list_add_rcu(&tg->list, &task_groups);
10399
10400 /* Root should already exist: */
10401 WARN_ON(!parent);
10402
10403 tg->parent = parent;
10404 INIT_LIST_HEAD(&tg->children);
10405 list_add_rcu(&tg->siblings, &parent->children);
10406 spin_unlock_irqrestore(&task_group_lock, flags);
10407
10408 online_fair_sched_group(tg);
10409 }
10410
10411 /* rcu callback to free various structures associated with a task group */
sched_unregister_group_rcu(struct rcu_head * rhp)10412 static void sched_unregister_group_rcu(struct rcu_head *rhp)
10413 {
10414 /* Now it should be safe to free those cfs_rqs: */
10415 sched_unregister_group(container_of(rhp, struct task_group, rcu));
10416 }
10417
sched_destroy_group(struct task_group * tg)10418 void sched_destroy_group(struct task_group *tg)
10419 {
10420 /* Wait for possible concurrent references to cfs_rqs complete: */
10421 call_rcu(&tg->rcu, sched_unregister_group_rcu);
10422 }
10423
sched_release_group(struct task_group * tg)10424 void sched_release_group(struct task_group *tg)
10425 {
10426 unsigned long flags;
10427
10428 /*
10429 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10430 * sched_cfs_period_timer()).
10431 *
10432 * For this to be effective, we have to wait for all pending users of
10433 * this task group to leave their RCU critical section to ensure no new
10434 * user will see our dying task group any more. Specifically ensure
10435 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10436 *
10437 * We therefore defer calling unregister_fair_sched_group() to
10438 * sched_unregister_group() which is guarantied to get called only after the
10439 * current RCU grace period has expired.
10440 */
10441 spin_lock_irqsave(&task_group_lock, flags);
10442 list_del_rcu(&tg->list);
10443 list_del_rcu(&tg->siblings);
10444 spin_unlock_irqrestore(&task_group_lock, flags);
10445 }
10446
sched_get_task_group(struct task_struct * tsk)10447 static struct task_group *sched_get_task_group(struct task_struct *tsk)
10448 {
10449 struct task_group *tg;
10450
10451 /*
10452 * All callers are synchronized by task_rq_lock(); we do not use RCU
10453 * which is pointless here. Thus, we pass "true" to task_css_check()
10454 * to prevent lockdep warnings.
10455 */
10456 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10457 struct task_group, css);
10458 tg = autogroup_task_group(tsk, tg);
10459
10460 return tg;
10461 }
10462
sched_change_group(struct task_struct * tsk,struct task_group * group)10463 static void sched_change_group(struct task_struct *tsk, struct task_group *group)
10464 {
10465 tsk->sched_task_group = group;
10466
10467 #ifdef CONFIG_FAIR_GROUP_SCHED
10468 if (tsk->sched_class->task_change_group)
10469 tsk->sched_class->task_change_group(tsk);
10470 else
10471 #endif
10472 set_task_rq(tsk, task_cpu(tsk));
10473 }
10474
10475 /*
10476 * Change task's runqueue when it moves between groups.
10477 *
10478 * The caller of this function should have put the task in its new group by
10479 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10480 * its new group.
10481 */
sched_move_task(struct task_struct * tsk)10482 void sched_move_task(struct task_struct *tsk)
10483 {
10484 int queued, running, queue_flags =
10485 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10486 struct task_group *group;
10487 struct rq *rq;
10488
10489 CLASS(task_rq_lock, rq_guard)(tsk);
10490 rq = rq_guard.rq;
10491
10492 /*
10493 * Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
10494 * group changes.
10495 */
10496 group = sched_get_task_group(tsk);
10497 if (group == tsk->sched_task_group)
10498 return;
10499
10500 update_rq_clock(rq);
10501
10502 running = task_current(rq, tsk);
10503 queued = task_on_rq_queued(tsk);
10504
10505 if (queued)
10506 dequeue_task(rq, tsk, queue_flags);
10507 if (running)
10508 put_prev_task(rq, tsk);
10509
10510 sched_change_group(tsk, group);
10511
10512 if (queued)
10513 enqueue_task(rq, tsk, queue_flags);
10514 if (running) {
10515 set_next_task(rq, tsk);
10516 /*
10517 * After changing group, the running task may have joined a
10518 * throttled one but it's still the running task. Trigger a
10519 * resched to make sure that task can still run.
10520 */
10521 resched_curr(rq);
10522 }
10523 }
10524
css_tg(struct cgroup_subsys_state * css)10525 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10526 {
10527 return css ? container_of(css, struct task_group, css) : NULL;
10528 }
10529
10530 static struct cgroup_subsys_state *
cpu_cgroup_css_alloc(struct cgroup_subsys_state * parent_css)10531 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10532 {
10533 struct task_group *parent = css_tg(parent_css);
10534 struct task_group *tg;
10535
10536 if (!parent) {
10537 /* This is early initialization for the top cgroup */
10538 return &root_task_group.css;
10539 }
10540
10541 tg = sched_create_group(parent);
10542 if (IS_ERR(tg))
10543 return ERR_PTR(-ENOMEM);
10544
10545 return &tg->css;
10546 }
10547
10548 /* Expose task group only after completing cgroup initialization */
cpu_cgroup_css_online(struct cgroup_subsys_state * css)10549 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10550 {
10551 struct task_group *tg = css_tg(css);
10552 struct task_group *parent = css_tg(css->parent);
10553
10554 if (parent)
10555 sched_online_group(tg, parent);
10556
10557 #ifdef CONFIG_UCLAMP_TASK_GROUP
10558 /* Propagate the effective uclamp value for the new group */
10559 guard(mutex)(&uclamp_mutex);
10560 guard(rcu)();
10561 cpu_util_update_eff(css);
10562 #endif
10563
10564 return 0;
10565 }
10566
cpu_cgroup_css_released(struct cgroup_subsys_state * css)10567 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10568 {
10569 struct task_group *tg = css_tg(css);
10570
10571 sched_release_group(tg);
10572 }
10573
cpu_cgroup_css_free(struct cgroup_subsys_state * css)10574 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10575 {
10576 struct task_group *tg = css_tg(css);
10577
10578 /*
10579 * Relies on the RCU grace period between css_released() and this.
10580 */
10581 sched_unregister_group(tg);
10582 }
10583
10584 #ifdef CONFIG_RT_GROUP_SCHED
cpu_cgroup_can_attach(struct cgroup_taskset * tset)10585 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10586 {
10587 struct task_struct *task;
10588 struct cgroup_subsys_state *css;
10589
10590 cgroup_taskset_for_each(task, css, tset) {
10591 if (!sched_rt_can_attach(css_tg(css), task))
10592 return -EINVAL;
10593 }
10594 return 0;
10595 }
10596 #endif
10597
cpu_cgroup_attach(struct cgroup_taskset * tset)10598 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10599 {
10600 struct task_struct *task;
10601 struct cgroup_subsys_state *css;
10602
10603 cgroup_taskset_for_each(task, css, tset)
10604 sched_move_task(task);
10605 }
10606
10607 #ifdef CONFIG_UCLAMP_TASK_GROUP
cpu_util_update_eff(struct cgroup_subsys_state * css)10608 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10609 {
10610 struct cgroup_subsys_state *top_css = css;
10611 struct uclamp_se *uc_parent = NULL;
10612 struct uclamp_se *uc_se = NULL;
10613 unsigned int eff[UCLAMP_CNT];
10614 enum uclamp_id clamp_id;
10615 unsigned int clamps;
10616
10617 lockdep_assert_held(&uclamp_mutex);
10618 SCHED_WARN_ON(!rcu_read_lock_held());
10619
10620 css_for_each_descendant_pre(css, top_css) {
10621 uc_parent = css_tg(css)->parent
10622 ? css_tg(css)->parent->uclamp : NULL;
10623
10624 for_each_clamp_id(clamp_id) {
10625 /* Assume effective clamps matches requested clamps */
10626 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10627 /* Cap effective clamps with parent's effective clamps */
10628 if (uc_parent &&
10629 eff[clamp_id] > uc_parent[clamp_id].value) {
10630 eff[clamp_id] = uc_parent[clamp_id].value;
10631 }
10632 }
10633 /* Ensure protection is always capped by limit */
10634 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10635
10636 /* Propagate most restrictive effective clamps */
10637 clamps = 0x0;
10638 uc_se = css_tg(css)->uclamp;
10639 for_each_clamp_id(clamp_id) {
10640 if (eff[clamp_id] == uc_se[clamp_id].value)
10641 continue;
10642 uc_se[clamp_id].value = eff[clamp_id];
10643 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10644 clamps |= (0x1 << clamp_id);
10645 }
10646 if (!clamps) {
10647 css = css_rightmost_descendant(css);
10648 continue;
10649 }
10650
10651 /* Immediately update descendants RUNNABLE tasks */
10652 uclamp_update_active_tasks(css);
10653 }
10654 }
10655
10656 /*
10657 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10658 * C expression. Since there is no way to convert a macro argument (N) into a
10659 * character constant, use two levels of macros.
10660 */
10661 #define _POW10(exp) ((unsigned int)1e##exp)
10662 #define POW10(exp) _POW10(exp)
10663
10664 struct uclamp_request {
10665 #define UCLAMP_PERCENT_SHIFT 2
10666 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10667 s64 percent;
10668 u64 util;
10669 int ret;
10670 };
10671
10672 static inline struct uclamp_request
capacity_from_percent(char * buf)10673 capacity_from_percent(char *buf)
10674 {
10675 struct uclamp_request req = {
10676 .percent = UCLAMP_PERCENT_SCALE,
10677 .util = SCHED_CAPACITY_SCALE,
10678 .ret = 0,
10679 };
10680
10681 buf = strim(buf);
10682 if (strcmp(buf, "max")) {
10683 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10684 &req.percent);
10685 if (req.ret)
10686 return req;
10687 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10688 req.ret = -ERANGE;
10689 return req;
10690 }
10691
10692 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10693 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10694 }
10695
10696 return req;
10697 }
10698
cpu_uclamp_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off,enum uclamp_id clamp_id)10699 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10700 size_t nbytes, loff_t off,
10701 enum uclamp_id clamp_id)
10702 {
10703 struct uclamp_request req;
10704 struct task_group *tg;
10705
10706 req = capacity_from_percent(buf);
10707 if (req.ret)
10708 return req.ret;
10709
10710 static_branch_enable(&sched_uclamp_used);
10711
10712 guard(mutex)(&uclamp_mutex);
10713 guard(rcu)();
10714
10715 tg = css_tg(of_css(of));
10716 if (tg->uclamp_req[clamp_id].value != req.util)
10717 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10718
10719 /*
10720 * Because of not recoverable conversion rounding we keep track of the
10721 * exact requested value
10722 */
10723 tg->uclamp_pct[clamp_id] = req.percent;
10724
10725 /* Update effective clamps to track the most restrictive value */
10726 cpu_util_update_eff(of_css(of));
10727
10728 return nbytes;
10729 }
10730
cpu_uclamp_min_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)10731 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10732 char *buf, size_t nbytes,
10733 loff_t off)
10734 {
10735 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10736 }
10737
cpu_uclamp_max_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)10738 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10739 char *buf, size_t nbytes,
10740 loff_t off)
10741 {
10742 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10743 }
10744
cpu_uclamp_print(struct seq_file * sf,enum uclamp_id clamp_id)10745 static inline void cpu_uclamp_print(struct seq_file *sf,
10746 enum uclamp_id clamp_id)
10747 {
10748 struct task_group *tg;
10749 u64 util_clamp;
10750 u64 percent;
10751 u32 rem;
10752
10753 scoped_guard (rcu) {
10754 tg = css_tg(seq_css(sf));
10755 util_clamp = tg->uclamp_req[clamp_id].value;
10756 }
10757
10758 if (util_clamp == SCHED_CAPACITY_SCALE) {
10759 seq_puts(sf, "max\n");
10760 return;
10761 }
10762
10763 percent = tg->uclamp_pct[clamp_id];
10764 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10765 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10766 }
10767
cpu_uclamp_min_show(struct seq_file * sf,void * v)10768 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10769 {
10770 cpu_uclamp_print(sf, UCLAMP_MIN);
10771 return 0;
10772 }
10773
cpu_uclamp_max_show(struct seq_file * sf,void * v)10774 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10775 {
10776 cpu_uclamp_print(sf, UCLAMP_MAX);
10777 return 0;
10778 }
10779 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10780
10781 #ifdef CONFIG_FAIR_GROUP_SCHED
cpu_shares_write_u64(struct cgroup_subsys_state * css,struct cftype * cftype,u64 shareval)10782 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10783 struct cftype *cftype, u64 shareval)
10784 {
10785 if (shareval > scale_load_down(ULONG_MAX))
10786 shareval = MAX_SHARES;
10787 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10788 }
10789
cpu_shares_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)10790 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10791 struct cftype *cft)
10792 {
10793 struct task_group *tg = css_tg(css);
10794
10795 return (u64) scale_load_down(tg->shares);
10796 }
10797
10798 #ifdef CONFIG_CFS_BANDWIDTH
10799 static DEFINE_MUTEX(cfs_constraints_mutex);
10800
10801 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10802 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10803 /* More than 203 days if BW_SHIFT equals 20. */
10804 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10805
10806 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10807
tg_set_cfs_bandwidth(struct task_group * tg,u64 period,u64 quota,u64 burst)10808 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10809 u64 burst)
10810 {
10811 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10812 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10813
10814 if (tg == &root_task_group)
10815 return -EINVAL;
10816
10817 /*
10818 * Ensure we have at some amount of bandwidth every period. This is
10819 * to prevent reaching a state of large arrears when throttled via
10820 * entity_tick() resulting in prolonged exit starvation.
10821 */
10822 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10823 return -EINVAL;
10824
10825 /*
10826 * Likewise, bound things on the other side by preventing insane quota
10827 * periods. This also allows us to normalize in computing quota
10828 * feasibility.
10829 */
10830 if (period > max_cfs_quota_period)
10831 return -EINVAL;
10832
10833 /*
10834 * Bound quota to defend quota against overflow during bandwidth shift.
10835 */
10836 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10837 return -EINVAL;
10838
10839 if (quota != RUNTIME_INF && (burst > quota ||
10840 burst + quota > max_cfs_runtime))
10841 return -EINVAL;
10842
10843 /*
10844 * Prevent race between setting of cfs_rq->runtime_enabled and
10845 * unthrottle_offline_cfs_rqs().
10846 */
10847 guard(cpus_read_lock)();
10848 guard(mutex)(&cfs_constraints_mutex);
10849
10850 ret = __cfs_schedulable(tg, period, quota);
10851 if (ret)
10852 return ret;
10853
10854 runtime_enabled = quota != RUNTIME_INF;
10855 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10856 /*
10857 * If we need to toggle cfs_bandwidth_used, off->on must occur
10858 * before making related changes, and on->off must occur afterwards
10859 */
10860 if (runtime_enabled && !runtime_was_enabled)
10861 cfs_bandwidth_usage_inc();
10862
10863 scoped_guard (raw_spinlock_irq, &cfs_b->lock) {
10864 cfs_b->period = ns_to_ktime(period);
10865 cfs_b->quota = quota;
10866 cfs_b->burst = burst;
10867
10868 __refill_cfs_bandwidth_runtime(cfs_b);
10869
10870 /*
10871 * Restart the period timer (if active) to handle new
10872 * period expiry:
10873 */
10874 if (runtime_enabled)
10875 start_cfs_bandwidth(cfs_b);
10876 }
10877
10878 for_each_online_cpu(i) {
10879 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10880 struct rq *rq = cfs_rq->rq;
10881
10882 guard(rq_lock_irq)(rq);
10883 cfs_rq->runtime_enabled = runtime_enabled;
10884 cfs_rq->runtime_remaining = 0;
10885
10886 if (cfs_rq->throttled)
10887 unthrottle_cfs_rq(cfs_rq);
10888 }
10889
10890 if (runtime_was_enabled && !runtime_enabled)
10891 cfs_bandwidth_usage_dec();
10892
10893 return 0;
10894 }
10895
tg_set_cfs_quota(struct task_group * tg,long cfs_quota_us)10896 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10897 {
10898 u64 quota, period, burst;
10899
10900 period = ktime_to_ns(tg->cfs_bandwidth.period);
10901 burst = tg->cfs_bandwidth.burst;
10902 if (cfs_quota_us < 0)
10903 quota = RUNTIME_INF;
10904 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10905 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10906 else
10907 return -EINVAL;
10908
10909 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10910 }
10911
tg_get_cfs_quota(struct task_group * tg)10912 static long tg_get_cfs_quota(struct task_group *tg)
10913 {
10914 u64 quota_us;
10915
10916 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10917 return -1;
10918
10919 quota_us = tg->cfs_bandwidth.quota;
10920 do_div(quota_us, NSEC_PER_USEC);
10921
10922 return quota_us;
10923 }
10924
tg_set_cfs_period(struct task_group * tg,long cfs_period_us)10925 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10926 {
10927 u64 quota, period, burst;
10928
10929 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10930 return -EINVAL;
10931
10932 period = (u64)cfs_period_us * NSEC_PER_USEC;
10933 quota = tg->cfs_bandwidth.quota;
10934 burst = tg->cfs_bandwidth.burst;
10935
10936 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10937 }
10938
tg_get_cfs_period(struct task_group * tg)10939 static long tg_get_cfs_period(struct task_group *tg)
10940 {
10941 u64 cfs_period_us;
10942
10943 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10944 do_div(cfs_period_us, NSEC_PER_USEC);
10945
10946 return cfs_period_us;
10947 }
10948
tg_set_cfs_burst(struct task_group * tg,long cfs_burst_us)10949 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10950 {
10951 u64 quota, period, burst;
10952
10953 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10954 return -EINVAL;
10955
10956 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10957 period = ktime_to_ns(tg->cfs_bandwidth.period);
10958 quota = tg->cfs_bandwidth.quota;
10959
10960 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10961 }
10962
tg_get_cfs_burst(struct task_group * tg)10963 static long tg_get_cfs_burst(struct task_group *tg)
10964 {
10965 u64 burst_us;
10966
10967 burst_us = tg->cfs_bandwidth.burst;
10968 do_div(burst_us, NSEC_PER_USEC);
10969
10970 return burst_us;
10971 }
10972
cpu_cfs_quota_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)10973 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10974 struct cftype *cft)
10975 {
10976 return tg_get_cfs_quota(css_tg(css));
10977 }
10978
cpu_cfs_quota_write_s64(struct cgroup_subsys_state * css,struct cftype * cftype,s64 cfs_quota_us)10979 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10980 struct cftype *cftype, s64 cfs_quota_us)
10981 {
10982 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10983 }
10984
cpu_cfs_period_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)10985 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10986 struct cftype *cft)
10987 {
10988 return tg_get_cfs_period(css_tg(css));
10989 }
10990
cpu_cfs_period_write_u64(struct cgroup_subsys_state * css,struct cftype * cftype,u64 cfs_period_us)10991 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10992 struct cftype *cftype, u64 cfs_period_us)
10993 {
10994 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10995 }
10996
cpu_cfs_burst_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)10997 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10998 struct cftype *cft)
10999 {
11000 return tg_get_cfs_burst(css_tg(css));
11001 }
11002
cpu_cfs_burst_write_u64(struct cgroup_subsys_state * css,struct cftype * cftype,u64 cfs_burst_us)11003 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
11004 struct cftype *cftype, u64 cfs_burst_us)
11005 {
11006 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
11007 }
11008
11009 struct cfs_schedulable_data {
11010 struct task_group *tg;
11011 u64 period, quota;
11012 };
11013
11014 /*
11015 * normalize group quota/period to be quota/max_period
11016 * note: units are usecs
11017 */
normalize_cfs_quota(struct task_group * tg,struct cfs_schedulable_data * d)11018 static u64 normalize_cfs_quota(struct task_group *tg,
11019 struct cfs_schedulable_data *d)
11020 {
11021 u64 quota, period;
11022
11023 if (tg == d->tg) {
11024 period = d->period;
11025 quota = d->quota;
11026 } else {
11027 period = tg_get_cfs_period(tg);
11028 quota = tg_get_cfs_quota(tg);
11029 }
11030
11031 /* note: these should typically be equivalent */
11032 if (quota == RUNTIME_INF || quota == -1)
11033 return RUNTIME_INF;
11034
11035 return to_ratio(period, quota);
11036 }
11037
tg_cfs_schedulable_down(struct task_group * tg,void * data)11038 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
11039 {
11040 struct cfs_schedulable_data *d = data;
11041 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11042 s64 quota = 0, parent_quota = -1;
11043
11044 if (!tg->parent) {
11045 quota = RUNTIME_INF;
11046 } else {
11047 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
11048
11049 quota = normalize_cfs_quota(tg, d);
11050 parent_quota = parent_b->hierarchical_quota;
11051
11052 /*
11053 * Ensure max(child_quota) <= parent_quota. On cgroup2,
11054 * always take the non-RUNTIME_INF min. On cgroup1, only
11055 * inherit when no limit is set. In both cases this is used
11056 * by the scheduler to determine if a given CFS task has a
11057 * bandwidth constraint at some higher level.
11058 */
11059 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
11060 if (quota == RUNTIME_INF)
11061 quota = parent_quota;
11062 else if (parent_quota != RUNTIME_INF)
11063 quota = min(quota, parent_quota);
11064 } else {
11065 if (quota == RUNTIME_INF)
11066 quota = parent_quota;
11067 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
11068 return -EINVAL;
11069 }
11070 }
11071 cfs_b->hierarchical_quota = quota;
11072
11073 return 0;
11074 }
11075
__cfs_schedulable(struct task_group * tg,u64 period,u64 quota)11076 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
11077 {
11078 struct cfs_schedulable_data data = {
11079 .tg = tg,
11080 .period = period,
11081 .quota = quota,
11082 };
11083
11084 if (quota != RUNTIME_INF) {
11085 do_div(data.period, NSEC_PER_USEC);
11086 do_div(data.quota, NSEC_PER_USEC);
11087 }
11088
11089 guard(rcu)();
11090 return walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
11091 }
11092
cpu_cfs_stat_show(struct seq_file * sf,void * v)11093 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
11094 {
11095 struct task_group *tg = css_tg(seq_css(sf));
11096 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11097
11098 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
11099 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
11100 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
11101
11102 if (schedstat_enabled() && tg != &root_task_group) {
11103 struct sched_statistics *stats;
11104 u64 ws = 0;
11105 int i;
11106
11107 for_each_possible_cpu(i) {
11108 stats = __schedstats_from_se(tg->se[i]);
11109 ws += schedstat_val(stats->wait_sum);
11110 }
11111
11112 seq_printf(sf, "wait_sum %llu\n", ws);
11113 }
11114
11115 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
11116 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
11117
11118 return 0;
11119 }
11120
throttled_time_self(struct task_group * tg)11121 static u64 throttled_time_self(struct task_group *tg)
11122 {
11123 int i;
11124 u64 total = 0;
11125
11126 for_each_possible_cpu(i) {
11127 total += READ_ONCE(tg->cfs_rq[i]->throttled_clock_self_time);
11128 }
11129
11130 return total;
11131 }
11132
cpu_cfs_local_stat_show(struct seq_file * sf,void * v)11133 static int cpu_cfs_local_stat_show(struct seq_file *sf, void *v)
11134 {
11135 struct task_group *tg = css_tg(seq_css(sf));
11136
11137 seq_printf(sf, "throttled_time %llu\n", throttled_time_self(tg));
11138
11139 return 0;
11140 }
11141 #endif /* CONFIG_CFS_BANDWIDTH */
11142 #endif /* CONFIG_FAIR_GROUP_SCHED */
11143
11144 #ifdef CONFIG_RT_GROUP_SCHED
cpu_rt_runtime_write(struct cgroup_subsys_state * css,struct cftype * cft,s64 val)11145 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
11146 struct cftype *cft, s64 val)
11147 {
11148 return sched_group_set_rt_runtime(css_tg(css), val);
11149 }
11150
cpu_rt_runtime_read(struct cgroup_subsys_state * css,struct cftype * cft)11151 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
11152 struct cftype *cft)
11153 {
11154 return sched_group_rt_runtime(css_tg(css));
11155 }
11156
cpu_rt_period_write_uint(struct cgroup_subsys_state * css,struct cftype * cftype,u64 rt_period_us)11157 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
11158 struct cftype *cftype, u64 rt_period_us)
11159 {
11160 return sched_group_set_rt_period(css_tg(css), rt_period_us);
11161 }
11162
cpu_rt_period_read_uint(struct cgroup_subsys_state * css,struct cftype * cft)11163 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11164 struct cftype *cft)
11165 {
11166 return sched_group_rt_period(css_tg(css));
11167 }
11168 #endif /* CONFIG_RT_GROUP_SCHED */
11169
11170 #ifdef CONFIG_FAIR_GROUP_SCHED
cpu_idle_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)11171 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11172 struct cftype *cft)
11173 {
11174 return css_tg(css)->idle;
11175 }
11176
cpu_idle_write_s64(struct cgroup_subsys_state * css,struct cftype * cft,s64 idle)11177 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11178 struct cftype *cft, s64 idle)
11179 {
11180 return sched_group_set_idle(css_tg(css), idle);
11181 }
11182 #endif
11183
11184 static struct cftype cpu_legacy_files[] = {
11185 #ifdef CONFIG_FAIR_GROUP_SCHED
11186 {
11187 .name = "shares",
11188 .read_u64 = cpu_shares_read_u64,
11189 .write_u64 = cpu_shares_write_u64,
11190 },
11191 {
11192 .name = "idle",
11193 .read_s64 = cpu_idle_read_s64,
11194 .write_s64 = cpu_idle_write_s64,
11195 },
11196 #endif
11197 #ifdef CONFIG_CFS_BANDWIDTH
11198 {
11199 .name = "cfs_quota_us",
11200 .read_s64 = cpu_cfs_quota_read_s64,
11201 .write_s64 = cpu_cfs_quota_write_s64,
11202 },
11203 {
11204 .name = "cfs_period_us",
11205 .read_u64 = cpu_cfs_period_read_u64,
11206 .write_u64 = cpu_cfs_period_write_u64,
11207 },
11208 {
11209 .name = "cfs_burst_us",
11210 .read_u64 = cpu_cfs_burst_read_u64,
11211 .write_u64 = cpu_cfs_burst_write_u64,
11212 },
11213 {
11214 .name = "stat",
11215 .seq_show = cpu_cfs_stat_show,
11216 },
11217 {
11218 .name = "stat.local",
11219 .seq_show = cpu_cfs_local_stat_show,
11220 },
11221 #endif
11222 #ifdef CONFIG_RT_GROUP_SCHED
11223 {
11224 .name = "rt_runtime_us",
11225 .read_s64 = cpu_rt_runtime_read,
11226 .write_s64 = cpu_rt_runtime_write,
11227 },
11228 {
11229 .name = "rt_period_us",
11230 .read_u64 = cpu_rt_period_read_uint,
11231 .write_u64 = cpu_rt_period_write_uint,
11232 },
11233 #endif
11234 #ifdef CONFIG_UCLAMP_TASK_GROUP
11235 {
11236 .name = "uclamp.min",
11237 .flags = CFTYPE_NOT_ON_ROOT,
11238 .seq_show = cpu_uclamp_min_show,
11239 .write = cpu_uclamp_min_write,
11240 },
11241 {
11242 .name = "uclamp.max",
11243 .flags = CFTYPE_NOT_ON_ROOT,
11244 .seq_show = cpu_uclamp_max_show,
11245 .write = cpu_uclamp_max_write,
11246 },
11247 #endif
11248 { } /* Terminate */
11249 };
11250
cpu_extra_stat_show(struct seq_file * sf,struct cgroup_subsys_state * css)11251 static int cpu_extra_stat_show(struct seq_file *sf,
11252 struct cgroup_subsys_state *css)
11253 {
11254 #ifdef CONFIG_CFS_BANDWIDTH
11255 {
11256 struct task_group *tg = css_tg(css);
11257 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11258 u64 throttled_usec, burst_usec;
11259
11260 throttled_usec = cfs_b->throttled_time;
11261 do_div(throttled_usec, NSEC_PER_USEC);
11262 burst_usec = cfs_b->burst_time;
11263 do_div(burst_usec, NSEC_PER_USEC);
11264
11265 seq_printf(sf, "nr_periods %d\n"
11266 "nr_throttled %d\n"
11267 "throttled_usec %llu\n"
11268 "nr_bursts %d\n"
11269 "burst_usec %llu\n",
11270 cfs_b->nr_periods, cfs_b->nr_throttled,
11271 throttled_usec, cfs_b->nr_burst, burst_usec);
11272 }
11273 #endif
11274 return 0;
11275 }
11276
cpu_local_stat_show(struct seq_file * sf,struct cgroup_subsys_state * css)11277 static int cpu_local_stat_show(struct seq_file *sf,
11278 struct cgroup_subsys_state *css)
11279 {
11280 #ifdef CONFIG_CFS_BANDWIDTH
11281 {
11282 struct task_group *tg = css_tg(css);
11283 u64 throttled_self_usec;
11284
11285 throttled_self_usec = throttled_time_self(tg);
11286 do_div(throttled_self_usec, NSEC_PER_USEC);
11287
11288 seq_printf(sf, "throttled_usec %llu\n",
11289 throttled_self_usec);
11290 }
11291 #endif
11292 return 0;
11293 }
11294
11295 #ifdef CONFIG_FAIR_GROUP_SCHED
cpu_weight_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)11296 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11297 struct cftype *cft)
11298 {
11299 struct task_group *tg = css_tg(css);
11300 u64 weight = scale_load_down(tg->shares);
11301
11302 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11303 }
11304
cpu_weight_write_u64(struct cgroup_subsys_state * css,struct cftype * cft,u64 weight)11305 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11306 struct cftype *cft, u64 weight)
11307 {
11308 /*
11309 * cgroup weight knobs should use the common MIN, DFL and MAX
11310 * values which are 1, 100 and 10000 respectively. While it loses
11311 * a bit of range on both ends, it maps pretty well onto the shares
11312 * value used by scheduler and the round-trip conversions preserve
11313 * the original value over the entire range.
11314 */
11315 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11316 return -ERANGE;
11317
11318 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11319
11320 return sched_group_set_shares(css_tg(css), scale_load(weight));
11321 }
11322
cpu_weight_nice_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)11323 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11324 struct cftype *cft)
11325 {
11326 unsigned long weight = scale_load_down(css_tg(css)->shares);
11327 int last_delta = INT_MAX;
11328 int prio, delta;
11329
11330 /* find the closest nice value to the current weight */
11331 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11332 delta = abs(sched_prio_to_weight[prio] - weight);
11333 if (delta >= last_delta)
11334 break;
11335 last_delta = delta;
11336 }
11337
11338 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11339 }
11340
cpu_weight_nice_write_s64(struct cgroup_subsys_state * css,struct cftype * cft,s64 nice)11341 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11342 struct cftype *cft, s64 nice)
11343 {
11344 unsigned long weight;
11345 int idx;
11346
11347 if (nice < MIN_NICE || nice > MAX_NICE)
11348 return -ERANGE;
11349
11350 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11351 idx = array_index_nospec(idx, 40);
11352 weight = sched_prio_to_weight[idx];
11353
11354 return sched_group_set_shares(css_tg(css), scale_load(weight));
11355 }
11356 #endif
11357
cpu_period_quota_print(struct seq_file * sf,long period,long quota)11358 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11359 long period, long quota)
11360 {
11361 if (quota < 0)
11362 seq_puts(sf, "max");
11363 else
11364 seq_printf(sf, "%ld", quota);
11365
11366 seq_printf(sf, " %ld\n", period);
11367 }
11368
11369 /* caller should put the current value in *@periodp before calling */
cpu_period_quota_parse(char * buf,u64 * periodp,u64 * quotap)11370 static int __maybe_unused cpu_period_quota_parse(char *buf,
11371 u64 *periodp, u64 *quotap)
11372 {
11373 char tok[21]; /* U64_MAX */
11374
11375 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11376 return -EINVAL;
11377
11378 *periodp *= NSEC_PER_USEC;
11379
11380 if (sscanf(tok, "%llu", quotap))
11381 *quotap *= NSEC_PER_USEC;
11382 else if (!strcmp(tok, "max"))
11383 *quotap = RUNTIME_INF;
11384 else
11385 return -EINVAL;
11386
11387 return 0;
11388 }
11389
11390 #ifdef CONFIG_CFS_BANDWIDTH
cpu_max_show(struct seq_file * sf,void * v)11391 static int cpu_max_show(struct seq_file *sf, void *v)
11392 {
11393 struct task_group *tg = css_tg(seq_css(sf));
11394
11395 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11396 return 0;
11397 }
11398
cpu_max_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)11399 static ssize_t cpu_max_write(struct kernfs_open_file *of,
11400 char *buf, size_t nbytes, loff_t off)
11401 {
11402 struct task_group *tg = css_tg(of_css(of));
11403 u64 period = tg_get_cfs_period(tg);
11404 u64 burst = tg->cfs_bandwidth.burst;
11405 u64 quota;
11406 int ret;
11407
11408 ret = cpu_period_quota_parse(buf, &period, "a);
11409 if (!ret)
11410 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11411 return ret ?: nbytes;
11412 }
11413 #endif
11414
11415 static struct cftype cpu_files[] = {
11416 #ifdef CONFIG_FAIR_GROUP_SCHED
11417 {
11418 .name = "weight",
11419 .flags = CFTYPE_NOT_ON_ROOT,
11420 .read_u64 = cpu_weight_read_u64,
11421 .write_u64 = cpu_weight_write_u64,
11422 },
11423 {
11424 .name = "weight.nice",
11425 .flags = CFTYPE_NOT_ON_ROOT,
11426 .read_s64 = cpu_weight_nice_read_s64,
11427 .write_s64 = cpu_weight_nice_write_s64,
11428 },
11429 {
11430 .name = "idle",
11431 .flags = CFTYPE_NOT_ON_ROOT,
11432 .read_s64 = cpu_idle_read_s64,
11433 .write_s64 = cpu_idle_write_s64,
11434 },
11435 #endif
11436 #ifdef CONFIG_CFS_BANDWIDTH
11437 {
11438 .name = "max",
11439 .flags = CFTYPE_NOT_ON_ROOT,
11440 .seq_show = cpu_max_show,
11441 .write = cpu_max_write,
11442 },
11443 {
11444 .name = "max.burst",
11445 .flags = CFTYPE_NOT_ON_ROOT,
11446 .read_u64 = cpu_cfs_burst_read_u64,
11447 .write_u64 = cpu_cfs_burst_write_u64,
11448 },
11449 #endif
11450 #ifdef CONFIG_UCLAMP_TASK_GROUP
11451 {
11452 .name = "uclamp.min",
11453 .flags = CFTYPE_NOT_ON_ROOT,
11454 .seq_show = cpu_uclamp_min_show,
11455 .write = cpu_uclamp_min_write,
11456 },
11457 {
11458 .name = "uclamp.max",
11459 .flags = CFTYPE_NOT_ON_ROOT,
11460 .seq_show = cpu_uclamp_max_show,
11461 .write = cpu_uclamp_max_write,
11462 },
11463 #endif
11464 { } /* terminate */
11465 };
11466
11467 struct cgroup_subsys cpu_cgrp_subsys = {
11468 .css_alloc = cpu_cgroup_css_alloc,
11469 .css_online = cpu_cgroup_css_online,
11470 .css_released = cpu_cgroup_css_released,
11471 .css_free = cpu_cgroup_css_free,
11472 .css_extra_stat_show = cpu_extra_stat_show,
11473 .css_local_stat_show = cpu_local_stat_show,
11474 #ifdef CONFIG_RT_GROUP_SCHED
11475 .can_attach = cpu_cgroup_can_attach,
11476 #endif
11477 .attach = cpu_cgroup_attach,
11478 .legacy_cftypes = cpu_legacy_files,
11479 .dfl_cftypes = cpu_files,
11480 .early_init = true,
11481 .threaded = true,
11482 };
11483
11484 #endif /* CONFIG_CGROUP_SCHED */
11485
dump_cpu_task(int cpu)11486 void dump_cpu_task(int cpu)
11487 {
11488 if (cpu == smp_processor_id() && in_hardirq()) {
11489 struct pt_regs *regs;
11490
11491 regs = get_irq_regs();
11492 if (regs) {
11493 show_regs(regs);
11494 return;
11495 }
11496 }
11497
11498 if (trigger_single_cpu_backtrace(cpu))
11499 return;
11500
11501 pr_info("Task dump for CPU %d:\n", cpu);
11502 sched_show_task(cpu_curr(cpu));
11503 }
11504
11505 /*
11506 * Nice levels are multiplicative, with a gentle 10% change for every
11507 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11508 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11509 * that remained on nice 0.
11510 *
11511 * The "10% effect" is relative and cumulative: from _any_ nice level,
11512 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11513 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11514 * If a task goes up by ~10% and another task goes down by ~10% then
11515 * the relative distance between them is ~25%.)
11516 */
11517 const int sched_prio_to_weight[40] = {
11518 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11519 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11520 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11521 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11522 /* 0 */ 1024, 820, 655, 526, 423,
11523 /* 5 */ 335, 272, 215, 172, 137,
11524 /* 10 */ 110, 87, 70, 56, 45,
11525 /* 15 */ 36, 29, 23, 18, 15,
11526 };
11527
11528 /*
11529 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11530 *
11531 * In cases where the weight does not change often, we can use the
11532 * precalculated inverse to speed up arithmetics by turning divisions
11533 * into multiplications:
11534 */
11535 const u32 sched_prio_to_wmult[40] = {
11536 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11537 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11538 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11539 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11540 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11541 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11542 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11543 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11544 };
11545
call_trace_sched_update_nr_running(struct rq * rq,int count)11546 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11547 {
11548 trace_sched_update_nr_running_tp(rq, count);
11549 }
11550
11551 #ifdef CONFIG_SCHED_MM_CID
11552
11553 /*
11554 * @cid_lock: Guarantee forward-progress of cid allocation.
11555 *
11556 * Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
11557 * is only used when contention is detected by the lock-free allocation so
11558 * forward progress can be guaranteed.
11559 */
11560 DEFINE_RAW_SPINLOCK(cid_lock);
11561
11562 /*
11563 * @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
11564 *
11565 * When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
11566 * detected, it is set to 1 to ensure that all newly coming allocations are
11567 * serialized by @cid_lock until the allocation which detected contention
11568 * completes and sets @use_cid_lock back to 0. This guarantees forward progress
11569 * of a cid allocation.
11570 */
11571 int use_cid_lock;
11572
11573 /*
11574 * mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
11575 * concurrently with respect to the execution of the source runqueue context
11576 * switch.
11577 *
11578 * There is one basic properties we want to guarantee here:
11579 *
11580 * (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
11581 * used by a task. That would lead to concurrent allocation of the cid and
11582 * userspace corruption.
11583 *
11584 * Provide this guarantee by introducing a Dekker memory ordering to guarantee
11585 * that a pair of loads observe at least one of a pair of stores, which can be
11586 * shown as:
11587 *
11588 * X = Y = 0
11589 *
11590 * w[X]=1 w[Y]=1
11591 * MB MB
11592 * r[Y]=y r[X]=x
11593 *
11594 * Which guarantees that x==0 && y==0 is impossible. But rather than using
11595 * values 0 and 1, this algorithm cares about specific state transitions of the
11596 * runqueue current task (as updated by the scheduler context switch), and the
11597 * per-mm/cpu cid value.
11598 *
11599 * Let's introduce task (Y) which has task->mm == mm and task (N) which has
11600 * task->mm != mm for the rest of the discussion. There are two scheduler state
11601 * transitions on context switch we care about:
11602 *
11603 * (TSA) Store to rq->curr with transition from (N) to (Y)
11604 *
11605 * (TSB) Store to rq->curr with transition from (Y) to (N)
11606 *
11607 * On the remote-clear side, there is one transition we care about:
11608 *
11609 * (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
11610 *
11611 * There is also a transition to UNSET state which can be performed from all
11612 * sides (scheduler, remote-clear). It is always performed with a cmpxchg which
11613 * guarantees that only a single thread will succeed:
11614 *
11615 * (TMB) cmpxchg to *pcpu_cid to mark UNSET
11616 *
11617 * Just to be clear, what we do _not_ want to happen is a transition to UNSET
11618 * when a thread is actively using the cid (property (1)).
11619 *
11620 * Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
11621 *
11622 * Scenario A) (TSA)+(TMA) (from next task perspective)
11623 *
11624 * CPU0 CPU1
11625 *
11626 * Context switch CS-1 Remote-clear
11627 * - store to rq->curr: (N)->(Y) (TSA) - cmpxchg to *pcpu_id to LAZY (TMA)
11628 * (implied barrier after cmpxchg)
11629 * - switch_mm_cid()
11630 * - memory barrier (see switch_mm_cid()
11631 * comment explaining how this barrier
11632 * is combined with other scheduler
11633 * barriers)
11634 * - mm_cid_get (next)
11635 * - READ_ONCE(*pcpu_cid) - rcu_dereference(src_rq->curr)
11636 *
11637 * This Dekker ensures that either task (Y) is observed by the
11638 * rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
11639 * observed.
11640 *
11641 * If task (Y) store is observed by rcu_dereference(), it means that there is
11642 * still an active task on the cpu. Remote-clear will therefore not transition
11643 * to UNSET, which fulfills property (1).
11644 *
11645 * If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
11646 * it will move its state to UNSET, which clears the percpu cid perhaps
11647 * uselessly (which is not an issue for correctness). Because task (Y) is not
11648 * observed, CPU1 can move ahead to set the state to UNSET. Because moving
11649 * state to UNSET is done with a cmpxchg expecting that the old state has the
11650 * LAZY flag set, only one thread will successfully UNSET.
11651 *
11652 * If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
11653 * will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
11654 * CPU1 will observe task (Y) and do nothing more, which is fine.
11655 *
11656 * What we are effectively preventing with this Dekker is a scenario where
11657 * neither LAZY flag nor store (Y) are observed, which would fail property (1)
11658 * because this would UNSET a cid which is actively used.
11659 */
11660
sched_mm_cid_migrate_from(struct task_struct * t)11661 void sched_mm_cid_migrate_from(struct task_struct *t)
11662 {
11663 t->migrate_from_cpu = task_cpu(t);
11664 }
11665
11666 static
__sched_mm_cid_migrate_from_fetch_cid(struct rq * src_rq,struct task_struct * t,struct mm_cid * src_pcpu_cid)11667 int __sched_mm_cid_migrate_from_fetch_cid(struct rq *src_rq,
11668 struct task_struct *t,
11669 struct mm_cid *src_pcpu_cid)
11670 {
11671 struct mm_struct *mm = t->mm;
11672 struct task_struct *src_task;
11673 int src_cid, last_mm_cid;
11674
11675 if (!mm)
11676 return -1;
11677
11678 last_mm_cid = t->last_mm_cid;
11679 /*
11680 * If the migrated task has no last cid, or if the current
11681 * task on src rq uses the cid, it means the source cid does not need
11682 * to be moved to the destination cpu.
11683 */
11684 if (last_mm_cid == -1)
11685 return -1;
11686 src_cid = READ_ONCE(src_pcpu_cid->cid);
11687 if (!mm_cid_is_valid(src_cid) || last_mm_cid != src_cid)
11688 return -1;
11689
11690 /*
11691 * If we observe an active task using the mm on this rq, it means we
11692 * are not the last task to be migrated from this cpu for this mm, so
11693 * there is no need to move src_cid to the destination cpu.
11694 */
11695 guard(rcu)();
11696 src_task = rcu_dereference(src_rq->curr);
11697 if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11698 t->last_mm_cid = -1;
11699 return -1;
11700 }
11701
11702 return src_cid;
11703 }
11704
11705 static
__sched_mm_cid_migrate_from_try_steal_cid(struct rq * src_rq,struct task_struct * t,struct mm_cid * src_pcpu_cid,int src_cid)11706 int __sched_mm_cid_migrate_from_try_steal_cid(struct rq *src_rq,
11707 struct task_struct *t,
11708 struct mm_cid *src_pcpu_cid,
11709 int src_cid)
11710 {
11711 struct task_struct *src_task;
11712 struct mm_struct *mm = t->mm;
11713 int lazy_cid;
11714
11715 if (src_cid == -1)
11716 return -1;
11717
11718 /*
11719 * Attempt to clear the source cpu cid to move it to the destination
11720 * cpu.
11721 */
11722 lazy_cid = mm_cid_set_lazy_put(src_cid);
11723 if (!try_cmpxchg(&src_pcpu_cid->cid, &src_cid, lazy_cid))
11724 return -1;
11725
11726 /*
11727 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11728 * rq->curr->mm matches the scheduler barrier in context_switch()
11729 * between store to rq->curr and load of prev and next task's
11730 * per-mm/cpu cid.
11731 *
11732 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11733 * rq->curr->mm_cid_active matches the barrier in
11734 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11735 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11736 * load of per-mm/cpu cid.
11737 */
11738
11739 /*
11740 * If we observe an active task using the mm on this rq after setting
11741 * the lazy-put flag, this task will be responsible for transitioning
11742 * from lazy-put flag set to MM_CID_UNSET.
11743 */
11744 scoped_guard (rcu) {
11745 src_task = rcu_dereference(src_rq->curr);
11746 if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11747 /*
11748 * We observed an active task for this mm, there is therefore
11749 * no point in moving this cid to the destination cpu.
11750 */
11751 t->last_mm_cid = -1;
11752 return -1;
11753 }
11754 }
11755
11756 /*
11757 * The src_cid is unused, so it can be unset.
11758 */
11759 if (!try_cmpxchg(&src_pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11760 return -1;
11761 return src_cid;
11762 }
11763
11764 /*
11765 * Migration to dst cpu. Called with dst_rq lock held.
11766 * Interrupts are disabled, which keeps the window of cid ownership without the
11767 * source rq lock held small.
11768 */
sched_mm_cid_migrate_to(struct rq * dst_rq,struct task_struct * t)11769 void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t)
11770 {
11771 struct mm_cid *src_pcpu_cid, *dst_pcpu_cid;
11772 struct mm_struct *mm = t->mm;
11773 int src_cid, dst_cid, src_cpu;
11774 struct rq *src_rq;
11775
11776 lockdep_assert_rq_held(dst_rq);
11777
11778 if (!mm)
11779 return;
11780 src_cpu = t->migrate_from_cpu;
11781 if (src_cpu == -1) {
11782 t->last_mm_cid = -1;
11783 return;
11784 }
11785 /*
11786 * Move the src cid if the dst cid is unset. This keeps id
11787 * allocation closest to 0 in cases where few threads migrate around
11788 * many cpus.
11789 *
11790 * If destination cid is already set, we may have to just clear
11791 * the src cid to ensure compactness in frequent migrations
11792 * scenarios.
11793 *
11794 * It is not useful to clear the src cid when the number of threads is
11795 * greater or equal to the number of allowed cpus, because user-space
11796 * can expect that the number of allowed cids can reach the number of
11797 * allowed cpus.
11798 */
11799 dst_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(dst_rq));
11800 dst_cid = READ_ONCE(dst_pcpu_cid->cid);
11801 if (!mm_cid_is_unset(dst_cid) &&
11802 atomic_read(&mm->mm_users) >= t->nr_cpus_allowed)
11803 return;
11804 src_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, src_cpu);
11805 src_rq = cpu_rq(src_cpu);
11806 src_cid = __sched_mm_cid_migrate_from_fetch_cid(src_rq, t, src_pcpu_cid);
11807 if (src_cid == -1)
11808 return;
11809 src_cid = __sched_mm_cid_migrate_from_try_steal_cid(src_rq, t, src_pcpu_cid,
11810 src_cid);
11811 if (src_cid == -1)
11812 return;
11813 if (!mm_cid_is_unset(dst_cid)) {
11814 __mm_cid_put(mm, src_cid);
11815 return;
11816 }
11817 /* Move src_cid to dst cpu. */
11818 mm_cid_snapshot_time(dst_rq, mm);
11819 WRITE_ONCE(dst_pcpu_cid->cid, src_cid);
11820 }
11821
sched_mm_cid_remote_clear(struct mm_struct * mm,struct mm_cid * pcpu_cid,int cpu)11822 static void sched_mm_cid_remote_clear(struct mm_struct *mm, struct mm_cid *pcpu_cid,
11823 int cpu)
11824 {
11825 struct rq *rq = cpu_rq(cpu);
11826 struct task_struct *t;
11827 int cid, lazy_cid;
11828
11829 cid = READ_ONCE(pcpu_cid->cid);
11830 if (!mm_cid_is_valid(cid))
11831 return;
11832
11833 /*
11834 * Clear the cpu cid if it is set to keep cid allocation compact. If
11835 * there happens to be other tasks left on the source cpu using this
11836 * mm, the next task using this mm will reallocate its cid on context
11837 * switch.
11838 */
11839 lazy_cid = mm_cid_set_lazy_put(cid);
11840 if (!try_cmpxchg(&pcpu_cid->cid, &cid, lazy_cid))
11841 return;
11842
11843 /*
11844 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11845 * rq->curr->mm matches the scheduler barrier in context_switch()
11846 * between store to rq->curr and load of prev and next task's
11847 * per-mm/cpu cid.
11848 *
11849 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11850 * rq->curr->mm_cid_active matches the barrier in
11851 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11852 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11853 * load of per-mm/cpu cid.
11854 */
11855
11856 /*
11857 * If we observe an active task using the mm on this rq after setting
11858 * the lazy-put flag, that task will be responsible for transitioning
11859 * from lazy-put flag set to MM_CID_UNSET.
11860 */
11861 scoped_guard (rcu) {
11862 t = rcu_dereference(rq->curr);
11863 if (READ_ONCE(t->mm_cid_active) && t->mm == mm)
11864 return;
11865 }
11866
11867 /*
11868 * The cid is unused, so it can be unset.
11869 * Disable interrupts to keep the window of cid ownership without rq
11870 * lock small.
11871 */
11872 scoped_guard (irqsave) {
11873 if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11874 __mm_cid_put(mm, cid);
11875 }
11876 }
11877
sched_mm_cid_remote_clear_old(struct mm_struct * mm,int cpu)11878 static void sched_mm_cid_remote_clear_old(struct mm_struct *mm, int cpu)
11879 {
11880 struct rq *rq = cpu_rq(cpu);
11881 struct mm_cid *pcpu_cid;
11882 struct task_struct *curr;
11883 u64 rq_clock;
11884
11885 /*
11886 * rq->clock load is racy on 32-bit but one spurious clear once in a
11887 * while is irrelevant.
11888 */
11889 rq_clock = READ_ONCE(rq->clock);
11890 pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11891
11892 /*
11893 * In order to take care of infrequently scheduled tasks, bump the time
11894 * snapshot associated with this cid if an active task using the mm is
11895 * observed on this rq.
11896 */
11897 scoped_guard (rcu) {
11898 curr = rcu_dereference(rq->curr);
11899 if (READ_ONCE(curr->mm_cid_active) && curr->mm == mm) {
11900 WRITE_ONCE(pcpu_cid->time, rq_clock);
11901 return;
11902 }
11903 }
11904
11905 if (rq_clock < pcpu_cid->time + SCHED_MM_CID_PERIOD_NS)
11906 return;
11907 sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11908 }
11909
sched_mm_cid_remote_clear_weight(struct mm_struct * mm,int cpu,int weight)11910 static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
11911 int weight)
11912 {
11913 struct mm_cid *pcpu_cid;
11914 int cid;
11915
11916 pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11917 cid = READ_ONCE(pcpu_cid->cid);
11918 if (!mm_cid_is_valid(cid) || cid < weight)
11919 return;
11920 sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11921 }
11922
task_mm_cid_work(struct callback_head * work)11923 static void task_mm_cid_work(struct callback_head *work)
11924 {
11925 unsigned long now = jiffies, old_scan, next_scan;
11926 struct task_struct *t = current;
11927 struct cpumask *cidmask;
11928 struct mm_struct *mm;
11929 int weight, cpu;
11930
11931 SCHED_WARN_ON(t != container_of(work, struct task_struct, cid_work));
11932
11933 work->next = work; /* Prevent double-add */
11934 if (t->flags & PF_EXITING)
11935 return;
11936 mm = t->mm;
11937 if (!mm)
11938 return;
11939 old_scan = READ_ONCE(mm->mm_cid_next_scan);
11940 next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11941 if (!old_scan) {
11942 unsigned long res;
11943
11944 res = cmpxchg(&mm->mm_cid_next_scan, old_scan, next_scan);
11945 if (res != old_scan)
11946 old_scan = res;
11947 else
11948 old_scan = next_scan;
11949 }
11950 if (time_before(now, old_scan))
11951 return;
11952 if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
11953 return;
11954 cidmask = mm_cidmask(mm);
11955 /* Clear cids that were not recently used. */
11956 for_each_possible_cpu(cpu)
11957 sched_mm_cid_remote_clear_old(mm, cpu);
11958 weight = cpumask_weight(cidmask);
11959 /*
11960 * Clear cids that are greater or equal to the cidmask weight to
11961 * recompact it.
11962 */
11963 for_each_possible_cpu(cpu)
11964 sched_mm_cid_remote_clear_weight(mm, cpu, weight);
11965 }
11966
init_sched_mm_cid(struct task_struct * t)11967 void init_sched_mm_cid(struct task_struct *t)
11968 {
11969 struct mm_struct *mm = t->mm;
11970 int mm_users = 0;
11971
11972 if (mm) {
11973 mm_users = atomic_read(&mm->mm_users);
11974 if (mm_users == 1)
11975 mm->mm_cid_next_scan = jiffies + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11976 }
11977 t->cid_work.next = &t->cid_work; /* Protect against double add */
11978 init_task_work(&t->cid_work, task_mm_cid_work);
11979 }
11980
task_tick_mm_cid(struct rq * rq,struct task_struct * curr)11981 void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
11982 {
11983 struct callback_head *work = &curr->cid_work;
11984 unsigned long now = jiffies;
11985
11986 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
11987 work->next != work)
11988 return;
11989 if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
11990 return;
11991 task_work_add(curr, work, TWA_RESUME);
11992 }
11993
sched_mm_cid_exit_signals(struct task_struct * t)11994 void sched_mm_cid_exit_signals(struct task_struct *t)
11995 {
11996 struct mm_struct *mm = t->mm;
11997 struct rq *rq;
11998
11999 if (!mm)
12000 return;
12001
12002 preempt_disable();
12003 rq = this_rq();
12004 guard(rq_lock_irqsave)(rq);
12005 preempt_enable_no_resched(); /* holding spinlock */
12006 WRITE_ONCE(t->mm_cid_active, 0);
12007 /*
12008 * Store t->mm_cid_active before loading per-mm/cpu cid.
12009 * Matches barrier in sched_mm_cid_remote_clear_old().
12010 */
12011 smp_mb();
12012 mm_cid_put(mm);
12013 t->last_mm_cid = t->mm_cid = -1;
12014 }
12015
sched_mm_cid_before_execve(struct task_struct * t)12016 void sched_mm_cid_before_execve(struct task_struct *t)
12017 {
12018 struct mm_struct *mm = t->mm;
12019 struct rq *rq;
12020
12021 if (!mm)
12022 return;
12023
12024 preempt_disable();
12025 rq = this_rq();
12026 guard(rq_lock_irqsave)(rq);
12027 preempt_enable_no_resched(); /* holding spinlock */
12028 WRITE_ONCE(t->mm_cid_active, 0);
12029 /*
12030 * Store t->mm_cid_active before loading per-mm/cpu cid.
12031 * Matches barrier in sched_mm_cid_remote_clear_old().
12032 */
12033 smp_mb();
12034 mm_cid_put(mm);
12035 t->last_mm_cid = t->mm_cid = -1;
12036 }
12037
sched_mm_cid_after_execve(struct task_struct * t)12038 void sched_mm_cid_after_execve(struct task_struct *t)
12039 {
12040 struct mm_struct *mm = t->mm;
12041 struct rq *rq;
12042
12043 if (!mm)
12044 return;
12045
12046 preempt_disable();
12047 rq = this_rq();
12048 scoped_guard (rq_lock_irqsave, rq) {
12049 preempt_enable_no_resched(); /* holding spinlock */
12050 WRITE_ONCE(t->mm_cid_active, 1);
12051 /*
12052 * Store t->mm_cid_active before loading per-mm/cpu cid.
12053 * Matches barrier in sched_mm_cid_remote_clear_old().
12054 */
12055 smp_mb();
12056 t->last_mm_cid = t->mm_cid = mm_cid_get(rq, mm);
12057 }
12058 rseq_set_notify_resume(t);
12059 }
12060
sched_mm_cid_fork(struct task_struct * t)12061 void sched_mm_cid_fork(struct task_struct *t)
12062 {
12063 WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
12064 t->mm_cid_active = 1;
12065 }
12066 #endif
12067