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