1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  kernel/sched/core.c
4  *
5  *  Core kernel scheduler code and related syscalls
6  *
7  *  Copyright (C) 1991-2002  Linus Torvalds
8  */
9 #include <linux/highmem.h>
10 #include <linux/hrtimer_api.h>
11 #include <linux/ktime_api.h>
12 #include <linux/sched/signal.h>
13 #include <linux/syscalls_api.h>
14 #include <linux/debug_locks.h>
15 #include <linux/prefetch.h>
16 #include <linux/capability.h>
17 #include <linux/pgtable_api.h>
18 #include <linux/wait_bit.h>
19 #include <linux/jiffies.h>
20 #include <linux/spinlock_api.h>
21 #include <linux/cpumask_api.h>
22 #include <linux/lockdep_api.h>
23 #include <linux/hardirq.h>
24 #include <linux/softirq.h>
25 #include <linux/refcount_api.h>
26 #include <linux/topology.h>
27 #include <linux/sched/clock.h>
28 #include <linux/sched/cond_resched.h>
29 #include <linux/sched/cputime.h>
30 #include <linux/sched/debug.h>
31 #include <linux/sched/hotplug.h>
32 #include <linux/sched/init.h>
33 #include <linux/sched/isolation.h>
34 #include <linux/sched/loadavg.h>
35 #include <linux/sched/mm.h>
36 #include <linux/sched/nohz.h>
37 #include <linux/sched/rseq_api.h>
38 #include <linux/sched/rt.h>
39 
40 #include <linux/blkdev.h>
41 #include <linux/context_tracking.h>
42 #include <linux/cpuset.h>
43 #include <linux/delayacct.h>
44 #include <linux/init_task.h>
45 #include <linux/interrupt.h>
46 #include <linux/ioprio.h>
47 #include <linux/kallsyms.h>
48 #include <linux/kcov.h>
49 #include <linux/kprobes.h>
50 #include <linux/llist_api.h>
51 #include <linux/mmu_context.h>
52 #include <linux/mmzone.h>
53 #include <linux/mutex_api.h>
54 #include <linux/nmi.h>
55 #include <linux/nospec.h>
56 #include <linux/perf_event_api.h>
57 #include <linux/profile.h>
58 #include <linux/psi.h>
59 #include <linux/rcuwait_api.h>
60 #include <linux/sched/wake_q.h>
61 #include <linux/scs.h>
62 #include <linux/slab.h>
63 #include <linux/syscalls.h>
64 #include <linux/vtime.h>
65 #include <linux/wait_api.h>
66 #include <linux/workqueue_api.h>
67 
68 #ifdef CONFIG_PREEMPT_DYNAMIC
69 # ifdef CONFIG_GENERIC_ENTRY
70 #  include <linux/entry-common.h>
71 # endif
72 #endif
73 
74 #include <uapi/linux/sched/types.h>
75 
76 #include <asm/irq_regs.h>
77 #include <asm/switch_to.h>
78 #include <asm/tlb.h>
79 
80 #define CREATE_TRACE_POINTS
81 #include <linux/sched/rseq_api.h>
82 #include <trace/events/sched.h>
83 #include <trace/events/ipi.h>
84 #undef CREATE_TRACE_POINTS
85 
86 #include "sched.h"
87 #include "stats.h"
88 #include "autogroup.h"
89 
90 #include "autogroup.h"
91 #include "pelt.h"
92 #include "smp.h"
93 #include "stats.h"
94 
95 #include "../workqueue_internal.h"
96 #include "../../io_uring/io-wq.h"
97 #include "../smpboot.h"
98 
99 EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpu);
100 EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpumask);
101 
102 /*
103  * Export tracepoints that act as a bare tracehook (ie: have no trace event
104  * associated with them) to allow external modules to probe them.
105  */
106 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
107 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
108 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
109 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
110 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
111 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
112 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
113 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
114 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
115 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
116 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
117 
118 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
119 
120 #ifdef CONFIG_SCHED_DEBUG
121 /*
122  * Debugging: various feature bits
123  *
124  * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
125  * sysctl_sched_features, defined in sched.h, to allow constants propagation
126  * at compile time and compiler optimization based on features default.
127  */
128 #define SCHED_FEAT(name, enabled)	\
129 	(1UL << __SCHED_FEAT_##name) * enabled |
130 const_debug unsigned int sysctl_sched_features =
131 #include "features.h"
132 	0;
133 #undef SCHED_FEAT
134 
135 /*
136  * Print a warning if need_resched is set for the given duration (if
137  * LATENCY_WARN is enabled).
138  *
139  * If sysctl_resched_latency_warn_once is set, only one warning will be shown
140  * per boot.
141  */
142 __read_mostly int sysctl_resched_latency_warn_ms = 100;
143 __read_mostly int sysctl_resched_latency_warn_once = 1;
144 #endif /* CONFIG_SCHED_DEBUG */
145 
146 /*
147  * Number of tasks to iterate in a single balance run.
148  * Limited because this is done with IRQs disabled.
149  */
150 const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
151 
152 __read_mostly int scheduler_running;
153 
154 #ifdef CONFIG_SCHED_CORE
155 
156 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
157 
158 /* kernel prio, less is more */
__task_prio(const struct task_struct * p)159 static inline int __task_prio(const struct task_struct *p)
160 {
161 	if (p->sched_class == &stop_sched_class) /* trumps deadline */
162 		return -2;
163 
164 	if (rt_prio(p->prio)) /* includes deadline */
165 		return p->prio; /* [-1, 99] */
166 
167 	if (p->sched_class == &idle_sched_class)
168 		return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
169 
170 	return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
171 }
172 
173 /*
174  * l(a,b)
175  * le(a,b) := !l(b,a)
176  * g(a,b)  := l(b,a)
177  * ge(a,b) := !l(a,b)
178  */
179 
180 /* real prio, less is less */
prio_less(const struct task_struct * a,const struct task_struct * b,bool in_fi)181 static inline bool prio_less(const struct task_struct *a,
182 			     const struct task_struct *b, bool in_fi)
183 {
184 
185 	int pa = __task_prio(a), pb = __task_prio(b);
186 
187 	if (-pa < -pb)
188 		return true;
189 
190 	if (-pb < -pa)
191 		return false;
192 
193 	if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
194 		return !dl_time_before(a->dl.deadline, b->dl.deadline);
195 
196 	if (pa == MAX_RT_PRIO + MAX_NICE)	/* fair */
197 		return cfs_prio_less(a, b, in_fi);
198 
199 	return false;
200 }
201 
__sched_core_less(const struct task_struct * a,const struct task_struct * b)202 static inline bool __sched_core_less(const struct task_struct *a,
203 				     const struct task_struct *b)
204 {
205 	if (a->core_cookie < b->core_cookie)
206 		return true;
207 
208 	if (a->core_cookie > b->core_cookie)
209 		return false;
210 
211 	/* flip prio, so high prio is leftmost */
212 	if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
213 		return true;
214 
215 	return false;
216 }
217 
218 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
219 
rb_sched_core_less(struct rb_node * a,const struct rb_node * b)220 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
221 {
222 	return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
223 }
224 
rb_sched_core_cmp(const void * key,const struct rb_node * node)225 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
226 {
227 	const struct task_struct *p = __node_2_sc(node);
228 	unsigned long cookie = (unsigned long)key;
229 
230 	if (cookie < p->core_cookie)
231 		return -1;
232 
233 	if (cookie > p->core_cookie)
234 		return 1;
235 
236 	return 0;
237 }
238 
sched_core_enqueue(struct rq * rq,struct task_struct * p)239 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
240 {
241 	rq->core->core_task_seq++;
242 
243 	if (!p->core_cookie)
244 		return;
245 
246 	rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
247 }
248 
sched_core_dequeue(struct rq * rq,struct task_struct * p,int flags)249 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
250 {
251 	rq->core->core_task_seq++;
252 
253 	if (sched_core_enqueued(p)) {
254 		rb_erase(&p->core_node, &rq->core_tree);
255 		RB_CLEAR_NODE(&p->core_node);
256 	}
257 
258 	/*
259 	 * Migrating the last task off the cpu, with the cpu in forced idle
260 	 * state. Reschedule to create an accounting edge for forced idle,
261 	 * and re-examine whether the core is still in forced idle state.
262 	 */
263 	if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
264 	    rq->core->core_forceidle_count && rq->curr == rq->idle)
265 		resched_curr(rq);
266 }
267 
sched_task_is_throttled(struct task_struct * p,int cpu)268 static int sched_task_is_throttled(struct task_struct *p, int cpu)
269 {
270 	if (p->sched_class->task_is_throttled)
271 		return p->sched_class->task_is_throttled(p, cpu);
272 
273 	return 0;
274 }
275 
sched_core_next(struct task_struct * p,unsigned long cookie)276 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
277 {
278 	struct rb_node *node = &p->core_node;
279 	int cpu = task_cpu(p);
280 
281 	do {
282 		node = rb_next(node);
283 		if (!node)
284 			return NULL;
285 
286 		p = __node_2_sc(node);
287 		if (p->core_cookie != cookie)
288 			return NULL;
289 
290 	} while (sched_task_is_throttled(p, cpu));
291 
292 	return p;
293 }
294 
295 /*
296  * Find left-most (aka, highest priority) and unthrottled task matching @cookie.
297  * If no suitable task is found, NULL will be returned.
298  */
sched_core_find(struct rq * rq,unsigned long cookie)299 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
300 {
301 	struct task_struct *p;
302 	struct rb_node *node;
303 
304 	node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
305 	if (!node)
306 		return NULL;
307 
308 	p = __node_2_sc(node);
309 	if (!sched_task_is_throttled(p, rq->cpu))
310 		return p;
311 
312 	return sched_core_next(p, cookie);
313 }
314 
315 /*
316  * Magic required such that:
317  *
318  *	raw_spin_rq_lock(rq);
319  *	...
320  *	raw_spin_rq_unlock(rq);
321  *
322  * ends up locking and unlocking the _same_ lock, and all CPUs
323  * always agree on what rq has what lock.
324  *
325  * XXX entirely possible to selectively enable cores, don't bother for now.
326  */
327 
328 static DEFINE_MUTEX(sched_core_mutex);
329 static atomic_t sched_core_count;
330 static struct cpumask sched_core_mask;
331 
sched_core_lock(int cpu,unsigned long * flags)332 static void sched_core_lock(int cpu, unsigned long *flags)
333 {
334 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
335 	int t, i = 0;
336 
337 	local_irq_save(*flags);
338 	for_each_cpu(t, smt_mask)
339 		raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
340 }
341 
sched_core_unlock(int cpu,unsigned long * flags)342 static void sched_core_unlock(int cpu, unsigned long *flags)
343 {
344 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
345 	int t;
346 
347 	for_each_cpu(t, smt_mask)
348 		raw_spin_unlock(&cpu_rq(t)->__lock);
349 	local_irq_restore(*flags);
350 }
351 
__sched_core_flip(bool enabled)352 static void __sched_core_flip(bool enabled)
353 {
354 	unsigned long flags;
355 	int cpu, t;
356 
357 	cpus_read_lock();
358 
359 	/*
360 	 * Toggle the online cores, one by one.
361 	 */
362 	cpumask_copy(&sched_core_mask, cpu_online_mask);
363 	for_each_cpu(cpu, &sched_core_mask) {
364 		const struct cpumask *smt_mask = cpu_smt_mask(cpu);
365 
366 		sched_core_lock(cpu, &flags);
367 
368 		for_each_cpu(t, smt_mask)
369 			cpu_rq(t)->core_enabled = enabled;
370 
371 		cpu_rq(cpu)->core->core_forceidle_start = 0;
372 
373 		sched_core_unlock(cpu, &flags);
374 
375 		cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
376 	}
377 
378 	/*
379 	 * Toggle the offline CPUs.
380 	 */
381 	for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
382 		cpu_rq(cpu)->core_enabled = enabled;
383 
384 	cpus_read_unlock();
385 }
386 
sched_core_assert_empty(void)387 static void sched_core_assert_empty(void)
388 {
389 	int cpu;
390 
391 	for_each_possible_cpu(cpu)
392 		WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
393 }
394 
__sched_core_enable(void)395 static void __sched_core_enable(void)
396 {
397 	static_branch_enable(&__sched_core_enabled);
398 	/*
399 	 * Ensure all previous instances of raw_spin_rq_*lock() have finished
400 	 * and future ones will observe !sched_core_disabled().
401 	 */
402 	synchronize_rcu();
403 	__sched_core_flip(true);
404 	sched_core_assert_empty();
405 }
406 
__sched_core_disable(void)407 static void __sched_core_disable(void)
408 {
409 	sched_core_assert_empty();
410 	__sched_core_flip(false);
411 	static_branch_disable(&__sched_core_enabled);
412 }
413 
sched_core_get(void)414 void sched_core_get(void)
415 {
416 	if (atomic_inc_not_zero(&sched_core_count))
417 		return;
418 
419 	mutex_lock(&sched_core_mutex);
420 	if (!atomic_read(&sched_core_count))
421 		__sched_core_enable();
422 
423 	smp_mb__before_atomic();
424 	atomic_inc(&sched_core_count);
425 	mutex_unlock(&sched_core_mutex);
426 }
427 
__sched_core_put(struct work_struct * work)428 static void __sched_core_put(struct work_struct *work)
429 {
430 	if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
431 		__sched_core_disable();
432 		mutex_unlock(&sched_core_mutex);
433 	}
434 }
435 
sched_core_put(void)436 void sched_core_put(void)
437 {
438 	static DECLARE_WORK(_work, __sched_core_put);
439 
440 	/*
441 	 * "There can be only one"
442 	 *
443 	 * Either this is the last one, or we don't actually need to do any
444 	 * 'work'. If it is the last *again*, we rely on
445 	 * WORK_STRUCT_PENDING_BIT.
446 	 */
447 	if (!atomic_add_unless(&sched_core_count, -1, 1))
448 		schedule_work(&_work);
449 }
450 
451 #else /* !CONFIG_SCHED_CORE */
452 
sched_core_enqueue(struct rq * rq,struct task_struct * p)453 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
454 static inline void
sched_core_dequeue(struct rq * rq,struct task_struct * p,int flags)455 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
456 
457 #endif /* CONFIG_SCHED_CORE */
458 
459 /*
460  * Serialization rules:
461  *
462  * Lock order:
463  *
464  *   p->pi_lock
465  *     rq->lock
466  *       hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
467  *
468  *  rq1->lock
469  *    rq2->lock  where: rq1 < rq2
470  *
471  * Regular state:
472  *
473  * Normal scheduling state is serialized by rq->lock. __schedule() takes the
474  * local CPU's rq->lock, it optionally removes the task from the runqueue and
475  * always looks at the local rq data structures to find the most eligible task
476  * to run next.
477  *
478  * Task enqueue is also under rq->lock, possibly taken from another CPU.
479  * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
480  * the local CPU to avoid bouncing the runqueue state around [ see
481  * ttwu_queue_wakelist() ]
482  *
483  * Task wakeup, specifically wakeups that involve migration, are horribly
484  * complicated to avoid having to take two rq->locks.
485  *
486  * Special state:
487  *
488  * System-calls and anything external will use task_rq_lock() which acquires
489  * both p->pi_lock and rq->lock. As a consequence the state they change is
490  * stable while holding either lock:
491  *
492  *  - sched_setaffinity()/
493  *    set_cpus_allowed_ptr():	p->cpus_ptr, p->nr_cpus_allowed
494  *  - set_user_nice():		p->se.load, p->*prio
495  *  - __sched_setscheduler():	p->sched_class, p->policy, p->*prio,
496  *				p->se.load, p->rt_priority,
497  *				p->dl.dl_{runtime, deadline, period, flags, bw, density}
498  *  - sched_setnuma():		p->numa_preferred_nid
499  *  - sched_move_task():	p->sched_task_group
500  *  - uclamp_update_active()	p->uclamp*
501  *
502  * p->state <- TASK_*:
503  *
504  *   is changed locklessly using set_current_state(), __set_current_state() or
505  *   set_special_state(), see their respective comments, or by
506  *   try_to_wake_up(). This latter uses p->pi_lock to serialize against
507  *   concurrent self.
508  *
509  * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
510  *
511  *   is set by activate_task() and cleared by deactivate_task(), under
512  *   rq->lock. Non-zero indicates the task is runnable, the special
513  *   ON_RQ_MIGRATING state is used for migration without holding both
514  *   rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
515  *
516  * p->on_cpu <- { 0, 1 }:
517  *
518  *   is set by prepare_task() and cleared by finish_task() such that it will be
519  *   set before p is scheduled-in and cleared after p is scheduled-out, both
520  *   under rq->lock. Non-zero indicates the task is running on its CPU.
521  *
522  *   [ The astute reader will observe that it is possible for two tasks on one
523  *     CPU to have ->on_cpu = 1 at the same time. ]
524  *
525  * task_cpu(p): is changed by set_task_cpu(), the rules are:
526  *
527  *  - Don't call set_task_cpu() on a blocked task:
528  *
529  *    We don't care what CPU we're not running on, this simplifies hotplug,
530  *    the CPU assignment of blocked tasks isn't required to be valid.
531  *
532  *  - for try_to_wake_up(), called under p->pi_lock:
533  *
534  *    This allows try_to_wake_up() to only take one rq->lock, see its comment.
535  *
536  *  - for migration called under rq->lock:
537  *    [ see task_on_rq_migrating() in task_rq_lock() ]
538  *
539  *    o move_queued_task()
540  *    o detach_task()
541  *
542  *  - for migration called under double_rq_lock():
543  *
544  *    o __migrate_swap_task()
545  *    o push_rt_task() / pull_rt_task()
546  *    o push_dl_task() / pull_dl_task()
547  *    o dl_task_offline_migration()
548  *
549  */
550 
raw_spin_rq_lock_nested(struct rq * rq,int subclass)551 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
552 {
553 	raw_spinlock_t *lock;
554 
555 	/* Matches synchronize_rcu() in __sched_core_enable() */
556 	preempt_disable();
557 	if (sched_core_disabled()) {
558 		raw_spin_lock_nested(&rq->__lock, subclass);
559 		/* preempt_count *MUST* be > 1 */
560 		preempt_enable_no_resched();
561 		return;
562 	}
563 
564 	for (;;) {
565 		lock = __rq_lockp(rq);
566 		raw_spin_lock_nested(lock, subclass);
567 		if (likely(lock == __rq_lockp(rq))) {
568 			/* preempt_count *MUST* be > 1 */
569 			preempt_enable_no_resched();
570 			return;
571 		}
572 		raw_spin_unlock(lock);
573 	}
574 }
575 
raw_spin_rq_trylock(struct rq * rq)576 bool raw_spin_rq_trylock(struct rq *rq)
577 {
578 	raw_spinlock_t *lock;
579 	bool ret;
580 
581 	/* Matches synchronize_rcu() in __sched_core_enable() */
582 	preempt_disable();
583 	if (sched_core_disabled()) {
584 		ret = raw_spin_trylock(&rq->__lock);
585 		preempt_enable();
586 		return ret;
587 	}
588 
589 	for (;;) {
590 		lock = __rq_lockp(rq);
591 		ret = raw_spin_trylock(lock);
592 		if (!ret || (likely(lock == __rq_lockp(rq)))) {
593 			preempt_enable();
594 			return ret;
595 		}
596 		raw_spin_unlock(lock);
597 	}
598 }
599 
raw_spin_rq_unlock(struct rq * rq)600 void raw_spin_rq_unlock(struct rq *rq)
601 {
602 	raw_spin_unlock(rq_lockp(rq));
603 }
604 
605 #ifdef CONFIG_SMP
606 /*
607  * double_rq_lock - safely lock two runqueues
608  */
double_rq_lock(struct rq * rq1,struct rq * rq2)609 void double_rq_lock(struct rq *rq1, struct rq *rq2)
610 {
611 	lockdep_assert_irqs_disabled();
612 
613 	if (rq_order_less(rq2, rq1))
614 		swap(rq1, rq2);
615 
616 	raw_spin_rq_lock(rq1);
617 	if (__rq_lockp(rq1) != __rq_lockp(rq2))
618 		raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
619 
620 	double_rq_clock_clear_update(rq1, rq2);
621 }
622 #endif
623 
624 /*
625  * __task_rq_lock - lock the rq @p resides on.
626  */
__task_rq_lock(struct task_struct * p,struct rq_flags * rf)627 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
628 	__acquires(rq->lock)
629 {
630 	struct rq *rq;
631 
632 	lockdep_assert_held(&p->pi_lock);
633 
634 	for (;;) {
635 		rq = task_rq(p);
636 		raw_spin_rq_lock(rq);
637 		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
638 			rq_pin_lock(rq, rf);
639 			return rq;
640 		}
641 		raw_spin_rq_unlock(rq);
642 
643 		while (unlikely(task_on_rq_migrating(p)))
644 			cpu_relax();
645 	}
646 }
647 
648 /*
649  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
650  */
task_rq_lock(struct task_struct * p,struct rq_flags * rf)651 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
652 	__acquires(p->pi_lock)
653 	__acquires(rq->lock)
654 {
655 	struct rq *rq;
656 
657 	for (;;) {
658 		raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
659 		rq = task_rq(p);
660 		raw_spin_rq_lock(rq);
661 		/*
662 		 *	move_queued_task()		task_rq_lock()
663 		 *
664 		 *	ACQUIRE (rq->lock)
665 		 *	[S] ->on_rq = MIGRATING		[L] rq = task_rq()
666 		 *	WMB (__set_task_cpu())		ACQUIRE (rq->lock);
667 		 *	[S] ->cpu = new_cpu		[L] task_rq()
668 		 *					[L] ->on_rq
669 		 *	RELEASE (rq->lock)
670 		 *
671 		 * If we observe the old CPU in task_rq_lock(), the acquire of
672 		 * the old rq->lock will fully serialize against the stores.
673 		 *
674 		 * If we observe the new CPU in task_rq_lock(), the address
675 		 * dependency headed by '[L] rq = task_rq()' and the acquire
676 		 * will pair with the WMB to ensure we then also see migrating.
677 		 */
678 		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
679 			rq_pin_lock(rq, rf);
680 			return rq;
681 		}
682 		raw_spin_rq_unlock(rq);
683 		raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
684 
685 		while (unlikely(task_on_rq_migrating(p)))
686 			cpu_relax();
687 	}
688 }
689 
690 /*
691  * RQ-clock updating methods:
692  */
693 
update_rq_clock_task(struct rq * rq,s64 delta)694 static void update_rq_clock_task(struct rq *rq, s64 delta)
695 {
696 /*
697  * In theory, the compile should just see 0 here, and optimize out the call
698  * to sched_rt_avg_update. But I don't trust it...
699  */
700 	s64 __maybe_unused steal = 0, irq_delta = 0;
701 
702 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
703 	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
704 
705 	/*
706 	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
707 	 * this case when a previous update_rq_clock() happened inside a
708 	 * {soft,}irq region.
709 	 *
710 	 * When this happens, we stop ->clock_task and only update the
711 	 * prev_irq_time stamp to account for the part that fit, so that a next
712 	 * update will consume the rest. This ensures ->clock_task is
713 	 * monotonic.
714 	 *
715 	 * It does however cause some slight miss-attribution of {soft,}irq
716 	 * time, a more accurate solution would be to update the irq_time using
717 	 * the current rq->clock timestamp, except that would require using
718 	 * atomic ops.
719 	 */
720 	if (irq_delta > delta)
721 		irq_delta = delta;
722 
723 	rq->prev_irq_time += irq_delta;
724 	delta -= irq_delta;
725 	psi_account_irqtime(rq->curr, irq_delta);
726 	delayacct_irq(rq->curr, irq_delta);
727 #endif
728 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
729 	if (static_key_false((&paravirt_steal_rq_enabled))) {
730 		steal = paravirt_steal_clock(cpu_of(rq));
731 		steal -= rq->prev_steal_time_rq;
732 
733 		if (unlikely(steal > delta))
734 			steal = delta;
735 
736 		rq->prev_steal_time_rq += steal;
737 		delta -= steal;
738 	}
739 #endif
740 
741 	rq->clock_task += delta;
742 
743 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
744 	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
745 		update_irq_load_avg(rq, irq_delta + steal);
746 #endif
747 	update_rq_clock_pelt(rq, delta);
748 }
749 
update_rq_clock(struct rq * rq)750 void update_rq_clock(struct rq *rq)
751 {
752 	s64 delta;
753 
754 	lockdep_assert_rq_held(rq);
755 
756 	if (rq->clock_update_flags & RQCF_ACT_SKIP)
757 		return;
758 
759 #ifdef CONFIG_SCHED_DEBUG
760 	if (sched_feat(WARN_DOUBLE_CLOCK))
761 		SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
762 	rq->clock_update_flags |= RQCF_UPDATED;
763 #endif
764 
765 	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
766 	if (delta < 0)
767 		return;
768 	rq->clock += delta;
769 	update_rq_clock_task(rq, delta);
770 }
771 
772 #ifdef CONFIG_SCHED_HRTICK
773 /*
774  * Use HR-timers to deliver accurate preemption points.
775  */
776 
hrtick_clear(struct rq * rq)777 static void hrtick_clear(struct rq *rq)
778 {
779 	if (hrtimer_active(&rq->hrtick_timer))
780 		hrtimer_cancel(&rq->hrtick_timer);
781 }
782 
783 /*
784  * High-resolution timer tick.
785  * Runs from hardirq context with interrupts disabled.
786  */
hrtick(struct hrtimer * timer)787 static enum hrtimer_restart hrtick(struct hrtimer *timer)
788 {
789 	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
790 	struct rq_flags rf;
791 
792 	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
793 
794 	rq_lock(rq, &rf);
795 	update_rq_clock(rq);
796 	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
797 	rq_unlock(rq, &rf);
798 
799 	return HRTIMER_NORESTART;
800 }
801 
802 #ifdef CONFIG_SMP
803 
__hrtick_restart(struct rq * rq)804 static void __hrtick_restart(struct rq *rq)
805 {
806 	struct hrtimer *timer = &rq->hrtick_timer;
807 	ktime_t time = rq->hrtick_time;
808 
809 	hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
810 }
811 
812 /*
813  * called from hardirq (IPI) context
814  */
__hrtick_start(void * arg)815 static void __hrtick_start(void *arg)
816 {
817 	struct rq *rq = arg;
818 	struct rq_flags rf;
819 
820 	rq_lock(rq, &rf);
821 	__hrtick_restart(rq);
822 	rq_unlock(rq, &rf);
823 }
824 
825 /*
826  * Called to set the hrtick timer state.
827  *
828  * called with rq->lock held and irqs disabled
829  */
hrtick_start(struct rq * rq,u64 delay)830 void hrtick_start(struct rq *rq, u64 delay)
831 {
832 	struct hrtimer *timer = &rq->hrtick_timer;
833 	s64 delta;
834 
835 	/*
836 	 * Don't schedule slices shorter than 10000ns, that just
837 	 * doesn't make sense and can cause timer DoS.
838 	 */
839 	delta = max_t(s64, delay, 10000LL);
840 	rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
841 
842 	if (rq == this_rq())
843 		__hrtick_restart(rq);
844 	else
845 		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
846 }
847 
848 #else
849 /*
850  * Called to set the hrtick timer state.
851  *
852  * called with rq->lock held and irqs disabled
853  */
hrtick_start(struct rq * rq,u64 delay)854 void hrtick_start(struct rq *rq, u64 delay)
855 {
856 	/*
857 	 * Don't schedule slices shorter than 10000ns, that just
858 	 * doesn't make sense. Rely on vruntime for fairness.
859 	 */
860 	delay = max_t(u64, delay, 10000LL);
861 	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
862 		      HRTIMER_MODE_REL_PINNED_HARD);
863 }
864 
865 #endif /* CONFIG_SMP */
866 
hrtick_rq_init(struct rq * rq)867 static void hrtick_rq_init(struct rq *rq)
868 {
869 #ifdef CONFIG_SMP
870 	INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
871 #endif
872 	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
873 	rq->hrtick_timer.function = hrtick;
874 }
875 #else	/* CONFIG_SCHED_HRTICK */
hrtick_clear(struct rq * rq)876 static inline void hrtick_clear(struct rq *rq)
877 {
878 }
879 
hrtick_rq_init(struct rq * rq)880 static inline void hrtick_rq_init(struct rq *rq)
881 {
882 }
883 #endif	/* CONFIG_SCHED_HRTICK */
884 
885 /*
886  * cmpxchg based fetch_or, macro so it works for different integer types
887  */
888 #define fetch_or(ptr, mask)						\
889 	({								\
890 		typeof(ptr) _ptr = (ptr);				\
891 		typeof(mask) _mask = (mask);				\
892 		typeof(*_ptr) _val = *_ptr;				\
893 									\
894 		do {							\
895 		} while (!try_cmpxchg(_ptr, &_val, _val | _mask));	\
896 	_val;								\
897 })
898 
899 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
900 /*
901  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
902  * this avoids any races wrt polling state changes and thereby avoids
903  * spurious IPIs.
904  */
set_nr_and_not_polling(struct task_struct * p)905 static inline bool set_nr_and_not_polling(struct task_struct *p)
906 {
907 	struct thread_info *ti = task_thread_info(p);
908 	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
909 }
910 
911 /*
912  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
913  *
914  * If this returns true, then the idle task promises to call
915  * sched_ttwu_pending() and reschedule soon.
916  */
set_nr_if_polling(struct task_struct * p)917 static bool set_nr_if_polling(struct task_struct *p)
918 {
919 	struct thread_info *ti = task_thread_info(p);
920 	typeof(ti->flags) val = READ_ONCE(ti->flags);
921 
922 	for (;;) {
923 		if (!(val & _TIF_POLLING_NRFLAG))
924 			return false;
925 		if (val & _TIF_NEED_RESCHED)
926 			return true;
927 		if (try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED))
928 			break;
929 	}
930 	return true;
931 }
932 
933 #else
set_nr_and_not_polling(struct task_struct * p)934 static inline bool set_nr_and_not_polling(struct task_struct *p)
935 {
936 	set_tsk_need_resched(p);
937 	return true;
938 }
939 
940 #ifdef CONFIG_SMP
set_nr_if_polling(struct task_struct * p)941 static inline bool set_nr_if_polling(struct task_struct *p)
942 {
943 	return false;
944 }
945 #endif
946 #endif
947 
__wake_q_add(struct wake_q_head * head,struct task_struct * task)948 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
949 {
950 	struct wake_q_node *node = &task->wake_q;
951 
952 	/*
953 	 * Atomically grab the task, if ->wake_q is !nil already it means
954 	 * it's already queued (either by us or someone else) and will get the
955 	 * wakeup due to that.
956 	 *
957 	 * In order to ensure that a pending wakeup will observe our pending
958 	 * state, even in the failed case, an explicit smp_mb() must be used.
959 	 */
960 	smp_mb__before_atomic();
961 	if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
962 		return false;
963 
964 	/*
965 	 * The head is context local, there can be no concurrency.
966 	 */
967 	*head->lastp = node;
968 	head->lastp = &node->next;
969 	return true;
970 }
971 
972 /**
973  * wake_q_add() - queue a wakeup for 'later' waking.
974  * @head: the wake_q_head to add @task to
975  * @task: the task to queue for 'later' wakeup
976  *
977  * Queue a task for later wakeup, most likely by the wake_up_q() call in the
978  * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
979  * instantly.
980  *
981  * This function must be used as-if it were wake_up_process(); IOW the task
982  * must be ready to be woken at this location.
983  */
wake_q_add(struct wake_q_head * head,struct task_struct * task)984 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
985 {
986 	if (__wake_q_add(head, task))
987 		get_task_struct(task);
988 }
989 
990 /**
991  * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
992  * @head: the wake_q_head to add @task to
993  * @task: the task to queue for 'later' wakeup
994  *
995  * Queue a task for later wakeup, most likely by the wake_up_q() call in the
996  * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
997  * instantly.
998  *
999  * This function must be used as-if it were wake_up_process(); IOW the task
1000  * must be ready to be woken at this location.
1001  *
1002  * This function is essentially a task-safe equivalent to wake_q_add(). Callers
1003  * that already hold reference to @task can call the 'safe' version and trust
1004  * wake_q to do the right thing depending whether or not the @task is already
1005  * queued for wakeup.
1006  */
wake_q_add_safe(struct wake_q_head * head,struct task_struct * task)1007 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
1008 {
1009 	if (!__wake_q_add(head, task))
1010 		put_task_struct(task);
1011 }
1012 
wake_up_q(struct wake_q_head * head)1013 void wake_up_q(struct wake_q_head *head)
1014 {
1015 	struct wake_q_node *node = head->first;
1016 
1017 	while (node != WAKE_Q_TAIL) {
1018 		struct task_struct *task;
1019 
1020 		task = container_of(node, struct task_struct, wake_q);
1021 		/* Task can safely be re-inserted now: */
1022 		node = node->next;
1023 		task->wake_q.next = NULL;
1024 
1025 		/*
1026 		 * wake_up_process() executes a full barrier, which pairs with
1027 		 * the queueing in wake_q_add() so as not to miss wakeups.
1028 		 */
1029 		wake_up_process(task);
1030 		put_task_struct(task);
1031 	}
1032 }
1033 
1034 /*
1035  * resched_curr - mark rq's current task 'to be rescheduled now'.
1036  *
1037  * On UP this means the setting of the need_resched flag, on SMP it
1038  * might also involve a cross-CPU call to trigger the scheduler on
1039  * the target CPU.
1040  */
resched_curr(struct rq * rq)1041 void resched_curr(struct rq *rq)
1042 {
1043 	struct task_struct *curr = rq->curr;
1044 	int cpu;
1045 
1046 	lockdep_assert_rq_held(rq);
1047 
1048 	if (test_tsk_need_resched(curr))
1049 		return;
1050 
1051 	cpu = cpu_of(rq);
1052 
1053 	if (cpu == smp_processor_id()) {
1054 		set_tsk_need_resched(curr);
1055 		set_preempt_need_resched();
1056 		return;
1057 	}
1058 
1059 	if (set_nr_and_not_polling(curr))
1060 		smp_send_reschedule(cpu);
1061 	else
1062 		trace_sched_wake_idle_without_ipi(cpu);
1063 }
1064 
resched_cpu(int cpu)1065 void resched_cpu(int cpu)
1066 {
1067 	struct rq *rq = cpu_rq(cpu);
1068 	unsigned long flags;
1069 
1070 	raw_spin_rq_lock_irqsave(rq, flags);
1071 	if (cpu_online(cpu) || cpu == smp_processor_id())
1072 		resched_curr(rq);
1073 	raw_spin_rq_unlock_irqrestore(rq, flags);
1074 }
1075 
1076 #ifdef CONFIG_SMP
1077 #ifdef CONFIG_NO_HZ_COMMON
1078 /*
1079  * In the semi idle case, use the nearest busy CPU for migrating timers
1080  * from an idle CPU.  This is good for power-savings.
1081  *
1082  * We don't do similar optimization for completely idle system, as
1083  * selecting an idle CPU will add more delays to the timers than intended
1084  * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1085  */
get_nohz_timer_target(void)1086 int get_nohz_timer_target(void)
1087 {
1088 	int i, cpu = smp_processor_id(), default_cpu = -1;
1089 	struct sched_domain *sd;
1090 	const struct cpumask *hk_mask;
1091 
1092 	if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1093 		if (!idle_cpu(cpu))
1094 			return cpu;
1095 		default_cpu = cpu;
1096 	}
1097 
1098 	hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1099 
1100 	guard(rcu)();
1101 
1102 	for_each_domain(cpu, sd) {
1103 		for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1104 			if (cpu == i)
1105 				continue;
1106 
1107 			if (!idle_cpu(i))
1108 				return i;
1109 		}
1110 	}
1111 
1112 	if (default_cpu == -1)
1113 		default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1114 
1115 	return default_cpu;
1116 }
1117 
1118 /*
1119  * When add_timer_on() enqueues a timer into the timer wheel of an
1120  * idle CPU then this timer might expire before the next timer event
1121  * which is scheduled to wake up that CPU. In case of a completely
1122  * idle system the next event might even be infinite time into the
1123  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1124  * leaves the inner idle loop so the newly added timer is taken into
1125  * account when the CPU goes back to idle and evaluates the timer
1126  * wheel for the next timer event.
1127  */
wake_up_idle_cpu(int cpu)1128 static void wake_up_idle_cpu(int cpu)
1129 {
1130 	struct rq *rq = cpu_rq(cpu);
1131 
1132 	if (cpu == smp_processor_id())
1133 		return;
1134 
1135 	if (set_nr_and_not_polling(rq->idle))
1136 		smp_send_reschedule(cpu);
1137 	else
1138 		trace_sched_wake_idle_without_ipi(cpu);
1139 }
1140 
wake_up_full_nohz_cpu(int cpu)1141 static bool wake_up_full_nohz_cpu(int cpu)
1142 {
1143 	/*
1144 	 * We just need the target to call irq_exit() and re-evaluate
1145 	 * the next tick. The nohz full kick at least implies that.
1146 	 * If needed we can still optimize that later with an
1147 	 * empty IRQ.
1148 	 */
1149 	if (cpu_is_offline(cpu))
1150 		return true;  /* Don't try to wake offline CPUs. */
1151 	if (tick_nohz_full_cpu(cpu)) {
1152 		if (cpu != smp_processor_id() ||
1153 		    tick_nohz_tick_stopped())
1154 			tick_nohz_full_kick_cpu(cpu);
1155 		return true;
1156 	}
1157 
1158 	return false;
1159 }
1160 
1161 /*
1162  * Wake up the specified CPU.  If the CPU is going offline, it is the
1163  * caller's responsibility to deal with the lost wakeup, for example,
1164  * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1165  */
wake_up_nohz_cpu(int cpu)1166 void wake_up_nohz_cpu(int cpu)
1167 {
1168 	if (!wake_up_full_nohz_cpu(cpu))
1169 		wake_up_idle_cpu(cpu);
1170 }
1171 
nohz_csd_func(void * info)1172 static void nohz_csd_func(void *info)
1173 {
1174 	struct rq *rq = info;
1175 	int cpu = cpu_of(rq);
1176 	unsigned int flags;
1177 
1178 	/*
1179 	 * Release the rq::nohz_csd.
1180 	 */
1181 	flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1182 	WARN_ON(!(flags & NOHZ_KICK_MASK));
1183 
1184 	rq->idle_balance = idle_cpu(cpu);
1185 	if (rq->idle_balance && !need_resched()) {
1186 		rq->nohz_idle_balance = flags;
1187 		raise_softirq_irqoff(SCHED_SOFTIRQ);
1188 	}
1189 }
1190 
1191 #endif /* CONFIG_NO_HZ_COMMON */
1192 
1193 #ifdef CONFIG_NO_HZ_FULL
__need_bw_check(struct rq * rq,struct task_struct * p)1194 static inline bool __need_bw_check(struct rq *rq, struct task_struct *p)
1195 {
1196 	if (rq->nr_running != 1)
1197 		return false;
1198 
1199 	if (p->sched_class != &fair_sched_class)
1200 		return false;
1201 
1202 	if (!task_on_rq_queued(p))
1203 		return false;
1204 
1205 	return true;
1206 }
1207 
sched_can_stop_tick(struct rq * rq)1208 bool sched_can_stop_tick(struct rq *rq)
1209 {
1210 	int fifo_nr_running;
1211 
1212 	/* Deadline tasks, even if single, need the tick */
1213 	if (rq->dl.dl_nr_running)
1214 		return false;
1215 
1216 	/*
1217 	 * If there are more than one RR tasks, we need the tick to affect the
1218 	 * actual RR behaviour.
1219 	 */
1220 	if (rq->rt.rr_nr_running) {
1221 		if (rq->rt.rr_nr_running == 1)
1222 			return true;
1223 		else
1224 			return false;
1225 	}
1226 
1227 	/*
1228 	 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1229 	 * forced preemption between FIFO tasks.
1230 	 */
1231 	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1232 	if (fifo_nr_running)
1233 		return true;
1234 
1235 	/*
1236 	 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1237 	 * if there's more than one we need the tick for involuntary
1238 	 * preemption.
1239 	 */
1240 	if (rq->nr_running > 1)
1241 		return false;
1242 
1243 	/*
1244 	 * If there is one task and it has CFS runtime bandwidth constraints
1245 	 * and it's on the cpu now we don't want to stop the tick.
1246 	 * This check prevents clearing the bit if a newly enqueued task here is
1247 	 * dequeued by migrating while the constrained task continues to run.
1248 	 * E.g. going from 2->1 without going through pick_next_task().
1249 	 */
1250 	if (sched_feat(HZ_BW) && __need_bw_check(rq, rq->curr)) {
1251 		if (cfs_task_bw_constrained(rq->curr))
1252 			return false;
1253 	}
1254 
1255 	return true;
1256 }
1257 #endif /* CONFIG_NO_HZ_FULL */
1258 #endif /* CONFIG_SMP */
1259 
1260 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1261 			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1262 /*
1263  * Iterate task_group tree rooted at *from, calling @down when first entering a
1264  * node and @up when leaving it for the final time.
1265  *
1266  * Caller must hold rcu_lock or sufficient equivalent.
1267  */
walk_tg_tree_from(struct task_group * from,tg_visitor down,tg_visitor up,void * data)1268 int walk_tg_tree_from(struct task_group *from,
1269 			     tg_visitor down, tg_visitor up, void *data)
1270 {
1271 	struct task_group *parent, *child;
1272 	int ret;
1273 
1274 	parent = from;
1275 
1276 down:
1277 	ret = (*down)(parent, data);
1278 	if (ret)
1279 		goto out;
1280 	list_for_each_entry_rcu(child, &parent->children, siblings) {
1281 		parent = child;
1282 		goto down;
1283 
1284 up:
1285 		continue;
1286 	}
1287 	ret = (*up)(parent, data);
1288 	if (ret || parent == from)
1289 		goto out;
1290 
1291 	child = parent;
1292 	parent = parent->parent;
1293 	if (parent)
1294 		goto up;
1295 out:
1296 	return ret;
1297 }
1298 
tg_nop(struct task_group * tg,void * data)1299 int tg_nop(struct task_group *tg, void *data)
1300 {
1301 	return 0;
1302 }
1303 #endif
1304 
set_load_weight(struct task_struct * p,bool update_load)1305 static void set_load_weight(struct task_struct *p, bool update_load)
1306 {
1307 	int prio = p->static_prio - MAX_RT_PRIO;
1308 	struct load_weight *load = &p->se.load;
1309 
1310 	/*
1311 	 * SCHED_IDLE tasks get minimal weight:
1312 	 */
1313 	if (task_has_idle_policy(p)) {
1314 		load->weight = scale_load(WEIGHT_IDLEPRIO);
1315 		load->inv_weight = WMULT_IDLEPRIO;
1316 		return;
1317 	}
1318 
1319 	/*
1320 	 * SCHED_OTHER tasks have to update their load when changing their
1321 	 * weight
1322 	 */
1323 	if (update_load && p->sched_class == &fair_sched_class) {
1324 		reweight_task(p, prio);
1325 	} else {
1326 		load->weight = scale_load(sched_prio_to_weight[prio]);
1327 		load->inv_weight = sched_prio_to_wmult[prio];
1328 	}
1329 }
1330 
1331 #ifdef CONFIG_UCLAMP_TASK
1332 /*
1333  * Serializes updates of utilization clamp values
1334  *
1335  * The (slow-path) user-space triggers utilization clamp value updates which
1336  * can require updates on (fast-path) scheduler's data structures used to
1337  * support enqueue/dequeue operations.
1338  * While the per-CPU rq lock protects fast-path update operations, user-space
1339  * requests are serialized using a mutex to reduce the risk of conflicting
1340  * updates or API abuses.
1341  */
1342 static DEFINE_MUTEX(uclamp_mutex);
1343 
1344 /* Max allowed minimum utilization */
1345 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1346 
1347 /* Max allowed maximum utilization */
1348 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1349 
1350 /*
1351  * By default RT tasks run at the maximum performance point/capacity of the
1352  * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1353  * SCHED_CAPACITY_SCALE.
1354  *
1355  * This knob allows admins to change the default behavior when uclamp is being
1356  * used. In battery powered devices, particularly, running at the maximum
1357  * capacity and frequency will increase energy consumption and shorten the
1358  * battery life.
1359  *
1360  * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1361  *
1362  * This knob will not override the system default sched_util_clamp_min defined
1363  * above.
1364  */
1365 static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1366 
1367 /* All clamps are required to be less or equal than these values */
1368 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1369 
1370 /*
1371  * This static key is used to reduce the uclamp overhead in the fast path. It
1372  * primarily disables the call to uclamp_rq_{inc, dec}() in
1373  * enqueue/dequeue_task().
1374  *
1375  * This allows users to continue to enable uclamp in their kernel config with
1376  * minimum uclamp overhead in the fast path.
1377  *
1378  * As soon as userspace modifies any of the uclamp knobs, the static key is
1379  * enabled, since we have an actual users that make use of uclamp
1380  * functionality.
1381  *
1382  * The knobs that would enable this static key are:
1383  *
1384  *   * A task modifying its uclamp value with sched_setattr().
1385  *   * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1386  *   * An admin modifying the cgroup cpu.uclamp.{min, max}
1387  */
1388 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1389 
1390 /* Integer rounded range for each bucket */
1391 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1392 
1393 #define for_each_clamp_id(clamp_id) \
1394 	for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1395 
uclamp_bucket_id(unsigned int clamp_value)1396 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1397 {
1398 	return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1399 }
1400 
uclamp_none(enum uclamp_id clamp_id)1401 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1402 {
1403 	if (clamp_id == UCLAMP_MIN)
1404 		return 0;
1405 	return SCHED_CAPACITY_SCALE;
1406 }
1407 
uclamp_se_set(struct uclamp_se * uc_se,unsigned int value,bool user_defined)1408 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1409 				 unsigned int value, bool user_defined)
1410 {
1411 	uc_se->value = value;
1412 	uc_se->bucket_id = uclamp_bucket_id(value);
1413 	uc_se->user_defined = user_defined;
1414 }
1415 
1416 static inline unsigned int
uclamp_idle_value(struct rq * rq,enum uclamp_id clamp_id,unsigned int clamp_value)1417 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1418 		  unsigned int clamp_value)
1419 {
1420 	/*
1421 	 * Avoid blocked utilization pushing up the frequency when we go
1422 	 * idle (which drops the max-clamp) by retaining the last known
1423 	 * max-clamp.
1424 	 */
1425 	if (clamp_id == UCLAMP_MAX) {
1426 		rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1427 		return clamp_value;
1428 	}
1429 
1430 	return uclamp_none(UCLAMP_MIN);
1431 }
1432 
uclamp_idle_reset(struct rq * rq,enum uclamp_id clamp_id,unsigned int clamp_value)1433 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1434 				     unsigned int clamp_value)
1435 {
1436 	/* Reset max-clamp retention only on idle exit */
1437 	if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1438 		return;
1439 
1440 	uclamp_rq_set(rq, clamp_id, clamp_value);
1441 }
1442 
1443 static inline
uclamp_rq_max_value(struct rq * rq,enum uclamp_id clamp_id,unsigned int clamp_value)1444 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1445 				   unsigned int clamp_value)
1446 {
1447 	struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1448 	int bucket_id = UCLAMP_BUCKETS - 1;
1449 
1450 	/*
1451 	 * Since both min and max clamps are max aggregated, find the
1452 	 * top most bucket with tasks in.
1453 	 */
1454 	for ( ; bucket_id >= 0; bucket_id--) {
1455 		if (!bucket[bucket_id].tasks)
1456 			continue;
1457 		return bucket[bucket_id].value;
1458 	}
1459 
1460 	/* No tasks -- default clamp values */
1461 	return uclamp_idle_value(rq, clamp_id, clamp_value);
1462 }
1463 
__uclamp_update_util_min_rt_default(struct task_struct * p)1464 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1465 {
1466 	unsigned int default_util_min;
1467 	struct uclamp_se *uc_se;
1468 
1469 	lockdep_assert_held(&p->pi_lock);
1470 
1471 	uc_se = &p->uclamp_req[UCLAMP_MIN];
1472 
1473 	/* Only sync if user didn't override the default */
1474 	if (uc_se->user_defined)
1475 		return;
1476 
1477 	default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1478 	uclamp_se_set(uc_se, default_util_min, false);
1479 }
1480 
uclamp_update_util_min_rt_default(struct task_struct * p)1481 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1482 {
1483 	struct rq_flags rf;
1484 	struct rq *rq;
1485 
1486 	if (!rt_task(p))
1487 		return;
1488 
1489 	/* Protect updates to p->uclamp_* */
1490 	rq = task_rq_lock(p, &rf);
1491 	__uclamp_update_util_min_rt_default(p);
1492 	task_rq_unlock(rq, p, &rf);
1493 }
1494 
1495 static inline struct uclamp_se
uclamp_tg_restrict(struct task_struct * p,enum uclamp_id clamp_id)1496 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1497 {
1498 	/* Copy by value as we could modify it */
1499 	struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1500 #ifdef CONFIG_UCLAMP_TASK_GROUP
1501 	unsigned int tg_min, tg_max, value;
1502 
1503 	/*
1504 	 * Tasks in autogroups or root task group will be
1505 	 * restricted by system defaults.
1506 	 */
1507 	if (task_group_is_autogroup(task_group(p)))
1508 		return uc_req;
1509 	if (task_group(p) == &root_task_group)
1510 		return uc_req;
1511 
1512 	tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1513 	tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1514 	value = uc_req.value;
1515 	value = clamp(value, tg_min, tg_max);
1516 	uclamp_se_set(&uc_req, value, false);
1517 #endif
1518 
1519 	return uc_req;
1520 }
1521 
1522 /*
1523  * The effective clamp bucket index of a task depends on, by increasing
1524  * priority:
1525  * - the task specific clamp value, when explicitly requested from userspace
1526  * - the task group effective clamp value, for tasks not either in the root
1527  *   group or in an autogroup
1528  * - the system default clamp value, defined by the sysadmin
1529  */
1530 static inline struct uclamp_se
uclamp_eff_get(struct task_struct * p,enum uclamp_id clamp_id)1531 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1532 {
1533 	struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1534 	struct uclamp_se uc_max = uclamp_default[clamp_id];
1535 
1536 	/* System default restrictions always apply */
1537 	if (unlikely(uc_req.value > uc_max.value))
1538 		return uc_max;
1539 
1540 	return uc_req;
1541 }
1542 
uclamp_eff_value(struct task_struct * p,enum uclamp_id clamp_id)1543 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1544 {
1545 	struct uclamp_se uc_eff;
1546 
1547 	/* Task currently refcounted: use back-annotated (effective) value */
1548 	if (p->uclamp[clamp_id].active)
1549 		return (unsigned long)p->uclamp[clamp_id].value;
1550 
1551 	uc_eff = uclamp_eff_get(p, clamp_id);
1552 
1553 	return (unsigned long)uc_eff.value;
1554 }
1555 
1556 /*
1557  * When a task is enqueued on a rq, the clamp bucket currently defined by the
1558  * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1559  * updates the rq's clamp value if required.
1560  *
1561  * Tasks can have a task-specific value requested from user-space, track
1562  * within each bucket the maximum value for tasks refcounted in it.
1563  * This "local max aggregation" allows to track the exact "requested" value
1564  * for each bucket when all its RUNNABLE tasks require the same clamp.
1565  */
uclamp_rq_inc_id(struct rq * rq,struct task_struct * p,enum uclamp_id clamp_id)1566 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1567 				    enum uclamp_id clamp_id)
1568 {
1569 	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1570 	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1571 	struct uclamp_bucket *bucket;
1572 
1573 	lockdep_assert_rq_held(rq);
1574 
1575 	/* Update task effective clamp */
1576 	p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1577 
1578 	bucket = &uc_rq->bucket[uc_se->bucket_id];
1579 	bucket->tasks++;
1580 	uc_se->active = true;
1581 
1582 	uclamp_idle_reset(rq, clamp_id, uc_se->value);
1583 
1584 	/*
1585 	 * Local max aggregation: rq buckets always track the max
1586 	 * "requested" clamp value of its RUNNABLE tasks.
1587 	 */
1588 	if (bucket->tasks == 1 || uc_se->value > bucket->value)
1589 		bucket->value = uc_se->value;
1590 
1591 	if (uc_se->value > uclamp_rq_get(rq, clamp_id))
1592 		uclamp_rq_set(rq, clamp_id, uc_se->value);
1593 }
1594 
1595 /*
1596  * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1597  * is released. If this is the last task reference counting the rq's max
1598  * active clamp value, then the rq's clamp value is updated.
1599  *
1600  * Both refcounted tasks and rq's cached clamp values are expected to be
1601  * always valid. If it's detected they are not, as defensive programming,
1602  * enforce the expected state and warn.
1603  */
uclamp_rq_dec_id(struct rq * rq,struct task_struct * p,enum uclamp_id clamp_id)1604 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1605 				    enum uclamp_id clamp_id)
1606 {
1607 	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1608 	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1609 	struct uclamp_bucket *bucket;
1610 	unsigned int bkt_clamp;
1611 	unsigned int rq_clamp;
1612 
1613 	lockdep_assert_rq_held(rq);
1614 
1615 	/*
1616 	 * If sched_uclamp_used was enabled after task @p was enqueued,
1617 	 * we could end up with unbalanced call to uclamp_rq_dec_id().
1618 	 *
1619 	 * In this case the uc_se->active flag should be false since no uclamp
1620 	 * accounting was performed at enqueue time and we can just return
1621 	 * here.
1622 	 *
1623 	 * Need to be careful of the following enqueue/dequeue ordering
1624 	 * problem too
1625 	 *
1626 	 *	enqueue(taskA)
1627 	 *	// sched_uclamp_used gets enabled
1628 	 *	enqueue(taskB)
1629 	 *	dequeue(taskA)
1630 	 *	// Must not decrement bucket->tasks here
1631 	 *	dequeue(taskB)
1632 	 *
1633 	 * where we could end up with stale data in uc_se and
1634 	 * bucket[uc_se->bucket_id].
1635 	 *
1636 	 * The following check here eliminates the possibility of such race.
1637 	 */
1638 	if (unlikely(!uc_se->active))
1639 		return;
1640 
1641 	bucket = &uc_rq->bucket[uc_se->bucket_id];
1642 
1643 	SCHED_WARN_ON(!bucket->tasks);
1644 	if (likely(bucket->tasks))
1645 		bucket->tasks--;
1646 
1647 	uc_se->active = false;
1648 
1649 	/*
1650 	 * Keep "local max aggregation" simple and accept to (possibly)
1651 	 * overboost some RUNNABLE tasks in the same bucket.
1652 	 * The rq clamp bucket value is reset to its base value whenever
1653 	 * there are no more RUNNABLE tasks refcounting it.
1654 	 */
1655 	if (likely(bucket->tasks))
1656 		return;
1657 
1658 	rq_clamp = uclamp_rq_get(rq, clamp_id);
1659 	/*
1660 	 * Defensive programming: this should never happen. If it happens,
1661 	 * e.g. due to future modification, warn and fixup the expected value.
1662 	 */
1663 	SCHED_WARN_ON(bucket->value > rq_clamp);
1664 	if (bucket->value >= rq_clamp) {
1665 		bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1666 		uclamp_rq_set(rq, clamp_id, bkt_clamp);
1667 	}
1668 }
1669 
uclamp_rq_inc(struct rq * rq,struct task_struct * p)1670 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1671 {
1672 	enum uclamp_id clamp_id;
1673 
1674 	/*
1675 	 * Avoid any overhead until uclamp is actually used by the userspace.
1676 	 *
1677 	 * The condition is constructed such that a NOP is generated when
1678 	 * sched_uclamp_used is disabled.
1679 	 */
1680 	if (!static_branch_unlikely(&sched_uclamp_used))
1681 		return;
1682 
1683 	if (unlikely(!p->sched_class->uclamp_enabled))
1684 		return;
1685 
1686 	for_each_clamp_id(clamp_id)
1687 		uclamp_rq_inc_id(rq, p, clamp_id);
1688 
1689 	/* Reset clamp idle holding when there is one RUNNABLE task */
1690 	if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1691 		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1692 }
1693 
uclamp_rq_dec(struct rq * rq,struct task_struct * p)1694 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1695 {
1696 	enum uclamp_id clamp_id;
1697 
1698 	/*
1699 	 * Avoid any overhead until uclamp is actually used by the userspace.
1700 	 *
1701 	 * The condition is constructed such that a NOP is generated when
1702 	 * sched_uclamp_used is disabled.
1703 	 */
1704 	if (!static_branch_unlikely(&sched_uclamp_used))
1705 		return;
1706 
1707 	if (unlikely(!p->sched_class->uclamp_enabled))
1708 		return;
1709 
1710 	for_each_clamp_id(clamp_id)
1711 		uclamp_rq_dec_id(rq, p, clamp_id);
1712 }
1713 
uclamp_rq_reinc_id(struct rq * rq,struct task_struct * p,enum uclamp_id clamp_id)1714 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1715 				      enum uclamp_id clamp_id)
1716 {
1717 	if (!p->uclamp[clamp_id].active)
1718 		return;
1719 
1720 	uclamp_rq_dec_id(rq, p, clamp_id);
1721 	uclamp_rq_inc_id(rq, p, clamp_id);
1722 
1723 	/*
1724 	 * Make sure to clear the idle flag if we've transiently reached 0
1725 	 * active tasks on rq.
1726 	 */
1727 	if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1728 		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1729 }
1730 
1731 static inline void
uclamp_update_active(struct task_struct * p)1732 uclamp_update_active(struct task_struct *p)
1733 {
1734 	enum uclamp_id clamp_id;
1735 	struct rq_flags rf;
1736 	struct rq *rq;
1737 
1738 	/*
1739 	 * Lock the task and the rq where the task is (or was) queued.
1740 	 *
1741 	 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1742 	 * price to pay to safely serialize util_{min,max} updates with
1743 	 * enqueues, dequeues and migration operations.
1744 	 * This is the same locking schema used by __set_cpus_allowed_ptr().
1745 	 */
1746 	rq = task_rq_lock(p, &rf);
1747 
1748 	/*
1749 	 * Setting the clamp bucket is serialized by task_rq_lock().
1750 	 * If the task is not yet RUNNABLE and its task_struct is not
1751 	 * affecting a valid clamp bucket, the next time it's enqueued,
1752 	 * it will already see the updated clamp bucket value.
1753 	 */
1754 	for_each_clamp_id(clamp_id)
1755 		uclamp_rq_reinc_id(rq, p, clamp_id);
1756 
1757 	task_rq_unlock(rq, p, &rf);
1758 }
1759 
1760 #ifdef CONFIG_UCLAMP_TASK_GROUP
1761 static inline void
uclamp_update_active_tasks(struct cgroup_subsys_state * css)1762 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1763 {
1764 	struct css_task_iter it;
1765 	struct task_struct *p;
1766 
1767 	css_task_iter_start(css, 0, &it);
1768 	while ((p = css_task_iter_next(&it)))
1769 		uclamp_update_active(p);
1770 	css_task_iter_end(&it);
1771 }
1772 
1773 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1774 #endif
1775 
1776 #ifdef CONFIG_SYSCTL
1777 #ifdef CONFIG_UCLAMP_TASK
1778 #ifdef CONFIG_UCLAMP_TASK_GROUP
uclamp_update_root_tg(void)1779 static void uclamp_update_root_tg(void)
1780 {
1781 	struct task_group *tg = &root_task_group;
1782 
1783 	uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1784 		      sysctl_sched_uclamp_util_min, false);
1785 	uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1786 		      sysctl_sched_uclamp_util_max, false);
1787 
1788 	rcu_read_lock();
1789 	cpu_util_update_eff(&root_task_group.css);
1790 	rcu_read_unlock();
1791 }
1792 #else
uclamp_update_root_tg(void)1793 static void uclamp_update_root_tg(void) { }
1794 #endif
1795 
uclamp_sync_util_min_rt_default(void)1796 static void uclamp_sync_util_min_rt_default(void)
1797 {
1798 	struct task_struct *g, *p;
1799 
1800 	/*
1801 	 * copy_process()			sysctl_uclamp
1802 	 *					  uclamp_min_rt = X;
1803 	 *   write_lock(&tasklist_lock)		  read_lock(&tasklist_lock)
1804 	 *   // link thread			  smp_mb__after_spinlock()
1805 	 *   write_unlock(&tasklist_lock)	  read_unlock(&tasklist_lock);
1806 	 *   sched_post_fork()			  for_each_process_thread()
1807 	 *     __uclamp_sync_rt()		    __uclamp_sync_rt()
1808 	 *
1809 	 * Ensures that either sched_post_fork() will observe the new
1810 	 * uclamp_min_rt or for_each_process_thread() will observe the new
1811 	 * task.
1812 	 */
1813 	read_lock(&tasklist_lock);
1814 	smp_mb__after_spinlock();
1815 	read_unlock(&tasklist_lock);
1816 
1817 	rcu_read_lock();
1818 	for_each_process_thread(g, p)
1819 		uclamp_update_util_min_rt_default(p);
1820 	rcu_read_unlock();
1821 }
1822 
sysctl_sched_uclamp_handler(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)1823 static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1824 				void *buffer, size_t *lenp, loff_t *ppos)
1825 {
1826 	bool update_root_tg = false;
1827 	int old_min, old_max, old_min_rt;
1828 	int result;
1829 
1830 	guard(mutex)(&uclamp_mutex);
1831 
1832 	old_min = sysctl_sched_uclamp_util_min;
1833 	old_max = sysctl_sched_uclamp_util_max;
1834 	old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1835 
1836 	result = proc_dointvec(table, write, buffer, lenp, ppos);
1837 	if (result)
1838 		goto undo;
1839 	if (!write)
1840 		return 0;
1841 
1842 	if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1843 	    sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE	||
1844 	    sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1845 
1846 		result = -EINVAL;
1847 		goto undo;
1848 	}
1849 
1850 	if (old_min != sysctl_sched_uclamp_util_min) {
1851 		uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1852 			      sysctl_sched_uclamp_util_min, false);
1853 		update_root_tg = true;
1854 	}
1855 	if (old_max != sysctl_sched_uclamp_util_max) {
1856 		uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1857 			      sysctl_sched_uclamp_util_max, false);
1858 		update_root_tg = true;
1859 	}
1860 
1861 	if (update_root_tg) {
1862 		static_branch_enable(&sched_uclamp_used);
1863 		uclamp_update_root_tg();
1864 	}
1865 
1866 	if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1867 		static_branch_enable(&sched_uclamp_used);
1868 		uclamp_sync_util_min_rt_default();
1869 	}
1870 
1871 	/*
1872 	 * We update all RUNNABLE tasks only when task groups are in use.
1873 	 * Otherwise, keep it simple and do just a lazy update at each next
1874 	 * task enqueue time.
1875 	 */
1876 	return 0;
1877 
1878 undo:
1879 	sysctl_sched_uclamp_util_min = old_min;
1880 	sysctl_sched_uclamp_util_max = old_max;
1881 	sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1882 	return result;
1883 }
1884 #endif
1885 #endif
1886 
uclamp_validate(struct task_struct * p,const struct sched_attr * attr)1887 static int uclamp_validate(struct task_struct *p,
1888 			   const struct sched_attr *attr)
1889 {
1890 	int util_min = p->uclamp_req[UCLAMP_MIN].value;
1891 	int util_max = p->uclamp_req[UCLAMP_MAX].value;
1892 
1893 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1894 		util_min = attr->sched_util_min;
1895 
1896 		if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1897 			return -EINVAL;
1898 	}
1899 
1900 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1901 		util_max = attr->sched_util_max;
1902 
1903 		if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1904 			return -EINVAL;
1905 	}
1906 
1907 	if (util_min != -1 && util_max != -1 && util_min > util_max)
1908 		return -EINVAL;
1909 
1910 	/*
1911 	 * We have valid uclamp attributes; make sure uclamp is enabled.
1912 	 *
1913 	 * We need to do that here, because enabling static branches is a
1914 	 * blocking operation which obviously cannot be done while holding
1915 	 * scheduler locks.
1916 	 */
1917 	static_branch_enable(&sched_uclamp_used);
1918 
1919 	return 0;
1920 }
1921 
uclamp_reset(const struct sched_attr * attr,enum uclamp_id clamp_id,struct uclamp_se * uc_se)1922 static bool uclamp_reset(const struct sched_attr *attr,
1923 			 enum uclamp_id clamp_id,
1924 			 struct uclamp_se *uc_se)
1925 {
1926 	/* Reset on sched class change for a non user-defined clamp value. */
1927 	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1928 	    !uc_se->user_defined)
1929 		return true;
1930 
1931 	/* Reset on sched_util_{min,max} == -1. */
1932 	if (clamp_id == UCLAMP_MIN &&
1933 	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1934 	    attr->sched_util_min == -1) {
1935 		return true;
1936 	}
1937 
1938 	if (clamp_id == UCLAMP_MAX &&
1939 	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1940 	    attr->sched_util_max == -1) {
1941 		return true;
1942 	}
1943 
1944 	return false;
1945 }
1946 
__setscheduler_uclamp(struct task_struct * p,const struct sched_attr * attr)1947 static void __setscheduler_uclamp(struct task_struct *p,
1948 				  const struct sched_attr *attr)
1949 {
1950 	enum uclamp_id clamp_id;
1951 
1952 	for_each_clamp_id(clamp_id) {
1953 		struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1954 		unsigned int value;
1955 
1956 		if (!uclamp_reset(attr, clamp_id, uc_se))
1957 			continue;
1958 
1959 		/*
1960 		 * RT by default have a 100% boost value that could be modified
1961 		 * at runtime.
1962 		 */
1963 		if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1964 			value = sysctl_sched_uclamp_util_min_rt_default;
1965 		else
1966 			value = uclamp_none(clamp_id);
1967 
1968 		uclamp_se_set(uc_se, value, false);
1969 
1970 	}
1971 
1972 	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1973 		return;
1974 
1975 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1976 	    attr->sched_util_min != -1) {
1977 		uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1978 			      attr->sched_util_min, true);
1979 	}
1980 
1981 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1982 	    attr->sched_util_max != -1) {
1983 		uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1984 			      attr->sched_util_max, true);
1985 	}
1986 }
1987 
uclamp_fork(struct task_struct * p)1988 static void uclamp_fork(struct task_struct *p)
1989 {
1990 	enum uclamp_id clamp_id;
1991 
1992 	/*
1993 	 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1994 	 * as the task is still at its early fork stages.
1995 	 */
1996 	for_each_clamp_id(clamp_id)
1997 		p->uclamp[clamp_id].active = false;
1998 
1999 	if (likely(!p->sched_reset_on_fork))
2000 		return;
2001 
2002 	for_each_clamp_id(clamp_id) {
2003 		uclamp_se_set(&p->uclamp_req[clamp_id],
2004 			      uclamp_none(clamp_id), false);
2005 	}
2006 }
2007 
uclamp_post_fork(struct task_struct * p)2008 static void uclamp_post_fork(struct task_struct *p)
2009 {
2010 	uclamp_update_util_min_rt_default(p);
2011 }
2012 
init_uclamp_rq(struct rq * rq)2013 static void __init init_uclamp_rq(struct rq *rq)
2014 {
2015 	enum uclamp_id clamp_id;
2016 	struct uclamp_rq *uc_rq = rq->uclamp;
2017 
2018 	for_each_clamp_id(clamp_id) {
2019 		uc_rq[clamp_id] = (struct uclamp_rq) {
2020 			.value = uclamp_none(clamp_id)
2021 		};
2022 	}
2023 
2024 	rq->uclamp_flags = UCLAMP_FLAG_IDLE;
2025 }
2026 
init_uclamp(void)2027 static void __init init_uclamp(void)
2028 {
2029 	struct uclamp_se uc_max = {};
2030 	enum uclamp_id clamp_id;
2031 	int cpu;
2032 
2033 	for_each_possible_cpu(cpu)
2034 		init_uclamp_rq(cpu_rq(cpu));
2035 
2036 	for_each_clamp_id(clamp_id) {
2037 		uclamp_se_set(&init_task.uclamp_req[clamp_id],
2038 			      uclamp_none(clamp_id), false);
2039 	}
2040 
2041 	/* System defaults allow max clamp values for both indexes */
2042 	uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2043 	for_each_clamp_id(clamp_id) {
2044 		uclamp_default[clamp_id] = uc_max;
2045 #ifdef CONFIG_UCLAMP_TASK_GROUP
2046 		root_task_group.uclamp_req[clamp_id] = uc_max;
2047 		root_task_group.uclamp[clamp_id] = uc_max;
2048 #endif
2049 	}
2050 }
2051 
2052 #else /* CONFIG_UCLAMP_TASK */
uclamp_rq_inc(struct rq * rq,struct task_struct * p)2053 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
uclamp_rq_dec(struct rq * rq,struct task_struct * p)2054 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
uclamp_validate(struct task_struct * p,const struct sched_attr * attr)2055 static inline int uclamp_validate(struct task_struct *p,
2056 				  const struct sched_attr *attr)
2057 {
2058 	return -EOPNOTSUPP;
2059 }
__setscheduler_uclamp(struct task_struct * p,const struct sched_attr * attr)2060 static void __setscheduler_uclamp(struct task_struct *p,
2061 				  const struct sched_attr *attr) { }
uclamp_fork(struct task_struct * p)2062 static inline void uclamp_fork(struct task_struct *p) { }
uclamp_post_fork(struct task_struct * p)2063 static inline void uclamp_post_fork(struct task_struct *p) { }
init_uclamp(void)2064 static inline void init_uclamp(void) { }
2065 #endif /* CONFIG_UCLAMP_TASK */
2066 
sched_task_on_rq(struct task_struct * p)2067 bool sched_task_on_rq(struct task_struct *p)
2068 {
2069 	return task_on_rq_queued(p);
2070 }
2071 
get_wchan(struct task_struct * p)2072 unsigned long get_wchan(struct task_struct *p)
2073 {
2074 	unsigned long ip = 0;
2075 	unsigned int state;
2076 
2077 	if (!p || p == current)
2078 		return 0;
2079 
2080 	/* Only get wchan if task is blocked and we can keep it that way. */
2081 	raw_spin_lock_irq(&p->pi_lock);
2082 	state = READ_ONCE(p->__state);
2083 	smp_rmb(); /* see try_to_wake_up() */
2084 	if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2085 		ip = __get_wchan(p);
2086 	raw_spin_unlock_irq(&p->pi_lock);
2087 
2088 	return ip;
2089 }
2090 
enqueue_task(struct rq * rq,struct task_struct * p,int flags)2091 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2092 {
2093 	if (!(flags & ENQUEUE_NOCLOCK))
2094 		update_rq_clock(rq);
2095 
2096 	if (!(flags & ENQUEUE_RESTORE)) {
2097 		sched_info_enqueue(rq, p);
2098 		psi_enqueue(p, (flags & ENQUEUE_WAKEUP) && !(flags & ENQUEUE_MIGRATED));
2099 	}
2100 
2101 	uclamp_rq_inc(rq, p);
2102 	p->sched_class->enqueue_task(rq, p, flags);
2103 
2104 	if (sched_core_enabled(rq))
2105 		sched_core_enqueue(rq, p);
2106 }
2107 
dequeue_task(struct rq * rq,struct task_struct * p,int flags)2108 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2109 {
2110 	if (sched_core_enabled(rq))
2111 		sched_core_dequeue(rq, p, flags);
2112 
2113 	if (!(flags & DEQUEUE_NOCLOCK))
2114 		update_rq_clock(rq);
2115 
2116 	if (!(flags & DEQUEUE_SAVE)) {
2117 		sched_info_dequeue(rq, p);
2118 		psi_dequeue(p, flags & DEQUEUE_SLEEP);
2119 	}
2120 
2121 	uclamp_rq_dec(rq, p);
2122 	p->sched_class->dequeue_task(rq, p, flags);
2123 }
2124 
activate_task(struct rq * rq,struct task_struct * p,int flags)2125 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2126 {
2127 	if (task_on_rq_migrating(p))
2128 		flags |= ENQUEUE_MIGRATED;
2129 	if (flags & ENQUEUE_MIGRATED)
2130 		sched_mm_cid_migrate_to(rq, p);
2131 
2132 	enqueue_task(rq, p, flags);
2133 
2134 	p->on_rq = TASK_ON_RQ_QUEUED;
2135 }
2136 
deactivate_task(struct rq * rq,struct task_struct * p,int flags)2137 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2138 {
2139 	p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2140 
2141 	dequeue_task(rq, p, flags);
2142 }
2143 
__normal_prio(int policy,int rt_prio,int nice)2144 static inline int __normal_prio(int policy, int rt_prio, int nice)
2145 {
2146 	int prio;
2147 
2148 	if (dl_policy(policy))
2149 		prio = MAX_DL_PRIO - 1;
2150 	else if (rt_policy(policy))
2151 		prio = MAX_RT_PRIO - 1 - rt_prio;
2152 	else
2153 		prio = NICE_TO_PRIO(nice);
2154 
2155 	return prio;
2156 }
2157 
2158 /*
2159  * Calculate the expected normal priority: i.e. priority
2160  * without taking RT-inheritance into account. Might be
2161  * boosted by interactivity modifiers. Changes upon fork,
2162  * setprio syscalls, and whenever the interactivity
2163  * estimator recalculates.
2164  */
normal_prio(struct task_struct * p)2165 static inline int normal_prio(struct task_struct *p)
2166 {
2167 	return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2168 }
2169 
2170 /*
2171  * Calculate the current priority, i.e. the priority
2172  * taken into account by the scheduler. This value might
2173  * be boosted by RT tasks, or might be boosted by
2174  * interactivity modifiers. Will be RT if the task got
2175  * RT-boosted. If not then it returns p->normal_prio.
2176  */
effective_prio(struct task_struct * p)2177 static int effective_prio(struct task_struct *p)
2178 {
2179 	p->normal_prio = normal_prio(p);
2180 	/*
2181 	 * If we are RT tasks or we were boosted to RT priority,
2182 	 * keep the priority unchanged. Otherwise, update priority
2183 	 * to the normal priority:
2184 	 */
2185 	if (!rt_prio(p->prio))
2186 		return p->normal_prio;
2187 	return p->prio;
2188 }
2189 
2190 /**
2191  * task_curr - is this task currently executing on a CPU?
2192  * @p: the task in question.
2193  *
2194  * Return: 1 if the task is currently executing. 0 otherwise.
2195  */
task_curr(const struct task_struct * p)2196 inline int task_curr(const struct task_struct *p)
2197 {
2198 	return cpu_curr(task_cpu(p)) == p;
2199 }
2200 
2201 /*
2202  * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2203  * use the balance_callback list if you want balancing.
2204  *
2205  * this means any call to check_class_changed() must be followed by a call to
2206  * balance_callback().
2207  */
check_class_changed(struct rq * rq,struct task_struct * p,const struct sched_class * prev_class,int oldprio)2208 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2209 				       const struct sched_class *prev_class,
2210 				       int oldprio)
2211 {
2212 	if (prev_class != p->sched_class) {
2213 		if (prev_class->switched_from)
2214 			prev_class->switched_from(rq, p);
2215 
2216 		p->sched_class->switched_to(rq, p);
2217 	} else if (oldprio != p->prio || dl_task(p))
2218 		p->sched_class->prio_changed(rq, p, oldprio);
2219 }
2220 
check_preempt_curr(struct rq * rq,struct task_struct * p,int flags)2221 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2222 {
2223 	if (p->sched_class == rq->curr->sched_class)
2224 		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2225 	else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2226 		resched_curr(rq);
2227 
2228 	/*
2229 	 * A queue event has occurred, and we're going to schedule.  In
2230 	 * this case, we can save a useless back to back clock update.
2231 	 */
2232 	if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2233 		rq_clock_skip_update(rq);
2234 }
2235 
2236 static __always_inline
__task_state_match(struct task_struct * p,unsigned int state)2237 int __task_state_match(struct task_struct *p, unsigned int state)
2238 {
2239 	if (READ_ONCE(p->__state) & state)
2240 		return 1;
2241 
2242 #ifdef CONFIG_PREEMPT_RT
2243 	if (READ_ONCE(p->saved_state) & state)
2244 		return -1;
2245 #endif
2246 	return 0;
2247 }
2248 
2249 static __always_inline
task_state_match(struct task_struct * p,unsigned int state)2250 int task_state_match(struct task_struct *p, unsigned int state)
2251 {
2252 #ifdef CONFIG_PREEMPT_RT
2253 	int match;
2254 
2255 	/*
2256 	 * Serialize against current_save_and_set_rtlock_wait_state() and
2257 	 * current_restore_rtlock_saved_state().
2258 	 */
2259 	raw_spin_lock_irq(&p->pi_lock);
2260 	match = __task_state_match(p, state);
2261 	raw_spin_unlock_irq(&p->pi_lock);
2262 
2263 	return match;
2264 #else
2265 	return __task_state_match(p, state);
2266 #endif
2267 }
2268 
2269 /*
2270  * wait_task_inactive - wait for a thread to unschedule.
2271  *
2272  * Wait for the thread to block in any of the states set in @match_state.
2273  * If it changes, i.e. @p might have woken up, then return zero.  When we
2274  * succeed in waiting for @p to be off its CPU, we return a positive number
2275  * (its total switch count).  If a second call a short while later returns the
2276  * same number, the caller can be sure that @p has remained unscheduled the
2277  * whole time.
2278  *
2279  * The caller must ensure that the task *will* unschedule sometime soon,
2280  * else this function might spin for a *long* time. This function can't
2281  * be called with interrupts off, or it may introduce deadlock with
2282  * smp_call_function() if an IPI is sent by the same process we are
2283  * waiting to become inactive.
2284  */
wait_task_inactive(struct task_struct * p,unsigned int match_state)2285 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
2286 {
2287 	int running, queued, match;
2288 	struct rq_flags rf;
2289 	unsigned long ncsw;
2290 	struct rq *rq;
2291 
2292 	for (;;) {
2293 		/*
2294 		 * We do the initial early heuristics without holding
2295 		 * any task-queue locks at all. We'll only try to get
2296 		 * the runqueue lock when things look like they will
2297 		 * work out!
2298 		 */
2299 		rq = task_rq(p);
2300 
2301 		/*
2302 		 * If the task is actively running on another CPU
2303 		 * still, just relax and busy-wait without holding
2304 		 * any locks.
2305 		 *
2306 		 * NOTE! Since we don't hold any locks, it's not
2307 		 * even sure that "rq" stays as the right runqueue!
2308 		 * But we don't care, since "task_on_cpu()" will
2309 		 * return false if the runqueue has changed and p
2310 		 * is actually now running somewhere else!
2311 		 */
2312 		while (task_on_cpu(rq, p)) {
2313 			if (!task_state_match(p, match_state))
2314 				return 0;
2315 			cpu_relax();
2316 		}
2317 
2318 		/*
2319 		 * Ok, time to look more closely! We need the rq
2320 		 * lock now, to be *sure*. If we're wrong, we'll
2321 		 * just go back and repeat.
2322 		 */
2323 		rq = task_rq_lock(p, &rf);
2324 		trace_sched_wait_task(p);
2325 		running = task_on_cpu(rq, p);
2326 		queued = task_on_rq_queued(p);
2327 		ncsw = 0;
2328 		if ((match = __task_state_match(p, match_state))) {
2329 			/*
2330 			 * When matching on p->saved_state, consider this task
2331 			 * still queued so it will wait.
2332 			 */
2333 			if (match < 0)
2334 				queued = 1;
2335 			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2336 		}
2337 		task_rq_unlock(rq, p, &rf);
2338 
2339 		/*
2340 		 * If it changed from the expected state, bail out now.
2341 		 */
2342 		if (unlikely(!ncsw))
2343 			break;
2344 
2345 		/*
2346 		 * Was it really running after all now that we
2347 		 * checked with the proper locks actually held?
2348 		 *
2349 		 * Oops. Go back and try again..
2350 		 */
2351 		if (unlikely(running)) {
2352 			cpu_relax();
2353 			continue;
2354 		}
2355 
2356 		/*
2357 		 * It's not enough that it's not actively running,
2358 		 * it must be off the runqueue _entirely_, and not
2359 		 * preempted!
2360 		 *
2361 		 * So if it was still runnable (but just not actively
2362 		 * running right now), it's preempted, and we should
2363 		 * yield - it could be a while.
2364 		 */
2365 		if (unlikely(queued)) {
2366 			ktime_t to = NSEC_PER_SEC / HZ;
2367 
2368 			set_current_state(TASK_UNINTERRUPTIBLE);
2369 			schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
2370 			continue;
2371 		}
2372 
2373 		/*
2374 		 * Ahh, all good. It wasn't running, and it wasn't
2375 		 * runnable, which means that it will never become
2376 		 * running in the future either. We're all done!
2377 		 */
2378 		break;
2379 	}
2380 
2381 	return ncsw;
2382 }
2383 
2384 #ifdef CONFIG_SMP
2385 
2386 static void
2387 __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
2388 
2389 static int __set_cpus_allowed_ptr(struct task_struct *p,
2390 				  struct affinity_context *ctx);
2391 
migrate_disable_switch(struct rq * rq,struct task_struct * p)2392 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2393 {
2394 	struct affinity_context ac = {
2395 		.new_mask  = cpumask_of(rq->cpu),
2396 		.flags     = SCA_MIGRATE_DISABLE,
2397 	};
2398 
2399 	if (likely(!p->migration_disabled))
2400 		return;
2401 
2402 	if (p->cpus_ptr != &p->cpus_mask)
2403 		return;
2404 
2405 	/*
2406 	 * Violates locking rules! see comment in __do_set_cpus_allowed().
2407 	 */
2408 	__do_set_cpus_allowed(p, &ac);
2409 }
2410 
migrate_disable(void)2411 void migrate_disable(void)
2412 {
2413 	struct task_struct *p = current;
2414 
2415 	if (p->migration_disabled) {
2416 		p->migration_disabled++;
2417 		return;
2418 	}
2419 
2420 	preempt_disable();
2421 	this_rq()->nr_pinned++;
2422 	p->migration_disabled = 1;
2423 	preempt_enable();
2424 }
2425 EXPORT_SYMBOL_GPL(migrate_disable);
2426 
migrate_enable(void)2427 void migrate_enable(void)
2428 {
2429 	struct task_struct *p = current;
2430 	struct affinity_context ac = {
2431 		.new_mask  = &p->cpus_mask,
2432 		.flags     = SCA_MIGRATE_ENABLE,
2433 	};
2434 
2435 	if (p->migration_disabled > 1) {
2436 		p->migration_disabled--;
2437 		return;
2438 	}
2439 
2440 	if (WARN_ON_ONCE(!p->migration_disabled))
2441 		return;
2442 
2443 	/*
2444 	 * Ensure stop_task runs either before or after this, and that
2445 	 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2446 	 */
2447 	preempt_disable();
2448 	if (p->cpus_ptr != &p->cpus_mask)
2449 		__set_cpus_allowed_ptr(p, &ac);
2450 	/*
2451 	 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2452 	 * regular cpus_mask, otherwise things that race (eg.
2453 	 * select_fallback_rq) get confused.
2454 	 */
2455 	barrier();
2456 	p->migration_disabled = 0;
2457 	this_rq()->nr_pinned--;
2458 	preempt_enable();
2459 }
2460 EXPORT_SYMBOL_GPL(migrate_enable);
2461 
rq_has_pinned_tasks(struct rq * rq)2462 static inline bool rq_has_pinned_tasks(struct rq *rq)
2463 {
2464 	return rq->nr_pinned;
2465 }
2466 
2467 /*
2468  * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2469  * __set_cpus_allowed_ptr() and select_fallback_rq().
2470  */
is_cpu_allowed(struct task_struct * p,int cpu)2471 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2472 {
2473 	/* When not in the task's cpumask, no point in looking further. */
2474 	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2475 		return false;
2476 
2477 	/* migrate_disabled() must be allowed to finish. */
2478 	if (is_migration_disabled(p))
2479 		return cpu_online(cpu);
2480 
2481 	/* Non kernel threads are not allowed during either online or offline. */
2482 	if (!(p->flags & PF_KTHREAD))
2483 		return cpu_active(cpu) && task_cpu_possible(cpu, p);
2484 
2485 	/* KTHREAD_IS_PER_CPU is always allowed. */
2486 	if (kthread_is_per_cpu(p))
2487 		return cpu_online(cpu);
2488 
2489 	/* Regular kernel threads don't get to stay during offline. */
2490 	if (cpu_dying(cpu))
2491 		return false;
2492 
2493 	/* But are allowed during online. */
2494 	return cpu_online(cpu);
2495 }
2496 
2497 /*
2498  * This is how migration works:
2499  *
2500  * 1) we invoke migration_cpu_stop() on the target CPU using
2501  *    stop_one_cpu().
2502  * 2) stopper starts to run (implicitly forcing the migrated thread
2503  *    off the CPU)
2504  * 3) it checks whether the migrated task is still in the wrong runqueue.
2505  * 4) if it's in the wrong runqueue then the migration thread removes
2506  *    it and puts it into the right queue.
2507  * 5) stopper completes and stop_one_cpu() returns and the migration
2508  *    is done.
2509  */
2510 
2511 /*
2512  * move_queued_task - move a queued task to new rq.
2513  *
2514  * Returns (locked) new rq. Old rq's lock is released.
2515  */
move_queued_task(struct rq * rq,struct rq_flags * rf,struct task_struct * p,int new_cpu)2516 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2517 				   struct task_struct *p, int new_cpu)
2518 {
2519 	lockdep_assert_rq_held(rq);
2520 
2521 	deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2522 	set_task_cpu(p, new_cpu);
2523 	rq_unlock(rq, rf);
2524 
2525 	rq = cpu_rq(new_cpu);
2526 
2527 	rq_lock(rq, rf);
2528 	WARN_ON_ONCE(task_cpu(p) != new_cpu);
2529 	activate_task(rq, p, 0);
2530 	check_preempt_curr(rq, p, 0);
2531 
2532 	return rq;
2533 }
2534 
2535 struct migration_arg {
2536 	struct task_struct		*task;
2537 	int				dest_cpu;
2538 	struct set_affinity_pending	*pending;
2539 };
2540 
2541 /*
2542  * @refs: number of wait_for_completion()
2543  * @stop_pending: is @stop_work in use
2544  */
2545 struct set_affinity_pending {
2546 	refcount_t		refs;
2547 	unsigned int		stop_pending;
2548 	struct completion	done;
2549 	struct cpu_stop_work	stop_work;
2550 	struct migration_arg	arg;
2551 };
2552 
2553 /*
2554  * Move (not current) task off this CPU, onto the destination CPU. We're doing
2555  * this because either it can't run here any more (set_cpus_allowed()
2556  * away from this CPU, or CPU going down), or because we're
2557  * attempting to rebalance this task on exec (sched_exec).
2558  *
2559  * So we race with normal scheduler movements, but that's OK, as long
2560  * as the task is no longer on this CPU.
2561  */
__migrate_task(struct rq * rq,struct rq_flags * rf,struct task_struct * p,int dest_cpu)2562 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2563 				 struct task_struct *p, int dest_cpu)
2564 {
2565 	/* Affinity changed (again). */
2566 	if (!is_cpu_allowed(p, dest_cpu))
2567 		return rq;
2568 
2569 	rq = move_queued_task(rq, rf, p, dest_cpu);
2570 
2571 	return rq;
2572 }
2573 
2574 /*
2575  * migration_cpu_stop - this will be executed by a highprio stopper thread
2576  * and performs thread migration by bumping thread off CPU then
2577  * 'pushing' onto another runqueue.
2578  */
migration_cpu_stop(void * data)2579 static int migration_cpu_stop(void *data)
2580 {
2581 	struct migration_arg *arg = data;
2582 	struct set_affinity_pending *pending = arg->pending;
2583 	struct task_struct *p = arg->task;
2584 	struct rq *rq = this_rq();
2585 	bool complete = false;
2586 	struct rq_flags rf;
2587 
2588 	/*
2589 	 * The original target CPU might have gone down and we might
2590 	 * be on another CPU but it doesn't matter.
2591 	 */
2592 	local_irq_save(rf.flags);
2593 	/*
2594 	 * We need to explicitly wake pending tasks before running
2595 	 * __migrate_task() such that we will not miss enforcing cpus_ptr
2596 	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2597 	 */
2598 	flush_smp_call_function_queue();
2599 
2600 	raw_spin_lock(&p->pi_lock);
2601 	rq_lock(rq, &rf);
2602 
2603 	/*
2604 	 * If we were passed a pending, then ->stop_pending was set, thus
2605 	 * p->migration_pending must have remained stable.
2606 	 */
2607 	WARN_ON_ONCE(pending && pending != p->migration_pending);
2608 
2609 	/*
2610 	 * If task_rq(p) != rq, it cannot be migrated here, because we're
2611 	 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2612 	 * we're holding p->pi_lock.
2613 	 */
2614 	if (task_rq(p) == rq) {
2615 		if (is_migration_disabled(p))
2616 			goto out;
2617 
2618 		if (pending) {
2619 			p->migration_pending = NULL;
2620 			complete = true;
2621 
2622 			if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2623 				goto out;
2624 		}
2625 
2626 		if (task_on_rq_queued(p)) {
2627 			update_rq_clock(rq);
2628 			rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2629 		} else {
2630 			p->wake_cpu = arg->dest_cpu;
2631 		}
2632 
2633 		/*
2634 		 * XXX __migrate_task() can fail, at which point we might end
2635 		 * up running on a dodgy CPU, AFAICT this can only happen
2636 		 * during CPU hotplug, at which point we'll get pushed out
2637 		 * anyway, so it's probably not a big deal.
2638 		 */
2639 
2640 	} else if (pending) {
2641 		/*
2642 		 * This happens when we get migrated between migrate_enable()'s
2643 		 * preempt_enable() and scheduling the stopper task. At that
2644 		 * point we're a regular task again and not current anymore.
2645 		 *
2646 		 * A !PREEMPT kernel has a giant hole here, which makes it far
2647 		 * more likely.
2648 		 */
2649 
2650 		/*
2651 		 * The task moved before the stopper got to run. We're holding
2652 		 * ->pi_lock, so the allowed mask is stable - if it got
2653 		 * somewhere allowed, we're done.
2654 		 */
2655 		if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2656 			p->migration_pending = NULL;
2657 			complete = true;
2658 			goto out;
2659 		}
2660 
2661 		/*
2662 		 * When migrate_enable() hits a rq mis-match we can't reliably
2663 		 * determine is_migration_disabled() and so have to chase after
2664 		 * it.
2665 		 */
2666 		WARN_ON_ONCE(!pending->stop_pending);
2667 		preempt_disable();
2668 		task_rq_unlock(rq, p, &rf);
2669 		stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2670 				    &pending->arg, &pending->stop_work);
2671 		preempt_enable();
2672 		return 0;
2673 	}
2674 out:
2675 	if (pending)
2676 		pending->stop_pending = false;
2677 	task_rq_unlock(rq, p, &rf);
2678 
2679 	if (complete)
2680 		complete_all(&pending->done);
2681 
2682 	return 0;
2683 }
2684 
push_cpu_stop(void * arg)2685 int push_cpu_stop(void *arg)
2686 {
2687 	struct rq *lowest_rq = NULL, *rq = this_rq();
2688 	struct task_struct *p = arg;
2689 
2690 	raw_spin_lock_irq(&p->pi_lock);
2691 	raw_spin_rq_lock(rq);
2692 
2693 	if (task_rq(p) != rq)
2694 		goto out_unlock;
2695 
2696 	if (is_migration_disabled(p)) {
2697 		p->migration_flags |= MDF_PUSH;
2698 		goto out_unlock;
2699 	}
2700 
2701 	p->migration_flags &= ~MDF_PUSH;
2702 
2703 	if (p->sched_class->find_lock_rq)
2704 		lowest_rq = p->sched_class->find_lock_rq(p, rq);
2705 
2706 	if (!lowest_rq)
2707 		goto out_unlock;
2708 
2709 	// XXX validate p is still the highest prio task
2710 	if (task_rq(p) == rq) {
2711 		deactivate_task(rq, p, 0);
2712 		set_task_cpu(p, lowest_rq->cpu);
2713 		activate_task(lowest_rq, p, 0);
2714 		resched_curr(lowest_rq);
2715 	}
2716 
2717 	double_unlock_balance(rq, lowest_rq);
2718 
2719 out_unlock:
2720 	rq->push_busy = false;
2721 	raw_spin_rq_unlock(rq);
2722 	raw_spin_unlock_irq(&p->pi_lock);
2723 
2724 	put_task_struct(p);
2725 	return 0;
2726 }
2727 
2728 /*
2729  * sched_class::set_cpus_allowed must do the below, but is not required to
2730  * actually call this function.
2731  */
set_cpus_allowed_common(struct task_struct * p,struct affinity_context * ctx)2732 void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
2733 {
2734 	if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2735 		p->cpus_ptr = ctx->new_mask;
2736 		return;
2737 	}
2738 
2739 	cpumask_copy(&p->cpus_mask, ctx->new_mask);
2740 	p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
2741 
2742 	/*
2743 	 * Swap in a new user_cpus_ptr if SCA_USER flag set
2744 	 */
2745 	if (ctx->flags & SCA_USER)
2746 		swap(p->user_cpus_ptr, ctx->user_mask);
2747 }
2748 
2749 static void
__do_set_cpus_allowed(struct task_struct * p,struct affinity_context * ctx)2750 __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
2751 {
2752 	struct rq *rq = task_rq(p);
2753 	bool queued, running;
2754 
2755 	/*
2756 	 * This here violates the locking rules for affinity, since we're only
2757 	 * supposed to change these variables while holding both rq->lock and
2758 	 * p->pi_lock.
2759 	 *
2760 	 * HOWEVER, it magically works, because ttwu() is the only code that
2761 	 * accesses these variables under p->pi_lock and only does so after
2762 	 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2763 	 * before finish_task().
2764 	 *
2765 	 * XXX do further audits, this smells like something putrid.
2766 	 */
2767 	if (ctx->flags & SCA_MIGRATE_DISABLE)
2768 		SCHED_WARN_ON(!p->on_cpu);
2769 	else
2770 		lockdep_assert_held(&p->pi_lock);
2771 
2772 	queued = task_on_rq_queued(p);
2773 	running = task_current(rq, p);
2774 
2775 	if (queued) {
2776 		/*
2777 		 * Because __kthread_bind() calls this on blocked tasks without
2778 		 * holding rq->lock.
2779 		 */
2780 		lockdep_assert_rq_held(rq);
2781 		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2782 	}
2783 	if (running)
2784 		put_prev_task(rq, p);
2785 
2786 	p->sched_class->set_cpus_allowed(p, ctx);
2787 
2788 	if (queued)
2789 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2790 	if (running)
2791 		set_next_task(rq, p);
2792 }
2793 
2794 /*
2795  * Used for kthread_bind() and select_fallback_rq(), in both cases the user
2796  * affinity (if any) should be destroyed too.
2797  */
do_set_cpus_allowed(struct task_struct * p,const struct cpumask * new_mask)2798 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2799 {
2800 	struct affinity_context ac = {
2801 		.new_mask  = new_mask,
2802 		.user_mask = NULL,
2803 		.flags     = SCA_USER,	/* clear the user requested mask */
2804 	};
2805 	union cpumask_rcuhead {
2806 		cpumask_t cpumask;
2807 		struct rcu_head rcu;
2808 	};
2809 
2810 	__do_set_cpus_allowed(p, &ac);
2811 
2812 	/*
2813 	 * Because this is called with p->pi_lock held, it is not possible
2814 	 * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2815 	 * kfree_rcu().
2816 	 */
2817 	kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
2818 }
2819 
alloc_user_cpus_ptr(int node)2820 static cpumask_t *alloc_user_cpus_ptr(int node)
2821 {
2822 	/*
2823 	 * See do_set_cpus_allowed() above for the rcu_head usage.
2824 	 */
2825 	int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2826 
2827 	return kmalloc_node(size, GFP_KERNEL, node);
2828 }
2829 
dup_user_cpus_ptr(struct task_struct * dst,struct task_struct * src,int node)2830 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2831 		      int node)
2832 {
2833 	cpumask_t *user_mask;
2834 	unsigned long flags;
2835 
2836 	/*
2837 	 * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2838 	 * may differ by now due to racing.
2839 	 */
2840 	dst->user_cpus_ptr = NULL;
2841 
2842 	/*
2843 	 * This check is racy and losing the race is a valid situation.
2844 	 * It is not worth the extra overhead of taking the pi_lock on
2845 	 * every fork/clone.
2846 	 */
2847 	if (data_race(!src->user_cpus_ptr))
2848 		return 0;
2849 
2850 	user_mask = alloc_user_cpus_ptr(node);
2851 	if (!user_mask)
2852 		return -ENOMEM;
2853 
2854 	/*
2855 	 * Use pi_lock to protect content of user_cpus_ptr
2856 	 *
2857 	 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2858 	 * do_set_cpus_allowed().
2859 	 */
2860 	raw_spin_lock_irqsave(&src->pi_lock, flags);
2861 	if (src->user_cpus_ptr) {
2862 		swap(dst->user_cpus_ptr, user_mask);
2863 		cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2864 	}
2865 	raw_spin_unlock_irqrestore(&src->pi_lock, flags);
2866 
2867 	if (unlikely(user_mask))
2868 		kfree(user_mask);
2869 
2870 	return 0;
2871 }
2872 
clear_user_cpus_ptr(struct task_struct * p)2873 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2874 {
2875 	struct cpumask *user_mask = NULL;
2876 
2877 	swap(p->user_cpus_ptr, user_mask);
2878 
2879 	return user_mask;
2880 }
2881 
release_user_cpus_ptr(struct task_struct * p)2882 void release_user_cpus_ptr(struct task_struct *p)
2883 {
2884 	kfree(clear_user_cpus_ptr(p));
2885 }
2886 
2887 /*
2888  * This function is wildly self concurrent; here be dragons.
2889  *
2890  *
2891  * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2892  * designated task is enqueued on an allowed CPU. If that task is currently
2893  * running, we have to kick it out using the CPU stopper.
2894  *
2895  * Migrate-Disable comes along and tramples all over our nice sandcastle.
2896  * Consider:
2897  *
2898  *     Initial conditions: P0->cpus_mask = [0, 1]
2899  *
2900  *     P0@CPU0                  P1
2901  *
2902  *     migrate_disable();
2903  *     <preempted>
2904  *                              set_cpus_allowed_ptr(P0, [1]);
2905  *
2906  * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2907  * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2908  * This means we need the following scheme:
2909  *
2910  *     P0@CPU0                  P1
2911  *
2912  *     migrate_disable();
2913  *     <preempted>
2914  *                              set_cpus_allowed_ptr(P0, [1]);
2915  *                                <blocks>
2916  *     <resumes>
2917  *     migrate_enable();
2918  *       __set_cpus_allowed_ptr();
2919  *       <wakes local stopper>
2920  *                         `--> <woken on migration completion>
2921  *
2922  * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2923  * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2924  * task p are serialized by p->pi_lock, which we can leverage: the one that
2925  * should come into effect at the end of the Migrate-Disable region is the last
2926  * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2927  * but we still need to properly signal those waiting tasks at the appropriate
2928  * moment.
2929  *
2930  * This is implemented using struct set_affinity_pending. The first
2931  * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2932  * setup an instance of that struct and install it on the targeted task_struct.
2933  * Any and all further callers will reuse that instance. Those then wait for
2934  * a completion signaled at the tail of the CPU stopper callback (1), triggered
2935  * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2936  *
2937  *
2938  * (1) In the cases covered above. There is one more where the completion is
2939  * signaled within affine_move_task() itself: when a subsequent affinity request
2940  * occurs after the stopper bailed out due to the targeted task still being
2941  * Migrate-Disable. Consider:
2942  *
2943  *     Initial conditions: P0->cpus_mask = [0, 1]
2944  *
2945  *     CPU0		  P1				P2
2946  *     <P0>
2947  *       migrate_disable();
2948  *       <preempted>
2949  *                        set_cpus_allowed_ptr(P0, [1]);
2950  *                          <blocks>
2951  *     <migration/0>
2952  *       migration_cpu_stop()
2953  *         is_migration_disabled()
2954  *           <bails>
2955  *                                                       set_cpus_allowed_ptr(P0, [0, 1]);
2956  *                                                         <signal completion>
2957  *                          <awakes>
2958  *
2959  * Note that the above is safe vs a concurrent migrate_enable(), as any
2960  * pending affinity completion is preceded by an uninstallation of
2961  * p->migration_pending done with p->pi_lock held.
2962  */
affine_move_task(struct rq * rq,struct task_struct * p,struct rq_flags * rf,int dest_cpu,unsigned int flags)2963 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2964 			    int dest_cpu, unsigned int flags)
2965 	__releases(rq->lock)
2966 	__releases(p->pi_lock)
2967 {
2968 	struct set_affinity_pending my_pending = { }, *pending = NULL;
2969 	bool stop_pending, complete = false;
2970 
2971 	/* Can the task run on the task's current CPU? If so, we're done */
2972 	if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2973 		struct task_struct *push_task = NULL;
2974 
2975 		if ((flags & SCA_MIGRATE_ENABLE) &&
2976 		    (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2977 			rq->push_busy = true;
2978 			push_task = get_task_struct(p);
2979 		}
2980 
2981 		/*
2982 		 * If there are pending waiters, but no pending stop_work,
2983 		 * then complete now.
2984 		 */
2985 		pending = p->migration_pending;
2986 		if (pending && !pending->stop_pending) {
2987 			p->migration_pending = NULL;
2988 			complete = true;
2989 		}
2990 
2991 		preempt_disable();
2992 		task_rq_unlock(rq, p, rf);
2993 		if (push_task) {
2994 			stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2995 					    p, &rq->push_work);
2996 		}
2997 		preempt_enable();
2998 
2999 		if (complete)
3000 			complete_all(&pending->done);
3001 
3002 		return 0;
3003 	}
3004 
3005 	if (!(flags & SCA_MIGRATE_ENABLE)) {
3006 		/* serialized by p->pi_lock */
3007 		if (!p->migration_pending) {
3008 			/* Install the request */
3009 			refcount_set(&my_pending.refs, 1);
3010 			init_completion(&my_pending.done);
3011 			my_pending.arg = (struct migration_arg) {
3012 				.task = p,
3013 				.dest_cpu = dest_cpu,
3014 				.pending = &my_pending,
3015 			};
3016 
3017 			p->migration_pending = &my_pending;
3018 		} else {
3019 			pending = p->migration_pending;
3020 			refcount_inc(&pending->refs);
3021 			/*
3022 			 * Affinity has changed, but we've already installed a
3023 			 * pending. migration_cpu_stop() *must* see this, else
3024 			 * we risk a completion of the pending despite having a
3025 			 * task on a disallowed CPU.
3026 			 *
3027 			 * Serialized by p->pi_lock, so this is safe.
3028 			 */
3029 			pending->arg.dest_cpu = dest_cpu;
3030 		}
3031 	}
3032 	pending = p->migration_pending;
3033 	/*
3034 	 * - !MIGRATE_ENABLE:
3035 	 *   we'll have installed a pending if there wasn't one already.
3036 	 *
3037 	 * - MIGRATE_ENABLE:
3038 	 *   we're here because the current CPU isn't matching anymore,
3039 	 *   the only way that can happen is because of a concurrent
3040 	 *   set_cpus_allowed_ptr() call, which should then still be
3041 	 *   pending completion.
3042 	 *
3043 	 * Either way, we really should have a @pending here.
3044 	 */
3045 	if (WARN_ON_ONCE(!pending)) {
3046 		task_rq_unlock(rq, p, rf);
3047 		return -EINVAL;
3048 	}
3049 
3050 	if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
3051 		/*
3052 		 * MIGRATE_ENABLE gets here because 'p == current', but for
3053 		 * anything else we cannot do is_migration_disabled(), punt
3054 		 * and have the stopper function handle it all race-free.
3055 		 */
3056 		stop_pending = pending->stop_pending;
3057 		if (!stop_pending)
3058 			pending->stop_pending = true;
3059 
3060 		if (flags & SCA_MIGRATE_ENABLE)
3061 			p->migration_flags &= ~MDF_PUSH;
3062 
3063 		preempt_disable();
3064 		task_rq_unlock(rq, p, rf);
3065 		if (!stop_pending) {
3066 			stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
3067 					    &pending->arg, &pending->stop_work);
3068 		}
3069 		preempt_enable();
3070 
3071 		if (flags & SCA_MIGRATE_ENABLE)
3072 			return 0;
3073 	} else {
3074 
3075 		if (!is_migration_disabled(p)) {
3076 			if (task_on_rq_queued(p))
3077 				rq = move_queued_task(rq, rf, p, dest_cpu);
3078 
3079 			if (!pending->stop_pending) {
3080 				p->migration_pending = NULL;
3081 				complete = true;
3082 			}
3083 		}
3084 		task_rq_unlock(rq, p, rf);
3085 
3086 		if (complete)
3087 			complete_all(&pending->done);
3088 	}
3089 
3090 	wait_for_completion(&pending->done);
3091 
3092 	if (refcount_dec_and_test(&pending->refs))
3093 		wake_up_var(&pending->refs); /* No UaF, just an address */
3094 
3095 	/*
3096 	 * Block the original owner of &pending until all subsequent callers
3097 	 * have seen the completion and decremented the refcount
3098 	 */
3099 	wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
3100 
3101 	/* ARGH */
3102 	WARN_ON_ONCE(my_pending.stop_pending);
3103 
3104 	return 0;
3105 }
3106 
3107 /*
3108  * Called with both p->pi_lock and rq->lock held; drops both before returning.
3109  */
__set_cpus_allowed_ptr_locked(struct task_struct * p,struct affinity_context * ctx,struct rq * rq,struct rq_flags * rf)3110 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
3111 					 struct affinity_context *ctx,
3112 					 struct rq *rq,
3113 					 struct rq_flags *rf)
3114 	__releases(rq->lock)
3115 	__releases(p->pi_lock)
3116 {
3117 	const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
3118 	const struct cpumask *cpu_valid_mask = cpu_active_mask;
3119 	bool kthread = p->flags & PF_KTHREAD;
3120 	unsigned int dest_cpu;
3121 	int ret = 0;
3122 
3123 	update_rq_clock(rq);
3124 
3125 	if (kthread || is_migration_disabled(p)) {
3126 		/*
3127 		 * Kernel threads are allowed on online && !active CPUs,
3128 		 * however, during cpu-hot-unplug, even these might get pushed
3129 		 * away if not KTHREAD_IS_PER_CPU.
3130 		 *
3131 		 * Specifically, migration_disabled() tasks must not fail the
3132 		 * cpumask_any_and_distribute() pick below, esp. so on
3133 		 * SCA_MIGRATE_ENABLE, otherwise we'll not call
3134 		 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
3135 		 */
3136 		cpu_valid_mask = cpu_online_mask;
3137 	}
3138 
3139 	if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
3140 		ret = -EINVAL;
3141 		goto out;
3142 	}
3143 
3144 	/*
3145 	 * Must re-check here, to close a race against __kthread_bind(),
3146 	 * sched_setaffinity() is not guaranteed to observe the flag.
3147 	 */
3148 	if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
3149 		ret = -EINVAL;
3150 		goto out;
3151 	}
3152 
3153 	if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
3154 		if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) {
3155 			if (ctx->flags & SCA_USER)
3156 				swap(p->user_cpus_ptr, ctx->user_mask);
3157 			goto out;
3158 		}
3159 
3160 		if (WARN_ON_ONCE(p == current &&
3161 				 is_migration_disabled(p) &&
3162 				 !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
3163 			ret = -EBUSY;
3164 			goto out;
3165 		}
3166 	}
3167 
3168 	/*
3169 	 * Picking a ~random cpu helps in cases where we are changing affinity
3170 	 * for groups of tasks (ie. cpuset), so that load balancing is not
3171 	 * immediately required to distribute the tasks within their new mask.
3172 	 */
3173 	dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
3174 	if (dest_cpu >= nr_cpu_ids) {
3175 		ret = -EINVAL;
3176 		goto out;
3177 	}
3178 
3179 	__do_set_cpus_allowed(p, ctx);
3180 
3181 	return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
3182 
3183 out:
3184 	task_rq_unlock(rq, p, rf);
3185 
3186 	return ret;
3187 }
3188 
3189 /*
3190  * Change a given task's CPU affinity. Migrate the thread to a
3191  * proper CPU and schedule it away if the CPU it's executing on
3192  * is removed from the allowed bitmask.
3193  *
3194  * NOTE: the caller must have a valid reference to the task, the
3195  * task must not exit() & deallocate itself prematurely. The
3196  * call is not atomic; no spinlocks may be held.
3197  */
__set_cpus_allowed_ptr(struct task_struct * p,struct affinity_context * ctx)3198 static int __set_cpus_allowed_ptr(struct task_struct *p,
3199 				  struct affinity_context *ctx)
3200 {
3201 	struct rq_flags rf;
3202 	struct rq *rq;
3203 
3204 	rq = task_rq_lock(p, &rf);
3205 	/*
3206 	 * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3207 	 * flags are set.
3208 	 */
3209 	if (p->user_cpus_ptr &&
3210 	    !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
3211 	    cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
3212 		ctx->new_mask = rq->scratch_mask;
3213 
3214 	return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
3215 }
3216 
set_cpus_allowed_ptr(struct task_struct * p,const struct cpumask * new_mask)3217 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3218 {
3219 	struct affinity_context ac = {
3220 		.new_mask  = new_mask,
3221 		.flags     = 0,
3222 	};
3223 
3224 	return __set_cpus_allowed_ptr(p, &ac);
3225 }
3226 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
3227 
3228 /*
3229  * Change a given task's CPU affinity to the intersection of its current
3230  * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3231  * If user_cpus_ptr is defined, use it as the basis for restricting CPU
3232  * affinity or use cpu_online_mask instead.
3233  *
3234  * If the resulting mask is empty, leave the affinity unchanged and return
3235  * -EINVAL.
3236  */
restrict_cpus_allowed_ptr(struct task_struct * p,struct cpumask * new_mask,const struct cpumask * subset_mask)3237 static int restrict_cpus_allowed_ptr(struct task_struct *p,
3238 				     struct cpumask *new_mask,
3239 				     const struct cpumask *subset_mask)
3240 {
3241 	struct affinity_context ac = {
3242 		.new_mask  = new_mask,
3243 		.flags     = 0,
3244 	};
3245 	struct rq_flags rf;
3246 	struct rq *rq;
3247 	int err;
3248 
3249 	rq = task_rq_lock(p, &rf);
3250 
3251 	/*
3252 	 * Forcefully restricting the affinity of a deadline task is
3253 	 * likely to cause problems, so fail and noisily override the
3254 	 * mask entirely.
3255 	 */
3256 	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
3257 		err = -EPERM;
3258 		goto err_unlock;
3259 	}
3260 
3261 	if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
3262 		err = -EINVAL;
3263 		goto err_unlock;
3264 	}
3265 
3266 	return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
3267 
3268 err_unlock:
3269 	task_rq_unlock(rq, p, &rf);
3270 	return err;
3271 }
3272 
3273 /*
3274  * Restrict the CPU affinity of task @p so that it is a subset of
3275  * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
3276  * old affinity mask. If the resulting mask is empty, we warn and walk
3277  * up the cpuset hierarchy until we find a suitable mask.
3278  */
force_compatible_cpus_allowed_ptr(struct task_struct * p)3279 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3280 {
3281 	cpumask_var_t new_mask;
3282 	const struct cpumask *override_mask = task_cpu_possible_mask(p);
3283 
3284 	alloc_cpumask_var(&new_mask, GFP_KERNEL);
3285 
3286 	/*
3287 	 * __migrate_task() can fail silently in the face of concurrent
3288 	 * offlining of the chosen destination CPU, so take the hotplug
3289 	 * lock to ensure that the migration succeeds.
3290 	 */
3291 	cpus_read_lock();
3292 	if (!cpumask_available(new_mask))
3293 		goto out_set_mask;
3294 
3295 	if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3296 		goto out_free_mask;
3297 
3298 	/*
3299 	 * We failed to find a valid subset of the affinity mask for the
3300 	 * task, so override it based on its cpuset hierarchy.
3301 	 */
3302 	cpuset_cpus_allowed(p, new_mask);
3303 	override_mask = new_mask;
3304 
3305 out_set_mask:
3306 	if (printk_ratelimit()) {
3307 		printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3308 				task_pid_nr(p), p->comm,
3309 				cpumask_pr_args(override_mask));
3310 	}
3311 
3312 	WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3313 out_free_mask:
3314 	cpus_read_unlock();
3315 	free_cpumask_var(new_mask);
3316 }
3317 
3318 static int
3319 __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
3320 
3321 /*
3322  * Restore the affinity of a task @p which was previously restricted by a
3323  * call to force_compatible_cpus_allowed_ptr().
3324  *
3325  * It is the caller's responsibility to serialise this with any calls to
3326  * force_compatible_cpus_allowed_ptr(@p).
3327  */
relax_compatible_cpus_allowed_ptr(struct task_struct * p)3328 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3329 {
3330 	struct affinity_context ac = {
3331 		.new_mask  = task_user_cpus(p),
3332 		.flags     = 0,
3333 	};
3334 	int ret;
3335 
3336 	/*
3337 	 * Try to restore the old affinity mask with __sched_setaffinity().
3338 	 * Cpuset masking will be done there too.
3339 	 */
3340 	ret = __sched_setaffinity(p, &ac);
3341 	WARN_ON_ONCE(ret);
3342 }
3343 
set_task_cpu(struct task_struct * p,unsigned int new_cpu)3344 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3345 {
3346 #ifdef CONFIG_SCHED_DEBUG
3347 	unsigned int state = READ_ONCE(p->__state);
3348 
3349 	/*
3350 	 * We should never call set_task_cpu() on a blocked task,
3351 	 * ttwu() will sort out the placement.
3352 	 */
3353 	WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3354 
3355 	/*
3356 	 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3357 	 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3358 	 * time relying on p->on_rq.
3359 	 */
3360 	WARN_ON_ONCE(state == TASK_RUNNING &&
3361 		     p->sched_class == &fair_sched_class &&
3362 		     (p->on_rq && !task_on_rq_migrating(p)));
3363 
3364 #ifdef CONFIG_LOCKDEP
3365 	/*
3366 	 * The caller should hold either p->pi_lock or rq->lock, when changing
3367 	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3368 	 *
3369 	 * sched_move_task() holds both and thus holding either pins the cgroup,
3370 	 * see task_group().
3371 	 *
3372 	 * Furthermore, all task_rq users should acquire both locks, see
3373 	 * task_rq_lock().
3374 	 */
3375 	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3376 				      lockdep_is_held(__rq_lockp(task_rq(p)))));
3377 #endif
3378 	/*
3379 	 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3380 	 */
3381 	WARN_ON_ONCE(!cpu_online(new_cpu));
3382 
3383 	WARN_ON_ONCE(is_migration_disabled(p));
3384 #endif
3385 
3386 	trace_sched_migrate_task(p, new_cpu);
3387 
3388 	if (task_cpu(p) != new_cpu) {
3389 		if (p->sched_class->migrate_task_rq)
3390 			p->sched_class->migrate_task_rq(p, new_cpu);
3391 		p->se.nr_migrations++;
3392 		rseq_migrate(p);
3393 		sched_mm_cid_migrate_from(p);
3394 		perf_event_task_migrate(p);
3395 	}
3396 
3397 	__set_task_cpu(p, new_cpu);
3398 }
3399 
3400 #ifdef CONFIG_NUMA_BALANCING
__migrate_swap_task(struct task_struct * p,int cpu)3401 static void __migrate_swap_task(struct task_struct *p, int cpu)
3402 {
3403 	if (task_on_rq_queued(p)) {
3404 		struct rq *src_rq, *dst_rq;
3405 		struct rq_flags srf, drf;
3406 
3407 		src_rq = task_rq(p);
3408 		dst_rq = cpu_rq(cpu);
3409 
3410 		rq_pin_lock(src_rq, &srf);
3411 		rq_pin_lock(dst_rq, &drf);
3412 
3413 		deactivate_task(src_rq, p, 0);
3414 		set_task_cpu(p, cpu);
3415 		activate_task(dst_rq, p, 0);
3416 		check_preempt_curr(dst_rq, p, 0);
3417 
3418 		rq_unpin_lock(dst_rq, &drf);
3419 		rq_unpin_lock(src_rq, &srf);
3420 
3421 	} else {
3422 		/*
3423 		 * Task isn't running anymore; make it appear like we migrated
3424 		 * it before it went to sleep. This means on wakeup we make the
3425 		 * previous CPU our target instead of where it really is.
3426 		 */
3427 		p->wake_cpu = cpu;
3428 	}
3429 }
3430 
3431 struct migration_swap_arg {
3432 	struct task_struct *src_task, *dst_task;
3433 	int src_cpu, dst_cpu;
3434 };
3435 
migrate_swap_stop(void * data)3436 static int migrate_swap_stop(void *data)
3437 {
3438 	struct migration_swap_arg *arg = data;
3439 	struct rq *src_rq, *dst_rq;
3440 
3441 	if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3442 		return -EAGAIN;
3443 
3444 	src_rq = cpu_rq(arg->src_cpu);
3445 	dst_rq = cpu_rq(arg->dst_cpu);
3446 
3447 	guard(double_raw_spinlock)(&arg->src_task->pi_lock, &arg->dst_task->pi_lock);
3448 	guard(double_rq_lock)(src_rq, dst_rq);
3449 
3450 	if (task_cpu(arg->dst_task) != arg->dst_cpu)
3451 		return -EAGAIN;
3452 
3453 	if (task_cpu(arg->src_task) != arg->src_cpu)
3454 		return -EAGAIN;
3455 
3456 	if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3457 		return -EAGAIN;
3458 
3459 	if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3460 		return -EAGAIN;
3461 
3462 	__migrate_swap_task(arg->src_task, arg->dst_cpu);
3463 	__migrate_swap_task(arg->dst_task, arg->src_cpu);
3464 
3465 	return 0;
3466 }
3467 
3468 /*
3469  * Cross migrate two tasks
3470  */
migrate_swap(struct task_struct * cur,struct task_struct * p,int target_cpu,int curr_cpu)3471 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3472 		int target_cpu, int curr_cpu)
3473 {
3474 	struct migration_swap_arg arg;
3475 	int ret = -EINVAL;
3476 
3477 	arg = (struct migration_swap_arg){
3478 		.src_task = cur,
3479 		.src_cpu = curr_cpu,
3480 		.dst_task = p,
3481 		.dst_cpu = target_cpu,
3482 	};
3483 
3484 	if (arg.src_cpu == arg.dst_cpu)
3485 		goto out;
3486 
3487 	/*
3488 	 * These three tests are all lockless; this is OK since all of them
3489 	 * will be re-checked with proper locks held further down the line.
3490 	 */
3491 	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3492 		goto out;
3493 
3494 	if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3495 		goto out;
3496 
3497 	if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3498 		goto out;
3499 
3500 	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3501 	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3502 
3503 out:
3504 	return ret;
3505 }
3506 #endif /* CONFIG_NUMA_BALANCING */
3507 
3508 /***
3509  * kick_process - kick a running thread to enter/exit the kernel
3510  * @p: the to-be-kicked thread
3511  *
3512  * Cause a process which is running on another CPU to enter
3513  * kernel-mode, without any delay. (to get signals handled.)
3514  *
3515  * NOTE: this function doesn't have to take the runqueue lock,
3516  * because all it wants to ensure is that the remote task enters
3517  * the kernel. If the IPI races and the task has been migrated
3518  * to another CPU then no harm is done and the purpose has been
3519  * achieved as well.
3520  */
kick_process(struct task_struct * p)3521 void kick_process(struct task_struct *p)
3522 {
3523 	int cpu;
3524 
3525 	preempt_disable();
3526 	cpu = task_cpu(p);
3527 	if ((cpu != smp_processor_id()) && task_curr(p))
3528 		smp_send_reschedule(cpu);
3529 	preempt_enable();
3530 }
3531 EXPORT_SYMBOL_GPL(kick_process);
3532 
3533 /*
3534  * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3535  *
3536  * A few notes on cpu_active vs cpu_online:
3537  *
3538  *  - cpu_active must be a subset of cpu_online
3539  *
3540  *  - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3541  *    see __set_cpus_allowed_ptr(). At this point the newly online
3542  *    CPU isn't yet part of the sched domains, and balancing will not
3543  *    see it.
3544  *
3545  *  - on CPU-down we clear cpu_active() to mask the sched domains and
3546  *    avoid the load balancer to place new tasks on the to be removed
3547  *    CPU. Existing tasks will remain running there and will be taken
3548  *    off.
3549  *
3550  * This means that fallback selection must not select !active CPUs.
3551  * And can assume that any active CPU must be online. Conversely
3552  * select_task_rq() below may allow selection of !active CPUs in order
3553  * to satisfy the above rules.
3554  */
select_fallback_rq(int cpu,struct task_struct * p)3555 static int select_fallback_rq(int cpu, struct task_struct *p)
3556 {
3557 	int nid = cpu_to_node(cpu);
3558 	const struct cpumask *nodemask = NULL;
3559 	enum { cpuset, possible, fail } state = cpuset;
3560 	int dest_cpu;
3561 
3562 	/*
3563 	 * If the node that the CPU is on has been offlined, cpu_to_node()
3564 	 * will return -1. There is no CPU on the node, and we should
3565 	 * select the CPU on the other node.
3566 	 */
3567 	if (nid != -1) {
3568 		nodemask = cpumask_of_node(nid);
3569 
3570 		/* Look for allowed, online CPU in same node. */
3571 		for_each_cpu(dest_cpu, nodemask) {
3572 			if (is_cpu_allowed(p, dest_cpu))
3573 				return dest_cpu;
3574 		}
3575 	}
3576 
3577 	for (;;) {
3578 		/* Any allowed, online CPU? */
3579 		for_each_cpu(dest_cpu, p->cpus_ptr) {
3580 			if (!is_cpu_allowed(p, dest_cpu))
3581 				continue;
3582 
3583 			goto out;
3584 		}
3585 
3586 		/* No more Mr. Nice Guy. */
3587 		switch (state) {
3588 		case cpuset:
3589 			if (cpuset_cpus_allowed_fallback(p)) {
3590 				state = possible;
3591 				break;
3592 			}
3593 			fallthrough;
3594 		case possible:
3595 			/*
3596 			 * XXX When called from select_task_rq() we only
3597 			 * hold p->pi_lock and again violate locking order.
3598 			 *
3599 			 * More yuck to audit.
3600 			 */
3601 			do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3602 			state = fail;
3603 			break;
3604 		case fail:
3605 			BUG();
3606 			break;
3607 		}
3608 	}
3609 
3610 out:
3611 	if (state != cpuset) {
3612 		/*
3613 		 * Don't tell them about moving exiting tasks or
3614 		 * kernel threads (both mm NULL), since they never
3615 		 * leave kernel.
3616 		 */
3617 		if (p->mm && printk_ratelimit()) {
3618 			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3619 					task_pid_nr(p), p->comm, cpu);
3620 		}
3621 	}
3622 
3623 	return dest_cpu;
3624 }
3625 
3626 /*
3627  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3628  */
3629 static inline
select_task_rq(struct task_struct * p,int cpu,int wake_flags)3630 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3631 {
3632 	lockdep_assert_held(&p->pi_lock);
3633 
3634 	if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3635 		cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3636 	else
3637 		cpu = cpumask_any(p->cpus_ptr);
3638 
3639 	/*
3640 	 * In order not to call set_task_cpu() on a blocking task we need
3641 	 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3642 	 * CPU.
3643 	 *
3644 	 * Since this is common to all placement strategies, this lives here.
3645 	 *
3646 	 * [ this allows ->select_task() to simply return task_cpu(p) and
3647 	 *   not worry about this generic constraint ]
3648 	 */
3649 	if (unlikely(!is_cpu_allowed(p, cpu)))
3650 		cpu = select_fallback_rq(task_cpu(p), p);
3651 
3652 	return cpu;
3653 }
3654 
sched_set_stop_task(int cpu,struct task_struct * stop)3655 void sched_set_stop_task(int cpu, struct task_struct *stop)
3656 {
3657 	static struct lock_class_key stop_pi_lock;
3658 	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3659 	struct task_struct *old_stop = cpu_rq(cpu)->stop;
3660 
3661 	if (stop) {
3662 		/*
3663 		 * Make it appear like a SCHED_FIFO task, its something
3664 		 * userspace knows about and won't get confused about.
3665 		 *
3666 		 * Also, it will make PI more or less work without too
3667 		 * much confusion -- but then, stop work should not
3668 		 * rely on PI working anyway.
3669 		 */
3670 		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
3671 
3672 		stop->sched_class = &stop_sched_class;
3673 
3674 		/*
3675 		 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3676 		 * adjust the effective priority of a task. As a result,
3677 		 * rt_mutex_setprio() can trigger (RT) balancing operations,
3678 		 * which can then trigger wakeups of the stop thread to push
3679 		 * around the current task.
3680 		 *
3681 		 * The stop task itself will never be part of the PI-chain, it
3682 		 * never blocks, therefore that ->pi_lock recursion is safe.
3683 		 * Tell lockdep about this by placing the stop->pi_lock in its
3684 		 * own class.
3685 		 */
3686 		lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3687 	}
3688 
3689 	cpu_rq(cpu)->stop = stop;
3690 
3691 	if (old_stop) {
3692 		/*
3693 		 * Reset it back to a normal scheduling class so that
3694 		 * it can die in pieces.
3695 		 */
3696 		old_stop->sched_class = &rt_sched_class;
3697 	}
3698 }
3699 
3700 #else /* CONFIG_SMP */
3701 
__set_cpus_allowed_ptr(struct task_struct * p,struct affinity_context * ctx)3702 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3703 					 struct affinity_context *ctx)
3704 {
3705 	return set_cpus_allowed_ptr(p, ctx->new_mask);
3706 }
3707 
migrate_disable_switch(struct rq * rq,struct task_struct * p)3708 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3709 
rq_has_pinned_tasks(struct rq * rq)3710 static inline bool rq_has_pinned_tasks(struct rq *rq)
3711 {
3712 	return false;
3713 }
3714 
alloc_user_cpus_ptr(int node)3715 static inline cpumask_t *alloc_user_cpus_ptr(int node)
3716 {
3717 	return NULL;
3718 }
3719 
3720 #endif /* !CONFIG_SMP */
3721 
3722 static void
ttwu_stat(struct task_struct * p,int cpu,int wake_flags)3723 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3724 {
3725 	struct rq *rq;
3726 
3727 	if (!schedstat_enabled())
3728 		return;
3729 
3730 	rq = this_rq();
3731 
3732 #ifdef CONFIG_SMP
3733 	if (cpu == rq->cpu) {
3734 		__schedstat_inc(rq->ttwu_local);
3735 		__schedstat_inc(p->stats.nr_wakeups_local);
3736 	} else {
3737 		struct sched_domain *sd;
3738 
3739 		__schedstat_inc(p->stats.nr_wakeups_remote);
3740 
3741 		guard(rcu)();
3742 		for_each_domain(rq->cpu, sd) {
3743 			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3744 				__schedstat_inc(sd->ttwu_wake_remote);
3745 				break;
3746 			}
3747 		}
3748 	}
3749 
3750 	if (wake_flags & WF_MIGRATED)
3751 		__schedstat_inc(p->stats.nr_wakeups_migrate);
3752 #endif /* CONFIG_SMP */
3753 
3754 	__schedstat_inc(rq->ttwu_count);
3755 	__schedstat_inc(p->stats.nr_wakeups);
3756 
3757 	if (wake_flags & WF_SYNC)
3758 		__schedstat_inc(p->stats.nr_wakeups_sync);
3759 }
3760 
3761 /*
3762  * Mark the task runnable.
3763  */
ttwu_do_wakeup(struct task_struct * p)3764 static inline void ttwu_do_wakeup(struct task_struct *p)
3765 {
3766 	WRITE_ONCE(p->__state, TASK_RUNNING);
3767 	trace_sched_wakeup(p);
3768 }
3769 
3770 static void
ttwu_do_activate(struct rq * rq,struct task_struct * p,int wake_flags,struct rq_flags * rf)3771 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3772 		 struct rq_flags *rf)
3773 {
3774 	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3775 
3776 	lockdep_assert_rq_held(rq);
3777 
3778 	if (p->sched_contributes_to_load)
3779 		rq->nr_uninterruptible--;
3780 
3781 #ifdef CONFIG_SMP
3782 	if (wake_flags & WF_MIGRATED)
3783 		en_flags |= ENQUEUE_MIGRATED;
3784 	else
3785 #endif
3786 	if (p->in_iowait) {
3787 		delayacct_blkio_end(p);
3788 		atomic_dec(&task_rq(p)->nr_iowait);
3789 	}
3790 
3791 	activate_task(rq, p, en_flags);
3792 	check_preempt_curr(rq, p, wake_flags);
3793 
3794 	ttwu_do_wakeup(p);
3795 
3796 #ifdef CONFIG_SMP
3797 	if (p->sched_class->task_woken) {
3798 		/*
3799 		 * Our task @p is fully woken up and running; so it's safe to
3800 		 * drop the rq->lock, hereafter rq is only used for statistics.
3801 		 */
3802 		rq_unpin_lock(rq, rf);
3803 		p->sched_class->task_woken(rq, p);
3804 		rq_repin_lock(rq, rf);
3805 	}
3806 
3807 	if (rq->idle_stamp) {
3808 		u64 delta = rq_clock(rq) - rq->idle_stamp;
3809 		u64 max = 2*rq->max_idle_balance_cost;
3810 
3811 		update_avg(&rq->avg_idle, delta);
3812 
3813 		if (rq->avg_idle > max)
3814 			rq->avg_idle = max;
3815 
3816 		rq->wake_stamp = jiffies;
3817 		rq->wake_avg_idle = rq->avg_idle / 2;
3818 
3819 		rq->idle_stamp = 0;
3820 	}
3821 #endif
3822 }
3823 
3824 /*
3825  * Consider @p being inside a wait loop:
3826  *
3827  *   for (;;) {
3828  *      set_current_state(TASK_UNINTERRUPTIBLE);
3829  *
3830  *      if (CONDITION)
3831  *         break;
3832  *
3833  *      schedule();
3834  *   }
3835  *   __set_current_state(TASK_RUNNING);
3836  *
3837  * between set_current_state() and schedule(). In this case @p is still
3838  * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3839  * an atomic manner.
3840  *
3841  * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3842  * then schedule() must still happen and p->state can be changed to
3843  * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3844  * need to do a full wakeup with enqueue.
3845  *
3846  * Returns: %true when the wakeup is done,
3847  *          %false otherwise.
3848  */
ttwu_runnable(struct task_struct * p,int wake_flags)3849 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3850 {
3851 	struct rq_flags rf;
3852 	struct rq *rq;
3853 	int ret = 0;
3854 
3855 	rq = __task_rq_lock(p, &rf);
3856 	if (task_on_rq_queued(p)) {
3857 		if (!task_on_cpu(rq, p)) {
3858 			/*
3859 			 * When on_rq && !on_cpu the task is preempted, see if
3860 			 * it should preempt the task that is current now.
3861 			 */
3862 			update_rq_clock(rq);
3863 			check_preempt_curr(rq, p, wake_flags);
3864 		}
3865 		ttwu_do_wakeup(p);
3866 		ret = 1;
3867 	}
3868 	__task_rq_unlock(rq, &rf);
3869 
3870 	return ret;
3871 }
3872 
3873 #ifdef CONFIG_SMP
sched_ttwu_pending(void * arg)3874 void sched_ttwu_pending(void *arg)
3875 {
3876 	struct llist_node *llist = arg;
3877 	struct rq *rq = this_rq();
3878 	struct task_struct *p, *t;
3879 	struct rq_flags rf;
3880 
3881 	if (!llist)
3882 		return;
3883 
3884 	rq_lock_irqsave(rq, &rf);
3885 	update_rq_clock(rq);
3886 
3887 	llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3888 		if (WARN_ON_ONCE(p->on_cpu))
3889 			smp_cond_load_acquire(&p->on_cpu, !VAL);
3890 
3891 		if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3892 			set_task_cpu(p, cpu_of(rq));
3893 
3894 		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3895 	}
3896 
3897 	/*
3898 	 * Must be after enqueueing at least once task such that
3899 	 * idle_cpu() does not observe a false-negative -- if it does,
3900 	 * it is possible for select_idle_siblings() to stack a number
3901 	 * of tasks on this CPU during that window.
3902 	 *
3903 	 * It is ok to clear ttwu_pending when another task pending.
3904 	 * We will receive IPI after local irq enabled and then enqueue it.
3905 	 * Since now nr_running > 0, idle_cpu() will always get correct result.
3906 	 */
3907 	WRITE_ONCE(rq->ttwu_pending, 0);
3908 	rq_unlock_irqrestore(rq, &rf);
3909 }
3910 
3911 /*
3912  * Prepare the scene for sending an IPI for a remote smp_call
3913  *
3914  * Returns true if the caller can proceed with sending the IPI.
3915  * Returns false otherwise.
3916  */
call_function_single_prep_ipi(int cpu)3917 bool call_function_single_prep_ipi(int cpu)
3918 {
3919 	if (set_nr_if_polling(cpu_rq(cpu)->idle)) {
3920 		trace_sched_wake_idle_without_ipi(cpu);
3921 		return false;
3922 	}
3923 
3924 	return true;
3925 }
3926 
3927 /*
3928  * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3929  * necessary. The wakee CPU on receipt of the IPI will queue the task
3930  * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3931  * of the wakeup instead of the waker.
3932  */
__ttwu_queue_wakelist(struct task_struct * p,int cpu,int wake_flags)3933 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3934 {
3935 	struct rq *rq = cpu_rq(cpu);
3936 
3937 	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3938 
3939 	WRITE_ONCE(rq->ttwu_pending, 1);
3940 	__smp_call_single_queue(cpu, &p->wake_entry.llist);
3941 }
3942 
wake_up_if_idle(int cpu)3943 void wake_up_if_idle(int cpu)
3944 {
3945 	struct rq *rq = cpu_rq(cpu);
3946 
3947 	guard(rcu)();
3948 	if (is_idle_task(rcu_dereference(rq->curr))) {
3949 		guard(rq_lock_irqsave)(rq);
3950 		if (is_idle_task(rq->curr))
3951 			resched_curr(rq);
3952 	}
3953 }
3954 
cpus_share_cache(int this_cpu,int that_cpu)3955 bool cpus_share_cache(int this_cpu, int that_cpu)
3956 {
3957 	if (this_cpu == that_cpu)
3958 		return true;
3959 
3960 	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3961 }
3962 
ttwu_queue_cond(struct task_struct * p,int cpu)3963 static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3964 {
3965 	/*
3966 	 * Do not complicate things with the async wake_list while the CPU is
3967 	 * in hotplug state.
3968 	 */
3969 	if (!cpu_active(cpu))
3970 		return false;
3971 
3972 	/* Ensure the task will still be allowed to run on the CPU. */
3973 	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3974 		return false;
3975 
3976 	/*
3977 	 * If the CPU does not share cache, then queue the task on the
3978 	 * remote rqs wakelist to avoid accessing remote data.
3979 	 */
3980 	if (!cpus_share_cache(smp_processor_id(), cpu))
3981 		return true;
3982 
3983 	if (cpu == smp_processor_id())
3984 		return false;
3985 
3986 	/*
3987 	 * If the wakee cpu is idle, or the task is descheduling and the
3988 	 * only running task on the CPU, then use the wakelist to offload
3989 	 * the task activation to the idle (or soon-to-be-idle) CPU as
3990 	 * the current CPU is likely busy. nr_running is checked to
3991 	 * avoid unnecessary task stacking.
3992 	 *
3993 	 * Note that we can only get here with (wakee) p->on_rq=0,
3994 	 * p->on_cpu can be whatever, we've done the dequeue, so
3995 	 * the wakee has been accounted out of ->nr_running.
3996 	 */
3997 	if (!cpu_rq(cpu)->nr_running)
3998 		return true;
3999 
4000 	return false;
4001 }
4002 
ttwu_queue_wakelist(struct task_struct * p,int cpu,int wake_flags)4003 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4004 {
4005 	if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
4006 		sched_clock_cpu(cpu); /* Sync clocks across CPUs */
4007 		__ttwu_queue_wakelist(p, cpu, wake_flags);
4008 		return true;
4009 	}
4010 
4011 	return false;
4012 }
4013 
4014 #else /* !CONFIG_SMP */
4015 
ttwu_queue_wakelist(struct task_struct * p,int cpu,int wake_flags)4016 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4017 {
4018 	return false;
4019 }
4020 
4021 #endif /* CONFIG_SMP */
4022 
ttwu_queue(struct task_struct * p,int cpu,int wake_flags)4023 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
4024 {
4025 	struct rq *rq = cpu_rq(cpu);
4026 	struct rq_flags rf;
4027 
4028 	if (ttwu_queue_wakelist(p, cpu, wake_flags))
4029 		return;
4030 
4031 	rq_lock(rq, &rf);
4032 	update_rq_clock(rq);
4033 	ttwu_do_activate(rq, p, wake_flags, &rf);
4034 	rq_unlock(rq, &rf);
4035 }
4036 
4037 /*
4038  * Invoked from try_to_wake_up() to check whether the task can be woken up.
4039  *
4040  * The caller holds p::pi_lock if p != current or has preemption
4041  * disabled when p == current.
4042  *
4043  * The rules of PREEMPT_RT saved_state:
4044  *
4045  *   The related locking code always holds p::pi_lock when updating
4046  *   p::saved_state, which means the code is fully serialized in both cases.
4047  *
4048  *   The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
4049  *   bits set. This allows to distinguish all wakeup scenarios.
4050  */
4051 static __always_inline
ttwu_state_match(struct task_struct * p,unsigned int state,int * success)4052 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
4053 {
4054 	int match;
4055 
4056 	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
4057 		WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
4058 			     state != TASK_RTLOCK_WAIT);
4059 	}
4060 
4061 	*success = !!(match = __task_state_match(p, state));
4062 
4063 #ifdef CONFIG_PREEMPT_RT
4064 	/*
4065 	 * Saved state preserves the task state across blocking on
4066 	 * an RT lock.  If the state matches, set p::saved_state to
4067 	 * TASK_RUNNING, but do not wake the task because it waits
4068 	 * for a lock wakeup. Also indicate success because from
4069 	 * the regular waker's point of view this has succeeded.
4070 	 *
4071 	 * After acquiring the lock the task will restore p::__state
4072 	 * from p::saved_state which ensures that the regular
4073 	 * wakeup is not lost. The restore will also set
4074 	 * p::saved_state to TASK_RUNNING so any further tests will
4075 	 * not result in false positives vs. @success
4076 	 */
4077 	if (match < 0)
4078 		p->saved_state = TASK_RUNNING;
4079 #endif
4080 	return match > 0;
4081 }
4082 
4083 /*
4084  * Notes on Program-Order guarantees on SMP systems.
4085  *
4086  *  MIGRATION
4087  *
4088  * The basic program-order guarantee on SMP systems is that when a task [t]
4089  * migrates, all its activity on its old CPU [c0] happens-before any subsequent
4090  * execution on its new CPU [c1].
4091  *
4092  * For migration (of runnable tasks) this is provided by the following means:
4093  *
4094  *  A) UNLOCK of the rq(c0)->lock scheduling out task t
4095  *  B) migration for t is required to synchronize *both* rq(c0)->lock and
4096  *     rq(c1)->lock (if not at the same time, then in that order).
4097  *  C) LOCK of the rq(c1)->lock scheduling in task
4098  *
4099  * Release/acquire chaining guarantees that B happens after A and C after B.
4100  * Note: the CPU doing B need not be c0 or c1
4101  *
4102  * Example:
4103  *
4104  *   CPU0            CPU1            CPU2
4105  *
4106  *   LOCK rq(0)->lock
4107  *   sched-out X
4108  *   sched-in Y
4109  *   UNLOCK rq(0)->lock
4110  *
4111  *                                   LOCK rq(0)->lock // orders against CPU0
4112  *                                   dequeue X
4113  *                                   UNLOCK rq(0)->lock
4114  *
4115  *                                   LOCK rq(1)->lock
4116  *                                   enqueue X
4117  *                                   UNLOCK rq(1)->lock
4118  *
4119  *                   LOCK rq(1)->lock // orders against CPU2
4120  *                   sched-out Z
4121  *                   sched-in X
4122  *                   UNLOCK rq(1)->lock
4123  *
4124  *
4125  *  BLOCKING -- aka. SLEEP + WAKEUP
4126  *
4127  * For blocking we (obviously) need to provide the same guarantee as for
4128  * migration. However the means are completely different as there is no lock
4129  * chain to provide order. Instead we do:
4130  *
4131  *   1) smp_store_release(X->on_cpu, 0)   -- finish_task()
4132  *   2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
4133  *
4134  * Example:
4135  *
4136  *   CPU0 (schedule)  CPU1 (try_to_wake_up) CPU2 (schedule)
4137  *
4138  *   LOCK rq(0)->lock LOCK X->pi_lock
4139  *   dequeue X
4140  *   sched-out X
4141  *   smp_store_release(X->on_cpu, 0);
4142  *
4143  *                    smp_cond_load_acquire(&X->on_cpu, !VAL);
4144  *                    X->state = WAKING
4145  *                    set_task_cpu(X,2)
4146  *
4147  *                    LOCK rq(2)->lock
4148  *                    enqueue X
4149  *                    X->state = RUNNING
4150  *                    UNLOCK rq(2)->lock
4151  *
4152  *                                          LOCK rq(2)->lock // orders against CPU1
4153  *                                          sched-out Z
4154  *                                          sched-in X
4155  *                                          UNLOCK rq(2)->lock
4156  *
4157  *                    UNLOCK X->pi_lock
4158  *   UNLOCK rq(0)->lock
4159  *
4160  *
4161  * However, for wakeups there is a second guarantee we must provide, namely we
4162  * must ensure that CONDITION=1 done by the caller can not be reordered with
4163  * accesses to the task state; see try_to_wake_up() and set_current_state().
4164  */
4165 
4166 /**
4167  * try_to_wake_up - wake up a thread
4168  * @p: the thread to be awakened
4169  * @state: the mask of task states that can be woken
4170  * @wake_flags: wake modifier flags (WF_*)
4171  *
4172  * Conceptually does:
4173  *
4174  *   If (@state & @p->state) @p->state = TASK_RUNNING.
4175  *
4176  * If the task was not queued/runnable, also place it back on a runqueue.
4177  *
4178  * This function is atomic against schedule() which would dequeue the task.
4179  *
4180  * It issues a full memory barrier before accessing @p->state, see the comment
4181  * with set_current_state().
4182  *
4183  * Uses p->pi_lock to serialize against concurrent wake-ups.
4184  *
4185  * Relies on p->pi_lock stabilizing:
4186  *  - p->sched_class
4187  *  - p->cpus_ptr
4188  *  - p->sched_task_group
4189  * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4190  *
4191  * Tries really hard to only take one task_rq(p)->lock for performance.
4192  * Takes rq->lock in:
4193  *  - ttwu_runnable()    -- old rq, unavoidable, see comment there;
4194  *  - ttwu_queue()       -- new rq, for enqueue of the task;
4195  *  - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4196  *
4197  * As a consequence we race really badly with just about everything. See the
4198  * many memory barriers and their comments for details.
4199  *
4200  * Return: %true if @p->state changes (an actual wakeup was done),
4201  *	   %false otherwise.
4202  */
try_to_wake_up(struct task_struct * p,unsigned int state,int wake_flags)4203 int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4204 {
4205 	guard(preempt)();
4206 	int cpu, success = 0;
4207 
4208 	if (p == current) {
4209 		/*
4210 		 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4211 		 * == smp_processor_id()'. Together this means we can special
4212 		 * case the whole 'p->on_rq && ttwu_runnable()' case below
4213 		 * without taking any locks.
4214 		 *
4215 		 * In particular:
4216 		 *  - we rely on Program-Order guarantees for all the ordering,
4217 		 *  - we're serialized against set_special_state() by virtue of
4218 		 *    it disabling IRQs (this allows not taking ->pi_lock).
4219 		 */
4220 		if (!ttwu_state_match(p, state, &success))
4221 			goto out;
4222 
4223 		trace_sched_waking(p);
4224 		ttwu_do_wakeup(p);
4225 		goto out;
4226 	}
4227 
4228 	/*
4229 	 * If we are going to wake up a thread waiting for CONDITION we
4230 	 * need to ensure that CONDITION=1 done by the caller can not be
4231 	 * reordered with p->state check below. This pairs with smp_store_mb()
4232 	 * in set_current_state() that the waiting thread does.
4233 	 */
4234 	scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
4235 		smp_mb__after_spinlock();
4236 		if (!ttwu_state_match(p, state, &success))
4237 			break;
4238 
4239 		trace_sched_waking(p);
4240 
4241 		/*
4242 		 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4243 		 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4244 		 * in smp_cond_load_acquire() below.
4245 		 *
4246 		 * sched_ttwu_pending()			try_to_wake_up()
4247 		 *   STORE p->on_rq = 1			  LOAD p->state
4248 		 *   UNLOCK rq->lock
4249 		 *
4250 		 * __schedule() (switch to task 'p')
4251 		 *   LOCK rq->lock			  smp_rmb();
4252 		 *   smp_mb__after_spinlock();
4253 		 *   UNLOCK rq->lock
4254 		 *
4255 		 * [task p]
4256 		 *   STORE p->state = UNINTERRUPTIBLE	  LOAD p->on_rq
4257 		 *
4258 		 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4259 		 * __schedule().  See the comment for smp_mb__after_spinlock().
4260 		 *
4261 		 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4262 		 */
4263 		smp_rmb();
4264 		if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4265 			break;
4266 
4267 #ifdef CONFIG_SMP
4268 		/*
4269 		 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4270 		 * possible to, falsely, observe p->on_cpu == 0.
4271 		 *
4272 		 * One must be running (->on_cpu == 1) in order to remove oneself
4273 		 * from the runqueue.
4274 		 *
4275 		 * __schedule() (switch to task 'p')	try_to_wake_up()
4276 		 *   STORE p->on_cpu = 1		  LOAD p->on_rq
4277 		 *   UNLOCK rq->lock
4278 		 *
4279 		 * __schedule() (put 'p' to sleep)
4280 		 *   LOCK rq->lock			  smp_rmb();
4281 		 *   smp_mb__after_spinlock();
4282 		 *   STORE p->on_rq = 0			  LOAD p->on_cpu
4283 		 *
4284 		 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4285 		 * __schedule().  See the comment for smp_mb__after_spinlock().
4286 		 *
4287 		 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4288 		 * schedule()'s deactivate_task() has 'happened' and p will no longer
4289 		 * care about it's own p->state. See the comment in __schedule().
4290 		 */
4291 		smp_acquire__after_ctrl_dep();
4292 
4293 		/*
4294 		 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4295 		 * == 0), which means we need to do an enqueue, change p->state to
4296 		 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4297 		 * enqueue, such as ttwu_queue_wakelist().
4298 		 */
4299 		WRITE_ONCE(p->__state, TASK_WAKING);
4300 
4301 		/*
4302 		 * If the owning (remote) CPU is still in the middle of schedule() with
4303 		 * this task as prev, considering queueing p on the remote CPUs wake_list
4304 		 * which potentially sends an IPI instead of spinning on p->on_cpu to
4305 		 * let the waker make forward progress. This is safe because IRQs are
4306 		 * disabled and the IPI will deliver after on_cpu is cleared.
4307 		 *
4308 		 * Ensure we load task_cpu(p) after p->on_cpu:
4309 		 *
4310 		 * set_task_cpu(p, cpu);
4311 		 *   STORE p->cpu = @cpu
4312 		 * __schedule() (switch to task 'p')
4313 		 *   LOCK rq->lock
4314 		 *   smp_mb__after_spin_lock()		smp_cond_load_acquire(&p->on_cpu)
4315 		 *   STORE p->on_cpu = 1		LOAD p->cpu
4316 		 *
4317 		 * to ensure we observe the correct CPU on which the task is currently
4318 		 * scheduling.
4319 		 */
4320 		if (smp_load_acquire(&p->on_cpu) &&
4321 		    ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4322 			break;
4323 
4324 		/*
4325 		 * If the owning (remote) CPU is still in the middle of schedule() with
4326 		 * this task as prev, wait until it's done referencing the task.
4327 		 *
4328 		 * Pairs with the smp_store_release() in finish_task().
4329 		 *
4330 		 * This ensures that tasks getting woken will be fully ordered against
4331 		 * their previous state and preserve Program Order.
4332 		 */
4333 		smp_cond_load_acquire(&p->on_cpu, !VAL);
4334 
4335 		cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4336 		if (task_cpu(p) != cpu) {
4337 			if (p->in_iowait) {
4338 				delayacct_blkio_end(p);
4339 				atomic_dec(&task_rq(p)->nr_iowait);
4340 			}
4341 
4342 			wake_flags |= WF_MIGRATED;
4343 			psi_ttwu_dequeue(p);
4344 			set_task_cpu(p, cpu);
4345 		}
4346 #else
4347 		cpu = task_cpu(p);
4348 #endif /* CONFIG_SMP */
4349 
4350 		ttwu_queue(p, cpu, wake_flags);
4351 	}
4352 out:
4353 	if (success)
4354 		ttwu_stat(p, task_cpu(p), wake_flags);
4355 
4356 	return success;
4357 }
4358 
__task_needs_rq_lock(struct task_struct * p)4359 static bool __task_needs_rq_lock(struct task_struct *p)
4360 {
4361 	unsigned int state = READ_ONCE(p->__state);
4362 
4363 	/*
4364 	 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4365 	 * the task is blocked. Make sure to check @state since ttwu() can drop
4366 	 * locks at the end, see ttwu_queue_wakelist().
4367 	 */
4368 	if (state == TASK_RUNNING || state == TASK_WAKING)
4369 		return true;
4370 
4371 	/*
4372 	 * Ensure we load p->on_rq after p->__state, otherwise it would be
4373 	 * possible to, falsely, observe p->on_rq == 0.
4374 	 *
4375 	 * See try_to_wake_up() for a longer comment.
4376 	 */
4377 	smp_rmb();
4378 	if (p->on_rq)
4379 		return true;
4380 
4381 #ifdef CONFIG_SMP
4382 	/*
4383 	 * Ensure the task has finished __schedule() and will not be referenced
4384 	 * anymore. Again, see try_to_wake_up() for a longer comment.
4385 	 */
4386 	smp_rmb();
4387 	smp_cond_load_acquire(&p->on_cpu, !VAL);
4388 #endif
4389 
4390 	return false;
4391 }
4392 
4393 /**
4394  * task_call_func - Invoke a function on task in fixed state
4395  * @p: Process for which the function is to be invoked, can be @current.
4396  * @func: Function to invoke.
4397  * @arg: Argument to function.
4398  *
4399  * Fix the task in it's current state by avoiding wakeups and or rq operations
4400  * and call @func(@arg) on it.  This function can use ->on_rq and task_curr()
4401  * to work out what the state is, if required.  Given that @func can be invoked
4402  * with a runqueue lock held, it had better be quite lightweight.
4403  *
4404  * Returns:
4405  *   Whatever @func returns
4406  */
task_call_func(struct task_struct * p,task_call_f func,void * arg)4407 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4408 {
4409 	struct rq *rq = NULL;
4410 	struct rq_flags rf;
4411 	int ret;
4412 
4413 	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4414 
4415 	if (__task_needs_rq_lock(p))
4416 		rq = __task_rq_lock(p, &rf);
4417 
4418 	/*
4419 	 * At this point the task is pinned; either:
4420 	 *  - blocked and we're holding off wakeups	 (pi->lock)
4421 	 *  - woken, and we're holding off enqueue	 (rq->lock)
4422 	 *  - queued, and we're holding off schedule	 (rq->lock)
4423 	 *  - running, and we're holding off de-schedule (rq->lock)
4424 	 *
4425 	 * The called function (@func) can use: task_curr(), p->on_rq and
4426 	 * p->__state to differentiate between these states.
4427 	 */
4428 	ret = func(p, arg);
4429 
4430 	if (rq)
4431 		rq_unlock(rq, &rf);
4432 
4433 	raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4434 	return ret;
4435 }
4436 
4437 /**
4438  * cpu_curr_snapshot - Return a snapshot of the currently running task
4439  * @cpu: The CPU on which to snapshot the task.
4440  *
4441  * Returns the task_struct pointer of the task "currently" running on
4442  * the specified CPU.  If the same task is running on that CPU throughout,
4443  * the return value will be a pointer to that task's task_struct structure.
4444  * If the CPU did any context switches even vaguely concurrently with the
4445  * execution of this function, the return value will be a pointer to the
4446  * task_struct structure of a randomly chosen task that was running on
4447  * that CPU somewhere around the time that this function was executing.
4448  *
4449  * If the specified CPU was offline, the return value is whatever it
4450  * is, perhaps a pointer to the task_struct structure of that CPU's idle
4451  * task, but there is no guarantee.  Callers wishing a useful return
4452  * value must take some action to ensure that the specified CPU remains
4453  * online throughout.
4454  *
4455  * This function executes full memory barriers before and after fetching
4456  * the pointer, which permits the caller to confine this function's fetch
4457  * with respect to the caller's accesses to other shared variables.
4458  */
cpu_curr_snapshot(int cpu)4459 struct task_struct *cpu_curr_snapshot(int cpu)
4460 {
4461 	struct task_struct *t;
4462 
4463 	smp_mb(); /* Pairing determined by caller's synchronization design. */
4464 	t = rcu_dereference(cpu_curr(cpu));
4465 	smp_mb(); /* Pairing determined by caller's synchronization design. */
4466 	return t;
4467 }
4468 
4469 /**
4470  * wake_up_process - Wake up a specific process
4471  * @p: The process to be woken up.
4472  *
4473  * Attempt to wake up the nominated process and move it to the set of runnable
4474  * processes.
4475  *
4476  * Return: 1 if the process was woken up, 0 if it was already running.
4477  *
4478  * This function executes a full memory barrier before accessing the task state.
4479  */
wake_up_process(struct task_struct * p)4480 int wake_up_process(struct task_struct *p)
4481 {
4482 	return try_to_wake_up(p, TASK_NORMAL, 0);
4483 }
4484 EXPORT_SYMBOL(wake_up_process);
4485 
wake_up_state(struct task_struct * p,unsigned int state)4486 int wake_up_state(struct task_struct *p, unsigned int state)
4487 {
4488 	return try_to_wake_up(p, state, 0);
4489 }
4490 
4491 /*
4492  * Perform scheduler related setup for a newly forked process p.
4493  * p is forked by current.
4494  *
4495  * __sched_fork() is basic setup used by init_idle() too:
4496  */
__sched_fork(unsigned long clone_flags,struct task_struct * p)4497 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4498 {
4499 	p->on_rq			= 0;
4500 
4501 	p->se.on_rq			= 0;
4502 	p->se.exec_start		= 0;
4503 	p->se.sum_exec_runtime		= 0;
4504 	p->se.prev_sum_exec_runtime	= 0;
4505 	p->se.nr_migrations		= 0;
4506 	p->se.vruntime			= 0;
4507 	p->se.vlag			= 0;
4508 	p->se.slice			= sysctl_sched_base_slice;
4509 	INIT_LIST_HEAD(&p->se.group_node);
4510 
4511 #ifdef CONFIG_FAIR_GROUP_SCHED
4512 	p->se.cfs_rq			= NULL;
4513 #endif
4514 
4515 #ifdef CONFIG_SCHEDSTATS
4516 	/* Even if schedstat is disabled, there should not be garbage */
4517 	memset(&p->stats, 0, sizeof(p->stats));
4518 #endif
4519 
4520 	RB_CLEAR_NODE(&p->dl.rb_node);
4521 	init_dl_task_timer(&p->dl);
4522 	init_dl_inactive_task_timer(&p->dl);
4523 	__dl_clear_params(p);
4524 
4525 	INIT_LIST_HEAD(&p->rt.run_list);
4526 	p->rt.timeout		= 0;
4527 	p->rt.time_slice	= sched_rr_timeslice;
4528 	p->rt.on_rq		= 0;
4529 	p->rt.on_list		= 0;
4530 
4531 #ifdef CONFIG_PREEMPT_NOTIFIERS
4532 	INIT_HLIST_HEAD(&p->preempt_notifiers);
4533 #endif
4534 
4535 #ifdef CONFIG_COMPACTION
4536 	p->capture_control = NULL;
4537 #endif
4538 	init_numa_balancing(clone_flags, p);
4539 #ifdef CONFIG_SMP
4540 	p->wake_entry.u_flags = CSD_TYPE_TTWU;
4541 	p->migration_pending = NULL;
4542 #endif
4543 	init_sched_mm_cid(p);
4544 }
4545 
4546 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4547 
4548 #ifdef CONFIG_NUMA_BALANCING
4549 
4550 int sysctl_numa_balancing_mode;
4551 
__set_numabalancing_state(bool enabled)4552 static void __set_numabalancing_state(bool enabled)
4553 {
4554 	if (enabled)
4555 		static_branch_enable(&sched_numa_balancing);
4556 	else
4557 		static_branch_disable(&sched_numa_balancing);
4558 }
4559 
set_numabalancing_state(bool enabled)4560 void set_numabalancing_state(bool enabled)
4561 {
4562 	if (enabled)
4563 		sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4564 	else
4565 		sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4566 	__set_numabalancing_state(enabled);
4567 }
4568 
4569 #ifdef CONFIG_PROC_SYSCTL
reset_memory_tiering(void)4570 static void reset_memory_tiering(void)
4571 {
4572 	struct pglist_data *pgdat;
4573 
4574 	for_each_online_pgdat(pgdat) {
4575 		pgdat->nbp_threshold = 0;
4576 		pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
4577 		pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
4578 	}
4579 }
4580 
sysctl_numa_balancing(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)4581 static int sysctl_numa_balancing(struct ctl_table *table, int write,
4582 			  void *buffer, size_t *lenp, loff_t *ppos)
4583 {
4584 	struct ctl_table t;
4585 	int err;
4586 	int state = sysctl_numa_balancing_mode;
4587 
4588 	if (write && !capable(CAP_SYS_ADMIN))
4589 		return -EPERM;
4590 
4591 	t = *table;
4592 	t.data = &state;
4593 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4594 	if (err < 0)
4595 		return err;
4596 	if (write) {
4597 		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4598 		    (state & NUMA_BALANCING_MEMORY_TIERING))
4599 			reset_memory_tiering();
4600 		sysctl_numa_balancing_mode = state;
4601 		__set_numabalancing_state(state);
4602 	}
4603 	return err;
4604 }
4605 #endif
4606 #endif
4607 
4608 #ifdef CONFIG_SCHEDSTATS
4609 
4610 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4611 
set_schedstats(bool enabled)4612 static void set_schedstats(bool enabled)
4613 {
4614 	if (enabled)
4615 		static_branch_enable(&sched_schedstats);
4616 	else
4617 		static_branch_disable(&sched_schedstats);
4618 }
4619 
force_schedstat_enabled(void)4620 void force_schedstat_enabled(void)
4621 {
4622 	if (!schedstat_enabled()) {
4623 		pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4624 		static_branch_enable(&sched_schedstats);
4625 	}
4626 }
4627 
setup_schedstats(char * str)4628 static int __init setup_schedstats(char *str)
4629 {
4630 	int ret = 0;
4631 	if (!str)
4632 		goto out;
4633 
4634 	if (!strcmp(str, "enable")) {
4635 		set_schedstats(true);
4636 		ret = 1;
4637 	} else if (!strcmp(str, "disable")) {
4638 		set_schedstats(false);
4639 		ret = 1;
4640 	}
4641 out:
4642 	if (!ret)
4643 		pr_warn("Unable to parse schedstats=\n");
4644 
4645 	return ret;
4646 }
4647 __setup("schedstats=", setup_schedstats);
4648 
4649 #ifdef CONFIG_PROC_SYSCTL
sysctl_schedstats(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)4650 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4651 		size_t *lenp, loff_t *ppos)
4652 {
4653 	struct ctl_table t;
4654 	int err;
4655 	int state = static_branch_likely(&sched_schedstats);
4656 
4657 	if (write && !capable(CAP_SYS_ADMIN))
4658 		return -EPERM;
4659 
4660 	t = *table;
4661 	t.data = &state;
4662 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4663 	if (err < 0)
4664 		return err;
4665 	if (write)
4666 		set_schedstats(state);
4667 	return err;
4668 }
4669 #endif /* CONFIG_PROC_SYSCTL */
4670 #endif /* CONFIG_SCHEDSTATS */
4671 
4672 #ifdef CONFIG_SYSCTL
4673 static struct ctl_table sched_core_sysctls[] = {
4674 #ifdef CONFIG_SCHEDSTATS
4675 	{
4676 		.procname       = "sched_schedstats",
4677 		.data           = NULL,
4678 		.maxlen         = sizeof(unsigned int),
4679 		.mode           = 0644,
4680 		.proc_handler   = sysctl_schedstats,
4681 		.extra1         = SYSCTL_ZERO,
4682 		.extra2         = SYSCTL_ONE,
4683 	},
4684 #endif /* CONFIG_SCHEDSTATS */
4685 #ifdef CONFIG_UCLAMP_TASK
4686 	{
4687 		.procname       = "sched_util_clamp_min",
4688 		.data           = &sysctl_sched_uclamp_util_min,
4689 		.maxlen         = sizeof(unsigned int),
4690 		.mode           = 0644,
4691 		.proc_handler   = sysctl_sched_uclamp_handler,
4692 	},
4693 	{
4694 		.procname       = "sched_util_clamp_max",
4695 		.data           = &sysctl_sched_uclamp_util_max,
4696 		.maxlen         = sizeof(unsigned int),
4697 		.mode           = 0644,
4698 		.proc_handler   = sysctl_sched_uclamp_handler,
4699 	},
4700 	{
4701 		.procname       = "sched_util_clamp_min_rt_default",
4702 		.data           = &sysctl_sched_uclamp_util_min_rt_default,
4703 		.maxlen         = sizeof(unsigned int),
4704 		.mode           = 0644,
4705 		.proc_handler   = sysctl_sched_uclamp_handler,
4706 	},
4707 #endif /* CONFIG_UCLAMP_TASK */
4708 #ifdef CONFIG_NUMA_BALANCING
4709 	{
4710 		.procname	= "numa_balancing",
4711 		.data		= NULL, /* filled in by handler */
4712 		.maxlen		= sizeof(unsigned int),
4713 		.mode		= 0644,
4714 		.proc_handler	= sysctl_numa_balancing,
4715 		.extra1		= SYSCTL_ZERO,
4716 		.extra2		= SYSCTL_FOUR,
4717 	},
4718 #endif /* CONFIG_NUMA_BALANCING */
4719 	{}
4720 };
sched_core_sysctl_init(void)4721 static int __init sched_core_sysctl_init(void)
4722 {
4723 	register_sysctl_init("kernel", sched_core_sysctls);
4724 	return 0;
4725 }
4726 late_initcall(sched_core_sysctl_init);
4727 #endif /* CONFIG_SYSCTL */
4728 
4729 /*
4730  * fork()/clone()-time setup:
4731  */
sched_fork(unsigned long clone_flags,struct task_struct * p)4732 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4733 {
4734 	__sched_fork(clone_flags, p);
4735 	/*
4736 	 * We mark the process as NEW here. This guarantees that
4737 	 * nobody will actually run it, and a signal or other external
4738 	 * event cannot wake it up and insert it on the runqueue either.
4739 	 */
4740 	p->__state = TASK_NEW;
4741 
4742 	/*
4743 	 * Make sure we do not leak PI boosting priority to the child.
4744 	 */
4745 	p->prio = current->normal_prio;
4746 
4747 	uclamp_fork(p);
4748 
4749 	/*
4750 	 * Revert to default priority/policy on fork if requested.
4751 	 */
4752 	if (unlikely(p->sched_reset_on_fork)) {
4753 		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4754 			p->policy = SCHED_NORMAL;
4755 			p->static_prio = NICE_TO_PRIO(0);
4756 			p->rt_priority = 0;
4757 		} else if (PRIO_TO_NICE(p->static_prio) < 0)
4758 			p->static_prio = NICE_TO_PRIO(0);
4759 
4760 		p->prio = p->normal_prio = p->static_prio;
4761 		set_load_weight(p, false);
4762 
4763 		/*
4764 		 * We don't need the reset flag anymore after the fork. It has
4765 		 * fulfilled its duty:
4766 		 */
4767 		p->sched_reset_on_fork = 0;
4768 	}
4769 
4770 	if (dl_prio(p->prio))
4771 		return -EAGAIN;
4772 	else if (rt_prio(p->prio))
4773 		p->sched_class = &rt_sched_class;
4774 	else
4775 		p->sched_class = &fair_sched_class;
4776 
4777 	init_entity_runnable_average(&p->se);
4778 
4779 
4780 #ifdef CONFIG_SCHED_INFO
4781 	if (likely(sched_info_on()))
4782 		memset(&p->sched_info, 0, sizeof(p->sched_info));
4783 #endif
4784 #if defined(CONFIG_SMP)
4785 	p->on_cpu = 0;
4786 #endif
4787 	init_task_preempt_count(p);
4788 #ifdef CONFIG_SMP
4789 	plist_node_init(&p->pushable_tasks, MAX_PRIO);
4790 	RB_CLEAR_NODE(&p->pushable_dl_tasks);
4791 #endif
4792 	return 0;
4793 }
4794 
sched_cgroup_fork(struct task_struct * p,struct kernel_clone_args * kargs)4795 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4796 {
4797 	unsigned long flags;
4798 
4799 	/*
4800 	 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4801 	 * required yet, but lockdep gets upset if rules are violated.
4802 	 */
4803 	raw_spin_lock_irqsave(&p->pi_lock, flags);
4804 #ifdef CONFIG_CGROUP_SCHED
4805 	if (1) {
4806 		struct task_group *tg;
4807 		tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4808 				  struct task_group, css);
4809 		tg = autogroup_task_group(p, tg);
4810 		p->sched_task_group = tg;
4811 	}
4812 #endif
4813 	rseq_migrate(p);
4814 	/*
4815 	 * We're setting the CPU for the first time, we don't migrate,
4816 	 * so use __set_task_cpu().
4817 	 */
4818 	__set_task_cpu(p, smp_processor_id());
4819 	if (p->sched_class->task_fork)
4820 		p->sched_class->task_fork(p);
4821 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4822 }
4823 
sched_post_fork(struct task_struct * p)4824 void sched_post_fork(struct task_struct *p)
4825 {
4826 	uclamp_post_fork(p);
4827 }
4828 
to_ratio(u64 period,u64 runtime)4829 unsigned long to_ratio(u64 period, u64 runtime)
4830 {
4831 	if (runtime == RUNTIME_INF)
4832 		return BW_UNIT;
4833 
4834 	/*
4835 	 * Doing this here saves a lot of checks in all
4836 	 * the calling paths, and returning zero seems
4837 	 * safe for them anyway.
4838 	 */
4839 	if (period == 0)
4840 		return 0;
4841 
4842 	return div64_u64(runtime << BW_SHIFT, period);
4843 }
4844 
4845 /*
4846  * wake_up_new_task - wake up a newly created task for the first time.
4847  *
4848  * This function will do some initial scheduler statistics housekeeping
4849  * that must be done for every newly created context, then puts the task
4850  * on the runqueue and wakes it.
4851  */
wake_up_new_task(struct task_struct * p)4852 void wake_up_new_task(struct task_struct *p)
4853 {
4854 	struct rq_flags rf;
4855 	struct rq *rq;
4856 
4857 	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4858 	WRITE_ONCE(p->__state, TASK_RUNNING);
4859 #ifdef CONFIG_SMP
4860 	/*
4861 	 * Fork balancing, do it here and not earlier because:
4862 	 *  - cpus_ptr can change in the fork path
4863 	 *  - any previously selected CPU might disappear through hotplug
4864 	 *
4865 	 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4866 	 * as we're not fully set-up yet.
4867 	 */
4868 	p->recent_used_cpu = task_cpu(p);
4869 	rseq_migrate(p);
4870 	__set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4871 #endif
4872 	rq = __task_rq_lock(p, &rf);
4873 	update_rq_clock(rq);
4874 	post_init_entity_util_avg(p);
4875 
4876 	activate_task(rq, p, ENQUEUE_NOCLOCK);
4877 	trace_sched_wakeup_new(p);
4878 	check_preempt_curr(rq, p, WF_FORK);
4879 #ifdef CONFIG_SMP
4880 	if (p->sched_class->task_woken) {
4881 		/*
4882 		 * Nothing relies on rq->lock after this, so it's fine to
4883 		 * drop it.
4884 		 */
4885 		rq_unpin_lock(rq, &rf);
4886 		p->sched_class->task_woken(rq, p);
4887 		rq_repin_lock(rq, &rf);
4888 	}
4889 #endif
4890 	task_rq_unlock(rq, p, &rf);
4891 }
4892 
4893 #ifdef CONFIG_PREEMPT_NOTIFIERS
4894 
4895 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4896 
preempt_notifier_inc(void)4897 void preempt_notifier_inc(void)
4898 {
4899 	static_branch_inc(&preempt_notifier_key);
4900 }
4901 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4902 
preempt_notifier_dec(void)4903 void preempt_notifier_dec(void)
4904 {
4905 	static_branch_dec(&preempt_notifier_key);
4906 }
4907 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4908 
4909 /**
4910  * preempt_notifier_register - tell me when current is being preempted & rescheduled
4911  * @notifier: notifier struct to register
4912  */
preempt_notifier_register(struct preempt_notifier * notifier)4913 void preempt_notifier_register(struct preempt_notifier *notifier)
4914 {
4915 	if (!static_branch_unlikely(&preempt_notifier_key))
4916 		WARN(1, "registering preempt_notifier while notifiers disabled\n");
4917 
4918 	hlist_add_head(&notifier->link, &current->preempt_notifiers);
4919 }
4920 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4921 
4922 /**
4923  * preempt_notifier_unregister - no longer interested in preemption notifications
4924  * @notifier: notifier struct to unregister
4925  *
4926  * This is *not* safe to call from within a preemption notifier.
4927  */
preempt_notifier_unregister(struct preempt_notifier * notifier)4928 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4929 {
4930 	hlist_del(&notifier->link);
4931 }
4932 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4933 
__fire_sched_in_preempt_notifiers(struct task_struct * curr)4934 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4935 {
4936 	struct preempt_notifier *notifier;
4937 
4938 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4939 		notifier->ops->sched_in(notifier, raw_smp_processor_id());
4940 }
4941 
fire_sched_in_preempt_notifiers(struct task_struct * curr)4942 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4943 {
4944 	if (static_branch_unlikely(&preempt_notifier_key))
4945 		__fire_sched_in_preempt_notifiers(curr);
4946 }
4947 
4948 static void
__fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)4949 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4950 				   struct task_struct *next)
4951 {
4952 	struct preempt_notifier *notifier;
4953 
4954 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4955 		notifier->ops->sched_out(notifier, next);
4956 }
4957 
4958 static __always_inline void
fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)4959 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4960 				 struct task_struct *next)
4961 {
4962 	if (static_branch_unlikely(&preempt_notifier_key))
4963 		__fire_sched_out_preempt_notifiers(curr, next);
4964 }
4965 
4966 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4967 
fire_sched_in_preempt_notifiers(struct task_struct * curr)4968 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4969 {
4970 }
4971 
4972 static inline void
fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)4973 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4974 				 struct task_struct *next)
4975 {
4976 }
4977 
4978 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4979 
prepare_task(struct task_struct * next)4980 static inline void prepare_task(struct task_struct *next)
4981 {
4982 #ifdef CONFIG_SMP
4983 	/*
4984 	 * Claim the task as running, we do this before switching to it
4985 	 * such that any running task will have this set.
4986 	 *
4987 	 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4988 	 * its ordering comment.
4989 	 */
4990 	WRITE_ONCE(next->on_cpu, 1);
4991 #endif
4992 }
4993 
finish_task(struct task_struct * prev)4994 static inline void finish_task(struct task_struct *prev)
4995 {
4996 #ifdef CONFIG_SMP
4997 	/*
4998 	 * This must be the very last reference to @prev from this CPU. After
4999 	 * p->on_cpu is cleared, the task can be moved to a different CPU. We
5000 	 * must ensure this doesn't happen until the switch is completely
5001 	 * finished.
5002 	 *
5003 	 * In particular, the load of prev->state in finish_task_switch() must
5004 	 * happen before this.
5005 	 *
5006 	 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
5007 	 */
5008 	smp_store_release(&prev->on_cpu, 0);
5009 #endif
5010 }
5011 
5012 #ifdef CONFIG_SMP
5013 
do_balance_callbacks(struct rq * rq,struct balance_callback * head)5014 static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
5015 {
5016 	void (*func)(struct rq *rq);
5017 	struct balance_callback *next;
5018 
5019 	lockdep_assert_rq_held(rq);
5020 
5021 	while (head) {
5022 		func = (void (*)(struct rq *))head->func;
5023 		next = head->next;
5024 		head->next = NULL;
5025 		head = next;
5026 
5027 		func(rq);
5028 	}
5029 }
5030 
5031 static void balance_push(struct rq *rq);
5032 
5033 /*
5034  * balance_push_callback is a right abuse of the callback interface and plays
5035  * by significantly different rules.
5036  *
5037  * Where the normal balance_callback's purpose is to be ran in the same context
5038  * that queued it (only later, when it's safe to drop rq->lock again),
5039  * balance_push_callback is specifically targeted at __schedule().
5040  *
5041  * This abuse is tolerated because it places all the unlikely/odd cases behind
5042  * a single test, namely: rq->balance_callback == NULL.
5043  */
5044 struct balance_callback balance_push_callback = {
5045 	.next = NULL,
5046 	.func = balance_push,
5047 };
5048 
5049 static inline struct balance_callback *
__splice_balance_callbacks(struct rq * rq,bool split)5050 __splice_balance_callbacks(struct rq *rq, bool split)
5051 {
5052 	struct balance_callback *head = rq->balance_callback;
5053 
5054 	if (likely(!head))
5055 		return NULL;
5056 
5057 	lockdep_assert_rq_held(rq);
5058 	/*
5059 	 * Must not take balance_push_callback off the list when
5060 	 * splice_balance_callbacks() and balance_callbacks() are not
5061 	 * in the same rq->lock section.
5062 	 *
5063 	 * In that case it would be possible for __schedule() to interleave
5064 	 * and observe the list empty.
5065 	 */
5066 	if (split && head == &balance_push_callback)
5067 		head = NULL;
5068 	else
5069 		rq->balance_callback = NULL;
5070 
5071 	return head;
5072 }
5073 
splice_balance_callbacks(struct rq * rq)5074 static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5075 {
5076 	return __splice_balance_callbacks(rq, true);
5077 }
5078 
__balance_callbacks(struct rq * rq)5079 static void __balance_callbacks(struct rq *rq)
5080 {
5081 	do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
5082 }
5083 
balance_callbacks(struct rq * rq,struct balance_callback * head)5084 static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5085 {
5086 	unsigned long flags;
5087 
5088 	if (unlikely(head)) {
5089 		raw_spin_rq_lock_irqsave(rq, flags);
5090 		do_balance_callbacks(rq, head);
5091 		raw_spin_rq_unlock_irqrestore(rq, flags);
5092 	}
5093 }
5094 
5095 #else
5096 
__balance_callbacks(struct rq * rq)5097 static inline void __balance_callbacks(struct rq *rq)
5098 {
5099 }
5100 
splice_balance_callbacks(struct rq * rq)5101 static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5102 {
5103 	return NULL;
5104 }
5105 
balance_callbacks(struct rq * rq,struct balance_callback * head)5106 static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5107 {
5108 }
5109 
5110 #endif
5111 
5112 static inline void
prepare_lock_switch(struct rq * rq,struct task_struct * next,struct rq_flags * rf)5113 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
5114 {
5115 	/*
5116 	 * Since the runqueue lock will be released by the next
5117 	 * task (which is an invalid locking op but in the case
5118 	 * of the scheduler it's an obvious special-case), so we
5119 	 * do an early lockdep release here:
5120 	 */
5121 	rq_unpin_lock(rq, rf);
5122 	spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
5123 #ifdef CONFIG_DEBUG_SPINLOCK
5124 	/* this is a valid case when another task releases the spinlock */
5125 	rq_lockp(rq)->owner = next;
5126 #endif
5127 }
5128 
finish_lock_switch(struct rq * rq)5129 static inline void finish_lock_switch(struct rq *rq)
5130 {
5131 	/*
5132 	 * If we are tracking spinlock dependencies then we have to
5133 	 * fix up the runqueue lock - which gets 'carried over' from
5134 	 * prev into current:
5135 	 */
5136 	spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
5137 	__balance_callbacks(rq);
5138 	raw_spin_rq_unlock_irq(rq);
5139 }
5140 
5141 /*
5142  * NOP if the arch has not defined these:
5143  */
5144 
5145 #ifndef prepare_arch_switch
5146 # define prepare_arch_switch(next)	do { } while (0)
5147 #endif
5148 
5149 #ifndef finish_arch_post_lock_switch
5150 # define finish_arch_post_lock_switch()	do { } while (0)
5151 #endif
5152 
kmap_local_sched_out(void)5153 static inline void kmap_local_sched_out(void)
5154 {
5155 #ifdef CONFIG_KMAP_LOCAL
5156 	if (unlikely(current->kmap_ctrl.idx))
5157 		__kmap_local_sched_out();
5158 #endif
5159 }
5160 
kmap_local_sched_in(void)5161 static inline void kmap_local_sched_in(void)
5162 {
5163 #ifdef CONFIG_KMAP_LOCAL
5164 	if (unlikely(current->kmap_ctrl.idx))
5165 		__kmap_local_sched_in();
5166 #endif
5167 }
5168 
5169 /**
5170  * prepare_task_switch - prepare to switch tasks
5171  * @rq: the runqueue preparing to switch
5172  * @prev: the current task that is being switched out
5173  * @next: the task we are going to switch to.
5174  *
5175  * This is called with the rq lock held and interrupts off. It must
5176  * be paired with a subsequent finish_task_switch after the context
5177  * switch.
5178  *
5179  * prepare_task_switch sets up locking and calls architecture specific
5180  * hooks.
5181  */
5182 static inline void
prepare_task_switch(struct rq * rq,struct task_struct * prev,struct task_struct * next)5183 prepare_task_switch(struct rq *rq, struct task_struct *prev,
5184 		    struct task_struct *next)
5185 {
5186 	kcov_prepare_switch(prev);
5187 	sched_info_switch(rq, prev, next);
5188 	perf_event_task_sched_out(prev, next);
5189 	rseq_preempt(prev);
5190 	fire_sched_out_preempt_notifiers(prev, next);
5191 	kmap_local_sched_out();
5192 	prepare_task(next);
5193 	prepare_arch_switch(next);
5194 }
5195 
5196 /**
5197  * finish_task_switch - clean up after a task-switch
5198  * @prev: the thread we just switched away from.
5199  *
5200  * finish_task_switch must be called after the context switch, paired
5201  * with a prepare_task_switch call before the context switch.
5202  * finish_task_switch will reconcile locking set up by prepare_task_switch,
5203  * and do any other architecture-specific cleanup actions.
5204  *
5205  * Note that we may have delayed dropping an mm in context_switch(). If
5206  * so, we finish that here outside of the runqueue lock. (Doing it
5207  * with the lock held can cause deadlocks; see schedule() for
5208  * details.)
5209  *
5210  * The context switch have flipped the stack from under us and restored the
5211  * local variables which were saved when this task called schedule() in the
5212  * past. prev == current is still correct but we need to recalculate this_rq
5213  * because prev may have moved to another CPU.
5214  */
finish_task_switch(struct task_struct * prev)5215 static struct rq *finish_task_switch(struct task_struct *prev)
5216 	__releases(rq->lock)
5217 {
5218 	struct rq *rq = this_rq();
5219 	struct mm_struct *mm = rq->prev_mm;
5220 	unsigned int prev_state;
5221 
5222 	/*
5223 	 * The previous task will have left us with a preempt_count of 2
5224 	 * because it left us after:
5225 	 *
5226 	 *	schedule()
5227 	 *	  preempt_disable();			// 1
5228 	 *	  __schedule()
5229 	 *	    raw_spin_lock_irq(&rq->lock)	// 2
5230 	 *
5231 	 * Also, see FORK_PREEMPT_COUNT.
5232 	 */
5233 	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5234 		      "corrupted preempt_count: %s/%d/0x%x\n",
5235 		      current->comm, current->pid, preempt_count()))
5236 		preempt_count_set(FORK_PREEMPT_COUNT);
5237 
5238 	rq->prev_mm = NULL;
5239 
5240 	/*
5241 	 * A task struct has one reference for the use as "current".
5242 	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5243 	 * schedule one last time. The schedule call will never return, and
5244 	 * the scheduled task must drop that reference.
5245 	 *
5246 	 * We must observe prev->state before clearing prev->on_cpu (in
5247 	 * finish_task), otherwise a concurrent wakeup can get prev
5248 	 * running on another CPU and we could rave with its RUNNING -> DEAD
5249 	 * transition, resulting in a double drop.
5250 	 */
5251 	prev_state = READ_ONCE(prev->__state);
5252 	vtime_task_switch(prev);
5253 	perf_event_task_sched_in(prev, current);
5254 	finish_task(prev);
5255 	tick_nohz_task_switch();
5256 	finish_lock_switch(rq);
5257 	finish_arch_post_lock_switch();
5258 	kcov_finish_switch(current);
5259 	/*
5260 	 * kmap_local_sched_out() is invoked with rq::lock held and
5261 	 * interrupts disabled. There is no requirement for that, but the
5262 	 * sched out code does not have an interrupt enabled section.
5263 	 * Restoring the maps on sched in does not require interrupts being
5264 	 * disabled either.
5265 	 */
5266 	kmap_local_sched_in();
5267 
5268 	fire_sched_in_preempt_notifiers(current);
5269 	/*
5270 	 * When switching through a kernel thread, the loop in
5271 	 * membarrier_{private,global}_expedited() may have observed that
5272 	 * kernel thread and not issued an IPI. It is therefore possible to
5273 	 * schedule between user->kernel->user threads without passing though
5274 	 * switch_mm(). Membarrier requires a barrier after storing to
5275 	 * rq->curr, before returning to userspace, so provide them here:
5276 	 *
5277 	 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5278 	 *   provided by mmdrop_lazy_tlb(),
5279 	 * - a sync_core for SYNC_CORE.
5280 	 */
5281 	if (mm) {
5282 		membarrier_mm_sync_core_before_usermode(mm);
5283 		mmdrop_lazy_tlb_sched(mm);
5284 	}
5285 
5286 	if (unlikely(prev_state == TASK_DEAD)) {
5287 		if (prev->sched_class->task_dead)
5288 			prev->sched_class->task_dead(prev);
5289 
5290 		/* Task is done with its stack. */
5291 		put_task_stack(prev);
5292 
5293 		put_task_struct_rcu_user(prev);
5294 	}
5295 
5296 	return rq;
5297 }
5298 
5299 /**
5300  * schedule_tail - first thing a freshly forked thread must call.
5301  * @prev: the thread we just switched away from.
5302  */
schedule_tail(struct task_struct * prev)5303 asmlinkage __visible void schedule_tail(struct task_struct *prev)
5304 	__releases(rq->lock)
5305 {
5306 	/*
5307 	 * New tasks start with FORK_PREEMPT_COUNT, see there and
5308 	 * finish_task_switch() for details.
5309 	 *
5310 	 * finish_task_switch() will drop rq->lock() and lower preempt_count
5311 	 * and the preempt_enable() will end up enabling preemption (on
5312 	 * PREEMPT_COUNT kernels).
5313 	 */
5314 
5315 	finish_task_switch(prev);
5316 	preempt_enable();
5317 
5318 	if (current->set_child_tid)
5319 		put_user(task_pid_vnr(current), current->set_child_tid);
5320 
5321 	calculate_sigpending();
5322 }
5323 
5324 /*
5325  * context_switch - switch to the new MM and the new thread's register state.
5326  */
5327 static __always_inline struct rq *
context_switch(struct rq * rq,struct task_struct * prev,struct task_struct * next,struct rq_flags * rf)5328 context_switch(struct rq *rq, struct task_struct *prev,
5329 	       struct task_struct *next, struct rq_flags *rf)
5330 {
5331 	prepare_task_switch(rq, prev, next);
5332 
5333 	/*
5334 	 * For paravirt, this is coupled with an exit in switch_to to
5335 	 * combine the page table reload and the switch backend into
5336 	 * one hypercall.
5337 	 */
5338 	arch_start_context_switch(prev);
5339 
5340 	/*
5341 	 * kernel -> kernel   lazy + transfer active
5342 	 *   user -> kernel   lazy + mmgrab_lazy_tlb() active
5343 	 *
5344 	 * kernel ->   user   switch + mmdrop_lazy_tlb() active
5345 	 *   user ->   user   switch
5346 	 *
5347 	 * switch_mm_cid() needs to be updated if the barriers provided
5348 	 * by context_switch() are modified.
5349 	 */
5350 	if (!next->mm) {                                // to kernel
5351 		enter_lazy_tlb(prev->active_mm, next);
5352 
5353 		next->active_mm = prev->active_mm;
5354 		if (prev->mm)                           // from user
5355 			mmgrab_lazy_tlb(prev->active_mm);
5356 		else
5357 			prev->active_mm = NULL;
5358 	} else {                                        // to user
5359 		membarrier_switch_mm(rq, prev->active_mm, next->mm);
5360 		/*
5361 		 * sys_membarrier() requires an smp_mb() between setting
5362 		 * rq->curr / membarrier_switch_mm() and returning to userspace.
5363 		 *
5364 		 * The below provides this either through switch_mm(), or in
5365 		 * case 'prev->active_mm == next->mm' through
5366 		 * finish_task_switch()'s mmdrop().
5367 		 */
5368 		switch_mm_irqs_off(prev->active_mm, next->mm, next);
5369 		lru_gen_use_mm(next->mm);
5370 
5371 		if (!prev->mm) {                        // from kernel
5372 			/* will mmdrop_lazy_tlb() in finish_task_switch(). */
5373 			rq->prev_mm = prev->active_mm;
5374 			prev->active_mm = NULL;
5375 		}
5376 	}
5377 
5378 	/* switch_mm_cid() requires the memory barriers above. */
5379 	switch_mm_cid(rq, prev, next);
5380 
5381 	prepare_lock_switch(rq, next, rf);
5382 
5383 	/* Here we just switch the register state and the stack. */
5384 	switch_to(prev, next, prev);
5385 	barrier();
5386 
5387 	return finish_task_switch(prev);
5388 }
5389 
5390 /*
5391  * nr_running and nr_context_switches:
5392  *
5393  * externally visible scheduler statistics: current number of runnable
5394  * threads, total number of context switches performed since bootup.
5395  */
nr_running(void)5396 unsigned int nr_running(void)
5397 {
5398 	unsigned int i, sum = 0;
5399 
5400 	for_each_online_cpu(i)
5401 		sum += cpu_rq(i)->nr_running;
5402 
5403 	return sum;
5404 }
5405 
5406 /*
5407  * Check if only the current task is running on the CPU.
5408  *
5409  * Caution: this function does not check that the caller has disabled
5410  * preemption, thus the result might have a time-of-check-to-time-of-use
5411  * race.  The caller is responsible to use it correctly, for example:
5412  *
5413  * - from a non-preemptible section (of course)
5414  *
5415  * - from a thread that is bound to a single CPU
5416  *
5417  * - in a loop with very short iterations (e.g. a polling loop)
5418  */
single_task_running(void)5419 bool single_task_running(void)
5420 {
5421 	return raw_rq()->nr_running == 1;
5422 }
5423 EXPORT_SYMBOL(single_task_running);
5424 
nr_context_switches_cpu(int cpu)5425 unsigned long long nr_context_switches_cpu(int cpu)
5426 {
5427 	return cpu_rq(cpu)->nr_switches;
5428 }
5429 
nr_context_switches(void)5430 unsigned long long nr_context_switches(void)
5431 {
5432 	int i;
5433 	unsigned long long sum = 0;
5434 
5435 	for_each_possible_cpu(i)
5436 		sum += cpu_rq(i)->nr_switches;
5437 
5438 	return sum;
5439 }
5440 
5441 /*
5442  * Consumers of these two interfaces, like for example the cpuidle menu
5443  * governor, are using nonsensical data. Preferring shallow idle state selection
5444  * for a CPU that has IO-wait which might not even end up running the task when
5445  * it does become runnable.
5446  */
5447 
nr_iowait_cpu(int cpu)5448 unsigned int nr_iowait_cpu(int cpu)
5449 {
5450 	return atomic_read(&cpu_rq(cpu)->nr_iowait);
5451 }
5452 
5453 /*
5454  * IO-wait accounting, and how it's mostly bollocks (on SMP).
5455  *
5456  * The idea behind IO-wait account is to account the idle time that we could
5457  * have spend running if it were not for IO. That is, if we were to improve the
5458  * storage performance, we'd have a proportional reduction in IO-wait time.
5459  *
5460  * This all works nicely on UP, where, when a task blocks on IO, we account
5461  * idle time as IO-wait, because if the storage were faster, it could've been
5462  * running and we'd not be idle.
5463  *
5464  * This has been extended to SMP, by doing the same for each CPU. This however
5465  * is broken.
5466  *
5467  * Imagine for instance the case where two tasks block on one CPU, only the one
5468  * CPU will have IO-wait accounted, while the other has regular idle. Even
5469  * though, if the storage were faster, both could've ran at the same time,
5470  * utilising both CPUs.
5471  *
5472  * This means, that when looking globally, the current IO-wait accounting on
5473  * SMP is a lower bound, by reason of under accounting.
5474  *
5475  * Worse, since the numbers are provided per CPU, they are sometimes
5476  * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5477  * associated with any one particular CPU, it can wake to another CPU than it
5478  * blocked on. This means the per CPU IO-wait number is meaningless.
5479  *
5480  * Task CPU affinities can make all that even more 'interesting'.
5481  */
5482 
nr_iowait(void)5483 unsigned int nr_iowait(void)
5484 {
5485 	unsigned int i, sum = 0;
5486 
5487 	for_each_possible_cpu(i)
5488 		sum += nr_iowait_cpu(i);
5489 
5490 	return sum;
5491 }
5492 
5493 #ifdef CONFIG_SMP
5494 
5495 /*
5496  * sched_exec - execve() is a valuable balancing opportunity, because at
5497  * this point the task has the smallest effective memory and cache footprint.
5498  */
sched_exec(void)5499 void sched_exec(void)
5500 {
5501 	struct task_struct *p = current;
5502 	struct migration_arg arg;
5503 	int dest_cpu;
5504 
5505 	scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
5506 		dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5507 		if (dest_cpu == smp_processor_id())
5508 			return;
5509 
5510 		if (unlikely(!cpu_active(dest_cpu)))
5511 			return;
5512 
5513 		arg = (struct migration_arg){ p, dest_cpu };
5514 	}
5515 	stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5516 }
5517 
5518 #endif
5519 
5520 DEFINE_PER_CPU(struct kernel_stat, kstat);
5521 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5522 
5523 EXPORT_PER_CPU_SYMBOL(kstat);
5524 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5525 
5526 /*
5527  * The function fair_sched_class.update_curr accesses the struct curr
5528  * and its field curr->exec_start; when called from task_sched_runtime(),
5529  * we observe a high rate of cache misses in practice.
5530  * Prefetching this data results in improved performance.
5531  */
prefetch_curr_exec_start(struct task_struct * p)5532 static inline void prefetch_curr_exec_start(struct task_struct *p)
5533 {
5534 #ifdef CONFIG_FAIR_GROUP_SCHED
5535 	struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5536 #else
5537 	struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5538 #endif
5539 	prefetch(curr);
5540 	prefetch(&curr->exec_start);
5541 }
5542 
5543 /*
5544  * Return accounted runtime for the task.
5545  * In case the task is currently running, return the runtime plus current's
5546  * pending runtime that have not been accounted yet.
5547  */
task_sched_runtime(struct task_struct * p)5548 unsigned long long task_sched_runtime(struct task_struct *p)
5549 {
5550 	struct rq_flags rf;
5551 	struct rq *rq;
5552 	u64 ns;
5553 
5554 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5555 	/*
5556 	 * 64-bit doesn't need locks to atomically read a 64-bit value.
5557 	 * So we have a optimization chance when the task's delta_exec is 0.
5558 	 * Reading ->on_cpu is racy, but this is ok.
5559 	 *
5560 	 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5561 	 * If we race with it entering CPU, unaccounted time is 0. This is
5562 	 * indistinguishable from the read occurring a few cycles earlier.
5563 	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5564 	 * been accounted, so we're correct here as well.
5565 	 */
5566 	if (!p->on_cpu || !task_on_rq_queued(p))
5567 		return p->se.sum_exec_runtime;
5568 #endif
5569 
5570 	rq = task_rq_lock(p, &rf);
5571 	/*
5572 	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
5573 	 * project cycles that may never be accounted to this
5574 	 * thread, breaking clock_gettime().
5575 	 */
5576 	if (task_current(rq, p) && task_on_rq_queued(p)) {
5577 		prefetch_curr_exec_start(p);
5578 		update_rq_clock(rq);
5579 		p->sched_class->update_curr(rq);
5580 	}
5581 	ns = p->se.sum_exec_runtime;
5582 	task_rq_unlock(rq, p, &rf);
5583 
5584 	return ns;
5585 }
5586 
5587 #ifdef CONFIG_SCHED_DEBUG
cpu_resched_latency(struct rq * rq)5588 static u64 cpu_resched_latency(struct rq *rq)
5589 {
5590 	int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5591 	u64 resched_latency, now = rq_clock(rq);
5592 	static bool warned_once;
5593 
5594 	if (sysctl_resched_latency_warn_once && warned_once)
5595 		return 0;
5596 
5597 	if (!need_resched() || !latency_warn_ms)
5598 		return 0;
5599 
5600 	if (system_state == SYSTEM_BOOTING)
5601 		return 0;
5602 
5603 	if (!rq->last_seen_need_resched_ns) {
5604 		rq->last_seen_need_resched_ns = now;
5605 		rq->ticks_without_resched = 0;
5606 		return 0;
5607 	}
5608 
5609 	rq->ticks_without_resched++;
5610 	resched_latency = now - rq->last_seen_need_resched_ns;
5611 	if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5612 		return 0;
5613 
5614 	warned_once = true;
5615 
5616 	return resched_latency;
5617 }
5618 
setup_resched_latency_warn_ms(char * str)5619 static int __init setup_resched_latency_warn_ms(char *str)
5620 {
5621 	long val;
5622 
5623 	if ((kstrtol(str, 0, &val))) {
5624 		pr_warn("Unable to set resched_latency_warn_ms\n");
5625 		return 1;
5626 	}
5627 
5628 	sysctl_resched_latency_warn_ms = val;
5629 	return 1;
5630 }
5631 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5632 #else
cpu_resched_latency(struct rq * rq)5633 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5634 #endif /* CONFIG_SCHED_DEBUG */
5635 
5636 /*
5637  * This function gets called by the timer code, with HZ frequency.
5638  * We call it with interrupts disabled.
5639  */
scheduler_tick(void)5640 void scheduler_tick(void)
5641 {
5642 	int cpu = smp_processor_id();
5643 	struct rq *rq = cpu_rq(cpu);
5644 	struct task_struct *curr = rq->curr;
5645 	struct rq_flags rf;
5646 	unsigned long thermal_pressure;
5647 	u64 resched_latency;
5648 
5649 	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5650 		arch_scale_freq_tick();
5651 
5652 	sched_clock_tick();
5653 
5654 	rq_lock(rq, &rf);
5655 
5656 	update_rq_clock(rq);
5657 	thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5658 	update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5659 	curr->sched_class->task_tick(rq, curr, 0);
5660 	if (sched_feat(LATENCY_WARN))
5661 		resched_latency = cpu_resched_latency(rq);
5662 	calc_global_load_tick(rq);
5663 	sched_core_tick(rq);
5664 	task_tick_mm_cid(rq, curr);
5665 
5666 	rq_unlock(rq, &rf);
5667 
5668 	if (sched_feat(LATENCY_WARN) && resched_latency)
5669 		resched_latency_warn(cpu, resched_latency);
5670 
5671 	perf_event_task_tick();
5672 
5673 	if (curr->flags & PF_WQ_WORKER)
5674 		wq_worker_tick(curr);
5675 
5676 #ifdef CONFIG_SMP
5677 	rq->idle_balance = idle_cpu(cpu);
5678 	trigger_load_balance(rq);
5679 #endif
5680 }
5681 
5682 #ifdef CONFIG_NO_HZ_FULL
5683 
5684 struct tick_work {
5685 	int			cpu;
5686 	atomic_t		state;
5687 	struct delayed_work	work;
5688 };
5689 /* Values for ->state, see diagram below. */
5690 #define TICK_SCHED_REMOTE_OFFLINE	0
5691 #define TICK_SCHED_REMOTE_OFFLINING	1
5692 #define TICK_SCHED_REMOTE_RUNNING	2
5693 
5694 /*
5695  * State diagram for ->state:
5696  *
5697  *
5698  *          TICK_SCHED_REMOTE_OFFLINE
5699  *                    |   ^
5700  *                    |   |
5701  *                    |   | sched_tick_remote()
5702  *                    |   |
5703  *                    |   |
5704  *                    +--TICK_SCHED_REMOTE_OFFLINING
5705  *                    |   ^
5706  *                    |   |
5707  * sched_tick_start() |   | sched_tick_stop()
5708  *                    |   |
5709  *                    V   |
5710  *          TICK_SCHED_REMOTE_RUNNING
5711  *
5712  *
5713  * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5714  * and sched_tick_start() are happy to leave the state in RUNNING.
5715  */
5716 
5717 static struct tick_work __percpu *tick_work_cpu;
5718 
sched_tick_remote(struct work_struct * work)5719 static void sched_tick_remote(struct work_struct *work)
5720 {
5721 	struct delayed_work *dwork = to_delayed_work(work);
5722 	struct tick_work *twork = container_of(dwork, struct tick_work, work);
5723 	int cpu = twork->cpu;
5724 	struct rq *rq = cpu_rq(cpu);
5725 	int os;
5726 
5727 	/*
5728 	 * Handle the tick only if it appears the remote CPU is running in full
5729 	 * dynticks mode. The check is racy by nature, but missing a tick or
5730 	 * having one too much is no big deal because the scheduler tick updates
5731 	 * statistics and checks timeslices in a time-independent way, regardless
5732 	 * of when exactly it is running.
5733 	 */
5734 	if (tick_nohz_tick_stopped_cpu(cpu)) {
5735 		guard(rq_lock_irq)(rq);
5736 		struct task_struct *curr = rq->curr;
5737 
5738 		if (cpu_online(cpu)) {
5739 			update_rq_clock(rq);
5740 
5741 			if (!is_idle_task(curr)) {
5742 				/*
5743 				 * Make sure the next tick runs within a
5744 				 * reasonable amount of time.
5745 				 */
5746 				u64 delta = rq_clock_task(rq) - curr->se.exec_start;
5747 				WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5748 			}
5749 			curr->sched_class->task_tick(rq, curr, 0);
5750 
5751 			calc_load_nohz_remote(rq);
5752 		}
5753 	}
5754 
5755 	/*
5756 	 * Run the remote tick once per second (1Hz). This arbitrary
5757 	 * frequency is large enough to avoid overload but short enough
5758 	 * to keep scheduler internal stats reasonably up to date.  But
5759 	 * first update state to reflect hotplug activity if required.
5760 	 */
5761 	os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5762 	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5763 	if (os == TICK_SCHED_REMOTE_RUNNING)
5764 		queue_delayed_work(system_unbound_wq, dwork, HZ);
5765 }
5766 
sched_tick_start(int cpu)5767 static void sched_tick_start(int cpu)
5768 {
5769 	int os;
5770 	struct tick_work *twork;
5771 
5772 	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5773 		return;
5774 
5775 	WARN_ON_ONCE(!tick_work_cpu);
5776 
5777 	twork = per_cpu_ptr(tick_work_cpu, cpu);
5778 	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5779 	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5780 	if (os == TICK_SCHED_REMOTE_OFFLINE) {
5781 		twork->cpu = cpu;
5782 		INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5783 		queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5784 	}
5785 }
5786 
5787 #ifdef CONFIG_HOTPLUG_CPU
sched_tick_stop(int cpu)5788 static void sched_tick_stop(int cpu)
5789 {
5790 	struct tick_work *twork;
5791 	int os;
5792 
5793 	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5794 		return;
5795 
5796 	WARN_ON_ONCE(!tick_work_cpu);
5797 
5798 	twork = per_cpu_ptr(tick_work_cpu, cpu);
5799 	/* There cannot be competing actions, but don't rely on stop-machine. */
5800 	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5801 	WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5802 	/* Don't cancel, as this would mess up the state machine. */
5803 }
5804 #endif /* CONFIG_HOTPLUG_CPU */
5805 
sched_tick_offload_init(void)5806 int __init sched_tick_offload_init(void)
5807 {
5808 	tick_work_cpu = alloc_percpu(struct tick_work);
5809 	BUG_ON(!tick_work_cpu);
5810 	return 0;
5811 }
5812 
5813 #else /* !CONFIG_NO_HZ_FULL */
sched_tick_start(int cpu)5814 static inline void sched_tick_start(int cpu) { }
sched_tick_stop(int cpu)5815 static inline void sched_tick_stop(int cpu) { }
5816 #endif
5817 
5818 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5819 				defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5820 /*
5821  * If the value passed in is equal to the current preempt count
5822  * then we just disabled preemption. Start timing the latency.
5823  */
preempt_latency_start(int val)5824 static inline void preempt_latency_start(int val)
5825 {
5826 	if (preempt_count() == val) {
5827 		unsigned long ip = get_lock_parent_ip();
5828 #ifdef CONFIG_DEBUG_PREEMPT
5829 		current->preempt_disable_ip = ip;
5830 #endif
5831 		trace_preempt_off(CALLER_ADDR0, ip);
5832 	}
5833 }
5834 
preempt_count_add(int val)5835 void preempt_count_add(int val)
5836 {
5837 #ifdef CONFIG_DEBUG_PREEMPT
5838 	/*
5839 	 * Underflow?
5840 	 */
5841 	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5842 		return;
5843 #endif
5844 	__preempt_count_add(val);
5845 #ifdef CONFIG_DEBUG_PREEMPT
5846 	/*
5847 	 * Spinlock count overflowing soon?
5848 	 */
5849 	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5850 				PREEMPT_MASK - 10);
5851 #endif
5852 	preempt_latency_start(val);
5853 }
5854 EXPORT_SYMBOL(preempt_count_add);
5855 NOKPROBE_SYMBOL(preempt_count_add);
5856 
5857 /*
5858  * If the value passed in equals to the current preempt count
5859  * then we just enabled preemption. Stop timing the latency.
5860  */
preempt_latency_stop(int val)5861 static inline void preempt_latency_stop(int val)
5862 {
5863 	if (preempt_count() == val)
5864 		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5865 }
5866 
preempt_count_sub(int val)5867 void preempt_count_sub(int val)
5868 {
5869 #ifdef CONFIG_DEBUG_PREEMPT
5870 	/*
5871 	 * Underflow?
5872 	 */
5873 	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5874 		return;
5875 	/*
5876 	 * Is the spinlock portion underflowing?
5877 	 */
5878 	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5879 			!(preempt_count() & PREEMPT_MASK)))
5880 		return;
5881 #endif
5882 
5883 	preempt_latency_stop(val);
5884 	__preempt_count_sub(val);
5885 }
5886 EXPORT_SYMBOL(preempt_count_sub);
5887 NOKPROBE_SYMBOL(preempt_count_sub);
5888 
5889 #else
preempt_latency_start(int val)5890 static inline void preempt_latency_start(int val) { }
preempt_latency_stop(int val)5891 static inline void preempt_latency_stop(int val) { }
5892 #endif
5893 
get_preempt_disable_ip(struct task_struct * p)5894 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5895 {
5896 #ifdef CONFIG_DEBUG_PREEMPT
5897 	return p->preempt_disable_ip;
5898 #else
5899 	return 0;
5900 #endif
5901 }
5902 
5903 /*
5904  * Print scheduling while atomic bug:
5905  */
__schedule_bug(struct task_struct * prev)5906 static noinline void __schedule_bug(struct task_struct *prev)
5907 {
5908 	/* Save this before calling printk(), since that will clobber it */
5909 	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5910 
5911 	if (oops_in_progress)
5912 		return;
5913 
5914 	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5915 		prev->comm, prev->pid, preempt_count());
5916 
5917 	debug_show_held_locks(prev);
5918 	print_modules();
5919 	if (irqs_disabled())
5920 		print_irqtrace_events(prev);
5921 	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5922 	    && in_atomic_preempt_off()) {
5923 		pr_err("Preemption disabled at:");
5924 		print_ip_sym(KERN_ERR, preempt_disable_ip);
5925 	}
5926 	check_panic_on_warn("scheduling while atomic");
5927 
5928 	dump_stack();
5929 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5930 }
5931 
5932 /*
5933  * Various schedule()-time debugging checks and statistics:
5934  */
schedule_debug(struct task_struct * prev,bool preempt)5935 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5936 {
5937 #ifdef CONFIG_SCHED_STACK_END_CHECK
5938 	if (task_stack_end_corrupted(prev))
5939 		panic("corrupted stack end detected inside scheduler\n");
5940 
5941 	if (task_scs_end_corrupted(prev))
5942 		panic("corrupted shadow stack detected inside scheduler\n");
5943 #endif
5944 
5945 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5946 	if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5947 		printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5948 			prev->comm, prev->pid, prev->non_block_count);
5949 		dump_stack();
5950 		add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5951 	}
5952 #endif
5953 
5954 	if (unlikely(in_atomic_preempt_off())) {
5955 		__schedule_bug(prev);
5956 		preempt_count_set(PREEMPT_DISABLED);
5957 	}
5958 	rcu_sleep_check();
5959 	SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5960 
5961 	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5962 
5963 	schedstat_inc(this_rq()->sched_count);
5964 }
5965 
put_prev_task_balance(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)5966 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5967 				  struct rq_flags *rf)
5968 {
5969 #ifdef CONFIG_SMP
5970 	const struct sched_class *class;
5971 	/*
5972 	 * We must do the balancing pass before put_prev_task(), such
5973 	 * that when we release the rq->lock the task is in the same
5974 	 * state as before we took rq->lock.
5975 	 *
5976 	 * We can terminate the balance pass as soon as we know there is
5977 	 * a runnable task of @class priority or higher.
5978 	 */
5979 	for_class_range(class, prev->sched_class, &idle_sched_class) {
5980 		if (class->balance(rq, prev, rf))
5981 			break;
5982 	}
5983 #endif
5984 
5985 	put_prev_task(rq, prev);
5986 }
5987 
5988 /*
5989  * Pick up the highest-prio task:
5990  */
5991 static inline struct task_struct *
__pick_next_task(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)5992 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5993 {
5994 	const struct sched_class *class;
5995 	struct task_struct *p;
5996 
5997 	/*
5998 	 * Optimization: we know that if all tasks are in the fair class we can
5999 	 * call that function directly, but only if the @prev task wasn't of a
6000 	 * higher scheduling class, because otherwise those lose the
6001 	 * opportunity to pull in more work from other CPUs.
6002 	 */
6003 	if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
6004 		   rq->nr_running == rq->cfs.h_nr_running)) {
6005 
6006 		p = pick_next_task_fair(rq, prev, rf);
6007 		if (unlikely(p == RETRY_TASK))
6008 			goto restart;
6009 
6010 		/* Assume the next prioritized class is idle_sched_class */
6011 		if (!p) {
6012 			put_prev_task(rq, prev);
6013 			p = pick_next_task_idle(rq);
6014 		}
6015 
6016 		return p;
6017 	}
6018 
6019 restart:
6020 	put_prev_task_balance(rq, prev, rf);
6021 
6022 	for_each_class(class) {
6023 		p = class->pick_next_task(rq);
6024 		if (p)
6025 			return p;
6026 	}
6027 
6028 	BUG(); /* The idle class should always have a runnable task. */
6029 }
6030 
6031 #ifdef CONFIG_SCHED_CORE
is_task_rq_idle(struct task_struct * t)6032 static inline bool is_task_rq_idle(struct task_struct *t)
6033 {
6034 	return (task_rq(t)->idle == t);
6035 }
6036 
cookie_equals(struct task_struct * a,unsigned long cookie)6037 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
6038 {
6039 	return is_task_rq_idle(a) || (a->core_cookie == cookie);
6040 }
6041 
cookie_match(struct task_struct * a,struct task_struct * b)6042 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
6043 {
6044 	if (is_task_rq_idle(a) || is_task_rq_idle(b))
6045 		return true;
6046 
6047 	return a->core_cookie == b->core_cookie;
6048 }
6049 
pick_task(struct rq * rq)6050 static inline struct task_struct *pick_task(struct rq *rq)
6051 {
6052 	const struct sched_class *class;
6053 	struct task_struct *p;
6054 
6055 	for_each_class(class) {
6056 		p = class->pick_task(rq);
6057 		if (p)
6058 			return p;
6059 	}
6060 
6061 	BUG(); /* The idle class should always have a runnable task. */
6062 }
6063 
6064 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
6065 
6066 static void queue_core_balance(struct rq *rq);
6067 
6068 static struct task_struct *
pick_next_task(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)6069 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6070 {
6071 	struct task_struct *next, *p, *max = NULL;
6072 	const struct cpumask *smt_mask;
6073 	bool fi_before = false;
6074 	bool core_clock_updated = (rq == rq->core);
6075 	unsigned long cookie;
6076 	int i, cpu, occ = 0;
6077 	struct rq *rq_i;
6078 	bool need_sync;
6079 
6080 	if (!sched_core_enabled(rq))
6081 		return __pick_next_task(rq, prev, rf);
6082 
6083 	cpu = cpu_of(rq);
6084 
6085 	/* Stopper task is switching into idle, no need core-wide selection. */
6086 	if (cpu_is_offline(cpu)) {
6087 		/*
6088 		 * Reset core_pick so that we don't enter the fastpath when
6089 		 * coming online. core_pick would already be migrated to
6090 		 * another cpu during offline.
6091 		 */
6092 		rq->core_pick = NULL;
6093 		return __pick_next_task(rq, prev, rf);
6094 	}
6095 
6096 	/*
6097 	 * If there were no {en,de}queues since we picked (IOW, the task
6098 	 * pointers are all still valid), and we haven't scheduled the last
6099 	 * pick yet, do so now.
6100 	 *
6101 	 * rq->core_pick can be NULL if no selection was made for a CPU because
6102 	 * it was either offline or went offline during a sibling's core-wide
6103 	 * selection. In this case, do a core-wide selection.
6104 	 */
6105 	if (rq->core->core_pick_seq == rq->core->core_task_seq &&
6106 	    rq->core->core_pick_seq != rq->core_sched_seq &&
6107 	    rq->core_pick) {
6108 		WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
6109 
6110 		next = rq->core_pick;
6111 		if (next != prev) {
6112 			put_prev_task(rq, prev);
6113 			set_next_task(rq, next);
6114 		}
6115 
6116 		rq->core_pick = NULL;
6117 		goto out;
6118 	}
6119 
6120 	put_prev_task_balance(rq, prev, rf);
6121 
6122 	smt_mask = cpu_smt_mask(cpu);
6123 	need_sync = !!rq->core->core_cookie;
6124 
6125 	/* reset state */
6126 	rq->core->core_cookie = 0UL;
6127 	if (rq->core->core_forceidle_count) {
6128 		if (!core_clock_updated) {
6129 			update_rq_clock(rq->core);
6130 			core_clock_updated = true;
6131 		}
6132 		sched_core_account_forceidle(rq);
6133 		/* reset after accounting force idle */
6134 		rq->core->core_forceidle_start = 0;
6135 		rq->core->core_forceidle_count = 0;
6136 		rq->core->core_forceidle_occupation = 0;
6137 		need_sync = true;
6138 		fi_before = true;
6139 	}
6140 
6141 	/*
6142 	 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6143 	 *
6144 	 * @task_seq guards the task state ({en,de}queues)
6145 	 * @pick_seq is the @task_seq we did a selection on
6146 	 * @sched_seq is the @pick_seq we scheduled
6147 	 *
6148 	 * However, preemptions can cause multiple picks on the same task set.
6149 	 * 'Fix' this by also increasing @task_seq for every pick.
6150 	 */
6151 	rq->core->core_task_seq++;
6152 
6153 	/*
6154 	 * Optimize for common case where this CPU has no cookies
6155 	 * and there are no cookied tasks running on siblings.
6156 	 */
6157 	if (!need_sync) {
6158 		next = pick_task(rq);
6159 		if (!next->core_cookie) {
6160 			rq->core_pick = NULL;
6161 			/*
6162 			 * For robustness, update the min_vruntime_fi for
6163 			 * unconstrained picks as well.
6164 			 */
6165 			WARN_ON_ONCE(fi_before);
6166 			task_vruntime_update(rq, next, false);
6167 			goto out_set_next;
6168 		}
6169 	}
6170 
6171 	/*
6172 	 * For each thread: do the regular task pick and find the max prio task
6173 	 * amongst them.
6174 	 *
6175 	 * Tie-break prio towards the current CPU
6176 	 */
6177 	for_each_cpu_wrap(i, smt_mask, cpu) {
6178 		rq_i = cpu_rq(i);
6179 
6180 		/*
6181 		 * Current cpu always has its clock updated on entrance to
6182 		 * pick_next_task(). If the current cpu is not the core,
6183 		 * the core may also have been updated above.
6184 		 */
6185 		if (i != cpu && (rq_i != rq->core || !core_clock_updated))
6186 			update_rq_clock(rq_i);
6187 
6188 		p = rq_i->core_pick = pick_task(rq_i);
6189 		if (!max || prio_less(max, p, fi_before))
6190 			max = p;
6191 	}
6192 
6193 	cookie = rq->core->core_cookie = max->core_cookie;
6194 
6195 	/*
6196 	 * For each thread: try and find a runnable task that matches @max or
6197 	 * force idle.
6198 	 */
6199 	for_each_cpu(i, smt_mask) {
6200 		rq_i = cpu_rq(i);
6201 		p = rq_i->core_pick;
6202 
6203 		if (!cookie_equals(p, cookie)) {
6204 			p = NULL;
6205 			if (cookie)
6206 				p = sched_core_find(rq_i, cookie);
6207 			if (!p)
6208 				p = idle_sched_class.pick_task(rq_i);
6209 		}
6210 
6211 		rq_i->core_pick = p;
6212 
6213 		if (p == rq_i->idle) {
6214 			if (rq_i->nr_running) {
6215 				rq->core->core_forceidle_count++;
6216 				if (!fi_before)
6217 					rq->core->core_forceidle_seq++;
6218 			}
6219 		} else {
6220 			occ++;
6221 		}
6222 	}
6223 
6224 	if (schedstat_enabled() && rq->core->core_forceidle_count) {
6225 		rq->core->core_forceidle_start = rq_clock(rq->core);
6226 		rq->core->core_forceidle_occupation = occ;
6227 	}
6228 
6229 	rq->core->core_pick_seq = rq->core->core_task_seq;
6230 	next = rq->core_pick;
6231 	rq->core_sched_seq = rq->core->core_pick_seq;
6232 
6233 	/* Something should have been selected for current CPU */
6234 	WARN_ON_ONCE(!next);
6235 
6236 	/*
6237 	 * Reschedule siblings
6238 	 *
6239 	 * NOTE: L1TF -- at this point we're no longer running the old task and
6240 	 * sending an IPI (below) ensures the sibling will no longer be running
6241 	 * their task. This ensures there is no inter-sibling overlap between
6242 	 * non-matching user state.
6243 	 */
6244 	for_each_cpu(i, smt_mask) {
6245 		rq_i = cpu_rq(i);
6246 
6247 		/*
6248 		 * An online sibling might have gone offline before a task
6249 		 * could be picked for it, or it might be offline but later
6250 		 * happen to come online, but its too late and nothing was
6251 		 * picked for it.  That's Ok - it will pick tasks for itself,
6252 		 * so ignore it.
6253 		 */
6254 		if (!rq_i->core_pick)
6255 			continue;
6256 
6257 		/*
6258 		 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6259 		 * fi_before     fi      update?
6260 		 *  0            0       1
6261 		 *  0            1       1
6262 		 *  1            0       1
6263 		 *  1            1       0
6264 		 */
6265 		if (!(fi_before && rq->core->core_forceidle_count))
6266 			task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6267 
6268 		rq_i->core_pick->core_occupation = occ;
6269 
6270 		if (i == cpu) {
6271 			rq_i->core_pick = NULL;
6272 			continue;
6273 		}
6274 
6275 		/* Did we break L1TF mitigation requirements? */
6276 		WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6277 
6278 		if (rq_i->curr == rq_i->core_pick) {
6279 			rq_i->core_pick = NULL;
6280 			continue;
6281 		}
6282 
6283 		resched_curr(rq_i);
6284 	}
6285 
6286 out_set_next:
6287 	set_next_task(rq, next);
6288 out:
6289 	if (rq->core->core_forceidle_count && next == rq->idle)
6290 		queue_core_balance(rq);
6291 
6292 	return next;
6293 }
6294 
try_steal_cookie(int this,int that)6295 static bool try_steal_cookie(int this, int that)
6296 {
6297 	struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6298 	struct task_struct *p;
6299 	unsigned long cookie;
6300 	bool success = false;
6301 
6302 	guard(irq)();
6303 	guard(double_rq_lock)(dst, src);
6304 
6305 	cookie = dst->core->core_cookie;
6306 	if (!cookie)
6307 		return false;
6308 
6309 	if (dst->curr != dst->idle)
6310 		return false;
6311 
6312 	p = sched_core_find(src, cookie);
6313 	if (!p)
6314 		return false;
6315 
6316 	do {
6317 		if (p == src->core_pick || p == src->curr)
6318 			goto next;
6319 
6320 		if (!is_cpu_allowed(p, this))
6321 			goto next;
6322 
6323 		if (p->core_occupation > dst->idle->core_occupation)
6324 			goto next;
6325 		/*
6326 		 * sched_core_find() and sched_core_next() will ensure
6327 		 * that task @p is not throttled now, we also need to
6328 		 * check whether the runqueue of the destination CPU is
6329 		 * being throttled.
6330 		 */
6331 		if (sched_task_is_throttled(p, this))
6332 			goto next;
6333 
6334 		deactivate_task(src, p, 0);
6335 		set_task_cpu(p, this);
6336 		activate_task(dst, p, 0);
6337 
6338 		resched_curr(dst);
6339 
6340 		success = true;
6341 		break;
6342 
6343 next:
6344 		p = sched_core_next(p, cookie);
6345 	} while (p);
6346 
6347 	return success;
6348 }
6349 
steal_cookie_task(int cpu,struct sched_domain * sd)6350 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6351 {
6352 	int i;
6353 
6354 	for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
6355 		if (i == cpu)
6356 			continue;
6357 
6358 		if (need_resched())
6359 			break;
6360 
6361 		if (try_steal_cookie(cpu, i))
6362 			return true;
6363 	}
6364 
6365 	return false;
6366 }
6367 
sched_core_balance(struct rq * rq)6368 static void sched_core_balance(struct rq *rq)
6369 {
6370 	struct sched_domain *sd;
6371 	int cpu = cpu_of(rq);
6372 
6373 	preempt_disable();
6374 	rcu_read_lock();
6375 	raw_spin_rq_unlock_irq(rq);
6376 	for_each_domain(cpu, sd) {
6377 		if (need_resched())
6378 			break;
6379 
6380 		if (steal_cookie_task(cpu, sd))
6381 			break;
6382 	}
6383 	raw_spin_rq_lock_irq(rq);
6384 	rcu_read_unlock();
6385 	preempt_enable();
6386 }
6387 
6388 static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
6389 
queue_core_balance(struct rq * rq)6390 static void queue_core_balance(struct rq *rq)
6391 {
6392 	if (!sched_core_enabled(rq))
6393 		return;
6394 
6395 	if (!rq->core->core_cookie)
6396 		return;
6397 
6398 	if (!rq->nr_running) /* not forced idle */
6399 		return;
6400 
6401 	queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6402 }
6403 
6404 DEFINE_LOCK_GUARD_1(core_lock, int,
6405 		    sched_core_lock(*_T->lock, &_T->flags),
6406 		    sched_core_unlock(*_T->lock, &_T->flags),
6407 		    unsigned long flags)
6408 
sched_core_cpu_starting(unsigned int cpu)6409 static void sched_core_cpu_starting(unsigned int cpu)
6410 {
6411 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6412 	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6413 	int t;
6414 
6415 	guard(core_lock)(&cpu);
6416 
6417 	WARN_ON_ONCE(rq->core != rq);
6418 
6419 	/* if we're the first, we'll be our own leader */
6420 	if (cpumask_weight(smt_mask) == 1)
6421 		return;
6422 
6423 	/* find the leader */
6424 	for_each_cpu(t, smt_mask) {
6425 		if (t == cpu)
6426 			continue;
6427 		rq = cpu_rq(t);
6428 		if (rq->core == rq) {
6429 			core_rq = rq;
6430 			break;
6431 		}
6432 	}
6433 
6434 	if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6435 		return;
6436 
6437 	/* install and validate core_rq */
6438 	for_each_cpu(t, smt_mask) {
6439 		rq = cpu_rq(t);
6440 
6441 		if (t == cpu)
6442 			rq->core = core_rq;
6443 
6444 		WARN_ON_ONCE(rq->core != core_rq);
6445 	}
6446 }
6447 
sched_core_cpu_deactivate(unsigned int cpu)6448 static void sched_core_cpu_deactivate(unsigned int cpu)
6449 {
6450 	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6451 	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6452 	int t;
6453 
6454 	guard(core_lock)(&cpu);
6455 
6456 	/* if we're the last man standing, nothing to do */
6457 	if (cpumask_weight(smt_mask) == 1) {
6458 		WARN_ON_ONCE(rq->core != rq);
6459 		return;
6460 	}
6461 
6462 	/* if we're not the leader, nothing to do */
6463 	if (rq->core != rq)
6464 		return;
6465 
6466 	/* find a new leader */
6467 	for_each_cpu(t, smt_mask) {
6468 		if (t == cpu)
6469 			continue;
6470 		core_rq = cpu_rq(t);
6471 		break;
6472 	}
6473 
6474 	if (WARN_ON_ONCE(!core_rq)) /* impossible */
6475 		return;
6476 
6477 	/* copy the shared state to the new leader */
6478 	core_rq->core_task_seq             = rq->core_task_seq;
6479 	core_rq->core_pick_seq             = rq->core_pick_seq;
6480 	core_rq->core_cookie               = rq->core_cookie;
6481 	core_rq->core_forceidle_count      = rq->core_forceidle_count;
6482 	core_rq->core_forceidle_seq        = rq->core_forceidle_seq;
6483 	core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6484 
6485 	/*
6486 	 * Accounting edge for forced idle is handled in pick_next_task().
6487 	 * Don't need another one here, since the hotplug thread shouldn't
6488 	 * have a cookie.
6489 	 */
6490 	core_rq->core_forceidle_start = 0;
6491 
6492 	/* install new leader */
6493 	for_each_cpu(t, smt_mask) {
6494 		rq = cpu_rq(t);
6495 		rq->core = core_rq;
6496 	}
6497 }
6498 
sched_core_cpu_dying(unsigned int cpu)6499 static inline void sched_core_cpu_dying(unsigned int cpu)
6500 {
6501 	struct rq *rq = cpu_rq(cpu);
6502 
6503 	if (rq->core != rq)
6504 		rq->core = rq;
6505 }
6506 
6507 #else /* !CONFIG_SCHED_CORE */
6508 
sched_core_cpu_starting(unsigned int cpu)6509 static inline void sched_core_cpu_starting(unsigned int cpu) {}
sched_core_cpu_deactivate(unsigned int cpu)6510 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
sched_core_cpu_dying(unsigned int cpu)6511 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6512 
6513 static struct task_struct *
pick_next_task(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)6514 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6515 {
6516 	return __pick_next_task(rq, prev, rf);
6517 }
6518 
6519 #endif /* CONFIG_SCHED_CORE */
6520 
6521 /*
6522  * Constants for the sched_mode argument of __schedule().
6523  *
6524  * The mode argument allows RT enabled kernels to differentiate a
6525  * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6526  * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6527  * optimize the AND operation out and just check for zero.
6528  */
6529 #define SM_NONE			0x0
6530 #define SM_PREEMPT		0x1
6531 #define SM_RTLOCK_WAIT		0x2
6532 
6533 #ifndef CONFIG_PREEMPT_RT
6534 # define SM_MASK_PREEMPT	(~0U)
6535 #else
6536 # define SM_MASK_PREEMPT	SM_PREEMPT
6537 #endif
6538 
6539 /*
6540  * __schedule() is the main scheduler function.
6541  *
6542  * The main means of driving the scheduler and thus entering this function are:
6543  *
6544  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6545  *
6546  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6547  *      paths. For example, see arch/x86/entry_64.S.
6548  *
6549  *      To drive preemption between tasks, the scheduler sets the flag in timer
6550  *      interrupt handler scheduler_tick().
6551  *
6552  *   3. Wakeups don't really cause entry into schedule(). They add a
6553  *      task to the run-queue and that's it.
6554  *
6555  *      Now, if the new task added to the run-queue preempts the current
6556  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6557  *      called on the nearest possible occasion:
6558  *
6559  *       - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6560  *
6561  *         - in syscall or exception context, at the next outmost
6562  *           preempt_enable(). (this might be as soon as the wake_up()'s
6563  *           spin_unlock()!)
6564  *
6565  *         - in IRQ context, return from interrupt-handler to
6566  *           preemptible context
6567  *
6568  *       - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6569  *         then at the next:
6570  *
6571  *          - cond_resched() call
6572  *          - explicit schedule() call
6573  *          - return from syscall or exception to user-space
6574  *          - return from interrupt-handler to user-space
6575  *
6576  * WARNING: must be called with preemption disabled!
6577  */
__schedule(unsigned int sched_mode)6578 static void __sched notrace __schedule(unsigned int sched_mode)
6579 {
6580 	struct task_struct *prev, *next;
6581 	unsigned long *switch_count;
6582 	unsigned long prev_state;
6583 	struct rq_flags rf;
6584 	struct rq *rq;
6585 	int cpu;
6586 
6587 	cpu = smp_processor_id();
6588 	rq = cpu_rq(cpu);
6589 	prev = rq->curr;
6590 
6591 	schedule_debug(prev, !!sched_mode);
6592 
6593 	if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6594 		hrtick_clear(rq);
6595 
6596 	local_irq_disable();
6597 	rcu_note_context_switch(!!sched_mode);
6598 
6599 	/*
6600 	 * Make sure that signal_pending_state()->signal_pending() below
6601 	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6602 	 * done by the caller to avoid the race with signal_wake_up():
6603 	 *
6604 	 * __set_current_state(@state)		signal_wake_up()
6605 	 * schedule()				  set_tsk_thread_flag(p, TIF_SIGPENDING)
6606 	 *					  wake_up_state(p, state)
6607 	 *   LOCK rq->lock			    LOCK p->pi_state
6608 	 *   smp_mb__after_spinlock()		    smp_mb__after_spinlock()
6609 	 *     if (signal_pending_state())	    if (p->state & @state)
6610 	 *
6611 	 * Also, the membarrier system call requires a full memory barrier
6612 	 * after coming from user-space, before storing to rq->curr.
6613 	 */
6614 	rq_lock(rq, &rf);
6615 	smp_mb__after_spinlock();
6616 
6617 	/* Promote REQ to ACT */
6618 	rq->clock_update_flags <<= 1;
6619 	update_rq_clock(rq);
6620 	rq->clock_update_flags = RQCF_UPDATED;
6621 
6622 	switch_count = &prev->nivcsw;
6623 
6624 	/*
6625 	 * We must load prev->state once (task_struct::state is volatile), such
6626 	 * that we form a control dependency vs deactivate_task() below.
6627 	 */
6628 	prev_state = READ_ONCE(prev->__state);
6629 	if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6630 		if (signal_pending_state(prev_state, prev)) {
6631 			WRITE_ONCE(prev->__state, TASK_RUNNING);
6632 		} else {
6633 			prev->sched_contributes_to_load =
6634 				(prev_state & TASK_UNINTERRUPTIBLE) &&
6635 				!(prev_state & TASK_NOLOAD) &&
6636 				!(prev_state & TASK_FROZEN);
6637 
6638 			if (prev->sched_contributes_to_load)
6639 				rq->nr_uninterruptible++;
6640 
6641 			/*
6642 			 * __schedule()			ttwu()
6643 			 *   prev_state = prev->state;    if (p->on_rq && ...)
6644 			 *   if (prev_state)		    goto out;
6645 			 *     p->on_rq = 0;		  smp_acquire__after_ctrl_dep();
6646 			 *				  p->state = TASK_WAKING
6647 			 *
6648 			 * Where __schedule() and ttwu() have matching control dependencies.
6649 			 *
6650 			 * After this, schedule() must not care about p->state any more.
6651 			 */
6652 			deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6653 
6654 			if (prev->in_iowait) {
6655 				atomic_inc(&rq->nr_iowait);
6656 				delayacct_blkio_start();
6657 			}
6658 		}
6659 		switch_count = &prev->nvcsw;
6660 	}
6661 
6662 	next = pick_next_task(rq, prev, &rf);
6663 	clear_tsk_need_resched(prev);
6664 	clear_preempt_need_resched();
6665 #ifdef CONFIG_SCHED_DEBUG
6666 	rq->last_seen_need_resched_ns = 0;
6667 #endif
6668 
6669 	if (likely(prev != next)) {
6670 		rq->nr_switches++;
6671 		/*
6672 		 * RCU users of rcu_dereference(rq->curr) may not see
6673 		 * changes to task_struct made by pick_next_task().
6674 		 */
6675 		RCU_INIT_POINTER(rq->curr, next);
6676 		/*
6677 		 * The membarrier system call requires each architecture
6678 		 * to have a full memory barrier after updating
6679 		 * rq->curr, before returning to user-space.
6680 		 *
6681 		 * Here are the schemes providing that barrier on the
6682 		 * various architectures:
6683 		 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6684 		 *   switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6685 		 * - finish_lock_switch() for weakly-ordered
6686 		 *   architectures where spin_unlock is a full barrier,
6687 		 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6688 		 *   is a RELEASE barrier),
6689 		 */
6690 		++*switch_count;
6691 
6692 		migrate_disable_switch(rq, prev);
6693 		psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6694 
6695 		trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6696 
6697 		/* Also unlocks the rq: */
6698 		rq = context_switch(rq, prev, next, &rf);
6699 	} else {
6700 		rq_unpin_lock(rq, &rf);
6701 		__balance_callbacks(rq);
6702 		raw_spin_rq_unlock_irq(rq);
6703 	}
6704 }
6705 
do_task_dead(void)6706 void __noreturn do_task_dead(void)
6707 {
6708 	/* Causes final put_task_struct in finish_task_switch(): */
6709 	set_special_state(TASK_DEAD);
6710 
6711 	/* Tell freezer to ignore us: */
6712 	current->flags |= PF_NOFREEZE;
6713 
6714 	__schedule(SM_NONE);
6715 	BUG();
6716 
6717 	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6718 	for (;;)
6719 		cpu_relax();
6720 }
6721 
sched_submit_work(struct task_struct * tsk)6722 static inline void sched_submit_work(struct task_struct *tsk)
6723 {
6724 	unsigned int task_flags;
6725 
6726 	if (task_is_running(tsk))
6727 		return;
6728 
6729 	task_flags = tsk->flags;
6730 	/*
6731 	 * If a worker goes to sleep, notify and ask workqueue whether it
6732 	 * wants to wake up a task to maintain concurrency.
6733 	 */
6734 	if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6735 		if (task_flags & PF_WQ_WORKER)
6736 			wq_worker_sleeping(tsk);
6737 		else
6738 			io_wq_worker_sleeping(tsk);
6739 	}
6740 
6741 	/*
6742 	 * spinlock and rwlock must not flush block requests.  This will
6743 	 * deadlock if the callback attempts to acquire a lock which is
6744 	 * already acquired.
6745 	 */
6746 	SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6747 
6748 	/*
6749 	 * If we are going to sleep and we have plugged IO queued,
6750 	 * make sure to submit it to avoid deadlocks.
6751 	 */
6752 	blk_flush_plug(tsk->plug, true);
6753 }
6754 
sched_update_worker(struct task_struct * tsk)6755 static void sched_update_worker(struct task_struct *tsk)
6756 {
6757 	if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6758 		if (tsk->flags & PF_WQ_WORKER)
6759 			wq_worker_running(tsk);
6760 		else
6761 			io_wq_worker_running(tsk);
6762 	}
6763 }
6764 
schedule(void)6765 asmlinkage __visible void __sched schedule(void)
6766 {
6767 	struct task_struct *tsk = current;
6768 
6769 	sched_submit_work(tsk);
6770 	do {
6771 		preempt_disable();
6772 		__schedule(SM_NONE);
6773 		sched_preempt_enable_no_resched();
6774 	} while (need_resched());
6775 	sched_update_worker(tsk);
6776 }
6777 EXPORT_SYMBOL(schedule);
6778 
6779 /*
6780  * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6781  * state (have scheduled out non-voluntarily) by making sure that all
6782  * tasks have either left the run queue or have gone into user space.
6783  * As idle tasks do not do either, they must not ever be preempted
6784  * (schedule out non-voluntarily).
6785  *
6786  * schedule_idle() is similar to schedule_preempt_disable() except that it
6787  * never enables preemption because it does not call sched_submit_work().
6788  */
schedule_idle(void)6789 void __sched schedule_idle(void)
6790 {
6791 	/*
6792 	 * As this skips calling sched_submit_work(), which the idle task does
6793 	 * regardless because that function is a nop when the task is in a
6794 	 * TASK_RUNNING state, make sure this isn't used someplace that the
6795 	 * current task can be in any other state. Note, idle is always in the
6796 	 * TASK_RUNNING state.
6797 	 */
6798 	WARN_ON_ONCE(current->__state);
6799 	do {
6800 		__schedule(SM_NONE);
6801 	} while (need_resched());
6802 }
6803 
6804 #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
schedule_user(void)6805 asmlinkage __visible void __sched schedule_user(void)
6806 {
6807 	/*
6808 	 * If we come here after a random call to set_need_resched(),
6809 	 * or we have been woken up remotely but the IPI has not yet arrived,
6810 	 * we haven't yet exited the RCU idle mode. Do it here manually until
6811 	 * we find a better solution.
6812 	 *
6813 	 * NB: There are buggy callers of this function.  Ideally we
6814 	 * should warn if prev_state != CONTEXT_USER, but that will trigger
6815 	 * too frequently to make sense yet.
6816 	 */
6817 	enum ctx_state prev_state = exception_enter();
6818 	schedule();
6819 	exception_exit(prev_state);
6820 }
6821 #endif
6822 
6823 /**
6824  * schedule_preempt_disabled - called with preemption disabled
6825  *
6826  * Returns with preemption disabled. Note: preempt_count must be 1
6827  */
schedule_preempt_disabled(void)6828 void __sched schedule_preempt_disabled(void)
6829 {
6830 	sched_preempt_enable_no_resched();
6831 	schedule();
6832 	preempt_disable();
6833 }
6834 
6835 #ifdef CONFIG_PREEMPT_RT
schedule_rtlock(void)6836 void __sched notrace schedule_rtlock(void)
6837 {
6838 	do {
6839 		preempt_disable();
6840 		__schedule(SM_RTLOCK_WAIT);
6841 		sched_preempt_enable_no_resched();
6842 	} while (need_resched());
6843 }
6844 NOKPROBE_SYMBOL(schedule_rtlock);
6845 #endif
6846 
preempt_schedule_common(void)6847 static void __sched notrace preempt_schedule_common(void)
6848 {
6849 	do {
6850 		/*
6851 		 * Because the function tracer can trace preempt_count_sub()
6852 		 * and it also uses preempt_enable/disable_notrace(), if
6853 		 * NEED_RESCHED is set, the preempt_enable_notrace() called
6854 		 * by the function tracer will call this function again and
6855 		 * cause infinite recursion.
6856 		 *
6857 		 * Preemption must be disabled here before the function
6858 		 * tracer can trace. Break up preempt_disable() into two
6859 		 * calls. One to disable preemption without fear of being
6860 		 * traced. The other to still record the preemption latency,
6861 		 * which can also be traced by the function tracer.
6862 		 */
6863 		preempt_disable_notrace();
6864 		preempt_latency_start(1);
6865 		__schedule(SM_PREEMPT);
6866 		preempt_latency_stop(1);
6867 		preempt_enable_no_resched_notrace();
6868 
6869 		/*
6870 		 * Check again in case we missed a preemption opportunity
6871 		 * between schedule and now.
6872 		 */
6873 	} while (need_resched());
6874 }
6875 
6876 #ifdef CONFIG_PREEMPTION
6877 /*
6878  * This is the entry point to schedule() from in-kernel preemption
6879  * off of preempt_enable.
6880  */
preempt_schedule(void)6881 asmlinkage __visible void __sched notrace preempt_schedule(void)
6882 {
6883 	/*
6884 	 * If there is a non-zero preempt_count or interrupts are disabled,
6885 	 * we do not want to preempt the current task. Just return..
6886 	 */
6887 	if (likely(!preemptible()))
6888 		return;
6889 	preempt_schedule_common();
6890 }
6891 NOKPROBE_SYMBOL(preempt_schedule);
6892 EXPORT_SYMBOL(preempt_schedule);
6893 
6894 #ifdef CONFIG_PREEMPT_DYNAMIC
6895 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6896 #ifndef preempt_schedule_dynamic_enabled
6897 #define preempt_schedule_dynamic_enabled	preempt_schedule
6898 #define preempt_schedule_dynamic_disabled	NULL
6899 #endif
6900 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6901 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6902 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6903 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
dynamic_preempt_schedule(void)6904 void __sched notrace dynamic_preempt_schedule(void)
6905 {
6906 	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6907 		return;
6908 	preempt_schedule();
6909 }
6910 NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6911 EXPORT_SYMBOL(dynamic_preempt_schedule);
6912 #endif
6913 #endif
6914 
6915 /**
6916  * preempt_schedule_notrace - preempt_schedule called by tracing
6917  *
6918  * The tracing infrastructure uses preempt_enable_notrace to prevent
6919  * recursion and tracing preempt enabling caused by the tracing
6920  * infrastructure itself. But as tracing can happen in areas coming
6921  * from userspace or just about to enter userspace, a preempt enable
6922  * can occur before user_exit() is called. This will cause the scheduler
6923  * to be called when the system is still in usermode.
6924  *
6925  * To prevent this, the preempt_enable_notrace will use this function
6926  * instead of preempt_schedule() to exit user context if needed before
6927  * calling the scheduler.
6928  */
preempt_schedule_notrace(void)6929 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6930 {
6931 	enum ctx_state prev_ctx;
6932 
6933 	if (likely(!preemptible()))
6934 		return;
6935 
6936 	do {
6937 		/*
6938 		 * Because the function tracer can trace preempt_count_sub()
6939 		 * and it also uses preempt_enable/disable_notrace(), if
6940 		 * NEED_RESCHED is set, the preempt_enable_notrace() called
6941 		 * by the function tracer will call this function again and
6942 		 * cause infinite recursion.
6943 		 *
6944 		 * Preemption must be disabled here before the function
6945 		 * tracer can trace. Break up preempt_disable() into two
6946 		 * calls. One to disable preemption without fear of being
6947 		 * traced. The other to still record the preemption latency,
6948 		 * which can also be traced by the function tracer.
6949 		 */
6950 		preempt_disable_notrace();
6951 		preempt_latency_start(1);
6952 		/*
6953 		 * Needs preempt disabled in case user_exit() is traced
6954 		 * and the tracer calls preempt_enable_notrace() causing
6955 		 * an infinite recursion.
6956 		 */
6957 		prev_ctx = exception_enter();
6958 		__schedule(SM_PREEMPT);
6959 		exception_exit(prev_ctx);
6960 
6961 		preempt_latency_stop(1);
6962 		preempt_enable_no_resched_notrace();
6963 	} while (need_resched());
6964 }
6965 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6966 
6967 #ifdef CONFIG_PREEMPT_DYNAMIC
6968 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6969 #ifndef preempt_schedule_notrace_dynamic_enabled
6970 #define preempt_schedule_notrace_dynamic_enabled	preempt_schedule_notrace
6971 #define preempt_schedule_notrace_dynamic_disabled	NULL
6972 #endif
6973 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6974 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6975 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6976 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
dynamic_preempt_schedule_notrace(void)6977 void __sched notrace dynamic_preempt_schedule_notrace(void)
6978 {
6979 	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6980 		return;
6981 	preempt_schedule_notrace();
6982 }
6983 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6984 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6985 #endif
6986 #endif
6987 
6988 #endif /* CONFIG_PREEMPTION */
6989 
6990 /*
6991  * This is the entry point to schedule() from kernel preemption
6992  * off of irq context.
6993  * Note, that this is called and return with irqs disabled. This will
6994  * protect us against recursive calling from irq.
6995  */
preempt_schedule_irq(void)6996 asmlinkage __visible void __sched preempt_schedule_irq(void)
6997 {
6998 	enum ctx_state prev_state;
6999 
7000 	/* Catch callers which need to be fixed */
7001 	BUG_ON(preempt_count() || !irqs_disabled());
7002 
7003 	prev_state = exception_enter();
7004 
7005 	do {
7006 		preempt_disable();
7007 		local_irq_enable();
7008 		__schedule(SM_PREEMPT);
7009 		local_irq_disable();
7010 		sched_preempt_enable_no_resched();
7011 	} while (need_resched());
7012 
7013 	exception_exit(prev_state);
7014 }
7015 
default_wake_function(wait_queue_entry_t * curr,unsigned mode,int wake_flags,void * key)7016 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
7017 			  void *key)
7018 {
7019 	WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~(WF_SYNC|WF_CURRENT_CPU));
7020 	return try_to_wake_up(curr->private, mode, wake_flags);
7021 }
7022 EXPORT_SYMBOL(default_wake_function);
7023 
__setscheduler_prio(struct task_struct * p,int prio)7024 static void __setscheduler_prio(struct task_struct *p, int prio)
7025 {
7026 	if (dl_prio(prio))
7027 		p->sched_class = &dl_sched_class;
7028 	else if (rt_prio(prio))
7029 		p->sched_class = &rt_sched_class;
7030 	else
7031 		p->sched_class = &fair_sched_class;
7032 
7033 	p->prio = prio;
7034 }
7035 
7036 #ifdef CONFIG_RT_MUTEXES
7037 
__rt_effective_prio(struct task_struct * pi_task,int prio)7038 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
7039 {
7040 	if (pi_task)
7041 		prio = min(prio, pi_task->prio);
7042 
7043 	return prio;
7044 }
7045 
rt_effective_prio(struct task_struct * p,int prio)7046 static inline int rt_effective_prio(struct task_struct *p, int prio)
7047 {
7048 	struct task_struct *pi_task = rt_mutex_get_top_task(p);
7049 
7050 	return __rt_effective_prio(pi_task, prio);
7051 }
7052 
7053 /*
7054  * rt_mutex_setprio - set the current priority of a task
7055  * @p: task to boost
7056  * @pi_task: donor task
7057  *
7058  * This function changes the 'effective' priority of a task. It does
7059  * not touch ->normal_prio like __setscheduler().
7060  *
7061  * Used by the rt_mutex code to implement priority inheritance
7062  * logic. Call site only calls if the priority of the task changed.
7063  */
rt_mutex_setprio(struct task_struct * p,struct task_struct * pi_task)7064 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
7065 {
7066 	int prio, oldprio, queued, running, queue_flag =
7067 		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7068 	const struct sched_class *prev_class;
7069 	struct rq_flags rf;
7070 	struct rq *rq;
7071 
7072 	/* XXX used to be waiter->prio, not waiter->task->prio */
7073 	prio = __rt_effective_prio(pi_task, p->normal_prio);
7074 
7075 	/*
7076 	 * If nothing changed; bail early.
7077 	 */
7078 	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
7079 		return;
7080 
7081 	rq = __task_rq_lock(p, &rf);
7082 	update_rq_clock(rq);
7083 	/*
7084 	 * Set under pi_lock && rq->lock, such that the value can be used under
7085 	 * either lock.
7086 	 *
7087 	 * Note that there is loads of tricky to make this pointer cache work
7088 	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
7089 	 * ensure a task is de-boosted (pi_task is set to NULL) before the
7090 	 * task is allowed to run again (and can exit). This ensures the pointer
7091 	 * points to a blocked task -- which guarantees the task is present.
7092 	 */
7093 	p->pi_top_task = pi_task;
7094 
7095 	/*
7096 	 * For FIFO/RR we only need to set prio, if that matches we're done.
7097 	 */
7098 	if (prio == p->prio && !dl_prio(prio))
7099 		goto out_unlock;
7100 
7101 	/*
7102 	 * Idle task boosting is a nono in general. There is one
7103 	 * exception, when PREEMPT_RT and NOHZ is active:
7104 	 *
7105 	 * The idle task calls get_next_timer_interrupt() and holds
7106 	 * the timer wheel base->lock on the CPU and another CPU wants
7107 	 * to access the timer (probably to cancel it). We can safely
7108 	 * ignore the boosting request, as the idle CPU runs this code
7109 	 * with interrupts disabled and will complete the lock
7110 	 * protected section without being interrupted. So there is no
7111 	 * real need to boost.
7112 	 */
7113 	if (unlikely(p == rq->idle)) {
7114 		WARN_ON(p != rq->curr);
7115 		WARN_ON(p->pi_blocked_on);
7116 		goto out_unlock;
7117 	}
7118 
7119 	trace_sched_pi_setprio(p, pi_task);
7120 	oldprio = p->prio;
7121 
7122 	if (oldprio == prio)
7123 		queue_flag &= ~DEQUEUE_MOVE;
7124 
7125 	prev_class = p->sched_class;
7126 	queued = task_on_rq_queued(p);
7127 	running = task_current(rq, p);
7128 	if (queued)
7129 		dequeue_task(rq, p, queue_flag);
7130 	if (running)
7131 		put_prev_task(rq, p);
7132 
7133 	/*
7134 	 * Boosting condition are:
7135 	 * 1. -rt task is running and holds mutex A
7136 	 *      --> -dl task blocks on mutex A
7137 	 *
7138 	 * 2. -dl task is running and holds mutex A
7139 	 *      --> -dl task blocks on mutex A and could preempt the
7140 	 *          running task
7141 	 */
7142 	if (dl_prio(prio)) {
7143 		if (!dl_prio(p->normal_prio) ||
7144 		    (pi_task && dl_prio(pi_task->prio) &&
7145 		     dl_entity_preempt(&pi_task->dl, &p->dl))) {
7146 			p->dl.pi_se = pi_task->dl.pi_se;
7147 			queue_flag |= ENQUEUE_REPLENISH;
7148 		} else {
7149 			p->dl.pi_se = &p->dl;
7150 		}
7151 	} else if (rt_prio(prio)) {
7152 		if (dl_prio(oldprio))
7153 			p->dl.pi_se = &p->dl;
7154 		if (oldprio < prio)
7155 			queue_flag |= ENQUEUE_HEAD;
7156 	} else {
7157 		if (dl_prio(oldprio))
7158 			p->dl.pi_se = &p->dl;
7159 		if (rt_prio(oldprio))
7160 			p->rt.timeout = 0;
7161 	}
7162 
7163 	__setscheduler_prio(p, prio);
7164 
7165 	if (queued)
7166 		enqueue_task(rq, p, queue_flag);
7167 	if (running)
7168 		set_next_task(rq, p);
7169 
7170 	check_class_changed(rq, p, prev_class, oldprio);
7171 out_unlock:
7172 	/* Avoid rq from going away on us: */
7173 	preempt_disable();
7174 
7175 	rq_unpin_lock(rq, &rf);
7176 	__balance_callbacks(rq);
7177 	raw_spin_rq_unlock(rq);
7178 
7179 	preempt_enable();
7180 }
7181 #else
rt_effective_prio(struct task_struct * p,int prio)7182 static inline int rt_effective_prio(struct task_struct *p, int prio)
7183 {
7184 	return prio;
7185 }
7186 #endif
7187 
set_user_nice(struct task_struct * p,long nice)7188 void set_user_nice(struct task_struct *p, long nice)
7189 {
7190 	bool queued, running;
7191 	int old_prio;
7192 	struct rq_flags rf;
7193 	struct rq *rq;
7194 
7195 	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
7196 		return;
7197 	/*
7198 	 * We have to be careful, if called from sys_setpriority(),
7199 	 * the task might be in the middle of scheduling on another CPU.
7200 	 */
7201 	rq = task_rq_lock(p, &rf);
7202 	update_rq_clock(rq);
7203 
7204 	/*
7205 	 * The RT priorities are set via sched_setscheduler(), but we still
7206 	 * allow the 'normal' nice value to be set - but as expected
7207 	 * it won't have any effect on scheduling until the task is
7208 	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7209 	 */
7210 	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7211 		p->static_prio = NICE_TO_PRIO(nice);
7212 		goto out_unlock;
7213 	}
7214 	queued = task_on_rq_queued(p);
7215 	running = task_current(rq, p);
7216 	if (queued)
7217 		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7218 	if (running)
7219 		put_prev_task(rq, p);
7220 
7221 	p->static_prio = NICE_TO_PRIO(nice);
7222 	set_load_weight(p, true);
7223 	old_prio = p->prio;
7224 	p->prio = effective_prio(p);
7225 
7226 	if (queued)
7227 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7228 	if (running)
7229 		set_next_task(rq, p);
7230 
7231 	/*
7232 	 * If the task increased its priority or is running and
7233 	 * lowered its priority, then reschedule its CPU:
7234 	 */
7235 	p->sched_class->prio_changed(rq, p, old_prio);
7236 
7237 out_unlock:
7238 	task_rq_unlock(rq, p, &rf);
7239 }
7240 EXPORT_SYMBOL(set_user_nice);
7241 
7242 /*
7243  * is_nice_reduction - check if nice value is an actual reduction
7244  *
7245  * Similar to can_nice() but does not perform a capability check.
7246  *
7247  * @p: task
7248  * @nice: nice value
7249  */
is_nice_reduction(const struct task_struct * p,const int nice)7250 static bool is_nice_reduction(const struct task_struct *p, const int nice)
7251 {
7252 	/* Convert nice value [19,-20] to rlimit style value [1,40]: */
7253 	int nice_rlim = nice_to_rlimit(nice);
7254 
7255 	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7256 }
7257 
7258 /*
7259  * can_nice - check if a task can reduce its nice value
7260  * @p: task
7261  * @nice: nice value
7262  */
can_nice(const struct task_struct * p,const int nice)7263 int can_nice(const struct task_struct *p, const int nice)
7264 {
7265 	return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7266 }
7267 
7268 #ifdef __ARCH_WANT_SYS_NICE
7269 
7270 /*
7271  * sys_nice - change the priority of the current process.
7272  * @increment: priority increment
7273  *
7274  * sys_setpriority is a more generic, but much slower function that
7275  * does similar things.
7276  */
SYSCALL_DEFINE1(nice,int,increment)7277 SYSCALL_DEFINE1(nice, int, increment)
7278 {
7279 	long nice, retval;
7280 
7281 	/*
7282 	 * Setpriority might change our priority at the same moment.
7283 	 * We don't have to worry. Conceptually one call occurs first
7284 	 * and we have a single winner.
7285 	 */
7286 	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7287 	nice = task_nice(current) + increment;
7288 
7289 	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7290 	if (increment < 0 && !can_nice(current, nice))
7291 		return -EPERM;
7292 
7293 	retval = security_task_setnice(current, nice);
7294 	if (retval)
7295 		return retval;
7296 
7297 	set_user_nice(current, nice);
7298 	return 0;
7299 }
7300 
7301 #endif
7302 
7303 /**
7304  * task_prio - return the priority value of a given task.
7305  * @p: the task in question.
7306  *
7307  * Return: The priority value as seen by users in /proc.
7308  *
7309  * sched policy         return value   kernel prio    user prio/nice
7310  *
7311  * normal, batch, idle     [0 ... 39]  [100 ... 139]          0/[-20 ... 19]
7312  * fifo, rr             [-2 ... -100]     [98 ... 0]  [1 ... 99]
7313  * deadline                     -101             -1           0
7314  */
task_prio(const struct task_struct * p)7315 int task_prio(const struct task_struct *p)
7316 {
7317 	return p->prio - MAX_RT_PRIO;
7318 }
7319 
7320 /**
7321  * idle_cpu - is a given CPU idle currently?
7322  * @cpu: the processor in question.
7323  *
7324  * Return: 1 if the CPU is currently idle. 0 otherwise.
7325  */
idle_cpu(int cpu)7326 int idle_cpu(int cpu)
7327 {
7328 	struct rq *rq = cpu_rq(cpu);
7329 
7330 	if (rq->curr != rq->idle)
7331 		return 0;
7332 
7333 	if (rq->nr_running)
7334 		return 0;
7335 
7336 #ifdef CONFIG_SMP
7337 	if (rq->ttwu_pending)
7338 		return 0;
7339 #endif
7340 
7341 	return 1;
7342 }
7343 
7344 /**
7345  * available_idle_cpu - is a given CPU idle for enqueuing work.
7346  * @cpu: the CPU in question.
7347  *
7348  * Return: 1 if the CPU is currently idle. 0 otherwise.
7349  */
available_idle_cpu(int cpu)7350 int available_idle_cpu(int cpu)
7351 {
7352 	if (!idle_cpu(cpu))
7353 		return 0;
7354 
7355 	if (vcpu_is_preempted(cpu))
7356 		return 0;
7357 
7358 	return 1;
7359 }
7360 
7361 /**
7362  * idle_task - return the idle task for a given CPU.
7363  * @cpu: the processor in question.
7364  *
7365  * Return: The idle task for the CPU @cpu.
7366  */
idle_task(int cpu)7367 struct task_struct *idle_task(int cpu)
7368 {
7369 	return cpu_rq(cpu)->idle;
7370 }
7371 
7372 #ifdef CONFIG_SCHED_CORE
sched_core_idle_cpu(int cpu)7373 int sched_core_idle_cpu(int cpu)
7374 {
7375 	struct rq *rq = cpu_rq(cpu);
7376 
7377 	if (sched_core_enabled(rq) && rq->curr == rq->idle)
7378 		return 1;
7379 
7380 	return idle_cpu(cpu);
7381 }
7382 
7383 #endif
7384 
7385 #ifdef CONFIG_SMP
7386 /*
7387  * This function computes an effective utilization for the given CPU, to be
7388  * used for frequency selection given the linear relation: f = u * f_max.
7389  *
7390  * The scheduler tracks the following metrics:
7391  *
7392  *   cpu_util_{cfs,rt,dl,irq}()
7393  *   cpu_bw_dl()
7394  *
7395  * Where the cfs,rt and dl util numbers are tracked with the same metric and
7396  * synchronized windows and are thus directly comparable.
7397  *
7398  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7399  * which excludes things like IRQ and steal-time. These latter are then accrued
7400  * in the irq utilization.
7401  *
7402  * The DL bandwidth number otoh is not a measured metric but a value computed
7403  * based on the task model parameters and gives the minimal utilization
7404  * required to meet deadlines.
7405  */
effective_cpu_util(int cpu,unsigned long util_cfs,enum cpu_util_type type,struct task_struct * p)7406 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7407 				 enum cpu_util_type type,
7408 				 struct task_struct *p)
7409 {
7410 	unsigned long dl_util, util, irq, max;
7411 	struct rq *rq = cpu_rq(cpu);
7412 
7413 	max = arch_scale_cpu_capacity(cpu);
7414 
7415 	if (!uclamp_is_used() &&
7416 	    type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7417 		return max;
7418 	}
7419 
7420 	/*
7421 	 * Early check to see if IRQ/steal time saturates the CPU, can be
7422 	 * because of inaccuracies in how we track these -- see
7423 	 * update_irq_load_avg().
7424 	 */
7425 	irq = cpu_util_irq(rq);
7426 	if (unlikely(irq >= max))
7427 		return max;
7428 
7429 	/*
7430 	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7431 	 * CFS tasks and we use the same metric to track the effective
7432 	 * utilization (PELT windows are synchronized) we can directly add them
7433 	 * to obtain the CPU's actual utilization.
7434 	 *
7435 	 * CFS and RT utilization can be boosted or capped, depending on
7436 	 * utilization clamp constraints requested by currently RUNNABLE
7437 	 * tasks.
7438 	 * When there are no CFS RUNNABLE tasks, clamps are released and
7439 	 * frequency will be gracefully reduced with the utilization decay.
7440 	 */
7441 	util = util_cfs + cpu_util_rt(rq);
7442 	if (type == FREQUENCY_UTIL)
7443 		util = uclamp_rq_util_with(rq, util, p);
7444 
7445 	dl_util = cpu_util_dl(rq);
7446 
7447 	/*
7448 	 * For frequency selection we do not make cpu_util_dl() a permanent part
7449 	 * of this sum because we want to use cpu_bw_dl() later on, but we need
7450 	 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7451 	 * that we select f_max when there is no idle time.
7452 	 *
7453 	 * NOTE: numerical errors or stop class might cause us to not quite hit
7454 	 * saturation when we should -- something for later.
7455 	 */
7456 	if (util + dl_util >= max)
7457 		return max;
7458 
7459 	/*
7460 	 * OTOH, for energy computation we need the estimated running time, so
7461 	 * include util_dl and ignore dl_bw.
7462 	 */
7463 	if (type == ENERGY_UTIL)
7464 		util += dl_util;
7465 
7466 	/*
7467 	 * There is still idle time; further improve the number by using the
7468 	 * irq metric. Because IRQ/steal time is hidden from the task clock we
7469 	 * need to scale the task numbers:
7470 	 *
7471 	 *              max - irq
7472 	 *   U' = irq + --------- * U
7473 	 *                 max
7474 	 */
7475 	util = scale_irq_capacity(util, irq, max);
7476 	util += irq;
7477 
7478 	/*
7479 	 * Bandwidth required by DEADLINE must always be granted while, for
7480 	 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7481 	 * to gracefully reduce the frequency when no tasks show up for longer
7482 	 * periods of time.
7483 	 *
7484 	 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7485 	 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7486 	 * an interface. So, we only do the latter for now.
7487 	 */
7488 	if (type == FREQUENCY_UTIL)
7489 		util += cpu_bw_dl(rq);
7490 
7491 	return min(max, util);
7492 }
7493 
sched_cpu_util(int cpu)7494 unsigned long sched_cpu_util(int cpu)
7495 {
7496 	return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
7497 }
7498 #endif /* CONFIG_SMP */
7499 
7500 /**
7501  * find_process_by_pid - find a process with a matching PID value.
7502  * @pid: the pid in question.
7503  *
7504  * The task of @pid, if found. %NULL otherwise.
7505  */
find_process_by_pid(pid_t pid)7506 static struct task_struct *find_process_by_pid(pid_t pid)
7507 {
7508 	return pid ? find_task_by_vpid(pid) : current;
7509 }
7510 
7511 /*
7512  * sched_setparam() passes in -1 for its policy, to let the functions
7513  * it calls know not to change it.
7514  */
7515 #define SETPARAM_POLICY	-1
7516 
__setscheduler_params(struct task_struct * p,const struct sched_attr * attr)7517 static void __setscheduler_params(struct task_struct *p,
7518 		const struct sched_attr *attr)
7519 {
7520 	int policy = attr->sched_policy;
7521 
7522 	if (policy == SETPARAM_POLICY)
7523 		policy = p->policy;
7524 
7525 	p->policy = policy;
7526 
7527 	if (dl_policy(policy))
7528 		__setparam_dl(p, attr);
7529 	else if (fair_policy(policy))
7530 		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7531 
7532 	/*
7533 	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7534 	 * !rt_policy. Always setting this ensures that things like
7535 	 * getparam()/getattr() don't report silly values for !rt tasks.
7536 	 */
7537 	p->rt_priority = attr->sched_priority;
7538 	p->normal_prio = normal_prio(p);
7539 	set_load_weight(p, true);
7540 }
7541 
7542 /*
7543  * Check the target process has a UID that matches the current process's:
7544  */
check_same_owner(struct task_struct * p)7545 static bool check_same_owner(struct task_struct *p)
7546 {
7547 	const struct cred *cred = current_cred(), *pcred;
7548 	bool match;
7549 
7550 	rcu_read_lock();
7551 	pcred = __task_cred(p);
7552 	match = (uid_eq(cred->euid, pcred->euid) ||
7553 		 uid_eq(cred->euid, pcred->uid));
7554 	rcu_read_unlock();
7555 	return match;
7556 }
7557 
7558 /*
7559  * Allow unprivileged RT tasks to decrease priority.
7560  * Only issue a capable test if needed and only once to avoid an audit
7561  * event on permitted non-privileged operations:
7562  */
user_check_sched_setscheduler(struct task_struct * p,const struct sched_attr * attr,int policy,int reset_on_fork)7563 static int user_check_sched_setscheduler(struct task_struct *p,
7564 					 const struct sched_attr *attr,
7565 					 int policy, int reset_on_fork)
7566 {
7567 	if (fair_policy(policy)) {
7568 		if (attr->sched_nice < task_nice(p) &&
7569 		    !is_nice_reduction(p, attr->sched_nice))
7570 			goto req_priv;
7571 	}
7572 
7573 	if (rt_policy(policy)) {
7574 		unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7575 
7576 		/* Can't set/change the rt policy: */
7577 		if (policy != p->policy && !rlim_rtprio)
7578 			goto req_priv;
7579 
7580 		/* Can't increase priority: */
7581 		if (attr->sched_priority > p->rt_priority &&
7582 		    attr->sched_priority > rlim_rtprio)
7583 			goto req_priv;
7584 	}
7585 
7586 	/*
7587 	 * Can't set/change SCHED_DEADLINE policy at all for now
7588 	 * (safest behavior); in the future we would like to allow
7589 	 * unprivileged DL tasks to increase their relative deadline
7590 	 * or reduce their runtime (both ways reducing utilization)
7591 	 */
7592 	if (dl_policy(policy))
7593 		goto req_priv;
7594 
7595 	/*
7596 	 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7597 	 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7598 	 */
7599 	if (task_has_idle_policy(p) && !idle_policy(policy)) {
7600 		if (!is_nice_reduction(p, task_nice(p)))
7601 			goto req_priv;
7602 	}
7603 
7604 	/* Can't change other user's priorities: */
7605 	if (!check_same_owner(p))
7606 		goto req_priv;
7607 
7608 	/* Normal users shall not reset the sched_reset_on_fork flag: */
7609 	if (p->sched_reset_on_fork && !reset_on_fork)
7610 		goto req_priv;
7611 
7612 	return 0;
7613 
7614 req_priv:
7615 	if (!capable(CAP_SYS_NICE))
7616 		return -EPERM;
7617 
7618 	return 0;
7619 }
7620 
__sched_setscheduler(struct task_struct * p,const struct sched_attr * attr,bool user,bool pi)7621 static int __sched_setscheduler(struct task_struct *p,
7622 				const struct sched_attr *attr,
7623 				bool user, bool pi)
7624 {
7625 	int oldpolicy = -1, policy = attr->sched_policy;
7626 	int retval, oldprio, newprio, queued, running;
7627 	const struct sched_class *prev_class;
7628 	struct balance_callback *head;
7629 	struct rq_flags rf;
7630 	int reset_on_fork;
7631 	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7632 	struct rq *rq;
7633 	bool cpuset_locked = false;
7634 
7635 	/* The pi code expects interrupts enabled */
7636 	BUG_ON(pi && in_interrupt());
7637 recheck:
7638 	/* Double check policy once rq lock held: */
7639 	if (policy < 0) {
7640 		reset_on_fork = p->sched_reset_on_fork;
7641 		policy = oldpolicy = p->policy;
7642 	} else {
7643 		reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7644 
7645 		if (!valid_policy(policy))
7646 			return -EINVAL;
7647 	}
7648 
7649 	if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7650 		return -EINVAL;
7651 
7652 	/*
7653 	 * Valid priorities for SCHED_FIFO and SCHED_RR are
7654 	 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7655 	 * SCHED_BATCH and SCHED_IDLE is 0.
7656 	 */
7657 	if (attr->sched_priority > MAX_RT_PRIO-1)
7658 		return -EINVAL;
7659 	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7660 	    (rt_policy(policy) != (attr->sched_priority != 0)))
7661 		return -EINVAL;
7662 
7663 	if (user) {
7664 		retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7665 		if (retval)
7666 			return retval;
7667 
7668 		if (attr->sched_flags & SCHED_FLAG_SUGOV)
7669 			return -EINVAL;
7670 
7671 		retval = security_task_setscheduler(p);
7672 		if (retval)
7673 			return retval;
7674 	}
7675 
7676 	/* Update task specific "requested" clamps */
7677 	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7678 		retval = uclamp_validate(p, attr);
7679 		if (retval)
7680 			return retval;
7681 	}
7682 
7683 	/*
7684 	 * SCHED_DEADLINE bandwidth accounting relies on stable cpusets
7685 	 * information.
7686 	 */
7687 	if (dl_policy(policy) || dl_policy(p->policy)) {
7688 		cpuset_locked = true;
7689 		cpuset_lock();
7690 	}
7691 
7692 	/*
7693 	 * Make sure no PI-waiters arrive (or leave) while we are
7694 	 * changing the priority of the task:
7695 	 *
7696 	 * To be able to change p->policy safely, the appropriate
7697 	 * runqueue lock must be held.
7698 	 */
7699 	rq = task_rq_lock(p, &rf);
7700 	update_rq_clock(rq);
7701 
7702 	/*
7703 	 * Changing the policy of the stop threads its a very bad idea:
7704 	 */
7705 	if (p == rq->stop) {
7706 		retval = -EINVAL;
7707 		goto unlock;
7708 	}
7709 
7710 	/*
7711 	 * If not changing anything there's no need to proceed further,
7712 	 * but store a possible modification of reset_on_fork.
7713 	 */
7714 	if (unlikely(policy == p->policy)) {
7715 		if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7716 			goto change;
7717 		if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7718 			goto change;
7719 		if (dl_policy(policy) && dl_param_changed(p, attr))
7720 			goto change;
7721 		if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7722 			goto change;
7723 
7724 		p->sched_reset_on_fork = reset_on_fork;
7725 		retval = 0;
7726 		goto unlock;
7727 	}
7728 change:
7729 
7730 	if (user) {
7731 #ifdef CONFIG_RT_GROUP_SCHED
7732 		/*
7733 		 * Do not allow realtime tasks into groups that have no runtime
7734 		 * assigned.
7735 		 */
7736 		if (rt_bandwidth_enabled() && rt_policy(policy) &&
7737 				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7738 				!task_group_is_autogroup(task_group(p))) {
7739 			retval = -EPERM;
7740 			goto unlock;
7741 		}
7742 #endif
7743 #ifdef CONFIG_SMP
7744 		if (dl_bandwidth_enabled() && dl_policy(policy) &&
7745 				!(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7746 			cpumask_t *span = rq->rd->span;
7747 
7748 			/*
7749 			 * Don't allow tasks with an affinity mask smaller than
7750 			 * the entire root_domain to become SCHED_DEADLINE. We
7751 			 * will also fail if there's no bandwidth available.
7752 			 */
7753 			if (!cpumask_subset(span, p->cpus_ptr) ||
7754 			    rq->rd->dl_bw.bw == 0) {
7755 				retval = -EPERM;
7756 				goto unlock;
7757 			}
7758 		}
7759 #endif
7760 	}
7761 
7762 	/* Re-check policy now with rq lock held: */
7763 	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7764 		policy = oldpolicy = -1;
7765 		task_rq_unlock(rq, p, &rf);
7766 		if (cpuset_locked)
7767 			cpuset_unlock();
7768 		goto recheck;
7769 	}
7770 
7771 	/*
7772 	 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7773 	 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7774 	 * is available.
7775 	 */
7776 	if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7777 		retval = -EBUSY;
7778 		goto unlock;
7779 	}
7780 
7781 	p->sched_reset_on_fork = reset_on_fork;
7782 	oldprio = p->prio;
7783 
7784 	newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7785 	if (pi) {
7786 		/*
7787 		 * Take priority boosted tasks into account. If the new
7788 		 * effective priority is unchanged, we just store the new
7789 		 * normal parameters and do not touch the scheduler class and
7790 		 * the runqueue. This will be done when the task deboost
7791 		 * itself.
7792 		 */
7793 		newprio = rt_effective_prio(p, newprio);
7794 		if (newprio == oldprio)
7795 			queue_flags &= ~DEQUEUE_MOVE;
7796 	}
7797 
7798 	queued = task_on_rq_queued(p);
7799 	running = task_current(rq, p);
7800 	if (queued)
7801 		dequeue_task(rq, p, queue_flags);
7802 	if (running)
7803 		put_prev_task(rq, p);
7804 
7805 	prev_class = p->sched_class;
7806 
7807 	if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7808 		__setscheduler_params(p, attr);
7809 		__setscheduler_prio(p, newprio);
7810 	}
7811 	__setscheduler_uclamp(p, attr);
7812 
7813 	if (queued) {
7814 		/*
7815 		 * We enqueue to tail when the priority of a task is
7816 		 * increased (user space view).
7817 		 */
7818 		if (oldprio < p->prio)
7819 			queue_flags |= ENQUEUE_HEAD;
7820 
7821 		enqueue_task(rq, p, queue_flags);
7822 	}
7823 	if (running)
7824 		set_next_task(rq, p);
7825 
7826 	check_class_changed(rq, p, prev_class, oldprio);
7827 
7828 	/* Avoid rq from going away on us: */
7829 	preempt_disable();
7830 	head = splice_balance_callbacks(rq);
7831 	task_rq_unlock(rq, p, &rf);
7832 
7833 	if (pi) {
7834 		if (cpuset_locked)
7835 			cpuset_unlock();
7836 		rt_mutex_adjust_pi(p);
7837 	}
7838 
7839 	/* Run balance callbacks after we've adjusted the PI chain: */
7840 	balance_callbacks(rq, head);
7841 	preempt_enable();
7842 
7843 	return 0;
7844 
7845 unlock:
7846 	task_rq_unlock(rq, p, &rf);
7847 	if (cpuset_locked)
7848 		cpuset_unlock();
7849 	return retval;
7850 }
7851 
_sched_setscheduler(struct task_struct * p,int policy,const struct sched_param * param,bool check)7852 static int _sched_setscheduler(struct task_struct *p, int policy,
7853 			       const struct sched_param *param, bool check)
7854 {
7855 	struct sched_attr attr = {
7856 		.sched_policy   = policy,
7857 		.sched_priority = param->sched_priority,
7858 		.sched_nice	= PRIO_TO_NICE(p->static_prio),
7859 	};
7860 
7861 	/* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7862 	if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7863 		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7864 		policy &= ~SCHED_RESET_ON_FORK;
7865 		attr.sched_policy = policy;
7866 	}
7867 
7868 	return __sched_setscheduler(p, &attr, check, true);
7869 }
7870 /**
7871  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7872  * @p: the task in question.
7873  * @policy: new policy.
7874  * @param: structure containing the new RT priority.
7875  *
7876  * Use sched_set_fifo(), read its comment.
7877  *
7878  * Return: 0 on success. An error code otherwise.
7879  *
7880  * NOTE that the task may be already dead.
7881  */
sched_setscheduler(struct task_struct * p,int policy,const struct sched_param * param)7882 int sched_setscheduler(struct task_struct *p, int policy,
7883 		       const struct sched_param *param)
7884 {
7885 	return _sched_setscheduler(p, policy, param, true);
7886 }
7887 
sched_setattr(struct task_struct * p,const struct sched_attr * attr)7888 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7889 {
7890 	return __sched_setscheduler(p, attr, true, true);
7891 }
7892 
sched_setattr_nocheck(struct task_struct * p,const struct sched_attr * attr)7893 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7894 {
7895 	return __sched_setscheduler(p, attr, false, true);
7896 }
7897 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7898 
7899 /**
7900  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7901  * @p: the task in question.
7902  * @policy: new policy.
7903  * @param: structure containing the new RT priority.
7904  *
7905  * Just like sched_setscheduler, only don't bother checking if the
7906  * current context has permission.  For example, this is needed in
7907  * stop_machine(): we create temporary high priority worker threads,
7908  * but our caller might not have that capability.
7909  *
7910  * Return: 0 on success. An error code otherwise.
7911  */
sched_setscheduler_nocheck(struct task_struct * p,int policy,const struct sched_param * param)7912 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7913 			       const struct sched_param *param)
7914 {
7915 	return _sched_setscheduler(p, policy, param, false);
7916 }
7917 
7918 /*
7919  * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7920  * incapable of resource management, which is the one thing an OS really should
7921  * be doing.
7922  *
7923  * This is of course the reason it is limited to privileged users only.
7924  *
7925  * Worse still; it is fundamentally impossible to compose static priority
7926  * workloads. You cannot take two correctly working static prio workloads
7927  * and smash them together and still expect them to work.
7928  *
7929  * For this reason 'all' FIFO tasks the kernel creates are basically at:
7930  *
7931  *   MAX_RT_PRIO / 2
7932  *
7933  * The administrator _MUST_ configure the system, the kernel simply doesn't
7934  * know enough information to make a sensible choice.
7935  */
sched_set_fifo(struct task_struct * p)7936 void sched_set_fifo(struct task_struct *p)
7937 {
7938 	struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7939 	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7940 }
7941 EXPORT_SYMBOL_GPL(sched_set_fifo);
7942 
7943 /*
7944  * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7945  */
sched_set_fifo_low(struct task_struct * p)7946 void sched_set_fifo_low(struct task_struct *p)
7947 {
7948 	struct sched_param sp = { .sched_priority = 1 };
7949 	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7950 }
7951 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7952 
sched_set_normal(struct task_struct * p,int nice)7953 void sched_set_normal(struct task_struct *p, int nice)
7954 {
7955 	struct sched_attr attr = {
7956 		.sched_policy = SCHED_NORMAL,
7957 		.sched_nice = nice,
7958 	};
7959 	WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7960 }
7961 EXPORT_SYMBOL_GPL(sched_set_normal);
7962 
7963 static int
do_sched_setscheduler(pid_t pid,int policy,struct sched_param __user * param)7964 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7965 {
7966 	struct sched_param lparam;
7967 	struct task_struct *p;
7968 	int retval;
7969 
7970 	if (!param || pid < 0)
7971 		return -EINVAL;
7972 	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7973 		return -EFAULT;
7974 
7975 	rcu_read_lock();
7976 	retval = -ESRCH;
7977 	p = find_process_by_pid(pid);
7978 	if (likely(p))
7979 		get_task_struct(p);
7980 	rcu_read_unlock();
7981 
7982 	if (likely(p)) {
7983 		retval = sched_setscheduler(p, policy, &lparam);
7984 		put_task_struct(p);
7985 	}
7986 
7987 	return retval;
7988 }
7989 
7990 /*
7991  * Mimics kernel/events/core.c perf_copy_attr().
7992  */
sched_copy_attr(struct sched_attr __user * uattr,struct sched_attr * attr)7993 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7994 {
7995 	u32 size;
7996 	int ret;
7997 
7998 	/* Zero the full structure, so that a short copy will be nice: */
7999 	memset(attr, 0, sizeof(*attr));
8000 
8001 	ret = get_user(size, &uattr->size);
8002 	if (ret)
8003 		return ret;
8004 
8005 	/* ABI compatibility quirk: */
8006 	if (!size)
8007 		size = SCHED_ATTR_SIZE_VER0;
8008 	if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
8009 		goto err_size;
8010 
8011 	ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
8012 	if (ret) {
8013 		if (ret == -E2BIG)
8014 			goto err_size;
8015 		return ret;
8016 	}
8017 
8018 	if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
8019 	    size < SCHED_ATTR_SIZE_VER1)
8020 		return -EINVAL;
8021 
8022 	/*
8023 	 * XXX: Do we want to be lenient like existing syscalls; or do we want
8024 	 * to be strict and return an error on out-of-bounds values?
8025 	 */
8026 	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
8027 
8028 	return 0;
8029 
8030 err_size:
8031 	put_user(sizeof(*attr), &uattr->size);
8032 	return -E2BIG;
8033 }
8034 
get_params(struct task_struct * p,struct sched_attr * attr)8035 static void get_params(struct task_struct *p, struct sched_attr *attr)
8036 {
8037 	if (task_has_dl_policy(p))
8038 		__getparam_dl(p, attr);
8039 	else if (task_has_rt_policy(p))
8040 		attr->sched_priority = p->rt_priority;
8041 	else
8042 		attr->sched_nice = task_nice(p);
8043 }
8044 
8045 /**
8046  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
8047  * @pid: the pid in question.
8048  * @policy: new policy.
8049  * @param: structure containing the new RT priority.
8050  *
8051  * Return: 0 on success. An error code otherwise.
8052  */
SYSCALL_DEFINE3(sched_setscheduler,pid_t,pid,int,policy,struct sched_param __user *,param)8053 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
8054 {
8055 	if (policy < 0)
8056 		return -EINVAL;
8057 
8058 	return do_sched_setscheduler(pid, policy, param);
8059 }
8060 
8061 /**
8062  * sys_sched_setparam - set/change the RT priority of a thread
8063  * @pid: the pid in question.
8064  * @param: structure containing the new RT priority.
8065  *
8066  * Return: 0 on success. An error code otherwise.
8067  */
SYSCALL_DEFINE2(sched_setparam,pid_t,pid,struct sched_param __user *,param)8068 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
8069 {
8070 	return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
8071 }
8072 
8073 /**
8074  * sys_sched_setattr - same as above, but with extended sched_attr
8075  * @pid: the pid in question.
8076  * @uattr: structure containing the extended parameters.
8077  * @flags: for future extension.
8078  */
SYSCALL_DEFINE3(sched_setattr,pid_t,pid,struct sched_attr __user *,uattr,unsigned int,flags)8079 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
8080 			       unsigned int, flags)
8081 {
8082 	struct sched_attr attr;
8083 	struct task_struct *p;
8084 	int retval;
8085 
8086 	if (!uattr || pid < 0 || flags)
8087 		return -EINVAL;
8088 
8089 	retval = sched_copy_attr(uattr, &attr);
8090 	if (retval)
8091 		return retval;
8092 
8093 	if ((int)attr.sched_policy < 0)
8094 		return -EINVAL;
8095 	if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
8096 		attr.sched_policy = SETPARAM_POLICY;
8097 
8098 	rcu_read_lock();
8099 	retval = -ESRCH;
8100 	p = find_process_by_pid(pid);
8101 	if (likely(p))
8102 		get_task_struct(p);
8103 	rcu_read_unlock();
8104 
8105 	if (likely(p)) {
8106 		if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
8107 			get_params(p, &attr);
8108 		retval = sched_setattr(p, &attr);
8109 		put_task_struct(p);
8110 	}
8111 
8112 	return retval;
8113 }
8114 
8115 /**
8116  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
8117  * @pid: the pid in question.
8118  *
8119  * Return: On success, the policy of the thread. Otherwise, a negative error
8120  * code.
8121  */
SYSCALL_DEFINE1(sched_getscheduler,pid_t,pid)8122 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
8123 {
8124 	struct task_struct *p;
8125 	int retval;
8126 
8127 	if (pid < 0)
8128 		return -EINVAL;
8129 
8130 	retval = -ESRCH;
8131 	rcu_read_lock();
8132 	p = find_process_by_pid(pid);
8133 	if (p) {
8134 		retval = security_task_getscheduler(p);
8135 		if (!retval)
8136 			retval = p->policy
8137 				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
8138 	}
8139 	rcu_read_unlock();
8140 	return retval;
8141 }
8142 
8143 /**
8144  * sys_sched_getparam - get the RT priority of a thread
8145  * @pid: the pid in question.
8146  * @param: structure containing the RT priority.
8147  *
8148  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8149  * code.
8150  */
SYSCALL_DEFINE2(sched_getparam,pid_t,pid,struct sched_param __user *,param)8151 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
8152 {
8153 	struct sched_param lp = { .sched_priority = 0 };
8154 	struct task_struct *p;
8155 	int retval;
8156 
8157 	if (!param || pid < 0)
8158 		return -EINVAL;
8159 
8160 	rcu_read_lock();
8161 	p = find_process_by_pid(pid);
8162 	retval = -ESRCH;
8163 	if (!p)
8164 		goto out_unlock;
8165 
8166 	retval = security_task_getscheduler(p);
8167 	if (retval)
8168 		goto out_unlock;
8169 
8170 	if (task_has_rt_policy(p))
8171 		lp.sched_priority = p->rt_priority;
8172 	rcu_read_unlock();
8173 
8174 	/*
8175 	 * This one might sleep, we cannot do it with a spinlock held ...
8176 	 */
8177 	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
8178 
8179 	return retval;
8180 
8181 out_unlock:
8182 	rcu_read_unlock();
8183 	return retval;
8184 }
8185 
8186 /*
8187  * Copy the kernel size attribute structure (which might be larger
8188  * than what user-space knows about) to user-space.
8189  *
8190  * Note that all cases are valid: user-space buffer can be larger or
8191  * smaller than the kernel-space buffer. The usual case is that both
8192  * have the same size.
8193  */
8194 static int
sched_attr_copy_to_user(struct sched_attr __user * uattr,struct sched_attr * kattr,unsigned int usize)8195 sched_attr_copy_to_user(struct sched_attr __user *uattr,
8196 			struct sched_attr *kattr,
8197 			unsigned int usize)
8198 {
8199 	unsigned int ksize = sizeof(*kattr);
8200 
8201 	if (!access_ok(uattr, usize))
8202 		return -EFAULT;
8203 
8204 	/*
8205 	 * sched_getattr() ABI forwards and backwards compatibility:
8206 	 *
8207 	 * If usize == ksize then we just copy everything to user-space and all is good.
8208 	 *
8209 	 * If usize < ksize then we only copy as much as user-space has space for,
8210 	 * this keeps ABI compatibility as well. We skip the rest.
8211 	 *
8212 	 * If usize > ksize then user-space is using a newer version of the ABI,
8213 	 * which part the kernel doesn't know about. Just ignore it - tooling can
8214 	 * detect the kernel's knowledge of attributes from the attr->size value
8215 	 * which is set to ksize in this case.
8216 	 */
8217 	kattr->size = min(usize, ksize);
8218 
8219 	if (copy_to_user(uattr, kattr, kattr->size))
8220 		return -EFAULT;
8221 
8222 	return 0;
8223 }
8224 
8225 /**
8226  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8227  * @pid: the pid in question.
8228  * @uattr: structure containing the extended parameters.
8229  * @usize: sizeof(attr) for fwd/bwd comp.
8230  * @flags: for future extension.
8231  */
SYSCALL_DEFINE4(sched_getattr,pid_t,pid,struct sched_attr __user *,uattr,unsigned int,usize,unsigned int,flags)8232 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8233 		unsigned int, usize, unsigned int, flags)
8234 {
8235 	struct sched_attr kattr = { };
8236 	struct task_struct *p;
8237 	int retval;
8238 
8239 	if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8240 	    usize < SCHED_ATTR_SIZE_VER0 || flags)
8241 		return -EINVAL;
8242 
8243 	rcu_read_lock();
8244 	p = find_process_by_pid(pid);
8245 	retval = -ESRCH;
8246 	if (!p)
8247 		goto out_unlock;
8248 
8249 	retval = security_task_getscheduler(p);
8250 	if (retval)
8251 		goto out_unlock;
8252 
8253 	kattr.sched_policy = p->policy;
8254 	if (p->sched_reset_on_fork)
8255 		kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8256 	get_params(p, &kattr);
8257 	kattr.sched_flags &= SCHED_FLAG_ALL;
8258 
8259 #ifdef CONFIG_UCLAMP_TASK
8260 	/*
8261 	 * This could race with another potential updater, but this is fine
8262 	 * because it'll correctly read the old or the new value. We don't need
8263 	 * to guarantee who wins the race as long as it doesn't return garbage.
8264 	 */
8265 	kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8266 	kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8267 #endif
8268 
8269 	rcu_read_unlock();
8270 
8271 	return sched_attr_copy_to_user(uattr, &kattr, usize);
8272 
8273 out_unlock:
8274 	rcu_read_unlock();
8275 	return retval;
8276 }
8277 
8278 #ifdef CONFIG_SMP
dl_task_check_affinity(struct task_struct * p,const struct cpumask * mask)8279 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8280 {
8281 	int ret = 0;
8282 
8283 	/*
8284 	 * If the task isn't a deadline task or admission control is
8285 	 * disabled then we don't care about affinity changes.
8286 	 */
8287 	if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8288 		return 0;
8289 
8290 	/*
8291 	 * Since bandwidth control happens on root_domain basis,
8292 	 * if admission test is enabled, we only admit -deadline
8293 	 * tasks allowed to run on all the CPUs in the task's
8294 	 * root_domain.
8295 	 */
8296 	rcu_read_lock();
8297 	if (!cpumask_subset(task_rq(p)->rd->span, mask))
8298 		ret = -EBUSY;
8299 	rcu_read_unlock();
8300 	return ret;
8301 }
8302 #endif
8303 
8304 static int
__sched_setaffinity(struct task_struct * p,struct affinity_context * ctx)8305 __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
8306 {
8307 	int retval;
8308 	cpumask_var_t cpus_allowed, new_mask;
8309 
8310 	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8311 		return -ENOMEM;
8312 
8313 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8314 		retval = -ENOMEM;
8315 		goto out_free_cpus_allowed;
8316 	}
8317 
8318 	cpuset_cpus_allowed(p, cpus_allowed);
8319 	cpumask_and(new_mask, ctx->new_mask, cpus_allowed);
8320 
8321 	ctx->new_mask = new_mask;
8322 	ctx->flags |= SCA_CHECK;
8323 
8324 	retval = dl_task_check_affinity(p, new_mask);
8325 	if (retval)
8326 		goto out_free_new_mask;
8327 
8328 	retval = __set_cpus_allowed_ptr(p, ctx);
8329 	if (retval)
8330 		goto out_free_new_mask;
8331 
8332 	cpuset_cpus_allowed(p, cpus_allowed);
8333 	if (!cpumask_subset(new_mask, cpus_allowed)) {
8334 		/*
8335 		 * We must have raced with a concurrent cpuset update.
8336 		 * Just reset the cpumask to the cpuset's cpus_allowed.
8337 		 */
8338 		cpumask_copy(new_mask, cpus_allowed);
8339 
8340 		/*
8341 		 * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8342 		 * will restore the previous user_cpus_ptr value.
8343 		 *
8344 		 * In the unlikely event a previous user_cpus_ptr exists,
8345 		 * we need to further restrict the mask to what is allowed
8346 		 * by that old user_cpus_ptr.
8347 		 */
8348 		if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
8349 			bool empty = !cpumask_and(new_mask, new_mask,
8350 						  ctx->user_mask);
8351 
8352 			if (WARN_ON_ONCE(empty))
8353 				cpumask_copy(new_mask, cpus_allowed);
8354 		}
8355 		__set_cpus_allowed_ptr(p, ctx);
8356 		retval = -EINVAL;
8357 	}
8358 
8359 out_free_new_mask:
8360 	free_cpumask_var(new_mask);
8361 out_free_cpus_allowed:
8362 	free_cpumask_var(cpus_allowed);
8363 	return retval;
8364 }
8365 
sched_setaffinity(pid_t pid,const struct cpumask * in_mask)8366 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8367 {
8368 	struct affinity_context ac;
8369 	struct cpumask *user_mask;
8370 	struct task_struct *p;
8371 	int retval;
8372 
8373 	rcu_read_lock();
8374 
8375 	p = find_process_by_pid(pid);
8376 	if (!p) {
8377 		rcu_read_unlock();
8378 		return -ESRCH;
8379 	}
8380 
8381 	/* Prevent p going away */
8382 	get_task_struct(p);
8383 	rcu_read_unlock();
8384 
8385 	if (p->flags & PF_NO_SETAFFINITY) {
8386 		retval = -EINVAL;
8387 		goto out_put_task;
8388 	}
8389 
8390 	if (!check_same_owner(p)) {
8391 		rcu_read_lock();
8392 		if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8393 			rcu_read_unlock();
8394 			retval = -EPERM;
8395 			goto out_put_task;
8396 		}
8397 		rcu_read_unlock();
8398 	}
8399 
8400 	retval = security_task_setscheduler(p);
8401 	if (retval)
8402 		goto out_put_task;
8403 
8404 	/*
8405 	 * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8406 	 * alloc_user_cpus_ptr() returns NULL.
8407 	 */
8408 	user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
8409 	if (user_mask) {
8410 		cpumask_copy(user_mask, in_mask);
8411 	} else if (IS_ENABLED(CONFIG_SMP)) {
8412 		retval = -ENOMEM;
8413 		goto out_put_task;
8414 	}
8415 
8416 	ac = (struct affinity_context){
8417 		.new_mask  = in_mask,
8418 		.user_mask = user_mask,
8419 		.flags     = SCA_USER,
8420 	};
8421 
8422 	retval = __sched_setaffinity(p, &ac);
8423 	kfree(ac.user_mask);
8424 
8425 out_put_task:
8426 	put_task_struct(p);
8427 	return retval;
8428 }
8429 
get_user_cpu_mask(unsigned long __user * user_mask_ptr,unsigned len,struct cpumask * new_mask)8430 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8431 			     struct cpumask *new_mask)
8432 {
8433 	if (len < cpumask_size())
8434 		cpumask_clear(new_mask);
8435 	else if (len > cpumask_size())
8436 		len = cpumask_size();
8437 
8438 	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8439 }
8440 
8441 /**
8442  * sys_sched_setaffinity - set the CPU affinity of a process
8443  * @pid: pid of the process
8444  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8445  * @user_mask_ptr: user-space pointer to the new CPU mask
8446  *
8447  * Return: 0 on success. An error code otherwise.
8448  */
SYSCALL_DEFINE3(sched_setaffinity,pid_t,pid,unsigned int,len,unsigned long __user *,user_mask_ptr)8449 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8450 		unsigned long __user *, user_mask_ptr)
8451 {
8452 	cpumask_var_t new_mask;
8453 	int retval;
8454 
8455 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8456 		return -ENOMEM;
8457 
8458 	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8459 	if (retval == 0)
8460 		retval = sched_setaffinity(pid, new_mask);
8461 	free_cpumask_var(new_mask);
8462 	return retval;
8463 }
8464 
sched_getaffinity(pid_t pid,struct cpumask * mask)8465 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8466 {
8467 	struct task_struct *p;
8468 	unsigned long flags;
8469 	int retval;
8470 
8471 	rcu_read_lock();
8472 
8473 	retval = -ESRCH;
8474 	p = find_process_by_pid(pid);
8475 	if (!p)
8476 		goto out_unlock;
8477 
8478 	retval = security_task_getscheduler(p);
8479 	if (retval)
8480 		goto out_unlock;
8481 
8482 	raw_spin_lock_irqsave(&p->pi_lock, flags);
8483 	cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8484 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8485 
8486 out_unlock:
8487 	rcu_read_unlock();
8488 
8489 	return retval;
8490 }
8491 
8492 /**
8493  * sys_sched_getaffinity - get the CPU affinity of a process
8494  * @pid: pid of the process
8495  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8496  * @user_mask_ptr: user-space pointer to hold the current CPU mask
8497  *
8498  * Return: size of CPU mask copied to user_mask_ptr on success. An
8499  * error code otherwise.
8500  */
SYSCALL_DEFINE3(sched_getaffinity,pid_t,pid,unsigned int,len,unsigned long __user *,user_mask_ptr)8501 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8502 		unsigned long __user *, user_mask_ptr)
8503 {
8504 	int ret;
8505 	cpumask_var_t mask;
8506 
8507 	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8508 		return -EINVAL;
8509 	if (len & (sizeof(unsigned long)-1))
8510 		return -EINVAL;
8511 
8512 	if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
8513 		return -ENOMEM;
8514 
8515 	ret = sched_getaffinity(pid, mask);
8516 	if (ret == 0) {
8517 		unsigned int retlen = min(len, cpumask_size());
8518 
8519 		if (copy_to_user(user_mask_ptr, cpumask_bits(mask), retlen))
8520 			ret = -EFAULT;
8521 		else
8522 			ret = retlen;
8523 	}
8524 	free_cpumask_var(mask);
8525 
8526 	return ret;
8527 }
8528 
do_sched_yield(void)8529 static void do_sched_yield(void)
8530 {
8531 	struct rq_flags rf;
8532 	struct rq *rq;
8533 
8534 	rq = this_rq_lock_irq(&rf);
8535 
8536 	schedstat_inc(rq->yld_count);
8537 	current->sched_class->yield_task(rq);
8538 
8539 	preempt_disable();
8540 	rq_unlock_irq(rq, &rf);
8541 	sched_preempt_enable_no_resched();
8542 
8543 	schedule();
8544 }
8545 
8546 /**
8547  * sys_sched_yield - yield the current processor to other threads.
8548  *
8549  * This function yields the current CPU to other tasks. If there are no
8550  * other threads running on this CPU then this function will return.
8551  *
8552  * Return: 0.
8553  */
SYSCALL_DEFINE0(sched_yield)8554 SYSCALL_DEFINE0(sched_yield)
8555 {
8556 	do_sched_yield();
8557 	return 0;
8558 }
8559 
8560 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
__cond_resched(void)8561 int __sched __cond_resched(void)
8562 {
8563 	if (should_resched(0)) {
8564 		preempt_schedule_common();
8565 		return 1;
8566 	}
8567 	/*
8568 	 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8569 	 * whether the current CPU is in an RCU read-side critical section,
8570 	 * so the tick can report quiescent states even for CPUs looping
8571 	 * in kernel context.  In contrast, in non-preemptible kernels,
8572 	 * RCU readers leave no in-memory hints, which means that CPU-bound
8573 	 * processes executing in kernel context might never report an
8574 	 * RCU quiescent state.  Therefore, the following code causes
8575 	 * cond_resched() to report a quiescent state, but only when RCU
8576 	 * is in urgent need of one.
8577 	 */
8578 #ifndef CONFIG_PREEMPT_RCU
8579 	rcu_all_qs();
8580 #endif
8581 	return 0;
8582 }
8583 EXPORT_SYMBOL(__cond_resched);
8584 #endif
8585 
8586 #ifdef CONFIG_PREEMPT_DYNAMIC
8587 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8588 #define cond_resched_dynamic_enabled	__cond_resched
8589 #define cond_resched_dynamic_disabled	((void *)&__static_call_return0)
8590 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8591 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8592 
8593 #define might_resched_dynamic_enabled	__cond_resched
8594 #define might_resched_dynamic_disabled	((void *)&__static_call_return0)
8595 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8596 EXPORT_STATIC_CALL_TRAMP(might_resched);
8597 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8598 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
dynamic_cond_resched(void)8599 int __sched dynamic_cond_resched(void)
8600 {
8601 	klp_sched_try_switch();
8602 	if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8603 		return 0;
8604 	return __cond_resched();
8605 }
8606 EXPORT_SYMBOL(dynamic_cond_resched);
8607 
8608 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
dynamic_might_resched(void)8609 int __sched dynamic_might_resched(void)
8610 {
8611 	if (!static_branch_unlikely(&sk_dynamic_might_resched))
8612 		return 0;
8613 	return __cond_resched();
8614 }
8615 EXPORT_SYMBOL(dynamic_might_resched);
8616 #endif
8617 #endif
8618 
8619 /*
8620  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8621  * call schedule, and on return reacquire the lock.
8622  *
8623  * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8624  * operations here to prevent schedule() from being called twice (once via
8625  * spin_unlock(), once by hand).
8626  */
__cond_resched_lock(spinlock_t * lock)8627 int __cond_resched_lock(spinlock_t *lock)
8628 {
8629 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8630 	int ret = 0;
8631 
8632 	lockdep_assert_held(lock);
8633 
8634 	if (spin_needbreak(lock) || resched) {
8635 		spin_unlock(lock);
8636 		if (!_cond_resched())
8637 			cpu_relax();
8638 		ret = 1;
8639 		spin_lock(lock);
8640 	}
8641 	return ret;
8642 }
8643 EXPORT_SYMBOL(__cond_resched_lock);
8644 
__cond_resched_rwlock_read(rwlock_t * lock)8645 int __cond_resched_rwlock_read(rwlock_t *lock)
8646 {
8647 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8648 	int ret = 0;
8649 
8650 	lockdep_assert_held_read(lock);
8651 
8652 	if (rwlock_needbreak(lock) || resched) {
8653 		read_unlock(lock);
8654 		if (!_cond_resched())
8655 			cpu_relax();
8656 		ret = 1;
8657 		read_lock(lock);
8658 	}
8659 	return ret;
8660 }
8661 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8662 
__cond_resched_rwlock_write(rwlock_t * lock)8663 int __cond_resched_rwlock_write(rwlock_t *lock)
8664 {
8665 	int resched = should_resched(PREEMPT_LOCK_OFFSET);
8666 	int ret = 0;
8667 
8668 	lockdep_assert_held_write(lock);
8669 
8670 	if (rwlock_needbreak(lock) || resched) {
8671 		write_unlock(lock);
8672 		if (!_cond_resched())
8673 			cpu_relax();
8674 		ret = 1;
8675 		write_lock(lock);
8676 	}
8677 	return ret;
8678 }
8679 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8680 
8681 #ifdef CONFIG_PREEMPT_DYNAMIC
8682 
8683 #ifdef CONFIG_GENERIC_ENTRY
8684 #include <linux/entry-common.h>
8685 #endif
8686 
8687 /*
8688  * SC:cond_resched
8689  * SC:might_resched
8690  * SC:preempt_schedule
8691  * SC:preempt_schedule_notrace
8692  * SC:irqentry_exit_cond_resched
8693  *
8694  *
8695  * NONE:
8696  *   cond_resched               <- __cond_resched
8697  *   might_resched              <- RET0
8698  *   preempt_schedule           <- NOP
8699  *   preempt_schedule_notrace   <- NOP
8700  *   irqentry_exit_cond_resched <- NOP
8701  *
8702  * VOLUNTARY:
8703  *   cond_resched               <- __cond_resched
8704  *   might_resched              <- __cond_resched
8705  *   preempt_schedule           <- NOP
8706  *   preempt_schedule_notrace   <- NOP
8707  *   irqentry_exit_cond_resched <- NOP
8708  *
8709  * FULL:
8710  *   cond_resched               <- RET0
8711  *   might_resched              <- RET0
8712  *   preempt_schedule           <- preempt_schedule
8713  *   preempt_schedule_notrace   <- preempt_schedule_notrace
8714  *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8715  */
8716 
8717 enum {
8718 	preempt_dynamic_undefined = -1,
8719 	preempt_dynamic_none,
8720 	preempt_dynamic_voluntary,
8721 	preempt_dynamic_full,
8722 };
8723 
8724 int preempt_dynamic_mode = preempt_dynamic_undefined;
8725 
sched_dynamic_mode(const char * str)8726 int sched_dynamic_mode(const char *str)
8727 {
8728 	if (!strcmp(str, "none"))
8729 		return preempt_dynamic_none;
8730 
8731 	if (!strcmp(str, "voluntary"))
8732 		return preempt_dynamic_voluntary;
8733 
8734 	if (!strcmp(str, "full"))
8735 		return preempt_dynamic_full;
8736 
8737 	return -EINVAL;
8738 }
8739 
8740 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8741 #define preempt_dynamic_enable(f)	static_call_update(f, f##_dynamic_enabled)
8742 #define preempt_dynamic_disable(f)	static_call_update(f, f##_dynamic_disabled)
8743 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8744 #define preempt_dynamic_enable(f)	static_key_enable(&sk_dynamic_##f.key)
8745 #define preempt_dynamic_disable(f)	static_key_disable(&sk_dynamic_##f.key)
8746 #else
8747 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8748 #endif
8749 
8750 static DEFINE_MUTEX(sched_dynamic_mutex);
8751 static bool klp_override;
8752 
__sched_dynamic_update(int mode)8753 static void __sched_dynamic_update(int mode)
8754 {
8755 	/*
8756 	 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8757 	 * the ZERO state, which is invalid.
8758 	 */
8759 	if (!klp_override)
8760 		preempt_dynamic_enable(cond_resched);
8761 	preempt_dynamic_enable(might_resched);
8762 	preempt_dynamic_enable(preempt_schedule);
8763 	preempt_dynamic_enable(preempt_schedule_notrace);
8764 	preempt_dynamic_enable(irqentry_exit_cond_resched);
8765 
8766 	switch (mode) {
8767 	case preempt_dynamic_none:
8768 		if (!klp_override)
8769 			preempt_dynamic_enable(cond_resched);
8770 		preempt_dynamic_disable(might_resched);
8771 		preempt_dynamic_disable(preempt_schedule);
8772 		preempt_dynamic_disable(preempt_schedule_notrace);
8773 		preempt_dynamic_disable(irqentry_exit_cond_resched);
8774 		if (mode != preempt_dynamic_mode)
8775 			pr_info("Dynamic Preempt: none\n");
8776 		break;
8777 
8778 	case preempt_dynamic_voluntary:
8779 		if (!klp_override)
8780 			preempt_dynamic_enable(cond_resched);
8781 		preempt_dynamic_enable(might_resched);
8782 		preempt_dynamic_disable(preempt_schedule);
8783 		preempt_dynamic_disable(preempt_schedule_notrace);
8784 		preempt_dynamic_disable(irqentry_exit_cond_resched);
8785 		if (mode != preempt_dynamic_mode)
8786 			pr_info("Dynamic Preempt: voluntary\n");
8787 		break;
8788 
8789 	case preempt_dynamic_full:
8790 		if (!klp_override)
8791 			preempt_dynamic_disable(cond_resched);
8792 		preempt_dynamic_disable(might_resched);
8793 		preempt_dynamic_enable(preempt_schedule);
8794 		preempt_dynamic_enable(preempt_schedule_notrace);
8795 		preempt_dynamic_enable(irqentry_exit_cond_resched);
8796 		if (mode != preempt_dynamic_mode)
8797 			pr_info("Dynamic Preempt: full\n");
8798 		break;
8799 	}
8800 
8801 	preempt_dynamic_mode = mode;
8802 }
8803 
sched_dynamic_update(int mode)8804 void sched_dynamic_update(int mode)
8805 {
8806 	mutex_lock(&sched_dynamic_mutex);
8807 	__sched_dynamic_update(mode);
8808 	mutex_unlock(&sched_dynamic_mutex);
8809 }
8810 
8811 #ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
8812 
klp_cond_resched(void)8813 static int klp_cond_resched(void)
8814 {
8815 	__klp_sched_try_switch();
8816 	return __cond_resched();
8817 }
8818 
sched_dynamic_klp_enable(void)8819 void sched_dynamic_klp_enable(void)
8820 {
8821 	mutex_lock(&sched_dynamic_mutex);
8822 
8823 	klp_override = true;
8824 	static_call_update(cond_resched, klp_cond_resched);
8825 
8826 	mutex_unlock(&sched_dynamic_mutex);
8827 }
8828 
sched_dynamic_klp_disable(void)8829 void sched_dynamic_klp_disable(void)
8830 {
8831 	mutex_lock(&sched_dynamic_mutex);
8832 
8833 	klp_override = false;
8834 	__sched_dynamic_update(preempt_dynamic_mode);
8835 
8836 	mutex_unlock(&sched_dynamic_mutex);
8837 }
8838 
8839 #endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
8840 
setup_preempt_mode(char * str)8841 static int __init setup_preempt_mode(char *str)
8842 {
8843 	int mode = sched_dynamic_mode(str);
8844 	if (mode < 0) {
8845 		pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8846 		return 0;
8847 	}
8848 
8849 	sched_dynamic_update(mode);
8850 	return 1;
8851 }
8852 __setup("preempt=", setup_preempt_mode);
8853 
preempt_dynamic_init(void)8854 static void __init preempt_dynamic_init(void)
8855 {
8856 	if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8857 		if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8858 			sched_dynamic_update(preempt_dynamic_none);
8859 		} else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8860 			sched_dynamic_update(preempt_dynamic_voluntary);
8861 		} else {
8862 			/* Default static call setting, nothing to do */
8863 			WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8864 			preempt_dynamic_mode = preempt_dynamic_full;
8865 			pr_info("Dynamic Preempt: full\n");
8866 		}
8867 	}
8868 }
8869 
8870 #define PREEMPT_MODEL_ACCESSOR(mode) \
8871 	bool preempt_model_##mode(void)						 \
8872 	{									 \
8873 		WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8874 		return preempt_dynamic_mode == preempt_dynamic_##mode;		 \
8875 	}									 \
8876 	EXPORT_SYMBOL_GPL(preempt_model_##mode)
8877 
8878 PREEMPT_MODEL_ACCESSOR(none);
8879 PREEMPT_MODEL_ACCESSOR(voluntary);
8880 PREEMPT_MODEL_ACCESSOR(full);
8881 
8882 #else /* !CONFIG_PREEMPT_DYNAMIC */
8883 
preempt_dynamic_init(void)8884 static inline void preempt_dynamic_init(void) { }
8885 
8886 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8887 
8888 /**
8889  * yield - yield the current processor to other threads.
8890  *
8891  * Do not ever use this function, there's a 99% chance you're doing it wrong.
8892  *
8893  * The scheduler is at all times free to pick the calling task as the most
8894  * eligible task to run, if removing the yield() call from your code breaks
8895  * it, it's already broken.
8896  *
8897  * Typical broken usage is:
8898  *
8899  * while (!event)
8900  *	yield();
8901  *
8902  * where one assumes that yield() will let 'the other' process run that will
8903  * make event true. If the current task is a SCHED_FIFO task that will never
8904  * happen. Never use yield() as a progress guarantee!!
8905  *
8906  * If you want to use yield() to wait for something, use wait_event().
8907  * If you want to use yield() to be 'nice' for others, use cond_resched().
8908  * If you still want to use yield(), do not!
8909  */
yield(void)8910 void __sched yield(void)
8911 {
8912 	set_current_state(TASK_RUNNING);
8913 	do_sched_yield();
8914 }
8915 EXPORT_SYMBOL(yield);
8916 
8917 /**
8918  * yield_to - yield the current processor to another thread in
8919  * your thread group, or accelerate that thread toward the
8920  * processor it's on.
8921  * @p: target task
8922  * @preempt: whether task preemption is allowed or not
8923  *
8924  * It's the caller's job to ensure that the target task struct
8925  * can't go away on us before we can do any checks.
8926  *
8927  * Return:
8928  *	true (>0) if we indeed boosted the target task.
8929  *	false (0) if we failed to boost the target.
8930  *	-ESRCH if there's no task to yield to.
8931  */
yield_to(struct task_struct * p,bool preempt)8932 int __sched yield_to(struct task_struct *p, bool preempt)
8933 {
8934 	struct task_struct *curr = current;
8935 	struct rq *rq, *p_rq;
8936 	unsigned long flags;
8937 	int yielded = 0;
8938 
8939 	local_irq_save(flags);
8940 	rq = this_rq();
8941 
8942 again:
8943 	p_rq = task_rq(p);
8944 	/*
8945 	 * If we're the only runnable task on the rq and target rq also
8946 	 * has only one task, there's absolutely no point in yielding.
8947 	 */
8948 	if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8949 		yielded = -ESRCH;
8950 		goto out_irq;
8951 	}
8952 
8953 	double_rq_lock(rq, p_rq);
8954 	if (task_rq(p) != p_rq) {
8955 		double_rq_unlock(rq, p_rq);
8956 		goto again;
8957 	}
8958 
8959 	if (!curr->sched_class->yield_to_task)
8960 		goto out_unlock;
8961 
8962 	if (curr->sched_class != p->sched_class)
8963 		goto out_unlock;
8964 
8965 	if (task_on_cpu(p_rq, p) || !task_is_running(p))
8966 		goto out_unlock;
8967 
8968 	yielded = curr->sched_class->yield_to_task(rq, p);
8969 	if (yielded) {
8970 		schedstat_inc(rq->yld_count);
8971 		/*
8972 		 * Make p's CPU reschedule; pick_next_entity takes care of
8973 		 * fairness.
8974 		 */
8975 		if (preempt && rq != p_rq)
8976 			resched_curr(p_rq);
8977 	}
8978 
8979 out_unlock:
8980 	double_rq_unlock(rq, p_rq);
8981 out_irq:
8982 	local_irq_restore(flags);
8983 
8984 	if (yielded > 0)
8985 		schedule();
8986 
8987 	return yielded;
8988 }
8989 EXPORT_SYMBOL_GPL(yield_to);
8990 
io_schedule_prepare(void)8991 int io_schedule_prepare(void)
8992 {
8993 	int old_iowait = current->in_iowait;
8994 
8995 	current->in_iowait = 1;
8996 	blk_flush_plug(current->plug, true);
8997 	return old_iowait;
8998 }
8999 
io_schedule_finish(int token)9000 void io_schedule_finish(int token)
9001 {
9002 	current->in_iowait = token;
9003 }
9004 
9005 /*
9006  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
9007  * that process accounting knows that this is a task in IO wait state.
9008  */
io_schedule_timeout(long timeout)9009 long __sched io_schedule_timeout(long timeout)
9010 {
9011 	int token;
9012 	long ret;
9013 
9014 	token = io_schedule_prepare();
9015 	ret = schedule_timeout(timeout);
9016 	io_schedule_finish(token);
9017 
9018 	return ret;
9019 }
9020 EXPORT_SYMBOL(io_schedule_timeout);
9021 
io_schedule(void)9022 void __sched io_schedule(void)
9023 {
9024 	int token;
9025 
9026 	token = io_schedule_prepare();
9027 	schedule();
9028 	io_schedule_finish(token);
9029 }
9030 EXPORT_SYMBOL(io_schedule);
9031 
9032 /**
9033  * sys_sched_get_priority_max - return maximum RT priority.
9034  * @policy: scheduling class.
9035  *
9036  * Return: On success, this syscall returns the maximum
9037  * rt_priority that can be used by a given scheduling class.
9038  * On failure, a negative error code is returned.
9039  */
SYSCALL_DEFINE1(sched_get_priority_max,int,policy)9040 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
9041 {
9042 	int ret = -EINVAL;
9043 
9044 	switch (policy) {
9045 	case SCHED_FIFO:
9046 	case SCHED_RR:
9047 		ret = MAX_RT_PRIO-1;
9048 		break;
9049 	case SCHED_DEADLINE:
9050 	case SCHED_NORMAL:
9051 	case SCHED_BATCH:
9052 	case SCHED_IDLE:
9053 		ret = 0;
9054 		break;
9055 	}
9056 	return ret;
9057 }
9058 
9059 /**
9060  * sys_sched_get_priority_min - return minimum RT priority.
9061  * @policy: scheduling class.
9062  *
9063  * Return: On success, this syscall returns the minimum
9064  * rt_priority that can be used by a given scheduling class.
9065  * On failure, a negative error code is returned.
9066  */
SYSCALL_DEFINE1(sched_get_priority_min,int,policy)9067 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
9068 {
9069 	int ret = -EINVAL;
9070 
9071 	switch (policy) {
9072 	case SCHED_FIFO:
9073 	case SCHED_RR:
9074 		ret = 1;
9075 		break;
9076 	case SCHED_DEADLINE:
9077 	case SCHED_NORMAL:
9078 	case SCHED_BATCH:
9079 	case SCHED_IDLE:
9080 		ret = 0;
9081 	}
9082 	return ret;
9083 }
9084 
sched_rr_get_interval(pid_t pid,struct timespec64 * t)9085 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
9086 {
9087 	struct task_struct *p;
9088 	unsigned int time_slice;
9089 	struct rq_flags rf;
9090 	struct rq *rq;
9091 	int retval;
9092 
9093 	if (pid < 0)
9094 		return -EINVAL;
9095 
9096 	retval = -ESRCH;
9097 	rcu_read_lock();
9098 	p = find_process_by_pid(pid);
9099 	if (!p)
9100 		goto out_unlock;
9101 
9102 	retval = security_task_getscheduler(p);
9103 	if (retval)
9104 		goto out_unlock;
9105 
9106 	rq = task_rq_lock(p, &rf);
9107 	time_slice = 0;
9108 	if (p->sched_class->get_rr_interval)
9109 		time_slice = p->sched_class->get_rr_interval(rq, p);
9110 	task_rq_unlock(rq, p, &rf);
9111 
9112 	rcu_read_unlock();
9113 	jiffies_to_timespec64(time_slice, t);
9114 	return 0;
9115 
9116 out_unlock:
9117 	rcu_read_unlock();
9118 	return retval;
9119 }
9120 
9121 /**
9122  * sys_sched_rr_get_interval - return the default timeslice of a process.
9123  * @pid: pid of the process.
9124  * @interval: userspace pointer to the timeslice value.
9125  *
9126  * this syscall writes the default timeslice value of a given process
9127  * into the user-space timespec buffer. A value of '0' means infinity.
9128  *
9129  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
9130  * an error code.
9131  */
SYSCALL_DEFINE2(sched_rr_get_interval,pid_t,pid,struct __kernel_timespec __user *,interval)9132 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
9133 		struct __kernel_timespec __user *, interval)
9134 {
9135 	struct timespec64 t;
9136 	int retval = sched_rr_get_interval(pid, &t);
9137 
9138 	if (retval == 0)
9139 		retval = put_timespec64(&t, interval);
9140 
9141 	return retval;
9142 }
9143 
9144 #ifdef CONFIG_COMPAT_32BIT_TIME
SYSCALL_DEFINE2(sched_rr_get_interval_time32,pid_t,pid,struct old_timespec32 __user *,interval)9145 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
9146 		struct old_timespec32 __user *, interval)
9147 {
9148 	struct timespec64 t;
9149 	int retval = sched_rr_get_interval(pid, &t);
9150 
9151 	if (retval == 0)
9152 		retval = put_old_timespec32(&t, interval);
9153 	return retval;
9154 }
9155 #endif
9156 
sched_show_task(struct task_struct * p)9157 void sched_show_task(struct task_struct *p)
9158 {
9159 	unsigned long free = 0;
9160 	int ppid;
9161 
9162 	if (!try_get_task_stack(p))
9163 		return;
9164 
9165 	pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
9166 
9167 	if (task_is_running(p))
9168 		pr_cont("  running task    ");
9169 #ifdef CONFIG_DEBUG_STACK_USAGE
9170 	free = stack_not_used(p);
9171 #endif
9172 	ppid = 0;
9173 	rcu_read_lock();
9174 	if (pid_alive(p))
9175 		ppid = task_pid_nr(rcu_dereference(p->real_parent));
9176 	rcu_read_unlock();
9177 	pr_cont(" stack:%-5lu pid:%-5d ppid:%-6d flags:0x%08lx\n",
9178 		free, task_pid_nr(p), ppid,
9179 		read_task_thread_flags(p));
9180 
9181 	print_worker_info(KERN_INFO, p);
9182 	print_stop_info(KERN_INFO, p);
9183 	show_stack(p, NULL, KERN_INFO);
9184 	put_task_stack(p);
9185 }
9186 EXPORT_SYMBOL_GPL(sched_show_task);
9187 
9188 static inline bool
state_filter_match(unsigned long state_filter,struct task_struct * p)9189 state_filter_match(unsigned long state_filter, struct task_struct *p)
9190 {
9191 	unsigned int state = READ_ONCE(p->__state);
9192 
9193 	/* no filter, everything matches */
9194 	if (!state_filter)
9195 		return true;
9196 
9197 	/* filter, but doesn't match */
9198 	if (!(state & state_filter))
9199 		return false;
9200 
9201 	/*
9202 	 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9203 	 * TASK_KILLABLE).
9204 	 */
9205 	if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
9206 		return false;
9207 
9208 	return true;
9209 }
9210 
9211 
show_state_filter(unsigned int state_filter)9212 void show_state_filter(unsigned int state_filter)
9213 {
9214 	struct task_struct *g, *p;
9215 
9216 	rcu_read_lock();
9217 	for_each_process_thread(g, p) {
9218 		/*
9219 		 * reset the NMI-timeout, listing all files on a slow
9220 		 * console might take a lot of time:
9221 		 * Also, reset softlockup watchdogs on all CPUs, because
9222 		 * another CPU might be blocked waiting for us to process
9223 		 * an IPI.
9224 		 */
9225 		touch_nmi_watchdog();
9226 		touch_all_softlockup_watchdogs();
9227 		if (state_filter_match(state_filter, p))
9228 			sched_show_task(p);
9229 	}
9230 
9231 #ifdef CONFIG_SCHED_DEBUG
9232 	if (!state_filter)
9233 		sysrq_sched_debug_show();
9234 #endif
9235 	rcu_read_unlock();
9236 	/*
9237 	 * Only show locks if all tasks are dumped:
9238 	 */
9239 	if (!state_filter)
9240 		debug_show_all_locks();
9241 }
9242 
9243 /**
9244  * init_idle - set up an idle thread for a given CPU
9245  * @idle: task in question
9246  * @cpu: CPU the idle task belongs to
9247  *
9248  * NOTE: this function does not set the idle thread's NEED_RESCHED
9249  * flag, to make booting more robust.
9250  */
init_idle(struct task_struct * idle,int cpu)9251 void __init init_idle(struct task_struct *idle, int cpu)
9252 {
9253 #ifdef CONFIG_SMP
9254 	struct affinity_context ac = (struct affinity_context) {
9255 		.new_mask  = cpumask_of(cpu),
9256 		.flags     = 0,
9257 	};
9258 #endif
9259 	struct rq *rq = cpu_rq(cpu);
9260 	unsigned long flags;
9261 
9262 	__sched_fork(0, idle);
9263 
9264 	raw_spin_lock_irqsave(&idle->pi_lock, flags);
9265 	raw_spin_rq_lock(rq);
9266 
9267 	idle->__state = TASK_RUNNING;
9268 	idle->se.exec_start = sched_clock();
9269 	/*
9270 	 * PF_KTHREAD should already be set at this point; regardless, make it
9271 	 * look like a proper per-CPU kthread.
9272 	 */
9273 	idle->flags |= PF_KTHREAD | PF_NO_SETAFFINITY;
9274 	kthread_set_per_cpu(idle, cpu);
9275 
9276 #ifdef CONFIG_SMP
9277 	/*
9278 	 * It's possible that init_idle() gets called multiple times on a task,
9279 	 * in that case do_set_cpus_allowed() will not do the right thing.
9280 	 *
9281 	 * And since this is boot we can forgo the serialization.
9282 	 */
9283 	set_cpus_allowed_common(idle, &ac);
9284 #endif
9285 	/*
9286 	 * We're having a chicken and egg problem, even though we are
9287 	 * holding rq->lock, the CPU isn't yet set to this CPU so the
9288 	 * lockdep check in task_group() will fail.
9289 	 *
9290 	 * Similar case to sched_fork(). / Alternatively we could
9291 	 * use task_rq_lock() here and obtain the other rq->lock.
9292 	 *
9293 	 * Silence PROVE_RCU
9294 	 */
9295 	rcu_read_lock();
9296 	__set_task_cpu(idle, cpu);
9297 	rcu_read_unlock();
9298 
9299 	rq->idle = idle;
9300 	rcu_assign_pointer(rq->curr, idle);
9301 	idle->on_rq = TASK_ON_RQ_QUEUED;
9302 #ifdef CONFIG_SMP
9303 	idle->on_cpu = 1;
9304 #endif
9305 	raw_spin_rq_unlock(rq);
9306 	raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
9307 
9308 	/* Set the preempt count _outside_ the spinlocks! */
9309 	init_idle_preempt_count(idle, cpu);
9310 
9311 	/*
9312 	 * The idle tasks have their own, simple scheduling class:
9313 	 */
9314 	idle->sched_class = &idle_sched_class;
9315 	ftrace_graph_init_idle_task(idle, cpu);
9316 	vtime_init_idle(idle, cpu);
9317 #ifdef CONFIG_SMP
9318 	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9319 #endif
9320 }
9321 
9322 #ifdef CONFIG_SMP
9323 
cpuset_cpumask_can_shrink(const struct cpumask * cur,const struct cpumask * trial)9324 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9325 			      const struct cpumask *trial)
9326 {
9327 	int ret = 1;
9328 
9329 	if (cpumask_empty(cur))
9330 		return ret;
9331 
9332 	ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9333 
9334 	return ret;
9335 }
9336 
task_can_attach(struct task_struct * p)9337 int task_can_attach(struct task_struct *p)
9338 {
9339 	int ret = 0;
9340 
9341 	/*
9342 	 * Kthreads which disallow setaffinity shouldn't be moved
9343 	 * to a new cpuset; we don't want to change their CPU
9344 	 * affinity and isolating such threads by their set of
9345 	 * allowed nodes is unnecessary.  Thus, cpusets are not
9346 	 * applicable for such threads.  This prevents checking for
9347 	 * success of set_cpus_allowed_ptr() on all attached tasks
9348 	 * before cpus_mask may be changed.
9349 	 */
9350 	if (p->flags & PF_NO_SETAFFINITY)
9351 		ret = -EINVAL;
9352 
9353 	return ret;
9354 }
9355 
9356 bool sched_smp_initialized __read_mostly;
9357 
9358 #ifdef CONFIG_NUMA_BALANCING
9359 /* Migrate current task p to target_cpu */
migrate_task_to(struct task_struct * p,int target_cpu)9360 int migrate_task_to(struct task_struct *p, int target_cpu)
9361 {
9362 	struct migration_arg arg = { p, target_cpu };
9363 	int curr_cpu = task_cpu(p);
9364 
9365 	if (curr_cpu == target_cpu)
9366 		return 0;
9367 
9368 	if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9369 		return -EINVAL;
9370 
9371 	/* TODO: This is not properly updating schedstats */
9372 
9373 	trace_sched_move_numa(p, curr_cpu, target_cpu);
9374 	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9375 }
9376 
9377 /*
9378  * Requeue a task on a given node and accurately track the number of NUMA
9379  * tasks on the runqueues
9380  */
sched_setnuma(struct task_struct * p,int nid)9381 void sched_setnuma(struct task_struct *p, int nid)
9382 {
9383 	bool queued, running;
9384 	struct rq_flags rf;
9385 	struct rq *rq;
9386 
9387 	rq = task_rq_lock(p, &rf);
9388 	queued = task_on_rq_queued(p);
9389 	running = task_current(rq, p);
9390 
9391 	if (queued)
9392 		dequeue_task(rq, p, DEQUEUE_SAVE);
9393 	if (running)
9394 		put_prev_task(rq, p);
9395 
9396 	p->numa_preferred_nid = nid;
9397 
9398 	if (queued)
9399 		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9400 	if (running)
9401 		set_next_task(rq, p);
9402 	task_rq_unlock(rq, p, &rf);
9403 }
9404 #endif /* CONFIG_NUMA_BALANCING */
9405 
9406 #ifdef CONFIG_HOTPLUG_CPU
9407 /*
9408  * Ensure that the idle task is using init_mm right before its CPU goes
9409  * offline.
9410  */
idle_task_exit(void)9411 void idle_task_exit(void)
9412 {
9413 	struct mm_struct *mm = current->active_mm;
9414 
9415 	BUG_ON(cpu_online(smp_processor_id()));
9416 	BUG_ON(current != this_rq()->idle);
9417 
9418 	if (mm != &init_mm) {
9419 		switch_mm(mm, &init_mm, current);
9420 		finish_arch_post_lock_switch();
9421 	}
9422 
9423 	/* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9424 }
9425 
__balance_push_cpu_stop(void * arg)9426 static int __balance_push_cpu_stop(void *arg)
9427 {
9428 	struct task_struct *p = arg;
9429 	struct rq *rq = this_rq();
9430 	struct rq_flags rf;
9431 	int cpu;
9432 
9433 	raw_spin_lock_irq(&p->pi_lock);
9434 	rq_lock(rq, &rf);
9435 
9436 	update_rq_clock(rq);
9437 
9438 	if (task_rq(p) == rq && task_on_rq_queued(p)) {
9439 		cpu = select_fallback_rq(rq->cpu, p);
9440 		rq = __migrate_task(rq, &rf, p, cpu);
9441 	}
9442 
9443 	rq_unlock(rq, &rf);
9444 	raw_spin_unlock_irq(&p->pi_lock);
9445 
9446 	put_task_struct(p);
9447 
9448 	return 0;
9449 }
9450 
9451 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9452 
9453 /*
9454  * Ensure we only run per-cpu kthreads once the CPU goes !active.
9455  *
9456  * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9457  * effective when the hotplug motion is down.
9458  */
balance_push(struct rq * rq)9459 static void balance_push(struct rq *rq)
9460 {
9461 	struct task_struct *push_task = rq->curr;
9462 
9463 	lockdep_assert_rq_held(rq);
9464 
9465 	/*
9466 	 * Ensure the thing is persistent until balance_push_set(.on = false);
9467 	 */
9468 	rq->balance_callback = &balance_push_callback;
9469 
9470 	/*
9471 	 * Only active while going offline and when invoked on the outgoing
9472 	 * CPU.
9473 	 */
9474 	if (!cpu_dying(rq->cpu) || rq != this_rq())
9475 		return;
9476 
9477 	/*
9478 	 * Both the cpu-hotplug and stop task are in this case and are
9479 	 * required to complete the hotplug process.
9480 	 */
9481 	if (kthread_is_per_cpu(push_task) ||
9482 	    is_migration_disabled(push_task)) {
9483 
9484 		/*
9485 		 * If this is the idle task on the outgoing CPU try to wake
9486 		 * up the hotplug control thread which might wait for the
9487 		 * last task to vanish. The rcuwait_active() check is
9488 		 * accurate here because the waiter is pinned on this CPU
9489 		 * and can't obviously be running in parallel.
9490 		 *
9491 		 * On RT kernels this also has to check whether there are
9492 		 * pinned and scheduled out tasks on the runqueue. They
9493 		 * need to leave the migrate disabled section first.
9494 		 */
9495 		if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9496 		    rcuwait_active(&rq->hotplug_wait)) {
9497 			raw_spin_rq_unlock(rq);
9498 			rcuwait_wake_up(&rq->hotplug_wait);
9499 			raw_spin_rq_lock(rq);
9500 		}
9501 		return;
9502 	}
9503 
9504 	get_task_struct(push_task);
9505 	/*
9506 	 * Temporarily drop rq->lock such that we can wake-up the stop task.
9507 	 * Both preemption and IRQs are still disabled.
9508 	 */
9509 	preempt_disable();
9510 	raw_spin_rq_unlock(rq);
9511 	stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9512 			    this_cpu_ptr(&push_work));
9513 	preempt_enable();
9514 	/*
9515 	 * At this point need_resched() is true and we'll take the loop in
9516 	 * schedule(). The next pick is obviously going to be the stop task
9517 	 * which kthread_is_per_cpu() and will push this task away.
9518 	 */
9519 	raw_spin_rq_lock(rq);
9520 }
9521 
balance_push_set(int cpu,bool on)9522 static void balance_push_set(int cpu, bool on)
9523 {
9524 	struct rq *rq = cpu_rq(cpu);
9525 	struct rq_flags rf;
9526 
9527 	rq_lock_irqsave(rq, &rf);
9528 	if (on) {
9529 		WARN_ON_ONCE(rq->balance_callback);
9530 		rq->balance_callback = &balance_push_callback;
9531 	} else if (rq->balance_callback == &balance_push_callback) {
9532 		rq->balance_callback = NULL;
9533 	}
9534 	rq_unlock_irqrestore(rq, &rf);
9535 }
9536 
9537 /*
9538  * Invoked from a CPUs hotplug control thread after the CPU has been marked
9539  * inactive. All tasks which are not per CPU kernel threads are either
9540  * pushed off this CPU now via balance_push() or placed on a different CPU
9541  * during wakeup. Wait until the CPU is quiescent.
9542  */
balance_hotplug_wait(void)9543 static void balance_hotplug_wait(void)
9544 {
9545 	struct rq *rq = this_rq();
9546 
9547 	rcuwait_wait_event(&rq->hotplug_wait,
9548 			   rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9549 			   TASK_UNINTERRUPTIBLE);
9550 }
9551 
9552 #else
9553 
balance_push(struct rq * rq)9554 static inline void balance_push(struct rq *rq)
9555 {
9556 }
9557 
balance_push_set(int cpu,bool on)9558 static inline void balance_push_set(int cpu, bool on)
9559 {
9560 }
9561 
balance_hotplug_wait(void)9562 static inline void balance_hotplug_wait(void)
9563 {
9564 }
9565 
9566 #endif /* CONFIG_HOTPLUG_CPU */
9567 
set_rq_online(struct rq * rq)9568 void set_rq_online(struct rq *rq)
9569 {
9570 	if (!rq->online) {
9571 		const struct sched_class *class;
9572 
9573 		cpumask_set_cpu(rq->cpu, rq->rd->online);
9574 		rq->online = 1;
9575 
9576 		for_each_class(class) {
9577 			if (class->rq_online)
9578 				class->rq_online(rq);
9579 		}
9580 	}
9581 }
9582 
set_rq_offline(struct rq * rq)9583 void set_rq_offline(struct rq *rq)
9584 {
9585 	if (rq->online) {
9586 		const struct sched_class *class;
9587 
9588 		update_rq_clock(rq);
9589 		for_each_class(class) {
9590 			if (class->rq_offline)
9591 				class->rq_offline(rq);
9592 		}
9593 
9594 		cpumask_clear_cpu(rq->cpu, rq->rd->online);
9595 		rq->online = 0;
9596 	}
9597 }
9598 
9599 /*
9600  * used to mark begin/end of suspend/resume:
9601  */
9602 static int num_cpus_frozen;
9603 
9604 /*
9605  * Update cpusets according to cpu_active mask.  If cpusets are
9606  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9607  * around partition_sched_domains().
9608  *
9609  * If we come here as part of a suspend/resume, don't touch cpusets because we
9610  * want to restore it back to its original state upon resume anyway.
9611  */
cpuset_cpu_active(void)9612 static void cpuset_cpu_active(void)
9613 {
9614 	if (cpuhp_tasks_frozen) {
9615 		/*
9616 		 * num_cpus_frozen tracks how many CPUs are involved in suspend
9617 		 * resume sequence. As long as this is not the last online
9618 		 * operation in the resume sequence, just build a single sched
9619 		 * domain, ignoring cpusets.
9620 		 */
9621 		partition_sched_domains(1, NULL, NULL);
9622 		if (--num_cpus_frozen)
9623 			return;
9624 		/*
9625 		 * This is the last CPU online operation. So fall through and
9626 		 * restore the original sched domains by considering the
9627 		 * cpuset configurations.
9628 		 */
9629 		cpuset_force_rebuild();
9630 	}
9631 	cpuset_update_active_cpus();
9632 }
9633 
cpuset_cpu_inactive(unsigned int cpu)9634 static int cpuset_cpu_inactive(unsigned int cpu)
9635 {
9636 	if (!cpuhp_tasks_frozen) {
9637 		int ret = dl_bw_check_overflow(cpu);
9638 
9639 		if (ret)
9640 			return ret;
9641 		cpuset_update_active_cpus();
9642 	} else {
9643 		num_cpus_frozen++;
9644 		partition_sched_domains(1, NULL, NULL);
9645 	}
9646 	return 0;
9647 }
9648 
sched_cpu_activate(unsigned int cpu)9649 int sched_cpu_activate(unsigned int cpu)
9650 {
9651 	struct rq *rq = cpu_rq(cpu);
9652 	struct rq_flags rf;
9653 
9654 	/*
9655 	 * Clear the balance_push callback and prepare to schedule
9656 	 * regular tasks.
9657 	 */
9658 	balance_push_set(cpu, false);
9659 
9660 #ifdef CONFIG_SCHED_SMT
9661 	/*
9662 	 * When going up, increment the number of cores with SMT present.
9663 	 */
9664 	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9665 		static_branch_inc_cpuslocked(&sched_smt_present);
9666 #endif
9667 	set_cpu_active(cpu, true);
9668 
9669 	if (sched_smp_initialized) {
9670 		sched_update_numa(cpu, true);
9671 		sched_domains_numa_masks_set(cpu);
9672 		cpuset_cpu_active();
9673 	}
9674 
9675 	/*
9676 	 * Put the rq online, if not already. This happens:
9677 	 *
9678 	 * 1) In the early boot process, because we build the real domains
9679 	 *    after all CPUs have been brought up.
9680 	 *
9681 	 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9682 	 *    domains.
9683 	 */
9684 	rq_lock_irqsave(rq, &rf);
9685 	if (rq->rd) {
9686 		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9687 		set_rq_online(rq);
9688 	}
9689 	rq_unlock_irqrestore(rq, &rf);
9690 
9691 	return 0;
9692 }
9693 
sched_cpu_deactivate(unsigned int cpu)9694 int sched_cpu_deactivate(unsigned int cpu)
9695 {
9696 	struct rq *rq = cpu_rq(cpu);
9697 	struct rq_flags rf;
9698 	int ret;
9699 
9700 	/*
9701 	 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9702 	 * load balancing when not active
9703 	 */
9704 	nohz_balance_exit_idle(rq);
9705 
9706 	set_cpu_active(cpu, false);
9707 
9708 	/*
9709 	 * From this point forward, this CPU will refuse to run any task that
9710 	 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9711 	 * push those tasks away until this gets cleared, see
9712 	 * sched_cpu_dying().
9713 	 */
9714 	balance_push_set(cpu, true);
9715 
9716 	/*
9717 	 * We've cleared cpu_active_mask / set balance_push, wait for all
9718 	 * preempt-disabled and RCU users of this state to go away such that
9719 	 * all new such users will observe it.
9720 	 *
9721 	 * Specifically, we rely on ttwu to no longer target this CPU, see
9722 	 * ttwu_queue_cond() and is_cpu_allowed().
9723 	 *
9724 	 * Do sync before park smpboot threads to take care the rcu boost case.
9725 	 */
9726 	synchronize_rcu();
9727 
9728 	rq_lock_irqsave(rq, &rf);
9729 	if (rq->rd) {
9730 		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9731 		set_rq_offline(rq);
9732 	}
9733 	rq_unlock_irqrestore(rq, &rf);
9734 
9735 #ifdef CONFIG_SCHED_SMT
9736 	/*
9737 	 * When going down, decrement the number of cores with SMT present.
9738 	 */
9739 	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9740 		static_branch_dec_cpuslocked(&sched_smt_present);
9741 
9742 	sched_core_cpu_deactivate(cpu);
9743 #endif
9744 
9745 	if (!sched_smp_initialized)
9746 		return 0;
9747 
9748 	sched_update_numa(cpu, false);
9749 	ret = cpuset_cpu_inactive(cpu);
9750 	if (ret) {
9751 		balance_push_set(cpu, false);
9752 		set_cpu_active(cpu, true);
9753 		sched_update_numa(cpu, true);
9754 		return ret;
9755 	}
9756 	sched_domains_numa_masks_clear(cpu);
9757 	return 0;
9758 }
9759 
sched_rq_cpu_starting(unsigned int cpu)9760 static void sched_rq_cpu_starting(unsigned int cpu)
9761 {
9762 	struct rq *rq = cpu_rq(cpu);
9763 
9764 	rq->calc_load_update = calc_load_update;
9765 	update_max_interval();
9766 }
9767 
sched_cpu_starting(unsigned int cpu)9768 int sched_cpu_starting(unsigned int cpu)
9769 {
9770 	sched_core_cpu_starting(cpu);
9771 	sched_rq_cpu_starting(cpu);
9772 	sched_tick_start(cpu);
9773 	return 0;
9774 }
9775 
9776 #ifdef CONFIG_HOTPLUG_CPU
9777 
9778 /*
9779  * Invoked immediately before the stopper thread is invoked to bring the
9780  * CPU down completely. At this point all per CPU kthreads except the
9781  * hotplug thread (current) and the stopper thread (inactive) have been
9782  * either parked or have been unbound from the outgoing CPU. Ensure that
9783  * any of those which might be on the way out are gone.
9784  *
9785  * If after this point a bound task is being woken on this CPU then the
9786  * responsible hotplug callback has failed to do it's job.
9787  * sched_cpu_dying() will catch it with the appropriate fireworks.
9788  */
sched_cpu_wait_empty(unsigned int cpu)9789 int sched_cpu_wait_empty(unsigned int cpu)
9790 {
9791 	balance_hotplug_wait();
9792 	return 0;
9793 }
9794 
9795 /*
9796  * Since this CPU is going 'away' for a while, fold any nr_active delta we
9797  * might have. Called from the CPU stopper task after ensuring that the
9798  * stopper is the last running task on the CPU, so nr_active count is
9799  * stable. We need to take the teardown thread which is calling this into
9800  * account, so we hand in adjust = 1 to the load calculation.
9801  *
9802  * Also see the comment "Global load-average calculations".
9803  */
calc_load_migrate(struct rq * rq)9804 static void calc_load_migrate(struct rq *rq)
9805 {
9806 	long delta = calc_load_fold_active(rq, 1);
9807 
9808 	if (delta)
9809 		atomic_long_add(delta, &calc_load_tasks);
9810 }
9811 
dump_rq_tasks(struct rq * rq,const char * loglvl)9812 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9813 {
9814 	struct task_struct *g, *p;
9815 	int cpu = cpu_of(rq);
9816 
9817 	lockdep_assert_rq_held(rq);
9818 
9819 	printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9820 	for_each_process_thread(g, p) {
9821 		if (task_cpu(p) != cpu)
9822 			continue;
9823 
9824 		if (!task_on_rq_queued(p))
9825 			continue;
9826 
9827 		printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9828 	}
9829 }
9830 
sched_cpu_dying(unsigned int cpu)9831 int sched_cpu_dying(unsigned int cpu)
9832 {
9833 	struct rq *rq = cpu_rq(cpu);
9834 	struct rq_flags rf;
9835 
9836 	/* Handle pending wakeups and then migrate everything off */
9837 	sched_tick_stop(cpu);
9838 
9839 	rq_lock_irqsave(rq, &rf);
9840 	if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9841 		WARN(true, "Dying CPU not properly vacated!");
9842 		dump_rq_tasks(rq, KERN_WARNING);
9843 	}
9844 	rq_unlock_irqrestore(rq, &rf);
9845 
9846 	calc_load_migrate(rq);
9847 	update_max_interval();
9848 	hrtick_clear(rq);
9849 	sched_core_cpu_dying(cpu);
9850 	return 0;
9851 }
9852 #endif
9853 
sched_init_smp(void)9854 void __init sched_init_smp(void)
9855 {
9856 	sched_init_numa(NUMA_NO_NODE);
9857 
9858 	/*
9859 	 * There's no userspace yet to cause hotplug operations; hence all the
9860 	 * CPU masks are stable and all blatant races in the below code cannot
9861 	 * happen.
9862 	 */
9863 	mutex_lock(&sched_domains_mutex);
9864 	sched_init_domains(cpu_active_mask);
9865 	mutex_unlock(&sched_domains_mutex);
9866 
9867 	/* Move init over to a non-isolated CPU */
9868 	if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9869 		BUG();
9870 	current->flags &= ~PF_NO_SETAFFINITY;
9871 	sched_init_granularity();
9872 
9873 	init_sched_rt_class();
9874 	init_sched_dl_class();
9875 
9876 	sched_smp_initialized = true;
9877 }
9878 
migration_init(void)9879 static int __init migration_init(void)
9880 {
9881 	sched_cpu_starting(smp_processor_id());
9882 	return 0;
9883 }
9884 early_initcall(migration_init);
9885 
9886 #else
sched_init_smp(void)9887 void __init sched_init_smp(void)
9888 {
9889 	sched_init_granularity();
9890 }
9891 #endif /* CONFIG_SMP */
9892 
in_sched_functions(unsigned long addr)9893 int in_sched_functions(unsigned long addr)
9894 {
9895 	return in_lock_functions(addr) ||
9896 		(addr >= (unsigned long)__sched_text_start
9897 		&& addr < (unsigned long)__sched_text_end);
9898 }
9899 
9900 #ifdef CONFIG_CGROUP_SCHED
9901 /*
9902  * Default task group.
9903  * Every task in system belongs to this group at bootup.
9904  */
9905 struct task_group root_task_group;
9906 LIST_HEAD(task_groups);
9907 
9908 /* Cacheline aligned slab cache for task_group */
9909 static struct kmem_cache *task_group_cache __read_mostly;
9910 #endif
9911 
sched_init(void)9912 void __init sched_init(void)
9913 {
9914 	unsigned long ptr = 0;
9915 	int i;
9916 
9917 	/* Make sure the linker didn't screw up */
9918 	BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9919 	       &fair_sched_class != &rt_sched_class + 1 ||
9920 	       &rt_sched_class   != &dl_sched_class + 1);
9921 #ifdef CONFIG_SMP
9922 	BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9923 #endif
9924 
9925 	wait_bit_init();
9926 
9927 #ifdef CONFIG_FAIR_GROUP_SCHED
9928 	ptr += 2 * nr_cpu_ids * sizeof(void **);
9929 #endif
9930 #ifdef CONFIG_RT_GROUP_SCHED
9931 	ptr += 2 * nr_cpu_ids * sizeof(void **);
9932 #endif
9933 	if (ptr) {
9934 		ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9935 
9936 #ifdef CONFIG_FAIR_GROUP_SCHED
9937 		root_task_group.se = (struct sched_entity **)ptr;
9938 		ptr += nr_cpu_ids * sizeof(void **);
9939 
9940 		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9941 		ptr += nr_cpu_ids * sizeof(void **);
9942 
9943 		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9944 		init_cfs_bandwidth(&root_task_group.cfs_bandwidth, NULL);
9945 #endif /* CONFIG_FAIR_GROUP_SCHED */
9946 #ifdef CONFIG_RT_GROUP_SCHED
9947 		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9948 		ptr += nr_cpu_ids * sizeof(void **);
9949 
9950 		root_task_group.rt_rq = (struct rt_rq **)ptr;
9951 		ptr += nr_cpu_ids * sizeof(void **);
9952 
9953 #endif /* CONFIG_RT_GROUP_SCHED */
9954 	}
9955 
9956 	init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9957 
9958 #ifdef CONFIG_SMP
9959 	init_defrootdomain();
9960 #endif
9961 
9962 #ifdef CONFIG_RT_GROUP_SCHED
9963 	init_rt_bandwidth(&root_task_group.rt_bandwidth,
9964 			global_rt_period(), global_rt_runtime());
9965 #endif /* CONFIG_RT_GROUP_SCHED */
9966 
9967 #ifdef CONFIG_CGROUP_SCHED
9968 	task_group_cache = KMEM_CACHE(task_group, 0);
9969 
9970 	list_add(&root_task_group.list, &task_groups);
9971 	INIT_LIST_HEAD(&root_task_group.children);
9972 	INIT_LIST_HEAD(&root_task_group.siblings);
9973 	autogroup_init(&init_task);
9974 #endif /* CONFIG_CGROUP_SCHED */
9975 
9976 	for_each_possible_cpu(i) {
9977 		struct rq *rq;
9978 
9979 		rq = cpu_rq(i);
9980 		raw_spin_lock_init(&rq->__lock);
9981 		rq->nr_running = 0;
9982 		rq->calc_load_active = 0;
9983 		rq->calc_load_update = jiffies + LOAD_FREQ;
9984 		init_cfs_rq(&rq->cfs);
9985 		init_rt_rq(&rq->rt);
9986 		init_dl_rq(&rq->dl);
9987 #ifdef CONFIG_FAIR_GROUP_SCHED
9988 		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9989 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9990 		/*
9991 		 * How much CPU bandwidth does root_task_group get?
9992 		 *
9993 		 * In case of task-groups formed thr' the cgroup filesystem, it
9994 		 * gets 100% of the CPU resources in the system. This overall
9995 		 * system CPU resource is divided among the tasks of
9996 		 * root_task_group and its child task-groups in a fair manner,
9997 		 * based on each entity's (task or task-group's) weight
9998 		 * (se->load.weight).
9999 		 *
10000 		 * In other words, if root_task_group has 10 tasks of weight
10001 		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
10002 		 * then A0's share of the CPU resource is:
10003 		 *
10004 		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
10005 		 *
10006 		 * We achieve this by letting root_task_group's tasks sit
10007 		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
10008 		 */
10009 		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
10010 #endif /* CONFIG_FAIR_GROUP_SCHED */
10011 
10012 		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
10013 #ifdef CONFIG_RT_GROUP_SCHED
10014 		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
10015 #endif
10016 #ifdef CONFIG_SMP
10017 		rq->sd = NULL;
10018 		rq->rd = NULL;
10019 		rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
10020 		rq->balance_callback = &balance_push_callback;
10021 		rq->active_balance = 0;
10022 		rq->next_balance = jiffies;
10023 		rq->push_cpu = 0;
10024 		rq->cpu = i;
10025 		rq->online = 0;
10026 		rq->idle_stamp = 0;
10027 		rq->avg_idle = 2*sysctl_sched_migration_cost;
10028 		rq->wake_stamp = jiffies;
10029 		rq->wake_avg_idle = rq->avg_idle;
10030 		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
10031 
10032 		INIT_LIST_HEAD(&rq->cfs_tasks);
10033 
10034 		rq_attach_root(rq, &def_root_domain);
10035 #ifdef CONFIG_NO_HZ_COMMON
10036 		rq->last_blocked_load_update_tick = jiffies;
10037 		atomic_set(&rq->nohz_flags, 0);
10038 
10039 		INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
10040 #endif
10041 #ifdef CONFIG_HOTPLUG_CPU
10042 		rcuwait_init(&rq->hotplug_wait);
10043 #endif
10044 #endif /* CONFIG_SMP */
10045 		hrtick_rq_init(rq);
10046 		atomic_set(&rq->nr_iowait, 0);
10047 
10048 #ifdef CONFIG_SCHED_CORE
10049 		rq->core = rq;
10050 		rq->core_pick = NULL;
10051 		rq->core_enabled = 0;
10052 		rq->core_tree = RB_ROOT;
10053 		rq->core_forceidle_count = 0;
10054 		rq->core_forceidle_occupation = 0;
10055 		rq->core_forceidle_start = 0;
10056 
10057 		rq->core_cookie = 0UL;
10058 #endif
10059 		zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
10060 	}
10061 
10062 	set_load_weight(&init_task, false);
10063 
10064 	/*
10065 	 * The boot idle thread does lazy MMU switching as well:
10066 	 */
10067 	mmgrab_lazy_tlb(&init_mm);
10068 	enter_lazy_tlb(&init_mm, current);
10069 
10070 	/*
10071 	 * The idle task doesn't need the kthread struct to function, but it
10072 	 * is dressed up as a per-CPU kthread and thus needs to play the part
10073 	 * if we want to avoid special-casing it in code that deals with per-CPU
10074 	 * kthreads.
10075 	 */
10076 	WARN_ON(!set_kthread_struct(current));
10077 
10078 	/*
10079 	 * Make us the idle thread. Technically, schedule() should not be
10080 	 * called from this thread, however somewhere below it might be,
10081 	 * but because we are the idle thread, we just pick up running again
10082 	 * when this runqueue becomes "idle".
10083 	 */
10084 	init_idle(current, smp_processor_id());
10085 
10086 	calc_load_update = jiffies + LOAD_FREQ;
10087 
10088 #ifdef CONFIG_SMP
10089 	idle_thread_set_boot_cpu();
10090 	balance_push_set(smp_processor_id(), false);
10091 #endif
10092 	init_sched_fair_class();
10093 
10094 	psi_init();
10095 
10096 	init_uclamp();
10097 
10098 	preempt_dynamic_init();
10099 
10100 	scheduler_running = 1;
10101 }
10102 
10103 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
10104 
__might_sleep(const char * file,int line)10105 void __might_sleep(const char *file, int line)
10106 {
10107 	unsigned int state = get_current_state();
10108 	/*
10109 	 * Blocking primitives will set (and therefore destroy) current->state,
10110 	 * since we will exit with TASK_RUNNING make sure we enter with it,
10111 	 * otherwise we will destroy state.
10112 	 */
10113 	WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
10114 			"do not call blocking ops when !TASK_RUNNING; "
10115 			"state=%x set at [<%p>] %pS\n", state,
10116 			(void *)current->task_state_change,
10117 			(void *)current->task_state_change);
10118 
10119 	__might_resched(file, line, 0);
10120 }
10121 EXPORT_SYMBOL(__might_sleep);
10122 
print_preempt_disable_ip(int preempt_offset,unsigned long ip)10123 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
10124 {
10125 	if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
10126 		return;
10127 
10128 	if (preempt_count() == preempt_offset)
10129 		return;
10130 
10131 	pr_err("Preemption disabled at:");
10132 	print_ip_sym(KERN_ERR, ip);
10133 }
10134 
resched_offsets_ok(unsigned int offsets)10135 static inline bool resched_offsets_ok(unsigned int offsets)
10136 {
10137 	unsigned int nested = preempt_count();
10138 
10139 	nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
10140 
10141 	return nested == offsets;
10142 }
10143 
__might_resched(const char * file,int line,unsigned int offsets)10144 void __might_resched(const char *file, int line, unsigned int offsets)
10145 {
10146 	/* Ratelimiting timestamp: */
10147 	static unsigned long prev_jiffy;
10148 
10149 	unsigned long preempt_disable_ip;
10150 
10151 	/* WARN_ON_ONCE() by default, no rate limit required: */
10152 	rcu_sleep_check();
10153 
10154 	if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
10155 	     !is_idle_task(current) && !current->non_block_count) ||
10156 	    system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10157 	    oops_in_progress)
10158 		return;
10159 
10160 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10161 		return;
10162 	prev_jiffy = jiffies;
10163 
10164 	/* Save this before calling printk(), since that will clobber it: */
10165 	preempt_disable_ip = get_preempt_disable_ip(current);
10166 
10167 	pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10168 	       file, line);
10169 	pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10170 	       in_atomic(), irqs_disabled(), current->non_block_count,
10171 	       current->pid, current->comm);
10172 	pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10173 	       offsets & MIGHT_RESCHED_PREEMPT_MASK);
10174 
10175 	if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
10176 		pr_err("RCU nest depth: %d, expected: %u\n",
10177 		       rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
10178 	}
10179 
10180 	if (task_stack_end_corrupted(current))
10181 		pr_emerg("Thread overran stack, or stack corrupted\n");
10182 
10183 	debug_show_held_locks(current);
10184 	if (irqs_disabled())
10185 		print_irqtrace_events(current);
10186 
10187 	print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
10188 				 preempt_disable_ip);
10189 
10190 	dump_stack();
10191 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10192 }
10193 EXPORT_SYMBOL(__might_resched);
10194 
__cant_sleep(const char * file,int line,int preempt_offset)10195 void __cant_sleep(const char *file, int line, int preempt_offset)
10196 {
10197 	static unsigned long prev_jiffy;
10198 
10199 	if (irqs_disabled())
10200 		return;
10201 
10202 	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10203 		return;
10204 
10205 	if (preempt_count() > preempt_offset)
10206 		return;
10207 
10208 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10209 		return;
10210 	prev_jiffy = jiffies;
10211 
10212 	printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10213 	printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10214 			in_atomic(), irqs_disabled(),
10215 			current->pid, current->comm);
10216 
10217 	debug_show_held_locks(current);
10218 	dump_stack();
10219 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10220 }
10221 EXPORT_SYMBOL_GPL(__cant_sleep);
10222 
10223 #ifdef CONFIG_SMP
__cant_migrate(const char * file,int line)10224 void __cant_migrate(const char *file, int line)
10225 {
10226 	static unsigned long prev_jiffy;
10227 
10228 	if (irqs_disabled())
10229 		return;
10230 
10231 	if (is_migration_disabled(current))
10232 		return;
10233 
10234 	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10235 		return;
10236 
10237 	if (preempt_count() > 0)
10238 		return;
10239 
10240 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10241 		return;
10242 	prev_jiffy = jiffies;
10243 
10244 	pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10245 	pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10246 	       in_atomic(), irqs_disabled(), is_migration_disabled(current),
10247 	       current->pid, current->comm);
10248 
10249 	debug_show_held_locks(current);
10250 	dump_stack();
10251 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10252 }
10253 EXPORT_SYMBOL_GPL(__cant_migrate);
10254 #endif
10255 #endif
10256 
10257 #ifdef CONFIG_MAGIC_SYSRQ
normalize_rt_tasks(void)10258 void normalize_rt_tasks(void)
10259 {
10260 	struct task_struct *g, *p;
10261 	struct sched_attr attr = {
10262 		.sched_policy = SCHED_NORMAL,
10263 	};
10264 
10265 	read_lock(&tasklist_lock);
10266 	for_each_process_thread(g, p) {
10267 		/*
10268 		 * Only normalize user tasks:
10269 		 */
10270 		if (p->flags & PF_KTHREAD)
10271 			continue;
10272 
10273 		p->se.exec_start = 0;
10274 		schedstat_set(p->stats.wait_start,  0);
10275 		schedstat_set(p->stats.sleep_start, 0);
10276 		schedstat_set(p->stats.block_start, 0);
10277 
10278 		if (!dl_task(p) && !rt_task(p)) {
10279 			/*
10280 			 * Renice negative nice level userspace
10281 			 * tasks back to 0:
10282 			 */
10283 			if (task_nice(p) < 0)
10284 				set_user_nice(p, 0);
10285 			continue;
10286 		}
10287 
10288 		__sched_setscheduler(p, &attr, false, false);
10289 	}
10290 	read_unlock(&tasklist_lock);
10291 }
10292 
10293 #endif /* CONFIG_MAGIC_SYSRQ */
10294 
10295 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
10296 /*
10297  * These functions are only useful for the IA64 MCA handling, or kdb.
10298  *
10299  * They can only be called when the whole system has been
10300  * stopped - every CPU needs to be quiescent, and no scheduling
10301  * activity can take place. Using them for anything else would
10302  * be a serious bug, and as a result, they aren't even visible
10303  * under any other configuration.
10304  */
10305 
10306 /**
10307  * curr_task - return the current task for a given CPU.
10308  * @cpu: the processor in question.
10309  *
10310  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10311  *
10312  * Return: The current task for @cpu.
10313  */
curr_task(int cpu)10314 struct task_struct *curr_task(int cpu)
10315 {
10316 	return cpu_curr(cpu);
10317 }
10318 
10319 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10320 
10321 #ifdef CONFIG_IA64
10322 /**
10323  * ia64_set_curr_task - set the current task for a given CPU.
10324  * @cpu: the processor in question.
10325  * @p: the task pointer to set.
10326  *
10327  * Description: This function must only be used when non-maskable interrupts
10328  * are serviced on a separate stack. It allows the architecture to switch the
10329  * notion of the current task on a CPU in a non-blocking manner. This function
10330  * must be called with all CPU's synchronized, and interrupts disabled, the
10331  * and caller must save the original value of the current task (see
10332  * curr_task() above) and restore that value before reenabling interrupts and
10333  * re-starting the system.
10334  *
10335  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10336  */
ia64_set_curr_task(int cpu,struct task_struct * p)10337 void ia64_set_curr_task(int cpu, struct task_struct *p)
10338 {
10339 	cpu_curr(cpu) = p;
10340 }
10341 
10342 #endif
10343 
10344 #ifdef CONFIG_CGROUP_SCHED
10345 /* task_group_lock serializes the addition/removal of task groups */
10346 static DEFINE_SPINLOCK(task_group_lock);
10347 
alloc_uclamp_sched_group(struct task_group * tg,struct task_group * parent)10348 static inline void alloc_uclamp_sched_group(struct task_group *tg,
10349 					    struct task_group *parent)
10350 {
10351 #ifdef CONFIG_UCLAMP_TASK_GROUP
10352 	enum uclamp_id clamp_id;
10353 
10354 	for_each_clamp_id(clamp_id) {
10355 		uclamp_se_set(&tg->uclamp_req[clamp_id],
10356 			      uclamp_none(clamp_id), false);
10357 		tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10358 	}
10359 #endif
10360 }
10361 
sched_free_group(struct task_group * tg)10362 static void sched_free_group(struct task_group *tg)
10363 {
10364 	free_fair_sched_group(tg);
10365 	free_rt_sched_group(tg);
10366 	autogroup_free(tg);
10367 	kmem_cache_free(task_group_cache, tg);
10368 }
10369 
sched_free_group_rcu(struct rcu_head * rcu)10370 static void sched_free_group_rcu(struct rcu_head *rcu)
10371 {
10372 	sched_free_group(container_of(rcu, struct task_group, rcu));
10373 }
10374 
sched_unregister_group(struct task_group * tg)10375 static void sched_unregister_group(struct task_group *tg)
10376 {
10377 	unregister_fair_sched_group(tg);
10378 	unregister_rt_sched_group(tg);
10379 	/*
10380 	 * We have to wait for yet another RCU grace period to expire, as
10381 	 * print_cfs_stats() might run concurrently.
10382 	 */
10383 	call_rcu(&tg->rcu, sched_free_group_rcu);
10384 }
10385 
10386 /* allocate runqueue etc for a new task group */
sched_create_group(struct task_group * parent)10387 struct task_group *sched_create_group(struct task_group *parent)
10388 {
10389 	struct task_group *tg;
10390 
10391 	tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10392 	if (!tg)
10393 		return ERR_PTR(-ENOMEM);
10394 
10395 	if (!alloc_fair_sched_group(tg, parent))
10396 		goto err;
10397 
10398 	if (!alloc_rt_sched_group(tg, parent))
10399 		goto err;
10400 
10401 	alloc_uclamp_sched_group(tg, parent);
10402 
10403 	return tg;
10404 
10405 err:
10406 	sched_free_group(tg);
10407 	return ERR_PTR(-ENOMEM);
10408 }
10409 
sched_online_group(struct task_group * tg,struct task_group * parent)10410 void sched_online_group(struct task_group *tg, struct task_group *parent)
10411 {
10412 	unsigned long flags;
10413 
10414 	spin_lock_irqsave(&task_group_lock, flags);
10415 	list_add_rcu(&tg->list, &task_groups);
10416 
10417 	/* Root should already exist: */
10418 	WARN_ON(!parent);
10419 
10420 	tg->parent = parent;
10421 	INIT_LIST_HEAD(&tg->children);
10422 	list_add_rcu(&tg->siblings, &parent->children);
10423 	spin_unlock_irqrestore(&task_group_lock, flags);
10424 
10425 	online_fair_sched_group(tg);
10426 }
10427 
10428 /* rcu callback to free various structures associated with a task group */
sched_unregister_group_rcu(struct rcu_head * rhp)10429 static void sched_unregister_group_rcu(struct rcu_head *rhp)
10430 {
10431 	/* Now it should be safe to free those cfs_rqs: */
10432 	sched_unregister_group(container_of(rhp, struct task_group, rcu));
10433 }
10434 
sched_destroy_group(struct task_group * tg)10435 void sched_destroy_group(struct task_group *tg)
10436 {
10437 	/* Wait for possible concurrent references to cfs_rqs complete: */
10438 	call_rcu(&tg->rcu, sched_unregister_group_rcu);
10439 }
10440 
sched_release_group(struct task_group * tg)10441 void sched_release_group(struct task_group *tg)
10442 {
10443 	unsigned long flags;
10444 
10445 	/*
10446 	 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10447 	 * sched_cfs_period_timer()).
10448 	 *
10449 	 * For this to be effective, we have to wait for all pending users of
10450 	 * this task group to leave their RCU critical section to ensure no new
10451 	 * user will see our dying task group any more. Specifically ensure
10452 	 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10453 	 *
10454 	 * We therefore defer calling unregister_fair_sched_group() to
10455 	 * sched_unregister_group() which is guarantied to get called only after the
10456 	 * current RCU grace period has expired.
10457 	 */
10458 	spin_lock_irqsave(&task_group_lock, flags);
10459 	list_del_rcu(&tg->list);
10460 	list_del_rcu(&tg->siblings);
10461 	spin_unlock_irqrestore(&task_group_lock, flags);
10462 }
10463 
sched_get_task_group(struct task_struct * tsk)10464 static struct task_group *sched_get_task_group(struct task_struct *tsk)
10465 {
10466 	struct task_group *tg;
10467 
10468 	/*
10469 	 * All callers are synchronized by task_rq_lock(); we do not use RCU
10470 	 * which is pointless here. Thus, we pass "true" to task_css_check()
10471 	 * to prevent lockdep warnings.
10472 	 */
10473 	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10474 			  struct task_group, css);
10475 	tg = autogroup_task_group(tsk, tg);
10476 
10477 	return tg;
10478 }
10479 
sched_change_group(struct task_struct * tsk,struct task_group * group)10480 static void sched_change_group(struct task_struct *tsk, struct task_group *group)
10481 {
10482 	tsk->sched_task_group = group;
10483 
10484 #ifdef CONFIG_FAIR_GROUP_SCHED
10485 	if (tsk->sched_class->task_change_group)
10486 		tsk->sched_class->task_change_group(tsk);
10487 	else
10488 #endif
10489 		set_task_rq(tsk, task_cpu(tsk));
10490 }
10491 
10492 /*
10493  * Change task's runqueue when it moves between groups.
10494  *
10495  * The caller of this function should have put the task in its new group by
10496  * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10497  * its new group.
10498  */
sched_move_task(struct task_struct * tsk)10499 void sched_move_task(struct task_struct *tsk)
10500 {
10501 	int queued, running, queue_flags =
10502 		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10503 	struct task_group *group;
10504 	struct rq_flags rf;
10505 	struct rq *rq;
10506 
10507 	rq = task_rq_lock(tsk, &rf);
10508 	/*
10509 	 * Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
10510 	 * group changes.
10511 	 */
10512 	group = sched_get_task_group(tsk);
10513 	if (group == tsk->sched_task_group)
10514 		goto unlock;
10515 
10516 	update_rq_clock(rq);
10517 
10518 	running = task_current(rq, tsk);
10519 	queued = task_on_rq_queued(tsk);
10520 
10521 	if (queued)
10522 		dequeue_task(rq, tsk, queue_flags);
10523 	if (running)
10524 		put_prev_task(rq, tsk);
10525 
10526 	sched_change_group(tsk, group);
10527 
10528 	if (queued)
10529 		enqueue_task(rq, tsk, queue_flags);
10530 	if (running) {
10531 		set_next_task(rq, tsk);
10532 		/*
10533 		 * After changing group, the running task may have joined a
10534 		 * throttled one but it's still the running task. Trigger a
10535 		 * resched to make sure that task can still run.
10536 		 */
10537 		resched_curr(rq);
10538 	}
10539 
10540 unlock:
10541 	task_rq_unlock(rq, tsk, &rf);
10542 }
10543 
css_tg(struct cgroup_subsys_state * css)10544 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10545 {
10546 	return css ? container_of(css, struct task_group, css) : NULL;
10547 }
10548 
10549 static struct cgroup_subsys_state *
cpu_cgroup_css_alloc(struct cgroup_subsys_state * parent_css)10550 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10551 {
10552 	struct task_group *parent = css_tg(parent_css);
10553 	struct task_group *tg;
10554 
10555 	if (!parent) {
10556 		/* This is early initialization for the top cgroup */
10557 		return &root_task_group.css;
10558 	}
10559 
10560 	tg = sched_create_group(parent);
10561 	if (IS_ERR(tg))
10562 		return ERR_PTR(-ENOMEM);
10563 
10564 	return &tg->css;
10565 }
10566 
10567 /* Expose task group only after completing cgroup initialization */
cpu_cgroup_css_online(struct cgroup_subsys_state * css)10568 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10569 {
10570 	struct task_group *tg = css_tg(css);
10571 	struct task_group *parent = css_tg(css->parent);
10572 
10573 	if (parent)
10574 		sched_online_group(tg, parent);
10575 
10576 #ifdef CONFIG_UCLAMP_TASK_GROUP
10577 	/* Propagate the effective uclamp value for the new group */
10578 	mutex_lock(&uclamp_mutex);
10579 	rcu_read_lock();
10580 	cpu_util_update_eff(css);
10581 	rcu_read_unlock();
10582 	mutex_unlock(&uclamp_mutex);
10583 #endif
10584 
10585 	return 0;
10586 }
10587 
cpu_cgroup_css_released(struct cgroup_subsys_state * css)10588 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10589 {
10590 	struct task_group *tg = css_tg(css);
10591 
10592 	sched_release_group(tg);
10593 }
10594 
cpu_cgroup_css_free(struct cgroup_subsys_state * css)10595 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10596 {
10597 	struct task_group *tg = css_tg(css);
10598 
10599 	/*
10600 	 * Relies on the RCU grace period between css_released() and this.
10601 	 */
10602 	sched_unregister_group(tg);
10603 }
10604 
10605 #ifdef CONFIG_RT_GROUP_SCHED
cpu_cgroup_can_attach(struct cgroup_taskset * tset)10606 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10607 {
10608 	struct task_struct *task;
10609 	struct cgroup_subsys_state *css;
10610 
10611 	cgroup_taskset_for_each(task, css, tset) {
10612 		if (!sched_rt_can_attach(css_tg(css), task))
10613 			return -EINVAL;
10614 	}
10615 	return 0;
10616 }
10617 #endif
10618 
cpu_cgroup_attach(struct cgroup_taskset * tset)10619 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10620 {
10621 	struct task_struct *task;
10622 	struct cgroup_subsys_state *css;
10623 
10624 	cgroup_taskset_for_each(task, css, tset)
10625 		sched_move_task(task);
10626 }
10627 
10628 #ifdef CONFIG_UCLAMP_TASK_GROUP
cpu_util_update_eff(struct cgroup_subsys_state * css)10629 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10630 {
10631 	struct cgroup_subsys_state *top_css = css;
10632 	struct uclamp_se *uc_parent = NULL;
10633 	struct uclamp_se *uc_se = NULL;
10634 	unsigned int eff[UCLAMP_CNT];
10635 	enum uclamp_id clamp_id;
10636 	unsigned int clamps;
10637 
10638 	lockdep_assert_held(&uclamp_mutex);
10639 	SCHED_WARN_ON(!rcu_read_lock_held());
10640 
10641 	css_for_each_descendant_pre(css, top_css) {
10642 		uc_parent = css_tg(css)->parent
10643 			? css_tg(css)->parent->uclamp : NULL;
10644 
10645 		for_each_clamp_id(clamp_id) {
10646 			/* Assume effective clamps matches requested clamps */
10647 			eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10648 			/* Cap effective clamps with parent's effective clamps */
10649 			if (uc_parent &&
10650 			    eff[clamp_id] > uc_parent[clamp_id].value) {
10651 				eff[clamp_id] = uc_parent[clamp_id].value;
10652 			}
10653 		}
10654 		/* Ensure protection is always capped by limit */
10655 		eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10656 
10657 		/* Propagate most restrictive effective clamps */
10658 		clamps = 0x0;
10659 		uc_se = css_tg(css)->uclamp;
10660 		for_each_clamp_id(clamp_id) {
10661 			if (eff[clamp_id] == uc_se[clamp_id].value)
10662 				continue;
10663 			uc_se[clamp_id].value = eff[clamp_id];
10664 			uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10665 			clamps |= (0x1 << clamp_id);
10666 		}
10667 		if (!clamps) {
10668 			css = css_rightmost_descendant(css);
10669 			continue;
10670 		}
10671 
10672 		/* Immediately update descendants RUNNABLE tasks */
10673 		uclamp_update_active_tasks(css);
10674 	}
10675 }
10676 
10677 /*
10678  * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10679  * C expression. Since there is no way to convert a macro argument (N) into a
10680  * character constant, use two levels of macros.
10681  */
10682 #define _POW10(exp) ((unsigned int)1e##exp)
10683 #define POW10(exp) _POW10(exp)
10684 
10685 struct uclamp_request {
10686 #define UCLAMP_PERCENT_SHIFT	2
10687 #define UCLAMP_PERCENT_SCALE	(100 * POW10(UCLAMP_PERCENT_SHIFT))
10688 	s64 percent;
10689 	u64 util;
10690 	int ret;
10691 };
10692 
10693 static inline struct uclamp_request
capacity_from_percent(char * buf)10694 capacity_from_percent(char *buf)
10695 {
10696 	struct uclamp_request req = {
10697 		.percent = UCLAMP_PERCENT_SCALE,
10698 		.util = SCHED_CAPACITY_SCALE,
10699 		.ret = 0,
10700 	};
10701 
10702 	buf = strim(buf);
10703 	if (strcmp(buf, "max")) {
10704 		req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10705 					     &req.percent);
10706 		if (req.ret)
10707 			return req;
10708 		if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10709 			req.ret = -ERANGE;
10710 			return req;
10711 		}
10712 
10713 		req.util = req.percent << SCHED_CAPACITY_SHIFT;
10714 		req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10715 	}
10716 
10717 	return req;
10718 }
10719 
cpu_uclamp_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off,enum uclamp_id clamp_id)10720 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10721 				size_t nbytes, loff_t off,
10722 				enum uclamp_id clamp_id)
10723 {
10724 	struct uclamp_request req;
10725 	struct task_group *tg;
10726 
10727 	req = capacity_from_percent(buf);
10728 	if (req.ret)
10729 		return req.ret;
10730 
10731 	static_branch_enable(&sched_uclamp_used);
10732 
10733 	mutex_lock(&uclamp_mutex);
10734 	rcu_read_lock();
10735 
10736 	tg = css_tg(of_css(of));
10737 	if (tg->uclamp_req[clamp_id].value != req.util)
10738 		uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10739 
10740 	/*
10741 	 * Because of not recoverable conversion rounding we keep track of the
10742 	 * exact requested value
10743 	 */
10744 	tg->uclamp_pct[clamp_id] = req.percent;
10745 
10746 	/* Update effective clamps to track the most restrictive value */
10747 	cpu_util_update_eff(of_css(of));
10748 
10749 	rcu_read_unlock();
10750 	mutex_unlock(&uclamp_mutex);
10751 
10752 	return nbytes;
10753 }
10754 
cpu_uclamp_min_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)10755 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10756 				    char *buf, size_t nbytes,
10757 				    loff_t off)
10758 {
10759 	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10760 }
10761 
cpu_uclamp_max_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)10762 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10763 				    char *buf, size_t nbytes,
10764 				    loff_t off)
10765 {
10766 	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10767 }
10768 
cpu_uclamp_print(struct seq_file * sf,enum uclamp_id clamp_id)10769 static inline void cpu_uclamp_print(struct seq_file *sf,
10770 				    enum uclamp_id clamp_id)
10771 {
10772 	struct task_group *tg;
10773 	u64 util_clamp;
10774 	u64 percent;
10775 	u32 rem;
10776 
10777 	rcu_read_lock();
10778 	tg = css_tg(seq_css(sf));
10779 	util_clamp = tg->uclamp_req[clamp_id].value;
10780 	rcu_read_unlock();
10781 
10782 	if (util_clamp == SCHED_CAPACITY_SCALE) {
10783 		seq_puts(sf, "max\n");
10784 		return;
10785 	}
10786 
10787 	percent = tg->uclamp_pct[clamp_id];
10788 	percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10789 	seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10790 }
10791 
cpu_uclamp_min_show(struct seq_file * sf,void * v)10792 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10793 {
10794 	cpu_uclamp_print(sf, UCLAMP_MIN);
10795 	return 0;
10796 }
10797 
cpu_uclamp_max_show(struct seq_file * sf,void * v)10798 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10799 {
10800 	cpu_uclamp_print(sf, UCLAMP_MAX);
10801 	return 0;
10802 }
10803 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10804 
10805 #ifdef CONFIG_FAIR_GROUP_SCHED
cpu_shares_write_u64(struct cgroup_subsys_state * css,struct cftype * cftype,u64 shareval)10806 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10807 				struct cftype *cftype, u64 shareval)
10808 {
10809 	if (shareval > scale_load_down(ULONG_MAX))
10810 		shareval = MAX_SHARES;
10811 	return sched_group_set_shares(css_tg(css), scale_load(shareval));
10812 }
10813 
cpu_shares_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)10814 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10815 			       struct cftype *cft)
10816 {
10817 	struct task_group *tg = css_tg(css);
10818 
10819 	return (u64) scale_load_down(tg->shares);
10820 }
10821 
10822 #ifdef CONFIG_CFS_BANDWIDTH
10823 static DEFINE_MUTEX(cfs_constraints_mutex);
10824 
10825 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10826 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10827 /* More than 203 days if BW_SHIFT equals 20. */
10828 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10829 
10830 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10831 
tg_set_cfs_bandwidth(struct task_group * tg,u64 period,u64 quota,u64 burst)10832 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10833 				u64 burst)
10834 {
10835 	int i, ret = 0, runtime_enabled, runtime_was_enabled;
10836 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10837 
10838 	if (tg == &root_task_group)
10839 		return -EINVAL;
10840 
10841 	/*
10842 	 * Ensure we have at some amount of bandwidth every period.  This is
10843 	 * to prevent reaching a state of large arrears when throttled via
10844 	 * entity_tick() resulting in prolonged exit starvation.
10845 	 */
10846 	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10847 		return -EINVAL;
10848 
10849 	/*
10850 	 * Likewise, bound things on the other side by preventing insane quota
10851 	 * periods.  This also allows us to normalize in computing quota
10852 	 * feasibility.
10853 	 */
10854 	if (period > max_cfs_quota_period)
10855 		return -EINVAL;
10856 
10857 	/*
10858 	 * Bound quota to defend quota against overflow during bandwidth shift.
10859 	 */
10860 	if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10861 		return -EINVAL;
10862 
10863 	if (quota != RUNTIME_INF && (burst > quota ||
10864 				     burst + quota > max_cfs_runtime))
10865 		return -EINVAL;
10866 
10867 	/*
10868 	 * Prevent race between setting of cfs_rq->runtime_enabled and
10869 	 * unthrottle_offline_cfs_rqs().
10870 	 */
10871 	cpus_read_lock();
10872 	mutex_lock(&cfs_constraints_mutex);
10873 	ret = __cfs_schedulable(tg, period, quota);
10874 	if (ret)
10875 		goto out_unlock;
10876 
10877 	runtime_enabled = quota != RUNTIME_INF;
10878 	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10879 	/*
10880 	 * If we need to toggle cfs_bandwidth_used, off->on must occur
10881 	 * before making related changes, and on->off must occur afterwards
10882 	 */
10883 	if (runtime_enabled && !runtime_was_enabled)
10884 		cfs_bandwidth_usage_inc();
10885 	raw_spin_lock_irq(&cfs_b->lock);
10886 	cfs_b->period = ns_to_ktime(period);
10887 	cfs_b->quota = quota;
10888 	cfs_b->burst = burst;
10889 
10890 	__refill_cfs_bandwidth_runtime(cfs_b);
10891 
10892 	/* Restart the period timer (if active) to handle new period expiry: */
10893 	if (runtime_enabled)
10894 		start_cfs_bandwidth(cfs_b);
10895 
10896 	raw_spin_unlock_irq(&cfs_b->lock);
10897 
10898 	for_each_online_cpu(i) {
10899 		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10900 		struct rq *rq = cfs_rq->rq;
10901 		struct rq_flags rf;
10902 
10903 		rq_lock_irq(rq, &rf);
10904 		cfs_rq->runtime_enabled = runtime_enabled;
10905 		cfs_rq->runtime_remaining = 0;
10906 
10907 		if (cfs_rq->throttled)
10908 			unthrottle_cfs_rq(cfs_rq);
10909 		rq_unlock_irq(rq, &rf);
10910 	}
10911 	if (runtime_was_enabled && !runtime_enabled)
10912 		cfs_bandwidth_usage_dec();
10913 out_unlock:
10914 	mutex_unlock(&cfs_constraints_mutex);
10915 	cpus_read_unlock();
10916 
10917 	return ret;
10918 }
10919 
tg_set_cfs_quota(struct task_group * tg,long cfs_quota_us)10920 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10921 {
10922 	u64 quota, period, burst;
10923 
10924 	period = ktime_to_ns(tg->cfs_bandwidth.period);
10925 	burst = tg->cfs_bandwidth.burst;
10926 	if (cfs_quota_us < 0)
10927 		quota = RUNTIME_INF;
10928 	else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10929 		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10930 	else
10931 		return -EINVAL;
10932 
10933 	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10934 }
10935 
tg_get_cfs_quota(struct task_group * tg)10936 static long tg_get_cfs_quota(struct task_group *tg)
10937 {
10938 	u64 quota_us;
10939 
10940 	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10941 		return -1;
10942 
10943 	quota_us = tg->cfs_bandwidth.quota;
10944 	do_div(quota_us, NSEC_PER_USEC);
10945 
10946 	return quota_us;
10947 }
10948 
tg_set_cfs_period(struct task_group * tg,long cfs_period_us)10949 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10950 {
10951 	u64 quota, period, burst;
10952 
10953 	if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10954 		return -EINVAL;
10955 
10956 	period = (u64)cfs_period_us * NSEC_PER_USEC;
10957 	quota = tg->cfs_bandwidth.quota;
10958 	burst = tg->cfs_bandwidth.burst;
10959 
10960 	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10961 }
10962 
tg_get_cfs_period(struct task_group * tg)10963 static long tg_get_cfs_period(struct task_group *tg)
10964 {
10965 	u64 cfs_period_us;
10966 
10967 	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10968 	do_div(cfs_period_us, NSEC_PER_USEC);
10969 
10970 	return cfs_period_us;
10971 }
10972 
tg_set_cfs_burst(struct task_group * tg,long cfs_burst_us)10973 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10974 {
10975 	u64 quota, period, burst;
10976 
10977 	if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10978 		return -EINVAL;
10979 
10980 	burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10981 	period = ktime_to_ns(tg->cfs_bandwidth.period);
10982 	quota = tg->cfs_bandwidth.quota;
10983 
10984 	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10985 }
10986 
tg_get_cfs_burst(struct task_group * tg)10987 static long tg_get_cfs_burst(struct task_group *tg)
10988 {
10989 	u64 burst_us;
10990 
10991 	burst_us = tg->cfs_bandwidth.burst;
10992 	do_div(burst_us, NSEC_PER_USEC);
10993 
10994 	return burst_us;
10995 }
10996 
cpu_cfs_quota_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)10997 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10998 				  struct cftype *cft)
10999 {
11000 	return tg_get_cfs_quota(css_tg(css));
11001 }
11002 
cpu_cfs_quota_write_s64(struct cgroup_subsys_state * css,struct cftype * cftype,s64 cfs_quota_us)11003 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
11004 				   struct cftype *cftype, s64 cfs_quota_us)
11005 {
11006 	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
11007 }
11008 
cpu_cfs_period_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)11009 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
11010 				   struct cftype *cft)
11011 {
11012 	return tg_get_cfs_period(css_tg(css));
11013 }
11014 
cpu_cfs_period_write_u64(struct cgroup_subsys_state * css,struct cftype * cftype,u64 cfs_period_us)11015 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
11016 				    struct cftype *cftype, u64 cfs_period_us)
11017 {
11018 	return tg_set_cfs_period(css_tg(css), cfs_period_us);
11019 }
11020 
cpu_cfs_burst_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)11021 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
11022 				  struct cftype *cft)
11023 {
11024 	return tg_get_cfs_burst(css_tg(css));
11025 }
11026 
cpu_cfs_burst_write_u64(struct cgroup_subsys_state * css,struct cftype * cftype,u64 cfs_burst_us)11027 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
11028 				   struct cftype *cftype, u64 cfs_burst_us)
11029 {
11030 	return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
11031 }
11032 
11033 struct cfs_schedulable_data {
11034 	struct task_group *tg;
11035 	u64 period, quota;
11036 };
11037 
11038 /*
11039  * normalize group quota/period to be quota/max_period
11040  * note: units are usecs
11041  */
normalize_cfs_quota(struct task_group * tg,struct cfs_schedulable_data * d)11042 static u64 normalize_cfs_quota(struct task_group *tg,
11043 			       struct cfs_schedulable_data *d)
11044 {
11045 	u64 quota, period;
11046 
11047 	if (tg == d->tg) {
11048 		period = d->period;
11049 		quota = d->quota;
11050 	} else {
11051 		period = tg_get_cfs_period(tg);
11052 		quota = tg_get_cfs_quota(tg);
11053 	}
11054 
11055 	/* note: these should typically be equivalent */
11056 	if (quota == RUNTIME_INF || quota == -1)
11057 		return RUNTIME_INF;
11058 
11059 	return to_ratio(period, quota);
11060 }
11061 
tg_cfs_schedulable_down(struct task_group * tg,void * data)11062 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
11063 {
11064 	struct cfs_schedulable_data *d = data;
11065 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11066 	s64 quota = 0, parent_quota = -1;
11067 
11068 	if (!tg->parent) {
11069 		quota = RUNTIME_INF;
11070 	} else {
11071 		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
11072 
11073 		quota = normalize_cfs_quota(tg, d);
11074 		parent_quota = parent_b->hierarchical_quota;
11075 
11076 		/*
11077 		 * Ensure max(child_quota) <= parent_quota.  On cgroup2,
11078 		 * always take the non-RUNTIME_INF min.  On cgroup1, only
11079 		 * inherit when no limit is set. In both cases this is used
11080 		 * by the scheduler to determine if a given CFS task has a
11081 		 * bandwidth constraint at some higher level.
11082 		 */
11083 		if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
11084 			if (quota == RUNTIME_INF)
11085 				quota = parent_quota;
11086 			else if (parent_quota != RUNTIME_INF)
11087 				quota = min(quota, parent_quota);
11088 		} else {
11089 			if (quota == RUNTIME_INF)
11090 				quota = parent_quota;
11091 			else if (parent_quota != RUNTIME_INF && quota > parent_quota)
11092 				return -EINVAL;
11093 		}
11094 	}
11095 	cfs_b->hierarchical_quota = quota;
11096 
11097 	return 0;
11098 }
11099 
__cfs_schedulable(struct task_group * tg,u64 period,u64 quota)11100 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
11101 {
11102 	int ret;
11103 	struct cfs_schedulable_data data = {
11104 		.tg = tg,
11105 		.period = period,
11106 		.quota = quota,
11107 	};
11108 
11109 	if (quota != RUNTIME_INF) {
11110 		do_div(data.period, NSEC_PER_USEC);
11111 		do_div(data.quota, NSEC_PER_USEC);
11112 	}
11113 
11114 	rcu_read_lock();
11115 	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
11116 	rcu_read_unlock();
11117 
11118 	return ret;
11119 }
11120 
cpu_cfs_stat_show(struct seq_file * sf,void * v)11121 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
11122 {
11123 	struct task_group *tg = css_tg(seq_css(sf));
11124 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11125 
11126 	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
11127 	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
11128 	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
11129 
11130 	if (schedstat_enabled() && tg != &root_task_group) {
11131 		struct sched_statistics *stats;
11132 		u64 ws = 0;
11133 		int i;
11134 
11135 		for_each_possible_cpu(i) {
11136 			stats = __schedstats_from_se(tg->se[i]);
11137 			ws += schedstat_val(stats->wait_sum);
11138 		}
11139 
11140 		seq_printf(sf, "wait_sum %llu\n", ws);
11141 	}
11142 
11143 	seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
11144 	seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
11145 
11146 	return 0;
11147 }
11148 
throttled_time_self(struct task_group * tg)11149 static u64 throttled_time_self(struct task_group *tg)
11150 {
11151 	int i;
11152 	u64 total = 0;
11153 
11154 	for_each_possible_cpu(i) {
11155 		total += READ_ONCE(tg->cfs_rq[i]->throttled_clock_self_time);
11156 	}
11157 
11158 	return total;
11159 }
11160 
cpu_cfs_local_stat_show(struct seq_file * sf,void * v)11161 static int cpu_cfs_local_stat_show(struct seq_file *sf, void *v)
11162 {
11163 	struct task_group *tg = css_tg(seq_css(sf));
11164 
11165 	seq_printf(sf, "throttled_time %llu\n", throttled_time_self(tg));
11166 
11167 	return 0;
11168 }
11169 #endif /* CONFIG_CFS_BANDWIDTH */
11170 #endif /* CONFIG_FAIR_GROUP_SCHED */
11171 
11172 #ifdef CONFIG_RT_GROUP_SCHED
cpu_rt_runtime_write(struct cgroup_subsys_state * css,struct cftype * cft,s64 val)11173 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
11174 				struct cftype *cft, s64 val)
11175 {
11176 	return sched_group_set_rt_runtime(css_tg(css), val);
11177 }
11178 
cpu_rt_runtime_read(struct cgroup_subsys_state * css,struct cftype * cft)11179 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
11180 			       struct cftype *cft)
11181 {
11182 	return sched_group_rt_runtime(css_tg(css));
11183 }
11184 
cpu_rt_period_write_uint(struct cgroup_subsys_state * css,struct cftype * cftype,u64 rt_period_us)11185 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
11186 				    struct cftype *cftype, u64 rt_period_us)
11187 {
11188 	return sched_group_set_rt_period(css_tg(css), rt_period_us);
11189 }
11190 
cpu_rt_period_read_uint(struct cgroup_subsys_state * css,struct cftype * cft)11191 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11192 				   struct cftype *cft)
11193 {
11194 	return sched_group_rt_period(css_tg(css));
11195 }
11196 #endif /* CONFIG_RT_GROUP_SCHED */
11197 
11198 #ifdef CONFIG_FAIR_GROUP_SCHED
cpu_idle_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)11199 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11200 			       struct cftype *cft)
11201 {
11202 	return css_tg(css)->idle;
11203 }
11204 
cpu_idle_write_s64(struct cgroup_subsys_state * css,struct cftype * cft,s64 idle)11205 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11206 				struct cftype *cft, s64 idle)
11207 {
11208 	return sched_group_set_idle(css_tg(css), idle);
11209 }
11210 #endif
11211 
11212 static struct cftype cpu_legacy_files[] = {
11213 #ifdef CONFIG_FAIR_GROUP_SCHED
11214 	{
11215 		.name = "shares",
11216 		.read_u64 = cpu_shares_read_u64,
11217 		.write_u64 = cpu_shares_write_u64,
11218 	},
11219 	{
11220 		.name = "idle",
11221 		.read_s64 = cpu_idle_read_s64,
11222 		.write_s64 = cpu_idle_write_s64,
11223 	},
11224 #endif
11225 #ifdef CONFIG_CFS_BANDWIDTH
11226 	{
11227 		.name = "cfs_quota_us",
11228 		.read_s64 = cpu_cfs_quota_read_s64,
11229 		.write_s64 = cpu_cfs_quota_write_s64,
11230 	},
11231 	{
11232 		.name = "cfs_period_us",
11233 		.read_u64 = cpu_cfs_period_read_u64,
11234 		.write_u64 = cpu_cfs_period_write_u64,
11235 	},
11236 	{
11237 		.name = "cfs_burst_us",
11238 		.read_u64 = cpu_cfs_burst_read_u64,
11239 		.write_u64 = cpu_cfs_burst_write_u64,
11240 	},
11241 	{
11242 		.name = "stat",
11243 		.seq_show = cpu_cfs_stat_show,
11244 	},
11245 	{
11246 		.name = "stat.local",
11247 		.seq_show = cpu_cfs_local_stat_show,
11248 	},
11249 #endif
11250 #ifdef CONFIG_RT_GROUP_SCHED
11251 	{
11252 		.name = "rt_runtime_us",
11253 		.read_s64 = cpu_rt_runtime_read,
11254 		.write_s64 = cpu_rt_runtime_write,
11255 	},
11256 	{
11257 		.name = "rt_period_us",
11258 		.read_u64 = cpu_rt_period_read_uint,
11259 		.write_u64 = cpu_rt_period_write_uint,
11260 	},
11261 #endif
11262 #ifdef CONFIG_UCLAMP_TASK_GROUP
11263 	{
11264 		.name = "uclamp.min",
11265 		.flags = CFTYPE_NOT_ON_ROOT,
11266 		.seq_show = cpu_uclamp_min_show,
11267 		.write = cpu_uclamp_min_write,
11268 	},
11269 	{
11270 		.name = "uclamp.max",
11271 		.flags = CFTYPE_NOT_ON_ROOT,
11272 		.seq_show = cpu_uclamp_max_show,
11273 		.write = cpu_uclamp_max_write,
11274 	},
11275 #endif
11276 	{ }	/* Terminate */
11277 };
11278 
cpu_extra_stat_show(struct seq_file * sf,struct cgroup_subsys_state * css)11279 static int cpu_extra_stat_show(struct seq_file *sf,
11280 			       struct cgroup_subsys_state *css)
11281 {
11282 #ifdef CONFIG_CFS_BANDWIDTH
11283 	{
11284 		struct task_group *tg = css_tg(css);
11285 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11286 		u64 throttled_usec, burst_usec;
11287 
11288 		throttled_usec = cfs_b->throttled_time;
11289 		do_div(throttled_usec, NSEC_PER_USEC);
11290 		burst_usec = cfs_b->burst_time;
11291 		do_div(burst_usec, NSEC_PER_USEC);
11292 
11293 		seq_printf(sf, "nr_periods %d\n"
11294 			   "nr_throttled %d\n"
11295 			   "throttled_usec %llu\n"
11296 			   "nr_bursts %d\n"
11297 			   "burst_usec %llu\n",
11298 			   cfs_b->nr_periods, cfs_b->nr_throttled,
11299 			   throttled_usec, cfs_b->nr_burst, burst_usec);
11300 	}
11301 #endif
11302 	return 0;
11303 }
11304 
cpu_local_stat_show(struct seq_file * sf,struct cgroup_subsys_state * css)11305 static int cpu_local_stat_show(struct seq_file *sf,
11306 			       struct cgroup_subsys_state *css)
11307 {
11308 #ifdef CONFIG_CFS_BANDWIDTH
11309 	{
11310 		struct task_group *tg = css_tg(css);
11311 		u64 throttled_self_usec;
11312 
11313 		throttled_self_usec = throttled_time_self(tg);
11314 		do_div(throttled_self_usec, NSEC_PER_USEC);
11315 
11316 		seq_printf(sf, "throttled_usec %llu\n",
11317 			   throttled_self_usec);
11318 	}
11319 #endif
11320 	return 0;
11321 }
11322 
11323 #ifdef CONFIG_FAIR_GROUP_SCHED
cpu_weight_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)11324 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11325 			       struct cftype *cft)
11326 {
11327 	struct task_group *tg = css_tg(css);
11328 	u64 weight = scale_load_down(tg->shares);
11329 
11330 	return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11331 }
11332 
cpu_weight_write_u64(struct cgroup_subsys_state * css,struct cftype * cft,u64 weight)11333 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11334 				struct cftype *cft, u64 weight)
11335 {
11336 	/*
11337 	 * cgroup weight knobs should use the common MIN, DFL and MAX
11338 	 * values which are 1, 100 and 10000 respectively.  While it loses
11339 	 * a bit of range on both ends, it maps pretty well onto the shares
11340 	 * value used by scheduler and the round-trip conversions preserve
11341 	 * the original value over the entire range.
11342 	 */
11343 	if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11344 		return -ERANGE;
11345 
11346 	weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11347 
11348 	return sched_group_set_shares(css_tg(css), scale_load(weight));
11349 }
11350 
cpu_weight_nice_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)11351 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11352 				    struct cftype *cft)
11353 {
11354 	unsigned long weight = scale_load_down(css_tg(css)->shares);
11355 	int last_delta = INT_MAX;
11356 	int prio, delta;
11357 
11358 	/* find the closest nice value to the current weight */
11359 	for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11360 		delta = abs(sched_prio_to_weight[prio] - weight);
11361 		if (delta >= last_delta)
11362 			break;
11363 		last_delta = delta;
11364 	}
11365 
11366 	return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11367 }
11368 
cpu_weight_nice_write_s64(struct cgroup_subsys_state * css,struct cftype * cft,s64 nice)11369 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11370 				     struct cftype *cft, s64 nice)
11371 {
11372 	unsigned long weight;
11373 	int idx;
11374 
11375 	if (nice < MIN_NICE || nice > MAX_NICE)
11376 		return -ERANGE;
11377 
11378 	idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11379 	idx = array_index_nospec(idx, 40);
11380 	weight = sched_prio_to_weight[idx];
11381 
11382 	return sched_group_set_shares(css_tg(css), scale_load(weight));
11383 }
11384 #endif
11385 
cpu_period_quota_print(struct seq_file * sf,long period,long quota)11386 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11387 						  long period, long quota)
11388 {
11389 	if (quota < 0)
11390 		seq_puts(sf, "max");
11391 	else
11392 		seq_printf(sf, "%ld", quota);
11393 
11394 	seq_printf(sf, " %ld\n", period);
11395 }
11396 
11397 /* caller should put the current value in *@periodp before calling */
cpu_period_quota_parse(char * buf,u64 * periodp,u64 * quotap)11398 static int __maybe_unused cpu_period_quota_parse(char *buf,
11399 						 u64 *periodp, u64 *quotap)
11400 {
11401 	char tok[21];	/* U64_MAX */
11402 
11403 	if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11404 		return -EINVAL;
11405 
11406 	*periodp *= NSEC_PER_USEC;
11407 
11408 	if (sscanf(tok, "%llu", quotap))
11409 		*quotap *= NSEC_PER_USEC;
11410 	else if (!strcmp(tok, "max"))
11411 		*quotap = RUNTIME_INF;
11412 	else
11413 		return -EINVAL;
11414 
11415 	return 0;
11416 }
11417 
11418 #ifdef CONFIG_CFS_BANDWIDTH
cpu_max_show(struct seq_file * sf,void * v)11419 static int cpu_max_show(struct seq_file *sf, void *v)
11420 {
11421 	struct task_group *tg = css_tg(seq_css(sf));
11422 
11423 	cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11424 	return 0;
11425 }
11426 
cpu_max_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)11427 static ssize_t cpu_max_write(struct kernfs_open_file *of,
11428 			     char *buf, size_t nbytes, loff_t off)
11429 {
11430 	struct task_group *tg = css_tg(of_css(of));
11431 	u64 period = tg_get_cfs_period(tg);
11432 	u64 burst = tg_get_cfs_burst(tg);
11433 	u64 quota;
11434 	int ret;
11435 
11436 	ret = cpu_period_quota_parse(buf, &period, &quota);
11437 	if (!ret)
11438 		ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11439 	return ret ?: nbytes;
11440 }
11441 #endif
11442 
11443 static struct cftype cpu_files[] = {
11444 #ifdef CONFIG_FAIR_GROUP_SCHED
11445 	{
11446 		.name = "weight",
11447 		.flags = CFTYPE_NOT_ON_ROOT,
11448 		.read_u64 = cpu_weight_read_u64,
11449 		.write_u64 = cpu_weight_write_u64,
11450 	},
11451 	{
11452 		.name = "weight.nice",
11453 		.flags = CFTYPE_NOT_ON_ROOT,
11454 		.read_s64 = cpu_weight_nice_read_s64,
11455 		.write_s64 = cpu_weight_nice_write_s64,
11456 	},
11457 	{
11458 		.name = "idle",
11459 		.flags = CFTYPE_NOT_ON_ROOT,
11460 		.read_s64 = cpu_idle_read_s64,
11461 		.write_s64 = cpu_idle_write_s64,
11462 	},
11463 #endif
11464 #ifdef CONFIG_CFS_BANDWIDTH
11465 	{
11466 		.name = "max",
11467 		.flags = CFTYPE_NOT_ON_ROOT,
11468 		.seq_show = cpu_max_show,
11469 		.write = cpu_max_write,
11470 	},
11471 	{
11472 		.name = "max.burst",
11473 		.flags = CFTYPE_NOT_ON_ROOT,
11474 		.read_u64 = cpu_cfs_burst_read_u64,
11475 		.write_u64 = cpu_cfs_burst_write_u64,
11476 	},
11477 #endif
11478 #ifdef CONFIG_UCLAMP_TASK_GROUP
11479 	{
11480 		.name = "uclamp.min",
11481 		.flags = CFTYPE_NOT_ON_ROOT,
11482 		.seq_show = cpu_uclamp_min_show,
11483 		.write = cpu_uclamp_min_write,
11484 	},
11485 	{
11486 		.name = "uclamp.max",
11487 		.flags = CFTYPE_NOT_ON_ROOT,
11488 		.seq_show = cpu_uclamp_max_show,
11489 		.write = cpu_uclamp_max_write,
11490 	},
11491 #endif
11492 	{ }	/* terminate */
11493 };
11494 
11495 struct cgroup_subsys cpu_cgrp_subsys = {
11496 	.css_alloc	= cpu_cgroup_css_alloc,
11497 	.css_online	= cpu_cgroup_css_online,
11498 	.css_released	= cpu_cgroup_css_released,
11499 	.css_free	= cpu_cgroup_css_free,
11500 	.css_extra_stat_show = cpu_extra_stat_show,
11501 	.css_local_stat_show = cpu_local_stat_show,
11502 #ifdef CONFIG_RT_GROUP_SCHED
11503 	.can_attach	= cpu_cgroup_can_attach,
11504 #endif
11505 	.attach		= cpu_cgroup_attach,
11506 	.legacy_cftypes	= cpu_legacy_files,
11507 	.dfl_cftypes	= cpu_files,
11508 	.early_init	= true,
11509 	.threaded	= true,
11510 };
11511 
11512 #endif	/* CONFIG_CGROUP_SCHED */
11513 
dump_cpu_task(int cpu)11514 void dump_cpu_task(int cpu)
11515 {
11516 	if (cpu == smp_processor_id() && in_hardirq()) {
11517 		struct pt_regs *regs;
11518 
11519 		regs = get_irq_regs();
11520 		if (regs) {
11521 			show_regs(regs);
11522 			return;
11523 		}
11524 	}
11525 
11526 	if (trigger_single_cpu_backtrace(cpu))
11527 		return;
11528 
11529 	pr_info("Task dump for CPU %d:\n", cpu);
11530 	sched_show_task(cpu_curr(cpu));
11531 }
11532 
11533 /*
11534  * Nice levels are multiplicative, with a gentle 10% change for every
11535  * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11536  * nice 1, it will get ~10% less CPU time than another CPU-bound task
11537  * that remained on nice 0.
11538  *
11539  * The "10% effect" is relative and cumulative: from _any_ nice level,
11540  * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11541  * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11542  * If a task goes up by ~10% and another task goes down by ~10% then
11543  * the relative distance between them is ~25%.)
11544  */
11545 const int sched_prio_to_weight[40] = {
11546  /* -20 */     88761,     71755,     56483,     46273,     36291,
11547  /* -15 */     29154,     23254,     18705,     14949,     11916,
11548  /* -10 */      9548,      7620,      6100,      4904,      3906,
11549  /*  -5 */      3121,      2501,      1991,      1586,      1277,
11550  /*   0 */      1024,       820,       655,       526,       423,
11551  /*   5 */       335,       272,       215,       172,       137,
11552  /*  10 */       110,        87,        70,        56,        45,
11553  /*  15 */        36,        29,        23,        18,        15,
11554 };
11555 
11556 /*
11557  * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11558  *
11559  * In cases where the weight does not change often, we can use the
11560  * precalculated inverse to speed up arithmetics by turning divisions
11561  * into multiplications:
11562  */
11563 const u32 sched_prio_to_wmult[40] = {
11564  /* -20 */     48388,     59856,     76040,     92818,    118348,
11565  /* -15 */    147320,    184698,    229616,    287308,    360437,
11566  /* -10 */    449829,    563644,    704093,    875809,   1099582,
11567  /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
11568  /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
11569  /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
11570  /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
11571  /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11572 };
11573 
call_trace_sched_update_nr_running(struct rq * rq,int count)11574 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11575 {
11576         trace_sched_update_nr_running_tp(rq, count);
11577 }
11578 
11579 #ifdef CONFIG_SCHED_MM_CID
11580 
11581 /*
11582  * @cid_lock: Guarantee forward-progress of cid allocation.
11583  *
11584  * Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
11585  * is only used when contention is detected by the lock-free allocation so
11586  * forward progress can be guaranteed.
11587  */
11588 DEFINE_RAW_SPINLOCK(cid_lock);
11589 
11590 /*
11591  * @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
11592  *
11593  * When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
11594  * detected, it is set to 1 to ensure that all newly coming allocations are
11595  * serialized by @cid_lock until the allocation which detected contention
11596  * completes and sets @use_cid_lock back to 0. This guarantees forward progress
11597  * of a cid allocation.
11598  */
11599 int use_cid_lock;
11600 
11601 /*
11602  * mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
11603  * concurrently with respect to the execution of the source runqueue context
11604  * switch.
11605  *
11606  * There is one basic properties we want to guarantee here:
11607  *
11608  * (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
11609  * used by a task. That would lead to concurrent allocation of the cid and
11610  * userspace corruption.
11611  *
11612  * Provide this guarantee by introducing a Dekker memory ordering to guarantee
11613  * that a pair of loads observe at least one of a pair of stores, which can be
11614  * shown as:
11615  *
11616  *      X = Y = 0
11617  *
11618  *      w[X]=1          w[Y]=1
11619  *      MB              MB
11620  *      r[Y]=y          r[X]=x
11621  *
11622  * Which guarantees that x==0 && y==0 is impossible. But rather than using
11623  * values 0 and 1, this algorithm cares about specific state transitions of the
11624  * runqueue current task (as updated by the scheduler context switch), and the
11625  * per-mm/cpu cid value.
11626  *
11627  * Let's introduce task (Y) which has task->mm == mm and task (N) which has
11628  * task->mm != mm for the rest of the discussion. There are two scheduler state
11629  * transitions on context switch we care about:
11630  *
11631  * (TSA) Store to rq->curr with transition from (N) to (Y)
11632  *
11633  * (TSB) Store to rq->curr with transition from (Y) to (N)
11634  *
11635  * On the remote-clear side, there is one transition we care about:
11636  *
11637  * (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
11638  *
11639  * There is also a transition to UNSET state which can be performed from all
11640  * sides (scheduler, remote-clear). It is always performed with a cmpxchg which
11641  * guarantees that only a single thread will succeed:
11642  *
11643  * (TMB) cmpxchg to *pcpu_cid to mark UNSET
11644  *
11645  * Just to be clear, what we do _not_ want to happen is a transition to UNSET
11646  * when a thread is actively using the cid (property (1)).
11647  *
11648  * Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
11649  *
11650  * Scenario A) (TSA)+(TMA) (from next task perspective)
11651  *
11652  * CPU0                                      CPU1
11653  *
11654  * Context switch CS-1                       Remote-clear
11655  *   - store to rq->curr: (N)->(Y) (TSA)     - cmpxchg to *pcpu_id to LAZY (TMA)
11656  *                                             (implied barrier after cmpxchg)
11657  *   - switch_mm_cid()
11658  *     - memory barrier (see switch_mm_cid()
11659  *       comment explaining how this barrier
11660  *       is combined with other scheduler
11661  *       barriers)
11662  *     - mm_cid_get (next)
11663  *       - READ_ONCE(*pcpu_cid)              - rcu_dereference(src_rq->curr)
11664  *
11665  * This Dekker ensures that either task (Y) is observed by the
11666  * rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
11667  * observed.
11668  *
11669  * If task (Y) store is observed by rcu_dereference(), it means that there is
11670  * still an active task on the cpu. Remote-clear will therefore not transition
11671  * to UNSET, which fulfills property (1).
11672  *
11673  * If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
11674  * it will move its state to UNSET, which clears the percpu cid perhaps
11675  * uselessly (which is not an issue for correctness). Because task (Y) is not
11676  * observed, CPU1 can move ahead to set the state to UNSET. Because moving
11677  * state to UNSET is done with a cmpxchg expecting that the old state has the
11678  * LAZY flag set, only one thread will successfully UNSET.
11679  *
11680  * If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
11681  * will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
11682  * CPU1 will observe task (Y) and do nothing more, which is fine.
11683  *
11684  * What we are effectively preventing with this Dekker is a scenario where
11685  * neither LAZY flag nor store (Y) are observed, which would fail property (1)
11686  * because this would UNSET a cid which is actively used.
11687  */
11688 
sched_mm_cid_migrate_from(struct task_struct * t)11689 void sched_mm_cid_migrate_from(struct task_struct *t)
11690 {
11691 	t->migrate_from_cpu = task_cpu(t);
11692 }
11693 
11694 static
__sched_mm_cid_migrate_from_fetch_cid(struct rq * src_rq,struct task_struct * t,struct mm_cid * src_pcpu_cid)11695 int __sched_mm_cid_migrate_from_fetch_cid(struct rq *src_rq,
11696 					  struct task_struct *t,
11697 					  struct mm_cid *src_pcpu_cid)
11698 {
11699 	struct mm_struct *mm = t->mm;
11700 	struct task_struct *src_task;
11701 	int src_cid, last_mm_cid;
11702 
11703 	if (!mm)
11704 		return -1;
11705 
11706 	last_mm_cid = t->last_mm_cid;
11707 	/*
11708 	 * If the migrated task has no last cid, or if the current
11709 	 * task on src rq uses the cid, it means the source cid does not need
11710 	 * to be moved to the destination cpu.
11711 	 */
11712 	if (last_mm_cid == -1)
11713 		return -1;
11714 	src_cid = READ_ONCE(src_pcpu_cid->cid);
11715 	if (!mm_cid_is_valid(src_cid) || last_mm_cid != src_cid)
11716 		return -1;
11717 
11718 	/*
11719 	 * If we observe an active task using the mm on this rq, it means we
11720 	 * are not the last task to be migrated from this cpu for this mm, so
11721 	 * there is no need to move src_cid to the destination cpu.
11722 	 */
11723 	rcu_read_lock();
11724 	src_task = rcu_dereference(src_rq->curr);
11725 	if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11726 		rcu_read_unlock();
11727 		t->last_mm_cid = -1;
11728 		return -1;
11729 	}
11730 	rcu_read_unlock();
11731 
11732 	return src_cid;
11733 }
11734 
11735 static
__sched_mm_cid_migrate_from_try_steal_cid(struct rq * src_rq,struct task_struct * t,struct mm_cid * src_pcpu_cid,int src_cid)11736 int __sched_mm_cid_migrate_from_try_steal_cid(struct rq *src_rq,
11737 					      struct task_struct *t,
11738 					      struct mm_cid *src_pcpu_cid,
11739 					      int src_cid)
11740 {
11741 	struct task_struct *src_task;
11742 	struct mm_struct *mm = t->mm;
11743 	int lazy_cid;
11744 
11745 	if (src_cid == -1)
11746 		return -1;
11747 
11748 	/*
11749 	 * Attempt to clear the source cpu cid to move it to the destination
11750 	 * cpu.
11751 	 */
11752 	lazy_cid = mm_cid_set_lazy_put(src_cid);
11753 	if (!try_cmpxchg(&src_pcpu_cid->cid, &src_cid, lazy_cid))
11754 		return -1;
11755 
11756 	/*
11757 	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11758 	 * rq->curr->mm matches the scheduler barrier in context_switch()
11759 	 * between store to rq->curr and load of prev and next task's
11760 	 * per-mm/cpu cid.
11761 	 *
11762 	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11763 	 * rq->curr->mm_cid_active matches the barrier in
11764 	 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11765 	 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11766 	 * load of per-mm/cpu cid.
11767 	 */
11768 
11769 	/*
11770 	 * If we observe an active task using the mm on this rq after setting
11771 	 * the lazy-put flag, this task will be responsible for transitioning
11772 	 * from lazy-put flag set to MM_CID_UNSET.
11773 	 */
11774 	rcu_read_lock();
11775 	src_task = rcu_dereference(src_rq->curr);
11776 	if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11777 		rcu_read_unlock();
11778 		/*
11779 		 * We observed an active task for this mm, there is therefore
11780 		 * no point in moving this cid to the destination cpu.
11781 		 */
11782 		t->last_mm_cid = -1;
11783 		return -1;
11784 	}
11785 	rcu_read_unlock();
11786 
11787 	/*
11788 	 * The src_cid is unused, so it can be unset.
11789 	 */
11790 	if (!try_cmpxchg(&src_pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11791 		return -1;
11792 	return src_cid;
11793 }
11794 
11795 /*
11796  * Migration to dst cpu. Called with dst_rq lock held.
11797  * Interrupts are disabled, which keeps the window of cid ownership without the
11798  * source rq lock held small.
11799  */
sched_mm_cid_migrate_to(struct rq * dst_rq,struct task_struct * t)11800 void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t)
11801 {
11802 	struct mm_cid *src_pcpu_cid, *dst_pcpu_cid;
11803 	struct mm_struct *mm = t->mm;
11804 	int src_cid, dst_cid, src_cpu;
11805 	struct rq *src_rq;
11806 
11807 	lockdep_assert_rq_held(dst_rq);
11808 
11809 	if (!mm)
11810 		return;
11811 	src_cpu = t->migrate_from_cpu;
11812 	if (src_cpu == -1) {
11813 		t->last_mm_cid = -1;
11814 		return;
11815 	}
11816 	/*
11817 	 * Move the src cid if the dst cid is unset. This keeps id
11818 	 * allocation closest to 0 in cases where few threads migrate around
11819 	 * many cpus.
11820 	 *
11821 	 * If destination cid is already set, we may have to just clear
11822 	 * the src cid to ensure compactness in frequent migrations
11823 	 * scenarios.
11824 	 *
11825 	 * It is not useful to clear the src cid when the number of threads is
11826 	 * greater or equal to the number of allowed cpus, because user-space
11827 	 * can expect that the number of allowed cids can reach the number of
11828 	 * allowed cpus.
11829 	 */
11830 	dst_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(dst_rq));
11831 	dst_cid = READ_ONCE(dst_pcpu_cid->cid);
11832 	if (!mm_cid_is_unset(dst_cid) &&
11833 	    atomic_read(&mm->mm_users) >= t->nr_cpus_allowed)
11834 		return;
11835 	src_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, src_cpu);
11836 	src_rq = cpu_rq(src_cpu);
11837 	src_cid = __sched_mm_cid_migrate_from_fetch_cid(src_rq, t, src_pcpu_cid);
11838 	if (src_cid == -1)
11839 		return;
11840 	src_cid = __sched_mm_cid_migrate_from_try_steal_cid(src_rq, t, src_pcpu_cid,
11841 							    src_cid);
11842 	if (src_cid == -1)
11843 		return;
11844 	if (!mm_cid_is_unset(dst_cid)) {
11845 		__mm_cid_put(mm, src_cid);
11846 		return;
11847 	}
11848 	/* Move src_cid to dst cpu. */
11849 	mm_cid_snapshot_time(dst_rq, mm);
11850 	WRITE_ONCE(dst_pcpu_cid->cid, src_cid);
11851 }
11852 
sched_mm_cid_remote_clear(struct mm_struct * mm,struct mm_cid * pcpu_cid,int cpu)11853 static void sched_mm_cid_remote_clear(struct mm_struct *mm, struct mm_cid *pcpu_cid,
11854 				      int cpu)
11855 {
11856 	struct rq *rq = cpu_rq(cpu);
11857 	struct task_struct *t;
11858 	unsigned long flags;
11859 	int cid, lazy_cid;
11860 
11861 	cid = READ_ONCE(pcpu_cid->cid);
11862 	if (!mm_cid_is_valid(cid))
11863 		return;
11864 
11865 	/*
11866 	 * Clear the cpu cid if it is set to keep cid allocation compact.  If
11867 	 * there happens to be other tasks left on the source cpu using this
11868 	 * mm, the next task using this mm will reallocate its cid on context
11869 	 * switch.
11870 	 */
11871 	lazy_cid = mm_cid_set_lazy_put(cid);
11872 	if (!try_cmpxchg(&pcpu_cid->cid, &cid, lazy_cid))
11873 		return;
11874 
11875 	/*
11876 	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11877 	 * rq->curr->mm matches the scheduler barrier in context_switch()
11878 	 * between store to rq->curr and load of prev and next task's
11879 	 * per-mm/cpu cid.
11880 	 *
11881 	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11882 	 * rq->curr->mm_cid_active matches the barrier in
11883 	 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11884 	 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11885 	 * load of per-mm/cpu cid.
11886 	 */
11887 
11888 	/*
11889 	 * If we observe an active task using the mm on this rq after setting
11890 	 * the lazy-put flag, that task will be responsible for transitioning
11891 	 * from lazy-put flag set to MM_CID_UNSET.
11892 	 */
11893 	rcu_read_lock();
11894 	t = rcu_dereference(rq->curr);
11895 	if (READ_ONCE(t->mm_cid_active) && t->mm == mm) {
11896 		rcu_read_unlock();
11897 		return;
11898 	}
11899 	rcu_read_unlock();
11900 
11901 	/*
11902 	 * The cid is unused, so it can be unset.
11903 	 * Disable interrupts to keep the window of cid ownership without rq
11904 	 * lock small.
11905 	 */
11906 	local_irq_save(flags);
11907 	if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11908 		__mm_cid_put(mm, cid);
11909 	local_irq_restore(flags);
11910 }
11911 
sched_mm_cid_remote_clear_old(struct mm_struct * mm,int cpu)11912 static void sched_mm_cid_remote_clear_old(struct mm_struct *mm, int cpu)
11913 {
11914 	struct rq *rq = cpu_rq(cpu);
11915 	struct mm_cid *pcpu_cid;
11916 	struct task_struct *curr;
11917 	u64 rq_clock;
11918 
11919 	/*
11920 	 * rq->clock load is racy on 32-bit but one spurious clear once in a
11921 	 * while is irrelevant.
11922 	 */
11923 	rq_clock = READ_ONCE(rq->clock);
11924 	pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11925 
11926 	/*
11927 	 * In order to take care of infrequently scheduled tasks, bump the time
11928 	 * snapshot associated with this cid if an active task using the mm is
11929 	 * observed on this rq.
11930 	 */
11931 	rcu_read_lock();
11932 	curr = rcu_dereference(rq->curr);
11933 	if (READ_ONCE(curr->mm_cid_active) && curr->mm == mm) {
11934 		WRITE_ONCE(pcpu_cid->time, rq_clock);
11935 		rcu_read_unlock();
11936 		return;
11937 	}
11938 	rcu_read_unlock();
11939 
11940 	if (rq_clock < pcpu_cid->time + SCHED_MM_CID_PERIOD_NS)
11941 		return;
11942 	sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11943 }
11944 
sched_mm_cid_remote_clear_weight(struct mm_struct * mm,int cpu,int weight)11945 static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
11946 					     int weight)
11947 {
11948 	struct mm_cid *pcpu_cid;
11949 	int cid;
11950 
11951 	pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11952 	cid = READ_ONCE(pcpu_cid->cid);
11953 	if (!mm_cid_is_valid(cid) || cid < weight)
11954 		return;
11955 	sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11956 }
11957 
task_mm_cid_work(struct callback_head * work)11958 static void task_mm_cid_work(struct callback_head *work)
11959 {
11960 	unsigned long now = jiffies, old_scan, next_scan;
11961 	struct task_struct *t = current;
11962 	struct cpumask *cidmask;
11963 	struct mm_struct *mm;
11964 	int weight, cpu;
11965 
11966 	SCHED_WARN_ON(t != container_of(work, struct task_struct, cid_work));
11967 
11968 	work->next = work;	/* Prevent double-add */
11969 	if (t->flags & PF_EXITING)
11970 		return;
11971 	mm = t->mm;
11972 	if (!mm)
11973 		return;
11974 	old_scan = READ_ONCE(mm->mm_cid_next_scan);
11975 	next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11976 	if (!old_scan) {
11977 		unsigned long res;
11978 
11979 		res = cmpxchg(&mm->mm_cid_next_scan, old_scan, next_scan);
11980 		if (res != old_scan)
11981 			old_scan = res;
11982 		else
11983 			old_scan = next_scan;
11984 	}
11985 	if (time_before(now, old_scan))
11986 		return;
11987 	if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
11988 		return;
11989 	cidmask = mm_cidmask(mm);
11990 	/* Clear cids that were not recently used. */
11991 	for_each_possible_cpu(cpu)
11992 		sched_mm_cid_remote_clear_old(mm, cpu);
11993 	weight = cpumask_weight(cidmask);
11994 	/*
11995 	 * Clear cids that are greater or equal to the cidmask weight to
11996 	 * recompact it.
11997 	 */
11998 	for_each_possible_cpu(cpu)
11999 		sched_mm_cid_remote_clear_weight(mm, cpu, weight);
12000 }
12001 
init_sched_mm_cid(struct task_struct * t)12002 void init_sched_mm_cid(struct task_struct *t)
12003 {
12004 	struct mm_struct *mm = t->mm;
12005 	int mm_users = 0;
12006 
12007 	if (mm) {
12008 		mm_users = atomic_read(&mm->mm_users);
12009 		if (mm_users == 1)
12010 			mm->mm_cid_next_scan = jiffies + msecs_to_jiffies(MM_CID_SCAN_DELAY);
12011 	}
12012 	t->cid_work.next = &t->cid_work;	/* Protect against double add */
12013 	init_task_work(&t->cid_work, task_mm_cid_work);
12014 }
12015 
task_tick_mm_cid(struct rq * rq,struct task_struct * curr)12016 void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
12017 {
12018 	struct callback_head *work = &curr->cid_work;
12019 	unsigned long now = jiffies;
12020 
12021 	if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
12022 	    work->next != work)
12023 		return;
12024 	if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
12025 		return;
12026 	task_work_add(curr, work, TWA_RESUME);
12027 }
12028 
sched_mm_cid_exit_signals(struct task_struct * t)12029 void sched_mm_cid_exit_signals(struct task_struct *t)
12030 {
12031 	struct mm_struct *mm = t->mm;
12032 	struct rq_flags rf;
12033 	struct rq *rq;
12034 
12035 	if (!mm)
12036 		return;
12037 
12038 	preempt_disable();
12039 	rq = this_rq();
12040 	rq_lock_irqsave(rq, &rf);
12041 	preempt_enable_no_resched();	/* holding spinlock */
12042 	WRITE_ONCE(t->mm_cid_active, 0);
12043 	/*
12044 	 * Store t->mm_cid_active before loading per-mm/cpu cid.
12045 	 * Matches barrier in sched_mm_cid_remote_clear_old().
12046 	 */
12047 	smp_mb();
12048 	mm_cid_put(mm);
12049 	t->last_mm_cid = t->mm_cid = -1;
12050 	rq_unlock_irqrestore(rq, &rf);
12051 }
12052 
sched_mm_cid_before_execve(struct task_struct * t)12053 void sched_mm_cid_before_execve(struct task_struct *t)
12054 {
12055 	struct mm_struct *mm = t->mm;
12056 	struct rq_flags rf;
12057 	struct rq *rq;
12058 
12059 	if (!mm)
12060 		return;
12061 
12062 	preempt_disable();
12063 	rq = this_rq();
12064 	rq_lock_irqsave(rq, &rf);
12065 	preempt_enable_no_resched();	/* holding spinlock */
12066 	WRITE_ONCE(t->mm_cid_active, 0);
12067 	/*
12068 	 * Store t->mm_cid_active before loading per-mm/cpu cid.
12069 	 * Matches barrier in sched_mm_cid_remote_clear_old().
12070 	 */
12071 	smp_mb();
12072 	mm_cid_put(mm);
12073 	t->last_mm_cid = t->mm_cid = -1;
12074 	rq_unlock_irqrestore(rq, &rf);
12075 }
12076 
sched_mm_cid_after_execve(struct task_struct * t)12077 void sched_mm_cid_after_execve(struct task_struct *t)
12078 {
12079 	struct mm_struct *mm = t->mm;
12080 	struct rq_flags rf;
12081 	struct rq *rq;
12082 
12083 	if (!mm)
12084 		return;
12085 
12086 	preempt_disable();
12087 	rq = this_rq();
12088 	rq_lock_irqsave(rq, &rf);
12089 	preempt_enable_no_resched();	/* holding spinlock */
12090 	WRITE_ONCE(t->mm_cid_active, 1);
12091 	/*
12092 	 * Store t->mm_cid_active before loading per-mm/cpu cid.
12093 	 * Matches barrier in sched_mm_cid_remote_clear_old().
12094 	 */
12095 	smp_mb();
12096 	t->last_mm_cid = t->mm_cid = mm_cid_get(rq, mm);
12097 	rq_unlock_irqrestore(rq, &rf);
12098 	rseq_set_notify_resume(t);
12099 }
12100 
sched_mm_cid_fork(struct task_struct * t)12101 void sched_mm_cid_fork(struct task_struct *t)
12102 {
12103 	WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
12104 	t->mm_cid_active = 1;
12105 }
12106 #endif
12107