1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  kernel/sched/core.c
4  *
5  *  Core kernel scheduler code and related syscalls
6  *
7  *  Copyright (C) 1991-2002  Linus Torvalds
8  */
9 #include <linux/highmem.h>
10 #include <linux/hrtimer_api.h>
11 #include <linux/ktime_api.h>
12 #include <linux/sched/signal.h>
13 #include <linux/syscalls_api.h>
14 #include <linux/debug_locks.h>
15 #include <linux/prefetch.h>
16 #include <linux/capability.h>
17 #include <linux/pgtable_api.h>
18 #include <linux/wait_bit.h>
19 #include <linux/jiffies.h>
20 #include <linux/spinlock_api.h>
21 #include <linux/cpumask_api.h>
22 #include <linux/lockdep_api.h>
23 #include <linux/hardirq.h>
24 #include <linux/softirq.h>
25 #include <linux/refcount_api.h>
26 #include <linux/topology.h>
27 #include <linux/sched/clock.h>
28 #include <linux/sched/cond_resched.h>
29 #include <linux/sched/cputime.h>
30 #include <linux/sched/debug.h>
31 #include <linux/sched/hotplug.h>
32 #include <linux/sched/init.h>
33 #include <linux/sched/isolation.h>
34 #include <linux/sched/loadavg.h>
35 #include <linux/sched/mm.h>
36 #include <linux/sched/nohz.h>
37 #include <linux/sched/rseq_api.h>
38 #include <linux/sched/rt.h>
39 
40 #include <linux/blkdev.h>
41 #include <linux/context_tracking.h>
42 #include <linux/cpuset.h>
43 #include <linux/delayacct.h>
44 #include <linux/init_task.h>
45 #include <linux/interrupt.h>
46 #include <linux/ioprio.h>
47 #include <linux/kallsyms.h>
48 #include <linux/kcov.h>
49 #include <linux/kprobes.h>
50 #include <linux/llist_api.h>
51 #include <linux/mmu_context.h>
52 #include <linux/mmzone.h>
53 #include <linux/mutex_api.h>
54 #include <linux/nmi.h>
55 #include <linux/nospec.h>
56 #include <linux/perf_event_api.h>
57 #include <linux/profile.h>
58 #include <linux/psi.h>
59 #include <linux/rcuwait_api.h>
60 #include <linux/sched/wake_q.h>
61 #include <linux/scs.h>
62 #include <linux/slab.h>
63 #include <linux/syscalls.h>
64 #include <linux/vtime.h>
65 #include <linux/wait_api.h>
66 #include <linux/workqueue_api.h>
67 
68 #ifdef CONFIG_PREEMPT_DYNAMIC
69 # ifdef CONFIG_GENERIC_ENTRY
70 #  include <linux/entry-common.h>
71 # endif
72 #endif
73 
74 #include <uapi/linux/sched/types.h>
75 
76 #include <asm/switch_to.h>
77 #include <asm/tlb.h>
78 
79 #define CREATE_TRACE_POINTS
80 #include <linux/sched/rseq_api.h>
81 #include <trace/events/sched.h>
82 #undef CREATE_TRACE_POINTS
83 
84 #include "sched.h"
85 #include "stats.h"
86 #include "autogroup.h"
87 
88 #include "autogroup.h"
89 #include "pelt.h"
90 #include "smp.h"
91 #include "stats.h"
92 
93 #include "../workqueue_internal.h"
94 #include "../../io_uring/io-wq.h"
95 #include "../smpboot.h"
96 
97 /*
98  * Export tracepoints that act as a bare tracehook (ie: have no trace event
99  * associated with them) to allow external modules to probe them.
100  */
101 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
102 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
103 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
104 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
105 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
106 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
107 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
108 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
109 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
110 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
111 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
112 
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
114 
115 #ifdef CONFIG_SCHED_DEBUG
116 /*
117  * Debugging: various feature bits
118  *
119  * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
120  * sysctl_sched_features, defined in sched.h, to allow constants propagation
121  * at compile time and compiler optimization based on features default.
122  */
123 #define SCHED_FEAT(name, enabled)	\
124 	(1UL << __SCHED_FEAT_##name) * enabled |
125 const_debug unsigned int sysctl_sched_features =
126 #include "features.h"
127 	0;
128 #undef SCHED_FEAT
129 
130 /*
131  * Print a warning if need_resched is set for the given duration (if
132  * LATENCY_WARN is enabled).
133  *
134  * If sysctl_resched_latency_warn_once is set, only one warning will be shown
135  * per boot.
136  */
137 __read_mostly int sysctl_resched_latency_warn_ms = 100;
138 __read_mostly int sysctl_resched_latency_warn_once = 1;
139 #endif /* CONFIG_SCHED_DEBUG */
140 
141 /*
142  * Number of tasks to iterate in a single balance run.
143  * Limited because this is done with IRQs disabled.
144  */
145 #ifdef CONFIG_PREEMPT_RT
146 const_debug unsigned int sysctl_sched_nr_migrate = 8;
147 #else
148 const_debug unsigned int sysctl_sched_nr_migrate = 32;
149 #endif
150 
151 __read_mostly int scheduler_running;
152 
153 #ifdef CONFIG_SCHED_CORE
154 
155 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
156 
157 /* kernel prio, less is more */
__task_prio(struct task_struct * p)158 static inline int __task_prio(struct task_struct *p)
159 {
160 	if (p->sched_class == &stop_sched_class) /* trumps deadline */
161 		return -2;
162 
163 	if (rt_prio(p->prio)) /* includes deadline */
164 		return p->prio; /* [-1, 99] */
165 
166 	if (p->sched_class == &idle_sched_class)
167 		return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
168 
169 	return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
170 }
171 
172 /*
173  * l(a,b)
174  * le(a,b) := !l(b,a)
175  * g(a,b)  := l(b,a)
176  * ge(a,b) := !l(a,b)
177  */
178 
179 /* real prio, less is less */
prio_less(struct task_struct * a,struct task_struct * b,bool in_fi)180 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
181 {
182 
183 	int pa = __task_prio(a), pb = __task_prio(b);
184 
185 	if (-pa < -pb)
186 		return true;
187 
188 	if (-pb < -pa)
189 		return false;
190 
191 	if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
192 		return !dl_time_before(a->dl.deadline, b->dl.deadline);
193 
194 	if (pa == MAX_RT_PRIO + MAX_NICE)	/* fair */
195 		return cfs_prio_less(a, b, in_fi);
196 
197 	return false;
198 }
199 
__sched_core_less(struct task_struct * a,struct task_struct * b)200 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
201 {
202 	if (a->core_cookie < b->core_cookie)
203 		return true;
204 
205 	if (a->core_cookie > b->core_cookie)
206 		return false;
207 
208 	/* flip prio, so high prio is leftmost */
209 	if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
210 		return true;
211 
212 	return false;
213 }
214 
215 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
216 
rb_sched_core_less(struct rb_node * a,const struct rb_node * b)217 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
218 {
219 	return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
220 }
221 
rb_sched_core_cmp(const void * key,const struct rb_node * node)222 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
223 {
224 	const struct task_struct *p = __node_2_sc(node);
225 	unsigned long cookie = (unsigned long)key;
226 
227 	if (cookie < p->core_cookie)
228 		return -1;
229 
230 	if (cookie > p->core_cookie)
231 		return 1;
232 
233 	return 0;
234 }
235 
sched_core_enqueue(struct rq * rq,struct task_struct * p)236 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
237 {
238 	rq->core->core_task_seq++;
239 
240 	if (!p->core_cookie)
241 		return;
242 
243 	rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
244 }
245 
sched_core_dequeue(struct rq * rq,struct task_struct * p,int flags)246 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
247 {
248 	rq->core->core_task_seq++;
249 
250 	if (sched_core_enqueued(p)) {
251 		rb_erase(&p->core_node, &rq->core_tree);
252 		RB_CLEAR_NODE(&p->core_node);
253 	}
254 
255 	/*
256 	 * Migrating the last task off the cpu, with the cpu in forced idle
257 	 * state. Reschedule to create an accounting edge for forced idle,
258 	 * and re-examine whether the core is still in forced idle state.
259 	 */
260 	if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
261 	    rq->core->core_forceidle_count && rq->curr == rq->idle)
262 		resched_curr(rq);
263 }
264 
265 /*
266  * Find left-most (aka, highest priority) task matching @cookie.
267  */
sched_core_find(struct rq * rq,unsigned long cookie)268 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
269 {
270 	struct rb_node *node;
271 
272 	node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
273 	/*
274 	 * The idle task always matches any cookie!
275 	 */
276 	if (!node)
277 		return idle_sched_class.pick_task(rq);
278 
279 	return __node_2_sc(node);
280 }
281 
sched_core_next(struct task_struct * p,unsigned long cookie)282 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
283 {
284 	struct rb_node *node = &p->core_node;
285 
286 	node = rb_next(node);
287 	if (!node)
288 		return NULL;
289 
290 	p = container_of(node, struct task_struct, core_node);
291 	if (p->core_cookie != cookie)
292 		return NULL;
293 
294 	return p;
295 }
296 
297 /*
298  * Magic required such that:
299  *
300  *	raw_spin_rq_lock(rq);
301  *	...
302  *	raw_spin_rq_unlock(rq);
303  *
304  * ends up locking and unlocking the _same_ lock, and all CPUs
305  * always agree on what rq has what lock.
306  *
307  * XXX entirely possible to selectively enable cores, don't bother for now.
308  */
309 
310 static DEFINE_MUTEX(sched_core_mutex);
311 static atomic_t sched_core_count;
312 static struct cpumask sched_core_mask;
313 
sched_core_lock(int cpu,unsigned long * flags)314 static void sched_core_lock(int cpu, unsigned long *flags)
315 {
316 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
317 	int t, i = 0;
318 
319 	local_irq_save(*flags);
320 	for_each_cpu(t, smt_mask)
321 		raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
322 }
323 
sched_core_unlock(int cpu,unsigned long * flags)324 static void sched_core_unlock(int cpu, unsigned long *flags)
325 {
326 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
327 	int t;
328 
329 	for_each_cpu(t, smt_mask)
330 		raw_spin_unlock(&cpu_rq(t)->__lock);
331 	local_irq_restore(*flags);
332 }
333 
__sched_core_flip(bool enabled)334 static void __sched_core_flip(bool enabled)
335 {
336 	unsigned long flags;
337 	int cpu, t;
338 
339 	cpus_read_lock();
340 
341 	/*
342 	 * Toggle the online cores, one by one.
343 	 */
344 	cpumask_copy(&sched_core_mask, cpu_online_mask);
345 	for_each_cpu(cpu, &sched_core_mask) {
346 		const struct cpumask *smt_mask = cpu_smt_mask(cpu);
347 
348 		sched_core_lock(cpu, &flags);
349 
350 		for_each_cpu(t, smt_mask)
351 			cpu_rq(t)->core_enabled = enabled;
352 
353 		cpu_rq(cpu)->core->core_forceidle_start = 0;
354 
355 		sched_core_unlock(cpu, &flags);
356 
357 		cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
358 	}
359 
360 	/*
361 	 * Toggle the offline CPUs.
362 	 */
363 	cpumask_copy(&sched_core_mask, cpu_possible_mask);
364 	cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
365 
366 	for_each_cpu(cpu, &sched_core_mask)
367 		cpu_rq(cpu)->core_enabled = enabled;
368 
369 	cpus_read_unlock();
370 }
371 
sched_core_assert_empty(void)372 static void sched_core_assert_empty(void)
373 {
374 	int cpu;
375 
376 	for_each_possible_cpu(cpu)
377 		WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
378 }
379 
__sched_core_enable(void)380 static void __sched_core_enable(void)
381 {
382 	static_branch_enable(&__sched_core_enabled);
383 	/*
384 	 * Ensure all previous instances of raw_spin_rq_*lock() have finished
385 	 * and future ones will observe !sched_core_disabled().
386 	 */
387 	synchronize_rcu();
388 	__sched_core_flip(true);
389 	sched_core_assert_empty();
390 }
391 
__sched_core_disable(void)392 static void __sched_core_disable(void)
393 {
394 	sched_core_assert_empty();
395 	__sched_core_flip(false);
396 	static_branch_disable(&__sched_core_enabled);
397 }
398 
sched_core_get(void)399 void sched_core_get(void)
400 {
401 	if (atomic_inc_not_zero(&sched_core_count))
402 		return;
403 
404 	mutex_lock(&sched_core_mutex);
405 	if (!atomic_read(&sched_core_count))
406 		__sched_core_enable();
407 
408 	smp_mb__before_atomic();
409 	atomic_inc(&sched_core_count);
410 	mutex_unlock(&sched_core_mutex);
411 }
412 
__sched_core_put(struct work_struct * work)413 static void __sched_core_put(struct work_struct *work)
414 {
415 	if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
416 		__sched_core_disable();
417 		mutex_unlock(&sched_core_mutex);
418 	}
419 }
420 
sched_core_put(void)421 void sched_core_put(void)
422 {
423 	static DECLARE_WORK(_work, __sched_core_put);
424 
425 	/*
426 	 * "There can be only one"
427 	 *
428 	 * Either this is the last one, or we don't actually need to do any
429 	 * 'work'. If it is the last *again*, we rely on
430 	 * WORK_STRUCT_PENDING_BIT.
431 	 */
432 	if (!atomic_add_unless(&sched_core_count, -1, 1))
433 		schedule_work(&_work);
434 }
435 
436 #else /* !CONFIG_SCHED_CORE */
437 
sched_core_enqueue(struct rq * rq,struct task_struct * p)438 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
439 static inline void
sched_core_dequeue(struct rq * rq,struct task_struct * p,int flags)440 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
441 
442 #endif /* CONFIG_SCHED_CORE */
443 
444 /*
445  * Serialization rules:
446  *
447  * Lock order:
448  *
449  *   p->pi_lock
450  *     rq->lock
451  *       hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
452  *
453  *  rq1->lock
454  *    rq2->lock  where: rq1 < rq2
455  *
456  * Regular state:
457  *
458  * Normal scheduling state is serialized by rq->lock. __schedule() takes the
459  * local CPU's rq->lock, it optionally removes the task from the runqueue and
460  * always looks at the local rq data structures to find the most eligible task
461  * to run next.
462  *
463  * Task enqueue is also under rq->lock, possibly taken from another CPU.
464  * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
465  * the local CPU to avoid bouncing the runqueue state around [ see
466  * ttwu_queue_wakelist() ]
467  *
468  * Task wakeup, specifically wakeups that involve migration, are horribly
469  * complicated to avoid having to take two rq->locks.
470  *
471  * Special state:
472  *
473  * System-calls and anything external will use task_rq_lock() which acquires
474  * both p->pi_lock and rq->lock. As a consequence the state they change is
475  * stable while holding either lock:
476  *
477  *  - sched_setaffinity()/
478  *    set_cpus_allowed_ptr():	p->cpus_ptr, p->nr_cpus_allowed
479  *  - set_user_nice():		p->se.load, p->*prio
480  *  - __sched_setscheduler():	p->sched_class, p->policy, p->*prio,
481  *				p->se.load, p->rt_priority,
482  *				p->dl.dl_{runtime, deadline, period, flags, bw, density}
483  *  - sched_setnuma():		p->numa_preferred_nid
484  *  - sched_move_task()/
485  *    cpu_cgroup_fork():	p->sched_task_group
486  *  - uclamp_update_active()	p->uclamp*
487  *
488  * p->state <- TASK_*:
489  *
490  *   is changed locklessly using set_current_state(), __set_current_state() or
491  *   set_special_state(), see their respective comments, or by
492  *   try_to_wake_up(). This latter uses p->pi_lock to serialize against
493  *   concurrent self.
494  *
495  * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
496  *
497  *   is set by activate_task() and cleared by deactivate_task(), under
498  *   rq->lock. Non-zero indicates the task is runnable, the special
499  *   ON_RQ_MIGRATING state is used for migration without holding both
500  *   rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
501  *
502  * p->on_cpu <- { 0, 1 }:
503  *
504  *   is set by prepare_task() and cleared by finish_task() such that it will be
505  *   set before p is scheduled-in and cleared after p is scheduled-out, both
506  *   under rq->lock. Non-zero indicates the task is running on its CPU.
507  *
508  *   [ The astute reader will observe that it is possible for two tasks on one
509  *     CPU to have ->on_cpu = 1 at the same time. ]
510  *
511  * task_cpu(p): is changed by set_task_cpu(), the rules are:
512  *
513  *  - Don't call set_task_cpu() on a blocked task:
514  *
515  *    We don't care what CPU we're not running on, this simplifies hotplug,
516  *    the CPU assignment of blocked tasks isn't required to be valid.
517  *
518  *  - for try_to_wake_up(), called under p->pi_lock:
519  *
520  *    This allows try_to_wake_up() to only take one rq->lock, see its comment.
521  *
522  *  - for migration called under rq->lock:
523  *    [ see task_on_rq_migrating() in task_rq_lock() ]
524  *
525  *    o move_queued_task()
526  *    o detach_task()
527  *
528  *  - for migration called under double_rq_lock():
529  *
530  *    o __migrate_swap_task()
531  *    o push_rt_task() / pull_rt_task()
532  *    o push_dl_task() / pull_dl_task()
533  *    o dl_task_offline_migration()
534  *
535  */
536 
raw_spin_rq_lock_nested(struct rq * rq,int subclass)537 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
538 {
539 	raw_spinlock_t *lock;
540 
541 	/* Matches synchronize_rcu() in __sched_core_enable() */
542 	preempt_disable();
543 	if (sched_core_disabled()) {
544 		raw_spin_lock_nested(&rq->__lock, subclass);
545 		/* preempt_count *MUST* be > 1 */
546 		preempt_enable_no_resched();
547 		return;
548 	}
549 
550 	for (;;) {
551 		lock = __rq_lockp(rq);
552 		raw_spin_lock_nested(lock, subclass);
553 		if (likely(lock == __rq_lockp(rq))) {
554 			/* preempt_count *MUST* be > 1 */
555 			preempt_enable_no_resched();
556 			return;
557 		}
558 		raw_spin_unlock(lock);
559 	}
560 }
561 
raw_spin_rq_trylock(struct rq * rq)562 bool raw_spin_rq_trylock(struct rq *rq)
563 {
564 	raw_spinlock_t *lock;
565 	bool ret;
566 
567 	/* Matches synchronize_rcu() in __sched_core_enable() */
568 	preempt_disable();
569 	if (sched_core_disabled()) {
570 		ret = raw_spin_trylock(&rq->__lock);
571 		preempt_enable();
572 		return ret;
573 	}
574 
575 	for (;;) {
576 		lock = __rq_lockp(rq);
577 		ret = raw_spin_trylock(lock);
578 		if (!ret || (likely(lock == __rq_lockp(rq)))) {
579 			preempt_enable();
580 			return ret;
581 		}
582 		raw_spin_unlock(lock);
583 	}
584 }
585 
raw_spin_rq_unlock(struct rq * rq)586 void raw_spin_rq_unlock(struct rq *rq)
587 {
588 	raw_spin_unlock(rq_lockp(rq));
589 }
590 
591 #ifdef CONFIG_SMP
592 /*
593  * double_rq_lock - safely lock two runqueues
594  */
double_rq_lock(struct rq * rq1,struct rq * rq2)595 void double_rq_lock(struct rq *rq1, struct rq *rq2)
596 {
597 	lockdep_assert_irqs_disabled();
598 
599 	if (rq_order_less(rq2, rq1))
600 		swap(rq1, rq2);
601 
602 	raw_spin_rq_lock(rq1);
603 	if (__rq_lockp(rq1) != __rq_lockp(rq2))
604 		raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
605 
606 	double_rq_clock_clear_update(rq1, rq2);
607 }
608 #endif
609 
610 /*
611  * __task_rq_lock - lock the rq @p resides on.
612  */
__task_rq_lock(struct task_struct * p,struct rq_flags * rf)613 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
614 	__acquires(rq->lock)
615 {
616 	struct rq *rq;
617 
618 	lockdep_assert_held(&p->pi_lock);
619 
620 	for (;;) {
621 		rq = task_rq(p);
622 		raw_spin_rq_lock(rq);
623 		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
624 			rq_pin_lock(rq, rf);
625 			return rq;
626 		}
627 		raw_spin_rq_unlock(rq);
628 
629 		while (unlikely(task_on_rq_migrating(p)))
630 			cpu_relax();
631 	}
632 }
633 
634 /*
635  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
636  */
task_rq_lock(struct task_struct * p,struct rq_flags * rf)637 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
638 	__acquires(p->pi_lock)
639 	__acquires(rq->lock)
640 {
641 	struct rq *rq;
642 
643 	for (;;) {
644 		raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
645 		rq = task_rq(p);
646 		raw_spin_rq_lock(rq);
647 		/*
648 		 *	move_queued_task()		task_rq_lock()
649 		 *
650 		 *	ACQUIRE (rq->lock)
651 		 *	[S] ->on_rq = MIGRATING		[L] rq = task_rq()
652 		 *	WMB (__set_task_cpu())		ACQUIRE (rq->lock);
653 		 *	[S] ->cpu = new_cpu		[L] task_rq()
654 		 *					[L] ->on_rq
655 		 *	RELEASE (rq->lock)
656 		 *
657 		 * If we observe the old CPU in task_rq_lock(), the acquire of
658 		 * the old rq->lock will fully serialize against the stores.
659 		 *
660 		 * If we observe the new CPU in task_rq_lock(), the address
661 		 * dependency headed by '[L] rq = task_rq()' and the acquire
662 		 * will pair with the WMB to ensure we then also see migrating.
663 		 */
664 		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
665 			rq_pin_lock(rq, rf);
666 			return rq;
667 		}
668 		raw_spin_rq_unlock(rq);
669 		raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
670 
671 		while (unlikely(task_on_rq_migrating(p)))
672 			cpu_relax();
673 	}
674 }
675 
676 /*
677  * RQ-clock updating methods:
678  */
679 
update_rq_clock_task(struct rq * rq,s64 delta)680 static void update_rq_clock_task(struct rq *rq, s64 delta)
681 {
682 /*
683  * In theory, the compile should just see 0 here, and optimize out the call
684  * to sched_rt_avg_update. But I don't trust it...
685  */
686 	s64 __maybe_unused steal = 0, irq_delta = 0;
687 
688 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
689 	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
690 
691 	/*
692 	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
693 	 * this case when a previous update_rq_clock() happened inside a
694 	 * {soft,}irq region.
695 	 *
696 	 * When this happens, we stop ->clock_task and only update the
697 	 * prev_irq_time stamp to account for the part that fit, so that a next
698 	 * update will consume the rest. This ensures ->clock_task is
699 	 * monotonic.
700 	 *
701 	 * It does however cause some slight miss-attribution of {soft,}irq
702 	 * time, a more accurate solution would be to update the irq_time using
703 	 * the current rq->clock timestamp, except that would require using
704 	 * atomic ops.
705 	 */
706 	if (irq_delta > delta)
707 		irq_delta = delta;
708 
709 	rq->prev_irq_time += irq_delta;
710 	delta -= irq_delta;
711 #endif
712 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
713 	if (static_key_false((&paravirt_steal_rq_enabled))) {
714 		steal = paravirt_steal_clock(cpu_of(rq));
715 		steal -= rq->prev_steal_time_rq;
716 
717 		if (unlikely(steal > delta))
718 			steal = delta;
719 
720 		rq->prev_steal_time_rq += steal;
721 		delta -= steal;
722 	}
723 #endif
724 
725 	rq->clock_task += delta;
726 
727 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
728 	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
729 		update_irq_load_avg(rq, irq_delta + steal);
730 #endif
731 	update_rq_clock_pelt(rq, delta);
732 }
733 
update_rq_clock(struct rq * rq)734 void update_rq_clock(struct rq *rq)
735 {
736 	s64 delta;
737 
738 	lockdep_assert_rq_held(rq);
739 
740 	if (rq->clock_update_flags & RQCF_ACT_SKIP)
741 		return;
742 
743 #ifdef CONFIG_SCHED_DEBUG
744 	if (sched_feat(WARN_DOUBLE_CLOCK))
745 		SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
746 	rq->clock_update_flags |= RQCF_UPDATED;
747 #endif
748 
749 	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
750 	if (delta < 0)
751 		return;
752 	rq->clock += delta;
753 	update_rq_clock_task(rq, delta);
754 }
755 
756 #ifdef CONFIG_SCHED_HRTICK
757 /*
758  * Use HR-timers to deliver accurate preemption points.
759  */
760 
hrtick_clear(struct rq * rq)761 static void hrtick_clear(struct rq *rq)
762 {
763 	if (hrtimer_active(&rq->hrtick_timer))
764 		hrtimer_cancel(&rq->hrtick_timer);
765 }
766 
767 /*
768  * High-resolution timer tick.
769  * Runs from hardirq context with interrupts disabled.
770  */
hrtick(struct hrtimer * timer)771 static enum hrtimer_restart hrtick(struct hrtimer *timer)
772 {
773 	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
774 	struct rq_flags rf;
775 
776 	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
777 
778 	rq_lock(rq, &rf);
779 	update_rq_clock(rq);
780 	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
781 	rq_unlock(rq, &rf);
782 
783 	return HRTIMER_NORESTART;
784 }
785 
786 #ifdef CONFIG_SMP
787 
__hrtick_restart(struct rq * rq)788 static void __hrtick_restart(struct rq *rq)
789 {
790 	struct hrtimer *timer = &rq->hrtick_timer;
791 	ktime_t time = rq->hrtick_time;
792 
793 	hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
794 }
795 
796 /*
797  * called from hardirq (IPI) context
798  */
__hrtick_start(void * arg)799 static void __hrtick_start(void *arg)
800 {
801 	struct rq *rq = arg;
802 	struct rq_flags rf;
803 
804 	rq_lock(rq, &rf);
805 	__hrtick_restart(rq);
806 	rq_unlock(rq, &rf);
807 }
808 
809 /*
810  * Called to set the hrtick timer state.
811  *
812  * called with rq->lock held and irqs disabled
813  */
hrtick_start(struct rq * rq,u64 delay)814 void hrtick_start(struct rq *rq, u64 delay)
815 {
816 	struct hrtimer *timer = &rq->hrtick_timer;
817 	s64 delta;
818 
819 	/*
820 	 * Don't schedule slices shorter than 10000ns, that just
821 	 * doesn't make sense and can cause timer DoS.
822 	 */
823 	delta = max_t(s64, delay, 10000LL);
824 	rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
825 
826 	if (rq == this_rq())
827 		__hrtick_restart(rq);
828 	else
829 		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
830 }
831 
832 #else
833 /*
834  * Called to set the hrtick timer state.
835  *
836  * called with rq->lock held and irqs disabled
837  */
hrtick_start(struct rq * rq,u64 delay)838 void hrtick_start(struct rq *rq, u64 delay)
839 {
840 	/*
841 	 * Don't schedule slices shorter than 10000ns, that just
842 	 * doesn't make sense. Rely on vruntime for fairness.
843 	 */
844 	delay = max_t(u64, delay, 10000LL);
845 	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
846 		      HRTIMER_MODE_REL_PINNED_HARD);
847 }
848 
849 #endif /* CONFIG_SMP */
850 
hrtick_rq_init(struct rq * rq)851 static void hrtick_rq_init(struct rq *rq)
852 {
853 #ifdef CONFIG_SMP
854 	INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
855 #endif
856 	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
857 	rq->hrtick_timer.function = hrtick;
858 }
859 #else	/* CONFIG_SCHED_HRTICK */
hrtick_clear(struct rq * rq)860 static inline void hrtick_clear(struct rq *rq)
861 {
862 }
863 
hrtick_rq_init(struct rq * rq)864 static inline void hrtick_rq_init(struct rq *rq)
865 {
866 }
867 #endif	/* CONFIG_SCHED_HRTICK */
868 
869 /*
870  * cmpxchg based fetch_or, macro so it works for different integer types
871  */
872 #define fetch_or(ptr, mask)						\
873 	({								\
874 		typeof(ptr) _ptr = (ptr);				\
875 		typeof(mask) _mask = (mask);				\
876 		typeof(*_ptr) _old, _val = *_ptr;			\
877 									\
878 		for (;;) {						\
879 			_old = cmpxchg(_ptr, _val, _val | _mask);	\
880 			if (_old == _val)				\
881 				break;					\
882 			_val = _old;					\
883 		}							\
884 	_old;								\
885 })
886 
887 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
888 /*
889  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
890  * this avoids any races wrt polling state changes and thereby avoids
891  * spurious IPIs.
892  */
set_nr_and_not_polling(struct task_struct * p)893 static bool set_nr_and_not_polling(struct task_struct *p)
894 {
895 	struct thread_info *ti = task_thread_info(p);
896 	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
897 }
898 
899 /*
900  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
901  *
902  * If this returns true, then the idle task promises to call
903  * sched_ttwu_pending() and reschedule soon.
904  */
set_nr_if_polling(struct task_struct * p)905 static bool set_nr_if_polling(struct task_struct *p)
906 {
907 	struct thread_info *ti = task_thread_info(p);
908 	typeof(ti->flags) old, val = READ_ONCE(ti->flags);
909 
910 	for (;;) {
911 		if (!(val & _TIF_POLLING_NRFLAG))
912 			return false;
913 		if (val & _TIF_NEED_RESCHED)
914 			return true;
915 		old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
916 		if (old == val)
917 			break;
918 		val = old;
919 	}
920 	return true;
921 }
922 
923 #else
set_nr_and_not_polling(struct task_struct * p)924 static bool set_nr_and_not_polling(struct task_struct *p)
925 {
926 	set_tsk_need_resched(p);
927 	return true;
928 }
929 
930 #ifdef CONFIG_SMP
set_nr_if_polling(struct task_struct * p)931 static bool set_nr_if_polling(struct task_struct *p)
932 {
933 	return false;
934 }
935 #endif
936 #endif
937 
__wake_q_add(struct wake_q_head * head,struct task_struct * task)938 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
939 {
940 	struct wake_q_node *node = &task->wake_q;
941 
942 	/*
943 	 * Atomically grab the task, if ->wake_q is !nil already it means
944 	 * it's already queued (either by us or someone else) and will get the
945 	 * wakeup due to that.
946 	 *
947 	 * In order to ensure that a pending wakeup will observe our pending
948 	 * state, even in the failed case, an explicit smp_mb() must be used.
949 	 */
950 	smp_mb__before_atomic();
951 	if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
952 		return false;
953 
954 	/*
955 	 * The head is context local, there can be no concurrency.
956 	 */
957 	*head->lastp = node;
958 	head->lastp = &node->next;
959 	return true;
960 }
961 
962 /**
963  * wake_q_add() - queue a wakeup for 'later' waking.
964  * @head: the wake_q_head to add @task to
965  * @task: the task to queue for 'later' wakeup
966  *
967  * Queue a task for later wakeup, most likely by the wake_up_q() call in the
968  * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
969  * instantly.
970  *
971  * This function must be used as-if it were wake_up_process(); IOW the task
972  * must be ready to be woken at this location.
973  */
wake_q_add(struct wake_q_head * head,struct task_struct * task)974 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
975 {
976 	if (__wake_q_add(head, task))
977 		get_task_struct(task);
978 }
979 
980 /**
981  * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
982  * @head: the wake_q_head to add @task to
983  * @task: the task to queue for 'later' wakeup
984  *
985  * Queue a task for later wakeup, most likely by the wake_up_q() call in the
986  * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
987  * instantly.
988  *
989  * This function must be used as-if it were wake_up_process(); IOW the task
990  * must be ready to be woken at this location.
991  *
992  * This function is essentially a task-safe equivalent to wake_q_add(). Callers
993  * that already hold reference to @task can call the 'safe' version and trust
994  * wake_q to do the right thing depending whether or not the @task is already
995  * queued for wakeup.
996  */
wake_q_add_safe(struct wake_q_head * head,struct task_struct * task)997 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
998 {
999 	if (!__wake_q_add(head, task))
1000 		put_task_struct(task);
1001 }
1002 
wake_up_q(struct wake_q_head * head)1003 void wake_up_q(struct wake_q_head *head)
1004 {
1005 	struct wake_q_node *node = head->first;
1006 
1007 	while (node != WAKE_Q_TAIL) {
1008 		struct task_struct *task;
1009 
1010 		task = container_of(node, struct task_struct, wake_q);
1011 		/* Task can safely be re-inserted now: */
1012 		node = node->next;
1013 		task->wake_q.next = NULL;
1014 
1015 		/*
1016 		 * wake_up_process() executes a full barrier, which pairs with
1017 		 * the queueing in wake_q_add() so as not to miss wakeups.
1018 		 */
1019 		wake_up_process(task);
1020 		put_task_struct(task);
1021 	}
1022 }
1023 
1024 /*
1025  * resched_curr - mark rq's current task 'to be rescheduled now'.
1026  *
1027  * On UP this means the setting of the need_resched flag, on SMP it
1028  * might also involve a cross-CPU call to trigger the scheduler on
1029  * the target CPU.
1030  */
resched_curr(struct rq * rq)1031 void resched_curr(struct rq *rq)
1032 {
1033 	struct task_struct *curr = rq->curr;
1034 	int cpu;
1035 
1036 	lockdep_assert_rq_held(rq);
1037 
1038 	if (test_tsk_need_resched(curr))
1039 		return;
1040 
1041 	cpu = cpu_of(rq);
1042 
1043 	if (cpu == smp_processor_id()) {
1044 		set_tsk_need_resched(curr);
1045 		set_preempt_need_resched();
1046 		return;
1047 	}
1048 
1049 	if (set_nr_and_not_polling(curr))
1050 		smp_send_reschedule(cpu);
1051 	else
1052 		trace_sched_wake_idle_without_ipi(cpu);
1053 }
1054 
resched_cpu(int cpu)1055 void resched_cpu(int cpu)
1056 {
1057 	struct rq *rq = cpu_rq(cpu);
1058 	unsigned long flags;
1059 
1060 	raw_spin_rq_lock_irqsave(rq, flags);
1061 	if (cpu_online(cpu) || cpu == smp_processor_id())
1062 		resched_curr(rq);
1063 	raw_spin_rq_unlock_irqrestore(rq, flags);
1064 }
1065 
1066 #ifdef CONFIG_SMP
1067 #ifdef CONFIG_NO_HZ_COMMON
1068 /*
1069  * In the semi idle case, use the nearest busy CPU for migrating timers
1070  * from an idle CPU.  This is good for power-savings.
1071  *
1072  * We don't do similar optimization for completely idle system, as
1073  * selecting an idle CPU will add more delays to the timers than intended
1074  * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1075  */
get_nohz_timer_target(void)1076 int get_nohz_timer_target(void)
1077 {
1078 	int i, cpu = smp_processor_id(), default_cpu = -1;
1079 	struct sched_domain *sd;
1080 	const struct cpumask *hk_mask;
1081 
1082 	if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1083 		if (!idle_cpu(cpu))
1084 			return cpu;
1085 		default_cpu = cpu;
1086 	}
1087 
1088 	hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1089 
1090 	rcu_read_lock();
1091 	for_each_domain(cpu, sd) {
1092 		for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1093 			if (cpu == i)
1094 				continue;
1095 
1096 			if (!idle_cpu(i)) {
1097 				cpu = i;
1098 				goto unlock;
1099 			}
1100 		}
1101 	}
1102 
1103 	if (default_cpu == -1)
1104 		default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1105 	cpu = default_cpu;
1106 unlock:
1107 	rcu_read_unlock();
1108 	return cpu;
1109 }
1110 
1111 /*
1112  * When add_timer_on() enqueues a timer into the timer wheel of an
1113  * idle CPU then this timer might expire before the next timer event
1114  * which is scheduled to wake up that CPU. In case of a completely
1115  * idle system the next event might even be infinite time into the
1116  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1117  * leaves the inner idle loop so the newly added timer is taken into
1118  * account when the CPU goes back to idle and evaluates the timer
1119  * wheel for the next timer event.
1120  */
wake_up_idle_cpu(int cpu)1121 static void wake_up_idle_cpu(int cpu)
1122 {
1123 	struct rq *rq = cpu_rq(cpu);
1124 
1125 	if (cpu == smp_processor_id())
1126 		return;
1127 
1128 	if (set_nr_and_not_polling(rq->idle))
1129 		smp_send_reschedule(cpu);
1130 	else
1131 		trace_sched_wake_idle_without_ipi(cpu);
1132 }
1133 
wake_up_full_nohz_cpu(int cpu)1134 static bool wake_up_full_nohz_cpu(int cpu)
1135 {
1136 	/*
1137 	 * We just need the target to call irq_exit() and re-evaluate
1138 	 * the next tick. The nohz full kick at least implies that.
1139 	 * If needed we can still optimize that later with an
1140 	 * empty IRQ.
1141 	 */
1142 	if (cpu_is_offline(cpu))
1143 		return true;  /* Don't try to wake offline CPUs. */
1144 	if (tick_nohz_full_cpu(cpu)) {
1145 		if (cpu != smp_processor_id() ||
1146 		    tick_nohz_tick_stopped())
1147 			tick_nohz_full_kick_cpu(cpu);
1148 		return true;
1149 	}
1150 
1151 	return false;
1152 }
1153 
1154 /*
1155  * Wake up the specified CPU.  If the CPU is going offline, it is the
1156  * caller's responsibility to deal with the lost wakeup, for example,
1157  * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1158  */
wake_up_nohz_cpu(int cpu)1159 void wake_up_nohz_cpu(int cpu)
1160 {
1161 	if (!wake_up_full_nohz_cpu(cpu))
1162 		wake_up_idle_cpu(cpu);
1163 }
1164 
nohz_csd_func(void * info)1165 static void nohz_csd_func(void *info)
1166 {
1167 	struct rq *rq = info;
1168 	int cpu = cpu_of(rq);
1169 	unsigned int flags;
1170 
1171 	/*
1172 	 * Release the rq::nohz_csd.
1173 	 */
1174 	flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1175 	WARN_ON(!(flags & NOHZ_KICK_MASK));
1176 
1177 	rq->idle_balance = idle_cpu(cpu);
1178 	if (rq->idle_balance && !need_resched()) {
1179 		rq->nohz_idle_balance = flags;
1180 		raise_softirq_irqoff(SCHED_SOFTIRQ);
1181 	}
1182 }
1183 
1184 #endif /* CONFIG_NO_HZ_COMMON */
1185 
1186 #ifdef CONFIG_NO_HZ_FULL
sched_can_stop_tick(struct rq * rq)1187 bool sched_can_stop_tick(struct rq *rq)
1188 {
1189 	int fifo_nr_running;
1190 
1191 	/* Deadline tasks, even if single, need the tick */
1192 	if (rq->dl.dl_nr_running)
1193 		return false;
1194 
1195 	/*
1196 	 * If there are more than one RR tasks, we need the tick to affect the
1197 	 * actual RR behaviour.
1198 	 */
1199 	if (rq->rt.rr_nr_running) {
1200 		if (rq->rt.rr_nr_running == 1)
1201 			return true;
1202 		else
1203 			return false;
1204 	}
1205 
1206 	/*
1207 	 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1208 	 * forced preemption between FIFO tasks.
1209 	 */
1210 	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1211 	if (fifo_nr_running)
1212 		return true;
1213 
1214 	/*
1215 	 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1216 	 * if there's more than one we need the tick for involuntary
1217 	 * preemption.
1218 	 */
1219 	if (rq->nr_running > 1)
1220 		return false;
1221 
1222 	return true;
1223 }
1224 #endif /* CONFIG_NO_HZ_FULL */
1225 #endif /* CONFIG_SMP */
1226 
1227 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1228 			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1229 /*
1230  * Iterate task_group tree rooted at *from, calling @down when first entering a
1231  * node and @up when leaving it for the final time.
1232  *
1233  * Caller must hold rcu_lock or sufficient equivalent.
1234  */
walk_tg_tree_from(struct task_group * from,tg_visitor down,tg_visitor up,void * data)1235 int walk_tg_tree_from(struct task_group *from,
1236 			     tg_visitor down, tg_visitor up, void *data)
1237 {
1238 	struct task_group *parent, *child;
1239 	int ret;
1240 
1241 	parent = from;
1242 
1243 down:
1244 	ret = (*down)(parent, data);
1245 	if (ret)
1246 		goto out;
1247 	list_for_each_entry_rcu(child, &parent->children, siblings) {
1248 		parent = child;
1249 		goto down;
1250 
1251 up:
1252 		continue;
1253 	}
1254 	ret = (*up)(parent, data);
1255 	if (ret || parent == from)
1256 		goto out;
1257 
1258 	child = parent;
1259 	parent = parent->parent;
1260 	if (parent)
1261 		goto up;
1262 out:
1263 	return ret;
1264 }
1265 
tg_nop(struct task_group * tg,void * data)1266 int tg_nop(struct task_group *tg, void *data)
1267 {
1268 	return 0;
1269 }
1270 #endif
1271 
set_load_weight(struct task_struct * p,bool update_load)1272 static void set_load_weight(struct task_struct *p, bool update_load)
1273 {
1274 	int prio = p->static_prio - MAX_RT_PRIO;
1275 	struct load_weight *load = &p->se.load;
1276 
1277 	/*
1278 	 * SCHED_IDLE tasks get minimal weight:
1279 	 */
1280 	if (task_has_idle_policy(p)) {
1281 		load->weight = scale_load(WEIGHT_IDLEPRIO);
1282 		load->inv_weight = WMULT_IDLEPRIO;
1283 		return;
1284 	}
1285 
1286 	/*
1287 	 * SCHED_OTHER tasks have to update their load when changing their
1288 	 * weight
1289 	 */
1290 	if (update_load && p->sched_class == &fair_sched_class) {
1291 		reweight_task(p, prio);
1292 	} else {
1293 		load->weight = scale_load(sched_prio_to_weight[prio]);
1294 		load->inv_weight = sched_prio_to_wmult[prio];
1295 	}
1296 }
1297 
1298 #ifdef CONFIG_UCLAMP_TASK
1299 /*
1300  * Serializes updates of utilization clamp values
1301  *
1302  * The (slow-path) user-space triggers utilization clamp value updates which
1303  * can require updates on (fast-path) scheduler's data structures used to
1304  * support enqueue/dequeue operations.
1305  * While the per-CPU rq lock protects fast-path update operations, user-space
1306  * requests are serialized using a mutex to reduce the risk of conflicting
1307  * updates or API abuses.
1308  */
1309 static DEFINE_MUTEX(uclamp_mutex);
1310 
1311 /* Max allowed minimum utilization */
1312 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1313 
1314 /* Max allowed maximum utilization */
1315 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1316 
1317 /*
1318  * By default RT tasks run at the maximum performance point/capacity of the
1319  * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1320  * SCHED_CAPACITY_SCALE.
1321  *
1322  * This knob allows admins to change the default behavior when uclamp is being
1323  * used. In battery powered devices, particularly, running at the maximum
1324  * capacity and frequency will increase energy consumption and shorten the
1325  * battery life.
1326  *
1327  * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1328  *
1329  * This knob will not override the system default sched_util_clamp_min defined
1330  * above.
1331  */
1332 static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1333 
1334 /* All clamps are required to be less or equal than these values */
1335 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1336 
1337 /*
1338  * This static key is used to reduce the uclamp overhead in the fast path. It
1339  * primarily disables the call to uclamp_rq_{inc, dec}() in
1340  * enqueue/dequeue_task().
1341  *
1342  * This allows users to continue to enable uclamp in their kernel config with
1343  * minimum uclamp overhead in the fast path.
1344  *
1345  * As soon as userspace modifies any of the uclamp knobs, the static key is
1346  * enabled, since we have an actual users that make use of uclamp
1347  * functionality.
1348  *
1349  * The knobs that would enable this static key are:
1350  *
1351  *   * A task modifying its uclamp value with sched_setattr().
1352  *   * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1353  *   * An admin modifying the cgroup cpu.uclamp.{min, max}
1354  */
1355 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1356 
1357 /* Integer rounded range for each bucket */
1358 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1359 
1360 #define for_each_clamp_id(clamp_id) \
1361 	for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1362 
uclamp_bucket_id(unsigned int clamp_value)1363 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1364 {
1365 	return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1366 }
1367 
uclamp_none(enum uclamp_id clamp_id)1368 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1369 {
1370 	if (clamp_id == UCLAMP_MIN)
1371 		return 0;
1372 	return SCHED_CAPACITY_SCALE;
1373 }
1374 
uclamp_se_set(struct uclamp_se * uc_se,unsigned int value,bool user_defined)1375 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1376 				 unsigned int value, bool user_defined)
1377 {
1378 	uc_se->value = value;
1379 	uc_se->bucket_id = uclamp_bucket_id(value);
1380 	uc_se->user_defined = user_defined;
1381 }
1382 
1383 static inline unsigned int
uclamp_idle_value(struct rq * rq,enum uclamp_id clamp_id,unsigned int clamp_value)1384 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1385 		  unsigned int clamp_value)
1386 {
1387 	/*
1388 	 * Avoid blocked utilization pushing up the frequency when we go
1389 	 * idle (which drops the max-clamp) by retaining the last known
1390 	 * max-clamp.
1391 	 */
1392 	if (clamp_id == UCLAMP_MAX) {
1393 		rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1394 		return clamp_value;
1395 	}
1396 
1397 	return uclamp_none(UCLAMP_MIN);
1398 }
1399 
uclamp_idle_reset(struct rq * rq,enum uclamp_id clamp_id,unsigned int clamp_value)1400 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1401 				     unsigned int clamp_value)
1402 {
1403 	/* Reset max-clamp retention only on idle exit */
1404 	if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1405 		return;
1406 
1407 	WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1408 }
1409 
1410 static inline
uclamp_rq_max_value(struct rq * rq,enum uclamp_id clamp_id,unsigned int clamp_value)1411 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1412 				   unsigned int clamp_value)
1413 {
1414 	struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1415 	int bucket_id = UCLAMP_BUCKETS - 1;
1416 
1417 	/*
1418 	 * Since both min and max clamps are max aggregated, find the
1419 	 * top most bucket with tasks in.
1420 	 */
1421 	for ( ; bucket_id >= 0; bucket_id--) {
1422 		if (!bucket[bucket_id].tasks)
1423 			continue;
1424 		return bucket[bucket_id].value;
1425 	}
1426 
1427 	/* No tasks -- default clamp values */
1428 	return uclamp_idle_value(rq, clamp_id, clamp_value);
1429 }
1430 
__uclamp_update_util_min_rt_default(struct task_struct * p)1431 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1432 {
1433 	unsigned int default_util_min;
1434 	struct uclamp_se *uc_se;
1435 
1436 	lockdep_assert_held(&p->pi_lock);
1437 
1438 	uc_se = &p->uclamp_req[UCLAMP_MIN];
1439 
1440 	/* Only sync if user didn't override the default */
1441 	if (uc_se->user_defined)
1442 		return;
1443 
1444 	default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1445 	uclamp_se_set(uc_se, default_util_min, false);
1446 }
1447 
uclamp_update_util_min_rt_default(struct task_struct * p)1448 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1449 {
1450 	struct rq_flags rf;
1451 	struct rq *rq;
1452 
1453 	if (!rt_task(p))
1454 		return;
1455 
1456 	/* Protect updates to p->uclamp_* */
1457 	rq = task_rq_lock(p, &rf);
1458 	__uclamp_update_util_min_rt_default(p);
1459 	task_rq_unlock(rq, p, &rf);
1460 }
1461 
1462 static inline struct uclamp_se
uclamp_tg_restrict(struct task_struct * p,enum uclamp_id clamp_id)1463 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1464 {
1465 	/* Copy by value as we could modify it */
1466 	struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1467 #ifdef CONFIG_UCLAMP_TASK_GROUP
1468 	unsigned int tg_min, tg_max, value;
1469 
1470 	/*
1471 	 * Tasks in autogroups or root task group will be
1472 	 * restricted by system defaults.
1473 	 */
1474 	if (task_group_is_autogroup(task_group(p)))
1475 		return uc_req;
1476 	if (task_group(p) == &root_task_group)
1477 		return uc_req;
1478 
1479 	tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1480 	tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1481 	value = uc_req.value;
1482 	value = clamp(value, tg_min, tg_max);
1483 	uclamp_se_set(&uc_req, value, false);
1484 #endif
1485 
1486 	return uc_req;
1487 }
1488 
1489 /*
1490  * The effective clamp bucket index of a task depends on, by increasing
1491  * priority:
1492  * - the task specific clamp value, when explicitly requested from userspace
1493  * - the task group effective clamp value, for tasks not either in the root
1494  *   group or in an autogroup
1495  * - the system default clamp value, defined by the sysadmin
1496  */
1497 static inline struct uclamp_se
uclamp_eff_get(struct task_struct * p,enum uclamp_id clamp_id)1498 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1499 {
1500 	struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1501 	struct uclamp_se uc_max = uclamp_default[clamp_id];
1502 
1503 	/* System default restrictions always apply */
1504 	if (unlikely(uc_req.value > uc_max.value))
1505 		return uc_max;
1506 
1507 	return uc_req;
1508 }
1509 
uclamp_eff_value(struct task_struct * p,enum uclamp_id clamp_id)1510 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1511 {
1512 	struct uclamp_se uc_eff;
1513 
1514 	/* Task currently refcounted: use back-annotated (effective) value */
1515 	if (p->uclamp[clamp_id].active)
1516 		return (unsigned long)p->uclamp[clamp_id].value;
1517 
1518 	uc_eff = uclamp_eff_get(p, clamp_id);
1519 
1520 	return (unsigned long)uc_eff.value;
1521 }
1522 
1523 /*
1524  * When a task is enqueued on a rq, the clamp bucket currently defined by the
1525  * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1526  * updates the rq's clamp value if required.
1527  *
1528  * Tasks can have a task-specific value requested from user-space, track
1529  * within each bucket the maximum value for tasks refcounted in it.
1530  * This "local max aggregation" allows to track the exact "requested" value
1531  * for each bucket when all its RUNNABLE tasks require the same clamp.
1532  */
uclamp_rq_inc_id(struct rq * rq,struct task_struct * p,enum uclamp_id clamp_id)1533 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1534 				    enum uclamp_id clamp_id)
1535 {
1536 	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1537 	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1538 	struct uclamp_bucket *bucket;
1539 
1540 	lockdep_assert_rq_held(rq);
1541 
1542 	/* Update task effective clamp */
1543 	p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1544 
1545 	bucket = &uc_rq->bucket[uc_se->bucket_id];
1546 	bucket->tasks++;
1547 	uc_se->active = true;
1548 
1549 	uclamp_idle_reset(rq, clamp_id, uc_se->value);
1550 
1551 	/*
1552 	 * Local max aggregation: rq buckets always track the max
1553 	 * "requested" clamp value of its RUNNABLE tasks.
1554 	 */
1555 	if (bucket->tasks == 1 || uc_se->value > bucket->value)
1556 		bucket->value = uc_se->value;
1557 
1558 	if (uc_se->value > READ_ONCE(uc_rq->value))
1559 		WRITE_ONCE(uc_rq->value, uc_se->value);
1560 }
1561 
1562 /*
1563  * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1564  * is released. If this is the last task reference counting the rq's max
1565  * active clamp value, then the rq's clamp value is updated.
1566  *
1567  * Both refcounted tasks and rq's cached clamp values are expected to be
1568  * always valid. If it's detected they are not, as defensive programming,
1569  * enforce the expected state and warn.
1570  */
uclamp_rq_dec_id(struct rq * rq,struct task_struct * p,enum uclamp_id clamp_id)1571 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1572 				    enum uclamp_id clamp_id)
1573 {
1574 	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1575 	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1576 	struct uclamp_bucket *bucket;
1577 	unsigned int bkt_clamp;
1578 	unsigned int rq_clamp;
1579 
1580 	lockdep_assert_rq_held(rq);
1581 
1582 	/*
1583 	 * If sched_uclamp_used was enabled after task @p was enqueued,
1584 	 * we could end up with unbalanced call to uclamp_rq_dec_id().
1585 	 *
1586 	 * In this case the uc_se->active flag should be false since no uclamp
1587 	 * accounting was performed at enqueue time and we can just return
1588 	 * here.
1589 	 *
1590 	 * Need to be careful of the following enqueue/dequeue ordering
1591 	 * problem too
1592 	 *
1593 	 *	enqueue(taskA)
1594 	 *	// sched_uclamp_used gets enabled
1595 	 *	enqueue(taskB)
1596 	 *	dequeue(taskA)
1597 	 *	// Must not decrement bucket->tasks here
1598 	 *	dequeue(taskB)
1599 	 *
1600 	 * where we could end up with stale data in uc_se and
1601 	 * bucket[uc_se->bucket_id].
1602 	 *
1603 	 * The following check here eliminates the possibility of such race.
1604 	 */
1605 	if (unlikely(!uc_se->active))
1606 		return;
1607 
1608 	bucket = &uc_rq->bucket[uc_se->bucket_id];
1609 
1610 	SCHED_WARN_ON(!bucket->tasks);
1611 	if (likely(bucket->tasks))
1612 		bucket->tasks--;
1613 
1614 	uc_se->active = false;
1615 
1616 	/*
1617 	 * Keep "local max aggregation" simple and accept to (possibly)
1618 	 * overboost some RUNNABLE tasks in the same bucket.
1619 	 * The rq clamp bucket value is reset to its base value whenever
1620 	 * there are no more RUNNABLE tasks refcounting it.
1621 	 */
1622 	if (likely(bucket->tasks))
1623 		return;
1624 
1625 	rq_clamp = READ_ONCE(uc_rq->value);
1626 	/*
1627 	 * Defensive programming: this should never happen. If it happens,
1628 	 * e.g. due to future modification, warn and fixup the expected value.
1629 	 */
1630 	SCHED_WARN_ON(bucket->value > rq_clamp);
1631 	if (bucket->value >= rq_clamp) {
1632 		bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1633 		WRITE_ONCE(uc_rq->value, bkt_clamp);
1634 	}
1635 }
1636 
uclamp_rq_inc(struct rq * rq,struct task_struct * p)1637 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1638 {
1639 	enum uclamp_id clamp_id;
1640 
1641 	/*
1642 	 * Avoid any overhead until uclamp is actually used by the userspace.
1643 	 *
1644 	 * The condition is constructed such that a NOP is generated when
1645 	 * sched_uclamp_used is disabled.
1646 	 */
1647 	if (!static_branch_unlikely(&sched_uclamp_used))
1648 		return;
1649 
1650 	if (unlikely(!p->sched_class->uclamp_enabled))
1651 		return;
1652 
1653 	for_each_clamp_id(clamp_id)
1654 		uclamp_rq_inc_id(rq, p, clamp_id);
1655 
1656 	/* Reset clamp idle holding when there is one RUNNABLE task */
1657 	if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1658 		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1659 }
1660 
uclamp_rq_dec(struct rq * rq,struct task_struct * p)1661 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1662 {
1663 	enum uclamp_id clamp_id;
1664 
1665 	/*
1666 	 * Avoid any overhead until uclamp is actually used by the userspace.
1667 	 *
1668 	 * The condition is constructed such that a NOP is generated when
1669 	 * sched_uclamp_used is disabled.
1670 	 */
1671 	if (!static_branch_unlikely(&sched_uclamp_used))
1672 		return;
1673 
1674 	if (unlikely(!p->sched_class->uclamp_enabled))
1675 		return;
1676 
1677 	for_each_clamp_id(clamp_id)
1678 		uclamp_rq_dec_id(rq, p, clamp_id);
1679 }
1680 
uclamp_rq_reinc_id(struct rq * rq,struct task_struct * p,enum uclamp_id clamp_id)1681 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1682 				      enum uclamp_id clamp_id)
1683 {
1684 	if (!p->uclamp[clamp_id].active)
1685 		return;
1686 
1687 	uclamp_rq_dec_id(rq, p, clamp_id);
1688 	uclamp_rq_inc_id(rq, p, clamp_id);
1689 
1690 	/*
1691 	 * Make sure to clear the idle flag if we've transiently reached 0
1692 	 * active tasks on rq.
1693 	 */
1694 	if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1695 		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1696 }
1697 
1698 static inline void
uclamp_update_active(struct task_struct * p)1699 uclamp_update_active(struct task_struct *p)
1700 {
1701 	enum uclamp_id clamp_id;
1702 	struct rq_flags rf;
1703 	struct rq *rq;
1704 
1705 	/*
1706 	 * Lock the task and the rq where the task is (or was) queued.
1707 	 *
1708 	 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1709 	 * price to pay to safely serialize util_{min,max} updates with
1710 	 * enqueues, dequeues and migration operations.
1711 	 * This is the same locking schema used by __set_cpus_allowed_ptr().
1712 	 */
1713 	rq = task_rq_lock(p, &rf);
1714 
1715 	/*
1716 	 * Setting the clamp bucket is serialized by task_rq_lock().
1717 	 * If the task is not yet RUNNABLE and its task_struct is not
1718 	 * affecting a valid clamp bucket, the next time it's enqueued,
1719 	 * it will already see the updated clamp bucket value.
1720 	 */
1721 	for_each_clamp_id(clamp_id)
1722 		uclamp_rq_reinc_id(rq, p, clamp_id);
1723 
1724 	task_rq_unlock(rq, p, &rf);
1725 }
1726 
1727 #ifdef CONFIG_UCLAMP_TASK_GROUP
1728 static inline void
uclamp_update_active_tasks(struct cgroup_subsys_state * css)1729 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1730 {
1731 	struct css_task_iter it;
1732 	struct task_struct *p;
1733 
1734 	css_task_iter_start(css, 0, &it);
1735 	while ((p = css_task_iter_next(&it)))
1736 		uclamp_update_active(p);
1737 	css_task_iter_end(&it);
1738 }
1739 
1740 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1741 #endif
1742 
1743 #ifdef CONFIG_SYSCTL
1744 #ifdef CONFIG_UCLAMP_TASK
1745 #ifdef CONFIG_UCLAMP_TASK_GROUP
uclamp_update_root_tg(void)1746 static void uclamp_update_root_tg(void)
1747 {
1748 	struct task_group *tg = &root_task_group;
1749 
1750 	uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1751 		      sysctl_sched_uclamp_util_min, false);
1752 	uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1753 		      sysctl_sched_uclamp_util_max, false);
1754 
1755 	rcu_read_lock();
1756 	cpu_util_update_eff(&root_task_group.css);
1757 	rcu_read_unlock();
1758 }
1759 #else
uclamp_update_root_tg(void)1760 static void uclamp_update_root_tg(void) { }
1761 #endif
1762 
uclamp_sync_util_min_rt_default(void)1763 static void uclamp_sync_util_min_rt_default(void)
1764 {
1765 	struct task_struct *g, *p;
1766 
1767 	/*
1768 	 * copy_process()			sysctl_uclamp
1769 	 *					  uclamp_min_rt = X;
1770 	 *   write_lock(&tasklist_lock)		  read_lock(&tasklist_lock)
1771 	 *   // link thread			  smp_mb__after_spinlock()
1772 	 *   write_unlock(&tasklist_lock)	  read_unlock(&tasklist_lock);
1773 	 *   sched_post_fork()			  for_each_process_thread()
1774 	 *     __uclamp_sync_rt()		    __uclamp_sync_rt()
1775 	 *
1776 	 * Ensures that either sched_post_fork() will observe the new
1777 	 * uclamp_min_rt or for_each_process_thread() will observe the new
1778 	 * task.
1779 	 */
1780 	read_lock(&tasklist_lock);
1781 	smp_mb__after_spinlock();
1782 	read_unlock(&tasklist_lock);
1783 
1784 	rcu_read_lock();
1785 	for_each_process_thread(g, p)
1786 		uclamp_update_util_min_rt_default(p);
1787 	rcu_read_unlock();
1788 }
1789 
sysctl_sched_uclamp_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)1790 static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1791 				void *buffer, size_t *lenp, loff_t *ppos)
1792 {
1793 	bool update_root_tg = false;
1794 	int old_min, old_max, old_min_rt;
1795 	int result;
1796 
1797 	mutex_lock(&uclamp_mutex);
1798 	old_min = sysctl_sched_uclamp_util_min;
1799 	old_max = sysctl_sched_uclamp_util_max;
1800 	old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1801 
1802 	result = proc_dointvec(table, write, buffer, lenp, ppos);
1803 	if (result)
1804 		goto undo;
1805 	if (!write)
1806 		goto done;
1807 
1808 	if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1809 	    sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE	||
1810 	    sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1811 
1812 		result = -EINVAL;
1813 		goto undo;
1814 	}
1815 
1816 	if (old_min != sysctl_sched_uclamp_util_min) {
1817 		uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1818 			      sysctl_sched_uclamp_util_min, false);
1819 		update_root_tg = true;
1820 	}
1821 	if (old_max != sysctl_sched_uclamp_util_max) {
1822 		uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1823 			      sysctl_sched_uclamp_util_max, false);
1824 		update_root_tg = true;
1825 	}
1826 
1827 	if (update_root_tg) {
1828 		static_branch_enable(&sched_uclamp_used);
1829 		uclamp_update_root_tg();
1830 	}
1831 
1832 	if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1833 		static_branch_enable(&sched_uclamp_used);
1834 		uclamp_sync_util_min_rt_default();
1835 	}
1836 
1837 	/*
1838 	 * We update all RUNNABLE tasks only when task groups are in use.
1839 	 * Otherwise, keep it simple and do just a lazy update at each next
1840 	 * task enqueue time.
1841 	 */
1842 
1843 	goto done;
1844 
1845 undo:
1846 	sysctl_sched_uclamp_util_min = old_min;
1847 	sysctl_sched_uclamp_util_max = old_max;
1848 	sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1849 done:
1850 	mutex_unlock(&uclamp_mutex);
1851 
1852 	return result;
1853 }
1854 #endif
1855 #endif
1856 
uclamp_validate(struct task_struct * p,const struct sched_attr * attr)1857 static int uclamp_validate(struct task_struct *p,
1858 			   const struct sched_attr *attr)
1859 {
1860 	int util_min = p->uclamp_req[UCLAMP_MIN].value;
1861 	int util_max = p->uclamp_req[UCLAMP_MAX].value;
1862 
1863 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1864 		util_min = attr->sched_util_min;
1865 
1866 		if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1867 			return -EINVAL;
1868 	}
1869 
1870 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1871 		util_max = attr->sched_util_max;
1872 
1873 		if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1874 			return -EINVAL;
1875 	}
1876 
1877 	if (util_min != -1 && util_max != -1 && util_min > util_max)
1878 		return -EINVAL;
1879 
1880 	/*
1881 	 * We have valid uclamp attributes; make sure uclamp is enabled.
1882 	 *
1883 	 * We need to do that here, because enabling static branches is a
1884 	 * blocking operation which obviously cannot be done while holding
1885 	 * scheduler locks.
1886 	 */
1887 	static_branch_enable(&sched_uclamp_used);
1888 
1889 	return 0;
1890 }
1891 
uclamp_reset(const struct sched_attr * attr,enum uclamp_id clamp_id,struct uclamp_se * uc_se)1892 static bool uclamp_reset(const struct sched_attr *attr,
1893 			 enum uclamp_id clamp_id,
1894 			 struct uclamp_se *uc_se)
1895 {
1896 	/* Reset on sched class change for a non user-defined clamp value. */
1897 	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1898 	    !uc_se->user_defined)
1899 		return true;
1900 
1901 	/* Reset on sched_util_{min,max} == -1. */
1902 	if (clamp_id == UCLAMP_MIN &&
1903 	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1904 	    attr->sched_util_min == -1) {
1905 		return true;
1906 	}
1907 
1908 	if (clamp_id == UCLAMP_MAX &&
1909 	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1910 	    attr->sched_util_max == -1) {
1911 		return true;
1912 	}
1913 
1914 	return false;
1915 }
1916 
__setscheduler_uclamp(struct task_struct * p,const struct sched_attr * attr)1917 static void __setscheduler_uclamp(struct task_struct *p,
1918 				  const struct sched_attr *attr)
1919 {
1920 	enum uclamp_id clamp_id;
1921 
1922 	for_each_clamp_id(clamp_id) {
1923 		struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1924 		unsigned int value;
1925 
1926 		if (!uclamp_reset(attr, clamp_id, uc_se))
1927 			continue;
1928 
1929 		/*
1930 		 * RT by default have a 100% boost value that could be modified
1931 		 * at runtime.
1932 		 */
1933 		if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1934 			value = sysctl_sched_uclamp_util_min_rt_default;
1935 		else
1936 			value = uclamp_none(clamp_id);
1937 
1938 		uclamp_se_set(uc_se, value, false);
1939 
1940 	}
1941 
1942 	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1943 		return;
1944 
1945 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1946 	    attr->sched_util_min != -1) {
1947 		uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1948 			      attr->sched_util_min, true);
1949 	}
1950 
1951 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1952 	    attr->sched_util_max != -1) {
1953 		uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1954 			      attr->sched_util_max, true);
1955 	}
1956 }
1957 
uclamp_fork(struct task_struct * p)1958 static void uclamp_fork(struct task_struct *p)
1959 {
1960 	enum uclamp_id clamp_id;
1961 
1962 	/*
1963 	 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1964 	 * as the task is still at its early fork stages.
1965 	 */
1966 	for_each_clamp_id(clamp_id)
1967 		p->uclamp[clamp_id].active = false;
1968 
1969 	if (likely(!p->sched_reset_on_fork))
1970 		return;
1971 
1972 	for_each_clamp_id(clamp_id) {
1973 		uclamp_se_set(&p->uclamp_req[clamp_id],
1974 			      uclamp_none(clamp_id), false);
1975 	}
1976 }
1977 
uclamp_post_fork(struct task_struct * p)1978 static void uclamp_post_fork(struct task_struct *p)
1979 {
1980 	uclamp_update_util_min_rt_default(p);
1981 }
1982 
init_uclamp_rq(struct rq * rq)1983 static void __init init_uclamp_rq(struct rq *rq)
1984 {
1985 	enum uclamp_id clamp_id;
1986 	struct uclamp_rq *uc_rq = rq->uclamp;
1987 
1988 	for_each_clamp_id(clamp_id) {
1989 		uc_rq[clamp_id] = (struct uclamp_rq) {
1990 			.value = uclamp_none(clamp_id)
1991 		};
1992 	}
1993 
1994 	rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1995 }
1996 
init_uclamp(void)1997 static void __init init_uclamp(void)
1998 {
1999 	struct uclamp_se uc_max = {};
2000 	enum uclamp_id clamp_id;
2001 	int cpu;
2002 
2003 	for_each_possible_cpu(cpu)
2004 		init_uclamp_rq(cpu_rq(cpu));
2005 
2006 	for_each_clamp_id(clamp_id) {
2007 		uclamp_se_set(&init_task.uclamp_req[clamp_id],
2008 			      uclamp_none(clamp_id), false);
2009 	}
2010 
2011 	/* System defaults allow max clamp values for both indexes */
2012 	uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2013 	for_each_clamp_id(clamp_id) {
2014 		uclamp_default[clamp_id] = uc_max;
2015 #ifdef CONFIG_UCLAMP_TASK_GROUP
2016 		root_task_group.uclamp_req[clamp_id] = uc_max;
2017 		root_task_group.uclamp[clamp_id] = uc_max;
2018 #endif
2019 	}
2020 }
2021 
2022 #else /* CONFIG_UCLAMP_TASK */
uclamp_rq_inc(struct rq * rq,struct task_struct * p)2023 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
uclamp_rq_dec(struct rq * rq,struct task_struct * p)2024 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
uclamp_validate(struct task_struct * p,const struct sched_attr * attr)2025 static inline int uclamp_validate(struct task_struct *p,
2026 				  const struct sched_attr *attr)
2027 {
2028 	return -EOPNOTSUPP;
2029 }
__setscheduler_uclamp(struct task_struct * p,const struct sched_attr * attr)2030 static void __setscheduler_uclamp(struct task_struct *p,
2031 				  const struct sched_attr *attr) { }
uclamp_fork(struct task_struct * p)2032 static inline void uclamp_fork(struct task_struct *p) { }
uclamp_post_fork(struct task_struct * p)2033 static inline void uclamp_post_fork(struct task_struct *p) { }
init_uclamp(void)2034 static inline void init_uclamp(void) { }
2035 #endif /* CONFIG_UCLAMP_TASK */
2036 
sched_task_on_rq(struct task_struct * p)2037 bool sched_task_on_rq(struct task_struct *p)
2038 {
2039 	return task_on_rq_queued(p);
2040 }
2041 
get_wchan(struct task_struct * p)2042 unsigned long get_wchan(struct task_struct *p)
2043 {
2044 	unsigned long ip = 0;
2045 	unsigned int state;
2046 
2047 	if (!p || p == current)
2048 		return 0;
2049 
2050 	/* Only get wchan if task is blocked and we can keep it that way. */
2051 	raw_spin_lock_irq(&p->pi_lock);
2052 	state = READ_ONCE(p->__state);
2053 	smp_rmb(); /* see try_to_wake_up() */
2054 	if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2055 		ip = __get_wchan(p);
2056 	raw_spin_unlock_irq(&p->pi_lock);
2057 
2058 	return ip;
2059 }
2060 
enqueue_task(struct rq * rq,struct task_struct * p,int flags)2061 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2062 {
2063 	if (!(flags & ENQUEUE_NOCLOCK))
2064 		update_rq_clock(rq);
2065 
2066 	if (!(flags & ENQUEUE_RESTORE)) {
2067 		sched_info_enqueue(rq, p);
2068 		psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2069 	}
2070 
2071 	uclamp_rq_inc(rq, p);
2072 	p->sched_class->enqueue_task(rq, p, flags);
2073 
2074 	if (sched_core_enabled(rq))
2075 		sched_core_enqueue(rq, p);
2076 }
2077 
dequeue_task(struct rq * rq,struct task_struct * p,int flags)2078 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2079 {
2080 	if (sched_core_enabled(rq))
2081 		sched_core_dequeue(rq, p, flags);
2082 
2083 	if (!(flags & DEQUEUE_NOCLOCK))
2084 		update_rq_clock(rq);
2085 
2086 	if (!(flags & DEQUEUE_SAVE)) {
2087 		sched_info_dequeue(rq, p);
2088 		psi_dequeue(p, flags & DEQUEUE_SLEEP);
2089 	}
2090 
2091 	uclamp_rq_dec(rq, p);
2092 	p->sched_class->dequeue_task(rq, p, flags);
2093 }
2094 
activate_task(struct rq * rq,struct task_struct * p,int flags)2095 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2096 {
2097 	enqueue_task(rq, p, flags);
2098 
2099 	p->on_rq = TASK_ON_RQ_QUEUED;
2100 }
2101 
deactivate_task(struct rq * rq,struct task_struct * p,int flags)2102 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2103 {
2104 	p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2105 
2106 	dequeue_task(rq, p, flags);
2107 }
2108 
__normal_prio(int policy,int rt_prio,int nice)2109 static inline int __normal_prio(int policy, int rt_prio, int nice)
2110 {
2111 	int prio;
2112 
2113 	if (dl_policy(policy))
2114 		prio = MAX_DL_PRIO - 1;
2115 	else if (rt_policy(policy))
2116 		prio = MAX_RT_PRIO - 1 - rt_prio;
2117 	else
2118 		prio = NICE_TO_PRIO(nice);
2119 
2120 	return prio;
2121 }
2122 
2123 /*
2124  * Calculate the expected normal priority: i.e. priority
2125  * without taking RT-inheritance into account. Might be
2126  * boosted by interactivity modifiers. Changes upon fork,
2127  * setprio syscalls, and whenever the interactivity
2128  * estimator recalculates.
2129  */
normal_prio(struct task_struct * p)2130 static inline int normal_prio(struct task_struct *p)
2131 {
2132 	return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2133 }
2134 
2135 /*
2136  * Calculate the current priority, i.e. the priority
2137  * taken into account by the scheduler. This value might
2138  * be boosted by RT tasks, or might be boosted by
2139  * interactivity modifiers. Will be RT if the task got
2140  * RT-boosted. If not then it returns p->normal_prio.
2141  */
effective_prio(struct task_struct * p)2142 static int effective_prio(struct task_struct *p)
2143 {
2144 	p->normal_prio = normal_prio(p);
2145 	/*
2146 	 * If we are RT tasks or we were boosted to RT priority,
2147 	 * keep the priority unchanged. Otherwise, update priority
2148 	 * to the normal priority:
2149 	 */
2150 	if (!rt_prio(p->prio))
2151 		return p->normal_prio;
2152 	return p->prio;
2153 }
2154 
2155 /**
2156  * task_curr - is this task currently executing on a CPU?
2157  * @p: the task in question.
2158  *
2159  * Return: 1 if the task is currently executing. 0 otherwise.
2160  */
task_curr(const struct task_struct * p)2161 inline int task_curr(const struct task_struct *p)
2162 {
2163 	return cpu_curr(task_cpu(p)) == p;
2164 }
2165 
2166 /*
2167  * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2168  * use the balance_callback list if you want balancing.
2169  *
2170  * this means any call to check_class_changed() must be followed by a call to
2171  * balance_callback().
2172  */
check_class_changed(struct rq * rq,struct task_struct * p,const struct sched_class * prev_class,int oldprio)2173 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2174 				       const struct sched_class *prev_class,
2175 				       int oldprio)
2176 {
2177 	if (prev_class != p->sched_class) {
2178 		if (prev_class->switched_from)
2179 			prev_class->switched_from(rq, p);
2180 
2181 		p->sched_class->switched_to(rq, p);
2182 	} else if (oldprio != p->prio || dl_task(p))
2183 		p->sched_class->prio_changed(rq, p, oldprio);
2184 }
2185 
check_preempt_curr(struct rq * rq,struct task_struct * p,int flags)2186 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2187 {
2188 	if (p->sched_class == rq->curr->sched_class)
2189 		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2190 	else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2191 		resched_curr(rq);
2192 
2193 	/*
2194 	 * A queue event has occurred, and we're going to schedule.  In
2195 	 * this case, we can save a useless back to back clock update.
2196 	 */
2197 	if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2198 		rq_clock_skip_update(rq);
2199 }
2200 
2201 #ifdef CONFIG_SMP
2202 
2203 static void
2204 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2205 
2206 static int __set_cpus_allowed_ptr(struct task_struct *p,
2207 				  const struct cpumask *new_mask,
2208 				  u32 flags);
2209 
migrate_disable_switch(struct rq * rq,struct task_struct * p)2210 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2211 {
2212 	if (likely(!p->migration_disabled))
2213 		return;
2214 
2215 	if (p->cpus_ptr != &p->cpus_mask)
2216 		return;
2217 
2218 	/*
2219 	 * Violates locking rules! see comment in __do_set_cpus_allowed().
2220 	 */
2221 	__do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2222 }
2223 
migrate_disable(void)2224 void migrate_disable(void)
2225 {
2226 	struct task_struct *p = current;
2227 
2228 	if (p->migration_disabled) {
2229 		p->migration_disabled++;
2230 		return;
2231 	}
2232 
2233 	preempt_disable();
2234 	this_rq()->nr_pinned++;
2235 	p->migration_disabled = 1;
2236 	preempt_enable();
2237 }
2238 EXPORT_SYMBOL_GPL(migrate_disable);
2239 
migrate_enable(void)2240 void migrate_enable(void)
2241 {
2242 	struct task_struct *p = current;
2243 
2244 	if (p->migration_disabled > 1) {
2245 		p->migration_disabled--;
2246 		return;
2247 	}
2248 
2249 	if (WARN_ON_ONCE(!p->migration_disabled))
2250 		return;
2251 
2252 	/*
2253 	 * Ensure stop_task runs either before or after this, and that
2254 	 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2255 	 */
2256 	preempt_disable();
2257 	if (p->cpus_ptr != &p->cpus_mask)
2258 		__set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2259 	/*
2260 	 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2261 	 * regular cpus_mask, otherwise things that race (eg.
2262 	 * select_fallback_rq) get confused.
2263 	 */
2264 	barrier();
2265 	p->migration_disabled = 0;
2266 	this_rq()->nr_pinned--;
2267 	preempt_enable();
2268 }
2269 EXPORT_SYMBOL_GPL(migrate_enable);
2270 
rq_has_pinned_tasks(struct rq * rq)2271 static inline bool rq_has_pinned_tasks(struct rq *rq)
2272 {
2273 	return rq->nr_pinned;
2274 }
2275 
2276 /*
2277  * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2278  * __set_cpus_allowed_ptr() and select_fallback_rq().
2279  */
is_cpu_allowed(struct task_struct * p,int cpu)2280 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2281 {
2282 	/* When not in the task's cpumask, no point in looking further. */
2283 	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2284 		return false;
2285 
2286 	/* migrate_disabled() must be allowed to finish. */
2287 	if (is_migration_disabled(p))
2288 		return cpu_online(cpu);
2289 
2290 	/* Non kernel threads are not allowed during either online or offline. */
2291 	if (!(p->flags & PF_KTHREAD))
2292 		return cpu_active(cpu) && task_cpu_possible(cpu, p);
2293 
2294 	/* KTHREAD_IS_PER_CPU is always allowed. */
2295 	if (kthread_is_per_cpu(p))
2296 		return cpu_online(cpu);
2297 
2298 	/* Regular kernel threads don't get to stay during offline. */
2299 	if (cpu_dying(cpu))
2300 		return false;
2301 
2302 	/* But are allowed during online. */
2303 	return cpu_online(cpu);
2304 }
2305 
2306 /*
2307  * This is how migration works:
2308  *
2309  * 1) we invoke migration_cpu_stop() on the target CPU using
2310  *    stop_one_cpu().
2311  * 2) stopper starts to run (implicitly forcing the migrated thread
2312  *    off the CPU)
2313  * 3) it checks whether the migrated task is still in the wrong runqueue.
2314  * 4) if it's in the wrong runqueue then the migration thread removes
2315  *    it and puts it into the right queue.
2316  * 5) stopper completes and stop_one_cpu() returns and the migration
2317  *    is done.
2318  */
2319 
2320 /*
2321  * move_queued_task - move a queued task to new rq.
2322  *
2323  * Returns (locked) new rq. Old rq's lock is released.
2324  */
move_queued_task(struct rq * rq,struct rq_flags * rf,struct task_struct * p,int new_cpu)2325 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2326 				   struct task_struct *p, int new_cpu)
2327 {
2328 	lockdep_assert_rq_held(rq);
2329 
2330 	deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2331 	set_task_cpu(p, new_cpu);
2332 	rq_unlock(rq, rf);
2333 
2334 	rq = cpu_rq(new_cpu);
2335 
2336 	rq_lock(rq, rf);
2337 	BUG_ON(task_cpu(p) != new_cpu);
2338 	activate_task(rq, p, 0);
2339 	check_preempt_curr(rq, p, 0);
2340 
2341 	return rq;
2342 }
2343 
2344 struct migration_arg {
2345 	struct task_struct		*task;
2346 	int				dest_cpu;
2347 	struct set_affinity_pending	*pending;
2348 };
2349 
2350 /*
2351  * @refs: number of wait_for_completion()
2352  * @stop_pending: is @stop_work in use
2353  */
2354 struct set_affinity_pending {
2355 	refcount_t		refs;
2356 	unsigned int		stop_pending;
2357 	struct completion	done;
2358 	struct cpu_stop_work	stop_work;
2359 	struct migration_arg	arg;
2360 };
2361 
2362 /*
2363  * Move (not current) task off this CPU, onto the destination CPU. We're doing
2364  * this because either it can't run here any more (set_cpus_allowed()
2365  * away from this CPU, or CPU going down), or because we're
2366  * attempting to rebalance this task on exec (sched_exec).
2367  *
2368  * So we race with normal scheduler movements, but that's OK, as long
2369  * as the task is no longer on this CPU.
2370  */
__migrate_task(struct rq * rq,struct rq_flags * rf,struct task_struct * p,int dest_cpu)2371 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2372 				 struct task_struct *p, int dest_cpu)
2373 {
2374 	/* Affinity changed (again). */
2375 	if (!is_cpu_allowed(p, dest_cpu))
2376 		return rq;
2377 
2378 	update_rq_clock(rq);
2379 	rq = move_queued_task(rq, rf, p, dest_cpu);
2380 
2381 	return rq;
2382 }
2383 
2384 /*
2385  * migration_cpu_stop - this will be executed by a highprio stopper thread
2386  * and performs thread migration by bumping thread off CPU then
2387  * 'pushing' onto another runqueue.
2388  */
migration_cpu_stop(void * data)2389 static int migration_cpu_stop(void *data)
2390 {
2391 	struct migration_arg *arg = data;
2392 	struct set_affinity_pending *pending = arg->pending;
2393 	struct task_struct *p = arg->task;
2394 	struct rq *rq = this_rq();
2395 	bool complete = false;
2396 	struct rq_flags rf;
2397 
2398 	/*
2399 	 * The original target CPU might have gone down and we might
2400 	 * be on another CPU but it doesn't matter.
2401 	 */
2402 	local_irq_save(rf.flags);
2403 	/*
2404 	 * We need to explicitly wake pending tasks before running
2405 	 * __migrate_task() such that we will not miss enforcing cpus_ptr
2406 	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2407 	 */
2408 	flush_smp_call_function_queue();
2409 
2410 	raw_spin_lock(&p->pi_lock);
2411 	rq_lock(rq, &rf);
2412 
2413 	/*
2414 	 * If we were passed a pending, then ->stop_pending was set, thus
2415 	 * p->migration_pending must have remained stable.
2416 	 */
2417 	WARN_ON_ONCE(pending && pending != p->migration_pending);
2418 
2419 	/*
2420 	 * If task_rq(p) != rq, it cannot be migrated here, because we're
2421 	 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2422 	 * we're holding p->pi_lock.
2423 	 */
2424 	if (task_rq(p) == rq) {
2425 		if (is_migration_disabled(p))
2426 			goto out;
2427 
2428 		if (pending) {
2429 			p->migration_pending = NULL;
2430 			complete = true;
2431 
2432 			if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2433 				goto out;
2434 		}
2435 
2436 		if (task_on_rq_queued(p))
2437 			rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2438 		else
2439 			p->wake_cpu = arg->dest_cpu;
2440 
2441 		/*
2442 		 * XXX __migrate_task() can fail, at which point we might end
2443 		 * up running on a dodgy CPU, AFAICT this can only happen
2444 		 * during CPU hotplug, at which point we'll get pushed out
2445 		 * anyway, so it's probably not a big deal.
2446 		 */
2447 
2448 	} else if (pending) {
2449 		/*
2450 		 * This happens when we get migrated between migrate_enable()'s
2451 		 * preempt_enable() and scheduling the stopper task. At that
2452 		 * point we're a regular task again and not current anymore.
2453 		 *
2454 		 * A !PREEMPT kernel has a giant hole here, which makes it far
2455 		 * more likely.
2456 		 */
2457 
2458 		/*
2459 		 * The task moved before the stopper got to run. We're holding
2460 		 * ->pi_lock, so the allowed mask is stable - if it got
2461 		 * somewhere allowed, we're done.
2462 		 */
2463 		if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2464 			p->migration_pending = NULL;
2465 			complete = true;
2466 			goto out;
2467 		}
2468 
2469 		/*
2470 		 * When migrate_enable() hits a rq mis-match we can't reliably
2471 		 * determine is_migration_disabled() and so have to chase after
2472 		 * it.
2473 		 */
2474 		WARN_ON_ONCE(!pending->stop_pending);
2475 		task_rq_unlock(rq, p, &rf);
2476 		stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2477 				    &pending->arg, &pending->stop_work);
2478 		return 0;
2479 	}
2480 out:
2481 	if (pending)
2482 		pending->stop_pending = false;
2483 	task_rq_unlock(rq, p, &rf);
2484 
2485 	if (complete)
2486 		complete_all(&pending->done);
2487 
2488 	return 0;
2489 }
2490 
push_cpu_stop(void * arg)2491 int push_cpu_stop(void *arg)
2492 {
2493 	struct rq *lowest_rq = NULL, *rq = this_rq();
2494 	struct task_struct *p = arg;
2495 
2496 	raw_spin_lock_irq(&p->pi_lock);
2497 	raw_spin_rq_lock(rq);
2498 
2499 	if (task_rq(p) != rq)
2500 		goto out_unlock;
2501 
2502 	if (is_migration_disabled(p)) {
2503 		p->migration_flags |= MDF_PUSH;
2504 		goto out_unlock;
2505 	}
2506 
2507 	p->migration_flags &= ~MDF_PUSH;
2508 
2509 	if (p->sched_class->find_lock_rq)
2510 		lowest_rq = p->sched_class->find_lock_rq(p, rq);
2511 
2512 	if (!lowest_rq)
2513 		goto out_unlock;
2514 
2515 	// XXX validate p is still the highest prio task
2516 	if (task_rq(p) == rq) {
2517 		deactivate_task(rq, p, 0);
2518 		set_task_cpu(p, lowest_rq->cpu);
2519 		activate_task(lowest_rq, p, 0);
2520 		resched_curr(lowest_rq);
2521 	}
2522 
2523 	double_unlock_balance(rq, lowest_rq);
2524 
2525 out_unlock:
2526 	rq->push_busy = false;
2527 	raw_spin_rq_unlock(rq);
2528 	raw_spin_unlock_irq(&p->pi_lock);
2529 
2530 	put_task_struct(p);
2531 	return 0;
2532 }
2533 
2534 /*
2535  * sched_class::set_cpus_allowed must do the below, but is not required to
2536  * actually call this function.
2537  */
set_cpus_allowed_common(struct task_struct * p,const struct cpumask * new_mask,u32 flags)2538 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2539 {
2540 	if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2541 		p->cpus_ptr = new_mask;
2542 		return;
2543 	}
2544 
2545 	cpumask_copy(&p->cpus_mask, new_mask);
2546 	p->nr_cpus_allowed = cpumask_weight(new_mask);
2547 }
2548 
2549 static void
__do_set_cpus_allowed(struct task_struct * p,const struct cpumask * new_mask,u32 flags)2550 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2551 {
2552 	struct rq *rq = task_rq(p);
2553 	bool queued, running;
2554 
2555 	/*
2556 	 * This here violates the locking rules for affinity, since we're only
2557 	 * supposed to change these variables while holding both rq->lock and
2558 	 * p->pi_lock.
2559 	 *
2560 	 * HOWEVER, it magically works, because ttwu() is the only code that
2561 	 * accesses these variables under p->pi_lock and only does so after
2562 	 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2563 	 * before finish_task().
2564 	 *
2565 	 * XXX do further audits, this smells like something putrid.
2566 	 */
2567 	if (flags & SCA_MIGRATE_DISABLE)
2568 		SCHED_WARN_ON(!p->on_cpu);
2569 	else
2570 		lockdep_assert_held(&p->pi_lock);
2571 
2572 	queued = task_on_rq_queued(p);
2573 	running = task_current(rq, p);
2574 
2575 	if (queued) {
2576 		/*
2577 		 * Because __kthread_bind() calls this on blocked tasks without
2578 		 * holding rq->lock.
2579 		 */
2580 		lockdep_assert_rq_held(rq);
2581 		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2582 	}
2583 	if (running)
2584 		put_prev_task(rq, p);
2585 
2586 	p->sched_class->set_cpus_allowed(p, new_mask, flags);
2587 
2588 	if (queued)
2589 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2590 	if (running)
2591 		set_next_task(rq, p);
2592 }
2593 
do_set_cpus_allowed(struct task_struct * p,const struct cpumask * new_mask)2594 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2595 {
2596 	__do_set_cpus_allowed(p, new_mask, 0);
2597 }
2598 
dup_user_cpus_ptr(struct task_struct * dst,struct task_struct * src,int node)2599 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2600 		      int node)
2601 {
2602 	if (!src->user_cpus_ptr)
2603 		return 0;
2604 
2605 	dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2606 	if (!dst->user_cpus_ptr)
2607 		return -ENOMEM;
2608 
2609 	cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2610 	return 0;
2611 }
2612 
clear_user_cpus_ptr(struct task_struct * p)2613 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2614 {
2615 	struct cpumask *user_mask = NULL;
2616 
2617 	swap(p->user_cpus_ptr, user_mask);
2618 
2619 	return user_mask;
2620 }
2621 
release_user_cpus_ptr(struct task_struct * p)2622 void release_user_cpus_ptr(struct task_struct *p)
2623 {
2624 	kfree(clear_user_cpus_ptr(p));
2625 }
2626 
2627 /*
2628  * This function is wildly self concurrent; here be dragons.
2629  *
2630  *
2631  * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2632  * designated task is enqueued on an allowed CPU. If that task is currently
2633  * running, we have to kick it out using the CPU stopper.
2634  *
2635  * Migrate-Disable comes along and tramples all over our nice sandcastle.
2636  * Consider:
2637  *
2638  *     Initial conditions: P0->cpus_mask = [0, 1]
2639  *
2640  *     P0@CPU0                  P1
2641  *
2642  *     migrate_disable();
2643  *     <preempted>
2644  *                              set_cpus_allowed_ptr(P0, [1]);
2645  *
2646  * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2647  * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2648  * This means we need the following scheme:
2649  *
2650  *     P0@CPU0                  P1
2651  *
2652  *     migrate_disable();
2653  *     <preempted>
2654  *                              set_cpus_allowed_ptr(P0, [1]);
2655  *                                <blocks>
2656  *     <resumes>
2657  *     migrate_enable();
2658  *       __set_cpus_allowed_ptr();
2659  *       <wakes local stopper>
2660  *                         `--> <woken on migration completion>
2661  *
2662  * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2663  * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2664  * task p are serialized by p->pi_lock, which we can leverage: the one that
2665  * should come into effect at the end of the Migrate-Disable region is the last
2666  * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2667  * but we still need to properly signal those waiting tasks at the appropriate
2668  * moment.
2669  *
2670  * This is implemented using struct set_affinity_pending. The first
2671  * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2672  * setup an instance of that struct and install it on the targeted task_struct.
2673  * Any and all further callers will reuse that instance. Those then wait for
2674  * a completion signaled at the tail of the CPU stopper callback (1), triggered
2675  * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2676  *
2677  *
2678  * (1) In the cases covered above. There is one more where the completion is
2679  * signaled within affine_move_task() itself: when a subsequent affinity request
2680  * occurs after the stopper bailed out due to the targeted task still being
2681  * Migrate-Disable. Consider:
2682  *
2683  *     Initial conditions: P0->cpus_mask = [0, 1]
2684  *
2685  *     CPU0		  P1				P2
2686  *     <P0>
2687  *       migrate_disable();
2688  *       <preempted>
2689  *                        set_cpus_allowed_ptr(P0, [1]);
2690  *                          <blocks>
2691  *     <migration/0>
2692  *       migration_cpu_stop()
2693  *         is_migration_disabled()
2694  *           <bails>
2695  *                                                       set_cpus_allowed_ptr(P0, [0, 1]);
2696  *                                                         <signal completion>
2697  *                          <awakes>
2698  *
2699  * Note that the above is safe vs a concurrent migrate_enable(), as any
2700  * pending affinity completion is preceded by an uninstallation of
2701  * p->migration_pending done with p->pi_lock held.
2702  */
affine_move_task(struct rq * rq,struct task_struct * p,struct rq_flags * rf,int dest_cpu,unsigned int flags)2703 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2704 			    int dest_cpu, unsigned int flags)
2705 {
2706 	struct set_affinity_pending my_pending = { }, *pending = NULL;
2707 	bool stop_pending, complete = false;
2708 
2709 	/* Can the task run on the task's current CPU? If so, we're done */
2710 	if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2711 		struct task_struct *push_task = NULL;
2712 
2713 		if ((flags & SCA_MIGRATE_ENABLE) &&
2714 		    (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2715 			rq->push_busy = true;
2716 			push_task = get_task_struct(p);
2717 		}
2718 
2719 		/*
2720 		 * If there are pending waiters, but no pending stop_work,
2721 		 * then complete now.
2722 		 */
2723 		pending = p->migration_pending;
2724 		if (pending && !pending->stop_pending) {
2725 			p->migration_pending = NULL;
2726 			complete = true;
2727 		}
2728 
2729 		task_rq_unlock(rq, p, rf);
2730 
2731 		if (push_task) {
2732 			stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2733 					    p, &rq->push_work);
2734 		}
2735 
2736 		if (complete)
2737 			complete_all(&pending->done);
2738 
2739 		return 0;
2740 	}
2741 
2742 	if (!(flags & SCA_MIGRATE_ENABLE)) {
2743 		/* serialized by p->pi_lock */
2744 		if (!p->migration_pending) {
2745 			/* Install the request */
2746 			refcount_set(&my_pending.refs, 1);
2747 			init_completion(&my_pending.done);
2748 			my_pending.arg = (struct migration_arg) {
2749 				.task = p,
2750 				.dest_cpu = dest_cpu,
2751 				.pending = &my_pending,
2752 			};
2753 
2754 			p->migration_pending = &my_pending;
2755 		} else {
2756 			pending = p->migration_pending;
2757 			refcount_inc(&pending->refs);
2758 			/*
2759 			 * Affinity has changed, but we've already installed a
2760 			 * pending. migration_cpu_stop() *must* see this, else
2761 			 * we risk a completion of the pending despite having a
2762 			 * task on a disallowed CPU.
2763 			 *
2764 			 * Serialized by p->pi_lock, so this is safe.
2765 			 */
2766 			pending->arg.dest_cpu = dest_cpu;
2767 		}
2768 	}
2769 	pending = p->migration_pending;
2770 	/*
2771 	 * - !MIGRATE_ENABLE:
2772 	 *   we'll have installed a pending if there wasn't one already.
2773 	 *
2774 	 * - MIGRATE_ENABLE:
2775 	 *   we're here because the current CPU isn't matching anymore,
2776 	 *   the only way that can happen is because of a concurrent
2777 	 *   set_cpus_allowed_ptr() call, which should then still be
2778 	 *   pending completion.
2779 	 *
2780 	 * Either way, we really should have a @pending here.
2781 	 */
2782 	if (WARN_ON_ONCE(!pending)) {
2783 		task_rq_unlock(rq, p, rf);
2784 		return -EINVAL;
2785 	}
2786 
2787 	if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2788 		/*
2789 		 * MIGRATE_ENABLE gets here because 'p == current', but for
2790 		 * anything else we cannot do is_migration_disabled(), punt
2791 		 * and have the stopper function handle it all race-free.
2792 		 */
2793 		stop_pending = pending->stop_pending;
2794 		if (!stop_pending)
2795 			pending->stop_pending = true;
2796 
2797 		if (flags & SCA_MIGRATE_ENABLE)
2798 			p->migration_flags &= ~MDF_PUSH;
2799 
2800 		task_rq_unlock(rq, p, rf);
2801 
2802 		if (!stop_pending) {
2803 			stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2804 					    &pending->arg, &pending->stop_work);
2805 		}
2806 
2807 		if (flags & SCA_MIGRATE_ENABLE)
2808 			return 0;
2809 	} else {
2810 
2811 		if (!is_migration_disabled(p)) {
2812 			if (task_on_rq_queued(p))
2813 				rq = move_queued_task(rq, rf, p, dest_cpu);
2814 
2815 			if (!pending->stop_pending) {
2816 				p->migration_pending = NULL;
2817 				complete = true;
2818 			}
2819 		}
2820 		task_rq_unlock(rq, p, rf);
2821 
2822 		if (complete)
2823 			complete_all(&pending->done);
2824 	}
2825 
2826 	wait_for_completion(&pending->done);
2827 
2828 	if (refcount_dec_and_test(&pending->refs))
2829 		wake_up_var(&pending->refs); /* No UaF, just an address */
2830 
2831 	/*
2832 	 * Block the original owner of &pending until all subsequent callers
2833 	 * have seen the completion and decremented the refcount
2834 	 */
2835 	wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2836 
2837 	/* ARGH */
2838 	WARN_ON_ONCE(my_pending.stop_pending);
2839 
2840 	return 0;
2841 }
2842 
2843 /*
2844  * Called with both p->pi_lock and rq->lock held; drops both before returning.
2845  */
__set_cpus_allowed_ptr_locked(struct task_struct * p,const struct cpumask * new_mask,u32 flags,struct rq * rq,struct rq_flags * rf)2846 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2847 					 const struct cpumask *new_mask,
2848 					 u32 flags,
2849 					 struct rq *rq,
2850 					 struct rq_flags *rf)
2851 	__releases(rq->lock)
2852 	__releases(p->pi_lock)
2853 {
2854 	const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2855 	const struct cpumask *cpu_valid_mask = cpu_active_mask;
2856 	bool kthread = p->flags & PF_KTHREAD;
2857 	struct cpumask *user_mask = NULL;
2858 	unsigned int dest_cpu;
2859 	int ret = 0;
2860 
2861 	update_rq_clock(rq);
2862 
2863 	if (kthread || is_migration_disabled(p)) {
2864 		/*
2865 		 * Kernel threads are allowed on online && !active CPUs,
2866 		 * however, during cpu-hot-unplug, even these might get pushed
2867 		 * away if not KTHREAD_IS_PER_CPU.
2868 		 *
2869 		 * Specifically, migration_disabled() tasks must not fail the
2870 		 * cpumask_any_and_distribute() pick below, esp. so on
2871 		 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2872 		 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2873 		 */
2874 		cpu_valid_mask = cpu_online_mask;
2875 	}
2876 
2877 	if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2878 		ret = -EINVAL;
2879 		goto out;
2880 	}
2881 
2882 	/*
2883 	 * Must re-check here, to close a race against __kthread_bind(),
2884 	 * sched_setaffinity() is not guaranteed to observe the flag.
2885 	 */
2886 	if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2887 		ret = -EINVAL;
2888 		goto out;
2889 	}
2890 
2891 	if (!(flags & SCA_MIGRATE_ENABLE)) {
2892 		if (cpumask_equal(&p->cpus_mask, new_mask))
2893 			goto out;
2894 
2895 		if (WARN_ON_ONCE(p == current &&
2896 				 is_migration_disabled(p) &&
2897 				 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2898 			ret = -EBUSY;
2899 			goto out;
2900 		}
2901 	}
2902 
2903 	/*
2904 	 * Picking a ~random cpu helps in cases where we are changing affinity
2905 	 * for groups of tasks (ie. cpuset), so that load balancing is not
2906 	 * immediately required to distribute the tasks within their new mask.
2907 	 */
2908 	dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2909 	if (dest_cpu >= nr_cpu_ids) {
2910 		ret = -EINVAL;
2911 		goto out;
2912 	}
2913 
2914 	__do_set_cpus_allowed(p, new_mask, flags);
2915 
2916 	if (flags & SCA_USER)
2917 		user_mask = clear_user_cpus_ptr(p);
2918 
2919 	ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2920 
2921 	kfree(user_mask);
2922 
2923 	return ret;
2924 
2925 out:
2926 	task_rq_unlock(rq, p, rf);
2927 
2928 	return ret;
2929 }
2930 
2931 /*
2932  * Change a given task's CPU affinity. Migrate the thread to a
2933  * proper CPU and schedule it away if the CPU it's executing on
2934  * is removed from the allowed bitmask.
2935  *
2936  * NOTE: the caller must have a valid reference to the task, the
2937  * task must not exit() & deallocate itself prematurely. The
2938  * call is not atomic; no spinlocks may be held.
2939  */
__set_cpus_allowed_ptr(struct task_struct * p,const struct cpumask * new_mask,u32 flags)2940 static int __set_cpus_allowed_ptr(struct task_struct *p,
2941 				  const struct cpumask *new_mask, u32 flags)
2942 {
2943 	struct rq_flags rf;
2944 	struct rq *rq;
2945 
2946 	rq = task_rq_lock(p, &rf);
2947 	return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2948 }
2949 
set_cpus_allowed_ptr(struct task_struct * p,const struct cpumask * new_mask)2950 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2951 {
2952 	return __set_cpus_allowed_ptr(p, new_mask, 0);
2953 }
2954 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2955 
2956 /*
2957  * Change a given task's CPU affinity to the intersection of its current
2958  * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2959  * and pointing @p->user_cpus_ptr to a copy of the old mask.
2960  * If the resulting mask is empty, leave the affinity unchanged and return
2961  * -EINVAL.
2962  */
restrict_cpus_allowed_ptr(struct task_struct * p,struct cpumask * new_mask,const struct cpumask * subset_mask)2963 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2964 				     struct cpumask *new_mask,
2965 				     const struct cpumask *subset_mask)
2966 {
2967 	struct cpumask *user_mask = NULL;
2968 	struct rq_flags rf;
2969 	struct rq *rq;
2970 	int err;
2971 
2972 	if (!p->user_cpus_ptr) {
2973 		user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2974 		if (!user_mask)
2975 			return -ENOMEM;
2976 	}
2977 
2978 	rq = task_rq_lock(p, &rf);
2979 
2980 	/*
2981 	 * Forcefully restricting the affinity of a deadline task is
2982 	 * likely to cause problems, so fail and noisily override the
2983 	 * mask entirely.
2984 	 */
2985 	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2986 		err = -EPERM;
2987 		goto err_unlock;
2988 	}
2989 
2990 	if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2991 		err = -EINVAL;
2992 		goto err_unlock;
2993 	}
2994 
2995 	/*
2996 	 * We're about to butcher the task affinity, so keep track of what
2997 	 * the user asked for in case we're able to restore it later on.
2998 	 */
2999 	if (user_mask) {
3000 		cpumask_copy(user_mask, p->cpus_ptr);
3001 		p->user_cpus_ptr = user_mask;
3002 	}
3003 
3004 	return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
3005 
3006 err_unlock:
3007 	task_rq_unlock(rq, p, &rf);
3008 	kfree(user_mask);
3009 	return err;
3010 }
3011 
3012 /*
3013  * Restrict the CPU affinity of task @p so that it is a subset of
3014  * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
3015  * old affinity mask. If the resulting mask is empty, we warn and walk
3016  * up the cpuset hierarchy until we find a suitable mask.
3017  */
force_compatible_cpus_allowed_ptr(struct task_struct * p)3018 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3019 {
3020 	cpumask_var_t new_mask;
3021 	const struct cpumask *override_mask = task_cpu_possible_mask(p);
3022 
3023 	alloc_cpumask_var(&new_mask, GFP_KERNEL);
3024 
3025 	/*
3026 	 * __migrate_task() can fail silently in the face of concurrent
3027 	 * offlining of the chosen destination CPU, so take the hotplug
3028 	 * lock to ensure that the migration succeeds.
3029 	 */
3030 	cpus_read_lock();
3031 	if (!cpumask_available(new_mask))
3032 		goto out_set_mask;
3033 
3034 	if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3035 		goto out_free_mask;
3036 
3037 	/*
3038 	 * We failed to find a valid subset of the affinity mask for the
3039 	 * task, so override it based on its cpuset hierarchy.
3040 	 */
3041 	cpuset_cpus_allowed(p, new_mask);
3042 	override_mask = new_mask;
3043 
3044 out_set_mask:
3045 	if (printk_ratelimit()) {
3046 		printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3047 				task_pid_nr(p), p->comm,
3048 				cpumask_pr_args(override_mask));
3049 	}
3050 
3051 	WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3052 out_free_mask:
3053 	cpus_read_unlock();
3054 	free_cpumask_var(new_mask);
3055 }
3056 
3057 static int
3058 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
3059 
3060 /*
3061  * Restore the affinity of a task @p which was previously restricted by a
3062  * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3063  * @p->user_cpus_ptr.
3064  *
3065  * It is the caller's responsibility to serialise this with any calls to
3066  * force_compatible_cpus_allowed_ptr(@p).
3067  */
relax_compatible_cpus_allowed_ptr(struct task_struct * p)3068 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3069 {
3070 	struct cpumask *user_mask = p->user_cpus_ptr;
3071 	unsigned long flags;
3072 
3073 	/*
3074 	 * Try to restore the old affinity mask. If this fails, then
3075 	 * we free the mask explicitly to avoid it being inherited across
3076 	 * a subsequent fork().
3077 	 */
3078 	if (!user_mask || !__sched_setaffinity(p, user_mask))
3079 		return;
3080 
3081 	raw_spin_lock_irqsave(&p->pi_lock, flags);
3082 	user_mask = clear_user_cpus_ptr(p);
3083 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3084 
3085 	kfree(user_mask);
3086 }
3087 
set_task_cpu(struct task_struct * p,unsigned int new_cpu)3088 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3089 {
3090 #ifdef CONFIG_SCHED_DEBUG
3091 	unsigned int state = READ_ONCE(p->__state);
3092 
3093 	/*
3094 	 * We should never call set_task_cpu() on a blocked task,
3095 	 * ttwu() will sort out the placement.
3096 	 */
3097 	WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3098 
3099 	/*
3100 	 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3101 	 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3102 	 * time relying on p->on_rq.
3103 	 */
3104 	WARN_ON_ONCE(state == TASK_RUNNING &&
3105 		     p->sched_class == &fair_sched_class &&
3106 		     (p->on_rq && !task_on_rq_migrating(p)));
3107 
3108 #ifdef CONFIG_LOCKDEP
3109 	/*
3110 	 * The caller should hold either p->pi_lock or rq->lock, when changing
3111 	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3112 	 *
3113 	 * sched_move_task() holds both and thus holding either pins the cgroup,
3114 	 * see task_group().
3115 	 *
3116 	 * Furthermore, all task_rq users should acquire both locks, see
3117 	 * task_rq_lock().
3118 	 */
3119 	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3120 				      lockdep_is_held(__rq_lockp(task_rq(p)))));
3121 #endif
3122 	/*
3123 	 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3124 	 */
3125 	WARN_ON_ONCE(!cpu_online(new_cpu));
3126 
3127 	WARN_ON_ONCE(is_migration_disabled(p));
3128 #endif
3129 
3130 	trace_sched_migrate_task(p, new_cpu);
3131 
3132 	if (task_cpu(p) != new_cpu) {
3133 		if (p->sched_class->migrate_task_rq)
3134 			p->sched_class->migrate_task_rq(p, new_cpu);
3135 		p->se.nr_migrations++;
3136 		rseq_migrate(p);
3137 		perf_event_task_migrate(p);
3138 	}
3139 
3140 	__set_task_cpu(p, new_cpu);
3141 }
3142 
3143 #ifdef CONFIG_NUMA_BALANCING
__migrate_swap_task(struct task_struct * p,int cpu)3144 static void __migrate_swap_task(struct task_struct *p, int cpu)
3145 {
3146 	if (task_on_rq_queued(p)) {
3147 		struct rq *src_rq, *dst_rq;
3148 		struct rq_flags srf, drf;
3149 
3150 		src_rq = task_rq(p);
3151 		dst_rq = cpu_rq(cpu);
3152 
3153 		rq_pin_lock(src_rq, &srf);
3154 		rq_pin_lock(dst_rq, &drf);
3155 
3156 		deactivate_task(src_rq, p, 0);
3157 		set_task_cpu(p, cpu);
3158 		activate_task(dst_rq, p, 0);
3159 		check_preempt_curr(dst_rq, p, 0);
3160 
3161 		rq_unpin_lock(dst_rq, &drf);
3162 		rq_unpin_lock(src_rq, &srf);
3163 
3164 	} else {
3165 		/*
3166 		 * Task isn't running anymore; make it appear like we migrated
3167 		 * it before it went to sleep. This means on wakeup we make the
3168 		 * previous CPU our target instead of where it really is.
3169 		 */
3170 		p->wake_cpu = cpu;
3171 	}
3172 }
3173 
3174 struct migration_swap_arg {
3175 	struct task_struct *src_task, *dst_task;
3176 	int src_cpu, dst_cpu;
3177 };
3178 
migrate_swap_stop(void * data)3179 static int migrate_swap_stop(void *data)
3180 {
3181 	struct migration_swap_arg *arg = data;
3182 	struct rq *src_rq, *dst_rq;
3183 	int ret = -EAGAIN;
3184 
3185 	if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3186 		return -EAGAIN;
3187 
3188 	src_rq = cpu_rq(arg->src_cpu);
3189 	dst_rq = cpu_rq(arg->dst_cpu);
3190 
3191 	double_raw_lock(&arg->src_task->pi_lock,
3192 			&arg->dst_task->pi_lock);
3193 	double_rq_lock(src_rq, dst_rq);
3194 
3195 	if (task_cpu(arg->dst_task) != arg->dst_cpu)
3196 		goto unlock;
3197 
3198 	if (task_cpu(arg->src_task) != arg->src_cpu)
3199 		goto unlock;
3200 
3201 	if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3202 		goto unlock;
3203 
3204 	if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3205 		goto unlock;
3206 
3207 	__migrate_swap_task(arg->src_task, arg->dst_cpu);
3208 	__migrate_swap_task(arg->dst_task, arg->src_cpu);
3209 
3210 	ret = 0;
3211 
3212 unlock:
3213 	double_rq_unlock(src_rq, dst_rq);
3214 	raw_spin_unlock(&arg->dst_task->pi_lock);
3215 	raw_spin_unlock(&arg->src_task->pi_lock);
3216 
3217 	return ret;
3218 }
3219 
3220 /*
3221  * Cross migrate two tasks
3222  */
migrate_swap(struct task_struct * cur,struct task_struct * p,int target_cpu,int curr_cpu)3223 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3224 		int target_cpu, int curr_cpu)
3225 {
3226 	struct migration_swap_arg arg;
3227 	int ret = -EINVAL;
3228 
3229 	arg = (struct migration_swap_arg){
3230 		.src_task = cur,
3231 		.src_cpu = curr_cpu,
3232 		.dst_task = p,
3233 		.dst_cpu = target_cpu,
3234 	};
3235 
3236 	if (arg.src_cpu == arg.dst_cpu)
3237 		goto out;
3238 
3239 	/*
3240 	 * These three tests are all lockless; this is OK since all of them
3241 	 * will be re-checked with proper locks held further down the line.
3242 	 */
3243 	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3244 		goto out;
3245 
3246 	if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3247 		goto out;
3248 
3249 	if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3250 		goto out;
3251 
3252 	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3253 	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3254 
3255 out:
3256 	return ret;
3257 }
3258 #endif /* CONFIG_NUMA_BALANCING */
3259 
3260 /*
3261  * wait_task_inactive - wait for a thread to unschedule.
3262  *
3263  * If @match_state is nonzero, it's the @p->state value just checked and
3264  * not expected to change.  If it changes, i.e. @p might have woken up,
3265  * then return zero.  When we succeed in waiting for @p to be off its CPU,
3266  * we return a positive number (its total switch count).  If a second call
3267  * a short while later returns the same number, the caller can be sure that
3268  * @p has remained unscheduled the whole time.
3269  *
3270  * The caller must ensure that the task *will* unschedule sometime soon,
3271  * else this function might spin for a *long* time. This function can't
3272  * be called with interrupts off, or it may introduce deadlock with
3273  * smp_call_function() if an IPI is sent by the same process we are
3274  * waiting to become inactive.
3275  */
wait_task_inactive(struct task_struct * p,unsigned int match_state)3276 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3277 {
3278 	int running, queued;
3279 	struct rq_flags rf;
3280 	unsigned long ncsw;
3281 	struct rq *rq;
3282 
3283 	for (;;) {
3284 		/*
3285 		 * We do the initial early heuristics without holding
3286 		 * any task-queue locks at all. We'll only try to get
3287 		 * the runqueue lock when things look like they will
3288 		 * work out!
3289 		 */
3290 		rq = task_rq(p);
3291 
3292 		/*
3293 		 * If the task is actively running on another CPU
3294 		 * still, just relax and busy-wait without holding
3295 		 * any locks.
3296 		 *
3297 		 * NOTE! Since we don't hold any locks, it's not
3298 		 * even sure that "rq" stays as the right runqueue!
3299 		 * But we don't care, since "task_running()" will
3300 		 * return false if the runqueue has changed and p
3301 		 * is actually now running somewhere else!
3302 		 */
3303 		while (task_running(rq, p)) {
3304 			if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3305 				return 0;
3306 			cpu_relax();
3307 		}
3308 
3309 		/*
3310 		 * Ok, time to look more closely! We need the rq
3311 		 * lock now, to be *sure*. If we're wrong, we'll
3312 		 * just go back and repeat.
3313 		 */
3314 		rq = task_rq_lock(p, &rf);
3315 		trace_sched_wait_task(p);
3316 		running = task_running(rq, p);
3317 		queued = task_on_rq_queued(p);
3318 		ncsw = 0;
3319 		if (!match_state || READ_ONCE(p->__state) == match_state)
3320 			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3321 		task_rq_unlock(rq, p, &rf);
3322 
3323 		/*
3324 		 * If it changed from the expected state, bail out now.
3325 		 */
3326 		if (unlikely(!ncsw))
3327 			break;
3328 
3329 		/*
3330 		 * Was it really running after all now that we
3331 		 * checked with the proper locks actually held?
3332 		 *
3333 		 * Oops. Go back and try again..
3334 		 */
3335 		if (unlikely(running)) {
3336 			cpu_relax();
3337 			continue;
3338 		}
3339 
3340 		/*
3341 		 * It's not enough that it's not actively running,
3342 		 * it must be off the runqueue _entirely_, and not
3343 		 * preempted!
3344 		 *
3345 		 * So if it was still runnable (but just not actively
3346 		 * running right now), it's preempted, and we should
3347 		 * yield - it could be a while.
3348 		 */
3349 		if (unlikely(queued)) {
3350 			ktime_t to = NSEC_PER_SEC / HZ;
3351 
3352 			set_current_state(TASK_UNINTERRUPTIBLE);
3353 			schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3354 			continue;
3355 		}
3356 
3357 		/*
3358 		 * Ahh, all good. It wasn't running, and it wasn't
3359 		 * runnable, which means that it will never become
3360 		 * running in the future either. We're all done!
3361 		 */
3362 		break;
3363 	}
3364 
3365 	return ncsw;
3366 }
3367 
3368 /***
3369  * kick_process - kick a running thread to enter/exit the kernel
3370  * @p: the to-be-kicked thread
3371  *
3372  * Cause a process which is running on another CPU to enter
3373  * kernel-mode, without any delay. (to get signals handled.)
3374  *
3375  * NOTE: this function doesn't have to take the runqueue lock,
3376  * because all it wants to ensure is that the remote task enters
3377  * the kernel. If the IPI races and the task has been migrated
3378  * to another CPU then no harm is done and the purpose has been
3379  * achieved as well.
3380  */
kick_process(struct task_struct * p)3381 void kick_process(struct task_struct *p)
3382 {
3383 	int cpu;
3384 
3385 	preempt_disable();
3386 	cpu = task_cpu(p);
3387 	if ((cpu != smp_processor_id()) && task_curr(p))
3388 		smp_send_reschedule(cpu);
3389 	preempt_enable();
3390 }
3391 EXPORT_SYMBOL_GPL(kick_process);
3392 
3393 /*
3394  * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3395  *
3396  * A few notes on cpu_active vs cpu_online:
3397  *
3398  *  - cpu_active must be a subset of cpu_online
3399  *
3400  *  - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3401  *    see __set_cpus_allowed_ptr(). At this point the newly online
3402  *    CPU isn't yet part of the sched domains, and balancing will not
3403  *    see it.
3404  *
3405  *  - on CPU-down we clear cpu_active() to mask the sched domains and
3406  *    avoid the load balancer to place new tasks on the to be removed
3407  *    CPU. Existing tasks will remain running there and will be taken
3408  *    off.
3409  *
3410  * This means that fallback selection must not select !active CPUs.
3411  * And can assume that any active CPU must be online. Conversely
3412  * select_task_rq() below may allow selection of !active CPUs in order
3413  * to satisfy the above rules.
3414  */
select_fallback_rq(int cpu,struct task_struct * p)3415 static int select_fallback_rq(int cpu, struct task_struct *p)
3416 {
3417 	int nid = cpu_to_node(cpu);
3418 	const struct cpumask *nodemask = NULL;
3419 	enum { cpuset, possible, fail } state = cpuset;
3420 	int dest_cpu;
3421 
3422 	/*
3423 	 * If the node that the CPU is on has been offlined, cpu_to_node()
3424 	 * will return -1. There is no CPU on the node, and we should
3425 	 * select the CPU on the other node.
3426 	 */
3427 	if (nid != -1) {
3428 		nodemask = cpumask_of_node(nid);
3429 
3430 		/* Look for allowed, online CPU in same node. */
3431 		for_each_cpu(dest_cpu, nodemask) {
3432 			if (is_cpu_allowed(p, dest_cpu))
3433 				return dest_cpu;
3434 		}
3435 	}
3436 
3437 	for (;;) {
3438 		/* Any allowed, online CPU? */
3439 		for_each_cpu(dest_cpu, p->cpus_ptr) {
3440 			if (!is_cpu_allowed(p, dest_cpu))
3441 				continue;
3442 
3443 			goto out;
3444 		}
3445 
3446 		/* No more Mr. Nice Guy. */
3447 		switch (state) {
3448 		case cpuset:
3449 			if (cpuset_cpus_allowed_fallback(p)) {
3450 				state = possible;
3451 				break;
3452 			}
3453 			fallthrough;
3454 		case possible:
3455 			/*
3456 			 * XXX When called from select_task_rq() we only
3457 			 * hold p->pi_lock and again violate locking order.
3458 			 *
3459 			 * More yuck to audit.
3460 			 */
3461 			do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3462 			state = fail;
3463 			break;
3464 		case fail:
3465 			BUG();
3466 			break;
3467 		}
3468 	}
3469 
3470 out:
3471 	if (state != cpuset) {
3472 		/*
3473 		 * Don't tell them about moving exiting tasks or
3474 		 * kernel threads (both mm NULL), since they never
3475 		 * leave kernel.
3476 		 */
3477 		if (p->mm && printk_ratelimit()) {
3478 			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3479 					task_pid_nr(p), p->comm, cpu);
3480 		}
3481 	}
3482 
3483 	return dest_cpu;
3484 }
3485 
3486 /*
3487  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3488  */
3489 static inline
select_task_rq(struct task_struct * p,int cpu,int wake_flags)3490 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3491 {
3492 	lockdep_assert_held(&p->pi_lock);
3493 
3494 	if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3495 		cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3496 	else
3497 		cpu = cpumask_any(p->cpus_ptr);
3498 
3499 	/*
3500 	 * In order not to call set_task_cpu() on a blocking task we need
3501 	 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3502 	 * CPU.
3503 	 *
3504 	 * Since this is common to all placement strategies, this lives here.
3505 	 *
3506 	 * [ this allows ->select_task() to simply return task_cpu(p) and
3507 	 *   not worry about this generic constraint ]
3508 	 */
3509 	if (unlikely(!is_cpu_allowed(p, cpu)))
3510 		cpu = select_fallback_rq(task_cpu(p), p);
3511 
3512 	return cpu;
3513 }
3514 
sched_set_stop_task(int cpu,struct task_struct * stop)3515 void sched_set_stop_task(int cpu, struct task_struct *stop)
3516 {
3517 	static struct lock_class_key stop_pi_lock;
3518 	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3519 	struct task_struct *old_stop = cpu_rq(cpu)->stop;
3520 
3521 	if (stop) {
3522 		/*
3523 		 * Make it appear like a SCHED_FIFO task, its something
3524 		 * userspace knows about and won't get confused about.
3525 		 *
3526 		 * Also, it will make PI more or less work without too
3527 		 * much confusion -- but then, stop work should not
3528 		 * rely on PI working anyway.
3529 		 */
3530 		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
3531 
3532 		stop->sched_class = &stop_sched_class;
3533 
3534 		/*
3535 		 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3536 		 * adjust the effective priority of a task. As a result,
3537 		 * rt_mutex_setprio() can trigger (RT) balancing operations,
3538 		 * which can then trigger wakeups of the stop thread to push
3539 		 * around the current task.
3540 		 *
3541 		 * The stop task itself will never be part of the PI-chain, it
3542 		 * never blocks, therefore that ->pi_lock recursion is safe.
3543 		 * Tell lockdep about this by placing the stop->pi_lock in its
3544 		 * own class.
3545 		 */
3546 		lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3547 	}
3548 
3549 	cpu_rq(cpu)->stop = stop;
3550 
3551 	if (old_stop) {
3552 		/*
3553 		 * Reset it back to a normal scheduling class so that
3554 		 * it can die in pieces.
3555 		 */
3556 		old_stop->sched_class = &rt_sched_class;
3557 	}
3558 }
3559 
3560 #else /* CONFIG_SMP */
3561 
__set_cpus_allowed_ptr(struct task_struct * p,const struct cpumask * new_mask,u32 flags)3562 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3563 					 const struct cpumask *new_mask,
3564 					 u32 flags)
3565 {
3566 	return set_cpus_allowed_ptr(p, new_mask);
3567 }
3568 
migrate_disable_switch(struct rq * rq,struct task_struct * p)3569 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3570 
rq_has_pinned_tasks(struct rq * rq)3571 static inline bool rq_has_pinned_tasks(struct rq *rq)
3572 {
3573 	return false;
3574 }
3575 
3576 #endif /* !CONFIG_SMP */
3577 
3578 static void
ttwu_stat(struct task_struct * p,int cpu,int wake_flags)3579 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3580 {
3581 	struct rq *rq;
3582 
3583 	if (!schedstat_enabled())
3584 		return;
3585 
3586 	rq = this_rq();
3587 
3588 #ifdef CONFIG_SMP
3589 	if (cpu == rq->cpu) {
3590 		__schedstat_inc(rq->ttwu_local);
3591 		__schedstat_inc(p->stats.nr_wakeups_local);
3592 	} else {
3593 		struct sched_domain *sd;
3594 
3595 		__schedstat_inc(p->stats.nr_wakeups_remote);
3596 		rcu_read_lock();
3597 		for_each_domain(rq->cpu, sd) {
3598 			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3599 				__schedstat_inc(sd->ttwu_wake_remote);
3600 				break;
3601 			}
3602 		}
3603 		rcu_read_unlock();
3604 	}
3605 
3606 	if (wake_flags & WF_MIGRATED)
3607 		__schedstat_inc(p->stats.nr_wakeups_migrate);
3608 #endif /* CONFIG_SMP */
3609 
3610 	__schedstat_inc(rq->ttwu_count);
3611 	__schedstat_inc(p->stats.nr_wakeups);
3612 
3613 	if (wake_flags & WF_SYNC)
3614 		__schedstat_inc(p->stats.nr_wakeups_sync);
3615 }
3616 
3617 /*
3618  * Mark the task runnable and perform wakeup-preemption.
3619  */
ttwu_do_wakeup(struct rq * rq,struct task_struct * p,int wake_flags,struct rq_flags * rf)3620 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3621 			   struct rq_flags *rf)
3622 {
3623 	check_preempt_curr(rq, p, wake_flags);
3624 	WRITE_ONCE(p->__state, TASK_RUNNING);
3625 	trace_sched_wakeup(p);
3626 
3627 #ifdef CONFIG_SMP
3628 	if (p->sched_class->task_woken) {
3629 		/*
3630 		 * Our task @p is fully woken up and running; so it's safe to
3631 		 * drop the rq->lock, hereafter rq is only used for statistics.
3632 		 */
3633 		rq_unpin_lock(rq, rf);
3634 		p->sched_class->task_woken(rq, p);
3635 		rq_repin_lock(rq, rf);
3636 	}
3637 
3638 	if (rq->idle_stamp) {
3639 		u64 delta = rq_clock(rq) - rq->idle_stamp;
3640 		u64 max = 2*rq->max_idle_balance_cost;
3641 
3642 		update_avg(&rq->avg_idle, delta);
3643 
3644 		if (rq->avg_idle > max)
3645 			rq->avg_idle = max;
3646 
3647 		rq->wake_stamp = jiffies;
3648 		rq->wake_avg_idle = rq->avg_idle / 2;
3649 
3650 		rq->idle_stamp = 0;
3651 	}
3652 #endif
3653 }
3654 
3655 static void
ttwu_do_activate(struct rq * rq,struct task_struct * p,int wake_flags,struct rq_flags * rf)3656 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3657 		 struct rq_flags *rf)
3658 {
3659 	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3660 
3661 	lockdep_assert_rq_held(rq);
3662 
3663 	if (p->sched_contributes_to_load)
3664 		rq->nr_uninterruptible--;
3665 
3666 #ifdef CONFIG_SMP
3667 	if (wake_flags & WF_MIGRATED)
3668 		en_flags |= ENQUEUE_MIGRATED;
3669 	else
3670 #endif
3671 	if (p->in_iowait) {
3672 		delayacct_blkio_end(p);
3673 		atomic_dec(&task_rq(p)->nr_iowait);
3674 	}
3675 
3676 	activate_task(rq, p, en_flags);
3677 	ttwu_do_wakeup(rq, p, wake_flags, rf);
3678 }
3679 
3680 /*
3681  * Consider @p being inside a wait loop:
3682  *
3683  *   for (;;) {
3684  *      set_current_state(TASK_UNINTERRUPTIBLE);
3685  *
3686  *      if (CONDITION)
3687  *         break;
3688  *
3689  *      schedule();
3690  *   }
3691  *   __set_current_state(TASK_RUNNING);
3692  *
3693  * between set_current_state() and schedule(). In this case @p is still
3694  * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3695  * an atomic manner.
3696  *
3697  * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3698  * then schedule() must still happen and p->state can be changed to
3699  * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3700  * need to do a full wakeup with enqueue.
3701  *
3702  * Returns: %true when the wakeup is done,
3703  *          %false otherwise.
3704  */
ttwu_runnable(struct task_struct * p,int wake_flags)3705 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3706 {
3707 	struct rq_flags rf;
3708 	struct rq *rq;
3709 	int ret = 0;
3710 
3711 	rq = __task_rq_lock(p, &rf);
3712 	if (task_on_rq_queued(p)) {
3713 		/* check_preempt_curr() may use rq clock */
3714 		update_rq_clock(rq);
3715 		ttwu_do_wakeup(rq, p, wake_flags, &rf);
3716 		ret = 1;
3717 	}
3718 	__task_rq_unlock(rq, &rf);
3719 
3720 	return ret;
3721 }
3722 
3723 #ifdef CONFIG_SMP
sched_ttwu_pending(void * arg)3724 void sched_ttwu_pending(void *arg)
3725 {
3726 	struct llist_node *llist = arg;
3727 	struct rq *rq = this_rq();
3728 	struct task_struct *p, *t;
3729 	struct rq_flags rf;
3730 
3731 	if (!llist)
3732 		return;
3733 
3734 	/*
3735 	 * rq::ttwu_pending racy indication of out-standing wakeups.
3736 	 * Races such that false-negatives are possible, since they
3737 	 * are shorter lived that false-positives would be.
3738 	 */
3739 	WRITE_ONCE(rq->ttwu_pending, 0);
3740 
3741 	rq_lock_irqsave(rq, &rf);
3742 	update_rq_clock(rq);
3743 
3744 	llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3745 		if (WARN_ON_ONCE(p->on_cpu))
3746 			smp_cond_load_acquire(&p->on_cpu, !VAL);
3747 
3748 		if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3749 			set_task_cpu(p, cpu_of(rq));
3750 
3751 		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3752 	}
3753 
3754 	rq_unlock_irqrestore(rq, &rf);
3755 }
3756 
send_call_function_single_ipi(int cpu)3757 void send_call_function_single_ipi(int cpu)
3758 {
3759 	struct rq *rq = cpu_rq(cpu);
3760 
3761 	if (!set_nr_if_polling(rq->idle))
3762 		arch_send_call_function_single_ipi(cpu);
3763 	else
3764 		trace_sched_wake_idle_without_ipi(cpu);
3765 }
3766 
3767 /*
3768  * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3769  * necessary. The wakee CPU on receipt of the IPI will queue the task
3770  * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3771  * of the wakeup instead of the waker.
3772  */
__ttwu_queue_wakelist(struct task_struct * p,int cpu,int wake_flags)3773 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3774 {
3775 	struct rq *rq = cpu_rq(cpu);
3776 
3777 	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3778 
3779 	WRITE_ONCE(rq->ttwu_pending, 1);
3780 	__smp_call_single_queue(cpu, &p->wake_entry.llist);
3781 }
3782 
wake_up_if_idle(int cpu)3783 void wake_up_if_idle(int cpu)
3784 {
3785 	struct rq *rq = cpu_rq(cpu);
3786 	struct rq_flags rf;
3787 
3788 	rcu_read_lock();
3789 
3790 	if (!is_idle_task(rcu_dereference(rq->curr)))
3791 		goto out;
3792 
3793 	rq_lock_irqsave(rq, &rf);
3794 	if (is_idle_task(rq->curr))
3795 		resched_curr(rq);
3796 	/* Else CPU is not idle, do nothing here: */
3797 	rq_unlock_irqrestore(rq, &rf);
3798 
3799 out:
3800 	rcu_read_unlock();
3801 }
3802 
cpus_share_cache(int this_cpu,int that_cpu)3803 bool cpus_share_cache(int this_cpu, int that_cpu)
3804 {
3805 	if (this_cpu == that_cpu)
3806 		return true;
3807 
3808 	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3809 }
3810 
ttwu_queue_cond(struct task_struct * p,int cpu)3811 static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3812 {
3813 	/*
3814 	 * Do not complicate things with the async wake_list while the CPU is
3815 	 * in hotplug state.
3816 	 */
3817 	if (!cpu_active(cpu))
3818 		return false;
3819 
3820 	/* Ensure the task will still be allowed to run on the CPU. */
3821 	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3822 		return false;
3823 
3824 	/*
3825 	 * If the CPU does not share cache, then queue the task on the
3826 	 * remote rqs wakelist to avoid accessing remote data.
3827 	 */
3828 	if (!cpus_share_cache(smp_processor_id(), cpu))
3829 		return true;
3830 
3831 	if (cpu == smp_processor_id())
3832 		return false;
3833 
3834 	/*
3835 	 * If the wakee cpu is idle, or the task is descheduling and the
3836 	 * only running task on the CPU, then use the wakelist to offload
3837 	 * the task activation to the idle (or soon-to-be-idle) CPU as
3838 	 * the current CPU is likely busy. nr_running is checked to
3839 	 * avoid unnecessary task stacking.
3840 	 *
3841 	 * Note that we can only get here with (wakee) p->on_rq=0,
3842 	 * p->on_cpu can be whatever, we've done the dequeue, so
3843 	 * the wakee has been accounted out of ->nr_running.
3844 	 */
3845 	if (!cpu_rq(cpu)->nr_running)
3846 		return true;
3847 
3848 	return false;
3849 }
3850 
ttwu_queue_wakelist(struct task_struct * p,int cpu,int wake_flags)3851 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3852 {
3853 	if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
3854 		sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3855 		__ttwu_queue_wakelist(p, cpu, wake_flags);
3856 		return true;
3857 	}
3858 
3859 	return false;
3860 }
3861 
3862 #else /* !CONFIG_SMP */
3863 
ttwu_queue_wakelist(struct task_struct * p,int cpu,int wake_flags)3864 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3865 {
3866 	return false;
3867 }
3868 
3869 #endif /* CONFIG_SMP */
3870 
ttwu_queue(struct task_struct * p,int cpu,int wake_flags)3871 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3872 {
3873 	struct rq *rq = cpu_rq(cpu);
3874 	struct rq_flags rf;
3875 
3876 	if (ttwu_queue_wakelist(p, cpu, wake_flags))
3877 		return;
3878 
3879 	rq_lock(rq, &rf);
3880 	update_rq_clock(rq);
3881 	ttwu_do_activate(rq, p, wake_flags, &rf);
3882 	rq_unlock(rq, &rf);
3883 }
3884 
3885 /*
3886  * Invoked from try_to_wake_up() to check whether the task can be woken up.
3887  *
3888  * The caller holds p::pi_lock if p != current or has preemption
3889  * disabled when p == current.
3890  *
3891  * The rules of PREEMPT_RT saved_state:
3892  *
3893  *   The related locking code always holds p::pi_lock when updating
3894  *   p::saved_state, which means the code is fully serialized in both cases.
3895  *
3896  *   The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3897  *   bits set. This allows to distinguish all wakeup scenarios.
3898  */
3899 static __always_inline
ttwu_state_match(struct task_struct * p,unsigned int state,int * success)3900 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3901 {
3902 	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3903 		WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3904 			     state != TASK_RTLOCK_WAIT);
3905 	}
3906 
3907 	if (READ_ONCE(p->__state) & state) {
3908 		*success = 1;
3909 		return true;
3910 	}
3911 
3912 #ifdef CONFIG_PREEMPT_RT
3913 	/*
3914 	 * Saved state preserves the task state across blocking on
3915 	 * an RT lock.  If the state matches, set p::saved_state to
3916 	 * TASK_RUNNING, but do not wake the task because it waits
3917 	 * for a lock wakeup. Also indicate success because from
3918 	 * the regular waker's point of view this has succeeded.
3919 	 *
3920 	 * After acquiring the lock the task will restore p::__state
3921 	 * from p::saved_state which ensures that the regular
3922 	 * wakeup is not lost. The restore will also set
3923 	 * p::saved_state to TASK_RUNNING so any further tests will
3924 	 * not result in false positives vs. @success
3925 	 */
3926 	if (p->saved_state & state) {
3927 		p->saved_state = TASK_RUNNING;
3928 		*success = 1;
3929 	}
3930 #endif
3931 	return false;
3932 }
3933 
3934 /*
3935  * Notes on Program-Order guarantees on SMP systems.
3936  *
3937  *  MIGRATION
3938  *
3939  * The basic program-order guarantee on SMP systems is that when a task [t]
3940  * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3941  * execution on its new CPU [c1].
3942  *
3943  * For migration (of runnable tasks) this is provided by the following means:
3944  *
3945  *  A) UNLOCK of the rq(c0)->lock scheduling out task t
3946  *  B) migration for t is required to synchronize *both* rq(c0)->lock and
3947  *     rq(c1)->lock (if not at the same time, then in that order).
3948  *  C) LOCK of the rq(c1)->lock scheduling in task
3949  *
3950  * Release/acquire chaining guarantees that B happens after A and C after B.
3951  * Note: the CPU doing B need not be c0 or c1
3952  *
3953  * Example:
3954  *
3955  *   CPU0            CPU1            CPU2
3956  *
3957  *   LOCK rq(0)->lock
3958  *   sched-out X
3959  *   sched-in Y
3960  *   UNLOCK rq(0)->lock
3961  *
3962  *                                   LOCK rq(0)->lock // orders against CPU0
3963  *                                   dequeue X
3964  *                                   UNLOCK rq(0)->lock
3965  *
3966  *                                   LOCK rq(1)->lock
3967  *                                   enqueue X
3968  *                                   UNLOCK rq(1)->lock
3969  *
3970  *                   LOCK rq(1)->lock // orders against CPU2
3971  *                   sched-out Z
3972  *                   sched-in X
3973  *                   UNLOCK rq(1)->lock
3974  *
3975  *
3976  *  BLOCKING -- aka. SLEEP + WAKEUP
3977  *
3978  * For blocking we (obviously) need to provide the same guarantee as for
3979  * migration. However the means are completely different as there is no lock
3980  * chain to provide order. Instead we do:
3981  *
3982  *   1) smp_store_release(X->on_cpu, 0)   -- finish_task()
3983  *   2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3984  *
3985  * Example:
3986  *
3987  *   CPU0 (schedule)  CPU1 (try_to_wake_up) CPU2 (schedule)
3988  *
3989  *   LOCK rq(0)->lock LOCK X->pi_lock
3990  *   dequeue X
3991  *   sched-out X
3992  *   smp_store_release(X->on_cpu, 0);
3993  *
3994  *                    smp_cond_load_acquire(&X->on_cpu, !VAL);
3995  *                    X->state = WAKING
3996  *                    set_task_cpu(X,2)
3997  *
3998  *                    LOCK rq(2)->lock
3999  *                    enqueue X
4000  *                    X->state = RUNNING
4001  *                    UNLOCK rq(2)->lock
4002  *
4003  *                                          LOCK rq(2)->lock // orders against CPU1
4004  *                                          sched-out Z
4005  *                                          sched-in X
4006  *                                          UNLOCK rq(2)->lock
4007  *
4008  *                    UNLOCK X->pi_lock
4009  *   UNLOCK rq(0)->lock
4010  *
4011  *
4012  * However, for wakeups there is a second guarantee we must provide, namely we
4013  * must ensure that CONDITION=1 done by the caller can not be reordered with
4014  * accesses to the task state; see try_to_wake_up() and set_current_state().
4015  */
4016 
4017 /**
4018  * try_to_wake_up - wake up a thread
4019  * @p: the thread to be awakened
4020  * @state: the mask of task states that can be woken
4021  * @wake_flags: wake modifier flags (WF_*)
4022  *
4023  * Conceptually does:
4024  *
4025  *   If (@state & @p->state) @p->state = TASK_RUNNING.
4026  *
4027  * If the task was not queued/runnable, also place it back on a runqueue.
4028  *
4029  * This function is atomic against schedule() which would dequeue the task.
4030  *
4031  * It issues a full memory barrier before accessing @p->state, see the comment
4032  * with set_current_state().
4033  *
4034  * Uses p->pi_lock to serialize against concurrent wake-ups.
4035  *
4036  * Relies on p->pi_lock stabilizing:
4037  *  - p->sched_class
4038  *  - p->cpus_ptr
4039  *  - p->sched_task_group
4040  * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4041  *
4042  * Tries really hard to only take one task_rq(p)->lock for performance.
4043  * Takes rq->lock in:
4044  *  - ttwu_runnable()    -- old rq, unavoidable, see comment there;
4045  *  - ttwu_queue()       -- new rq, for enqueue of the task;
4046  *  - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4047  *
4048  * As a consequence we race really badly with just about everything. See the
4049  * many memory barriers and their comments for details.
4050  *
4051  * Return: %true if @p->state changes (an actual wakeup was done),
4052  *	   %false otherwise.
4053  */
4054 static int
try_to_wake_up(struct task_struct * p,unsigned int state,int wake_flags)4055 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4056 {
4057 	unsigned long flags;
4058 	int cpu, success = 0;
4059 
4060 	preempt_disable();
4061 	if (p == current) {
4062 		/*
4063 		 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4064 		 * == smp_processor_id()'. Together this means we can special
4065 		 * case the whole 'p->on_rq && ttwu_runnable()' case below
4066 		 * without taking any locks.
4067 		 *
4068 		 * In particular:
4069 		 *  - we rely on Program-Order guarantees for all the ordering,
4070 		 *  - we're serialized against set_special_state() by virtue of
4071 		 *    it disabling IRQs (this allows not taking ->pi_lock).
4072 		 */
4073 		if (!ttwu_state_match(p, state, &success))
4074 			goto out;
4075 
4076 		trace_sched_waking(p);
4077 		WRITE_ONCE(p->__state, TASK_RUNNING);
4078 		trace_sched_wakeup(p);
4079 		goto out;
4080 	}
4081 
4082 	/*
4083 	 * If we are going to wake up a thread waiting for CONDITION we
4084 	 * need to ensure that CONDITION=1 done by the caller can not be
4085 	 * reordered with p->state check below. This pairs with smp_store_mb()
4086 	 * in set_current_state() that the waiting thread does.
4087 	 */
4088 	raw_spin_lock_irqsave(&p->pi_lock, flags);
4089 	smp_mb__after_spinlock();
4090 	if (!ttwu_state_match(p, state, &success))
4091 		goto unlock;
4092 
4093 	trace_sched_waking(p);
4094 
4095 	/*
4096 	 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4097 	 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4098 	 * in smp_cond_load_acquire() below.
4099 	 *
4100 	 * sched_ttwu_pending()			try_to_wake_up()
4101 	 *   STORE p->on_rq = 1			  LOAD p->state
4102 	 *   UNLOCK rq->lock
4103 	 *
4104 	 * __schedule() (switch to task 'p')
4105 	 *   LOCK rq->lock			  smp_rmb();
4106 	 *   smp_mb__after_spinlock();
4107 	 *   UNLOCK rq->lock
4108 	 *
4109 	 * [task p]
4110 	 *   STORE p->state = UNINTERRUPTIBLE	  LOAD p->on_rq
4111 	 *
4112 	 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4113 	 * __schedule().  See the comment for smp_mb__after_spinlock().
4114 	 *
4115 	 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4116 	 */
4117 	smp_rmb();
4118 	if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4119 		goto unlock;
4120 
4121 #ifdef CONFIG_SMP
4122 	/*
4123 	 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4124 	 * possible to, falsely, observe p->on_cpu == 0.
4125 	 *
4126 	 * One must be running (->on_cpu == 1) in order to remove oneself
4127 	 * from the runqueue.
4128 	 *
4129 	 * __schedule() (switch to task 'p')	try_to_wake_up()
4130 	 *   STORE p->on_cpu = 1		  LOAD p->on_rq
4131 	 *   UNLOCK rq->lock
4132 	 *
4133 	 * __schedule() (put 'p' to sleep)
4134 	 *   LOCK rq->lock			  smp_rmb();
4135 	 *   smp_mb__after_spinlock();
4136 	 *   STORE p->on_rq = 0			  LOAD p->on_cpu
4137 	 *
4138 	 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4139 	 * __schedule().  See the comment for smp_mb__after_spinlock().
4140 	 *
4141 	 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4142 	 * schedule()'s deactivate_task() has 'happened' and p will no longer
4143 	 * care about it's own p->state. See the comment in __schedule().
4144 	 */
4145 	smp_acquire__after_ctrl_dep();
4146 
4147 	/*
4148 	 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4149 	 * == 0), which means we need to do an enqueue, change p->state to
4150 	 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4151 	 * enqueue, such as ttwu_queue_wakelist().
4152 	 */
4153 	WRITE_ONCE(p->__state, TASK_WAKING);
4154 
4155 	/*
4156 	 * If the owning (remote) CPU is still in the middle of schedule() with
4157 	 * this task as prev, considering queueing p on the remote CPUs wake_list
4158 	 * which potentially sends an IPI instead of spinning on p->on_cpu to
4159 	 * let the waker make forward progress. This is safe because IRQs are
4160 	 * disabled and the IPI will deliver after on_cpu is cleared.
4161 	 *
4162 	 * Ensure we load task_cpu(p) after p->on_cpu:
4163 	 *
4164 	 * set_task_cpu(p, cpu);
4165 	 *   STORE p->cpu = @cpu
4166 	 * __schedule() (switch to task 'p')
4167 	 *   LOCK rq->lock
4168 	 *   smp_mb__after_spin_lock()		smp_cond_load_acquire(&p->on_cpu)
4169 	 *   STORE p->on_cpu = 1		LOAD p->cpu
4170 	 *
4171 	 * to ensure we observe the correct CPU on which the task is currently
4172 	 * scheduling.
4173 	 */
4174 	if (smp_load_acquire(&p->on_cpu) &&
4175 	    ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4176 		goto unlock;
4177 
4178 	/*
4179 	 * If the owning (remote) CPU is still in the middle of schedule() with
4180 	 * this task as prev, wait until it's done referencing the task.
4181 	 *
4182 	 * Pairs with the smp_store_release() in finish_task().
4183 	 *
4184 	 * This ensures that tasks getting woken will be fully ordered against
4185 	 * their previous state and preserve Program Order.
4186 	 */
4187 	smp_cond_load_acquire(&p->on_cpu, !VAL);
4188 
4189 	cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4190 	if (task_cpu(p) != cpu) {
4191 		if (p->in_iowait) {
4192 			delayacct_blkio_end(p);
4193 			atomic_dec(&task_rq(p)->nr_iowait);
4194 		}
4195 
4196 		wake_flags |= WF_MIGRATED;
4197 		psi_ttwu_dequeue(p);
4198 		set_task_cpu(p, cpu);
4199 	}
4200 #else
4201 	cpu = task_cpu(p);
4202 #endif /* CONFIG_SMP */
4203 
4204 	ttwu_queue(p, cpu, wake_flags);
4205 unlock:
4206 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4207 out:
4208 	if (success)
4209 		ttwu_stat(p, task_cpu(p), wake_flags);
4210 	preempt_enable();
4211 
4212 	return success;
4213 }
4214 
4215 /**
4216  * task_call_func - Invoke a function on task in fixed state
4217  * @p: Process for which the function is to be invoked, can be @current.
4218  * @func: Function to invoke.
4219  * @arg: Argument to function.
4220  *
4221  * Fix the task in it's current state by avoiding wakeups and or rq operations
4222  * and call @func(@arg) on it.  This function can use ->on_rq and task_curr()
4223  * to work out what the state is, if required.  Given that @func can be invoked
4224  * with a runqueue lock held, it had better be quite lightweight.
4225  *
4226  * Returns:
4227  *   Whatever @func returns
4228  */
task_call_func(struct task_struct * p,task_call_f func,void * arg)4229 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4230 {
4231 	struct rq *rq = NULL;
4232 	unsigned int state;
4233 	struct rq_flags rf;
4234 	int ret;
4235 
4236 	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4237 
4238 	state = READ_ONCE(p->__state);
4239 
4240 	/*
4241 	 * Ensure we load p->on_rq after p->__state, otherwise it would be
4242 	 * possible to, falsely, observe p->on_rq == 0.
4243 	 *
4244 	 * See try_to_wake_up() for a longer comment.
4245 	 */
4246 	smp_rmb();
4247 
4248 	/*
4249 	 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4250 	 * the task is blocked. Make sure to check @state since ttwu() can drop
4251 	 * locks at the end, see ttwu_queue_wakelist().
4252 	 */
4253 	if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4254 		rq = __task_rq_lock(p, &rf);
4255 
4256 	/*
4257 	 * At this point the task is pinned; either:
4258 	 *  - blocked and we're holding off wakeups	 (pi->lock)
4259 	 *  - woken, and we're holding off enqueue	 (rq->lock)
4260 	 *  - queued, and we're holding off schedule	 (rq->lock)
4261 	 *  - running, and we're holding off de-schedule (rq->lock)
4262 	 *
4263 	 * The called function (@func) can use: task_curr(), p->on_rq and
4264 	 * p->__state to differentiate between these states.
4265 	 */
4266 	ret = func(p, arg);
4267 
4268 	if (rq)
4269 		rq_unlock(rq, &rf);
4270 
4271 	raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4272 	return ret;
4273 }
4274 
4275 /**
4276  * wake_up_process - Wake up a specific process
4277  * @p: The process to be woken up.
4278  *
4279  * Attempt to wake up the nominated process and move it to the set of runnable
4280  * processes.
4281  *
4282  * Return: 1 if the process was woken up, 0 if it was already running.
4283  *
4284  * This function executes a full memory barrier before accessing the task state.
4285  */
wake_up_process(struct task_struct * p)4286 int wake_up_process(struct task_struct *p)
4287 {
4288 	return try_to_wake_up(p, TASK_NORMAL, 0);
4289 }
4290 EXPORT_SYMBOL(wake_up_process);
4291 
wake_up_state(struct task_struct * p,unsigned int state)4292 int wake_up_state(struct task_struct *p, unsigned int state)
4293 {
4294 	return try_to_wake_up(p, state, 0);
4295 }
4296 
4297 /*
4298  * Perform scheduler related setup for a newly forked process p.
4299  * p is forked by current.
4300  *
4301  * __sched_fork() is basic setup used by init_idle() too:
4302  */
__sched_fork(unsigned long clone_flags,struct task_struct * p)4303 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4304 {
4305 	p->on_rq			= 0;
4306 
4307 	p->se.on_rq			= 0;
4308 	p->se.exec_start		= 0;
4309 	p->se.sum_exec_runtime		= 0;
4310 	p->se.prev_sum_exec_runtime	= 0;
4311 	p->se.nr_migrations		= 0;
4312 	p->se.vruntime			= 0;
4313 	INIT_LIST_HEAD(&p->se.group_node);
4314 
4315 #ifdef CONFIG_FAIR_GROUP_SCHED
4316 	p->se.cfs_rq			= NULL;
4317 #endif
4318 
4319 #ifdef CONFIG_SCHEDSTATS
4320 	/* Even if schedstat is disabled, there should not be garbage */
4321 	memset(&p->stats, 0, sizeof(p->stats));
4322 #endif
4323 
4324 	RB_CLEAR_NODE(&p->dl.rb_node);
4325 	init_dl_task_timer(&p->dl);
4326 	init_dl_inactive_task_timer(&p->dl);
4327 	__dl_clear_params(p);
4328 
4329 	INIT_LIST_HEAD(&p->rt.run_list);
4330 	p->rt.timeout		= 0;
4331 	p->rt.time_slice	= sched_rr_timeslice;
4332 	p->rt.on_rq		= 0;
4333 	p->rt.on_list		= 0;
4334 
4335 #ifdef CONFIG_PREEMPT_NOTIFIERS
4336 	INIT_HLIST_HEAD(&p->preempt_notifiers);
4337 #endif
4338 
4339 #ifdef CONFIG_COMPACTION
4340 	p->capture_control = NULL;
4341 #endif
4342 	init_numa_balancing(clone_flags, p);
4343 #ifdef CONFIG_SMP
4344 	p->wake_entry.u_flags = CSD_TYPE_TTWU;
4345 	p->migration_pending = NULL;
4346 #endif
4347 }
4348 
4349 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4350 
4351 #ifdef CONFIG_NUMA_BALANCING
4352 
4353 int sysctl_numa_balancing_mode;
4354 
__set_numabalancing_state(bool enabled)4355 static void __set_numabalancing_state(bool enabled)
4356 {
4357 	if (enabled)
4358 		static_branch_enable(&sched_numa_balancing);
4359 	else
4360 		static_branch_disable(&sched_numa_balancing);
4361 }
4362 
set_numabalancing_state(bool enabled)4363 void set_numabalancing_state(bool enabled)
4364 {
4365 	if (enabled)
4366 		sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4367 	else
4368 		sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4369 	__set_numabalancing_state(enabled);
4370 }
4371 
4372 #ifdef CONFIG_PROC_SYSCTL
sysctl_numa_balancing(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)4373 int sysctl_numa_balancing(struct ctl_table *table, int write,
4374 			  void *buffer, size_t *lenp, loff_t *ppos)
4375 {
4376 	struct ctl_table t;
4377 	int err;
4378 	int state = sysctl_numa_balancing_mode;
4379 
4380 	if (write && !capable(CAP_SYS_ADMIN))
4381 		return -EPERM;
4382 
4383 	t = *table;
4384 	t.data = &state;
4385 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4386 	if (err < 0)
4387 		return err;
4388 	if (write) {
4389 		sysctl_numa_balancing_mode = state;
4390 		__set_numabalancing_state(state);
4391 	}
4392 	return err;
4393 }
4394 #endif
4395 #endif
4396 
4397 #ifdef CONFIG_SCHEDSTATS
4398 
4399 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4400 
set_schedstats(bool enabled)4401 static void set_schedstats(bool enabled)
4402 {
4403 	if (enabled)
4404 		static_branch_enable(&sched_schedstats);
4405 	else
4406 		static_branch_disable(&sched_schedstats);
4407 }
4408 
force_schedstat_enabled(void)4409 void force_schedstat_enabled(void)
4410 {
4411 	if (!schedstat_enabled()) {
4412 		pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4413 		static_branch_enable(&sched_schedstats);
4414 	}
4415 }
4416 
setup_schedstats(char * str)4417 static int __init setup_schedstats(char *str)
4418 {
4419 	int ret = 0;
4420 	if (!str)
4421 		goto out;
4422 
4423 	if (!strcmp(str, "enable")) {
4424 		set_schedstats(true);
4425 		ret = 1;
4426 	} else if (!strcmp(str, "disable")) {
4427 		set_schedstats(false);
4428 		ret = 1;
4429 	}
4430 out:
4431 	if (!ret)
4432 		pr_warn("Unable to parse schedstats=\n");
4433 
4434 	return ret;
4435 }
4436 __setup("schedstats=", setup_schedstats);
4437 
4438 #ifdef CONFIG_PROC_SYSCTL
sysctl_schedstats(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)4439 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4440 		size_t *lenp, loff_t *ppos)
4441 {
4442 	struct ctl_table t;
4443 	int err;
4444 	int state = static_branch_likely(&sched_schedstats);
4445 
4446 	if (write && !capable(CAP_SYS_ADMIN))
4447 		return -EPERM;
4448 
4449 	t = *table;
4450 	t.data = &state;
4451 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4452 	if (err < 0)
4453 		return err;
4454 	if (write)
4455 		set_schedstats(state);
4456 	return err;
4457 }
4458 #endif /* CONFIG_PROC_SYSCTL */
4459 #endif /* CONFIG_SCHEDSTATS */
4460 
4461 #ifdef CONFIG_SYSCTL
4462 static struct ctl_table sched_core_sysctls[] = {
4463 #ifdef CONFIG_SCHEDSTATS
4464 	{
4465 		.procname       = "sched_schedstats",
4466 		.data           = NULL,
4467 		.maxlen         = sizeof(unsigned int),
4468 		.mode           = 0644,
4469 		.proc_handler   = sysctl_schedstats,
4470 		.extra1         = SYSCTL_ZERO,
4471 		.extra2         = SYSCTL_ONE,
4472 	},
4473 #endif /* CONFIG_SCHEDSTATS */
4474 #ifdef CONFIG_UCLAMP_TASK
4475 	{
4476 		.procname       = "sched_util_clamp_min",
4477 		.data           = &sysctl_sched_uclamp_util_min,
4478 		.maxlen         = sizeof(unsigned int),
4479 		.mode           = 0644,
4480 		.proc_handler   = sysctl_sched_uclamp_handler,
4481 	},
4482 	{
4483 		.procname       = "sched_util_clamp_max",
4484 		.data           = &sysctl_sched_uclamp_util_max,
4485 		.maxlen         = sizeof(unsigned int),
4486 		.mode           = 0644,
4487 		.proc_handler   = sysctl_sched_uclamp_handler,
4488 	},
4489 	{
4490 		.procname       = "sched_util_clamp_min_rt_default",
4491 		.data           = &sysctl_sched_uclamp_util_min_rt_default,
4492 		.maxlen         = sizeof(unsigned int),
4493 		.mode           = 0644,
4494 		.proc_handler   = sysctl_sched_uclamp_handler,
4495 	},
4496 #endif /* CONFIG_UCLAMP_TASK */
4497 	{}
4498 };
sched_core_sysctl_init(void)4499 static int __init sched_core_sysctl_init(void)
4500 {
4501 	register_sysctl_init("kernel", sched_core_sysctls);
4502 	return 0;
4503 }
4504 late_initcall(sched_core_sysctl_init);
4505 #endif /* CONFIG_SYSCTL */
4506 
4507 /*
4508  * fork()/clone()-time setup:
4509  */
sched_fork(unsigned long clone_flags,struct task_struct * p)4510 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4511 {
4512 	__sched_fork(clone_flags, p);
4513 	/*
4514 	 * We mark the process as NEW here. This guarantees that
4515 	 * nobody will actually run it, and a signal or other external
4516 	 * event cannot wake it up and insert it on the runqueue either.
4517 	 */
4518 	p->__state = TASK_NEW;
4519 
4520 	/*
4521 	 * Make sure we do not leak PI boosting priority to the child.
4522 	 */
4523 	p->prio = current->normal_prio;
4524 
4525 	uclamp_fork(p);
4526 
4527 	/*
4528 	 * Revert to default priority/policy on fork if requested.
4529 	 */
4530 	if (unlikely(p->sched_reset_on_fork)) {
4531 		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4532 			p->policy = SCHED_NORMAL;
4533 			p->static_prio = NICE_TO_PRIO(0);
4534 			p->rt_priority = 0;
4535 		} else if (PRIO_TO_NICE(p->static_prio) < 0)
4536 			p->static_prio = NICE_TO_PRIO(0);
4537 
4538 		p->prio = p->normal_prio = p->static_prio;
4539 		set_load_weight(p, false);
4540 
4541 		/*
4542 		 * We don't need the reset flag anymore after the fork. It has
4543 		 * fulfilled its duty:
4544 		 */
4545 		p->sched_reset_on_fork = 0;
4546 	}
4547 
4548 	if (dl_prio(p->prio))
4549 		return -EAGAIN;
4550 	else if (rt_prio(p->prio))
4551 		p->sched_class = &rt_sched_class;
4552 	else
4553 		p->sched_class = &fair_sched_class;
4554 
4555 	init_entity_runnable_average(&p->se);
4556 
4557 
4558 #ifdef CONFIG_SCHED_INFO
4559 	if (likely(sched_info_on()))
4560 		memset(&p->sched_info, 0, sizeof(p->sched_info));
4561 #endif
4562 #if defined(CONFIG_SMP)
4563 	p->on_cpu = 0;
4564 #endif
4565 	init_task_preempt_count(p);
4566 #ifdef CONFIG_SMP
4567 	plist_node_init(&p->pushable_tasks, MAX_PRIO);
4568 	RB_CLEAR_NODE(&p->pushable_dl_tasks);
4569 #endif
4570 	return 0;
4571 }
4572 
sched_cgroup_fork(struct task_struct * p,struct kernel_clone_args * kargs)4573 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4574 {
4575 	unsigned long flags;
4576 
4577 	/*
4578 	 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4579 	 * required yet, but lockdep gets upset if rules are violated.
4580 	 */
4581 	raw_spin_lock_irqsave(&p->pi_lock, flags);
4582 #ifdef CONFIG_CGROUP_SCHED
4583 	if (1) {
4584 		struct task_group *tg;
4585 		tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4586 				  struct task_group, css);
4587 		tg = autogroup_task_group(p, tg);
4588 		p->sched_task_group = tg;
4589 	}
4590 #endif
4591 	rseq_migrate(p);
4592 	/*
4593 	 * We're setting the CPU for the first time, we don't migrate,
4594 	 * so use __set_task_cpu().
4595 	 */
4596 	__set_task_cpu(p, smp_processor_id());
4597 	if (p->sched_class->task_fork)
4598 		p->sched_class->task_fork(p);
4599 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4600 }
4601 
sched_post_fork(struct task_struct * p)4602 void sched_post_fork(struct task_struct *p)
4603 {
4604 	uclamp_post_fork(p);
4605 }
4606 
to_ratio(u64 period,u64 runtime)4607 unsigned long to_ratio(u64 period, u64 runtime)
4608 {
4609 	if (runtime == RUNTIME_INF)
4610 		return BW_UNIT;
4611 
4612 	/*
4613 	 * Doing this here saves a lot of checks in all
4614 	 * the calling paths, and returning zero seems
4615 	 * safe for them anyway.
4616 	 */
4617 	if (period == 0)
4618 		return 0;
4619 
4620 	return div64_u64(runtime << BW_SHIFT, period);
4621 }
4622 
4623 /*
4624  * wake_up_new_task - wake up a newly created task for the first time.
4625  *
4626  * This function will do some initial scheduler statistics housekeeping
4627  * that must be done for every newly created context, then puts the task
4628  * on the runqueue and wakes it.
4629  */
wake_up_new_task(struct task_struct * p)4630 void wake_up_new_task(struct task_struct *p)
4631 {
4632 	struct rq_flags rf;
4633 	struct rq *rq;
4634 
4635 	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4636 	WRITE_ONCE(p->__state, TASK_RUNNING);
4637 #ifdef CONFIG_SMP
4638 	/*
4639 	 * Fork balancing, do it here and not earlier because:
4640 	 *  - cpus_ptr can change in the fork path
4641 	 *  - any previously selected CPU might disappear through hotplug
4642 	 *
4643 	 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4644 	 * as we're not fully set-up yet.
4645 	 */
4646 	p->recent_used_cpu = task_cpu(p);
4647 	rseq_migrate(p);
4648 	__set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4649 #endif
4650 	rq = __task_rq_lock(p, &rf);
4651 	update_rq_clock(rq);
4652 	post_init_entity_util_avg(p);
4653 
4654 	activate_task(rq, p, ENQUEUE_NOCLOCK);
4655 	trace_sched_wakeup_new(p);
4656 	check_preempt_curr(rq, p, WF_FORK);
4657 #ifdef CONFIG_SMP
4658 	if (p->sched_class->task_woken) {
4659 		/*
4660 		 * Nothing relies on rq->lock after this, so it's fine to
4661 		 * drop it.
4662 		 */
4663 		rq_unpin_lock(rq, &rf);
4664 		p->sched_class->task_woken(rq, p);
4665 		rq_repin_lock(rq, &rf);
4666 	}
4667 #endif
4668 	task_rq_unlock(rq, p, &rf);
4669 }
4670 
4671 #ifdef CONFIG_PREEMPT_NOTIFIERS
4672 
4673 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4674 
preempt_notifier_inc(void)4675 void preempt_notifier_inc(void)
4676 {
4677 	static_branch_inc(&preempt_notifier_key);
4678 }
4679 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4680 
preempt_notifier_dec(void)4681 void preempt_notifier_dec(void)
4682 {
4683 	static_branch_dec(&preempt_notifier_key);
4684 }
4685 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4686 
4687 /**
4688  * preempt_notifier_register - tell me when current is being preempted & rescheduled
4689  * @notifier: notifier struct to register
4690  */
preempt_notifier_register(struct preempt_notifier * notifier)4691 void preempt_notifier_register(struct preempt_notifier *notifier)
4692 {
4693 	if (!static_branch_unlikely(&preempt_notifier_key))
4694 		WARN(1, "registering preempt_notifier while notifiers disabled\n");
4695 
4696 	hlist_add_head(&notifier->link, &current->preempt_notifiers);
4697 }
4698 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4699 
4700 /**
4701  * preempt_notifier_unregister - no longer interested in preemption notifications
4702  * @notifier: notifier struct to unregister
4703  *
4704  * This is *not* safe to call from within a preemption notifier.
4705  */
preempt_notifier_unregister(struct preempt_notifier * notifier)4706 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4707 {
4708 	hlist_del(&notifier->link);
4709 }
4710 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4711 
__fire_sched_in_preempt_notifiers(struct task_struct * curr)4712 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4713 {
4714 	struct preempt_notifier *notifier;
4715 
4716 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4717 		notifier->ops->sched_in(notifier, raw_smp_processor_id());
4718 }
4719 
fire_sched_in_preempt_notifiers(struct task_struct * curr)4720 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4721 {
4722 	if (static_branch_unlikely(&preempt_notifier_key))
4723 		__fire_sched_in_preempt_notifiers(curr);
4724 }
4725 
4726 static void
__fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)4727 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4728 				   struct task_struct *next)
4729 {
4730 	struct preempt_notifier *notifier;
4731 
4732 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4733 		notifier->ops->sched_out(notifier, next);
4734 }
4735 
4736 static __always_inline void
fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)4737 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4738 				 struct task_struct *next)
4739 {
4740 	if (static_branch_unlikely(&preempt_notifier_key))
4741 		__fire_sched_out_preempt_notifiers(curr, next);
4742 }
4743 
4744 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4745 
fire_sched_in_preempt_notifiers(struct task_struct * curr)4746 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4747 {
4748 }
4749 
4750 static inline void
fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)4751 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4752 				 struct task_struct *next)
4753 {
4754 }
4755 
4756 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4757 
prepare_task(struct task_struct * next)4758 static inline void prepare_task(struct task_struct *next)
4759 {
4760 #ifdef CONFIG_SMP
4761 	/*
4762 	 * Claim the task as running, we do this before switching to it
4763 	 * such that any running task will have this set.
4764 	 *
4765 	 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4766 	 * its ordering comment.
4767 	 */
4768 	WRITE_ONCE(next->on_cpu, 1);
4769 #endif
4770 }
4771 
finish_task(struct task_struct * prev)4772 static inline void finish_task(struct task_struct *prev)
4773 {
4774 #ifdef CONFIG_SMP
4775 	/*
4776 	 * This must be the very last reference to @prev from this CPU. After
4777 	 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4778 	 * must ensure this doesn't happen until the switch is completely
4779 	 * finished.
4780 	 *
4781 	 * In particular, the load of prev->state in finish_task_switch() must
4782 	 * happen before this.
4783 	 *
4784 	 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4785 	 */
4786 	smp_store_release(&prev->on_cpu, 0);
4787 #endif
4788 }
4789 
4790 #ifdef CONFIG_SMP
4791 
do_balance_callbacks(struct rq * rq,struct callback_head * head)4792 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4793 {
4794 	void (*func)(struct rq *rq);
4795 	struct callback_head *next;
4796 
4797 	lockdep_assert_rq_held(rq);
4798 
4799 	while (head) {
4800 		func = (void (*)(struct rq *))head->func;
4801 		next = head->next;
4802 		head->next = NULL;
4803 		head = next;
4804 
4805 		func(rq);
4806 	}
4807 }
4808 
4809 static void balance_push(struct rq *rq);
4810 
4811 /*
4812  * balance_push_callback is a right abuse of the callback interface and plays
4813  * by significantly different rules.
4814  *
4815  * Where the normal balance_callback's purpose is to be ran in the same context
4816  * that queued it (only later, when it's safe to drop rq->lock again),
4817  * balance_push_callback is specifically targeted at __schedule().
4818  *
4819  * This abuse is tolerated because it places all the unlikely/odd cases behind
4820  * a single test, namely: rq->balance_callback == NULL.
4821  */
4822 struct callback_head balance_push_callback = {
4823 	.next = NULL,
4824 	.func = (void (*)(struct callback_head *))balance_push,
4825 };
4826 
4827 static inline struct callback_head *
__splice_balance_callbacks(struct rq * rq,bool split)4828 __splice_balance_callbacks(struct rq *rq, bool split)
4829 {
4830 	struct callback_head *head = rq->balance_callback;
4831 
4832 	if (likely(!head))
4833 		return NULL;
4834 
4835 	lockdep_assert_rq_held(rq);
4836 	/*
4837 	 * Must not take balance_push_callback off the list when
4838 	 * splice_balance_callbacks() and balance_callbacks() are not
4839 	 * in the same rq->lock section.
4840 	 *
4841 	 * In that case it would be possible for __schedule() to interleave
4842 	 * and observe the list empty.
4843 	 */
4844 	if (split && head == &balance_push_callback)
4845 		head = NULL;
4846 	else
4847 		rq->balance_callback = NULL;
4848 
4849 	return head;
4850 }
4851 
splice_balance_callbacks(struct rq * rq)4852 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4853 {
4854 	return __splice_balance_callbacks(rq, true);
4855 }
4856 
__balance_callbacks(struct rq * rq)4857 static void __balance_callbacks(struct rq *rq)
4858 {
4859 	do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
4860 }
4861 
balance_callbacks(struct rq * rq,struct callback_head * head)4862 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4863 {
4864 	unsigned long flags;
4865 
4866 	if (unlikely(head)) {
4867 		raw_spin_rq_lock_irqsave(rq, flags);
4868 		do_balance_callbacks(rq, head);
4869 		raw_spin_rq_unlock_irqrestore(rq, flags);
4870 	}
4871 }
4872 
4873 #else
4874 
__balance_callbacks(struct rq * rq)4875 static inline void __balance_callbacks(struct rq *rq)
4876 {
4877 }
4878 
splice_balance_callbacks(struct rq * rq)4879 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4880 {
4881 	return NULL;
4882 }
4883 
balance_callbacks(struct rq * rq,struct callback_head * head)4884 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4885 {
4886 }
4887 
4888 #endif
4889 
4890 static inline void
prepare_lock_switch(struct rq * rq,struct task_struct * next,struct rq_flags * rf)4891 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4892 {
4893 	/*
4894 	 * Since the runqueue lock will be released by the next
4895 	 * task (which is an invalid locking op but in the case
4896 	 * of the scheduler it's an obvious special-case), so we
4897 	 * do an early lockdep release here:
4898 	 */
4899 	rq_unpin_lock(rq, rf);
4900 	spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4901 #ifdef CONFIG_DEBUG_SPINLOCK
4902 	/* this is a valid case when another task releases the spinlock */
4903 	rq_lockp(rq)->owner = next;
4904 #endif
4905 }
4906 
finish_lock_switch(struct rq * rq)4907 static inline void finish_lock_switch(struct rq *rq)
4908 {
4909 	/*
4910 	 * If we are tracking spinlock dependencies then we have to
4911 	 * fix up the runqueue lock - which gets 'carried over' from
4912 	 * prev into current:
4913 	 */
4914 	spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4915 	__balance_callbacks(rq);
4916 	raw_spin_rq_unlock_irq(rq);
4917 }
4918 
4919 /*
4920  * NOP if the arch has not defined these:
4921  */
4922 
4923 #ifndef prepare_arch_switch
4924 # define prepare_arch_switch(next)	do { } while (0)
4925 #endif
4926 
4927 #ifndef finish_arch_post_lock_switch
4928 # define finish_arch_post_lock_switch()	do { } while (0)
4929 #endif
4930 
kmap_local_sched_out(void)4931 static inline void kmap_local_sched_out(void)
4932 {
4933 #ifdef CONFIG_KMAP_LOCAL
4934 	if (unlikely(current->kmap_ctrl.idx))
4935 		__kmap_local_sched_out();
4936 #endif
4937 }
4938 
kmap_local_sched_in(void)4939 static inline void kmap_local_sched_in(void)
4940 {
4941 #ifdef CONFIG_KMAP_LOCAL
4942 	if (unlikely(current->kmap_ctrl.idx))
4943 		__kmap_local_sched_in();
4944 #endif
4945 }
4946 
4947 /**
4948  * prepare_task_switch - prepare to switch tasks
4949  * @rq: the runqueue preparing to switch
4950  * @prev: the current task that is being switched out
4951  * @next: the task we are going to switch to.
4952  *
4953  * This is called with the rq lock held and interrupts off. It must
4954  * be paired with a subsequent finish_task_switch after the context
4955  * switch.
4956  *
4957  * prepare_task_switch sets up locking and calls architecture specific
4958  * hooks.
4959  */
4960 static inline void
prepare_task_switch(struct rq * rq,struct task_struct * prev,struct task_struct * next)4961 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4962 		    struct task_struct *next)
4963 {
4964 	kcov_prepare_switch(prev);
4965 	sched_info_switch(rq, prev, next);
4966 	perf_event_task_sched_out(prev, next);
4967 	rseq_preempt(prev);
4968 	fire_sched_out_preempt_notifiers(prev, next);
4969 	kmap_local_sched_out();
4970 	prepare_task(next);
4971 	prepare_arch_switch(next);
4972 }
4973 
4974 /**
4975  * finish_task_switch - clean up after a task-switch
4976  * @prev: the thread we just switched away from.
4977  *
4978  * finish_task_switch must be called after the context switch, paired
4979  * with a prepare_task_switch call before the context switch.
4980  * finish_task_switch will reconcile locking set up by prepare_task_switch,
4981  * and do any other architecture-specific cleanup actions.
4982  *
4983  * Note that we may have delayed dropping an mm in context_switch(). If
4984  * so, we finish that here outside of the runqueue lock. (Doing it
4985  * with the lock held can cause deadlocks; see schedule() for
4986  * details.)
4987  *
4988  * The context switch have flipped the stack from under us and restored the
4989  * local variables which were saved when this task called schedule() in the
4990  * past. prev == current is still correct but we need to recalculate this_rq
4991  * because prev may have moved to another CPU.
4992  */
finish_task_switch(struct task_struct * prev)4993 static struct rq *finish_task_switch(struct task_struct *prev)
4994 	__releases(rq->lock)
4995 {
4996 	struct rq *rq = this_rq();
4997 	struct mm_struct *mm = rq->prev_mm;
4998 	unsigned int prev_state;
4999 
5000 	/*
5001 	 * The previous task will have left us with a preempt_count of 2
5002 	 * because it left us after:
5003 	 *
5004 	 *	schedule()
5005 	 *	  preempt_disable();			// 1
5006 	 *	  __schedule()
5007 	 *	    raw_spin_lock_irq(&rq->lock)	// 2
5008 	 *
5009 	 * Also, see FORK_PREEMPT_COUNT.
5010 	 */
5011 	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5012 		      "corrupted preempt_count: %s/%d/0x%x\n",
5013 		      current->comm, current->pid, preempt_count()))
5014 		preempt_count_set(FORK_PREEMPT_COUNT);
5015 
5016 	rq->prev_mm = NULL;
5017 
5018 	/*
5019 	 * A task struct has one reference for the use as "current".
5020 	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5021 	 * schedule one last time. The schedule call will never return, and
5022 	 * the scheduled task must drop that reference.
5023 	 *
5024 	 * We must observe prev->state before clearing prev->on_cpu (in
5025 	 * finish_task), otherwise a concurrent wakeup can get prev
5026 	 * running on another CPU and we could rave with its RUNNING -> DEAD
5027 	 * transition, resulting in a double drop.
5028 	 */
5029 	prev_state = READ_ONCE(prev->__state);
5030 	vtime_task_switch(prev);
5031 	perf_event_task_sched_in(prev, current);
5032 	finish_task(prev);
5033 	tick_nohz_task_switch();
5034 	finish_lock_switch(rq);
5035 	finish_arch_post_lock_switch();
5036 	kcov_finish_switch(current);
5037 	/*
5038 	 * kmap_local_sched_out() is invoked with rq::lock held and
5039 	 * interrupts disabled. There is no requirement for that, but the
5040 	 * sched out code does not have an interrupt enabled section.
5041 	 * Restoring the maps on sched in does not require interrupts being
5042 	 * disabled either.
5043 	 */
5044 	kmap_local_sched_in();
5045 
5046 	fire_sched_in_preempt_notifiers(current);
5047 	/*
5048 	 * When switching through a kernel thread, the loop in
5049 	 * membarrier_{private,global}_expedited() may have observed that
5050 	 * kernel thread and not issued an IPI. It is therefore possible to
5051 	 * schedule between user->kernel->user threads without passing though
5052 	 * switch_mm(). Membarrier requires a barrier after storing to
5053 	 * rq->curr, before returning to userspace, so provide them here:
5054 	 *
5055 	 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5056 	 *   provided by mmdrop(),
5057 	 * - a sync_core for SYNC_CORE.
5058 	 */
5059 	if (mm) {
5060 		membarrier_mm_sync_core_before_usermode(mm);
5061 		mmdrop_sched(mm);
5062 	}
5063 	if (unlikely(prev_state == TASK_DEAD)) {
5064 		if (prev->sched_class->task_dead)
5065 			prev->sched_class->task_dead(prev);
5066 
5067 		/* Task is done with its stack. */
5068 		put_task_stack(prev);
5069 
5070 		put_task_struct_rcu_user(prev);
5071 	}
5072 
5073 	return rq;
5074 }
5075 
5076 /**
5077  * schedule_tail - first thing a freshly forked thread must call.
5078  * @prev: the thread we just switched away from.
5079  */
schedule_tail(struct task_struct * prev)5080 asmlinkage __visible void schedule_tail(struct task_struct *prev)
5081 	__releases(rq->lock)
5082 {
5083 	/*
5084 	 * New tasks start with FORK_PREEMPT_COUNT, see there and
5085 	 * finish_task_switch() for details.
5086 	 *
5087 	 * finish_task_switch() will drop rq->lock() and lower preempt_count
5088 	 * and the preempt_enable() will end up enabling preemption (on
5089 	 * PREEMPT_COUNT kernels).
5090 	 */
5091 
5092 	finish_task_switch(prev);
5093 	preempt_enable();
5094 
5095 	if (current->set_child_tid)
5096 		put_user(task_pid_vnr(current), current->set_child_tid);
5097 
5098 	calculate_sigpending();
5099 }
5100 
5101 /*
5102  * context_switch - switch to the new MM and the new thread's register state.
5103  */
5104 static __always_inline struct rq *
context_switch(struct rq * rq,struct task_struct * prev,struct task_struct * next,struct rq_flags * rf)5105 context_switch(struct rq *rq, struct task_struct *prev,
5106 	       struct task_struct *next, struct rq_flags *rf)
5107 {
5108 	prepare_task_switch(rq, prev, next);
5109 
5110 	/*
5111 	 * For paravirt, this is coupled with an exit in switch_to to
5112 	 * combine the page table reload and the switch backend into
5113 	 * one hypercall.
5114 	 */
5115 	arch_start_context_switch(prev);
5116 
5117 	/*
5118 	 * kernel -> kernel   lazy + transfer active
5119 	 *   user -> kernel   lazy + mmgrab() active
5120 	 *
5121 	 * kernel ->   user   switch + mmdrop() active
5122 	 *   user ->   user   switch
5123 	 */
5124 	if (!next->mm) {                                // to kernel
5125 		enter_lazy_tlb(prev->active_mm, next);
5126 
5127 		next->active_mm = prev->active_mm;
5128 		if (prev->mm)                           // from user
5129 			mmgrab(prev->active_mm);
5130 		else
5131 			prev->active_mm = NULL;
5132 	} else {                                        // to user
5133 		membarrier_switch_mm(rq, prev->active_mm, next->mm);
5134 		/*
5135 		 * sys_membarrier() requires an smp_mb() between setting
5136 		 * rq->curr / membarrier_switch_mm() and returning to userspace.
5137 		 *
5138 		 * The below provides this either through switch_mm(), or in
5139 		 * case 'prev->active_mm == next->mm' through
5140 		 * finish_task_switch()'s mmdrop().
5141 		 */
5142 		switch_mm_irqs_off(prev->active_mm, next->mm, next);
5143 
5144 		if (!prev->mm) {                        // from kernel
5145 			/* will mmdrop() in finish_task_switch(). */
5146 			rq->prev_mm = prev->active_mm;
5147 			prev->active_mm = NULL;
5148 		}
5149 	}
5150 
5151 	rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5152 
5153 	prepare_lock_switch(rq, next, rf);
5154 
5155 	/* Here we just switch the register state and the stack. */
5156 	switch_to(prev, next, prev);
5157 	barrier();
5158 
5159 	return finish_task_switch(prev);
5160 }
5161 
5162 /*
5163  * nr_running and nr_context_switches:
5164  *
5165  * externally visible scheduler statistics: current number of runnable
5166  * threads, total number of context switches performed since bootup.
5167  */
nr_running(void)5168 unsigned int nr_running(void)
5169 {
5170 	unsigned int i, sum = 0;
5171 
5172 	for_each_online_cpu(i)
5173 		sum += cpu_rq(i)->nr_running;
5174 
5175 	return sum;
5176 }
5177 
5178 /*
5179  * Check if only the current task is running on the CPU.
5180  *
5181  * Caution: this function does not check that the caller has disabled
5182  * preemption, thus the result might have a time-of-check-to-time-of-use
5183  * race.  The caller is responsible to use it correctly, for example:
5184  *
5185  * - from a non-preemptible section (of course)
5186  *
5187  * - from a thread that is bound to a single CPU
5188  *
5189  * - in a loop with very short iterations (e.g. a polling loop)
5190  */
single_task_running(void)5191 bool single_task_running(void)
5192 {
5193 	return raw_rq()->nr_running == 1;
5194 }
5195 EXPORT_SYMBOL(single_task_running);
5196 
nr_context_switches(void)5197 unsigned long long nr_context_switches(void)
5198 {
5199 	int i;
5200 	unsigned long long sum = 0;
5201 
5202 	for_each_possible_cpu(i)
5203 		sum += cpu_rq(i)->nr_switches;
5204 
5205 	return sum;
5206 }
5207 
5208 /*
5209  * Consumers of these two interfaces, like for example the cpuidle menu
5210  * governor, are using nonsensical data. Preferring shallow idle state selection
5211  * for a CPU that has IO-wait which might not even end up running the task when
5212  * it does become runnable.
5213  */
5214 
nr_iowait_cpu(int cpu)5215 unsigned int nr_iowait_cpu(int cpu)
5216 {
5217 	return atomic_read(&cpu_rq(cpu)->nr_iowait);
5218 }
5219 
5220 /*
5221  * IO-wait accounting, and how it's mostly bollocks (on SMP).
5222  *
5223  * The idea behind IO-wait account is to account the idle time that we could
5224  * have spend running if it were not for IO. That is, if we were to improve the
5225  * storage performance, we'd have a proportional reduction in IO-wait time.
5226  *
5227  * This all works nicely on UP, where, when a task blocks on IO, we account
5228  * idle time as IO-wait, because if the storage were faster, it could've been
5229  * running and we'd not be idle.
5230  *
5231  * This has been extended to SMP, by doing the same for each CPU. This however
5232  * is broken.
5233  *
5234  * Imagine for instance the case where two tasks block on one CPU, only the one
5235  * CPU will have IO-wait accounted, while the other has regular idle. Even
5236  * though, if the storage were faster, both could've ran at the same time,
5237  * utilising both CPUs.
5238  *
5239  * This means, that when looking globally, the current IO-wait accounting on
5240  * SMP is a lower bound, by reason of under accounting.
5241  *
5242  * Worse, since the numbers are provided per CPU, they are sometimes
5243  * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5244  * associated with any one particular CPU, it can wake to another CPU than it
5245  * blocked on. This means the per CPU IO-wait number is meaningless.
5246  *
5247  * Task CPU affinities can make all that even more 'interesting'.
5248  */
5249 
nr_iowait(void)5250 unsigned int nr_iowait(void)
5251 {
5252 	unsigned int i, sum = 0;
5253 
5254 	for_each_possible_cpu(i)
5255 		sum += nr_iowait_cpu(i);
5256 
5257 	return sum;
5258 }
5259 
5260 #ifdef CONFIG_SMP
5261 
5262 /*
5263  * sched_exec - execve() is a valuable balancing opportunity, because at
5264  * this point the task has the smallest effective memory and cache footprint.
5265  */
sched_exec(void)5266 void sched_exec(void)
5267 {
5268 	struct task_struct *p = current;
5269 	unsigned long flags;
5270 	int dest_cpu;
5271 
5272 	raw_spin_lock_irqsave(&p->pi_lock, flags);
5273 	dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5274 	if (dest_cpu == smp_processor_id())
5275 		goto unlock;
5276 
5277 	if (likely(cpu_active(dest_cpu))) {
5278 		struct migration_arg arg = { p, dest_cpu };
5279 
5280 		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5281 		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5282 		return;
5283 	}
5284 unlock:
5285 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5286 }
5287 
5288 #endif
5289 
5290 DEFINE_PER_CPU(struct kernel_stat, kstat);
5291 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5292 
5293 EXPORT_PER_CPU_SYMBOL(kstat);
5294 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5295 
5296 /*
5297  * The function fair_sched_class.update_curr accesses the struct curr
5298  * and its field curr->exec_start; when called from task_sched_runtime(),
5299  * we observe a high rate of cache misses in practice.
5300  * Prefetching this data results in improved performance.
5301  */
prefetch_curr_exec_start(struct task_struct * p)5302 static inline void prefetch_curr_exec_start(struct task_struct *p)
5303 {
5304 #ifdef CONFIG_FAIR_GROUP_SCHED
5305 	struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5306 #else
5307 	struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5308 #endif
5309 	prefetch(curr);
5310 	prefetch(&curr->exec_start);
5311 }
5312 
5313 /*
5314  * Return accounted runtime for the task.
5315  * In case the task is currently running, return the runtime plus current's
5316  * pending runtime that have not been accounted yet.
5317  */
task_sched_runtime(struct task_struct * p)5318 unsigned long long task_sched_runtime(struct task_struct *p)
5319 {
5320 	struct rq_flags rf;
5321 	struct rq *rq;
5322 	u64 ns;
5323 
5324 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5325 	/*
5326 	 * 64-bit doesn't need locks to atomically read a 64-bit value.
5327 	 * So we have a optimization chance when the task's delta_exec is 0.
5328 	 * Reading ->on_cpu is racy, but this is ok.
5329 	 *
5330 	 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5331 	 * If we race with it entering CPU, unaccounted time is 0. This is
5332 	 * indistinguishable from the read occurring a few cycles earlier.
5333 	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5334 	 * been accounted, so we're correct here as well.
5335 	 */
5336 	if (!p->on_cpu || !task_on_rq_queued(p))
5337 		return p->se.sum_exec_runtime;
5338 #endif
5339 
5340 	rq = task_rq_lock(p, &rf);
5341 	/*
5342 	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
5343 	 * project cycles that may never be accounted to this
5344 	 * thread, breaking clock_gettime().
5345 	 */
5346 	if (task_current(rq, p) && task_on_rq_queued(p)) {
5347 		prefetch_curr_exec_start(p);
5348 		update_rq_clock(rq);
5349 		p->sched_class->update_curr(rq);
5350 	}
5351 	ns = p->se.sum_exec_runtime;
5352 	task_rq_unlock(rq, p, &rf);
5353 
5354 	return ns;
5355 }
5356 
5357 #ifdef CONFIG_SCHED_DEBUG
cpu_resched_latency(struct rq * rq)5358 static u64 cpu_resched_latency(struct rq *rq)
5359 {
5360 	int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5361 	u64 resched_latency, now = rq_clock(rq);
5362 	static bool warned_once;
5363 
5364 	if (sysctl_resched_latency_warn_once && warned_once)
5365 		return 0;
5366 
5367 	if (!need_resched() || !latency_warn_ms)
5368 		return 0;
5369 
5370 	if (system_state == SYSTEM_BOOTING)
5371 		return 0;
5372 
5373 	if (!rq->last_seen_need_resched_ns) {
5374 		rq->last_seen_need_resched_ns = now;
5375 		rq->ticks_without_resched = 0;
5376 		return 0;
5377 	}
5378 
5379 	rq->ticks_without_resched++;
5380 	resched_latency = now - rq->last_seen_need_resched_ns;
5381 	if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5382 		return 0;
5383 
5384 	warned_once = true;
5385 
5386 	return resched_latency;
5387 }
5388 
setup_resched_latency_warn_ms(char * str)5389 static int __init setup_resched_latency_warn_ms(char *str)
5390 {
5391 	long val;
5392 
5393 	if ((kstrtol(str, 0, &val))) {
5394 		pr_warn("Unable to set resched_latency_warn_ms\n");
5395 		return 1;
5396 	}
5397 
5398 	sysctl_resched_latency_warn_ms = val;
5399 	return 1;
5400 }
5401 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5402 #else
cpu_resched_latency(struct rq * rq)5403 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5404 #endif /* CONFIG_SCHED_DEBUG */
5405 
5406 /*
5407  * This function gets called by the timer code, with HZ frequency.
5408  * We call it with interrupts disabled.
5409  */
scheduler_tick(void)5410 void scheduler_tick(void)
5411 {
5412 	int cpu = smp_processor_id();
5413 	struct rq *rq = cpu_rq(cpu);
5414 	struct task_struct *curr = rq->curr;
5415 	struct rq_flags rf;
5416 	unsigned long thermal_pressure;
5417 	u64 resched_latency;
5418 
5419 	arch_scale_freq_tick();
5420 	sched_clock_tick();
5421 
5422 	rq_lock(rq, &rf);
5423 
5424 	update_rq_clock(rq);
5425 	thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5426 	update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5427 	curr->sched_class->task_tick(rq, curr, 0);
5428 	if (sched_feat(LATENCY_WARN))
5429 		resched_latency = cpu_resched_latency(rq);
5430 	calc_global_load_tick(rq);
5431 	sched_core_tick(rq);
5432 
5433 	rq_unlock(rq, &rf);
5434 
5435 	if (sched_feat(LATENCY_WARN) && resched_latency)
5436 		resched_latency_warn(cpu, resched_latency);
5437 
5438 	perf_event_task_tick();
5439 
5440 #ifdef CONFIG_SMP
5441 	rq->idle_balance = idle_cpu(cpu);
5442 	trigger_load_balance(rq);
5443 #endif
5444 }
5445 
5446 #ifdef CONFIG_NO_HZ_FULL
5447 
5448 struct tick_work {
5449 	int			cpu;
5450 	atomic_t		state;
5451 	struct delayed_work	work;
5452 };
5453 /* Values for ->state, see diagram below. */
5454 #define TICK_SCHED_REMOTE_OFFLINE	0
5455 #define TICK_SCHED_REMOTE_OFFLINING	1
5456 #define TICK_SCHED_REMOTE_RUNNING	2
5457 
5458 /*
5459  * State diagram for ->state:
5460  *
5461  *
5462  *          TICK_SCHED_REMOTE_OFFLINE
5463  *                    |   ^
5464  *                    |   |
5465  *                    |   | sched_tick_remote()
5466  *                    |   |
5467  *                    |   |
5468  *                    +--TICK_SCHED_REMOTE_OFFLINING
5469  *                    |   ^
5470  *                    |   |
5471  * sched_tick_start() |   | sched_tick_stop()
5472  *                    |   |
5473  *                    V   |
5474  *          TICK_SCHED_REMOTE_RUNNING
5475  *
5476  *
5477  * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5478  * and sched_tick_start() are happy to leave the state in RUNNING.
5479  */
5480 
5481 static struct tick_work __percpu *tick_work_cpu;
5482 
sched_tick_remote(struct work_struct * work)5483 static void sched_tick_remote(struct work_struct *work)
5484 {
5485 	struct delayed_work *dwork = to_delayed_work(work);
5486 	struct tick_work *twork = container_of(dwork, struct tick_work, work);
5487 	int cpu = twork->cpu;
5488 	struct rq *rq = cpu_rq(cpu);
5489 	struct task_struct *curr;
5490 	struct rq_flags rf;
5491 	u64 delta;
5492 	int os;
5493 
5494 	/*
5495 	 * Handle the tick only if it appears the remote CPU is running in full
5496 	 * dynticks mode. The check is racy by nature, but missing a tick or
5497 	 * having one too much is no big deal because the scheduler tick updates
5498 	 * statistics and checks timeslices in a time-independent way, regardless
5499 	 * of when exactly it is running.
5500 	 */
5501 	if (!tick_nohz_tick_stopped_cpu(cpu))
5502 		goto out_requeue;
5503 
5504 	rq_lock_irq(rq, &rf);
5505 	curr = rq->curr;
5506 	if (cpu_is_offline(cpu))
5507 		goto out_unlock;
5508 
5509 	update_rq_clock(rq);
5510 
5511 	if (!is_idle_task(curr)) {
5512 		/*
5513 		 * Make sure the next tick runs within a reasonable
5514 		 * amount of time.
5515 		 */
5516 		delta = rq_clock_task(rq) - curr->se.exec_start;
5517 		WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5518 	}
5519 	curr->sched_class->task_tick(rq, curr, 0);
5520 
5521 	calc_load_nohz_remote(rq);
5522 out_unlock:
5523 	rq_unlock_irq(rq, &rf);
5524 out_requeue:
5525 
5526 	/*
5527 	 * Run the remote tick once per second (1Hz). This arbitrary
5528 	 * frequency is large enough to avoid overload but short enough
5529 	 * to keep scheduler internal stats reasonably up to date.  But
5530 	 * first update state to reflect hotplug activity if required.
5531 	 */
5532 	os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5533 	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5534 	if (os == TICK_SCHED_REMOTE_RUNNING)
5535 		queue_delayed_work(system_unbound_wq, dwork, HZ);
5536 }
5537 
sched_tick_start(int cpu)5538 static void sched_tick_start(int cpu)
5539 {
5540 	int os;
5541 	struct tick_work *twork;
5542 
5543 	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5544 		return;
5545 
5546 	WARN_ON_ONCE(!tick_work_cpu);
5547 
5548 	twork = per_cpu_ptr(tick_work_cpu, cpu);
5549 	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5550 	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5551 	if (os == TICK_SCHED_REMOTE_OFFLINE) {
5552 		twork->cpu = cpu;
5553 		INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5554 		queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5555 	}
5556 }
5557 
5558 #ifdef CONFIG_HOTPLUG_CPU
sched_tick_stop(int cpu)5559 static void sched_tick_stop(int cpu)
5560 {
5561 	struct tick_work *twork;
5562 	int os;
5563 
5564 	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5565 		return;
5566 
5567 	WARN_ON_ONCE(!tick_work_cpu);
5568 
5569 	twork = per_cpu_ptr(tick_work_cpu, cpu);
5570 	/* There cannot be competing actions, but don't rely on stop-machine. */
5571 	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5572 	WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5573 	/* Don't cancel, as this would mess up the state machine. */
5574 }
5575 #endif /* CONFIG_HOTPLUG_CPU */
5576 
sched_tick_offload_init(void)5577 int __init sched_tick_offload_init(void)
5578 {
5579 	tick_work_cpu = alloc_percpu(struct tick_work);
5580 	BUG_ON(!tick_work_cpu);
5581 	return 0;
5582 }
5583 
5584 #else /* !CONFIG_NO_HZ_FULL */
sched_tick_start(int cpu)5585 static inline void sched_tick_start(int cpu) { }
sched_tick_stop(int cpu)5586 static inline void sched_tick_stop(int cpu) { }
5587 #endif
5588 
5589 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5590 				defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5591 /*
5592  * If the value passed in is equal to the current preempt count
5593  * then we just disabled preemption. Start timing the latency.
5594  */
preempt_latency_start(int val)5595 static inline void preempt_latency_start(int val)
5596 {
5597 	if (preempt_count() == val) {
5598 		unsigned long ip = get_lock_parent_ip();
5599 #ifdef CONFIG_DEBUG_PREEMPT
5600 		current->preempt_disable_ip = ip;
5601 #endif
5602 		trace_preempt_off(CALLER_ADDR0, ip);
5603 	}
5604 }
5605 
preempt_count_add(int val)5606 void preempt_count_add(int val)
5607 {
5608 #ifdef CONFIG_DEBUG_PREEMPT
5609 	/*
5610 	 * Underflow?
5611 	 */
5612 	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5613 		return;
5614 #endif
5615 	__preempt_count_add(val);
5616 #ifdef CONFIG_DEBUG_PREEMPT
5617 	/*
5618 	 * Spinlock count overflowing soon?
5619 	 */
5620 	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5621 				PREEMPT_MASK - 10);
5622 #endif
5623 	preempt_latency_start(val);
5624 }
5625 EXPORT_SYMBOL(preempt_count_add);
5626 NOKPROBE_SYMBOL(preempt_count_add);
5627 
5628 /*
5629  * If the value passed in equals to the current preempt count
5630  * then we just enabled preemption. Stop timing the latency.
5631  */
preempt_latency_stop(int val)5632 static inline void preempt_latency_stop(int val)
5633 {
5634 	if (preempt_count() == val)
5635 		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5636 }
5637 
preempt_count_sub(int val)5638 void preempt_count_sub(int val)
5639 {
5640 #ifdef CONFIG_DEBUG_PREEMPT
5641 	/*
5642 	 * Underflow?
5643 	 */
5644 	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5645 		return;
5646 	/*
5647 	 * Is the spinlock portion underflowing?
5648 	 */
5649 	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5650 			!(preempt_count() & PREEMPT_MASK)))
5651 		return;
5652 #endif
5653 
5654 	preempt_latency_stop(val);
5655 	__preempt_count_sub(val);
5656 }
5657 EXPORT_SYMBOL(preempt_count_sub);
5658 NOKPROBE_SYMBOL(preempt_count_sub);
5659 
5660 #else
preempt_latency_start(int val)5661 static inline void preempt_latency_start(int val) { }
preempt_latency_stop(int val)5662 static inline void preempt_latency_stop(int val) { }
5663 #endif
5664 
get_preempt_disable_ip(struct task_struct * p)5665 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5666 {
5667 #ifdef CONFIG_DEBUG_PREEMPT
5668 	return p->preempt_disable_ip;
5669 #else
5670 	return 0;
5671 #endif
5672 }
5673 
5674 /*
5675  * Print scheduling while atomic bug:
5676  */
__schedule_bug(struct task_struct * prev)5677 static noinline void __schedule_bug(struct task_struct *prev)
5678 {
5679 	/* Save this before calling printk(), since that will clobber it */
5680 	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5681 
5682 	if (oops_in_progress)
5683 		return;
5684 
5685 	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5686 		prev->comm, prev->pid, preempt_count());
5687 
5688 	debug_show_held_locks(prev);
5689 	print_modules();
5690 	if (irqs_disabled())
5691 		print_irqtrace_events(prev);
5692 	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5693 	    && in_atomic_preempt_off()) {
5694 		pr_err("Preemption disabled at:");
5695 		print_ip_sym(KERN_ERR, preempt_disable_ip);
5696 	}
5697 	if (panic_on_warn)
5698 		panic("scheduling while atomic\n");
5699 
5700 	dump_stack();
5701 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5702 }
5703 
5704 /*
5705  * Various schedule()-time debugging checks and statistics:
5706  */
schedule_debug(struct task_struct * prev,bool preempt)5707 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5708 {
5709 #ifdef CONFIG_SCHED_STACK_END_CHECK
5710 	if (task_stack_end_corrupted(prev))
5711 		panic("corrupted stack end detected inside scheduler\n");
5712 
5713 	if (task_scs_end_corrupted(prev))
5714 		panic("corrupted shadow stack detected inside scheduler\n");
5715 #endif
5716 
5717 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5718 	if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5719 		printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5720 			prev->comm, prev->pid, prev->non_block_count);
5721 		dump_stack();
5722 		add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5723 	}
5724 #endif
5725 
5726 	if (unlikely(in_atomic_preempt_off())) {
5727 		__schedule_bug(prev);
5728 		preempt_count_set(PREEMPT_DISABLED);
5729 	}
5730 	rcu_sleep_check();
5731 	SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5732 
5733 	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5734 
5735 	schedstat_inc(this_rq()->sched_count);
5736 }
5737 
put_prev_task_balance(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)5738 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5739 				  struct rq_flags *rf)
5740 {
5741 #ifdef CONFIG_SMP
5742 	const struct sched_class *class;
5743 	/*
5744 	 * We must do the balancing pass before put_prev_task(), such
5745 	 * that when we release the rq->lock the task is in the same
5746 	 * state as before we took rq->lock.
5747 	 *
5748 	 * We can terminate the balance pass as soon as we know there is
5749 	 * a runnable task of @class priority or higher.
5750 	 */
5751 	for_class_range(class, prev->sched_class, &idle_sched_class) {
5752 		if (class->balance(rq, prev, rf))
5753 			break;
5754 	}
5755 #endif
5756 
5757 	put_prev_task(rq, prev);
5758 }
5759 
5760 /*
5761  * Pick up the highest-prio task:
5762  */
5763 static inline struct task_struct *
__pick_next_task(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)5764 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5765 {
5766 	const struct sched_class *class;
5767 	struct task_struct *p;
5768 
5769 	/*
5770 	 * Optimization: we know that if all tasks are in the fair class we can
5771 	 * call that function directly, but only if the @prev task wasn't of a
5772 	 * higher scheduling class, because otherwise those lose the
5773 	 * opportunity to pull in more work from other CPUs.
5774 	 */
5775 	if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
5776 		   rq->nr_running == rq->cfs.h_nr_running)) {
5777 
5778 		p = pick_next_task_fair(rq, prev, rf);
5779 		if (unlikely(p == RETRY_TASK))
5780 			goto restart;
5781 
5782 		/* Assume the next prioritized class is idle_sched_class */
5783 		if (!p) {
5784 			put_prev_task(rq, prev);
5785 			p = pick_next_task_idle(rq);
5786 		}
5787 
5788 		return p;
5789 	}
5790 
5791 restart:
5792 	put_prev_task_balance(rq, prev, rf);
5793 
5794 	for_each_class(class) {
5795 		p = class->pick_next_task(rq);
5796 		if (p)
5797 			return p;
5798 	}
5799 
5800 	BUG(); /* The idle class should always have a runnable task. */
5801 }
5802 
5803 #ifdef CONFIG_SCHED_CORE
is_task_rq_idle(struct task_struct * t)5804 static inline bool is_task_rq_idle(struct task_struct *t)
5805 {
5806 	return (task_rq(t)->idle == t);
5807 }
5808 
cookie_equals(struct task_struct * a,unsigned long cookie)5809 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5810 {
5811 	return is_task_rq_idle(a) || (a->core_cookie == cookie);
5812 }
5813 
cookie_match(struct task_struct * a,struct task_struct * b)5814 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5815 {
5816 	if (is_task_rq_idle(a) || is_task_rq_idle(b))
5817 		return true;
5818 
5819 	return a->core_cookie == b->core_cookie;
5820 }
5821 
pick_task(struct rq * rq)5822 static inline struct task_struct *pick_task(struct rq *rq)
5823 {
5824 	const struct sched_class *class;
5825 	struct task_struct *p;
5826 
5827 	for_each_class(class) {
5828 		p = class->pick_task(rq);
5829 		if (p)
5830 			return p;
5831 	}
5832 
5833 	BUG(); /* The idle class should always have a runnable task. */
5834 }
5835 
5836 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5837 
5838 static void queue_core_balance(struct rq *rq);
5839 
5840 static struct task_struct *
pick_next_task(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)5841 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5842 {
5843 	struct task_struct *next, *p, *max = NULL;
5844 	const struct cpumask *smt_mask;
5845 	bool fi_before = false;
5846 	bool core_clock_updated = (rq == rq->core);
5847 	unsigned long cookie;
5848 	int i, cpu, occ = 0;
5849 	struct rq *rq_i;
5850 	bool need_sync;
5851 
5852 	if (!sched_core_enabled(rq))
5853 		return __pick_next_task(rq, prev, rf);
5854 
5855 	cpu = cpu_of(rq);
5856 
5857 	/* Stopper task is switching into idle, no need core-wide selection. */
5858 	if (cpu_is_offline(cpu)) {
5859 		/*
5860 		 * Reset core_pick so that we don't enter the fastpath when
5861 		 * coming online. core_pick would already be migrated to
5862 		 * another cpu during offline.
5863 		 */
5864 		rq->core_pick = NULL;
5865 		return __pick_next_task(rq, prev, rf);
5866 	}
5867 
5868 	/*
5869 	 * If there were no {en,de}queues since we picked (IOW, the task
5870 	 * pointers are all still valid), and we haven't scheduled the last
5871 	 * pick yet, do so now.
5872 	 *
5873 	 * rq->core_pick can be NULL if no selection was made for a CPU because
5874 	 * it was either offline or went offline during a sibling's core-wide
5875 	 * selection. In this case, do a core-wide selection.
5876 	 */
5877 	if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5878 	    rq->core->core_pick_seq != rq->core_sched_seq &&
5879 	    rq->core_pick) {
5880 		WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5881 
5882 		next = rq->core_pick;
5883 		if (next != prev) {
5884 			put_prev_task(rq, prev);
5885 			set_next_task(rq, next);
5886 		}
5887 
5888 		rq->core_pick = NULL;
5889 		goto out;
5890 	}
5891 
5892 	put_prev_task_balance(rq, prev, rf);
5893 
5894 	smt_mask = cpu_smt_mask(cpu);
5895 	need_sync = !!rq->core->core_cookie;
5896 
5897 	/* reset state */
5898 	rq->core->core_cookie = 0UL;
5899 	if (rq->core->core_forceidle_count) {
5900 		if (!core_clock_updated) {
5901 			update_rq_clock(rq->core);
5902 			core_clock_updated = true;
5903 		}
5904 		sched_core_account_forceidle(rq);
5905 		/* reset after accounting force idle */
5906 		rq->core->core_forceidle_start = 0;
5907 		rq->core->core_forceidle_count = 0;
5908 		rq->core->core_forceidle_occupation = 0;
5909 		need_sync = true;
5910 		fi_before = true;
5911 	}
5912 
5913 	/*
5914 	 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5915 	 *
5916 	 * @task_seq guards the task state ({en,de}queues)
5917 	 * @pick_seq is the @task_seq we did a selection on
5918 	 * @sched_seq is the @pick_seq we scheduled
5919 	 *
5920 	 * However, preemptions can cause multiple picks on the same task set.
5921 	 * 'Fix' this by also increasing @task_seq for every pick.
5922 	 */
5923 	rq->core->core_task_seq++;
5924 
5925 	/*
5926 	 * Optimize for common case where this CPU has no cookies
5927 	 * and there are no cookied tasks running on siblings.
5928 	 */
5929 	if (!need_sync) {
5930 		next = pick_task(rq);
5931 		if (!next->core_cookie) {
5932 			rq->core_pick = NULL;
5933 			/*
5934 			 * For robustness, update the min_vruntime_fi for
5935 			 * unconstrained picks as well.
5936 			 */
5937 			WARN_ON_ONCE(fi_before);
5938 			task_vruntime_update(rq, next, false);
5939 			goto out_set_next;
5940 		}
5941 	}
5942 
5943 	/*
5944 	 * For each thread: do the regular task pick and find the max prio task
5945 	 * amongst them.
5946 	 *
5947 	 * Tie-break prio towards the current CPU
5948 	 */
5949 	for_each_cpu_wrap(i, smt_mask, cpu) {
5950 		rq_i = cpu_rq(i);
5951 
5952 		/*
5953 		 * Current cpu always has its clock updated on entrance to
5954 		 * pick_next_task(). If the current cpu is not the core,
5955 		 * the core may also have been updated above.
5956 		 */
5957 		if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5958 			update_rq_clock(rq_i);
5959 
5960 		p = rq_i->core_pick = pick_task(rq_i);
5961 		if (!max || prio_less(max, p, fi_before))
5962 			max = p;
5963 	}
5964 
5965 	cookie = rq->core->core_cookie = max->core_cookie;
5966 
5967 	/*
5968 	 * For each thread: try and find a runnable task that matches @max or
5969 	 * force idle.
5970 	 */
5971 	for_each_cpu(i, smt_mask) {
5972 		rq_i = cpu_rq(i);
5973 		p = rq_i->core_pick;
5974 
5975 		if (!cookie_equals(p, cookie)) {
5976 			p = NULL;
5977 			if (cookie)
5978 				p = sched_core_find(rq_i, cookie);
5979 			if (!p)
5980 				p = idle_sched_class.pick_task(rq_i);
5981 		}
5982 
5983 		rq_i->core_pick = p;
5984 
5985 		if (p == rq_i->idle) {
5986 			if (rq_i->nr_running) {
5987 				rq->core->core_forceidle_count++;
5988 				if (!fi_before)
5989 					rq->core->core_forceidle_seq++;
5990 			}
5991 		} else {
5992 			occ++;
5993 		}
5994 	}
5995 
5996 	if (schedstat_enabled() && rq->core->core_forceidle_count) {
5997 		rq->core->core_forceidle_start = rq_clock(rq->core);
5998 		rq->core->core_forceidle_occupation = occ;
5999 	}
6000 
6001 	rq->core->core_pick_seq = rq->core->core_task_seq;
6002 	next = rq->core_pick;
6003 	rq->core_sched_seq = rq->core->core_pick_seq;
6004 
6005 	/* Something should have been selected for current CPU */
6006 	WARN_ON_ONCE(!next);
6007 
6008 	/*
6009 	 * Reschedule siblings
6010 	 *
6011 	 * NOTE: L1TF -- at this point we're no longer running the old task and
6012 	 * sending an IPI (below) ensures the sibling will no longer be running
6013 	 * their task. This ensures there is no inter-sibling overlap between
6014 	 * non-matching user state.
6015 	 */
6016 	for_each_cpu(i, smt_mask) {
6017 		rq_i = cpu_rq(i);
6018 
6019 		/*
6020 		 * An online sibling might have gone offline before a task
6021 		 * could be picked for it, or it might be offline but later
6022 		 * happen to come online, but its too late and nothing was
6023 		 * picked for it.  That's Ok - it will pick tasks for itself,
6024 		 * so ignore it.
6025 		 */
6026 		if (!rq_i->core_pick)
6027 			continue;
6028 
6029 		/*
6030 		 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6031 		 * fi_before     fi      update?
6032 		 *  0            0       1
6033 		 *  0            1       1
6034 		 *  1            0       1
6035 		 *  1            1       0
6036 		 */
6037 		if (!(fi_before && rq->core->core_forceidle_count))
6038 			task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6039 
6040 		rq_i->core_pick->core_occupation = occ;
6041 
6042 		if (i == cpu) {
6043 			rq_i->core_pick = NULL;
6044 			continue;
6045 		}
6046 
6047 		/* Did we break L1TF mitigation requirements? */
6048 		WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6049 
6050 		if (rq_i->curr == rq_i->core_pick) {
6051 			rq_i->core_pick = NULL;
6052 			continue;
6053 		}
6054 
6055 		resched_curr(rq_i);
6056 	}
6057 
6058 out_set_next:
6059 	set_next_task(rq, next);
6060 out:
6061 	if (rq->core->core_forceidle_count && next == rq->idle)
6062 		queue_core_balance(rq);
6063 
6064 	return next;
6065 }
6066 
try_steal_cookie(int this,int that)6067 static bool try_steal_cookie(int this, int that)
6068 {
6069 	struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6070 	struct task_struct *p;
6071 	unsigned long cookie;
6072 	bool success = false;
6073 
6074 	local_irq_disable();
6075 	double_rq_lock(dst, src);
6076 
6077 	cookie = dst->core->core_cookie;
6078 	if (!cookie)
6079 		goto unlock;
6080 
6081 	if (dst->curr != dst->idle)
6082 		goto unlock;
6083 
6084 	p = sched_core_find(src, cookie);
6085 	if (p == src->idle)
6086 		goto unlock;
6087 
6088 	do {
6089 		if (p == src->core_pick || p == src->curr)
6090 			goto next;
6091 
6092 		if (!is_cpu_allowed(p, this))
6093 			goto next;
6094 
6095 		if (p->core_occupation > dst->idle->core_occupation)
6096 			goto next;
6097 
6098 		deactivate_task(src, p, 0);
6099 		set_task_cpu(p, this);
6100 		activate_task(dst, p, 0);
6101 
6102 		resched_curr(dst);
6103 
6104 		success = true;
6105 		break;
6106 
6107 next:
6108 		p = sched_core_next(p, cookie);
6109 	} while (p);
6110 
6111 unlock:
6112 	double_rq_unlock(dst, src);
6113 	local_irq_enable();
6114 
6115 	return success;
6116 }
6117 
steal_cookie_task(int cpu,struct sched_domain * sd)6118 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6119 {
6120 	int i;
6121 
6122 	for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
6123 		if (i == cpu)
6124 			continue;
6125 
6126 		if (need_resched())
6127 			break;
6128 
6129 		if (try_steal_cookie(cpu, i))
6130 			return true;
6131 	}
6132 
6133 	return false;
6134 }
6135 
sched_core_balance(struct rq * rq)6136 static void sched_core_balance(struct rq *rq)
6137 {
6138 	struct sched_domain *sd;
6139 	int cpu = cpu_of(rq);
6140 
6141 	preempt_disable();
6142 	rcu_read_lock();
6143 	raw_spin_rq_unlock_irq(rq);
6144 	for_each_domain(cpu, sd) {
6145 		if (need_resched())
6146 			break;
6147 
6148 		if (steal_cookie_task(cpu, sd))
6149 			break;
6150 	}
6151 	raw_spin_rq_lock_irq(rq);
6152 	rcu_read_unlock();
6153 	preempt_enable();
6154 }
6155 
6156 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
6157 
queue_core_balance(struct rq * rq)6158 static void queue_core_balance(struct rq *rq)
6159 {
6160 	if (!sched_core_enabled(rq))
6161 		return;
6162 
6163 	if (!rq->core->core_cookie)
6164 		return;
6165 
6166 	if (!rq->nr_running) /* not forced idle */
6167 		return;
6168 
6169 	queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6170 }
6171 
sched_core_cpu_starting(unsigned int cpu)6172 static void sched_core_cpu_starting(unsigned int cpu)
6173 {
6174 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6175 	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6176 	unsigned long flags;
6177 	int t;
6178 
6179 	sched_core_lock(cpu, &flags);
6180 
6181 	WARN_ON_ONCE(rq->core != rq);
6182 
6183 	/* if we're the first, we'll be our own leader */
6184 	if (cpumask_weight(smt_mask) == 1)
6185 		goto unlock;
6186 
6187 	/* find the leader */
6188 	for_each_cpu(t, smt_mask) {
6189 		if (t == cpu)
6190 			continue;
6191 		rq = cpu_rq(t);
6192 		if (rq->core == rq) {
6193 			core_rq = rq;
6194 			break;
6195 		}
6196 	}
6197 
6198 	if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6199 		goto unlock;
6200 
6201 	/* install and validate core_rq */
6202 	for_each_cpu(t, smt_mask) {
6203 		rq = cpu_rq(t);
6204 
6205 		if (t == cpu)
6206 			rq->core = core_rq;
6207 
6208 		WARN_ON_ONCE(rq->core != core_rq);
6209 	}
6210 
6211 unlock:
6212 	sched_core_unlock(cpu, &flags);
6213 }
6214 
sched_core_cpu_deactivate(unsigned int cpu)6215 static void sched_core_cpu_deactivate(unsigned int cpu)
6216 {
6217 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6218 	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6219 	unsigned long flags;
6220 	int t;
6221 
6222 	sched_core_lock(cpu, &flags);
6223 
6224 	/* if we're the last man standing, nothing to do */
6225 	if (cpumask_weight(smt_mask) == 1) {
6226 		WARN_ON_ONCE(rq->core != rq);
6227 		goto unlock;
6228 	}
6229 
6230 	/* if we're not the leader, nothing to do */
6231 	if (rq->core != rq)
6232 		goto unlock;
6233 
6234 	/* find a new leader */
6235 	for_each_cpu(t, smt_mask) {
6236 		if (t == cpu)
6237 			continue;
6238 		core_rq = cpu_rq(t);
6239 		break;
6240 	}
6241 
6242 	if (WARN_ON_ONCE(!core_rq)) /* impossible */
6243 		goto unlock;
6244 
6245 	/* copy the shared state to the new leader */
6246 	core_rq->core_task_seq             = rq->core_task_seq;
6247 	core_rq->core_pick_seq             = rq->core_pick_seq;
6248 	core_rq->core_cookie               = rq->core_cookie;
6249 	core_rq->core_forceidle_count      = rq->core_forceidle_count;
6250 	core_rq->core_forceidle_seq        = rq->core_forceidle_seq;
6251 	core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6252 
6253 	/*
6254 	 * Accounting edge for forced idle is handled in pick_next_task().
6255 	 * Don't need another one here, since the hotplug thread shouldn't
6256 	 * have a cookie.
6257 	 */
6258 	core_rq->core_forceidle_start = 0;
6259 
6260 	/* install new leader */
6261 	for_each_cpu(t, smt_mask) {
6262 		rq = cpu_rq(t);
6263 		rq->core = core_rq;
6264 	}
6265 
6266 unlock:
6267 	sched_core_unlock(cpu, &flags);
6268 }
6269 
sched_core_cpu_dying(unsigned int cpu)6270 static inline void sched_core_cpu_dying(unsigned int cpu)
6271 {
6272 	struct rq *rq = cpu_rq(cpu);
6273 
6274 	if (rq->core != rq)
6275 		rq->core = rq;
6276 }
6277 
6278 #else /* !CONFIG_SCHED_CORE */
6279 
sched_core_cpu_starting(unsigned int cpu)6280 static inline void sched_core_cpu_starting(unsigned int cpu) {}
sched_core_cpu_deactivate(unsigned int cpu)6281 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
sched_core_cpu_dying(unsigned int cpu)6282 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6283 
6284 static struct task_struct *
pick_next_task(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)6285 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6286 {
6287 	return __pick_next_task(rq, prev, rf);
6288 }
6289 
6290 #endif /* CONFIG_SCHED_CORE */
6291 
6292 /*
6293  * Constants for the sched_mode argument of __schedule().
6294  *
6295  * The mode argument allows RT enabled kernels to differentiate a
6296  * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6297  * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6298  * optimize the AND operation out and just check for zero.
6299  */
6300 #define SM_NONE			0x0
6301 #define SM_PREEMPT		0x1
6302 #define SM_RTLOCK_WAIT		0x2
6303 
6304 #ifndef CONFIG_PREEMPT_RT
6305 # define SM_MASK_PREEMPT	(~0U)
6306 #else
6307 # define SM_MASK_PREEMPT	SM_PREEMPT
6308 #endif
6309 
6310 /*
6311  * __schedule() is the main scheduler function.
6312  *
6313  * The main means of driving the scheduler and thus entering this function are:
6314  *
6315  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6316  *
6317  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6318  *      paths. For example, see arch/x86/entry_64.S.
6319  *
6320  *      To drive preemption between tasks, the scheduler sets the flag in timer
6321  *      interrupt handler scheduler_tick().
6322  *
6323  *   3. Wakeups don't really cause entry into schedule(). They add a
6324  *      task to the run-queue and that's it.
6325  *
6326  *      Now, if the new task added to the run-queue preempts the current
6327  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6328  *      called on the nearest possible occasion:
6329  *
6330  *       - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6331  *
6332  *         - in syscall or exception context, at the next outmost
6333  *           preempt_enable(). (this might be as soon as the wake_up()'s
6334  *           spin_unlock()!)
6335  *
6336  *         - in IRQ context, return from interrupt-handler to
6337  *           preemptible context
6338  *
6339  *       - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6340  *         then at the next:
6341  *
6342  *          - cond_resched() call
6343  *          - explicit schedule() call
6344  *          - return from syscall or exception to user-space
6345  *          - return from interrupt-handler to user-space
6346  *
6347  * WARNING: must be called with preemption disabled!
6348  */
__schedule(unsigned int sched_mode)6349 static void __sched notrace __schedule(unsigned int sched_mode)
6350 {
6351 	struct task_struct *prev, *next;
6352 	unsigned long *switch_count;
6353 	unsigned long prev_state;
6354 	struct rq_flags rf;
6355 	struct rq *rq;
6356 	int cpu;
6357 
6358 	cpu = smp_processor_id();
6359 	rq = cpu_rq(cpu);
6360 	prev = rq->curr;
6361 
6362 	schedule_debug(prev, !!sched_mode);
6363 
6364 	if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6365 		hrtick_clear(rq);
6366 
6367 	local_irq_disable();
6368 	rcu_note_context_switch(!!sched_mode);
6369 
6370 	/*
6371 	 * Make sure that signal_pending_state()->signal_pending() below
6372 	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6373 	 * done by the caller to avoid the race with signal_wake_up():
6374 	 *
6375 	 * __set_current_state(@state)		signal_wake_up()
6376 	 * schedule()				  set_tsk_thread_flag(p, TIF_SIGPENDING)
6377 	 *					  wake_up_state(p, state)
6378 	 *   LOCK rq->lock			    LOCK p->pi_state
6379 	 *   smp_mb__after_spinlock()		    smp_mb__after_spinlock()
6380 	 *     if (signal_pending_state())	    if (p->state & @state)
6381 	 *
6382 	 * Also, the membarrier system call requires a full memory barrier
6383 	 * after coming from user-space, before storing to rq->curr.
6384 	 */
6385 	rq_lock(rq, &rf);
6386 	smp_mb__after_spinlock();
6387 
6388 	/* Promote REQ to ACT */
6389 	rq->clock_update_flags <<= 1;
6390 	update_rq_clock(rq);
6391 
6392 	switch_count = &prev->nivcsw;
6393 
6394 	/*
6395 	 * We must load prev->state once (task_struct::state is volatile), such
6396 	 * that we form a control dependency vs deactivate_task() below.
6397 	 */
6398 	prev_state = READ_ONCE(prev->__state);
6399 	if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6400 		if (signal_pending_state(prev_state, prev)) {
6401 			WRITE_ONCE(prev->__state, TASK_RUNNING);
6402 		} else {
6403 			prev->sched_contributes_to_load =
6404 				(prev_state & TASK_UNINTERRUPTIBLE) &&
6405 				!(prev_state & TASK_NOLOAD) &&
6406 				!(prev->flags & PF_FROZEN);
6407 
6408 			if (prev->sched_contributes_to_load)
6409 				rq->nr_uninterruptible++;
6410 
6411 			/*
6412 			 * __schedule()			ttwu()
6413 			 *   prev_state = prev->state;    if (p->on_rq && ...)
6414 			 *   if (prev_state)		    goto out;
6415 			 *     p->on_rq = 0;		  smp_acquire__after_ctrl_dep();
6416 			 *				  p->state = TASK_WAKING
6417 			 *
6418 			 * Where __schedule() and ttwu() have matching control dependencies.
6419 			 *
6420 			 * After this, schedule() must not care about p->state any more.
6421 			 */
6422 			deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6423 
6424 			if (prev->in_iowait) {
6425 				atomic_inc(&rq->nr_iowait);
6426 				delayacct_blkio_start();
6427 			}
6428 		}
6429 		switch_count = &prev->nvcsw;
6430 	}
6431 
6432 	next = pick_next_task(rq, prev, &rf);
6433 	clear_tsk_need_resched(prev);
6434 	clear_preempt_need_resched();
6435 #ifdef CONFIG_SCHED_DEBUG
6436 	rq->last_seen_need_resched_ns = 0;
6437 #endif
6438 
6439 	if (likely(prev != next)) {
6440 		rq->nr_switches++;
6441 		/*
6442 		 * RCU users of rcu_dereference(rq->curr) may not see
6443 		 * changes to task_struct made by pick_next_task().
6444 		 */
6445 		RCU_INIT_POINTER(rq->curr, next);
6446 		/*
6447 		 * The membarrier system call requires each architecture
6448 		 * to have a full memory barrier after updating
6449 		 * rq->curr, before returning to user-space.
6450 		 *
6451 		 * Here are the schemes providing that barrier on the
6452 		 * various architectures:
6453 		 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6454 		 *   switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6455 		 * - finish_lock_switch() for weakly-ordered
6456 		 *   architectures where spin_unlock is a full barrier,
6457 		 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6458 		 *   is a RELEASE barrier),
6459 		 */
6460 		++*switch_count;
6461 
6462 		migrate_disable_switch(rq, prev);
6463 		psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6464 
6465 		trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6466 
6467 		/* Also unlocks the rq: */
6468 		rq = context_switch(rq, prev, next, &rf);
6469 	} else {
6470 		rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6471 
6472 		rq_unpin_lock(rq, &rf);
6473 		__balance_callbacks(rq);
6474 		raw_spin_rq_unlock_irq(rq);
6475 	}
6476 }
6477 
do_task_dead(void)6478 void __noreturn do_task_dead(void)
6479 {
6480 	/* Causes final put_task_struct in finish_task_switch(): */
6481 	set_special_state(TASK_DEAD);
6482 
6483 	/* Tell freezer to ignore us: */
6484 	current->flags |= PF_NOFREEZE;
6485 
6486 	__schedule(SM_NONE);
6487 	BUG();
6488 
6489 	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6490 	for (;;)
6491 		cpu_relax();
6492 }
6493 
sched_submit_work(struct task_struct * tsk)6494 static inline void sched_submit_work(struct task_struct *tsk)
6495 {
6496 	unsigned int task_flags;
6497 
6498 	if (task_is_running(tsk))
6499 		return;
6500 
6501 	task_flags = tsk->flags;
6502 	/*
6503 	 * If a worker goes to sleep, notify and ask workqueue whether it
6504 	 * wants to wake up a task to maintain concurrency.
6505 	 */
6506 	if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6507 		if (task_flags & PF_WQ_WORKER)
6508 			wq_worker_sleeping(tsk);
6509 		else
6510 			io_wq_worker_sleeping(tsk);
6511 	}
6512 
6513 	/*
6514 	 * spinlock and rwlock must not flush block requests.  This will
6515 	 * deadlock if the callback attempts to acquire a lock which is
6516 	 * already acquired.
6517 	 */
6518 	SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6519 
6520 	/*
6521 	 * If we are going to sleep and we have plugged IO queued,
6522 	 * make sure to submit it to avoid deadlocks.
6523 	 */
6524 	blk_flush_plug(tsk->plug, true);
6525 }
6526 
sched_update_worker(struct task_struct * tsk)6527 static void sched_update_worker(struct task_struct *tsk)
6528 {
6529 	if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6530 		if (tsk->flags & PF_WQ_WORKER)
6531 			wq_worker_running(tsk);
6532 		else
6533 			io_wq_worker_running(tsk);
6534 	}
6535 }
6536 
schedule(void)6537 asmlinkage __visible void __sched schedule(void)
6538 {
6539 	struct task_struct *tsk = current;
6540 
6541 	sched_submit_work(tsk);
6542 	do {
6543 		preempt_disable();
6544 		__schedule(SM_NONE);
6545 		sched_preempt_enable_no_resched();
6546 	} while (need_resched());
6547 	sched_update_worker(tsk);
6548 }
6549 EXPORT_SYMBOL(schedule);
6550 
6551 /*
6552  * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6553  * state (have scheduled out non-voluntarily) by making sure that all
6554  * tasks have either left the run queue or have gone into user space.
6555  * As idle tasks do not do either, they must not ever be preempted
6556  * (schedule out non-voluntarily).
6557  *
6558  * schedule_idle() is similar to schedule_preempt_disable() except that it
6559  * never enables preemption because it does not call sched_submit_work().
6560  */
schedule_idle(void)6561 void __sched schedule_idle(void)
6562 {
6563 	/*
6564 	 * As this skips calling sched_submit_work(), which the idle task does
6565 	 * regardless because that function is a nop when the task is in a
6566 	 * TASK_RUNNING state, make sure this isn't used someplace that the
6567 	 * current task can be in any other state. Note, idle is always in the
6568 	 * TASK_RUNNING state.
6569 	 */
6570 	WARN_ON_ONCE(current->__state);
6571 	do {
6572 		__schedule(SM_NONE);
6573 	} while (need_resched());
6574 }
6575 
6576 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
schedule_user(void)6577 asmlinkage __visible void __sched schedule_user(void)
6578 {
6579 	/*
6580 	 * If we come here after a random call to set_need_resched(),
6581 	 * or we have been woken up remotely but the IPI has not yet arrived,
6582 	 * we haven't yet exited the RCU idle mode. Do it here manually until
6583 	 * we find a better solution.
6584 	 *
6585 	 * NB: There are buggy callers of this function.  Ideally we
6586 	 * should warn if prev_state != CONTEXT_USER, but that will trigger
6587 	 * too frequently to make sense yet.
6588 	 */
6589 	enum ctx_state prev_state = exception_enter();
6590 	schedule();
6591 	exception_exit(prev_state);
6592 }
6593 #endif
6594 
6595 /**
6596  * schedule_preempt_disabled - called with preemption disabled
6597  *
6598  * Returns with preemption disabled. Note: preempt_count must be 1
6599  */
schedule_preempt_disabled(void)6600 void __sched schedule_preempt_disabled(void)
6601 {
6602 	sched_preempt_enable_no_resched();
6603 	schedule();
6604 	preempt_disable();
6605 }
6606 
6607 #ifdef CONFIG_PREEMPT_RT
schedule_rtlock(void)6608 void __sched notrace schedule_rtlock(void)
6609 {
6610 	do {
6611 		preempt_disable();
6612 		__schedule(SM_RTLOCK_WAIT);
6613 		sched_preempt_enable_no_resched();
6614 	} while (need_resched());
6615 }
6616 NOKPROBE_SYMBOL(schedule_rtlock);
6617 #endif
6618 
preempt_schedule_common(void)6619 static void __sched notrace preempt_schedule_common(void)
6620 {
6621 	do {
6622 		/*
6623 		 * Because the function tracer can trace preempt_count_sub()
6624 		 * and it also uses preempt_enable/disable_notrace(), if
6625 		 * NEED_RESCHED is set, the preempt_enable_notrace() called
6626 		 * by the function tracer will call this function again and
6627 		 * cause infinite recursion.
6628 		 *
6629 		 * Preemption must be disabled here before the function
6630 		 * tracer can trace. Break up preempt_disable() into two
6631 		 * calls. One to disable preemption without fear of being
6632 		 * traced. The other to still record the preemption latency,
6633 		 * which can also be traced by the function tracer.
6634 		 */
6635 		preempt_disable_notrace();
6636 		preempt_latency_start(1);
6637 		__schedule(SM_PREEMPT);
6638 		preempt_latency_stop(1);
6639 		preempt_enable_no_resched_notrace();
6640 
6641 		/*
6642 		 * Check again in case we missed a preemption opportunity
6643 		 * between schedule and now.
6644 		 */
6645 	} while (need_resched());
6646 }
6647 
6648 #ifdef CONFIG_PREEMPTION
6649 /*
6650  * This is the entry point to schedule() from in-kernel preemption
6651  * off of preempt_enable.
6652  */
preempt_schedule(void)6653 asmlinkage __visible void __sched notrace preempt_schedule(void)
6654 {
6655 	/*
6656 	 * If there is a non-zero preempt_count or interrupts are disabled,
6657 	 * we do not want to preempt the current task. Just return..
6658 	 */
6659 	if (likely(!preemptible()))
6660 		return;
6661 	preempt_schedule_common();
6662 }
6663 NOKPROBE_SYMBOL(preempt_schedule);
6664 EXPORT_SYMBOL(preempt_schedule);
6665 
6666 #ifdef CONFIG_PREEMPT_DYNAMIC
6667 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6668 #ifndef preempt_schedule_dynamic_enabled
6669 #define preempt_schedule_dynamic_enabled	preempt_schedule
6670 #define preempt_schedule_dynamic_disabled	NULL
6671 #endif
6672 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6673 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6674 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6675 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
dynamic_preempt_schedule(void)6676 void __sched notrace dynamic_preempt_schedule(void)
6677 {
6678 	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6679 		return;
6680 	preempt_schedule();
6681 }
6682 NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6683 EXPORT_SYMBOL(dynamic_preempt_schedule);
6684 #endif
6685 #endif
6686 
6687 /**
6688  * preempt_schedule_notrace - preempt_schedule called by tracing
6689  *
6690  * The tracing infrastructure uses preempt_enable_notrace to prevent
6691  * recursion and tracing preempt enabling caused by the tracing
6692  * infrastructure itself. But as tracing can happen in areas coming
6693  * from userspace or just about to enter userspace, a preempt enable
6694  * can occur before user_exit() is called. This will cause the scheduler
6695  * to be called when the system is still in usermode.
6696  *
6697  * To prevent this, the preempt_enable_notrace will use this function
6698  * instead of preempt_schedule() to exit user context if needed before
6699  * calling the scheduler.
6700  */
preempt_schedule_notrace(void)6701 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6702 {
6703 	enum ctx_state prev_ctx;
6704 
6705 	if (likely(!preemptible()))
6706 		return;
6707 
6708 	do {
6709 		/*
6710 		 * Because the function tracer can trace preempt_count_sub()
6711 		 * and it also uses preempt_enable/disable_notrace(), if
6712 		 * NEED_RESCHED is set, the preempt_enable_notrace() called
6713 		 * by the function tracer will call this function again and
6714 		 * cause infinite recursion.
6715 		 *
6716 		 * Preemption must be disabled here before the function
6717 		 * tracer can trace. Break up preempt_disable() into two
6718 		 * calls. One to disable preemption without fear of being
6719 		 * traced. The other to still record the preemption latency,
6720 		 * which can also be traced by the function tracer.
6721 		 */
6722 		preempt_disable_notrace();
6723 		preempt_latency_start(1);
6724 		/*
6725 		 * Needs preempt disabled in case user_exit() is traced
6726 		 * and the tracer calls preempt_enable_notrace() causing
6727 		 * an infinite recursion.
6728 		 */
6729 		prev_ctx = exception_enter();
6730 		__schedule(SM_PREEMPT);
6731 		exception_exit(prev_ctx);
6732 
6733 		preempt_latency_stop(1);
6734 		preempt_enable_no_resched_notrace();
6735 	} while (need_resched());
6736 }
6737 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6738 
6739 #ifdef CONFIG_PREEMPT_DYNAMIC
6740 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6741 #ifndef preempt_schedule_notrace_dynamic_enabled
6742 #define preempt_schedule_notrace_dynamic_enabled	preempt_schedule_notrace
6743 #define preempt_schedule_notrace_dynamic_disabled	NULL
6744 #endif
6745 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6746 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6747 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6748 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
dynamic_preempt_schedule_notrace(void)6749 void __sched notrace dynamic_preempt_schedule_notrace(void)
6750 {
6751 	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6752 		return;
6753 	preempt_schedule_notrace();
6754 }
6755 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6756 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6757 #endif
6758 #endif
6759 
6760 #endif /* CONFIG_PREEMPTION */
6761 
6762 /*
6763  * This is the entry point to schedule() from kernel preemption
6764  * off of irq context.
6765  * Note, that this is called and return with irqs disabled. This will
6766  * protect us against recursive calling from irq.
6767  */
preempt_schedule_irq(void)6768 asmlinkage __visible void __sched preempt_schedule_irq(void)
6769 {
6770 	enum ctx_state prev_state;
6771 
6772 	/* Catch callers which need to be fixed */
6773 	BUG_ON(preempt_count() || !irqs_disabled());
6774 
6775 	prev_state = exception_enter();
6776 
6777 	do {
6778 		preempt_disable();
6779 		local_irq_enable();
6780 		__schedule(SM_PREEMPT);
6781 		local_irq_disable();
6782 		sched_preempt_enable_no_resched();
6783 	} while (need_resched());
6784 
6785 	exception_exit(prev_state);
6786 }
6787 
default_wake_function(wait_queue_entry_t * curr,unsigned mode,int wake_flags,void * key)6788 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6789 			  void *key)
6790 {
6791 	WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6792 	return try_to_wake_up(curr->private, mode, wake_flags);
6793 }
6794 EXPORT_SYMBOL(default_wake_function);
6795 
__setscheduler_prio(struct task_struct * p,int prio)6796 static void __setscheduler_prio(struct task_struct *p, int prio)
6797 {
6798 	if (dl_prio(prio))
6799 		p->sched_class = &dl_sched_class;
6800 	else if (rt_prio(prio))
6801 		p->sched_class = &rt_sched_class;
6802 	else
6803 		p->sched_class = &fair_sched_class;
6804 
6805 	p->prio = prio;
6806 }
6807 
6808 #ifdef CONFIG_RT_MUTEXES
6809 
__rt_effective_prio(struct task_struct * pi_task,int prio)6810 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6811 {
6812 	if (pi_task)
6813 		prio = min(prio, pi_task->prio);
6814 
6815 	return prio;
6816 }
6817 
rt_effective_prio(struct task_struct * p,int prio)6818 static inline int rt_effective_prio(struct task_struct *p, int prio)
6819 {
6820 	struct task_struct *pi_task = rt_mutex_get_top_task(p);
6821 
6822 	return __rt_effective_prio(pi_task, prio);
6823 }
6824 
6825 /*
6826  * rt_mutex_setprio - set the current priority of a task
6827  * @p: task to boost
6828  * @pi_task: donor task
6829  *
6830  * This function changes the 'effective' priority of a task. It does
6831  * not touch ->normal_prio like __setscheduler().
6832  *
6833  * Used by the rt_mutex code to implement priority inheritance
6834  * logic. Call site only calls if the priority of the task changed.
6835  */
rt_mutex_setprio(struct task_struct * p,struct task_struct * pi_task)6836 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6837 {
6838 	int prio, oldprio, queued, running, queue_flag =
6839 		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6840 	const struct sched_class *prev_class;
6841 	struct rq_flags rf;
6842 	struct rq *rq;
6843 
6844 	/* XXX used to be waiter->prio, not waiter->task->prio */
6845 	prio = __rt_effective_prio(pi_task, p->normal_prio);
6846 
6847 	/*
6848 	 * If nothing changed; bail early.
6849 	 */
6850 	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6851 		return;
6852 
6853 	rq = __task_rq_lock(p, &rf);
6854 	update_rq_clock(rq);
6855 	/*
6856 	 * Set under pi_lock && rq->lock, such that the value can be used under
6857 	 * either lock.
6858 	 *
6859 	 * Note that there is loads of tricky to make this pointer cache work
6860 	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6861 	 * ensure a task is de-boosted (pi_task is set to NULL) before the
6862 	 * task is allowed to run again (and can exit). This ensures the pointer
6863 	 * points to a blocked task -- which guarantees the task is present.
6864 	 */
6865 	p->pi_top_task = pi_task;
6866 
6867 	/*
6868 	 * For FIFO/RR we only need to set prio, if that matches we're done.
6869 	 */
6870 	if (prio == p->prio && !dl_prio(prio))
6871 		goto out_unlock;
6872 
6873 	/*
6874 	 * Idle task boosting is a nono in general. There is one
6875 	 * exception, when PREEMPT_RT and NOHZ is active:
6876 	 *
6877 	 * The idle task calls get_next_timer_interrupt() and holds
6878 	 * the timer wheel base->lock on the CPU and another CPU wants
6879 	 * to access the timer (probably to cancel it). We can safely
6880 	 * ignore the boosting request, as the idle CPU runs this code
6881 	 * with interrupts disabled and will complete the lock
6882 	 * protected section without being interrupted. So there is no
6883 	 * real need to boost.
6884 	 */
6885 	if (unlikely(p == rq->idle)) {
6886 		WARN_ON(p != rq->curr);
6887 		WARN_ON(p->pi_blocked_on);
6888 		goto out_unlock;
6889 	}
6890 
6891 	trace_sched_pi_setprio(p, pi_task);
6892 	oldprio = p->prio;
6893 
6894 	if (oldprio == prio)
6895 		queue_flag &= ~DEQUEUE_MOVE;
6896 
6897 	prev_class = p->sched_class;
6898 	queued = task_on_rq_queued(p);
6899 	running = task_current(rq, p);
6900 	if (queued)
6901 		dequeue_task(rq, p, queue_flag);
6902 	if (running)
6903 		put_prev_task(rq, p);
6904 
6905 	/*
6906 	 * Boosting condition are:
6907 	 * 1. -rt task is running and holds mutex A
6908 	 *      --> -dl task blocks on mutex A
6909 	 *
6910 	 * 2. -dl task is running and holds mutex A
6911 	 *      --> -dl task blocks on mutex A and could preempt the
6912 	 *          running task
6913 	 */
6914 	if (dl_prio(prio)) {
6915 		if (!dl_prio(p->normal_prio) ||
6916 		    (pi_task && dl_prio(pi_task->prio) &&
6917 		     dl_entity_preempt(&pi_task->dl, &p->dl))) {
6918 			p->dl.pi_se = pi_task->dl.pi_se;
6919 			queue_flag |= ENQUEUE_REPLENISH;
6920 		} else {
6921 			p->dl.pi_se = &p->dl;
6922 		}
6923 	} else if (rt_prio(prio)) {
6924 		if (dl_prio(oldprio))
6925 			p->dl.pi_se = &p->dl;
6926 		if (oldprio < prio)
6927 			queue_flag |= ENQUEUE_HEAD;
6928 	} else {
6929 		if (dl_prio(oldprio))
6930 			p->dl.pi_se = &p->dl;
6931 		if (rt_prio(oldprio))
6932 			p->rt.timeout = 0;
6933 	}
6934 
6935 	__setscheduler_prio(p, prio);
6936 
6937 	if (queued)
6938 		enqueue_task(rq, p, queue_flag);
6939 	if (running)
6940 		set_next_task(rq, p);
6941 
6942 	check_class_changed(rq, p, prev_class, oldprio);
6943 out_unlock:
6944 	/* Avoid rq from going away on us: */
6945 	preempt_disable();
6946 
6947 	rq_unpin_lock(rq, &rf);
6948 	__balance_callbacks(rq);
6949 	raw_spin_rq_unlock(rq);
6950 
6951 	preempt_enable();
6952 }
6953 #else
rt_effective_prio(struct task_struct * p,int prio)6954 static inline int rt_effective_prio(struct task_struct *p, int prio)
6955 {
6956 	return prio;
6957 }
6958 #endif
6959 
set_user_nice(struct task_struct * p,long nice)6960 void set_user_nice(struct task_struct *p, long nice)
6961 {
6962 	bool queued, running;
6963 	int old_prio;
6964 	struct rq_flags rf;
6965 	struct rq *rq;
6966 
6967 	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6968 		return;
6969 	/*
6970 	 * We have to be careful, if called from sys_setpriority(),
6971 	 * the task might be in the middle of scheduling on another CPU.
6972 	 */
6973 	rq = task_rq_lock(p, &rf);
6974 	update_rq_clock(rq);
6975 
6976 	/*
6977 	 * The RT priorities are set via sched_setscheduler(), but we still
6978 	 * allow the 'normal' nice value to be set - but as expected
6979 	 * it won't have any effect on scheduling until the task is
6980 	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6981 	 */
6982 	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6983 		p->static_prio = NICE_TO_PRIO(nice);
6984 		goto out_unlock;
6985 	}
6986 	queued = task_on_rq_queued(p);
6987 	running = task_current(rq, p);
6988 	if (queued)
6989 		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6990 	if (running)
6991 		put_prev_task(rq, p);
6992 
6993 	p->static_prio = NICE_TO_PRIO(nice);
6994 	set_load_weight(p, true);
6995 	old_prio = p->prio;
6996 	p->prio = effective_prio(p);
6997 
6998 	if (queued)
6999 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7000 	if (running)
7001 		set_next_task(rq, p);
7002 
7003 	/*
7004 	 * If the task increased its priority or is running and
7005 	 * lowered its priority, then reschedule its CPU:
7006 	 */
7007 	p->sched_class->prio_changed(rq, p, old_prio);
7008 
7009 out_unlock:
7010 	task_rq_unlock(rq, p, &rf);
7011 }
7012 EXPORT_SYMBOL(set_user_nice);
7013 
7014 /*
7015  * is_nice_reduction - check if nice value is an actual reduction
7016  *
7017  * Similar to can_nice() but does not perform a capability check.
7018  *
7019  * @p: task
7020  * @nice: nice value
7021  */
is_nice_reduction(const struct task_struct * p,const int nice)7022 static bool is_nice_reduction(const struct task_struct *p, const int nice)
7023 {
7024 	/* Convert nice value [19,-20] to rlimit style value [1,40]: */
7025 	int nice_rlim = nice_to_rlimit(nice);
7026 
7027 	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7028 }
7029 
7030 /*
7031  * can_nice - check if a task can reduce its nice value
7032  * @p: task
7033  * @nice: nice value
7034  */
can_nice(const struct task_struct * p,const int nice)7035 int can_nice(const struct task_struct *p, const int nice)
7036 {
7037 	return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7038 }
7039 
7040 #ifdef __ARCH_WANT_SYS_NICE
7041 
7042 /*
7043  * sys_nice - change the priority of the current process.
7044  * @increment: priority increment
7045  *
7046  * sys_setpriority is a more generic, but much slower function that
7047  * does similar things.
7048  */
SYSCALL_DEFINE1(nice,int,increment)7049 SYSCALL_DEFINE1(nice, int, increment)
7050 {
7051 	long nice, retval;
7052 
7053 	/*
7054 	 * Setpriority might change our priority at the same moment.
7055 	 * We don't have to worry. Conceptually one call occurs first
7056 	 * and we have a single winner.
7057 	 */
7058 	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7059 	nice = task_nice(current) + increment;
7060 
7061 	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7062 	if (increment < 0 && !can_nice(current, nice))
7063 		return -EPERM;
7064 
7065 	retval = security_task_setnice(current, nice);
7066 	if (retval)
7067 		return retval;
7068 
7069 	set_user_nice(current, nice);
7070 	return 0;
7071 }
7072 
7073 #endif
7074 
7075 /**
7076  * task_prio - return the priority value of a given task.
7077  * @p: the task in question.
7078  *
7079  * Return: The priority value as seen by users in /proc.
7080  *
7081  * sched policy         return value   kernel prio    user prio/nice
7082  *
7083  * normal, batch, idle     [0 ... 39]  [100 ... 139]          0/[-20 ... 19]
7084  * fifo, rr             [-2 ... -100]     [98 ... 0]  [1 ... 99]
7085  * deadline                     -101             -1           0
7086  */
task_prio(const struct task_struct * p)7087 int task_prio(const struct task_struct *p)
7088 {
7089 	return p->prio - MAX_RT_PRIO;
7090 }
7091 
7092 /**
7093  * idle_cpu - is a given CPU idle currently?
7094  * @cpu: the processor in question.
7095  *
7096  * Return: 1 if the CPU is currently idle. 0 otherwise.
7097  */
idle_cpu(int cpu)7098 int idle_cpu(int cpu)
7099 {
7100 	struct rq *rq = cpu_rq(cpu);
7101 
7102 	if (rq->curr != rq->idle)
7103 		return 0;
7104 
7105 	if (rq->nr_running)
7106 		return 0;
7107 
7108 #ifdef CONFIG_SMP
7109 	if (rq->ttwu_pending)
7110 		return 0;
7111 #endif
7112 
7113 	return 1;
7114 }
7115 
7116 /**
7117  * available_idle_cpu - is a given CPU idle for enqueuing work.
7118  * @cpu: the CPU in question.
7119  *
7120  * Return: 1 if the CPU is currently idle. 0 otherwise.
7121  */
available_idle_cpu(int cpu)7122 int available_idle_cpu(int cpu)
7123 {
7124 	if (!idle_cpu(cpu))
7125 		return 0;
7126 
7127 	if (vcpu_is_preempted(cpu))
7128 		return 0;
7129 
7130 	return 1;
7131 }
7132 
7133 /**
7134  * idle_task - return the idle task for a given CPU.
7135  * @cpu: the processor in question.
7136  *
7137  * Return: The idle task for the CPU @cpu.
7138  */
idle_task(int cpu)7139 struct task_struct *idle_task(int cpu)
7140 {
7141 	return cpu_rq(cpu)->idle;
7142 }
7143 
7144 #ifdef CONFIG_SMP
7145 /*
7146  * This function computes an effective utilization for the given CPU, to be
7147  * used for frequency selection given the linear relation: f = u * f_max.
7148  *
7149  * The scheduler tracks the following metrics:
7150  *
7151  *   cpu_util_{cfs,rt,dl,irq}()
7152  *   cpu_bw_dl()
7153  *
7154  * Where the cfs,rt and dl util numbers are tracked with the same metric and
7155  * synchronized windows and are thus directly comparable.
7156  *
7157  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7158  * which excludes things like IRQ and steal-time. These latter are then accrued
7159  * in the irq utilization.
7160  *
7161  * The DL bandwidth number otoh is not a measured metric but a value computed
7162  * based on the task model parameters and gives the minimal utilization
7163  * required to meet deadlines.
7164  */
effective_cpu_util(int cpu,unsigned long util_cfs,unsigned long max,enum cpu_util_type type,struct task_struct * p)7165 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7166 				 unsigned long max, enum cpu_util_type type,
7167 				 struct task_struct *p)
7168 {
7169 	unsigned long dl_util, util, irq;
7170 	struct rq *rq = cpu_rq(cpu);
7171 
7172 	if (!uclamp_is_used() &&
7173 	    type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7174 		return max;
7175 	}
7176 
7177 	/*
7178 	 * Early check to see if IRQ/steal time saturates the CPU, can be
7179 	 * because of inaccuracies in how we track these -- see
7180 	 * update_irq_load_avg().
7181 	 */
7182 	irq = cpu_util_irq(rq);
7183 	if (unlikely(irq >= max))
7184 		return max;
7185 
7186 	/*
7187 	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7188 	 * CFS tasks and we use the same metric to track the effective
7189 	 * utilization (PELT windows are synchronized) we can directly add them
7190 	 * to obtain the CPU's actual utilization.
7191 	 *
7192 	 * CFS and RT utilization can be boosted or capped, depending on
7193 	 * utilization clamp constraints requested by currently RUNNABLE
7194 	 * tasks.
7195 	 * When there are no CFS RUNNABLE tasks, clamps are released and
7196 	 * frequency will be gracefully reduced with the utilization decay.
7197 	 */
7198 	util = util_cfs + cpu_util_rt(rq);
7199 	if (type == FREQUENCY_UTIL)
7200 		util = uclamp_rq_util_with(rq, util, p);
7201 
7202 	dl_util = cpu_util_dl(rq);
7203 
7204 	/*
7205 	 * For frequency selection we do not make cpu_util_dl() a permanent part
7206 	 * of this sum because we want to use cpu_bw_dl() later on, but we need
7207 	 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7208 	 * that we select f_max when there is no idle time.
7209 	 *
7210 	 * NOTE: numerical errors or stop class might cause us to not quite hit
7211 	 * saturation when we should -- something for later.
7212 	 */
7213 	if (util + dl_util >= max)
7214 		return max;
7215 
7216 	/*
7217 	 * OTOH, for energy computation we need the estimated running time, so
7218 	 * include util_dl and ignore dl_bw.
7219 	 */
7220 	if (type == ENERGY_UTIL)
7221 		util += dl_util;
7222 
7223 	/*
7224 	 * There is still idle time; further improve the number by using the
7225 	 * irq metric. Because IRQ/steal time is hidden from the task clock we
7226 	 * need to scale the task numbers:
7227 	 *
7228 	 *              max - irq
7229 	 *   U' = irq + --------- * U
7230 	 *                 max
7231 	 */
7232 	util = scale_irq_capacity(util, irq, max);
7233 	util += irq;
7234 
7235 	/*
7236 	 * Bandwidth required by DEADLINE must always be granted while, for
7237 	 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7238 	 * to gracefully reduce the frequency when no tasks show up for longer
7239 	 * periods of time.
7240 	 *
7241 	 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7242 	 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7243 	 * an interface. So, we only do the latter for now.
7244 	 */
7245 	if (type == FREQUENCY_UTIL)
7246 		util += cpu_bw_dl(rq);
7247 
7248 	return min(max, util);
7249 }
7250 
sched_cpu_util(int cpu,unsigned long max)7251 unsigned long sched_cpu_util(int cpu, unsigned long max)
7252 {
7253 	return effective_cpu_util(cpu, cpu_util_cfs(cpu), max,
7254 				  ENERGY_UTIL, NULL);
7255 }
7256 #endif /* CONFIG_SMP */
7257 
7258 /**
7259  * find_process_by_pid - find a process with a matching PID value.
7260  * @pid: the pid in question.
7261  *
7262  * The task of @pid, if found. %NULL otherwise.
7263  */
find_process_by_pid(pid_t pid)7264 static struct task_struct *find_process_by_pid(pid_t pid)
7265 {
7266 	return pid ? find_task_by_vpid(pid) : current;
7267 }
7268 
7269 /*
7270  * sched_setparam() passes in -1 for its policy, to let the functions
7271  * it calls know not to change it.
7272  */
7273 #define SETPARAM_POLICY	-1
7274 
__setscheduler_params(struct task_struct * p,const struct sched_attr * attr)7275 static void __setscheduler_params(struct task_struct *p,
7276 		const struct sched_attr *attr)
7277 {
7278 	int policy = attr->sched_policy;
7279 
7280 	if (policy == SETPARAM_POLICY)
7281 		policy = p->policy;
7282 
7283 	p->policy = policy;
7284 
7285 	if (dl_policy(policy))
7286 		__setparam_dl(p, attr);
7287 	else if (fair_policy(policy))
7288 		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7289 
7290 	/*
7291 	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7292 	 * !rt_policy. Always setting this ensures that things like
7293 	 * getparam()/getattr() don't report silly values for !rt tasks.
7294 	 */
7295 	p->rt_priority = attr->sched_priority;
7296 	p->normal_prio = normal_prio(p);
7297 	set_load_weight(p, true);
7298 }
7299 
7300 /*
7301  * Check the target process has a UID that matches the current process's:
7302  */
check_same_owner(struct task_struct * p)7303 static bool check_same_owner(struct task_struct *p)
7304 {
7305 	const struct cred *cred = current_cred(), *pcred;
7306 	bool match;
7307 
7308 	rcu_read_lock();
7309 	pcred = __task_cred(p);
7310 	match = (uid_eq(cred->euid, pcred->euid) ||
7311 		 uid_eq(cred->euid, pcred->uid));
7312 	rcu_read_unlock();
7313 	return match;
7314 }
7315 
7316 /*
7317  * Allow unprivileged RT tasks to decrease priority.
7318  * Only issue a capable test if needed and only once to avoid an audit
7319  * event on permitted non-privileged operations:
7320  */
user_check_sched_setscheduler(struct task_struct * p,const struct sched_attr * attr,int policy,int reset_on_fork)7321 static int user_check_sched_setscheduler(struct task_struct *p,
7322 					 const struct sched_attr *attr,
7323 					 int policy, int reset_on_fork)
7324 {
7325 	if (fair_policy(policy)) {
7326 		if (attr->sched_nice < task_nice(p) &&
7327 		    !is_nice_reduction(p, attr->sched_nice))
7328 			goto req_priv;
7329 	}
7330 
7331 	if (rt_policy(policy)) {
7332 		unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7333 
7334 		/* Can't set/change the rt policy: */
7335 		if (policy != p->policy && !rlim_rtprio)
7336 			goto req_priv;
7337 
7338 		/* Can't increase priority: */
7339 		if (attr->sched_priority > p->rt_priority &&
7340 		    attr->sched_priority > rlim_rtprio)
7341 			goto req_priv;
7342 	}
7343 
7344 	/*
7345 	 * Can't set/change SCHED_DEADLINE policy at all for now
7346 	 * (safest behavior); in the future we would like to allow
7347 	 * unprivileged DL tasks to increase their relative deadline
7348 	 * or reduce their runtime (both ways reducing utilization)
7349 	 */
7350 	if (dl_policy(policy))
7351 		goto req_priv;
7352 
7353 	/*
7354 	 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7355 	 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7356 	 */
7357 	if (task_has_idle_policy(p) && !idle_policy(policy)) {
7358 		if (!is_nice_reduction(p, task_nice(p)))
7359 			goto req_priv;
7360 	}
7361 
7362 	/* Can't change other user's priorities: */
7363 	if (!check_same_owner(p))
7364 		goto req_priv;
7365 
7366 	/* Normal users shall not reset the sched_reset_on_fork flag: */
7367 	if (p->sched_reset_on_fork && !reset_on_fork)
7368 		goto req_priv;
7369 
7370 	return 0;
7371 
7372 req_priv:
7373 	if (!capable(CAP_SYS_NICE))
7374 		return -EPERM;
7375 
7376 	return 0;
7377 }
7378 
__sched_setscheduler(struct task_struct * p,const struct sched_attr * attr,bool user,bool pi)7379 static int __sched_setscheduler(struct task_struct *p,
7380 				const struct sched_attr *attr,
7381 				bool user, bool pi)
7382 {
7383 	int oldpolicy = -1, policy = attr->sched_policy;
7384 	int retval, oldprio, newprio, queued, running;
7385 	const struct sched_class *prev_class;
7386 	struct callback_head *head;
7387 	struct rq_flags rf;
7388 	int reset_on_fork;
7389 	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7390 	struct rq *rq;
7391 
7392 	/* The pi code expects interrupts enabled */
7393 	BUG_ON(pi && in_interrupt());
7394 recheck:
7395 	/* Double check policy once rq lock held: */
7396 	if (policy < 0) {
7397 		reset_on_fork = p->sched_reset_on_fork;
7398 		policy = oldpolicy = p->policy;
7399 	} else {
7400 		reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7401 
7402 		if (!valid_policy(policy))
7403 			return -EINVAL;
7404 	}
7405 
7406 	if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7407 		return -EINVAL;
7408 
7409 	/*
7410 	 * Valid priorities for SCHED_FIFO and SCHED_RR are
7411 	 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7412 	 * SCHED_BATCH and SCHED_IDLE is 0.
7413 	 */
7414 	if (attr->sched_priority > MAX_RT_PRIO-1)
7415 		return -EINVAL;
7416 	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7417 	    (rt_policy(policy) != (attr->sched_priority != 0)))
7418 		return -EINVAL;
7419 
7420 	if (user) {
7421 		retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7422 		if (retval)
7423 			return retval;
7424 
7425 		if (attr->sched_flags & SCHED_FLAG_SUGOV)
7426 			return -EINVAL;
7427 
7428 		retval = security_task_setscheduler(p);
7429 		if (retval)
7430 			return retval;
7431 	}
7432 
7433 	/* Update task specific "requested" clamps */
7434 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7435 		retval = uclamp_validate(p, attr);
7436 		if (retval)
7437 			return retval;
7438 	}
7439 
7440 	if (pi)
7441 		cpuset_read_lock();
7442 
7443 	/*
7444 	 * Make sure no PI-waiters arrive (or leave) while we are
7445 	 * changing the priority of the task:
7446 	 *
7447 	 * To be able to change p->policy safely, the appropriate
7448 	 * runqueue lock must be held.
7449 	 */
7450 	rq = task_rq_lock(p, &rf);
7451 	update_rq_clock(rq);
7452 
7453 	/*
7454 	 * Changing the policy of the stop threads its a very bad idea:
7455 	 */
7456 	if (p == rq->stop) {
7457 		retval = -EINVAL;
7458 		goto unlock;
7459 	}
7460 
7461 	/*
7462 	 * If not changing anything there's no need to proceed further,
7463 	 * but store a possible modification of reset_on_fork.
7464 	 */
7465 	if (unlikely(policy == p->policy)) {
7466 		if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7467 			goto change;
7468 		if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7469 			goto change;
7470 		if (dl_policy(policy) && dl_param_changed(p, attr))
7471 			goto change;
7472 		if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7473 			goto change;
7474 
7475 		p->sched_reset_on_fork = reset_on_fork;
7476 		retval = 0;
7477 		goto unlock;
7478 	}
7479 change:
7480 
7481 	if (user) {
7482 #ifdef CONFIG_RT_GROUP_SCHED
7483 		/*
7484 		 * Do not allow realtime tasks into groups that have no runtime
7485 		 * assigned.
7486 		 */
7487 		if (rt_bandwidth_enabled() && rt_policy(policy) &&
7488 				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7489 				!task_group_is_autogroup(task_group(p))) {
7490 			retval = -EPERM;
7491 			goto unlock;
7492 		}
7493 #endif
7494 #ifdef CONFIG_SMP
7495 		if (dl_bandwidth_enabled() && dl_policy(policy) &&
7496 				!(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7497 			cpumask_t *span = rq->rd->span;
7498 
7499 			/*
7500 			 * Don't allow tasks with an affinity mask smaller than
7501 			 * the entire root_domain to become SCHED_DEADLINE. We
7502 			 * will also fail if there's no bandwidth available.
7503 			 */
7504 			if (!cpumask_subset(span, p->cpus_ptr) ||
7505 			    rq->rd->dl_bw.bw == 0) {
7506 				retval = -EPERM;
7507 				goto unlock;
7508 			}
7509 		}
7510 #endif
7511 	}
7512 
7513 	/* Re-check policy now with rq lock held: */
7514 	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7515 		policy = oldpolicy = -1;
7516 		task_rq_unlock(rq, p, &rf);
7517 		if (pi)
7518 			cpuset_read_unlock();
7519 		goto recheck;
7520 	}
7521 
7522 	/*
7523 	 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7524 	 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7525 	 * is available.
7526 	 */
7527 	if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7528 		retval = -EBUSY;
7529 		goto unlock;
7530 	}
7531 
7532 	p->sched_reset_on_fork = reset_on_fork;
7533 	oldprio = p->prio;
7534 
7535 	newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7536 	if (pi) {
7537 		/*
7538 		 * Take priority boosted tasks into account. If the new
7539 		 * effective priority is unchanged, we just store the new
7540 		 * normal parameters and do not touch the scheduler class and
7541 		 * the runqueue. This will be done when the task deboost
7542 		 * itself.
7543 		 */
7544 		newprio = rt_effective_prio(p, newprio);
7545 		if (newprio == oldprio)
7546 			queue_flags &= ~DEQUEUE_MOVE;
7547 	}
7548 
7549 	queued = task_on_rq_queued(p);
7550 	running = task_current(rq, p);
7551 	if (queued)
7552 		dequeue_task(rq, p, queue_flags);
7553 	if (running)
7554 		put_prev_task(rq, p);
7555 
7556 	prev_class = p->sched_class;
7557 
7558 	if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7559 		__setscheduler_params(p, attr);
7560 		__setscheduler_prio(p, newprio);
7561 	}
7562 	__setscheduler_uclamp(p, attr);
7563 
7564 	if (queued) {
7565 		/*
7566 		 * We enqueue to tail when the priority of a task is
7567 		 * increased (user space view).
7568 		 */
7569 		if (oldprio < p->prio)
7570 			queue_flags |= ENQUEUE_HEAD;
7571 
7572 		enqueue_task(rq, p, queue_flags);
7573 	}
7574 	if (running)
7575 		set_next_task(rq, p);
7576 
7577 	check_class_changed(rq, p, prev_class, oldprio);
7578 
7579 	/* Avoid rq from going away on us: */
7580 	preempt_disable();
7581 	head = splice_balance_callbacks(rq);
7582 	task_rq_unlock(rq, p, &rf);
7583 
7584 	if (pi) {
7585 		cpuset_read_unlock();
7586 		rt_mutex_adjust_pi(p);
7587 	}
7588 
7589 	/* Run balance callbacks after we've adjusted the PI chain: */
7590 	balance_callbacks(rq, head);
7591 	preempt_enable();
7592 
7593 	return 0;
7594 
7595 unlock:
7596 	task_rq_unlock(rq, p, &rf);
7597 	if (pi)
7598 		cpuset_read_unlock();
7599 	return retval;
7600 }
7601 
_sched_setscheduler(struct task_struct * p,int policy,const struct sched_param * param,bool check)7602 static int _sched_setscheduler(struct task_struct *p, int policy,
7603 			       const struct sched_param *param, bool check)
7604 {
7605 	struct sched_attr attr = {
7606 		.sched_policy   = policy,
7607 		.sched_priority = param->sched_priority,
7608 		.sched_nice	= PRIO_TO_NICE(p->static_prio),
7609 	};
7610 
7611 	/* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7612 	if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7613 		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7614 		policy &= ~SCHED_RESET_ON_FORK;
7615 		attr.sched_policy = policy;
7616 	}
7617 
7618 	return __sched_setscheduler(p, &attr, check, true);
7619 }
7620 /**
7621  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7622  * @p: the task in question.
7623  * @policy: new policy.
7624  * @param: structure containing the new RT priority.
7625  *
7626  * Use sched_set_fifo(), read its comment.
7627  *
7628  * Return: 0 on success. An error code otherwise.
7629  *
7630  * NOTE that the task may be already dead.
7631  */
sched_setscheduler(struct task_struct * p,int policy,const struct sched_param * param)7632 int sched_setscheduler(struct task_struct *p, int policy,
7633 		       const struct sched_param *param)
7634 {
7635 	return _sched_setscheduler(p, policy, param, true);
7636 }
7637 
sched_setattr(struct task_struct * p,const struct sched_attr * attr)7638 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7639 {
7640 	return __sched_setscheduler(p, attr, true, true);
7641 }
7642 
sched_setattr_nocheck(struct task_struct * p,const struct sched_attr * attr)7643 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7644 {
7645 	return __sched_setscheduler(p, attr, false, true);
7646 }
7647 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7648 
7649 /**
7650  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7651  * @p: the task in question.
7652  * @policy: new policy.
7653  * @param: structure containing the new RT priority.
7654  *
7655  * Just like sched_setscheduler, only don't bother checking if the
7656  * current context has permission.  For example, this is needed in
7657  * stop_machine(): we create temporary high priority worker threads,
7658  * but our caller might not have that capability.
7659  *
7660  * Return: 0 on success. An error code otherwise.
7661  */
sched_setscheduler_nocheck(struct task_struct * p,int policy,const struct sched_param * param)7662 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7663 			       const struct sched_param *param)
7664 {
7665 	return _sched_setscheduler(p, policy, param, false);
7666 }
7667 
7668 /*
7669  * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7670  * incapable of resource management, which is the one thing an OS really should
7671  * be doing.
7672  *
7673  * This is of course the reason it is limited to privileged users only.
7674  *
7675  * Worse still; it is fundamentally impossible to compose static priority
7676  * workloads. You cannot take two correctly working static prio workloads
7677  * and smash them together and still expect them to work.
7678  *
7679  * For this reason 'all' FIFO tasks the kernel creates are basically at:
7680  *
7681  *   MAX_RT_PRIO / 2
7682  *
7683  * The administrator _MUST_ configure the system, the kernel simply doesn't
7684  * know enough information to make a sensible choice.
7685  */
sched_set_fifo(struct task_struct * p)7686 void sched_set_fifo(struct task_struct *p)
7687 {
7688 	struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7689 	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7690 }
7691 EXPORT_SYMBOL_GPL(sched_set_fifo);
7692 
7693 /*
7694  * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7695  */
sched_set_fifo_low(struct task_struct * p)7696 void sched_set_fifo_low(struct task_struct *p)
7697 {
7698 	struct sched_param sp = { .sched_priority = 1 };
7699 	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7700 }
7701 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7702 
sched_set_normal(struct task_struct * p,int nice)7703 void sched_set_normal(struct task_struct *p, int nice)
7704 {
7705 	struct sched_attr attr = {
7706 		.sched_policy = SCHED_NORMAL,
7707 		.sched_nice = nice,
7708 	};
7709 	WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7710 }
7711 EXPORT_SYMBOL_GPL(sched_set_normal);
7712 
7713 static int
do_sched_setscheduler(pid_t pid,int policy,struct sched_param __user * param)7714 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7715 {
7716 	struct sched_param lparam;
7717 	struct task_struct *p;
7718 	int retval;
7719 
7720 	if (!param || pid < 0)
7721 		return -EINVAL;
7722 	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7723 		return -EFAULT;
7724 
7725 	rcu_read_lock();
7726 	retval = -ESRCH;
7727 	p = find_process_by_pid(pid);
7728 	if (likely(p))
7729 		get_task_struct(p);
7730 	rcu_read_unlock();
7731 
7732 	if (likely(p)) {
7733 		retval = sched_setscheduler(p, policy, &lparam);
7734 		put_task_struct(p);
7735 	}
7736 
7737 	return retval;
7738 }
7739 
7740 /*
7741  * Mimics kernel/events/core.c perf_copy_attr().
7742  */
sched_copy_attr(struct sched_attr __user * uattr,struct sched_attr * attr)7743 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7744 {
7745 	u32 size;
7746 	int ret;
7747 
7748 	/* Zero the full structure, so that a short copy will be nice: */
7749 	memset(attr, 0, sizeof(*attr));
7750 
7751 	ret = get_user(size, &uattr->size);
7752 	if (ret)
7753 		return ret;
7754 
7755 	/* ABI compatibility quirk: */
7756 	if (!size)
7757 		size = SCHED_ATTR_SIZE_VER0;
7758 	if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7759 		goto err_size;
7760 
7761 	ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7762 	if (ret) {
7763 		if (ret == -E2BIG)
7764 			goto err_size;
7765 		return ret;
7766 	}
7767 
7768 	if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7769 	    size < SCHED_ATTR_SIZE_VER1)
7770 		return -EINVAL;
7771 
7772 	/*
7773 	 * XXX: Do we want to be lenient like existing syscalls; or do we want
7774 	 * to be strict and return an error on out-of-bounds values?
7775 	 */
7776 	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7777 
7778 	return 0;
7779 
7780 err_size:
7781 	put_user(sizeof(*attr), &uattr->size);
7782 	return -E2BIG;
7783 }
7784 
get_params(struct task_struct * p,struct sched_attr * attr)7785 static void get_params(struct task_struct *p, struct sched_attr *attr)
7786 {
7787 	if (task_has_dl_policy(p))
7788 		__getparam_dl(p, attr);
7789 	else if (task_has_rt_policy(p))
7790 		attr->sched_priority = p->rt_priority;
7791 	else
7792 		attr->sched_nice = task_nice(p);
7793 }
7794 
7795 /**
7796  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7797  * @pid: the pid in question.
7798  * @policy: new policy.
7799  * @param: structure containing the new RT priority.
7800  *
7801  * Return: 0 on success. An error code otherwise.
7802  */
SYSCALL_DEFINE3(sched_setscheduler,pid_t,pid,int,policy,struct sched_param __user *,param)7803 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7804 {
7805 	if (policy < 0)
7806 		return -EINVAL;
7807 
7808 	return do_sched_setscheduler(pid, policy, param);
7809 }
7810 
7811 /**
7812  * sys_sched_setparam - set/change the RT priority of a thread
7813  * @pid: the pid in question.
7814  * @param: structure containing the new RT priority.
7815  *
7816  * Return: 0 on success. An error code otherwise.
7817  */
SYSCALL_DEFINE2(sched_setparam,pid_t,pid,struct sched_param __user *,param)7818 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7819 {
7820 	return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7821 }
7822 
7823 /**
7824  * sys_sched_setattr - same as above, but with extended sched_attr
7825  * @pid: the pid in question.
7826  * @uattr: structure containing the extended parameters.
7827  * @flags: for future extension.
7828  */
SYSCALL_DEFINE3(sched_setattr,pid_t,pid,struct sched_attr __user *,uattr,unsigned int,flags)7829 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7830 			       unsigned int, flags)
7831 {
7832 	struct sched_attr attr;
7833 	struct task_struct *p;
7834 	int retval;
7835 
7836 	if (!uattr || pid < 0 || flags)
7837 		return -EINVAL;
7838 
7839 	retval = sched_copy_attr(uattr, &attr);
7840 	if (retval)
7841 		return retval;
7842 
7843 	if ((int)attr.sched_policy < 0)
7844 		return -EINVAL;
7845 	if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7846 		attr.sched_policy = SETPARAM_POLICY;
7847 
7848 	rcu_read_lock();
7849 	retval = -ESRCH;
7850 	p = find_process_by_pid(pid);
7851 	if (likely(p))
7852 		get_task_struct(p);
7853 	rcu_read_unlock();
7854 
7855 	if (likely(p)) {
7856 		if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7857 			get_params(p, &attr);
7858 		retval = sched_setattr(p, &attr);
7859 		put_task_struct(p);
7860 	}
7861 
7862 	return retval;
7863 }
7864 
7865 /**
7866  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7867  * @pid: the pid in question.
7868  *
7869  * Return: On success, the policy of the thread. Otherwise, a negative error
7870  * code.
7871  */
SYSCALL_DEFINE1(sched_getscheduler,pid_t,pid)7872 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7873 {
7874 	struct task_struct *p;
7875 	int retval;
7876 
7877 	if (pid < 0)
7878 		return -EINVAL;
7879 
7880 	retval = -ESRCH;
7881 	rcu_read_lock();
7882 	p = find_process_by_pid(pid);
7883 	if (p) {
7884 		retval = security_task_getscheduler(p);
7885 		if (!retval)
7886 			retval = p->policy
7887 				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7888 	}
7889 	rcu_read_unlock();
7890 	return retval;
7891 }
7892 
7893 /**
7894  * sys_sched_getparam - get the RT priority of a thread
7895  * @pid: the pid in question.
7896  * @param: structure containing the RT priority.
7897  *
7898  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7899  * code.
7900  */
SYSCALL_DEFINE2(sched_getparam,pid_t,pid,struct sched_param __user *,param)7901 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7902 {
7903 	struct sched_param lp = { .sched_priority = 0 };
7904 	struct task_struct *p;
7905 	int retval;
7906 
7907 	if (!param || pid < 0)
7908 		return -EINVAL;
7909 
7910 	rcu_read_lock();
7911 	p = find_process_by_pid(pid);
7912 	retval = -ESRCH;
7913 	if (!p)
7914 		goto out_unlock;
7915 
7916 	retval = security_task_getscheduler(p);
7917 	if (retval)
7918 		goto out_unlock;
7919 
7920 	if (task_has_rt_policy(p))
7921 		lp.sched_priority = p->rt_priority;
7922 	rcu_read_unlock();
7923 
7924 	/*
7925 	 * This one might sleep, we cannot do it with a spinlock held ...
7926 	 */
7927 	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7928 
7929 	return retval;
7930 
7931 out_unlock:
7932 	rcu_read_unlock();
7933 	return retval;
7934 }
7935 
7936 /*
7937  * Copy the kernel size attribute structure (which might be larger
7938  * than what user-space knows about) to user-space.
7939  *
7940  * Note that all cases are valid: user-space buffer can be larger or
7941  * smaller than the kernel-space buffer. The usual case is that both
7942  * have the same size.
7943  */
7944 static int
sched_attr_copy_to_user(struct sched_attr __user * uattr,struct sched_attr * kattr,unsigned int usize)7945 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7946 			struct sched_attr *kattr,
7947 			unsigned int usize)
7948 {
7949 	unsigned int ksize = sizeof(*kattr);
7950 
7951 	if (!access_ok(uattr, usize))
7952 		return -EFAULT;
7953 
7954 	/*
7955 	 * sched_getattr() ABI forwards and backwards compatibility:
7956 	 *
7957 	 * If usize == ksize then we just copy everything to user-space and all is good.
7958 	 *
7959 	 * If usize < ksize then we only copy as much as user-space has space for,
7960 	 * this keeps ABI compatibility as well. We skip the rest.
7961 	 *
7962 	 * If usize > ksize then user-space is using a newer version of the ABI,
7963 	 * which part the kernel doesn't know about. Just ignore it - tooling can
7964 	 * detect the kernel's knowledge of attributes from the attr->size value
7965 	 * which is set to ksize in this case.
7966 	 */
7967 	kattr->size = min(usize, ksize);
7968 
7969 	if (copy_to_user(uattr, kattr, kattr->size))
7970 		return -EFAULT;
7971 
7972 	return 0;
7973 }
7974 
7975 /**
7976  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7977  * @pid: the pid in question.
7978  * @uattr: structure containing the extended parameters.
7979  * @usize: sizeof(attr) for fwd/bwd comp.
7980  * @flags: for future extension.
7981  */
SYSCALL_DEFINE4(sched_getattr,pid_t,pid,struct sched_attr __user *,uattr,unsigned int,usize,unsigned int,flags)7982 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7983 		unsigned int, usize, unsigned int, flags)
7984 {
7985 	struct sched_attr kattr = { };
7986 	struct task_struct *p;
7987 	int retval;
7988 
7989 	if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7990 	    usize < SCHED_ATTR_SIZE_VER0 || flags)
7991 		return -EINVAL;
7992 
7993 	rcu_read_lock();
7994 	p = find_process_by_pid(pid);
7995 	retval = -ESRCH;
7996 	if (!p)
7997 		goto out_unlock;
7998 
7999 	retval = security_task_getscheduler(p);
8000 	if (retval)
8001 		goto out_unlock;
8002 
8003 	kattr.sched_policy = p->policy;
8004 	if (p->sched_reset_on_fork)
8005 		kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8006 	get_params(p, &kattr);
8007 	kattr.sched_flags &= SCHED_FLAG_ALL;
8008 
8009 #ifdef CONFIG_UCLAMP_TASK
8010 	/*
8011 	 * This could race with another potential updater, but this is fine
8012 	 * because it'll correctly read the old or the new value. We don't need
8013 	 * to guarantee who wins the race as long as it doesn't return garbage.
8014 	 */
8015 	kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8016 	kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8017 #endif
8018 
8019 	rcu_read_unlock();
8020 
8021 	return sched_attr_copy_to_user(uattr, &kattr, usize);
8022 
8023 out_unlock:
8024 	rcu_read_unlock();
8025 	return retval;
8026 }
8027 
8028 #ifdef CONFIG_SMP
dl_task_check_affinity(struct task_struct * p,const struct cpumask * mask)8029 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8030 {
8031 	int ret = 0;
8032 
8033 	/*
8034 	 * If the task isn't a deadline task or admission control is
8035 	 * disabled then we don't care about affinity changes.
8036 	 */
8037 	if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8038 		return 0;
8039 
8040 	/*
8041 	 * Since bandwidth control happens on root_domain basis,
8042 	 * if admission test is enabled, we only admit -deadline
8043 	 * tasks allowed to run on all the CPUs in the task's
8044 	 * root_domain.
8045 	 */
8046 	rcu_read_lock();
8047 	if (!cpumask_subset(task_rq(p)->rd->span, mask))
8048 		ret = -EBUSY;
8049 	rcu_read_unlock();
8050 	return ret;
8051 }
8052 #endif
8053 
8054 static int
__sched_setaffinity(struct task_struct * p,const struct cpumask * mask)8055 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
8056 {
8057 	int retval;
8058 	cpumask_var_t cpus_allowed, new_mask;
8059 
8060 	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8061 		return -ENOMEM;
8062 
8063 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8064 		retval = -ENOMEM;
8065 		goto out_free_cpus_allowed;
8066 	}
8067 
8068 	cpuset_cpus_allowed(p, cpus_allowed);
8069 	cpumask_and(new_mask, mask, cpus_allowed);
8070 
8071 	retval = dl_task_check_affinity(p, new_mask);
8072 	if (retval)
8073 		goto out_free_new_mask;
8074 again:
8075 	retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
8076 	if (retval)
8077 		goto out_free_new_mask;
8078 
8079 	cpuset_cpus_allowed(p, cpus_allowed);
8080 	if (!cpumask_subset(new_mask, cpus_allowed)) {
8081 		/*
8082 		 * We must have raced with a concurrent cpuset update.
8083 		 * Just reset the cpumask to the cpuset's cpus_allowed.
8084 		 */
8085 		cpumask_copy(new_mask, cpus_allowed);
8086 		goto again;
8087 	}
8088 
8089 out_free_new_mask:
8090 	free_cpumask_var(new_mask);
8091 out_free_cpus_allowed:
8092 	free_cpumask_var(cpus_allowed);
8093 	return retval;
8094 }
8095 
sched_setaffinity(pid_t pid,const struct cpumask * in_mask)8096 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8097 {
8098 	struct task_struct *p;
8099 	int retval;
8100 
8101 	rcu_read_lock();
8102 
8103 	p = find_process_by_pid(pid);
8104 	if (!p) {
8105 		rcu_read_unlock();
8106 		return -ESRCH;
8107 	}
8108 
8109 	/* Prevent p going away */
8110 	get_task_struct(p);
8111 	rcu_read_unlock();
8112 
8113 	if (p->flags & PF_NO_SETAFFINITY) {
8114 		retval = -EINVAL;
8115 		goto out_put_task;
8116 	}
8117 
8118 	if (!check_same_owner(p)) {
8119 		rcu_read_lock();
8120 		if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8121 			rcu_read_unlock();
8122 			retval = -EPERM;
8123 			goto out_put_task;
8124 		}
8125 		rcu_read_unlock();
8126 	}
8127 
8128 	retval = security_task_setscheduler(p);
8129 	if (retval)
8130 		goto out_put_task;
8131 
8132 	retval = __sched_setaffinity(p, in_mask);
8133 out_put_task:
8134 	put_task_struct(p);
8135 	return retval;
8136 }
8137 
get_user_cpu_mask(unsigned long __user * user_mask_ptr,unsigned len,struct cpumask * new_mask)8138 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8139 			     struct cpumask *new_mask)
8140 {
8141 	if (len < cpumask_size())
8142 		cpumask_clear(new_mask);
8143 	else if (len > cpumask_size())
8144 		len = cpumask_size();
8145 
8146 	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8147 }
8148 
8149 /**
8150  * sys_sched_setaffinity - set the CPU affinity of a process
8151  * @pid: pid of the process
8152  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8153  * @user_mask_ptr: user-space pointer to the new CPU mask
8154  *
8155  * Return: 0 on success. An error code otherwise.
8156  */
SYSCALL_DEFINE3(sched_setaffinity,pid_t,pid,unsigned int,len,unsigned long __user *,user_mask_ptr)8157 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8158 		unsigned long __user *, user_mask_ptr)
8159 {
8160 	cpumask_var_t new_mask;
8161 	int retval;
8162 
8163 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8164 		return -ENOMEM;
8165 
8166 	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8167 	if (retval == 0)
8168 		retval = sched_setaffinity(pid, new_mask);
8169 	free_cpumask_var(new_mask);
8170 	return retval;
8171 }
8172 
sched_getaffinity(pid_t pid,struct cpumask * mask)8173 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8174 {
8175 	struct task_struct *p;
8176 	unsigned long flags;
8177 	int retval;
8178 
8179 	rcu_read_lock();
8180 
8181 	retval = -ESRCH;
8182 	p = find_process_by_pid(pid);
8183 	if (!p)
8184 		goto out_unlock;
8185 
8186 	retval = security_task_getscheduler(p);
8187 	if (retval)
8188 		goto out_unlock;
8189 
8190 	raw_spin_lock_irqsave(&p->pi_lock, flags);
8191 	cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8192 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8193 
8194 out_unlock:
8195 	rcu_read_unlock();
8196 
8197 	return retval;
8198 }
8199 
8200 /**
8201  * sys_sched_getaffinity - get the CPU affinity of a process
8202  * @pid: pid of the process
8203  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8204  * @user_mask_ptr: user-space pointer to hold the current CPU mask
8205  *
8206  * Return: size of CPU mask copied to user_mask_ptr on success. An
8207  * error code otherwise.
8208  */
SYSCALL_DEFINE3(sched_getaffinity,pid_t,pid,unsigned int,len,unsigned long __user *,user_mask_ptr)8209 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8210 		unsigned long __user *, user_mask_ptr)
8211 {
8212 	int ret;
8213 	cpumask_var_t mask;
8214 
8215 	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8216 		return -EINVAL;
8217 	if (len & (sizeof(unsigned long)-1))
8218 		return -EINVAL;
8219 
8220 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8221 		return -ENOMEM;
8222 
8223 	ret = sched_getaffinity(pid, mask);
8224 	if (ret == 0) {
8225 		unsigned int retlen = min(len, cpumask_size());
8226 
8227 		if (copy_to_user(user_mask_ptr, mask, retlen))
8228 			ret = -EFAULT;
8229 		else
8230 			ret = retlen;
8231 	}
8232 	free_cpumask_var(mask);
8233 
8234 	return ret;
8235 }
8236 
do_sched_yield(void)8237 static void do_sched_yield(void)
8238 {
8239 	struct rq_flags rf;
8240 	struct rq *rq;
8241 
8242 	rq = this_rq_lock_irq(&rf);
8243 
8244 	schedstat_inc(rq->yld_count);
8245 	current->sched_class->yield_task(rq);
8246 
8247 	preempt_disable();
8248 	rq_unlock_irq(rq, &rf);
8249 	sched_preempt_enable_no_resched();
8250 
8251 	schedule();
8252 }
8253 
8254 /**
8255  * sys_sched_yield - yield the current processor to other threads.
8256  *
8257  * This function yields the current CPU to other tasks. If there are no
8258  * other threads running on this CPU then this function will return.
8259  *
8260  * Return: 0.
8261  */
SYSCALL_DEFINE0(sched_yield)8262 SYSCALL_DEFINE0(sched_yield)
8263 {
8264 	do_sched_yield();
8265 	return 0;
8266 }
8267 
8268 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
__cond_resched(void)8269 int __sched __cond_resched(void)
8270 {
8271 	if (should_resched(0)) {
8272 		preempt_schedule_common();
8273 		return 1;
8274 	}
8275 	/*
8276 	 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8277 	 * whether the current CPU is in an RCU read-side critical section,
8278 	 * so the tick can report quiescent states even for CPUs looping
8279 	 * in kernel context.  In contrast, in non-preemptible kernels,
8280 	 * RCU readers leave no in-memory hints, which means that CPU-bound
8281 	 * processes executing in kernel context might never report an
8282 	 * RCU quiescent state.  Therefore, the following code causes
8283 	 * cond_resched() to report a quiescent state, but only when RCU
8284 	 * is in urgent need of one.
8285 	 */
8286 #ifndef CONFIG_PREEMPT_RCU
8287 	rcu_all_qs();
8288 #endif
8289 	return 0;
8290 }
8291 EXPORT_SYMBOL(__cond_resched);
8292 #endif
8293 
8294 #ifdef CONFIG_PREEMPT_DYNAMIC
8295 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8296 #define cond_resched_dynamic_enabled	__cond_resched
8297 #define cond_resched_dynamic_disabled	((void *)&__static_call_return0)
8298 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8299 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8300 
8301 #define might_resched_dynamic_enabled	__cond_resched
8302 #define might_resched_dynamic_disabled	((void *)&__static_call_return0)
8303 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8304 EXPORT_STATIC_CALL_TRAMP(might_resched);
8305 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8306 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
dynamic_cond_resched(void)8307 int __sched dynamic_cond_resched(void)
8308 {
8309 	if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8310 		return 0;
8311 	return __cond_resched();
8312 }
8313 EXPORT_SYMBOL(dynamic_cond_resched);
8314 
8315 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
dynamic_might_resched(void)8316 int __sched dynamic_might_resched(void)
8317 {
8318 	if (!static_branch_unlikely(&sk_dynamic_might_resched))
8319 		return 0;
8320 	return __cond_resched();
8321 }
8322 EXPORT_SYMBOL(dynamic_might_resched);
8323 #endif
8324 #endif
8325 
8326 /*
8327  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8328  * call schedule, and on return reacquire the lock.
8329  *
8330  * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8331  * operations here to prevent schedule() from being called twice (once via
8332  * spin_unlock(), once by hand).
8333  */
__cond_resched_lock(spinlock_t * lock)8334 int __cond_resched_lock(spinlock_t *lock)
8335 {
8336 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8337 	int ret = 0;
8338 
8339 	lockdep_assert_held(lock);
8340 
8341 	if (spin_needbreak(lock) || resched) {
8342 		spin_unlock(lock);
8343 		if (!_cond_resched())
8344 			cpu_relax();
8345 		ret = 1;
8346 		spin_lock(lock);
8347 	}
8348 	return ret;
8349 }
8350 EXPORT_SYMBOL(__cond_resched_lock);
8351 
__cond_resched_rwlock_read(rwlock_t * lock)8352 int __cond_resched_rwlock_read(rwlock_t *lock)
8353 {
8354 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8355 	int ret = 0;
8356 
8357 	lockdep_assert_held_read(lock);
8358 
8359 	if (rwlock_needbreak(lock) || resched) {
8360 		read_unlock(lock);
8361 		if (!_cond_resched())
8362 			cpu_relax();
8363 		ret = 1;
8364 		read_lock(lock);
8365 	}
8366 	return ret;
8367 }
8368 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8369 
__cond_resched_rwlock_write(rwlock_t * lock)8370 int __cond_resched_rwlock_write(rwlock_t *lock)
8371 {
8372 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8373 	int ret = 0;
8374 
8375 	lockdep_assert_held_write(lock);
8376 
8377 	if (rwlock_needbreak(lock) || resched) {
8378 		write_unlock(lock);
8379 		if (!_cond_resched())
8380 			cpu_relax();
8381 		ret = 1;
8382 		write_lock(lock);
8383 	}
8384 	return ret;
8385 }
8386 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8387 
8388 #ifdef CONFIG_PREEMPT_DYNAMIC
8389 
8390 #ifdef CONFIG_GENERIC_ENTRY
8391 #include <linux/entry-common.h>
8392 #endif
8393 
8394 /*
8395  * SC:cond_resched
8396  * SC:might_resched
8397  * SC:preempt_schedule
8398  * SC:preempt_schedule_notrace
8399  * SC:irqentry_exit_cond_resched
8400  *
8401  *
8402  * NONE:
8403  *   cond_resched               <- __cond_resched
8404  *   might_resched              <- RET0
8405  *   preempt_schedule           <- NOP
8406  *   preempt_schedule_notrace   <- NOP
8407  *   irqentry_exit_cond_resched <- NOP
8408  *
8409  * VOLUNTARY:
8410  *   cond_resched               <- __cond_resched
8411  *   might_resched              <- __cond_resched
8412  *   preempt_schedule           <- NOP
8413  *   preempt_schedule_notrace   <- NOP
8414  *   irqentry_exit_cond_resched <- NOP
8415  *
8416  * FULL:
8417  *   cond_resched               <- RET0
8418  *   might_resched              <- RET0
8419  *   preempt_schedule           <- preempt_schedule
8420  *   preempt_schedule_notrace   <- preempt_schedule_notrace
8421  *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8422  */
8423 
8424 enum {
8425 	preempt_dynamic_undefined = -1,
8426 	preempt_dynamic_none,
8427 	preempt_dynamic_voluntary,
8428 	preempt_dynamic_full,
8429 };
8430 
8431 int preempt_dynamic_mode = preempt_dynamic_undefined;
8432 
sched_dynamic_mode(const char * str)8433 int sched_dynamic_mode(const char *str)
8434 {
8435 	if (!strcmp(str, "none"))
8436 		return preempt_dynamic_none;
8437 
8438 	if (!strcmp(str, "voluntary"))
8439 		return preempt_dynamic_voluntary;
8440 
8441 	if (!strcmp(str, "full"))
8442 		return preempt_dynamic_full;
8443 
8444 	return -EINVAL;
8445 }
8446 
8447 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8448 #define preempt_dynamic_enable(f)	static_call_update(f, f##_dynamic_enabled)
8449 #define preempt_dynamic_disable(f)	static_call_update(f, f##_dynamic_disabled)
8450 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8451 #define preempt_dynamic_enable(f)	static_key_enable(&sk_dynamic_##f.key)
8452 #define preempt_dynamic_disable(f)	static_key_disable(&sk_dynamic_##f.key)
8453 #else
8454 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8455 #endif
8456 
sched_dynamic_update(int mode)8457 void sched_dynamic_update(int mode)
8458 {
8459 	/*
8460 	 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8461 	 * the ZERO state, which is invalid.
8462 	 */
8463 	preempt_dynamic_enable(cond_resched);
8464 	preempt_dynamic_enable(might_resched);
8465 	preempt_dynamic_enable(preempt_schedule);
8466 	preempt_dynamic_enable(preempt_schedule_notrace);
8467 	preempt_dynamic_enable(irqentry_exit_cond_resched);
8468 
8469 	switch (mode) {
8470 	case preempt_dynamic_none:
8471 		preempt_dynamic_enable(cond_resched);
8472 		preempt_dynamic_disable(might_resched);
8473 		preempt_dynamic_disable(preempt_schedule);
8474 		preempt_dynamic_disable(preempt_schedule_notrace);
8475 		preempt_dynamic_disable(irqentry_exit_cond_resched);
8476 		pr_info("Dynamic Preempt: none\n");
8477 		break;
8478 
8479 	case preempt_dynamic_voluntary:
8480 		preempt_dynamic_enable(cond_resched);
8481 		preempt_dynamic_enable(might_resched);
8482 		preempt_dynamic_disable(preempt_schedule);
8483 		preempt_dynamic_disable(preempt_schedule_notrace);
8484 		preempt_dynamic_disable(irqentry_exit_cond_resched);
8485 		pr_info("Dynamic Preempt: voluntary\n");
8486 		break;
8487 
8488 	case preempt_dynamic_full:
8489 		preempt_dynamic_disable(cond_resched);
8490 		preempt_dynamic_disable(might_resched);
8491 		preempt_dynamic_enable(preempt_schedule);
8492 		preempt_dynamic_enable(preempt_schedule_notrace);
8493 		preempt_dynamic_enable(irqentry_exit_cond_resched);
8494 		pr_info("Dynamic Preempt: full\n");
8495 		break;
8496 	}
8497 
8498 	preempt_dynamic_mode = mode;
8499 }
8500 
setup_preempt_mode(char * str)8501 static int __init setup_preempt_mode(char *str)
8502 {
8503 	int mode = sched_dynamic_mode(str);
8504 	if (mode < 0) {
8505 		pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8506 		return 0;
8507 	}
8508 
8509 	sched_dynamic_update(mode);
8510 	return 1;
8511 }
8512 __setup("preempt=", setup_preempt_mode);
8513 
preempt_dynamic_init(void)8514 static void __init preempt_dynamic_init(void)
8515 {
8516 	if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8517 		if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8518 			sched_dynamic_update(preempt_dynamic_none);
8519 		} else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8520 			sched_dynamic_update(preempt_dynamic_voluntary);
8521 		} else {
8522 			/* Default static call setting, nothing to do */
8523 			WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8524 			preempt_dynamic_mode = preempt_dynamic_full;
8525 			pr_info("Dynamic Preempt: full\n");
8526 		}
8527 	}
8528 }
8529 
8530 #define PREEMPT_MODEL_ACCESSOR(mode) \
8531 	bool preempt_model_##mode(void)						 \
8532 	{									 \
8533 		WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8534 		return preempt_dynamic_mode == preempt_dynamic_##mode;		 \
8535 	}									 \
8536 	EXPORT_SYMBOL_GPL(preempt_model_##mode)
8537 
8538 PREEMPT_MODEL_ACCESSOR(none);
8539 PREEMPT_MODEL_ACCESSOR(voluntary);
8540 PREEMPT_MODEL_ACCESSOR(full);
8541 
8542 #else /* !CONFIG_PREEMPT_DYNAMIC */
8543 
preempt_dynamic_init(void)8544 static inline void preempt_dynamic_init(void) { }
8545 
8546 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8547 
8548 /**
8549  * yield - yield the current processor to other threads.
8550  *
8551  * Do not ever use this function, there's a 99% chance you're doing it wrong.
8552  *
8553  * The scheduler is at all times free to pick the calling task as the most
8554  * eligible task to run, if removing the yield() call from your code breaks
8555  * it, it's already broken.
8556  *
8557  * Typical broken usage is:
8558  *
8559  * while (!event)
8560  *	yield();
8561  *
8562  * where one assumes that yield() will let 'the other' process run that will
8563  * make event true. If the current task is a SCHED_FIFO task that will never
8564  * happen. Never use yield() as a progress guarantee!!
8565  *
8566  * If you want to use yield() to wait for something, use wait_event().
8567  * If you want to use yield() to be 'nice' for others, use cond_resched().
8568  * If you still want to use yield(), do not!
8569  */
yield(void)8570 void __sched yield(void)
8571 {
8572 	set_current_state(TASK_RUNNING);
8573 	do_sched_yield();
8574 }
8575 EXPORT_SYMBOL(yield);
8576 
8577 /**
8578  * yield_to - yield the current processor to another thread in
8579  * your thread group, or accelerate that thread toward the
8580  * processor it's on.
8581  * @p: target task
8582  * @preempt: whether task preemption is allowed or not
8583  *
8584  * It's the caller's job to ensure that the target task struct
8585  * can't go away on us before we can do any checks.
8586  *
8587  * Return:
8588  *	true (>0) if we indeed boosted the target task.
8589  *	false (0) if we failed to boost the target.
8590  *	-ESRCH if there's no task to yield to.
8591  */
yield_to(struct task_struct * p,bool preempt)8592 int __sched yield_to(struct task_struct *p, bool preempt)
8593 {
8594 	struct task_struct *curr = current;
8595 	struct rq *rq, *p_rq;
8596 	unsigned long flags;
8597 	int yielded = 0;
8598 
8599 	local_irq_save(flags);
8600 	rq = this_rq();
8601 
8602 again:
8603 	p_rq = task_rq(p);
8604 	/*
8605 	 * If we're the only runnable task on the rq and target rq also
8606 	 * has only one task, there's absolutely no point in yielding.
8607 	 */
8608 	if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8609 		yielded = -ESRCH;
8610 		goto out_irq;
8611 	}
8612 
8613 	double_rq_lock(rq, p_rq);
8614 	if (task_rq(p) != p_rq) {
8615 		double_rq_unlock(rq, p_rq);
8616 		goto again;
8617 	}
8618 
8619 	if (!curr->sched_class->yield_to_task)
8620 		goto out_unlock;
8621 
8622 	if (curr->sched_class != p->sched_class)
8623 		goto out_unlock;
8624 
8625 	if (task_running(p_rq, p) || !task_is_running(p))
8626 		goto out_unlock;
8627 
8628 	yielded = curr->sched_class->yield_to_task(rq, p);
8629 	if (yielded) {
8630 		schedstat_inc(rq->yld_count);
8631 		/*
8632 		 * Make p's CPU reschedule; pick_next_entity takes care of
8633 		 * fairness.
8634 		 */
8635 		if (preempt && rq != p_rq)
8636 			resched_curr(p_rq);
8637 	}
8638 
8639 out_unlock:
8640 	double_rq_unlock(rq, p_rq);
8641 out_irq:
8642 	local_irq_restore(flags);
8643 
8644 	if (yielded > 0)
8645 		schedule();
8646 
8647 	return yielded;
8648 }
8649 EXPORT_SYMBOL_GPL(yield_to);
8650 
io_schedule_prepare(void)8651 int io_schedule_prepare(void)
8652 {
8653 	int old_iowait = current->in_iowait;
8654 
8655 	current->in_iowait = 1;
8656 	blk_flush_plug(current->plug, true);
8657 	return old_iowait;
8658 }
8659 
io_schedule_finish(int token)8660 void io_schedule_finish(int token)
8661 {
8662 	current->in_iowait = token;
8663 }
8664 
8665 /*
8666  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8667  * that process accounting knows that this is a task in IO wait state.
8668  */
io_schedule_timeout(long timeout)8669 long __sched io_schedule_timeout(long timeout)
8670 {
8671 	int token;
8672 	long ret;
8673 
8674 	token = io_schedule_prepare();
8675 	ret = schedule_timeout(timeout);
8676 	io_schedule_finish(token);
8677 
8678 	return ret;
8679 }
8680 EXPORT_SYMBOL(io_schedule_timeout);
8681 
io_schedule(void)8682 void __sched io_schedule(void)
8683 {
8684 	int token;
8685 
8686 	token = io_schedule_prepare();
8687 	schedule();
8688 	io_schedule_finish(token);
8689 }
8690 EXPORT_SYMBOL(io_schedule);
8691 
8692 /**
8693  * sys_sched_get_priority_max - return maximum RT priority.
8694  * @policy: scheduling class.
8695  *
8696  * Return: On success, this syscall returns the maximum
8697  * rt_priority that can be used by a given scheduling class.
8698  * On failure, a negative error code is returned.
8699  */
SYSCALL_DEFINE1(sched_get_priority_max,int,policy)8700 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8701 {
8702 	int ret = -EINVAL;
8703 
8704 	switch (policy) {
8705 	case SCHED_FIFO:
8706 	case SCHED_RR:
8707 		ret = MAX_RT_PRIO-1;
8708 		break;
8709 	case SCHED_DEADLINE:
8710 	case SCHED_NORMAL:
8711 	case SCHED_BATCH:
8712 	case SCHED_IDLE:
8713 		ret = 0;
8714 		break;
8715 	}
8716 	return ret;
8717 }
8718 
8719 /**
8720  * sys_sched_get_priority_min - return minimum RT priority.
8721  * @policy: scheduling class.
8722  *
8723  * Return: On success, this syscall returns the minimum
8724  * rt_priority that can be used by a given scheduling class.
8725  * On failure, a negative error code is returned.
8726  */
SYSCALL_DEFINE1(sched_get_priority_min,int,policy)8727 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8728 {
8729 	int ret = -EINVAL;
8730 
8731 	switch (policy) {
8732 	case SCHED_FIFO:
8733 	case SCHED_RR:
8734 		ret = 1;
8735 		break;
8736 	case SCHED_DEADLINE:
8737 	case SCHED_NORMAL:
8738 	case SCHED_BATCH:
8739 	case SCHED_IDLE:
8740 		ret = 0;
8741 	}
8742 	return ret;
8743 }
8744 
sched_rr_get_interval(pid_t pid,struct timespec64 * t)8745 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8746 {
8747 	struct task_struct *p;
8748 	unsigned int time_slice;
8749 	struct rq_flags rf;
8750 	struct rq *rq;
8751 	int retval;
8752 
8753 	if (pid < 0)
8754 		return -EINVAL;
8755 
8756 	retval = -ESRCH;
8757 	rcu_read_lock();
8758 	p = find_process_by_pid(pid);
8759 	if (!p)
8760 		goto out_unlock;
8761 
8762 	retval = security_task_getscheduler(p);
8763 	if (retval)
8764 		goto out_unlock;
8765 
8766 	rq = task_rq_lock(p, &rf);
8767 	time_slice = 0;
8768 	if (p->sched_class->get_rr_interval)
8769 		time_slice = p->sched_class->get_rr_interval(rq, p);
8770 	task_rq_unlock(rq, p, &rf);
8771 
8772 	rcu_read_unlock();
8773 	jiffies_to_timespec64(time_slice, t);
8774 	return 0;
8775 
8776 out_unlock:
8777 	rcu_read_unlock();
8778 	return retval;
8779 }
8780 
8781 /**
8782  * sys_sched_rr_get_interval - return the default timeslice of a process.
8783  * @pid: pid of the process.
8784  * @interval: userspace pointer to the timeslice value.
8785  *
8786  * this syscall writes the default timeslice value of a given process
8787  * into the user-space timespec buffer. A value of '0' means infinity.
8788  *
8789  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8790  * an error code.
8791  */
SYSCALL_DEFINE2(sched_rr_get_interval,pid_t,pid,struct __kernel_timespec __user *,interval)8792 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8793 		struct __kernel_timespec __user *, interval)
8794 {
8795 	struct timespec64 t;
8796 	int retval = sched_rr_get_interval(pid, &t);
8797 
8798 	if (retval == 0)
8799 		retval = put_timespec64(&t, interval);
8800 
8801 	return retval;
8802 }
8803 
8804 #ifdef CONFIG_COMPAT_32BIT_TIME
SYSCALL_DEFINE2(sched_rr_get_interval_time32,pid_t,pid,struct old_timespec32 __user *,interval)8805 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8806 		struct old_timespec32 __user *, interval)
8807 {
8808 	struct timespec64 t;
8809 	int retval = sched_rr_get_interval(pid, &t);
8810 
8811 	if (retval == 0)
8812 		retval = put_old_timespec32(&t, interval);
8813 	return retval;
8814 }
8815 #endif
8816 
sched_show_task(struct task_struct * p)8817 void sched_show_task(struct task_struct *p)
8818 {
8819 	unsigned long free = 0;
8820 	int ppid;
8821 
8822 	if (!try_get_task_stack(p))
8823 		return;
8824 
8825 	pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8826 
8827 	if (task_is_running(p))
8828 		pr_cont("  running task    ");
8829 #ifdef CONFIG_DEBUG_STACK_USAGE
8830 	free = stack_not_used(p);
8831 #endif
8832 	ppid = 0;
8833 	rcu_read_lock();
8834 	if (pid_alive(p))
8835 		ppid = task_pid_nr(rcu_dereference(p->real_parent));
8836 	rcu_read_unlock();
8837 	pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8838 		free, task_pid_nr(p), ppid,
8839 		read_task_thread_flags(p));
8840 
8841 	print_worker_info(KERN_INFO, p);
8842 	print_stop_info(KERN_INFO, p);
8843 	show_stack(p, NULL, KERN_INFO);
8844 	put_task_stack(p);
8845 }
8846 EXPORT_SYMBOL_GPL(sched_show_task);
8847 
8848 static inline bool
state_filter_match(unsigned long state_filter,struct task_struct * p)8849 state_filter_match(unsigned long state_filter, struct task_struct *p)
8850 {
8851 	unsigned int state = READ_ONCE(p->__state);
8852 
8853 	/* no filter, everything matches */
8854 	if (!state_filter)
8855 		return true;
8856 
8857 	/* filter, but doesn't match */
8858 	if (!(state & state_filter))
8859 		return false;
8860 
8861 	/*
8862 	 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8863 	 * TASK_KILLABLE).
8864 	 */
8865 	if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8866 		return false;
8867 
8868 	return true;
8869 }
8870 
8871 
show_state_filter(unsigned int state_filter)8872 void show_state_filter(unsigned int state_filter)
8873 {
8874 	struct task_struct *g, *p;
8875 
8876 	rcu_read_lock();
8877 	for_each_process_thread(g, p) {
8878 		/*
8879 		 * reset the NMI-timeout, listing all files on a slow
8880 		 * console might take a lot of time:
8881 		 * Also, reset softlockup watchdogs on all CPUs, because
8882 		 * another CPU might be blocked waiting for us to process
8883 		 * an IPI.
8884 		 */
8885 		touch_nmi_watchdog();
8886 		touch_all_softlockup_watchdogs();
8887 		if (state_filter_match(state_filter, p))
8888 			sched_show_task(p);
8889 	}
8890 
8891 #ifdef CONFIG_SCHED_DEBUG
8892 	if (!state_filter)
8893 		sysrq_sched_debug_show();
8894 #endif
8895 	rcu_read_unlock();
8896 	/*
8897 	 * Only show locks if all tasks are dumped:
8898 	 */
8899 	if (!state_filter)
8900 		debug_show_all_locks();
8901 }
8902 
8903 /**
8904  * init_idle - set up an idle thread for a given CPU
8905  * @idle: task in question
8906  * @cpu: CPU the idle task belongs to
8907  *
8908  * NOTE: this function does not set the idle thread's NEED_RESCHED
8909  * flag, to make booting more robust.
8910  */
init_idle(struct task_struct * idle,int cpu)8911 void __init init_idle(struct task_struct *idle, int cpu)
8912 {
8913 	struct rq *rq = cpu_rq(cpu);
8914 	unsigned long flags;
8915 
8916 	__sched_fork(0, idle);
8917 
8918 	raw_spin_lock_irqsave(&idle->pi_lock, flags);
8919 	raw_spin_rq_lock(rq);
8920 
8921 	idle->__state = TASK_RUNNING;
8922 	idle->se.exec_start = sched_clock();
8923 	/*
8924 	 * PF_KTHREAD should already be set at this point; regardless, make it
8925 	 * look like a proper per-CPU kthread.
8926 	 */
8927 	idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8928 	kthread_set_per_cpu(idle, cpu);
8929 
8930 #ifdef CONFIG_SMP
8931 	/*
8932 	 * It's possible that init_idle() gets called multiple times on a task,
8933 	 * in that case do_set_cpus_allowed() will not do the right thing.
8934 	 *
8935 	 * And since this is boot we can forgo the serialization.
8936 	 */
8937 	set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8938 #endif
8939 	/*
8940 	 * We're having a chicken and egg problem, even though we are
8941 	 * holding rq->lock, the CPU isn't yet set to this CPU so the
8942 	 * lockdep check in task_group() will fail.
8943 	 *
8944 	 * Similar case to sched_fork(). / Alternatively we could
8945 	 * use task_rq_lock() here and obtain the other rq->lock.
8946 	 *
8947 	 * Silence PROVE_RCU
8948 	 */
8949 	rcu_read_lock();
8950 	__set_task_cpu(idle, cpu);
8951 	rcu_read_unlock();
8952 
8953 	rq->idle = idle;
8954 	rcu_assign_pointer(rq->curr, idle);
8955 	idle->on_rq = TASK_ON_RQ_QUEUED;
8956 #ifdef CONFIG_SMP
8957 	idle->on_cpu = 1;
8958 #endif
8959 	raw_spin_rq_unlock(rq);
8960 	raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8961 
8962 	/* Set the preempt count _outside_ the spinlocks! */
8963 	init_idle_preempt_count(idle, cpu);
8964 
8965 	/*
8966 	 * The idle tasks have their own, simple scheduling class:
8967 	 */
8968 	idle->sched_class = &idle_sched_class;
8969 	ftrace_graph_init_idle_task(idle, cpu);
8970 	vtime_init_idle(idle, cpu);
8971 #ifdef CONFIG_SMP
8972 	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8973 #endif
8974 }
8975 
8976 #ifdef CONFIG_SMP
8977 
cpuset_cpumask_can_shrink(const struct cpumask * cur,const struct cpumask * trial)8978 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8979 			      const struct cpumask *trial)
8980 {
8981 	int ret = 1;
8982 
8983 	if (cpumask_empty(cur))
8984 		return ret;
8985 
8986 	ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8987 
8988 	return ret;
8989 }
8990 
task_can_attach(struct task_struct * p,const struct cpumask * cs_effective_cpus)8991 int task_can_attach(struct task_struct *p,
8992 		    const struct cpumask *cs_effective_cpus)
8993 {
8994 	int ret = 0;
8995 
8996 	/*
8997 	 * Kthreads which disallow setaffinity shouldn't be moved
8998 	 * to a new cpuset; we don't want to change their CPU
8999 	 * affinity and isolating such threads by their set of
9000 	 * allowed nodes is unnecessary.  Thus, cpusets are not
9001 	 * applicable for such threads.  This prevents checking for
9002 	 * success of set_cpus_allowed_ptr() on all attached tasks
9003 	 * before cpus_mask may be changed.
9004 	 */
9005 	if (p->flags & PF_NO_SETAFFINITY) {
9006 		ret = -EINVAL;
9007 		goto out;
9008 	}
9009 
9010 	if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
9011 					      cs_effective_cpus)) {
9012 		int cpu = cpumask_any_and(cpu_active_mask, cs_effective_cpus);
9013 
9014 		if (unlikely(cpu >= nr_cpu_ids))
9015 			return -EINVAL;
9016 		ret = dl_cpu_busy(cpu, p);
9017 	}
9018 
9019 out:
9020 	return ret;
9021 }
9022 
9023 bool sched_smp_initialized __read_mostly;
9024 
9025 #ifdef CONFIG_NUMA_BALANCING
9026 /* Migrate current task p to target_cpu */
migrate_task_to(struct task_struct * p,int target_cpu)9027 int migrate_task_to(struct task_struct *p, int target_cpu)
9028 {
9029 	struct migration_arg arg = { p, target_cpu };
9030 	int curr_cpu = task_cpu(p);
9031 
9032 	if (curr_cpu == target_cpu)
9033 		return 0;
9034 
9035 	if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9036 		return -EINVAL;
9037 
9038 	/* TODO: This is not properly updating schedstats */
9039 
9040 	trace_sched_move_numa(p, curr_cpu, target_cpu);
9041 	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9042 }
9043 
9044 /*
9045  * Requeue a task on a given node and accurately track the number of NUMA
9046  * tasks on the runqueues
9047  */
sched_setnuma(struct task_struct * p,int nid)9048 void sched_setnuma(struct task_struct *p, int nid)
9049 {
9050 	bool queued, running;
9051 	struct rq_flags rf;
9052 	struct rq *rq;
9053 
9054 	rq = task_rq_lock(p, &rf);
9055 	queued = task_on_rq_queued(p);
9056 	running = task_current(rq, p);
9057 
9058 	if (queued)
9059 		dequeue_task(rq, p, DEQUEUE_SAVE);
9060 	if (running)
9061 		put_prev_task(rq, p);
9062 
9063 	p->numa_preferred_nid = nid;
9064 
9065 	if (queued)
9066 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9067 	if (running)
9068 		set_next_task(rq, p);
9069 	task_rq_unlock(rq, p, &rf);
9070 }
9071 #endif /* CONFIG_NUMA_BALANCING */
9072 
9073 #ifdef CONFIG_HOTPLUG_CPU
9074 /*
9075  * Ensure that the idle task is using init_mm right before its CPU goes
9076  * offline.
9077  */
idle_task_exit(void)9078 void idle_task_exit(void)
9079 {
9080 	struct mm_struct *mm = current->active_mm;
9081 
9082 	BUG_ON(cpu_online(smp_processor_id()));
9083 	BUG_ON(current != this_rq()->idle);
9084 
9085 	if (mm != &init_mm) {
9086 		switch_mm(mm, &init_mm, current);
9087 		finish_arch_post_lock_switch();
9088 	}
9089 
9090 	/* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9091 }
9092 
__balance_push_cpu_stop(void * arg)9093 static int __balance_push_cpu_stop(void *arg)
9094 {
9095 	struct task_struct *p = arg;
9096 	struct rq *rq = this_rq();
9097 	struct rq_flags rf;
9098 	int cpu;
9099 
9100 	raw_spin_lock_irq(&p->pi_lock);
9101 	rq_lock(rq, &rf);
9102 
9103 	update_rq_clock(rq);
9104 
9105 	if (task_rq(p) == rq && task_on_rq_queued(p)) {
9106 		cpu = select_fallback_rq(rq->cpu, p);
9107 		rq = __migrate_task(rq, &rf, p, cpu);
9108 	}
9109 
9110 	rq_unlock(rq, &rf);
9111 	raw_spin_unlock_irq(&p->pi_lock);
9112 
9113 	put_task_struct(p);
9114 
9115 	return 0;
9116 }
9117 
9118 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9119 
9120 /*
9121  * Ensure we only run per-cpu kthreads once the CPU goes !active.
9122  *
9123  * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9124  * effective when the hotplug motion is down.
9125  */
balance_push(struct rq * rq)9126 static void balance_push(struct rq *rq)
9127 {
9128 	struct task_struct *push_task = rq->curr;
9129 
9130 	lockdep_assert_rq_held(rq);
9131 
9132 	/*
9133 	 * Ensure the thing is persistent until balance_push_set(.on = false);
9134 	 */
9135 	rq->balance_callback = &balance_push_callback;
9136 
9137 	/*
9138 	 * Only active while going offline and when invoked on the outgoing
9139 	 * CPU.
9140 	 */
9141 	if (!cpu_dying(rq->cpu) || rq != this_rq())
9142 		return;
9143 
9144 	/*
9145 	 * Both the cpu-hotplug and stop task are in this case and are
9146 	 * required to complete the hotplug process.
9147 	 */
9148 	if (kthread_is_per_cpu(push_task) ||
9149 	    is_migration_disabled(push_task)) {
9150 
9151 		/*
9152 		 * If this is the idle task on the outgoing CPU try to wake
9153 		 * up the hotplug control thread which might wait for the
9154 		 * last task to vanish. The rcuwait_active() check is
9155 		 * accurate here because the waiter is pinned on this CPU
9156 		 * and can't obviously be running in parallel.
9157 		 *
9158 		 * On RT kernels this also has to check whether there are
9159 		 * pinned and scheduled out tasks on the runqueue. They
9160 		 * need to leave the migrate disabled section first.
9161 		 */
9162 		if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9163 		    rcuwait_active(&rq->hotplug_wait)) {
9164 			raw_spin_rq_unlock(rq);
9165 			rcuwait_wake_up(&rq->hotplug_wait);
9166 			raw_spin_rq_lock(rq);
9167 		}
9168 		return;
9169 	}
9170 
9171 	get_task_struct(push_task);
9172 	/*
9173 	 * Temporarily drop rq->lock such that we can wake-up the stop task.
9174 	 * Both preemption and IRQs are still disabled.
9175 	 */
9176 	raw_spin_rq_unlock(rq);
9177 	stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9178 			    this_cpu_ptr(&push_work));
9179 	/*
9180 	 * At this point need_resched() is true and we'll take the loop in
9181 	 * schedule(). The next pick is obviously going to be the stop task
9182 	 * which kthread_is_per_cpu() and will push this task away.
9183 	 */
9184 	raw_spin_rq_lock(rq);
9185 }
9186 
balance_push_set(int cpu,bool on)9187 static void balance_push_set(int cpu, bool on)
9188 {
9189 	struct rq *rq = cpu_rq(cpu);
9190 	struct rq_flags rf;
9191 
9192 	rq_lock_irqsave(rq, &rf);
9193 	if (on) {
9194 		WARN_ON_ONCE(rq->balance_callback);
9195 		rq->balance_callback = &balance_push_callback;
9196 	} else if (rq->balance_callback == &balance_push_callback) {
9197 		rq->balance_callback = NULL;
9198 	}
9199 	rq_unlock_irqrestore(rq, &rf);
9200 }
9201 
9202 /*
9203  * Invoked from a CPUs hotplug control thread after the CPU has been marked
9204  * inactive. All tasks which are not per CPU kernel threads are either
9205  * pushed off this CPU now via balance_push() or placed on a different CPU
9206  * during wakeup. Wait until the CPU is quiescent.
9207  */
balance_hotplug_wait(void)9208 static void balance_hotplug_wait(void)
9209 {
9210 	struct rq *rq = this_rq();
9211 
9212 	rcuwait_wait_event(&rq->hotplug_wait,
9213 			   rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9214 			   TASK_UNINTERRUPTIBLE);
9215 }
9216 
9217 #else
9218 
balance_push(struct rq * rq)9219 static inline void balance_push(struct rq *rq)
9220 {
9221 }
9222 
balance_push_set(int cpu,bool on)9223 static inline void balance_push_set(int cpu, bool on)
9224 {
9225 }
9226 
balance_hotplug_wait(void)9227 static inline void balance_hotplug_wait(void)
9228 {
9229 }
9230 
9231 #endif /* CONFIG_HOTPLUG_CPU */
9232 
set_rq_online(struct rq * rq)9233 void set_rq_online(struct rq *rq)
9234 {
9235 	if (!rq->online) {
9236 		const struct sched_class *class;
9237 
9238 		cpumask_set_cpu(rq->cpu, rq->rd->online);
9239 		rq->online = 1;
9240 
9241 		for_each_class(class) {
9242 			if (class->rq_online)
9243 				class->rq_online(rq);
9244 		}
9245 	}
9246 }
9247 
set_rq_offline(struct rq * rq)9248 void set_rq_offline(struct rq *rq)
9249 {
9250 	if (rq->online) {
9251 		const struct sched_class *class;
9252 
9253 		for_each_class(class) {
9254 			if (class->rq_offline)
9255 				class->rq_offline(rq);
9256 		}
9257 
9258 		cpumask_clear_cpu(rq->cpu, rq->rd->online);
9259 		rq->online = 0;
9260 	}
9261 }
9262 
9263 /*
9264  * used to mark begin/end of suspend/resume:
9265  */
9266 static int num_cpus_frozen;
9267 
9268 /*
9269  * Update cpusets according to cpu_active mask.  If cpusets are
9270  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9271  * around partition_sched_domains().
9272  *
9273  * If we come here as part of a suspend/resume, don't touch cpusets because we
9274  * want to restore it back to its original state upon resume anyway.
9275  */
cpuset_cpu_active(void)9276 static void cpuset_cpu_active(void)
9277 {
9278 	if (cpuhp_tasks_frozen) {
9279 		/*
9280 		 * num_cpus_frozen tracks how many CPUs are involved in suspend
9281 		 * resume sequence. As long as this is not the last online
9282 		 * operation in the resume sequence, just build a single sched
9283 		 * domain, ignoring cpusets.
9284 		 */
9285 		partition_sched_domains(1, NULL, NULL);
9286 		if (--num_cpus_frozen)
9287 			return;
9288 		/*
9289 		 * This is the last CPU online operation. So fall through and
9290 		 * restore the original sched domains by considering the
9291 		 * cpuset configurations.
9292 		 */
9293 		cpuset_force_rebuild();
9294 	}
9295 	cpuset_update_active_cpus();
9296 }
9297 
cpuset_cpu_inactive(unsigned int cpu)9298 static int cpuset_cpu_inactive(unsigned int cpu)
9299 {
9300 	if (!cpuhp_tasks_frozen) {
9301 		int ret = dl_cpu_busy(cpu, NULL);
9302 
9303 		if (ret)
9304 			return ret;
9305 		cpuset_update_active_cpus();
9306 	} else {
9307 		num_cpus_frozen++;
9308 		partition_sched_domains(1, NULL, NULL);
9309 	}
9310 	return 0;
9311 }
9312 
sched_cpu_activate(unsigned int cpu)9313 int sched_cpu_activate(unsigned int cpu)
9314 {
9315 	struct rq *rq = cpu_rq(cpu);
9316 	struct rq_flags rf;
9317 
9318 	/*
9319 	 * Clear the balance_push callback and prepare to schedule
9320 	 * regular tasks.
9321 	 */
9322 	balance_push_set(cpu, false);
9323 
9324 #ifdef CONFIG_SCHED_SMT
9325 	/*
9326 	 * When going up, increment the number of cores with SMT present.
9327 	 */
9328 	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9329 		static_branch_inc_cpuslocked(&sched_smt_present);
9330 #endif
9331 	set_cpu_active(cpu, true);
9332 
9333 	if (sched_smp_initialized) {
9334 		sched_update_numa(cpu, true);
9335 		sched_domains_numa_masks_set(cpu);
9336 		cpuset_cpu_active();
9337 	}
9338 
9339 	/*
9340 	 * Put the rq online, if not already. This happens:
9341 	 *
9342 	 * 1) In the early boot process, because we build the real domains
9343 	 *    after all CPUs have been brought up.
9344 	 *
9345 	 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9346 	 *    domains.
9347 	 */
9348 	rq_lock_irqsave(rq, &rf);
9349 	if (rq->rd) {
9350 		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9351 		set_rq_online(rq);
9352 	}
9353 	rq_unlock_irqrestore(rq, &rf);
9354 
9355 	return 0;
9356 }
9357 
sched_cpu_deactivate(unsigned int cpu)9358 int sched_cpu_deactivate(unsigned int cpu)
9359 {
9360 	struct rq *rq = cpu_rq(cpu);
9361 	struct rq_flags rf;
9362 	int ret;
9363 
9364 	/*
9365 	 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9366 	 * load balancing when not active
9367 	 */
9368 	nohz_balance_exit_idle(rq);
9369 
9370 	set_cpu_active(cpu, false);
9371 
9372 	/*
9373 	 * From this point forward, this CPU will refuse to run any task that
9374 	 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9375 	 * push those tasks away until this gets cleared, see
9376 	 * sched_cpu_dying().
9377 	 */
9378 	balance_push_set(cpu, true);
9379 
9380 	/*
9381 	 * We've cleared cpu_active_mask / set balance_push, wait for all
9382 	 * preempt-disabled and RCU users of this state to go away such that
9383 	 * all new such users will observe it.
9384 	 *
9385 	 * Specifically, we rely on ttwu to no longer target this CPU, see
9386 	 * ttwu_queue_cond() and is_cpu_allowed().
9387 	 *
9388 	 * Do sync before park smpboot threads to take care the rcu boost case.
9389 	 */
9390 	synchronize_rcu();
9391 
9392 	rq_lock_irqsave(rq, &rf);
9393 	if (rq->rd) {
9394 		update_rq_clock(rq);
9395 		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9396 		set_rq_offline(rq);
9397 	}
9398 	rq_unlock_irqrestore(rq, &rf);
9399 
9400 #ifdef CONFIG_SCHED_SMT
9401 	/*
9402 	 * When going down, decrement the number of cores with SMT present.
9403 	 */
9404 	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9405 		static_branch_dec_cpuslocked(&sched_smt_present);
9406 
9407 	sched_core_cpu_deactivate(cpu);
9408 #endif
9409 
9410 	if (!sched_smp_initialized)
9411 		return 0;
9412 
9413 	sched_update_numa(cpu, false);
9414 	ret = cpuset_cpu_inactive(cpu);
9415 	if (ret) {
9416 		balance_push_set(cpu, false);
9417 		set_cpu_active(cpu, true);
9418 		sched_update_numa(cpu, true);
9419 		return ret;
9420 	}
9421 	sched_domains_numa_masks_clear(cpu);
9422 	return 0;
9423 }
9424 
sched_rq_cpu_starting(unsigned int cpu)9425 static void sched_rq_cpu_starting(unsigned int cpu)
9426 {
9427 	struct rq *rq = cpu_rq(cpu);
9428 
9429 	rq->calc_load_update = calc_load_update;
9430 	update_max_interval();
9431 }
9432 
sched_cpu_starting(unsigned int cpu)9433 int sched_cpu_starting(unsigned int cpu)
9434 {
9435 	sched_core_cpu_starting(cpu);
9436 	sched_rq_cpu_starting(cpu);
9437 	sched_tick_start(cpu);
9438 	return 0;
9439 }
9440 
9441 #ifdef CONFIG_HOTPLUG_CPU
9442 
9443 /*
9444  * Invoked immediately before the stopper thread is invoked to bring the
9445  * CPU down completely. At this point all per CPU kthreads except the
9446  * hotplug thread (current) and the stopper thread (inactive) have been
9447  * either parked or have been unbound from the outgoing CPU. Ensure that
9448  * any of those which might be on the way out are gone.
9449  *
9450  * If after this point a bound task is being woken on this CPU then the
9451  * responsible hotplug callback has failed to do it's job.
9452  * sched_cpu_dying() will catch it with the appropriate fireworks.
9453  */
sched_cpu_wait_empty(unsigned int cpu)9454 int sched_cpu_wait_empty(unsigned int cpu)
9455 {
9456 	balance_hotplug_wait();
9457 	return 0;
9458 }
9459 
9460 /*
9461  * Since this CPU is going 'away' for a while, fold any nr_active delta we
9462  * might have. Called from the CPU stopper task after ensuring that the
9463  * stopper is the last running task on the CPU, so nr_active count is
9464  * stable. We need to take the teardown thread which is calling this into
9465  * account, so we hand in adjust = 1 to the load calculation.
9466  *
9467  * Also see the comment "Global load-average calculations".
9468  */
calc_load_migrate(struct rq * rq)9469 static void calc_load_migrate(struct rq *rq)
9470 {
9471 	long delta = calc_load_fold_active(rq, 1);
9472 
9473 	if (delta)
9474 		atomic_long_add(delta, &calc_load_tasks);
9475 }
9476 
dump_rq_tasks(struct rq * rq,const char * loglvl)9477 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9478 {
9479 	struct task_struct *g, *p;
9480 	int cpu = cpu_of(rq);
9481 
9482 	lockdep_assert_rq_held(rq);
9483 
9484 	printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9485 	for_each_process_thread(g, p) {
9486 		if (task_cpu(p) != cpu)
9487 			continue;
9488 
9489 		if (!task_on_rq_queued(p))
9490 			continue;
9491 
9492 		printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9493 	}
9494 }
9495 
sched_cpu_dying(unsigned int cpu)9496 int sched_cpu_dying(unsigned int cpu)
9497 {
9498 	struct rq *rq = cpu_rq(cpu);
9499 	struct rq_flags rf;
9500 
9501 	/* Handle pending wakeups and then migrate everything off */
9502 	sched_tick_stop(cpu);
9503 
9504 	rq_lock_irqsave(rq, &rf);
9505 	if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9506 		WARN(true, "Dying CPU not properly vacated!");
9507 		dump_rq_tasks(rq, KERN_WARNING);
9508 	}
9509 	rq_unlock_irqrestore(rq, &rf);
9510 
9511 	calc_load_migrate(rq);
9512 	update_max_interval();
9513 	hrtick_clear(rq);
9514 	sched_core_cpu_dying(cpu);
9515 	return 0;
9516 }
9517 #endif
9518 
sched_init_smp(void)9519 void __init sched_init_smp(void)
9520 {
9521 	sched_init_numa(NUMA_NO_NODE);
9522 
9523 	/*
9524 	 * There's no userspace yet to cause hotplug operations; hence all the
9525 	 * CPU masks are stable and all blatant races in the below code cannot
9526 	 * happen.
9527 	 */
9528 	mutex_lock(&sched_domains_mutex);
9529 	sched_init_domains(cpu_active_mask);
9530 	mutex_unlock(&sched_domains_mutex);
9531 
9532 	/* Move init over to a non-isolated CPU */
9533 	if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9534 		BUG();
9535 	current->flags &= ~PF_NO_SETAFFINITY;
9536 	sched_init_granularity();
9537 
9538 	init_sched_rt_class();
9539 	init_sched_dl_class();
9540 
9541 	sched_smp_initialized = true;
9542 }
9543 
migration_init(void)9544 static int __init migration_init(void)
9545 {
9546 	sched_cpu_starting(smp_processor_id());
9547 	return 0;
9548 }
9549 early_initcall(migration_init);
9550 
9551 #else
sched_init_smp(void)9552 void __init sched_init_smp(void)
9553 {
9554 	sched_init_granularity();
9555 }
9556 #endif /* CONFIG_SMP */
9557 
in_sched_functions(unsigned long addr)9558 int in_sched_functions(unsigned long addr)
9559 {
9560 	return in_lock_functions(addr) ||
9561 		(addr >= (unsigned long)__sched_text_start
9562 		&& addr < (unsigned long)__sched_text_end);
9563 }
9564 
9565 #ifdef CONFIG_CGROUP_SCHED
9566 /*
9567  * Default task group.
9568  * Every task in system belongs to this group at bootup.
9569  */
9570 struct task_group root_task_group;
9571 LIST_HEAD(task_groups);
9572 
9573 /* Cacheline aligned slab cache for task_group */
9574 static struct kmem_cache *task_group_cache __read_mostly;
9575 #endif
9576 
9577 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9578 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
9579 
sched_init(void)9580 void __init sched_init(void)
9581 {
9582 	unsigned long ptr = 0;
9583 	int i;
9584 
9585 	/* Make sure the linker didn't screw up */
9586 	BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9587 	       &fair_sched_class != &rt_sched_class + 1 ||
9588 	       &rt_sched_class   != &dl_sched_class + 1);
9589 #ifdef CONFIG_SMP
9590 	BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9591 #endif
9592 
9593 	wait_bit_init();
9594 
9595 #ifdef CONFIG_FAIR_GROUP_SCHED
9596 	ptr += 2 * nr_cpu_ids * sizeof(void **);
9597 #endif
9598 #ifdef CONFIG_RT_GROUP_SCHED
9599 	ptr += 2 * nr_cpu_ids * sizeof(void **);
9600 #endif
9601 	if (ptr) {
9602 		ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9603 
9604 #ifdef CONFIG_FAIR_GROUP_SCHED
9605 		root_task_group.se = (struct sched_entity **)ptr;
9606 		ptr += nr_cpu_ids * sizeof(void **);
9607 
9608 		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9609 		ptr += nr_cpu_ids * sizeof(void **);
9610 
9611 		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9612 		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9613 #endif /* CONFIG_FAIR_GROUP_SCHED */
9614 #ifdef CONFIG_RT_GROUP_SCHED
9615 		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9616 		ptr += nr_cpu_ids * sizeof(void **);
9617 
9618 		root_task_group.rt_rq = (struct rt_rq **)ptr;
9619 		ptr += nr_cpu_ids * sizeof(void **);
9620 
9621 #endif /* CONFIG_RT_GROUP_SCHED */
9622 	}
9623 #ifdef CONFIG_CPUMASK_OFFSTACK
9624 	for_each_possible_cpu(i) {
9625 		per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9626 			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9627 		per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
9628 			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9629 	}
9630 #endif /* CONFIG_CPUMASK_OFFSTACK */
9631 
9632 	init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9633 
9634 #ifdef CONFIG_SMP
9635 	init_defrootdomain();
9636 #endif
9637 
9638 #ifdef CONFIG_RT_GROUP_SCHED
9639 	init_rt_bandwidth(&root_task_group.rt_bandwidth,
9640 			global_rt_period(), global_rt_runtime());
9641 #endif /* CONFIG_RT_GROUP_SCHED */
9642 
9643 #ifdef CONFIG_CGROUP_SCHED
9644 	task_group_cache = KMEM_CACHE(task_group, 0);
9645 
9646 	list_add(&root_task_group.list, &task_groups);
9647 	INIT_LIST_HEAD(&root_task_group.children);
9648 	INIT_LIST_HEAD(&root_task_group.siblings);
9649 	autogroup_init(&init_task);
9650 #endif /* CONFIG_CGROUP_SCHED */
9651 
9652 	for_each_possible_cpu(i) {
9653 		struct rq *rq;
9654 
9655 		rq = cpu_rq(i);
9656 		raw_spin_lock_init(&rq->__lock);
9657 		rq->nr_running = 0;
9658 		rq->calc_load_active = 0;
9659 		rq->calc_load_update = jiffies + LOAD_FREQ;
9660 		init_cfs_rq(&rq->cfs);
9661 		init_rt_rq(&rq->rt);
9662 		init_dl_rq(&rq->dl);
9663 #ifdef CONFIG_FAIR_GROUP_SCHED
9664 		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9665 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9666 		/*
9667 		 * How much CPU bandwidth does root_task_group get?
9668 		 *
9669 		 * In case of task-groups formed thr' the cgroup filesystem, it
9670 		 * gets 100% of the CPU resources in the system. This overall
9671 		 * system CPU resource is divided among the tasks of
9672 		 * root_task_group and its child task-groups in a fair manner,
9673 		 * based on each entity's (task or task-group's) weight
9674 		 * (se->load.weight).
9675 		 *
9676 		 * In other words, if root_task_group has 10 tasks of weight
9677 		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9678 		 * then A0's share of the CPU resource is:
9679 		 *
9680 		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9681 		 *
9682 		 * We achieve this by letting root_task_group's tasks sit
9683 		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9684 		 */
9685 		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9686 #endif /* CONFIG_FAIR_GROUP_SCHED */
9687 
9688 		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9689 #ifdef CONFIG_RT_GROUP_SCHED
9690 		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9691 #endif
9692 #ifdef CONFIG_SMP
9693 		rq->sd = NULL;
9694 		rq->rd = NULL;
9695 		rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9696 		rq->balance_callback = &balance_push_callback;
9697 		rq->active_balance = 0;
9698 		rq->next_balance = jiffies;
9699 		rq->push_cpu = 0;
9700 		rq->cpu = i;
9701 		rq->online = 0;
9702 		rq->idle_stamp = 0;
9703 		rq->avg_idle = 2*sysctl_sched_migration_cost;
9704 		rq->wake_stamp = jiffies;
9705 		rq->wake_avg_idle = rq->avg_idle;
9706 		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9707 
9708 		INIT_LIST_HEAD(&rq->cfs_tasks);
9709 
9710 		rq_attach_root(rq, &def_root_domain);
9711 #ifdef CONFIG_NO_HZ_COMMON
9712 		rq->last_blocked_load_update_tick = jiffies;
9713 		atomic_set(&rq->nohz_flags, 0);
9714 
9715 		INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9716 #endif
9717 #ifdef CONFIG_HOTPLUG_CPU
9718 		rcuwait_init(&rq->hotplug_wait);
9719 #endif
9720 #endif /* CONFIG_SMP */
9721 		hrtick_rq_init(rq);
9722 		atomic_set(&rq->nr_iowait, 0);
9723 
9724 #ifdef CONFIG_SCHED_CORE
9725 		rq->core = rq;
9726 		rq->core_pick = NULL;
9727 		rq->core_enabled = 0;
9728 		rq->core_tree = RB_ROOT;
9729 		rq->core_forceidle_count = 0;
9730 		rq->core_forceidle_occupation = 0;
9731 		rq->core_forceidle_start = 0;
9732 
9733 		rq->core_cookie = 0UL;
9734 #endif
9735 	}
9736 
9737 	set_load_weight(&init_task, false);
9738 
9739 	/*
9740 	 * The boot idle thread does lazy MMU switching as well:
9741 	 */
9742 	mmgrab(&init_mm);
9743 	enter_lazy_tlb(&init_mm, current);
9744 
9745 	/*
9746 	 * The idle task doesn't need the kthread struct to function, but it
9747 	 * is dressed up as a per-CPU kthread and thus needs to play the part
9748 	 * if we want to avoid special-casing it in code that deals with per-CPU
9749 	 * kthreads.
9750 	 */
9751 	WARN_ON(!set_kthread_struct(current));
9752 
9753 	/*
9754 	 * Make us the idle thread. Technically, schedule() should not be
9755 	 * called from this thread, however somewhere below it might be,
9756 	 * but because we are the idle thread, we just pick up running again
9757 	 * when this runqueue becomes "idle".
9758 	 */
9759 	init_idle(current, smp_processor_id());
9760 
9761 	calc_load_update = jiffies + LOAD_FREQ;
9762 
9763 #ifdef CONFIG_SMP
9764 	idle_thread_set_boot_cpu();
9765 	balance_push_set(smp_processor_id(), false);
9766 #endif
9767 	init_sched_fair_class();
9768 
9769 	psi_init();
9770 
9771 	init_uclamp();
9772 
9773 	preempt_dynamic_init();
9774 
9775 	scheduler_running = 1;
9776 }
9777 
9778 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9779 
__might_sleep(const char * file,int line)9780 void __might_sleep(const char *file, int line)
9781 {
9782 	unsigned int state = get_current_state();
9783 	/*
9784 	 * Blocking primitives will set (and therefore destroy) current->state,
9785 	 * since we will exit with TASK_RUNNING make sure we enter with it,
9786 	 * otherwise we will destroy state.
9787 	 */
9788 	WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9789 			"do not call blocking ops when !TASK_RUNNING; "
9790 			"state=%x set at [<%p>] %pS\n", state,
9791 			(void *)current->task_state_change,
9792 			(void *)current->task_state_change);
9793 
9794 	__might_resched(file, line, 0);
9795 }
9796 EXPORT_SYMBOL(__might_sleep);
9797 
print_preempt_disable_ip(int preempt_offset,unsigned long ip)9798 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9799 {
9800 	if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9801 		return;
9802 
9803 	if (preempt_count() == preempt_offset)
9804 		return;
9805 
9806 	pr_err("Preemption disabled at:");
9807 	print_ip_sym(KERN_ERR, ip);
9808 }
9809 
resched_offsets_ok(unsigned int offsets)9810 static inline bool resched_offsets_ok(unsigned int offsets)
9811 {
9812 	unsigned int nested = preempt_count();
9813 
9814 	nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9815 
9816 	return nested == offsets;
9817 }
9818 
__might_resched(const char * file,int line,unsigned int offsets)9819 void __might_resched(const char *file, int line, unsigned int offsets)
9820 {
9821 	/* Ratelimiting timestamp: */
9822 	static unsigned long prev_jiffy;
9823 
9824 	unsigned long preempt_disable_ip;
9825 
9826 	/* WARN_ON_ONCE() by default, no rate limit required: */
9827 	rcu_sleep_check();
9828 
9829 	if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9830 	     !is_idle_task(current) && !current->non_block_count) ||
9831 	    system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9832 	    oops_in_progress)
9833 		return;
9834 
9835 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9836 		return;
9837 	prev_jiffy = jiffies;
9838 
9839 	/* Save this before calling printk(), since that will clobber it: */
9840 	preempt_disable_ip = get_preempt_disable_ip(current);
9841 
9842 	pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9843 	       file, line);
9844 	pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9845 	       in_atomic(), irqs_disabled(), current->non_block_count,
9846 	       current->pid, current->comm);
9847 	pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9848 	       offsets & MIGHT_RESCHED_PREEMPT_MASK);
9849 
9850 	if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9851 		pr_err("RCU nest depth: %d, expected: %u\n",
9852 		       rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9853 	}
9854 
9855 	if (task_stack_end_corrupted(current))
9856 		pr_emerg("Thread overran stack, or stack corrupted\n");
9857 
9858 	debug_show_held_locks(current);
9859 	if (irqs_disabled())
9860 		print_irqtrace_events(current);
9861 
9862 	print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9863 				 preempt_disable_ip);
9864 
9865 	dump_stack();
9866 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9867 }
9868 EXPORT_SYMBOL(__might_resched);
9869 
__cant_sleep(const char * file,int line,int preempt_offset)9870 void __cant_sleep(const char *file, int line, int preempt_offset)
9871 {
9872 	static unsigned long prev_jiffy;
9873 
9874 	if (irqs_disabled())
9875 		return;
9876 
9877 	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9878 		return;
9879 
9880 	if (preempt_count() > preempt_offset)
9881 		return;
9882 
9883 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9884 		return;
9885 	prev_jiffy = jiffies;
9886 
9887 	printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9888 	printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9889 			in_atomic(), irqs_disabled(),
9890 			current->pid, current->comm);
9891 
9892 	debug_show_held_locks(current);
9893 	dump_stack();
9894 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9895 }
9896 EXPORT_SYMBOL_GPL(__cant_sleep);
9897 
9898 #ifdef CONFIG_SMP
__cant_migrate(const char * file,int line)9899 void __cant_migrate(const char *file, int line)
9900 {
9901 	static unsigned long prev_jiffy;
9902 
9903 	if (irqs_disabled())
9904 		return;
9905 
9906 	if (is_migration_disabled(current))
9907 		return;
9908 
9909 	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9910 		return;
9911 
9912 	if (preempt_count() > 0)
9913 		return;
9914 
9915 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9916 		return;
9917 	prev_jiffy = jiffies;
9918 
9919 	pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9920 	pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9921 	       in_atomic(), irqs_disabled(), is_migration_disabled(current),
9922 	       current->pid, current->comm);
9923 
9924 	debug_show_held_locks(current);
9925 	dump_stack();
9926 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9927 }
9928 EXPORT_SYMBOL_GPL(__cant_migrate);
9929 #endif
9930 #endif
9931 
9932 #ifdef CONFIG_MAGIC_SYSRQ
normalize_rt_tasks(void)9933 void normalize_rt_tasks(void)
9934 {
9935 	struct task_struct *g, *p;
9936 	struct sched_attr attr = {
9937 		.sched_policy = SCHED_NORMAL,
9938 	};
9939 
9940 	read_lock(&tasklist_lock);
9941 	for_each_process_thread(g, p) {
9942 		/*
9943 		 * Only normalize user tasks:
9944 		 */
9945 		if (p->flags & PF_KTHREAD)
9946 			continue;
9947 
9948 		p->se.exec_start = 0;
9949 		schedstat_set(p->stats.wait_start,  0);
9950 		schedstat_set(p->stats.sleep_start, 0);
9951 		schedstat_set(p->stats.block_start, 0);
9952 
9953 		if (!dl_task(p) && !rt_task(p)) {
9954 			/*
9955 			 * Renice negative nice level userspace
9956 			 * tasks back to 0:
9957 			 */
9958 			if (task_nice(p) < 0)
9959 				set_user_nice(p, 0);
9960 			continue;
9961 		}
9962 
9963 		__sched_setscheduler(p, &attr, false, false);
9964 	}
9965 	read_unlock(&tasklist_lock);
9966 }
9967 
9968 #endif /* CONFIG_MAGIC_SYSRQ */
9969 
9970 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9971 /*
9972  * These functions are only useful for the IA64 MCA handling, or kdb.
9973  *
9974  * They can only be called when the whole system has been
9975  * stopped - every CPU needs to be quiescent, and no scheduling
9976  * activity can take place. Using them for anything else would
9977  * be a serious bug, and as a result, they aren't even visible
9978  * under any other configuration.
9979  */
9980 
9981 /**
9982  * curr_task - return the current task for a given CPU.
9983  * @cpu: the processor in question.
9984  *
9985  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9986  *
9987  * Return: The current task for @cpu.
9988  */
curr_task(int cpu)9989 struct task_struct *curr_task(int cpu)
9990 {
9991 	return cpu_curr(cpu);
9992 }
9993 
9994 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9995 
9996 #ifdef CONFIG_IA64
9997 /**
9998  * ia64_set_curr_task - set the current task for a given CPU.
9999  * @cpu: the processor in question.
10000  * @p: the task pointer to set.
10001  *
10002  * Description: This function must only be used when non-maskable interrupts
10003  * are serviced on a separate stack. It allows the architecture to switch the
10004  * notion of the current task on a CPU in a non-blocking manner. This function
10005  * must be called with all CPU's synchronized, and interrupts disabled, the
10006  * and caller must save the original value of the current task (see
10007  * curr_task() above) and restore that value before reenabling interrupts and
10008  * re-starting the system.
10009  *
10010  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10011  */
ia64_set_curr_task(int cpu,struct task_struct * p)10012 void ia64_set_curr_task(int cpu, struct task_struct *p)
10013 {
10014 	cpu_curr(cpu) = p;
10015 }
10016 
10017 #endif
10018 
10019 #ifdef CONFIG_CGROUP_SCHED
10020 /* task_group_lock serializes the addition/removal of task groups */
10021 static DEFINE_SPINLOCK(task_group_lock);
10022 
alloc_uclamp_sched_group(struct task_group * tg,struct task_group * parent)10023 static inline void alloc_uclamp_sched_group(struct task_group *tg,
10024 					    struct task_group *parent)
10025 {
10026 #ifdef CONFIG_UCLAMP_TASK_GROUP
10027 	enum uclamp_id clamp_id;
10028 
10029 	for_each_clamp_id(clamp_id) {
10030 		uclamp_se_set(&tg->uclamp_req[clamp_id],
10031 			      uclamp_none(clamp_id), false);
10032 		tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10033 	}
10034 #endif
10035 }
10036 
sched_free_group(struct task_group * tg)10037 static void sched_free_group(struct task_group *tg)
10038 {
10039 	free_fair_sched_group(tg);
10040 	free_rt_sched_group(tg);
10041 	autogroup_free(tg);
10042 	kmem_cache_free(task_group_cache, tg);
10043 }
10044 
sched_free_group_rcu(struct rcu_head * rcu)10045 static void sched_free_group_rcu(struct rcu_head *rcu)
10046 {
10047 	sched_free_group(container_of(rcu, struct task_group, rcu));
10048 }
10049 
sched_unregister_group(struct task_group * tg)10050 static void sched_unregister_group(struct task_group *tg)
10051 {
10052 	unregister_fair_sched_group(tg);
10053 	unregister_rt_sched_group(tg);
10054 	/*
10055 	 * We have to wait for yet another RCU grace period to expire, as
10056 	 * print_cfs_stats() might run concurrently.
10057 	 */
10058 	call_rcu(&tg->rcu, sched_free_group_rcu);
10059 }
10060 
10061 /* allocate runqueue etc for a new task group */
sched_create_group(struct task_group * parent)10062 struct task_group *sched_create_group(struct task_group *parent)
10063 {
10064 	struct task_group *tg;
10065 
10066 	tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10067 	if (!tg)
10068 		return ERR_PTR(-ENOMEM);
10069 
10070 	if (!alloc_fair_sched_group(tg, parent))
10071 		goto err;
10072 
10073 	if (!alloc_rt_sched_group(tg, parent))
10074 		goto err;
10075 
10076 	alloc_uclamp_sched_group(tg, parent);
10077 
10078 	return tg;
10079 
10080 err:
10081 	sched_free_group(tg);
10082 	return ERR_PTR(-ENOMEM);
10083 }
10084 
sched_online_group(struct task_group * tg,struct task_group * parent)10085 void sched_online_group(struct task_group *tg, struct task_group *parent)
10086 {
10087 	unsigned long flags;
10088 
10089 	spin_lock_irqsave(&task_group_lock, flags);
10090 	list_add_rcu(&tg->list, &task_groups);
10091 
10092 	/* Root should already exist: */
10093 	WARN_ON(!parent);
10094 
10095 	tg->parent = parent;
10096 	INIT_LIST_HEAD(&tg->children);
10097 	list_add_rcu(&tg->siblings, &parent->children);
10098 	spin_unlock_irqrestore(&task_group_lock, flags);
10099 
10100 	online_fair_sched_group(tg);
10101 }
10102 
10103 /* rcu callback to free various structures associated with a task group */
sched_unregister_group_rcu(struct rcu_head * rhp)10104 static void sched_unregister_group_rcu(struct rcu_head *rhp)
10105 {
10106 	/* Now it should be safe to free those cfs_rqs: */
10107 	sched_unregister_group(container_of(rhp, struct task_group, rcu));
10108 }
10109 
sched_destroy_group(struct task_group * tg)10110 void sched_destroy_group(struct task_group *tg)
10111 {
10112 	/* Wait for possible concurrent references to cfs_rqs complete: */
10113 	call_rcu(&tg->rcu, sched_unregister_group_rcu);
10114 }
10115 
sched_release_group(struct task_group * tg)10116 void sched_release_group(struct task_group *tg)
10117 {
10118 	unsigned long flags;
10119 
10120 	/*
10121 	 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10122 	 * sched_cfs_period_timer()).
10123 	 *
10124 	 * For this to be effective, we have to wait for all pending users of
10125 	 * this task group to leave their RCU critical section to ensure no new
10126 	 * user will see our dying task group any more. Specifically ensure
10127 	 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10128 	 *
10129 	 * We therefore defer calling unregister_fair_sched_group() to
10130 	 * sched_unregister_group() which is guarantied to get called only after the
10131 	 * current RCU grace period has expired.
10132 	 */
10133 	spin_lock_irqsave(&task_group_lock, flags);
10134 	list_del_rcu(&tg->list);
10135 	list_del_rcu(&tg->siblings);
10136 	spin_unlock_irqrestore(&task_group_lock, flags);
10137 }
10138 
sched_change_group(struct task_struct * tsk,int type)10139 static void sched_change_group(struct task_struct *tsk, int type)
10140 {
10141 	struct task_group *tg;
10142 
10143 	/*
10144 	 * All callers are synchronized by task_rq_lock(); we do not use RCU
10145 	 * which is pointless here. Thus, we pass "true" to task_css_check()
10146 	 * to prevent lockdep warnings.
10147 	 */
10148 	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10149 			  struct task_group, css);
10150 	tg = autogroup_task_group(tsk, tg);
10151 	tsk->sched_task_group = tg;
10152 
10153 #ifdef CONFIG_FAIR_GROUP_SCHED
10154 	if (tsk->sched_class->task_change_group)
10155 		tsk->sched_class->task_change_group(tsk, type);
10156 	else
10157 #endif
10158 		set_task_rq(tsk, task_cpu(tsk));
10159 }
10160 
10161 /*
10162  * Change task's runqueue when it moves between groups.
10163  *
10164  * The caller of this function should have put the task in its new group by
10165  * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10166  * its new group.
10167  */
sched_move_task(struct task_struct * tsk)10168 void sched_move_task(struct task_struct *tsk)
10169 {
10170 	int queued, running, queue_flags =
10171 		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10172 	struct rq_flags rf;
10173 	struct rq *rq;
10174 
10175 	rq = task_rq_lock(tsk, &rf);
10176 	update_rq_clock(rq);
10177 
10178 	running = task_current(rq, tsk);
10179 	queued = task_on_rq_queued(tsk);
10180 
10181 	if (queued)
10182 		dequeue_task(rq, tsk, queue_flags);
10183 	if (running)
10184 		put_prev_task(rq, tsk);
10185 
10186 	sched_change_group(tsk, TASK_MOVE_GROUP);
10187 
10188 	if (queued)
10189 		enqueue_task(rq, tsk, queue_flags);
10190 	if (running) {
10191 		set_next_task(rq, tsk);
10192 		/*
10193 		 * After changing group, the running task may have joined a
10194 		 * throttled one but it's still the running task. Trigger a
10195 		 * resched to make sure that task can still run.
10196 		 */
10197 		resched_curr(rq);
10198 	}
10199 
10200 	task_rq_unlock(rq, tsk, &rf);
10201 }
10202 
css_tg(struct cgroup_subsys_state * css)10203 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10204 {
10205 	return css ? container_of(css, struct task_group, css) : NULL;
10206 }
10207 
10208 static struct cgroup_subsys_state *
cpu_cgroup_css_alloc(struct cgroup_subsys_state * parent_css)10209 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10210 {
10211 	struct task_group *parent = css_tg(parent_css);
10212 	struct task_group *tg;
10213 
10214 	if (!parent) {
10215 		/* This is early initialization for the top cgroup */
10216 		return &root_task_group.css;
10217 	}
10218 
10219 	tg = sched_create_group(parent);
10220 	if (IS_ERR(tg))
10221 		return ERR_PTR(-ENOMEM);
10222 
10223 	return &tg->css;
10224 }
10225 
10226 /* Expose task group only after completing cgroup initialization */
cpu_cgroup_css_online(struct cgroup_subsys_state * css)10227 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10228 {
10229 	struct task_group *tg = css_tg(css);
10230 	struct task_group *parent = css_tg(css->parent);
10231 
10232 	if (parent)
10233 		sched_online_group(tg, parent);
10234 
10235 #ifdef CONFIG_UCLAMP_TASK_GROUP
10236 	/* Propagate the effective uclamp value for the new group */
10237 	mutex_lock(&uclamp_mutex);
10238 	rcu_read_lock();
10239 	cpu_util_update_eff(css);
10240 	rcu_read_unlock();
10241 	mutex_unlock(&uclamp_mutex);
10242 #endif
10243 
10244 	return 0;
10245 }
10246 
cpu_cgroup_css_released(struct cgroup_subsys_state * css)10247 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10248 {
10249 	struct task_group *tg = css_tg(css);
10250 
10251 	sched_release_group(tg);
10252 }
10253 
cpu_cgroup_css_free(struct cgroup_subsys_state * css)10254 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10255 {
10256 	struct task_group *tg = css_tg(css);
10257 
10258 	/*
10259 	 * Relies on the RCU grace period between css_released() and this.
10260 	 */
10261 	sched_unregister_group(tg);
10262 }
10263 
10264 /*
10265  * This is called before wake_up_new_task(), therefore we really only
10266  * have to set its group bits, all the other stuff does not apply.
10267  */
cpu_cgroup_fork(struct task_struct * task)10268 static void cpu_cgroup_fork(struct task_struct *task)
10269 {
10270 	struct rq_flags rf;
10271 	struct rq *rq;
10272 
10273 	rq = task_rq_lock(task, &rf);
10274 
10275 	update_rq_clock(rq);
10276 	sched_change_group(task, TASK_SET_GROUP);
10277 
10278 	task_rq_unlock(rq, task, &rf);
10279 }
10280 
cpu_cgroup_can_attach(struct cgroup_taskset * tset)10281 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10282 {
10283 	struct task_struct *task;
10284 	struct cgroup_subsys_state *css;
10285 	int ret = 0;
10286 
10287 	cgroup_taskset_for_each(task, css, tset) {
10288 #ifdef CONFIG_RT_GROUP_SCHED
10289 		if (!sched_rt_can_attach(css_tg(css), task))
10290 			return -EINVAL;
10291 #endif
10292 		/*
10293 		 * Serialize against wake_up_new_task() such that if it's
10294 		 * running, we're sure to observe its full state.
10295 		 */
10296 		raw_spin_lock_irq(&task->pi_lock);
10297 		/*
10298 		 * Avoid calling sched_move_task() before wake_up_new_task()
10299 		 * has happened. This would lead to problems with PELT, due to
10300 		 * move wanting to detach+attach while we're not attached yet.
10301 		 */
10302 		if (READ_ONCE(task->__state) == TASK_NEW)
10303 			ret = -EINVAL;
10304 		raw_spin_unlock_irq(&task->pi_lock);
10305 
10306 		if (ret)
10307 			break;
10308 	}
10309 	return ret;
10310 }
10311 
cpu_cgroup_attach(struct cgroup_taskset * tset)10312 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10313 {
10314 	struct task_struct *task;
10315 	struct cgroup_subsys_state *css;
10316 
10317 	cgroup_taskset_for_each(task, css, tset)
10318 		sched_move_task(task);
10319 }
10320 
10321 #ifdef CONFIG_UCLAMP_TASK_GROUP
cpu_util_update_eff(struct cgroup_subsys_state * css)10322 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10323 {
10324 	struct cgroup_subsys_state *top_css = css;
10325 	struct uclamp_se *uc_parent = NULL;
10326 	struct uclamp_se *uc_se = NULL;
10327 	unsigned int eff[UCLAMP_CNT];
10328 	enum uclamp_id clamp_id;
10329 	unsigned int clamps;
10330 
10331 	lockdep_assert_held(&uclamp_mutex);
10332 	SCHED_WARN_ON(!rcu_read_lock_held());
10333 
10334 	css_for_each_descendant_pre(css, top_css) {
10335 		uc_parent = css_tg(css)->parent
10336 			? css_tg(css)->parent->uclamp : NULL;
10337 
10338 		for_each_clamp_id(clamp_id) {
10339 			/* Assume effective clamps matches requested clamps */
10340 			eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10341 			/* Cap effective clamps with parent's effective clamps */
10342 			if (uc_parent &&
10343 			    eff[clamp_id] > uc_parent[clamp_id].value) {
10344 				eff[clamp_id] = uc_parent[clamp_id].value;
10345 			}
10346 		}
10347 		/* Ensure protection is always capped by limit */
10348 		eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10349 
10350 		/* Propagate most restrictive effective clamps */
10351 		clamps = 0x0;
10352 		uc_se = css_tg(css)->uclamp;
10353 		for_each_clamp_id(clamp_id) {
10354 			if (eff[clamp_id] == uc_se[clamp_id].value)
10355 				continue;
10356 			uc_se[clamp_id].value = eff[clamp_id];
10357 			uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10358 			clamps |= (0x1 << clamp_id);
10359 		}
10360 		if (!clamps) {
10361 			css = css_rightmost_descendant(css);
10362 			continue;
10363 		}
10364 
10365 		/* Immediately update descendants RUNNABLE tasks */
10366 		uclamp_update_active_tasks(css);
10367 	}
10368 }
10369 
10370 /*
10371  * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10372  * C expression. Since there is no way to convert a macro argument (N) into a
10373  * character constant, use two levels of macros.
10374  */
10375 #define _POW10(exp) ((unsigned int)1e##exp)
10376 #define POW10(exp) _POW10(exp)
10377 
10378 struct uclamp_request {
10379 #define UCLAMP_PERCENT_SHIFT	2
10380 #define UCLAMP_PERCENT_SCALE	(100 * POW10(UCLAMP_PERCENT_SHIFT))
10381 	s64 percent;
10382 	u64 util;
10383 	int ret;
10384 };
10385 
10386 static inline struct uclamp_request
capacity_from_percent(char * buf)10387 capacity_from_percent(char *buf)
10388 {
10389 	struct uclamp_request req = {
10390 		.percent = UCLAMP_PERCENT_SCALE,
10391 		.util = SCHED_CAPACITY_SCALE,
10392 		.ret = 0,
10393 	};
10394 
10395 	buf = strim(buf);
10396 	if (strcmp(buf, "max")) {
10397 		req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10398 					     &req.percent);
10399 		if (req.ret)
10400 			return req;
10401 		if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10402 			req.ret = -ERANGE;
10403 			return req;
10404 		}
10405 
10406 		req.util = req.percent << SCHED_CAPACITY_SHIFT;
10407 		req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10408 	}
10409 
10410 	return req;
10411 }
10412 
cpu_uclamp_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off,enum uclamp_id clamp_id)10413 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10414 				size_t nbytes, loff_t off,
10415 				enum uclamp_id clamp_id)
10416 {
10417 	struct uclamp_request req;
10418 	struct task_group *tg;
10419 
10420 	req = capacity_from_percent(buf);
10421 	if (req.ret)
10422 		return req.ret;
10423 
10424 	static_branch_enable(&sched_uclamp_used);
10425 
10426 	mutex_lock(&uclamp_mutex);
10427 	rcu_read_lock();
10428 
10429 	tg = css_tg(of_css(of));
10430 	if (tg->uclamp_req[clamp_id].value != req.util)
10431 		uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10432 
10433 	/*
10434 	 * Because of not recoverable conversion rounding we keep track of the
10435 	 * exact requested value
10436 	 */
10437 	tg->uclamp_pct[clamp_id] = req.percent;
10438 
10439 	/* Update effective clamps to track the most restrictive value */
10440 	cpu_util_update_eff(of_css(of));
10441 
10442 	rcu_read_unlock();
10443 	mutex_unlock(&uclamp_mutex);
10444 
10445 	return nbytes;
10446 }
10447 
cpu_uclamp_min_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)10448 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10449 				    char *buf, size_t nbytes,
10450 				    loff_t off)
10451 {
10452 	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10453 }
10454 
cpu_uclamp_max_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)10455 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10456 				    char *buf, size_t nbytes,
10457 				    loff_t off)
10458 {
10459 	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10460 }
10461 
cpu_uclamp_print(struct seq_file * sf,enum uclamp_id clamp_id)10462 static inline void cpu_uclamp_print(struct seq_file *sf,
10463 				    enum uclamp_id clamp_id)
10464 {
10465 	struct task_group *tg;
10466 	u64 util_clamp;
10467 	u64 percent;
10468 	u32 rem;
10469 
10470 	rcu_read_lock();
10471 	tg = css_tg(seq_css(sf));
10472 	util_clamp = tg->uclamp_req[clamp_id].value;
10473 	rcu_read_unlock();
10474 
10475 	if (util_clamp == SCHED_CAPACITY_SCALE) {
10476 		seq_puts(sf, "max\n");
10477 		return;
10478 	}
10479 
10480 	percent = tg->uclamp_pct[clamp_id];
10481 	percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10482 	seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10483 }
10484 
cpu_uclamp_min_show(struct seq_file * sf,void * v)10485 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10486 {
10487 	cpu_uclamp_print(sf, UCLAMP_MIN);
10488 	return 0;
10489 }
10490 
cpu_uclamp_max_show(struct seq_file * sf,void * v)10491 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10492 {
10493 	cpu_uclamp_print(sf, UCLAMP_MAX);
10494 	return 0;
10495 }
10496 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10497 
10498 #ifdef CONFIG_FAIR_GROUP_SCHED
cpu_shares_write_u64(struct cgroup_subsys_state * css,struct cftype * cftype,u64 shareval)10499 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10500 				struct cftype *cftype, u64 shareval)
10501 {
10502 	if (shareval > scale_load_down(ULONG_MAX))
10503 		shareval = MAX_SHARES;
10504 	return sched_group_set_shares(css_tg(css), scale_load(shareval));
10505 }
10506 
cpu_shares_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)10507 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10508 			       struct cftype *cft)
10509 {
10510 	struct task_group *tg = css_tg(css);
10511 
10512 	return (u64) scale_load_down(tg->shares);
10513 }
10514 
10515 #ifdef CONFIG_CFS_BANDWIDTH
10516 static DEFINE_MUTEX(cfs_constraints_mutex);
10517 
10518 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10519 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10520 /* More than 203 days if BW_SHIFT equals 20. */
10521 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10522 
10523 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10524 
tg_set_cfs_bandwidth(struct task_group * tg,u64 period,u64 quota,u64 burst)10525 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10526 				u64 burst)
10527 {
10528 	int i, ret = 0, runtime_enabled, runtime_was_enabled;
10529 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10530 
10531 	if (tg == &root_task_group)
10532 		return -EINVAL;
10533 
10534 	/*
10535 	 * Ensure we have at some amount of bandwidth every period.  This is
10536 	 * to prevent reaching a state of large arrears when throttled via
10537 	 * entity_tick() resulting in prolonged exit starvation.
10538 	 */
10539 	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10540 		return -EINVAL;
10541 
10542 	/*
10543 	 * Likewise, bound things on the other side by preventing insane quota
10544 	 * periods.  This also allows us to normalize in computing quota
10545 	 * feasibility.
10546 	 */
10547 	if (period > max_cfs_quota_period)
10548 		return -EINVAL;
10549 
10550 	/*
10551 	 * Bound quota to defend quota against overflow during bandwidth shift.
10552 	 */
10553 	if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10554 		return -EINVAL;
10555 
10556 	if (quota != RUNTIME_INF && (burst > quota ||
10557 				     burst + quota > max_cfs_runtime))
10558 		return -EINVAL;
10559 
10560 	/*
10561 	 * Prevent race between setting of cfs_rq->runtime_enabled and
10562 	 * unthrottle_offline_cfs_rqs().
10563 	 */
10564 	cpus_read_lock();
10565 	mutex_lock(&cfs_constraints_mutex);
10566 	ret = __cfs_schedulable(tg, period, quota);
10567 	if (ret)
10568 		goto out_unlock;
10569 
10570 	runtime_enabled = quota != RUNTIME_INF;
10571 	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10572 	/*
10573 	 * If we need to toggle cfs_bandwidth_used, off->on must occur
10574 	 * before making related changes, and on->off must occur afterwards
10575 	 */
10576 	if (runtime_enabled && !runtime_was_enabled)
10577 		cfs_bandwidth_usage_inc();
10578 	raw_spin_lock_irq(&cfs_b->lock);
10579 	cfs_b->period = ns_to_ktime(period);
10580 	cfs_b->quota = quota;
10581 	cfs_b->burst = burst;
10582 
10583 	__refill_cfs_bandwidth_runtime(cfs_b);
10584 
10585 	/* Restart the period timer (if active) to handle new period expiry: */
10586 	if (runtime_enabled)
10587 		start_cfs_bandwidth(cfs_b);
10588 
10589 	raw_spin_unlock_irq(&cfs_b->lock);
10590 
10591 	for_each_online_cpu(i) {
10592 		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10593 		struct rq *rq = cfs_rq->rq;
10594 		struct rq_flags rf;
10595 
10596 		rq_lock_irq(rq, &rf);
10597 		cfs_rq->runtime_enabled = runtime_enabled;
10598 		cfs_rq->runtime_remaining = 0;
10599 
10600 		if (cfs_rq->throttled)
10601 			unthrottle_cfs_rq(cfs_rq);
10602 		rq_unlock_irq(rq, &rf);
10603 	}
10604 	if (runtime_was_enabled && !runtime_enabled)
10605 		cfs_bandwidth_usage_dec();
10606 out_unlock:
10607 	mutex_unlock(&cfs_constraints_mutex);
10608 	cpus_read_unlock();
10609 
10610 	return ret;
10611 }
10612 
tg_set_cfs_quota(struct task_group * tg,long cfs_quota_us)10613 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10614 {
10615 	u64 quota, period, burst;
10616 
10617 	period = ktime_to_ns(tg->cfs_bandwidth.period);
10618 	burst = tg->cfs_bandwidth.burst;
10619 	if (cfs_quota_us < 0)
10620 		quota = RUNTIME_INF;
10621 	else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10622 		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10623 	else
10624 		return -EINVAL;
10625 
10626 	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10627 }
10628 
tg_get_cfs_quota(struct task_group * tg)10629 static long tg_get_cfs_quota(struct task_group *tg)
10630 {
10631 	u64 quota_us;
10632 
10633 	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10634 		return -1;
10635 
10636 	quota_us = tg->cfs_bandwidth.quota;
10637 	do_div(quota_us, NSEC_PER_USEC);
10638 
10639 	return quota_us;
10640 }
10641 
tg_set_cfs_period(struct task_group * tg,long cfs_period_us)10642 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10643 {
10644 	u64 quota, period, burst;
10645 
10646 	if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10647 		return -EINVAL;
10648 
10649 	period = (u64)cfs_period_us * NSEC_PER_USEC;
10650 	quota = tg->cfs_bandwidth.quota;
10651 	burst = tg->cfs_bandwidth.burst;
10652 
10653 	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10654 }
10655 
tg_get_cfs_period(struct task_group * tg)10656 static long tg_get_cfs_period(struct task_group *tg)
10657 {
10658 	u64 cfs_period_us;
10659 
10660 	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10661 	do_div(cfs_period_us, NSEC_PER_USEC);
10662 
10663 	return cfs_period_us;
10664 }
10665 
tg_set_cfs_burst(struct task_group * tg,long cfs_burst_us)10666 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10667 {
10668 	u64 quota, period, burst;
10669 
10670 	if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10671 		return -EINVAL;
10672 
10673 	burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10674 	period = ktime_to_ns(tg->cfs_bandwidth.period);
10675 	quota = tg->cfs_bandwidth.quota;
10676 
10677 	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10678 }
10679 
tg_get_cfs_burst(struct task_group * tg)10680 static long tg_get_cfs_burst(struct task_group *tg)
10681 {
10682 	u64 burst_us;
10683 
10684 	burst_us = tg->cfs_bandwidth.burst;
10685 	do_div(burst_us, NSEC_PER_USEC);
10686 
10687 	return burst_us;
10688 }
10689 
cpu_cfs_quota_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)10690 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10691 				  struct cftype *cft)
10692 {
10693 	return tg_get_cfs_quota(css_tg(css));
10694 }
10695 
cpu_cfs_quota_write_s64(struct cgroup_subsys_state * css,struct cftype * cftype,s64 cfs_quota_us)10696 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10697 				   struct cftype *cftype, s64 cfs_quota_us)
10698 {
10699 	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10700 }
10701 
cpu_cfs_period_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)10702 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10703 				   struct cftype *cft)
10704 {
10705 	return tg_get_cfs_period(css_tg(css));
10706 }
10707 
cpu_cfs_period_write_u64(struct cgroup_subsys_state * css,struct cftype * cftype,u64 cfs_period_us)10708 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10709 				    struct cftype *cftype, u64 cfs_period_us)
10710 {
10711 	return tg_set_cfs_period(css_tg(css), cfs_period_us);
10712 }
10713 
cpu_cfs_burst_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)10714 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10715 				  struct cftype *cft)
10716 {
10717 	return tg_get_cfs_burst(css_tg(css));
10718 }
10719 
cpu_cfs_burst_write_u64(struct cgroup_subsys_state * css,struct cftype * cftype,u64 cfs_burst_us)10720 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10721 				   struct cftype *cftype, u64 cfs_burst_us)
10722 {
10723 	return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10724 }
10725 
10726 struct cfs_schedulable_data {
10727 	struct task_group *tg;
10728 	u64 period, quota;
10729 };
10730 
10731 /*
10732  * normalize group quota/period to be quota/max_period
10733  * note: units are usecs
10734  */
normalize_cfs_quota(struct task_group * tg,struct cfs_schedulable_data * d)10735 static u64 normalize_cfs_quota(struct task_group *tg,
10736 			       struct cfs_schedulable_data *d)
10737 {
10738 	u64 quota, period;
10739 
10740 	if (tg == d->tg) {
10741 		period = d->period;
10742 		quota = d->quota;
10743 	} else {
10744 		period = tg_get_cfs_period(tg);
10745 		quota = tg_get_cfs_quota(tg);
10746 	}
10747 
10748 	/* note: these should typically be equivalent */
10749 	if (quota == RUNTIME_INF || quota == -1)
10750 		return RUNTIME_INF;
10751 
10752 	return to_ratio(period, quota);
10753 }
10754 
tg_cfs_schedulable_down(struct task_group * tg,void * data)10755 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10756 {
10757 	struct cfs_schedulable_data *d = data;
10758 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10759 	s64 quota = 0, parent_quota = -1;
10760 
10761 	if (!tg->parent) {
10762 		quota = RUNTIME_INF;
10763 	} else {
10764 		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10765 
10766 		quota = normalize_cfs_quota(tg, d);
10767 		parent_quota = parent_b->hierarchical_quota;
10768 
10769 		/*
10770 		 * Ensure max(child_quota) <= parent_quota.  On cgroup2,
10771 		 * always take the min.  On cgroup1, only inherit when no
10772 		 * limit is set:
10773 		 */
10774 		if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10775 			quota = min(quota, parent_quota);
10776 		} else {
10777 			if (quota == RUNTIME_INF)
10778 				quota = parent_quota;
10779 			else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10780 				return -EINVAL;
10781 		}
10782 	}
10783 	cfs_b->hierarchical_quota = quota;
10784 
10785 	return 0;
10786 }
10787 
__cfs_schedulable(struct task_group * tg,u64 period,u64 quota)10788 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10789 {
10790 	int ret;
10791 	struct cfs_schedulable_data data = {
10792 		.tg = tg,
10793 		.period = period,
10794 		.quota = quota,
10795 	};
10796 
10797 	if (quota != RUNTIME_INF) {
10798 		do_div(data.period, NSEC_PER_USEC);
10799 		do_div(data.quota, NSEC_PER_USEC);
10800 	}
10801 
10802 	rcu_read_lock();
10803 	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10804 	rcu_read_unlock();
10805 
10806 	return ret;
10807 }
10808 
cpu_cfs_stat_show(struct seq_file * sf,void * v)10809 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10810 {
10811 	struct task_group *tg = css_tg(seq_css(sf));
10812 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10813 
10814 	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10815 	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10816 	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10817 
10818 	if (schedstat_enabled() && tg != &root_task_group) {
10819 		struct sched_statistics *stats;
10820 		u64 ws = 0;
10821 		int i;
10822 
10823 		for_each_possible_cpu(i) {
10824 			stats = __schedstats_from_se(tg->se[i]);
10825 			ws += schedstat_val(stats->wait_sum);
10826 		}
10827 
10828 		seq_printf(sf, "wait_sum %llu\n", ws);
10829 	}
10830 
10831 	seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10832 	seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10833 
10834 	return 0;
10835 }
10836 #endif /* CONFIG_CFS_BANDWIDTH */
10837 #endif /* CONFIG_FAIR_GROUP_SCHED */
10838 
10839 #ifdef CONFIG_RT_GROUP_SCHED
cpu_rt_runtime_write(struct cgroup_subsys_state * css,struct cftype * cft,s64 val)10840 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10841 				struct cftype *cft, s64 val)
10842 {
10843 	return sched_group_set_rt_runtime(css_tg(css), val);
10844 }
10845 
cpu_rt_runtime_read(struct cgroup_subsys_state * css,struct cftype * cft)10846 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10847 			       struct cftype *cft)
10848 {
10849 	return sched_group_rt_runtime(css_tg(css));
10850 }
10851 
cpu_rt_period_write_uint(struct cgroup_subsys_state * css,struct cftype * cftype,u64 rt_period_us)10852 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10853 				    struct cftype *cftype, u64 rt_period_us)
10854 {
10855 	return sched_group_set_rt_period(css_tg(css), rt_period_us);
10856 }
10857 
cpu_rt_period_read_uint(struct cgroup_subsys_state * css,struct cftype * cft)10858 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10859 				   struct cftype *cft)
10860 {
10861 	return sched_group_rt_period(css_tg(css));
10862 }
10863 #endif /* CONFIG_RT_GROUP_SCHED */
10864 
10865 #ifdef CONFIG_FAIR_GROUP_SCHED
cpu_idle_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)10866 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10867 			       struct cftype *cft)
10868 {
10869 	return css_tg(css)->idle;
10870 }
10871 
cpu_idle_write_s64(struct cgroup_subsys_state * css,struct cftype * cft,s64 idle)10872 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10873 				struct cftype *cft, s64 idle)
10874 {
10875 	return sched_group_set_idle(css_tg(css), idle);
10876 }
10877 #endif
10878 
10879 static struct cftype cpu_legacy_files[] = {
10880 #ifdef CONFIG_FAIR_GROUP_SCHED
10881 	{
10882 		.name = "shares",
10883 		.read_u64 = cpu_shares_read_u64,
10884 		.write_u64 = cpu_shares_write_u64,
10885 	},
10886 	{
10887 		.name = "idle",
10888 		.read_s64 = cpu_idle_read_s64,
10889 		.write_s64 = cpu_idle_write_s64,
10890 	},
10891 #endif
10892 #ifdef CONFIG_CFS_BANDWIDTH
10893 	{
10894 		.name = "cfs_quota_us",
10895 		.read_s64 = cpu_cfs_quota_read_s64,
10896 		.write_s64 = cpu_cfs_quota_write_s64,
10897 	},
10898 	{
10899 		.name = "cfs_period_us",
10900 		.read_u64 = cpu_cfs_period_read_u64,
10901 		.write_u64 = cpu_cfs_period_write_u64,
10902 	},
10903 	{
10904 		.name = "cfs_burst_us",
10905 		.read_u64 = cpu_cfs_burst_read_u64,
10906 		.write_u64 = cpu_cfs_burst_write_u64,
10907 	},
10908 	{
10909 		.name = "stat",
10910 		.seq_show = cpu_cfs_stat_show,
10911 	},
10912 #endif
10913 #ifdef CONFIG_RT_GROUP_SCHED
10914 	{
10915 		.name = "rt_runtime_us",
10916 		.read_s64 = cpu_rt_runtime_read,
10917 		.write_s64 = cpu_rt_runtime_write,
10918 	},
10919 	{
10920 		.name = "rt_period_us",
10921 		.read_u64 = cpu_rt_period_read_uint,
10922 		.write_u64 = cpu_rt_period_write_uint,
10923 	},
10924 #endif
10925 #ifdef CONFIG_UCLAMP_TASK_GROUP
10926 	{
10927 		.name = "uclamp.min",
10928 		.flags = CFTYPE_NOT_ON_ROOT,
10929 		.seq_show = cpu_uclamp_min_show,
10930 		.write = cpu_uclamp_min_write,
10931 	},
10932 	{
10933 		.name = "uclamp.max",
10934 		.flags = CFTYPE_NOT_ON_ROOT,
10935 		.seq_show = cpu_uclamp_max_show,
10936 		.write = cpu_uclamp_max_write,
10937 	},
10938 #endif
10939 	{ }	/* Terminate */
10940 };
10941 
cpu_extra_stat_show(struct seq_file * sf,struct cgroup_subsys_state * css)10942 static int cpu_extra_stat_show(struct seq_file *sf,
10943 			       struct cgroup_subsys_state *css)
10944 {
10945 #ifdef CONFIG_CFS_BANDWIDTH
10946 	{
10947 		struct task_group *tg = css_tg(css);
10948 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10949 		u64 throttled_usec, burst_usec;
10950 
10951 		throttled_usec = cfs_b->throttled_time;
10952 		do_div(throttled_usec, NSEC_PER_USEC);
10953 		burst_usec = cfs_b->burst_time;
10954 		do_div(burst_usec, NSEC_PER_USEC);
10955 
10956 		seq_printf(sf, "nr_periods %d\n"
10957 			   "nr_throttled %d\n"
10958 			   "throttled_usec %llu\n"
10959 			   "nr_bursts %d\n"
10960 			   "burst_usec %llu\n",
10961 			   cfs_b->nr_periods, cfs_b->nr_throttled,
10962 			   throttled_usec, cfs_b->nr_burst, burst_usec);
10963 	}
10964 #endif
10965 	return 0;
10966 }
10967 
10968 #ifdef CONFIG_FAIR_GROUP_SCHED
cpu_weight_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)10969 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10970 			       struct cftype *cft)
10971 {
10972 	struct task_group *tg = css_tg(css);
10973 	u64 weight = scale_load_down(tg->shares);
10974 
10975 	return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10976 }
10977 
cpu_weight_write_u64(struct cgroup_subsys_state * css,struct cftype * cft,u64 weight)10978 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10979 				struct cftype *cft, u64 weight)
10980 {
10981 	/*
10982 	 * cgroup weight knobs should use the common MIN, DFL and MAX
10983 	 * values which are 1, 100 and 10000 respectively.  While it loses
10984 	 * a bit of range on both ends, it maps pretty well onto the shares
10985 	 * value used by scheduler and the round-trip conversions preserve
10986 	 * the original value over the entire range.
10987 	 */
10988 	if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10989 		return -ERANGE;
10990 
10991 	weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10992 
10993 	return sched_group_set_shares(css_tg(css), scale_load(weight));
10994 }
10995 
cpu_weight_nice_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)10996 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10997 				    struct cftype *cft)
10998 {
10999 	unsigned long weight = scale_load_down(css_tg(css)->shares);
11000 	int last_delta = INT_MAX;
11001 	int prio, delta;
11002 
11003 	/* find the closest nice value to the current weight */
11004 	for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11005 		delta = abs(sched_prio_to_weight[prio] - weight);
11006 		if (delta >= last_delta)
11007 			break;
11008 		last_delta = delta;
11009 	}
11010 
11011 	return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11012 }
11013 
cpu_weight_nice_write_s64(struct cgroup_subsys_state * css,struct cftype * cft,s64 nice)11014 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11015 				     struct cftype *cft, s64 nice)
11016 {
11017 	unsigned long weight;
11018 	int idx;
11019 
11020 	if (nice < MIN_NICE || nice > MAX_NICE)
11021 		return -ERANGE;
11022 
11023 	idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11024 	idx = array_index_nospec(idx, 40);
11025 	weight = sched_prio_to_weight[idx];
11026 
11027 	return sched_group_set_shares(css_tg(css), scale_load(weight));
11028 }
11029 #endif
11030 
cpu_period_quota_print(struct seq_file * sf,long period,long quota)11031 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11032 						  long period, long quota)
11033 {
11034 	if (quota < 0)
11035 		seq_puts(sf, "max");
11036 	else
11037 		seq_printf(sf, "%ld", quota);
11038 
11039 	seq_printf(sf, " %ld\n", period);
11040 }
11041 
11042 /* caller should put the current value in *@periodp before calling */
cpu_period_quota_parse(char * buf,u64 * periodp,u64 * quotap)11043 static int __maybe_unused cpu_period_quota_parse(char *buf,
11044 						 u64 *periodp, u64 *quotap)
11045 {
11046 	char tok[21];	/* U64_MAX */
11047 
11048 	if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11049 		return -EINVAL;
11050 
11051 	*periodp *= NSEC_PER_USEC;
11052 
11053 	if (sscanf(tok, "%llu", quotap))
11054 		*quotap *= NSEC_PER_USEC;
11055 	else if (!strcmp(tok, "max"))
11056 		*quotap = RUNTIME_INF;
11057 	else
11058 		return -EINVAL;
11059 
11060 	return 0;
11061 }
11062 
11063 #ifdef CONFIG_CFS_BANDWIDTH
cpu_max_show(struct seq_file * sf,void * v)11064 static int cpu_max_show(struct seq_file *sf, void *v)
11065 {
11066 	struct task_group *tg = css_tg(seq_css(sf));
11067 
11068 	cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11069 	return 0;
11070 }
11071 
cpu_max_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)11072 static ssize_t cpu_max_write(struct kernfs_open_file *of,
11073 			     char *buf, size_t nbytes, loff_t off)
11074 {
11075 	struct task_group *tg = css_tg(of_css(of));
11076 	u64 period = tg_get_cfs_period(tg);
11077 	u64 burst = tg_get_cfs_burst(tg);
11078 	u64 quota;
11079 	int ret;
11080 
11081 	ret = cpu_period_quota_parse(buf, &period, &quota);
11082 	if (!ret)
11083 		ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11084 	return ret ?: nbytes;
11085 }
11086 #endif
11087 
11088 static struct cftype cpu_files[] = {
11089 #ifdef CONFIG_FAIR_GROUP_SCHED
11090 	{
11091 		.name = "weight",
11092 		.flags = CFTYPE_NOT_ON_ROOT,
11093 		.read_u64 = cpu_weight_read_u64,
11094 		.write_u64 = cpu_weight_write_u64,
11095 	},
11096 	{
11097 		.name = "weight.nice",
11098 		.flags = CFTYPE_NOT_ON_ROOT,
11099 		.read_s64 = cpu_weight_nice_read_s64,
11100 		.write_s64 = cpu_weight_nice_write_s64,
11101 	},
11102 	{
11103 		.name = "idle",
11104 		.flags = CFTYPE_NOT_ON_ROOT,
11105 		.read_s64 = cpu_idle_read_s64,
11106 		.write_s64 = cpu_idle_write_s64,
11107 	},
11108 #endif
11109 #ifdef CONFIG_CFS_BANDWIDTH
11110 	{
11111 		.name = "max",
11112 		.flags = CFTYPE_NOT_ON_ROOT,
11113 		.seq_show = cpu_max_show,
11114 		.write = cpu_max_write,
11115 	},
11116 	{
11117 		.name = "max.burst",
11118 		.flags = CFTYPE_NOT_ON_ROOT,
11119 		.read_u64 = cpu_cfs_burst_read_u64,
11120 		.write_u64 = cpu_cfs_burst_write_u64,
11121 	},
11122 #endif
11123 #ifdef CONFIG_UCLAMP_TASK_GROUP
11124 	{
11125 		.name = "uclamp.min",
11126 		.flags = CFTYPE_NOT_ON_ROOT,
11127 		.seq_show = cpu_uclamp_min_show,
11128 		.write = cpu_uclamp_min_write,
11129 	},
11130 	{
11131 		.name = "uclamp.max",
11132 		.flags = CFTYPE_NOT_ON_ROOT,
11133 		.seq_show = cpu_uclamp_max_show,
11134 		.write = cpu_uclamp_max_write,
11135 	},
11136 #endif
11137 	{ }	/* terminate */
11138 };
11139 
11140 struct cgroup_subsys cpu_cgrp_subsys = {
11141 	.css_alloc	= cpu_cgroup_css_alloc,
11142 	.css_online	= cpu_cgroup_css_online,
11143 	.css_released	= cpu_cgroup_css_released,
11144 	.css_free	= cpu_cgroup_css_free,
11145 	.css_extra_stat_show = cpu_extra_stat_show,
11146 	.fork		= cpu_cgroup_fork,
11147 	.can_attach	= cpu_cgroup_can_attach,
11148 	.attach		= cpu_cgroup_attach,
11149 	.legacy_cftypes	= cpu_legacy_files,
11150 	.dfl_cftypes	= cpu_files,
11151 	.early_init	= true,
11152 	.threaded	= true,
11153 };
11154 
11155 #endif	/* CONFIG_CGROUP_SCHED */
11156 
dump_cpu_task(int cpu)11157 void dump_cpu_task(int cpu)
11158 {
11159 	pr_info("Task dump for CPU %d:\n", cpu);
11160 	sched_show_task(cpu_curr(cpu));
11161 }
11162 
11163 /*
11164  * Nice levels are multiplicative, with a gentle 10% change for every
11165  * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11166  * nice 1, it will get ~10% less CPU time than another CPU-bound task
11167  * that remained on nice 0.
11168  *
11169  * The "10% effect" is relative and cumulative: from _any_ nice level,
11170  * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11171  * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11172  * If a task goes up by ~10% and another task goes down by ~10% then
11173  * the relative distance between them is ~25%.)
11174  */
11175 const int sched_prio_to_weight[40] = {
11176  /* -20 */     88761,     71755,     56483,     46273,     36291,
11177  /* -15 */     29154,     23254,     18705,     14949,     11916,
11178  /* -10 */      9548,      7620,      6100,      4904,      3906,
11179  /*  -5 */      3121,      2501,      1991,      1586,      1277,
11180  /*   0 */      1024,       820,       655,       526,       423,
11181  /*   5 */       335,       272,       215,       172,       137,
11182  /*  10 */       110,        87,        70,        56,        45,
11183  /*  15 */        36,        29,        23,        18,        15,
11184 };
11185 
11186 /*
11187  * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11188  *
11189  * In cases where the weight does not change often, we can use the
11190  * precalculated inverse to speed up arithmetics by turning divisions
11191  * into multiplications:
11192  */
11193 const u32 sched_prio_to_wmult[40] = {
11194  /* -20 */     48388,     59856,     76040,     92818,    118348,
11195  /* -15 */    147320,    184698,    229616,    287308,    360437,
11196  /* -10 */    449829,    563644,    704093,    875809,   1099582,
11197  /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
11198  /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
11199  /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
11200  /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
11201  /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11202 };
11203 
call_trace_sched_update_nr_running(struct rq * rq,int count)11204 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11205 {
11206         trace_sched_update_nr_running_tp(rq, count);
11207 }
11208