xref: /linux/kernel/sched/psi.c (revision e724e7aaf9ca794670a4d4931af7a7e24e37fec3)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Pressure stall information for CPU, memory and IO
4  *
5  * Copyright (c) 2018 Facebook, Inc.
6  * Author: Johannes Weiner <hannes@cmpxchg.org>
7  *
8  * Polling support by Suren Baghdasaryan <surenb@google.com>
9  * Copyright (c) 2018 Google, Inc.
10  *
11  * When CPU, memory and IO are contended, tasks experience delays that
12  * reduce throughput and introduce latencies into the workload. Memory
13  * and IO contention, in addition, can cause a full loss of forward
14  * progress in which the CPU goes idle.
15  *
16  * This code aggregates individual task delays into resource pressure
17  * metrics that indicate problems with both workload health and
18  * resource utilization.
19  *
20  *			Model
21  *
22  * The time in which a task can execute on a CPU is our baseline for
23  * productivity. Pressure expresses the amount of time in which this
24  * potential cannot be realized due to resource contention.
25  *
26  * This concept of productivity has two components: the workload and
27  * the CPU. To measure the impact of pressure on both, we define two
28  * contention states for a resource: SOME and FULL.
29  *
30  * In the SOME state of a given resource, one or more tasks are
31  * delayed on that resource. This affects the workload's ability to
32  * perform work, but the CPU may still be executing other tasks.
33  *
34  * In the FULL state of a given resource, all non-idle tasks are
35  * delayed on that resource such that nobody is advancing and the CPU
36  * goes idle. This leaves both workload and CPU unproductive.
37  *
38  *	SOME = nr_delayed_tasks != 0
39  *	FULL = nr_delayed_tasks != 0 && nr_productive_tasks == 0
40  *
41  * What it means for a task to be productive is defined differently
42  * for each resource. For IO, productive means a running task. For
43  * memory, productive means a running task that isn't a reclaimer. For
44  * CPU, productive means an oncpu task.
45  *
46  * Naturally, the FULL state doesn't exist for the CPU resource at the
47  * system level, but exist at the cgroup level. At the cgroup level,
48  * FULL means all non-idle tasks in the cgroup are delayed on the CPU
49  * resource which is being used by others outside of the cgroup or
50  * throttled by the cgroup cpu.max configuration.
51  *
52  * The percentage of wallclock time spent in those compound stall
53  * states gives pressure numbers between 0 and 100 for each resource,
54  * where the SOME percentage indicates workload slowdowns and the FULL
55  * percentage indicates reduced CPU utilization:
56  *
57  *	%SOME = time(SOME) / period
58  *	%FULL = time(FULL) / period
59  *
60  *			Multiple CPUs
61  *
62  * The more tasks and available CPUs there are, the more work can be
63  * performed concurrently. This means that the potential that can go
64  * unrealized due to resource contention *also* scales with non-idle
65  * tasks and CPUs.
66  *
67  * Consider a scenario where 257 number crunching tasks are trying to
68  * run concurrently on 256 CPUs. If we simply aggregated the task
69  * states, we would have to conclude a CPU SOME pressure number of
70  * 100%, since *somebody* is waiting on a runqueue at all
71  * times. However, that is clearly not the amount of contention the
72  * workload is experiencing: only one out of 256 possible execution
73  * threads will be contended at any given time, or about 0.4%.
74  *
75  * Conversely, consider a scenario of 4 tasks and 4 CPUs where at any
76  * given time *one* of the tasks is delayed due to a lack of memory.
77  * Again, looking purely at the task state would yield a memory FULL
78  * pressure number of 0%, since *somebody* is always making forward
79  * progress. But again this wouldn't capture the amount of execution
80  * potential lost, which is 1 out of 4 CPUs, or 25%.
81  *
82  * To calculate wasted potential (pressure) with multiple processors,
83  * we have to base our calculation on the number of non-idle tasks in
84  * conjunction with the number of available CPUs, which is the number
85  * of potential execution threads. SOME becomes then the proportion of
86  * delayed tasks to possible threads, and FULL is the share of possible
87  * threads that are unproductive due to delays:
88  *
89  *	threads = min(nr_nonidle_tasks, nr_cpus)
90  *	   SOME = min(nr_delayed_tasks / threads, 1)
91  *	   FULL = (threads - min(nr_productive_tasks, threads)) / threads
92  *
93  * For the 257 number crunchers on 256 CPUs, this yields:
94  *
95  *	threads = min(257, 256)
96  *	   SOME = min(1 / 256, 1)             = 0.4%
97  *	   FULL = (256 - min(256, 256)) / 256 = 0%
98  *
99  * For the 1 out of 4 memory-delayed tasks, this yields:
100  *
101  *	threads = min(4, 4)
102  *	   SOME = min(1 / 4, 1)               = 25%
103  *	   FULL = (4 - min(3, 4)) / 4         = 25%
104  *
105  * [ Substitute nr_cpus with 1, and you can see that it's a natural
106  *   extension of the single-CPU model. ]
107  *
108  *			Implementation
109  *
110  * To assess the precise time spent in each such state, we would have
111  * to freeze the system on task changes and start/stop the state
112  * clocks accordingly. Obviously that doesn't scale in practice.
113  *
114  * Because the scheduler aims to distribute the compute load evenly
115  * among the available CPUs, we can track task state locally to each
116  * CPU and, at much lower frequency, extrapolate the global state for
117  * the cumulative stall times and the running averages.
118  *
119  * For each runqueue, we track:
120  *
121  *	   tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0)
122  *	   tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_productive_tasks[cpu])
123  *	tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0)
124  *
125  * and then periodically aggregate:
126  *
127  *	tNONIDLE = sum(tNONIDLE[i])
128  *
129  *	   tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE
130  *	   tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE
131  *
132  *	   %SOME = tSOME / period
133  *	   %FULL = tFULL / period
134  *
135  * This gives us an approximation of pressure that is practical
136  * cost-wise, yet way more sensitive and accurate than periodic
137  * sampling of the aggregate task states would be.
138  */
139 
140 static int psi_bug __read_mostly;
141 
142 DEFINE_STATIC_KEY_FALSE(psi_disabled);
143 DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled);
144 
145 #ifdef CONFIG_PSI_DEFAULT_DISABLED
146 static bool psi_enable;
147 #else
148 static bool psi_enable = true;
149 #endif
150 static int __init setup_psi(char *str)
151 {
152 	return kstrtobool(str, &psi_enable) == 0;
153 }
154 __setup("psi=", setup_psi);
155 
156 /* Running averages - we need to be higher-res than loadavg */
157 #define PSI_FREQ	(2*HZ+1)	/* 2 sec intervals */
158 #define EXP_10s		1677		/* 1/exp(2s/10s) as fixed-point */
159 #define EXP_60s		1981		/* 1/exp(2s/60s) */
160 #define EXP_300s	2034		/* 1/exp(2s/300s) */
161 
162 /* PSI trigger definitions */
163 #define WINDOW_MAX_US 10000000	/* Max window size is 10s */
164 #define UPDATES_PER_WINDOW 10	/* 10 updates per window */
165 
166 /* Sampling frequency in nanoseconds */
167 static u64 psi_period __read_mostly;
168 
169 /* System-level pressure and stall tracking */
170 static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
171 struct psi_group psi_system = {
172 	.pcpu = &system_group_pcpu,
173 };
174 
175 static void psi_avgs_work(struct work_struct *work);
176 
177 static void poll_timer_fn(struct timer_list *t);
178 
179 static void group_init(struct psi_group *group)
180 {
181 	int cpu;
182 
183 	group->enabled = true;
184 	for_each_possible_cpu(cpu)
185 		seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);
186 	group->avg_last_update = sched_clock();
187 	group->avg_next_update = group->avg_last_update + psi_period;
188 	mutex_init(&group->avgs_lock);
189 
190 	/* Init avg trigger-related members */
191 	INIT_LIST_HEAD(&group->avg_triggers);
192 	memset(group->avg_nr_triggers, 0, sizeof(group->avg_nr_triggers));
193 	INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
194 
195 	/* Init rtpoll trigger-related members */
196 	atomic_set(&group->rtpoll_scheduled, 0);
197 	mutex_init(&group->rtpoll_trigger_lock);
198 	INIT_LIST_HEAD(&group->rtpoll_triggers);
199 	group->rtpoll_min_period = U32_MAX;
200 	group->rtpoll_next_update = ULLONG_MAX;
201 	init_waitqueue_head(&group->rtpoll_wait);
202 	timer_setup(&group->rtpoll_timer, poll_timer_fn, 0);
203 	rcu_assign_pointer(group->rtpoll_task, NULL);
204 }
205 
206 void __init psi_init(void)
207 {
208 	if (!psi_enable) {
209 		static_branch_enable(&psi_disabled);
210 		static_branch_disable(&psi_cgroups_enabled);
211 		return;
212 	}
213 
214 	if (!cgroup_psi_enabled())
215 		static_branch_disable(&psi_cgroups_enabled);
216 
217 	psi_period = jiffies_to_nsecs(PSI_FREQ);
218 	group_init(&psi_system);
219 }
220 
221 static bool test_state(unsigned int *tasks, enum psi_states state, bool oncpu)
222 {
223 	switch (state) {
224 	case PSI_IO_SOME:
225 		return unlikely(tasks[NR_IOWAIT]);
226 	case PSI_IO_FULL:
227 		return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]);
228 	case PSI_MEM_SOME:
229 		return unlikely(tasks[NR_MEMSTALL]);
230 	case PSI_MEM_FULL:
231 		return unlikely(tasks[NR_MEMSTALL] &&
232 			tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]);
233 	case PSI_CPU_SOME:
234 		return unlikely(tasks[NR_RUNNING] > oncpu);
235 	case PSI_CPU_FULL:
236 		return unlikely(tasks[NR_RUNNING] && !oncpu);
237 	case PSI_NONIDLE:
238 		return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
239 			tasks[NR_RUNNING];
240 	default:
241 		return false;
242 	}
243 }
244 
245 static void get_recent_times(struct psi_group *group, int cpu,
246 			     enum psi_aggregators aggregator, u32 *times,
247 			     u32 *pchanged_states)
248 {
249 	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
250 	int current_cpu = raw_smp_processor_id();
251 	unsigned int tasks[NR_PSI_TASK_COUNTS];
252 	u64 now, state_start;
253 	enum psi_states s;
254 	unsigned int seq;
255 	u32 state_mask;
256 
257 	*pchanged_states = 0;
258 
259 	/* Snapshot a coherent view of the CPU state */
260 	do {
261 		seq = read_seqcount_begin(&groupc->seq);
262 		now = cpu_clock(cpu);
263 		memcpy(times, groupc->times, sizeof(groupc->times));
264 		state_mask = groupc->state_mask;
265 		state_start = groupc->state_start;
266 		if (cpu == current_cpu)
267 			memcpy(tasks, groupc->tasks, sizeof(groupc->tasks));
268 	} while (read_seqcount_retry(&groupc->seq, seq));
269 
270 	/* Calculate state time deltas against the previous snapshot */
271 	for (s = 0; s < NR_PSI_STATES; s++) {
272 		u32 delta;
273 		/*
274 		 * In addition to already concluded states, we also
275 		 * incorporate currently active states on the CPU,
276 		 * since states may last for many sampling periods.
277 		 *
278 		 * This way we keep our delta sampling buckets small
279 		 * (u32) and our reported pressure close to what's
280 		 * actually happening.
281 		 */
282 		if (state_mask & (1 << s))
283 			times[s] += now - state_start;
284 
285 		delta = times[s] - groupc->times_prev[aggregator][s];
286 		groupc->times_prev[aggregator][s] = times[s];
287 
288 		times[s] = delta;
289 		if (delta)
290 			*pchanged_states |= (1 << s);
291 	}
292 
293 	/*
294 	 * When collect_percpu_times() from the avgs_work, we don't want to
295 	 * re-arm avgs_work when all CPUs are IDLE. But the current CPU running
296 	 * this avgs_work is never IDLE, cause avgs_work can't be shut off.
297 	 * So for the current CPU, we need to re-arm avgs_work only when
298 	 * (NR_RUNNING > 1 || NR_IOWAIT > 0 || NR_MEMSTALL > 0), for other CPUs
299 	 * we can just check PSI_NONIDLE delta.
300 	 */
301 	if (current_work() == &group->avgs_work.work) {
302 		bool reschedule;
303 
304 		if (cpu == current_cpu)
305 			reschedule = tasks[NR_RUNNING] +
306 				     tasks[NR_IOWAIT] +
307 				     tasks[NR_MEMSTALL] > 1;
308 		else
309 			reschedule = *pchanged_states & (1 << PSI_NONIDLE);
310 
311 		if (reschedule)
312 			*pchanged_states |= PSI_STATE_RESCHEDULE;
313 	}
314 }
315 
316 static void calc_avgs(unsigned long avg[3], int missed_periods,
317 		      u64 time, u64 period)
318 {
319 	unsigned long pct;
320 
321 	/* Fill in zeroes for periods of no activity */
322 	if (missed_periods) {
323 		avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
324 		avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
325 		avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
326 	}
327 
328 	/* Sample the most recent active period */
329 	pct = div_u64(time * 100, period);
330 	pct *= FIXED_1;
331 	avg[0] = calc_load(avg[0], EXP_10s, pct);
332 	avg[1] = calc_load(avg[1], EXP_60s, pct);
333 	avg[2] = calc_load(avg[2], EXP_300s, pct);
334 }
335 
336 static void collect_percpu_times(struct psi_group *group,
337 				 enum psi_aggregators aggregator,
338 				 u32 *pchanged_states)
339 {
340 	u64 deltas[NR_PSI_STATES - 1] = { 0, };
341 	unsigned long nonidle_total = 0;
342 	u32 changed_states = 0;
343 	int cpu;
344 	int s;
345 
346 	/*
347 	 * Collect the per-cpu time buckets and average them into a
348 	 * single time sample that is normalized to wallclock time.
349 	 *
350 	 * For averaging, each CPU is weighted by its non-idle time in
351 	 * the sampling period. This eliminates artifacts from uneven
352 	 * loading, or even entirely idle CPUs.
353 	 */
354 	for_each_possible_cpu(cpu) {
355 		u32 times[NR_PSI_STATES];
356 		u32 nonidle;
357 		u32 cpu_changed_states;
358 
359 		get_recent_times(group, cpu, aggregator, times,
360 				&cpu_changed_states);
361 		changed_states |= cpu_changed_states;
362 
363 		nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
364 		nonidle_total += nonidle;
365 
366 		for (s = 0; s < PSI_NONIDLE; s++)
367 			deltas[s] += (u64)times[s] * nonidle;
368 	}
369 
370 	/*
371 	 * Integrate the sample into the running statistics that are
372 	 * reported to userspace: the cumulative stall times and the
373 	 * decaying averages.
374 	 *
375 	 * Pressure percentages are sampled at PSI_FREQ. We might be
376 	 * called more often when the user polls more frequently than
377 	 * that; we might be called less often when there is no task
378 	 * activity, thus no data, and clock ticks are sporadic. The
379 	 * below handles both.
380 	 */
381 
382 	/* total= */
383 	for (s = 0; s < NR_PSI_STATES - 1; s++)
384 		group->total[aggregator][s] +=
385 				div_u64(deltas[s], max(nonidle_total, 1UL));
386 
387 	if (pchanged_states)
388 		*pchanged_states = changed_states;
389 }
390 
391 /* Trigger tracking window manipulations */
392 static void window_reset(struct psi_window *win, u64 now, u64 value,
393 			 u64 prev_growth)
394 {
395 	win->start_time = now;
396 	win->start_value = value;
397 	win->prev_growth = prev_growth;
398 }
399 
400 /*
401  * PSI growth tracking window update and growth calculation routine.
402  *
403  * This approximates a sliding tracking window by interpolating
404  * partially elapsed windows using historical growth data from the
405  * previous intervals. This minimizes memory requirements (by not storing
406  * all the intermediate values in the previous window) and simplifies
407  * the calculations. It works well because PSI signal changes only in
408  * positive direction and over relatively small window sizes the growth
409  * is close to linear.
410  */
411 static u64 window_update(struct psi_window *win, u64 now, u64 value)
412 {
413 	u64 elapsed;
414 	u64 growth;
415 
416 	elapsed = now - win->start_time;
417 	growth = value - win->start_value;
418 	/*
419 	 * After each tracking window passes win->start_value and
420 	 * win->start_time get reset and win->prev_growth stores
421 	 * the average per-window growth of the previous window.
422 	 * win->prev_growth is then used to interpolate additional
423 	 * growth from the previous window assuming it was linear.
424 	 */
425 	if (elapsed > win->size)
426 		window_reset(win, now, value, growth);
427 	else {
428 		u32 remaining;
429 
430 		remaining = win->size - elapsed;
431 		growth += div64_u64(win->prev_growth * remaining, win->size);
432 	}
433 
434 	return growth;
435 }
436 
437 static u64 update_triggers(struct psi_group *group, u64 now, bool *update_total,
438 						   enum psi_aggregators aggregator)
439 {
440 	struct psi_trigger *t;
441 	u64 *total = group->total[aggregator];
442 	struct list_head *triggers;
443 	u64 *aggregator_total;
444 	*update_total = false;
445 
446 	if (aggregator == PSI_AVGS) {
447 		triggers = &group->avg_triggers;
448 		aggregator_total = group->avg_total;
449 	} else {
450 		triggers = &group->rtpoll_triggers;
451 		aggregator_total = group->rtpoll_total;
452 	}
453 
454 	/*
455 	 * On subsequent updates, calculate growth deltas and let
456 	 * watchers know when their specified thresholds are exceeded.
457 	 */
458 	list_for_each_entry(t, triggers, node) {
459 		u64 growth;
460 		bool new_stall;
461 
462 		new_stall = aggregator_total[t->state] != total[t->state];
463 
464 		/* Check for stall activity or a previous threshold breach */
465 		if (!new_stall && !t->pending_event)
466 			continue;
467 		/*
468 		 * Check for new stall activity, as well as deferred
469 		 * events that occurred in the last window after the
470 		 * trigger had already fired (we want to ratelimit
471 		 * events without dropping any).
472 		 */
473 		if (new_stall) {
474 			/*
475 			 * Multiple triggers might be looking at the same state,
476 			 * remember to update group->polling_total[] once we've
477 			 * been through all of them. Also remember to extend the
478 			 * polling time if we see new stall activity.
479 			 */
480 			*update_total = true;
481 
482 			/* Calculate growth since last update */
483 			growth = window_update(&t->win, now, total[t->state]);
484 			if (!t->pending_event) {
485 				if (growth < t->threshold)
486 					continue;
487 
488 				t->pending_event = true;
489 			}
490 		}
491 		/* Limit event signaling to once per window */
492 		if (now < t->last_event_time + t->win.size)
493 			continue;
494 
495 		/* Generate an event */
496 		if (cmpxchg(&t->event, 0, 1) == 0) {
497 			if (t->of)
498 				kernfs_notify(t->of->kn);
499 			else
500 				wake_up_interruptible(&t->event_wait);
501 		}
502 		t->last_event_time = now;
503 		/* Reset threshold breach flag once event got generated */
504 		t->pending_event = false;
505 	}
506 
507 	return now + group->rtpoll_min_period;
508 }
509 
510 static u64 update_averages(struct psi_group *group, u64 now)
511 {
512 	unsigned long missed_periods = 0;
513 	u64 expires, period;
514 	u64 avg_next_update;
515 	int s;
516 
517 	/* avgX= */
518 	expires = group->avg_next_update;
519 	if (now - expires >= psi_period)
520 		missed_periods = div_u64(now - expires, psi_period);
521 
522 	/*
523 	 * The periodic clock tick can get delayed for various
524 	 * reasons, especially on loaded systems. To avoid clock
525 	 * drift, we schedule the clock in fixed psi_period intervals.
526 	 * But the deltas we sample out of the per-cpu buckets above
527 	 * are based on the actual time elapsing between clock ticks.
528 	 */
529 	avg_next_update = expires + ((1 + missed_periods) * psi_period);
530 	period = now - (group->avg_last_update + (missed_periods * psi_period));
531 	group->avg_last_update = now;
532 
533 	for (s = 0; s < NR_PSI_STATES - 1; s++) {
534 		u32 sample;
535 
536 		sample = group->total[PSI_AVGS][s] - group->avg_total[s];
537 		/*
538 		 * Due to the lockless sampling of the time buckets,
539 		 * recorded time deltas can slip into the next period,
540 		 * which under full pressure can result in samples in
541 		 * excess of the period length.
542 		 *
543 		 * We don't want to report non-sensical pressures in
544 		 * excess of 100%, nor do we want to drop such events
545 		 * on the floor. Instead we punt any overage into the
546 		 * future until pressure subsides. By doing this we
547 		 * don't underreport the occurring pressure curve, we
548 		 * just report it delayed by one period length.
549 		 *
550 		 * The error isn't cumulative. As soon as another
551 		 * delta slips from a period P to P+1, by definition
552 		 * it frees up its time T in P.
553 		 */
554 		if (sample > period)
555 			sample = period;
556 		group->avg_total[s] += sample;
557 		calc_avgs(group->avg[s], missed_periods, sample, period);
558 	}
559 
560 	return avg_next_update;
561 }
562 
563 static void psi_avgs_work(struct work_struct *work)
564 {
565 	struct delayed_work *dwork;
566 	struct psi_group *group;
567 	u32 changed_states;
568 	bool update_total;
569 	u64 now;
570 
571 	dwork = to_delayed_work(work);
572 	group = container_of(dwork, struct psi_group, avgs_work);
573 
574 	mutex_lock(&group->avgs_lock);
575 
576 	now = sched_clock();
577 
578 	collect_percpu_times(group, PSI_AVGS, &changed_states);
579 	/*
580 	 * If there is task activity, periodically fold the per-cpu
581 	 * times and feed samples into the running averages. If things
582 	 * are idle and there is no data to process, stop the clock.
583 	 * Once restarted, we'll catch up the running averages in one
584 	 * go - see calc_avgs() and missed_periods.
585 	 */
586 	if (now >= group->avg_next_update) {
587 		update_triggers(group, now, &update_total, PSI_AVGS);
588 		group->avg_next_update = update_averages(group, now);
589 	}
590 
591 	if (changed_states & PSI_STATE_RESCHEDULE) {
592 		schedule_delayed_work(dwork, nsecs_to_jiffies(
593 				group->avg_next_update - now) + 1);
594 	}
595 
596 	mutex_unlock(&group->avgs_lock);
597 }
598 
599 static void init_rtpoll_triggers(struct psi_group *group, u64 now)
600 {
601 	struct psi_trigger *t;
602 
603 	list_for_each_entry(t, &group->rtpoll_triggers, node)
604 		window_reset(&t->win, now,
605 				group->total[PSI_POLL][t->state], 0);
606 	memcpy(group->rtpoll_total, group->total[PSI_POLL],
607 		   sizeof(group->rtpoll_total));
608 	group->rtpoll_next_update = now + group->rtpoll_min_period;
609 }
610 
611 /* Schedule polling if it's not already scheduled or forced. */
612 static void psi_schedule_rtpoll_work(struct psi_group *group, unsigned long delay,
613 				   bool force)
614 {
615 	struct task_struct *task;
616 
617 	/*
618 	 * atomic_xchg should be called even when !force to provide a
619 	 * full memory barrier (see the comment inside psi_rtpoll_work).
620 	 */
621 	if (atomic_xchg(&group->rtpoll_scheduled, 1) && !force)
622 		return;
623 
624 	rcu_read_lock();
625 
626 	task = rcu_dereference(group->rtpoll_task);
627 	/*
628 	 * kworker might be NULL in case psi_trigger_destroy races with
629 	 * psi_task_change (hotpath) which can't use locks
630 	 */
631 	if (likely(task))
632 		mod_timer(&group->rtpoll_timer, jiffies + delay);
633 	else
634 		atomic_set(&group->rtpoll_scheduled, 0);
635 
636 	rcu_read_unlock();
637 }
638 
639 static void psi_rtpoll_work(struct psi_group *group)
640 {
641 	bool force_reschedule = false;
642 	u32 changed_states;
643 	bool update_total;
644 	u64 now;
645 
646 	mutex_lock(&group->rtpoll_trigger_lock);
647 
648 	now = sched_clock();
649 
650 	if (now > group->rtpoll_until) {
651 		/*
652 		 * We are either about to start or might stop polling if no
653 		 * state change was recorded. Resetting poll_scheduled leaves
654 		 * a small window for psi_group_change to sneak in and schedule
655 		 * an immediate poll_work before we get to rescheduling. One
656 		 * potential extra wakeup at the end of the polling window
657 		 * should be negligible and polling_next_update still keeps
658 		 * updates correctly on schedule.
659 		 */
660 		atomic_set(&group->rtpoll_scheduled, 0);
661 		/*
662 		 * A task change can race with the poll worker that is supposed to
663 		 * report on it. To avoid missing events, ensure ordering between
664 		 * poll_scheduled and the task state accesses, such that if the poll
665 		 * worker misses the state update, the task change is guaranteed to
666 		 * reschedule the poll worker:
667 		 *
668 		 * poll worker:
669 		 *   atomic_set(poll_scheduled, 0)
670 		 *   smp_mb()
671 		 *   LOAD states
672 		 *
673 		 * task change:
674 		 *   STORE states
675 		 *   if atomic_xchg(poll_scheduled, 1) == 0:
676 		 *     schedule poll worker
677 		 *
678 		 * The atomic_xchg() implies a full barrier.
679 		 */
680 		smp_mb();
681 	} else {
682 		/* Polling window is not over, keep rescheduling */
683 		force_reschedule = true;
684 	}
685 
686 
687 	collect_percpu_times(group, PSI_POLL, &changed_states);
688 
689 	if (changed_states & group->rtpoll_states) {
690 		/* Initialize trigger windows when entering polling mode */
691 		if (now > group->rtpoll_until)
692 			init_rtpoll_triggers(group, now);
693 
694 		/*
695 		 * Keep the monitor active for at least the duration of the
696 		 * minimum tracking window as long as monitor states are
697 		 * changing.
698 		 */
699 		group->rtpoll_until = now +
700 			group->rtpoll_min_period * UPDATES_PER_WINDOW;
701 	}
702 
703 	if (now > group->rtpoll_until) {
704 		group->rtpoll_next_update = ULLONG_MAX;
705 		goto out;
706 	}
707 
708 	if (now >= group->rtpoll_next_update) {
709 		group->rtpoll_next_update = update_triggers(group, now, &update_total, PSI_POLL);
710 		if (update_total)
711 			memcpy(group->rtpoll_total, group->total[PSI_POLL],
712 				   sizeof(group->rtpoll_total));
713 	}
714 
715 	psi_schedule_rtpoll_work(group,
716 		nsecs_to_jiffies(group->rtpoll_next_update - now) + 1,
717 		force_reschedule);
718 
719 out:
720 	mutex_unlock(&group->rtpoll_trigger_lock);
721 }
722 
723 static int psi_rtpoll_worker(void *data)
724 {
725 	struct psi_group *group = (struct psi_group *)data;
726 
727 	sched_set_fifo_low(current);
728 
729 	while (true) {
730 		wait_event_interruptible(group->rtpoll_wait,
731 				atomic_cmpxchg(&group->rtpoll_wakeup, 1, 0) ||
732 				kthread_should_stop());
733 		if (kthread_should_stop())
734 			break;
735 
736 		psi_rtpoll_work(group);
737 	}
738 	return 0;
739 }
740 
741 static void poll_timer_fn(struct timer_list *t)
742 {
743 	struct psi_group *group = from_timer(group, t, rtpoll_timer);
744 
745 	atomic_set(&group->rtpoll_wakeup, 1);
746 	wake_up_interruptible(&group->rtpoll_wait);
747 }
748 
749 static void record_times(struct psi_group_cpu *groupc, u64 now)
750 {
751 	u32 delta;
752 
753 	delta = now - groupc->state_start;
754 	groupc->state_start = now;
755 
756 	if (groupc->state_mask & (1 << PSI_IO_SOME)) {
757 		groupc->times[PSI_IO_SOME] += delta;
758 		if (groupc->state_mask & (1 << PSI_IO_FULL))
759 			groupc->times[PSI_IO_FULL] += delta;
760 	}
761 
762 	if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
763 		groupc->times[PSI_MEM_SOME] += delta;
764 		if (groupc->state_mask & (1 << PSI_MEM_FULL))
765 			groupc->times[PSI_MEM_FULL] += delta;
766 	}
767 
768 	if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
769 		groupc->times[PSI_CPU_SOME] += delta;
770 		if (groupc->state_mask & (1 << PSI_CPU_FULL))
771 			groupc->times[PSI_CPU_FULL] += delta;
772 	}
773 
774 	if (groupc->state_mask & (1 << PSI_NONIDLE))
775 		groupc->times[PSI_NONIDLE] += delta;
776 }
777 
778 static void psi_group_change(struct psi_group *group, int cpu,
779 			     unsigned int clear, unsigned int set, u64 now,
780 			     bool wake_clock)
781 {
782 	struct psi_group_cpu *groupc;
783 	unsigned int t, m;
784 	enum psi_states s;
785 	u32 state_mask;
786 
787 	groupc = per_cpu_ptr(group->pcpu, cpu);
788 
789 	/*
790 	 * First we update the task counts according to the state
791 	 * change requested through the @clear and @set bits.
792 	 *
793 	 * Then if the cgroup PSI stats accounting enabled, we
794 	 * assess the aggregate resource states this CPU's tasks
795 	 * have been in since the last change, and account any
796 	 * SOME and FULL time these may have resulted in.
797 	 */
798 	write_seqcount_begin(&groupc->seq);
799 
800 	/*
801 	 * Start with TSK_ONCPU, which doesn't have a corresponding
802 	 * task count - it's just a boolean flag directly encoded in
803 	 * the state mask. Clear, set, or carry the current state if
804 	 * no changes are requested.
805 	 */
806 	if (unlikely(clear & TSK_ONCPU)) {
807 		state_mask = 0;
808 		clear &= ~TSK_ONCPU;
809 	} else if (unlikely(set & TSK_ONCPU)) {
810 		state_mask = PSI_ONCPU;
811 		set &= ~TSK_ONCPU;
812 	} else {
813 		state_mask = groupc->state_mask & PSI_ONCPU;
814 	}
815 
816 	/*
817 	 * The rest of the state mask is calculated based on the task
818 	 * counts. Update those first, then construct the mask.
819 	 */
820 	for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
821 		if (!(m & (1 << t)))
822 			continue;
823 		if (groupc->tasks[t]) {
824 			groupc->tasks[t]--;
825 		} else if (!psi_bug) {
826 			printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n",
827 					cpu, t, groupc->tasks[0],
828 					groupc->tasks[1], groupc->tasks[2],
829 					groupc->tasks[3], clear, set);
830 			psi_bug = 1;
831 		}
832 	}
833 
834 	for (t = 0; set; set &= ~(1 << t), t++)
835 		if (set & (1 << t))
836 			groupc->tasks[t]++;
837 
838 	if (!group->enabled) {
839 		/*
840 		 * On the first group change after disabling PSI, conclude
841 		 * the current state and flush its time. This is unlikely
842 		 * to matter to the user, but aggregation (get_recent_times)
843 		 * may have already incorporated the live state into times_prev;
844 		 * avoid a delta sample underflow when PSI is later re-enabled.
845 		 */
846 		if (unlikely(groupc->state_mask & (1 << PSI_NONIDLE)))
847 			record_times(groupc, now);
848 
849 		groupc->state_mask = state_mask;
850 
851 		write_seqcount_end(&groupc->seq);
852 		return;
853 	}
854 
855 	for (s = 0; s < NR_PSI_STATES; s++) {
856 		if (test_state(groupc->tasks, s, state_mask & PSI_ONCPU))
857 			state_mask |= (1 << s);
858 	}
859 
860 	/*
861 	 * Since we care about lost potential, a memstall is FULL
862 	 * when there are no other working tasks, but also when
863 	 * the CPU is actively reclaiming and nothing productive
864 	 * could run even if it were runnable. So when the current
865 	 * task in a cgroup is in_memstall, the corresponding groupc
866 	 * on that cpu is in PSI_MEM_FULL state.
867 	 */
868 	if (unlikely((state_mask & PSI_ONCPU) && cpu_curr(cpu)->in_memstall))
869 		state_mask |= (1 << PSI_MEM_FULL);
870 
871 	record_times(groupc, now);
872 
873 	groupc->state_mask = state_mask;
874 
875 	write_seqcount_end(&groupc->seq);
876 
877 	if (state_mask & group->rtpoll_states)
878 		psi_schedule_rtpoll_work(group, 1, false);
879 
880 	if (wake_clock && !delayed_work_pending(&group->avgs_work))
881 		schedule_delayed_work(&group->avgs_work, PSI_FREQ);
882 }
883 
884 static inline struct psi_group *task_psi_group(struct task_struct *task)
885 {
886 #ifdef CONFIG_CGROUPS
887 	if (static_branch_likely(&psi_cgroups_enabled))
888 		return cgroup_psi(task_dfl_cgroup(task));
889 #endif
890 	return &psi_system;
891 }
892 
893 static void psi_flags_change(struct task_struct *task, int clear, int set)
894 {
895 	if (((task->psi_flags & set) ||
896 	     (task->psi_flags & clear) != clear) &&
897 	    !psi_bug) {
898 		printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
899 				task->pid, task->comm, task_cpu(task),
900 				task->psi_flags, clear, set);
901 		psi_bug = 1;
902 	}
903 
904 	task->psi_flags &= ~clear;
905 	task->psi_flags |= set;
906 }
907 
908 void psi_task_change(struct task_struct *task, int clear, int set)
909 {
910 	int cpu = task_cpu(task);
911 	struct psi_group *group;
912 	u64 now;
913 
914 	if (!task->pid)
915 		return;
916 
917 	psi_flags_change(task, clear, set);
918 
919 	now = cpu_clock(cpu);
920 
921 	group = task_psi_group(task);
922 	do {
923 		psi_group_change(group, cpu, clear, set, now, true);
924 	} while ((group = group->parent));
925 }
926 
927 void psi_task_switch(struct task_struct *prev, struct task_struct *next,
928 		     bool sleep)
929 {
930 	struct psi_group *group, *common = NULL;
931 	int cpu = task_cpu(prev);
932 	u64 now = cpu_clock(cpu);
933 
934 	if (next->pid) {
935 		psi_flags_change(next, 0, TSK_ONCPU);
936 		/*
937 		 * Set TSK_ONCPU on @next's cgroups. If @next shares any
938 		 * ancestors with @prev, those will already have @prev's
939 		 * TSK_ONCPU bit set, and we can stop the iteration there.
940 		 */
941 		group = task_psi_group(next);
942 		do {
943 			if (per_cpu_ptr(group->pcpu, cpu)->state_mask &
944 			    PSI_ONCPU) {
945 				common = group;
946 				break;
947 			}
948 
949 			psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
950 		} while ((group = group->parent));
951 	}
952 
953 	if (prev->pid) {
954 		int clear = TSK_ONCPU, set = 0;
955 		bool wake_clock = true;
956 
957 		/*
958 		 * When we're going to sleep, psi_dequeue() lets us
959 		 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
960 		 * TSK_IOWAIT here, where we can combine it with
961 		 * TSK_ONCPU and save walking common ancestors twice.
962 		 */
963 		if (sleep) {
964 			clear |= TSK_RUNNING;
965 			if (prev->in_memstall)
966 				clear |= TSK_MEMSTALL_RUNNING;
967 			if (prev->in_iowait)
968 				set |= TSK_IOWAIT;
969 
970 			/*
971 			 * Periodic aggregation shuts off if there is a period of no
972 			 * task changes, so we wake it back up if necessary. However,
973 			 * don't do this if the task change is the aggregation worker
974 			 * itself going to sleep, or we'll ping-pong forever.
975 			 */
976 			if (unlikely((prev->flags & PF_WQ_WORKER) &&
977 				     wq_worker_last_func(prev) == psi_avgs_work))
978 				wake_clock = false;
979 		}
980 
981 		psi_flags_change(prev, clear, set);
982 
983 		group = task_psi_group(prev);
984 		do {
985 			if (group == common)
986 				break;
987 			psi_group_change(group, cpu, clear, set, now, wake_clock);
988 		} while ((group = group->parent));
989 
990 		/*
991 		 * TSK_ONCPU is handled up to the common ancestor. If there are
992 		 * any other differences between the two tasks (e.g. prev goes
993 		 * to sleep, or only one task is memstall), finish propagating
994 		 * those differences all the way up to the root.
995 		 */
996 		if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) {
997 			clear &= ~TSK_ONCPU;
998 			for (; group; group = group->parent)
999 				psi_group_change(group, cpu, clear, set, now, wake_clock);
1000 		}
1001 	}
1002 }
1003 
1004 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1005 void psi_account_irqtime(struct task_struct *task, u32 delta)
1006 {
1007 	int cpu = task_cpu(task);
1008 	struct psi_group *group;
1009 	struct psi_group_cpu *groupc;
1010 	u64 now;
1011 
1012 	if (!task->pid)
1013 		return;
1014 
1015 	now = cpu_clock(cpu);
1016 
1017 	group = task_psi_group(task);
1018 	do {
1019 		if (!group->enabled)
1020 			continue;
1021 
1022 		groupc = per_cpu_ptr(group->pcpu, cpu);
1023 
1024 		write_seqcount_begin(&groupc->seq);
1025 
1026 		record_times(groupc, now);
1027 		groupc->times[PSI_IRQ_FULL] += delta;
1028 
1029 		write_seqcount_end(&groupc->seq);
1030 
1031 		if (group->rtpoll_states & (1 << PSI_IRQ_FULL))
1032 			psi_schedule_rtpoll_work(group, 1, false);
1033 	} while ((group = group->parent));
1034 }
1035 #endif
1036 
1037 /**
1038  * psi_memstall_enter - mark the beginning of a memory stall section
1039  * @flags: flags to handle nested sections
1040  *
1041  * Marks the calling task as being stalled due to a lack of memory,
1042  * such as waiting for a refault or performing reclaim.
1043  */
1044 void psi_memstall_enter(unsigned long *flags)
1045 {
1046 	struct rq_flags rf;
1047 	struct rq *rq;
1048 
1049 	if (static_branch_likely(&psi_disabled))
1050 		return;
1051 
1052 	*flags = current->in_memstall;
1053 	if (*flags)
1054 		return;
1055 	/*
1056 	 * in_memstall setting & accounting needs to be atomic wrt
1057 	 * changes to the task's scheduling state, otherwise we can
1058 	 * race with CPU migration.
1059 	 */
1060 	rq = this_rq_lock_irq(&rf);
1061 
1062 	current->in_memstall = 1;
1063 	psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
1064 
1065 	rq_unlock_irq(rq, &rf);
1066 }
1067 EXPORT_SYMBOL_GPL(psi_memstall_enter);
1068 
1069 /**
1070  * psi_memstall_leave - mark the end of an memory stall section
1071  * @flags: flags to handle nested memdelay sections
1072  *
1073  * Marks the calling task as no longer stalled due to lack of memory.
1074  */
1075 void psi_memstall_leave(unsigned long *flags)
1076 {
1077 	struct rq_flags rf;
1078 	struct rq *rq;
1079 
1080 	if (static_branch_likely(&psi_disabled))
1081 		return;
1082 
1083 	if (*flags)
1084 		return;
1085 	/*
1086 	 * in_memstall clearing & accounting needs to be atomic wrt
1087 	 * changes to the task's scheduling state, otherwise we could
1088 	 * race with CPU migration.
1089 	 */
1090 	rq = this_rq_lock_irq(&rf);
1091 
1092 	current->in_memstall = 0;
1093 	psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0);
1094 
1095 	rq_unlock_irq(rq, &rf);
1096 }
1097 EXPORT_SYMBOL_GPL(psi_memstall_leave);
1098 
1099 #ifdef CONFIG_CGROUPS
1100 int psi_cgroup_alloc(struct cgroup *cgroup)
1101 {
1102 	if (!static_branch_likely(&psi_cgroups_enabled))
1103 		return 0;
1104 
1105 	cgroup->psi = kzalloc(sizeof(struct psi_group), GFP_KERNEL);
1106 	if (!cgroup->psi)
1107 		return -ENOMEM;
1108 
1109 	cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
1110 	if (!cgroup->psi->pcpu) {
1111 		kfree(cgroup->psi);
1112 		return -ENOMEM;
1113 	}
1114 	group_init(cgroup->psi);
1115 	cgroup->psi->parent = cgroup_psi(cgroup_parent(cgroup));
1116 	return 0;
1117 }
1118 
1119 void psi_cgroup_free(struct cgroup *cgroup)
1120 {
1121 	if (!static_branch_likely(&psi_cgroups_enabled))
1122 		return;
1123 
1124 	cancel_delayed_work_sync(&cgroup->psi->avgs_work);
1125 	free_percpu(cgroup->psi->pcpu);
1126 	/* All triggers must be removed by now */
1127 	WARN_ONCE(cgroup->psi->rtpoll_states, "psi: trigger leak\n");
1128 	kfree(cgroup->psi);
1129 }
1130 
1131 /**
1132  * cgroup_move_task - move task to a different cgroup
1133  * @task: the task
1134  * @to: the target css_set
1135  *
1136  * Move task to a new cgroup and safely migrate its associated stall
1137  * state between the different groups.
1138  *
1139  * This function acquires the task's rq lock to lock out concurrent
1140  * changes to the task's scheduling state and - in case the task is
1141  * running - concurrent changes to its stall state.
1142  */
1143 void cgroup_move_task(struct task_struct *task, struct css_set *to)
1144 {
1145 	unsigned int task_flags;
1146 	struct rq_flags rf;
1147 	struct rq *rq;
1148 
1149 	if (!static_branch_likely(&psi_cgroups_enabled)) {
1150 		/*
1151 		 * Lame to do this here, but the scheduler cannot be locked
1152 		 * from the outside, so we move cgroups from inside sched/.
1153 		 */
1154 		rcu_assign_pointer(task->cgroups, to);
1155 		return;
1156 	}
1157 
1158 	rq = task_rq_lock(task, &rf);
1159 
1160 	/*
1161 	 * We may race with schedule() dropping the rq lock between
1162 	 * deactivating prev and switching to next. Because the psi
1163 	 * updates from the deactivation are deferred to the switch
1164 	 * callback to save cgroup tree updates, the task's scheduling
1165 	 * state here is not coherent with its psi state:
1166 	 *
1167 	 * schedule()                   cgroup_move_task()
1168 	 *   rq_lock()
1169 	 *   deactivate_task()
1170 	 *     p->on_rq = 0
1171 	 *     psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1172 	 *   pick_next_task()
1173 	 *     rq_unlock()
1174 	 *                                rq_lock()
1175 	 *                                psi_task_change() // old cgroup
1176 	 *                                task->cgroups = to
1177 	 *                                psi_task_change() // new cgroup
1178 	 *                                rq_unlock()
1179 	 *     rq_lock()
1180 	 *   psi_sched_switch() // does deferred updates in new cgroup
1181 	 *
1182 	 * Don't rely on the scheduling state. Use psi_flags instead.
1183 	 */
1184 	task_flags = task->psi_flags;
1185 
1186 	if (task_flags)
1187 		psi_task_change(task, task_flags, 0);
1188 
1189 	/* See comment above */
1190 	rcu_assign_pointer(task->cgroups, to);
1191 
1192 	if (task_flags)
1193 		psi_task_change(task, 0, task_flags);
1194 
1195 	task_rq_unlock(rq, task, &rf);
1196 }
1197 
1198 void psi_cgroup_restart(struct psi_group *group)
1199 {
1200 	int cpu;
1201 
1202 	/*
1203 	 * After we disable psi_group->enabled, we don't actually
1204 	 * stop percpu tasks accounting in each psi_group_cpu,
1205 	 * instead only stop test_state() loop, record_times()
1206 	 * and averaging worker, see psi_group_change() for details.
1207 	 *
1208 	 * When disable cgroup PSI, this function has nothing to sync
1209 	 * since cgroup pressure files are hidden and percpu psi_group_cpu
1210 	 * would see !psi_group->enabled and only do task accounting.
1211 	 *
1212 	 * When re-enable cgroup PSI, this function use psi_group_change()
1213 	 * to get correct state mask from test_state() loop on tasks[],
1214 	 * and restart groupc->state_start from now, use .clear = .set = 0
1215 	 * here since no task status really changed.
1216 	 */
1217 	if (!group->enabled)
1218 		return;
1219 
1220 	for_each_possible_cpu(cpu) {
1221 		struct rq *rq = cpu_rq(cpu);
1222 		struct rq_flags rf;
1223 		u64 now;
1224 
1225 		rq_lock_irq(rq, &rf);
1226 		now = cpu_clock(cpu);
1227 		psi_group_change(group, cpu, 0, 0, now, true);
1228 		rq_unlock_irq(rq, &rf);
1229 	}
1230 }
1231 #endif /* CONFIG_CGROUPS */
1232 
1233 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1234 {
1235 	bool only_full = false;
1236 	int full;
1237 	u64 now;
1238 
1239 	if (static_branch_likely(&psi_disabled))
1240 		return -EOPNOTSUPP;
1241 
1242 	/* Update averages before reporting them */
1243 	mutex_lock(&group->avgs_lock);
1244 	now = sched_clock();
1245 	collect_percpu_times(group, PSI_AVGS, NULL);
1246 	if (now >= group->avg_next_update)
1247 		group->avg_next_update = update_averages(group, now);
1248 	mutex_unlock(&group->avgs_lock);
1249 
1250 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1251 	only_full = res == PSI_IRQ;
1252 #endif
1253 
1254 	for (full = 0; full < 2 - only_full; full++) {
1255 		unsigned long avg[3] = { 0, };
1256 		u64 total = 0;
1257 		int w;
1258 
1259 		/* CPU FULL is undefined at the system level */
1260 		if (!(group == &psi_system && res == PSI_CPU && full)) {
1261 			for (w = 0; w < 3; w++)
1262 				avg[w] = group->avg[res * 2 + full][w];
1263 			total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1264 					NSEC_PER_USEC);
1265 		}
1266 
1267 		seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1268 			   full || only_full ? "full" : "some",
1269 			   LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1270 			   LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1271 			   LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1272 			   total);
1273 	}
1274 
1275 	return 0;
1276 }
1277 
1278 struct psi_trigger *psi_trigger_create(struct psi_group *group, char *buf,
1279 				       enum psi_res res, struct file *file,
1280 				       struct kernfs_open_file *of)
1281 {
1282 	struct psi_trigger *t;
1283 	enum psi_states state;
1284 	u32 threshold_us;
1285 	bool privileged;
1286 	u32 window_us;
1287 
1288 	if (static_branch_likely(&psi_disabled))
1289 		return ERR_PTR(-EOPNOTSUPP);
1290 
1291 	/*
1292 	 * Checking the privilege here on file->f_cred implies that a privileged user
1293 	 * could open the file and delegate the write to an unprivileged one.
1294 	 */
1295 	privileged = cap_raised(file->f_cred->cap_effective, CAP_SYS_RESOURCE);
1296 
1297 	if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1298 		state = PSI_IO_SOME + res * 2;
1299 	else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1300 		state = PSI_IO_FULL + res * 2;
1301 	else
1302 		return ERR_PTR(-EINVAL);
1303 
1304 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1305 	if (res == PSI_IRQ && --state != PSI_IRQ_FULL)
1306 		return ERR_PTR(-EINVAL);
1307 #endif
1308 
1309 	if (state >= PSI_NONIDLE)
1310 		return ERR_PTR(-EINVAL);
1311 
1312 	if (window_us == 0 || window_us > WINDOW_MAX_US)
1313 		return ERR_PTR(-EINVAL);
1314 
1315 	/*
1316 	 * Unprivileged users can only use 2s windows so that averages aggregation
1317 	 * work is used, and no RT threads need to be spawned.
1318 	 */
1319 	if (!privileged && window_us % 2000000)
1320 		return ERR_PTR(-EINVAL);
1321 
1322 	/* Check threshold */
1323 	if (threshold_us == 0 || threshold_us > window_us)
1324 		return ERR_PTR(-EINVAL);
1325 
1326 	t = kmalloc(sizeof(*t), GFP_KERNEL);
1327 	if (!t)
1328 		return ERR_PTR(-ENOMEM);
1329 
1330 	t->group = group;
1331 	t->state = state;
1332 	t->threshold = threshold_us * NSEC_PER_USEC;
1333 	t->win.size = window_us * NSEC_PER_USEC;
1334 	window_reset(&t->win, sched_clock(),
1335 			group->total[PSI_POLL][t->state], 0);
1336 
1337 	t->event = 0;
1338 	t->last_event_time = 0;
1339 	t->of = of;
1340 	if (!of)
1341 		init_waitqueue_head(&t->event_wait);
1342 	t->pending_event = false;
1343 	t->aggregator = privileged ? PSI_POLL : PSI_AVGS;
1344 
1345 	if (privileged) {
1346 		mutex_lock(&group->rtpoll_trigger_lock);
1347 
1348 		if (!rcu_access_pointer(group->rtpoll_task)) {
1349 			struct task_struct *task;
1350 
1351 			task = kthread_create(psi_rtpoll_worker, group, "psimon");
1352 			if (IS_ERR(task)) {
1353 				kfree(t);
1354 				mutex_unlock(&group->rtpoll_trigger_lock);
1355 				return ERR_CAST(task);
1356 			}
1357 			atomic_set(&group->rtpoll_wakeup, 0);
1358 			wake_up_process(task);
1359 			rcu_assign_pointer(group->rtpoll_task, task);
1360 		}
1361 
1362 		list_add(&t->node, &group->rtpoll_triggers);
1363 		group->rtpoll_min_period = min(group->rtpoll_min_period,
1364 			div_u64(t->win.size, UPDATES_PER_WINDOW));
1365 		group->rtpoll_nr_triggers[t->state]++;
1366 		group->rtpoll_states |= (1 << t->state);
1367 
1368 		mutex_unlock(&group->rtpoll_trigger_lock);
1369 	} else {
1370 		mutex_lock(&group->avgs_lock);
1371 
1372 		list_add(&t->node, &group->avg_triggers);
1373 		group->avg_nr_triggers[t->state]++;
1374 
1375 		mutex_unlock(&group->avgs_lock);
1376 	}
1377 	return t;
1378 }
1379 
1380 void psi_trigger_destroy(struct psi_trigger *t)
1381 {
1382 	struct psi_group *group;
1383 	struct task_struct *task_to_destroy = NULL;
1384 
1385 	/*
1386 	 * We do not check psi_disabled since it might have been disabled after
1387 	 * the trigger got created.
1388 	 */
1389 	if (!t)
1390 		return;
1391 
1392 	group = t->group;
1393 	/*
1394 	 * Wakeup waiters to stop polling and clear the queue to prevent it from
1395 	 * being accessed later. Can happen if cgroup is deleted from under a
1396 	 * polling process.
1397 	 */
1398 	if (t->of)
1399 		kernfs_notify(t->of->kn);
1400 	else
1401 		wake_up_interruptible(&t->event_wait);
1402 
1403 	if (t->aggregator == PSI_AVGS) {
1404 		mutex_lock(&group->avgs_lock);
1405 		if (!list_empty(&t->node)) {
1406 			list_del(&t->node);
1407 			group->avg_nr_triggers[t->state]--;
1408 		}
1409 		mutex_unlock(&group->avgs_lock);
1410 	} else {
1411 		mutex_lock(&group->rtpoll_trigger_lock);
1412 		if (!list_empty(&t->node)) {
1413 			struct psi_trigger *tmp;
1414 			u64 period = ULLONG_MAX;
1415 
1416 			list_del(&t->node);
1417 			group->rtpoll_nr_triggers[t->state]--;
1418 			if (!group->rtpoll_nr_triggers[t->state])
1419 				group->rtpoll_states &= ~(1 << t->state);
1420 			/*
1421 			 * Reset min update period for the remaining triggers
1422 			 * iff the destroying trigger had the min window size.
1423 			 */
1424 			if (group->rtpoll_min_period == div_u64(t->win.size, UPDATES_PER_WINDOW)) {
1425 				list_for_each_entry(tmp, &group->rtpoll_triggers, node)
1426 					period = min(period, div_u64(tmp->win.size,
1427 							UPDATES_PER_WINDOW));
1428 				group->rtpoll_min_period = period;
1429 			}
1430 			/* Destroy rtpoll_task when the last trigger is destroyed */
1431 			if (group->rtpoll_states == 0) {
1432 				group->rtpoll_until = 0;
1433 				task_to_destroy = rcu_dereference_protected(
1434 						group->rtpoll_task,
1435 						lockdep_is_held(&group->rtpoll_trigger_lock));
1436 				rcu_assign_pointer(group->rtpoll_task, NULL);
1437 				del_timer(&group->rtpoll_timer);
1438 			}
1439 		}
1440 		mutex_unlock(&group->rtpoll_trigger_lock);
1441 	}
1442 
1443 	/*
1444 	 * Wait for psi_schedule_rtpoll_work RCU to complete its read-side
1445 	 * critical section before destroying the trigger and optionally the
1446 	 * rtpoll_task.
1447 	 */
1448 	synchronize_rcu();
1449 	/*
1450 	 * Stop kthread 'psimon' after releasing rtpoll_trigger_lock to prevent
1451 	 * a deadlock while waiting for psi_rtpoll_work to acquire
1452 	 * rtpoll_trigger_lock
1453 	 */
1454 	if (task_to_destroy) {
1455 		/*
1456 		 * After the RCU grace period has expired, the worker
1457 		 * can no longer be found through group->rtpoll_task.
1458 		 */
1459 		kthread_stop(task_to_destroy);
1460 		atomic_set(&group->rtpoll_scheduled, 0);
1461 	}
1462 	kfree(t);
1463 }
1464 
1465 __poll_t psi_trigger_poll(void **trigger_ptr,
1466 				struct file *file, poll_table *wait)
1467 {
1468 	__poll_t ret = DEFAULT_POLLMASK;
1469 	struct psi_trigger *t;
1470 
1471 	if (static_branch_likely(&psi_disabled))
1472 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1473 
1474 	t = smp_load_acquire(trigger_ptr);
1475 	if (!t)
1476 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1477 
1478 	if (t->of)
1479 		kernfs_generic_poll(t->of, wait);
1480 	else
1481 		poll_wait(file, &t->event_wait, wait);
1482 
1483 	if (cmpxchg(&t->event, 1, 0) == 1)
1484 		ret |= EPOLLPRI;
1485 
1486 	return ret;
1487 }
1488 
1489 #ifdef CONFIG_PROC_FS
1490 static int psi_io_show(struct seq_file *m, void *v)
1491 {
1492 	return psi_show(m, &psi_system, PSI_IO);
1493 }
1494 
1495 static int psi_memory_show(struct seq_file *m, void *v)
1496 {
1497 	return psi_show(m, &psi_system, PSI_MEM);
1498 }
1499 
1500 static int psi_cpu_show(struct seq_file *m, void *v)
1501 {
1502 	return psi_show(m, &psi_system, PSI_CPU);
1503 }
1504 
1505 static int psi_io_open(struct inode *inode, struct file *file)
1506 {
1507 	return single_open(file, psi_io_show, NULL);
1508 }
1509 
1510 static int psi_memory_open(struct inode *inode, struct file *file)
1511 {
1512 	return single_open(file, psi_memory_show, NULL);
1513 }
1514 
1515 static int psi_cpu_open(struct inode *inode, struct file *file)
1516 {
1517 	return single_open(file, psi_cpu_show, NULL);
1518 }
1519 
1520 static ssize_t psi_write(struct file *file, const char __user *user_buf,
1521 			 size_t nbytes, enum psi_res res)
1522 {
1523 	char buf[32];
1524 	size_t buf_size;
1525 	struct seq_file *seq;
1526 	struct psi_trigger *new;
1527 
1528 	if (static_branch_likely(&psi_disabled))
1529 		return -EOPNOTSUPP;
1530 
1531 	if (!nbytes)
1532 		return -EINVAL;
1533 
1534 	buf_size = min(nbytes, sizeof(buf));
1535 	if (copy_from_user(buf, user_buf, buf_size))
1536 		return -EFAULT;
1537 
1538 	buf[buf_size - 1] = '\0';
1539 
1540 	seq = file->private_data;
1541 
1542 	/* Take seq->lock to protect seq->private from concurrent writes */
1543 	mutex_lock(&seq->lock);
1544 
1545 	/* Allow only one trigger per file descriptor */
1546 	if (seq->private) {
1547 		mutex_unlock(&seq->lock);
1548 		return -EBUSY;
1549 	}
1550 
1551 	new = psi_trigger_create(&psi_system, buf, res, file, NULL);
1552 	if (IS_ERR(new)) {
1553 		mutex_unlock(&seq->lock);
1554 		return PTR_ERR(new);
1555 	}
1556 
1557 	smp_store_release(&seq->private, new);
1558 	mutex_unlock(&seq->lock);
1559 
1560 	return nbytes;
1561 }
1562 
1563 static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1564 			    size_t nbytes, loff_t *ppos)
1565 {
1566 	return psi_write(file, user_buf, nbytes, PSI_IO);
1567 }
1568 
1569 static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1570 				size_t nbytes, loff_t *ppos)
1571 {
1572 	return psi_write(file, user_buf, nbytes, PSI_MEM);
1573 }
1574 
1575 static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1576 			     size_t nbytes, loff_t *ppos)
1577 {
1578 	return psi_write(file, user_buf, nbytes, PSI_CPU);
1579 }
1580 
1581 static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1582 {
1583 	struct seq_file *seq = file->private_data;
1584 
1585 	return psi_trigger_poll(&seq->private, file, wait);
1586 }
1587 
1588 static int psi_fop_release(struct inode *inode, struct file *file)
1589 {
1590 	struct seq_file *seq = file->private_data;
1591 
1592 	psi_trigger_destroy(seq->private);
1593 	return single_release(inode, file);
1594 }
1595 
1596 static const struct proc_ops psi_io_proc_ops = {
1597 	.proc_open	= psi_io_open,
1598 	.proc_read	= seq_read,
1599 	.proc_lseek	= seq_lseek,
1600 	.proc_write	= psi_io_write,
1601 	.proc_poll	= psi_fop_poll,
1602 	.proc_release	= psi_fop_release,
1603 };
1604 
1605 static const struct proc_ops psi_memory_proc_ops = {
1606 	.proc_open	= psi_memory_open,
1607 	.proc_read	= seq_read,
1608 	.proc_lseek	= seq_lseek,
1609 	.proc_write	= psi_memory_write,
1610 	.proc_poll	= psi_fop_poll,
1611 	.proc_release	= psi_fop_release,
1612 };
1613 
1614 static const struct proc_ops psi_cpu_proc_ops = {
1615 	.proc_open	= psi_cpu_open,
1616 	.proc_read	= seq_read,
1617 	.proc_lseek	= seq_lseek,
1618 	.proc_write	= psi_cpu_write,
1619 	.proc_poll	= psi_fop_poll,
1620 	.proc_release	= psi_fop_release,
1621 };
1622 
1623 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1624 static int psi_irq_show(struct seq_file *m, void *v)
1625 {
1626 	return psi_show(m, &psi_system, PSI_IRQ);
1627 }
1628 
1629 static int psi_irq_open(struct inode *inode, struct file *file)
1630 {
1631 	return single_open(file, psi_irq_show, NULL);
1632 }
1633 
1634 static ssize_t psi_irq_write(struct file *file, const char __user *user_buf,
1635 			     size_t nbytes, loff_t *ppos)
1636 {
1637 	return psi_write(file, user_buf, nbytes, PSI_IRQ);
1638 }
1639 
1640 static const struct proc_ops psi_irq_proc_ops = {
1641 	.proc_open	= psi_irq_open,
1642 	.proc_read	= seq_read,
1643 	.proc_lseek	= seq_lseek,
1644 	.proc_write	= psi_irq_write,
1645 	.proc_poll	= psi_fop_poll,
1646 	.proc_release	= psi_fop_release,
1647 };
1648 #endif
1649 
1650 static int __init psi_proc_init(void)
1651 {
1652 	if (psi_enable) {
1653 		proc_mkdir("pressure", NULL);
1654 		proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops);
1655 		proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops);
1656 		proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops);
1657 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1658 		proc_create("pressure/irq", 0666, NULL, &psi_irq_proc_ops);
1659 #endif
1660 	}
1661 	return 0;
1662 }
1663 module_init(psi_proc_init);
1664 
1665 #endif /* CONFIG_PROC_FS */
1666