xref: /openbmc/linux/kernel/sched/psi.c (revision 4c8b18af)
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_MIN_US 500000	/* Min window size is 500ms */
164 #define WINDOW_MAX_US 10000000	/* Max window size is 10s */
165 #define UPDATES_PER_WINDOW 10	/* 10 updates per window */
166 
167 /* Sampling frequency in nanoseconds */
168 static u64 psi_period __read_mostly;
169 
170 /* System-level pressure and stall tracking */
171 static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
172 struct psi_group psi_system = {
173 	.pcpu = &system_group_pcpu,
174 };
175 
176 static void psi_avgs_work(struct work_struct *work);
177 
178 static void poll_timer_fn(struct timer_list *t);
179 
180 static void group_init(struct psi_group *group)
181 {
182 	int cpu;
183 
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 	INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
189 	mutex_init(&group->avgs_lock);
190 	/* Init trigger-related members */
191 	mutex_init(&group->trigger_lock);
192 	INIT_LIST_HEAD(&group->triggers);
193 	group->poll_min_period = U32_MAX;
194 	group->polling_next_update = ULLONG_MAX;
195 	init_waitqueue_head(&group->poll_wait);
196 	timer_setup(&group->poll_timer, poll_timer_fn, 0);
197 	rcu_assign_pointer(group->poll_task, NULL);
198 }
199 
200 void __init psi_init(void)
201 {
202 	if (!psi_enable) {
203 		static_branch_enable(&psi_disabled);
204 		return;
205 	}
206 
207 	if (!cgroup_psi_enabled())
208 		static_branch_disable(&psi_cgroups_enabled);
209 
210 	psi_period = jiffies_to_nsecs(PSI_FREQ);
211 	group_init(&psi_system);
212 }
213 
214 static bool test_state(unsigned int *tasks, enum psi_states state)
215 {
216 	switch (state) {
217 	case PSI_IO_SOME:
218 		return unlikely(tasks[NR_IOWAIT]);
219 	case PSI_IO_FULL:
220 		return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]);
221 	case PSI_MEM_SOME:
222 		return unlikely(tasks[NR_MEMSTALL]);
223 	case PSI_MEM_FULL:
224 		return unlikely(tasks[NR_MEMSTALL] &&
225 			tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]);
226 	case PSI_CPU_SOME:
227 		return unlikely(tasks[NR_RUNNING] > tasks[NR_ONCPU]);
228 	case PSI_CPU_FULL:
229 		return unlikely(tasks[NR_RUNNING] && !tasks[NR_ONCPU]);
230 	case PSI_NONIDLE:
231 		return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
232 			tasks[NR_RUNNING];
233 	default:
234 		return false;
235 	}
236 }
237 
238 static void get_recent_times(struct psi_group *group, int cpu,
239 			     enum psi_aggregators aggregator, u32 *times,
240 			     u32 *pchanged_states)
241 {
242 	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
243 	u64 now, state_start;
244 	enum psi_states s;
245 	unsigned int seq;
246 	u32 state_mask;
247 
248 	*pchanged_states = 0;
249 
250 	/* Snapshot a coherent view of the CPU state */
251 	do {
252 		seq = read_seqcount_begin(&groupc->seq);
253 		now = cpu_clock(cpu);
254 		memcpy(times, groupc->times, sizeof(groupc->times));
255 		state_mask = groupc->state_mask;
256 		state_start = groupc->state_start;
257 	} while (read_seqcount_retry(&groupc->seq, seq));
258 
259 	/* Calculate state time deltas against the previous snapshot */
260 	for (s = 0; s < NR_PSI_STATES; s++) {
261 		u32 delta;
262 		/*
263 		 * In addition to already concluded states, we also
264 		 * incorporate currently active states on the CPU,
265 		 * since states may last for many sampling periods.
266 		 *
267 		 * This way we keep our delta sampling buckets small
268 		 * (u32) and our reported pressure close to what's
269 		 * actually happening.
270 		 */
271 		if (state_mask & (1 << s))
272 			times[s] += now - state_start;
273 
274 		delta = times[s] - groupc->times_prev[aggregator][s];
275 		groupc->times_prev[aggregator][s] = times[s];
276 
277 		times[s] = delta;
278 		if (delta)
279 			*pchanged_states |= (1 << s);
280 	}
281 }
282 
283 static void calc_avgs(unsigned long avg[3], int missed_periods,
284 		      u64 time, u64 period)
285 {
286 	unsigned long pct;
287 
288 	/* Fill in zeroes for periods of no activity */
289 	if (missed_periods) {
290 		avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
291 		avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
292 		avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
293 	}
294 
295 	/* Sample the most recent active period */
296 	pct = div_u64(time * 100, period);
297 	pct *= FIXED_1;
298 	avg[0] = calc_load(avg[0], EXP_10s, pct);
299 	avg[1] = calc_load(avg[1], EXP_60s, pct);
300 	avg[2] = calc_load(avg[2], EXP_300s, pct);
301 }
302 
303 static void collect_percpu_times(struct psi_group *group,
304 				 enum psi_aggregators aggregator,
305 				 u32 *pchanged_states)
306 {
307 	u64 deltas[NR_PSI_STATES - 1] = { 0, };
308 	unsigned long nonidle_total = 0;
309 	u32 changed_states = 0;
310 	int cpu;
311 	int s;
312 
313 	/*
314 	 * Collect the per-cpu time buckets and average them into a
315 	 * single time sample that is normalized to wallclock time.
316 	 *
317 	 * For averaging, each CPU is weighted by its non-idle time in
318 	 * the sampling period. This eliminates artifacts from uneven
319 	 * loading, or even entirely idle CPUs.
320 	 */
321 	for_each_possible_cpu(cpu) {
322 		u32 times[NR_PSI_STATES];
323 		u32 nonidle;
324 		u32 cpu_changed_states;
325 
326 		get_recent_times(group, cpu, aggregator, times,
327 				&cpu_changed_states);
328 		changed_states |= cpu_changed_states;
329 
330 		nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
331 		nonidle_total += nonidle;
332 
333 		for (s = 0; s < PSI_NONIDLE; s++)
334 			deltas[s] += (u64)times[s] * nonidle;
335 	}
336 
337 	/*
338 	 * Integrate the sample into the running statistics that are
339 	 * reported to userspace: the cumulative stall times and the
340 	 * decaying averages.
341 	 *
342 	 * Pressure percentages are sampled at PSI_FREQ. We might be
343 	 * called more often when the user polls more frequently than
344 	 * that; we might be called less often when there is no task
345 	 * activity, thus no data, and clock ticks are sporadic. The
346 	 * below handles both.
347 	 */
348 
349 	/* total= */
350 	for (s = 0; s < NR_PSI_STATES - 1; s++)
351 		group->total[aggregator][s] +=
352 				div_u64(deltas[s], max(nonidle_total, 1UL));
353 
354 	if (pchanged_states)
355 		*pchanged_states = changed_states;
356 }
357 
358 static u64 update_averages(struct psi_group *group, u64 now)
359 {
360 	unsigned long missed_periods = 0;
361 	u64 expires, period;
362 	u64 avg_next_update;
363 	int s;
364 
365 	/* avgX= */
366 	expires = group->avg_next_update;
367 	if (now - expires >= psi_period)
368 		missed_periods = div_u64(now - expires, psi_period);
369 
370 	/*
371 	 * The periodic clock tick can get delayed for various
372 	 * reasons, especially on loaded systems. To avoid clock
373 	 * drift, we schedule the clock in fixed psi_period intervals.
374 	 * But the deltas we sample out of the per-cpu buckets above
375 	 * are based on the actual time elapsing between clock ticks.
376 	 */
377 	avg_next_update = expires + ((1 + missed_periods) * psi_period);
378 	period = now - (group->avg_last_update + (missed_periods * psi_period));
379 	group->avg_last_update = now;
380 
381 	for (s = 0; s < NR_PSI_STATES - 1; s++) {
382 		u32 sample;
383 
384 		sample = group->total[PSI_AVGS][s] - group->avg_total[s];
385 		/*
386 		 * Due to the lockless sampling of the time buckets,
387 		 * recorded time deltas can slip into the next period,
388 		 * which under full pressure can result in samples in
389 		 * excess of the period length.
390 		 *
391 		 * We don't want to report non-sensical pressures in
392 		 * excess of 100%, nor do we want to drop such events
393 		 * on the floor. Instead we punt any overage into the
394 		 * future until pressure subsides. By doing this we
395 		 * don't underreport the occurring pressure curve, we
396 		 * just report it delayed by one period length.
397 		 *
398 		 * The error isn't cumulative. As soon as another
399 		 * delta slips from a period P to P+1, by definition
400 		 * it frees up its time T in P.
401 		 */
402 		if (sample > period)
403 			sample = period;
404 		group->avg_total[s] += sample;
405 		calc_avgs(group->avg[s], missed_periods, sample, period);
406 	}
407 
408 	return avg_next_update;
409 }
410 
411 static void psi_avgs_work(struct work_struct *work)
412 {
413 	struct delayed_work *dwork;
414 	struct psi_group *group;
415 	u32 changed_states;
416 	bool nonidle;
417 	u64 now;
418 
419 	dwork = to_delayed_work(work);
420 	group = container_of(dwork, struct psi_group, avgs_work);
421 
422 	mutex_lock(&group->avgs_lock);
423 
424 	now = sched_clock();
425 
426 	collect_percpu_times(group, PSI_AVGS, &changed_states);
427 	nonidle = changed_states & (1 << PSI_NONIDLE);
428 	/*
429 	 * If there is task activity, periodically fold the per-cpu
430 	 * times and feed samples into the running averages. If things
431 	 * are idle and there is no data to process, stop the clock.
432 	 * Once restarted, we'll catch up the running averages in one
433 	 * go - see calc_avgs() and missed_periods.
434 	 */
435 	if (now >= group->avg_next_update)
436 		group->avg_next_update = update_averages(group, now);
437 
438 	if (nonidle) {
439 		schedule_delayed_work(dwork, nsecs_to_jiffies(
440 				group->avg_next_update - now) + 1);
441 	}
442 
443 	mutex_unlock(&group->avgs_lock);
444 }
445 
446 /* Trigger tracking window manipulations */
447 static void window_reset(struct psi_window *win, u64 now, u64 value,
448 			 u64 prev_growth)
449 {
450 	win->start_time = now;
451 	win->start_value = value;
452 	win->prev_growth = prev_growth;
453 }
454 
455 /*
456  * PSI growth tracking window update and growth calculation routine.
457  *
458  * This approximates a sliding tracking window by interpolating
459  * partially elapsed windows using historical growth data from the
460  * previous intervals. This minimizes memory requirements (by not storing
461  * all the intermediate values in the previous window) and simplifies
462  * the calculations. It works well because PSI signal changes only in
463  * positive direction and over relatively small window sizes the growth
464  * is close to linear.
465  */
466 static u64 window_update(struct psi_window *win, u64 now, u64 value)
467 {
468 	u64 elapsed;
469 	u64 growth;
470 
471 	elapsed = now - win->start_time;
472 	growth = value - win->start_value;
473 	/*
474 	 * After each tracking window passes win->start_value and
475 	 * win->start_time get reset and win->prev_growth stores
476 	 * the average per-window growth of the previous window.
477 	 * win->prev_growth is then used to interpolate additional
478 	 * growth from the previous window assuming it was linear.
479 	 */
480 	if (elapsed > win->size)
481 		window_reset(win, now, value, growth);
482 	else {
483 		u32 remaining;
484 
485 		remaining = win->size - elapsed;
486 		growth += div64_u64(win->prev_growth * remaining, win->size);
487 	}
488 
489 	return growth;
490 }
491 
492 static void init_triggers(struct psi_group *group, u64 now)
493 {
494 	struct psi_trigger *t;
495 
496 	list_for_each_entry(t, &group->triggers, node)
497 		window_reset(&t->win, now,
498 				group->total[PSI_POLL][t->state], 0);
499 	memcpy(group->polling_total, group->total[PSI_POLL],
500 		   sizeof(group->polling_total));
501 	group->polling_next_update = now + group->poll_min_period;
502 }
503 
504 static u64 update_triggers(struct psi_group *group, u64 now)
505 {
506 	struct psi_trigger *t;
507 	bool update_total = false;
508 	u64 *total = group->total[PSI_POLL];
509 
510 	/*
511 	 * On subsequent updates, calculate growth deltas and let
512 	 * watchers know when their specified thresholds are exceeded.
513 	 */
514 	list_for_each_entry(t, &group->triggers, node) {
515 		u64 growth;
516 		bool new_stall;
517 
518 		new_stall = group->polling_total[t->state] != total[t->state];
519 
520 		/* Check for stall activity or a previous threshold breach */
521 		if (!new_stall && !t->pending_event)
522 			continue;
523 		/*
524 		 * Check for new stall activity, as well as deferred
525 		 * events that occurred in the last window after the
526 		 * trigger had already fired (we want to ratelimit
527 		 * events without dropping any).
528 		 */
529 		if (new_stall) {
530 			/*
531 			 * Multiple triggers might be looking at the same state,
532 			 * remember to update group->polling_total[] once we've
533 			 * been through all of them. Also remember to extend the
534 			 * polling time if we see new stall activity.
535 			 */
536 			update_total = true;
537 
538 			/* Calculate growth since last update */
539 			growth = window_update(&t->win, now, total[t->state]);
540 			if (growth < t->threshold)
541 				continue;
542 
543 			t->pending_event = true;
544 		}
545 		/* Limit event signaling to once per window */
546 		if (now < t->last_event_time + t->win.size)
547 			continue;
548 
549 		/* Generate an event */
550 		if (cmpxchg(&t->event, 0, 1) == 0)
551 			wake_up_interruptible(&t->event_wait);
552 		t->last_event_time = now;
553 		/* Reset threshold breach flag once event got generated */
554 		t->pending_event = false;
555 	}
556 
557 	if (update_total)
558 		memcpy(group->polling_total, total,
559 				sizeof(group->polling_total));
560 
561 	return now + group->poll_min_period;
562 }
563 
564 /* Schedule polling if it's not already scheduled. */
565 static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay)
566 {
567 	struct task_struct *task;
568 
569 	/*
570 	 * Do not reschedule if already scheduled.
571 	 * Possible race with a timer scheduled after this check but before
572 	 * mod_timer below can be tolerated because group->polling_next_update
573 	 * will keep updates on schedule.
574 	 */
575 	if (timer_pending(&group->poll_timer))
576 		return;
577 
578 	rcu_read_lock();
579 
580 	task = rcu_dereference(group->poll_task);
581 	/*
582 	 * kworker might be NULL in case psi_trigger_destroy races with
583 	 * psi_task_change (hotpath) which can't use locks
584 	 */
585 	if (likely(task))
586 		mod_timer(&group->poll_timer, jiffies + delay);
587 
588 	rcu_read_unlock();
589 }
590 
591 static void psi_poll_work(struct psi_group *group)
592 {
593 	u32 changed_states;
594 	u64 now;
595 
596 	mutex_lock(&group->trigger_lock);
597 
598 	now = sched_clock();
599 
600 	collect_percpu_times(group, PSI_POLL, &changed_states);
601 
602 	if (changed_states & group->poll_states) {
603 		/* Initialize trigger windows when entering polling mode */
604 		if (now > group->polling_until)
605 			init_triggers(group, now);
606 
607 		/*
608 		 * Keep the monitor active for at least the duration of the
609 		 * minimum tracking window as long as monitor states are
610 		 * changing.
611 		 */
612 		group->polling_until = now +
613 			group->poll_min_period * UPDATES_PER_WINDOW;
614 	}
615 
616 	if (now > group->polling_until) {
617 		group->polling_next_update = ULLONG_MAX;
618 		goto out;
619 	}
620 
621 	if (now >= group->polling_next_update)
622 		group->polling_next_update = update_triggers(group, now);
623 
624 	psi_schedule_poll_work(group,
625 		nsecs_to_jiffies(group->polling_next_update - now) + 1);
626 
627 out:
628 	mutex_unlock(&group->trigger_lock);
629 }
630 
631 static int psi_poll_worker(void *data)
632 {
633 	struct psi_group *group = (struct psi_group *)data;
634 
635 	sched_set_fifo_low(current);
636 
637 	while (true) {
638 		wait_event_interruptible(group->poll_wait,
639 				atomic_cmpxchg(&group->poll_wakeup, 1, 0) ||
640 				kthread_should_stop());
641 		if (kthread_should_stop())
642 			break;
643 
644 		psi_poll_work(group);
645 	}
646 	return 0;
647 }
648 
649 static void poll_timer_fn(struct timer_list *t)
650 {
651 	struct psi_group *group = from_timer(group, t, poll_timer);
652 
653 	atomic_set(&group->poll_wakeup, 1);
654 	wake_up_interruptible(&group->poll_wait);
655 }
656 
657 static void record_times(struct psi_group_cpu *groupc, u64 now)
658 {
659 	u32 delta;
660 
661 	delta = now - groupc->state_start;
662 	groupc->state_start = now;
663 
664 	if (groupc->state_mask & (1 << PSI_IO_SOME)) {
665 		groupc->times[PSI_IO_SOME] += delta;
666 		if (groupc->state_mask & (1 << PSI_IO_FULL))
667 			groupc->times[PSI_IO_FULL] += delta;
668 	}
669 
670 	if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
671 		groupc->times[PSI_MEM_SOME] += delta;
672 		if (groupc->state_mask & (1 << PSI_MEM_FULL))
673 			groupc->times[PSI_MEM_FULL] += delta;
674 	}
675 
676 	if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
677 		groupc->times[PSI_CPU_SOME] += delta;
678 		if (groupc->state_mask & (1 << PSI_CPU_FULL))
679 			groupc->times[PSI_CPU_FULL] += delta;
680 	}
681 
682 	if (groupc->state_mask & (1 << PSI_NONIDLE))
683 		groupc->times[PSI_NONIDLE] += delta;
684 }
685 
686 static void psi_group_change(struct psi_group *group, int cpu,
687 			     unsigned int clear, unsigned int set, u64 now,
688 			     bool wake_clock)
689 {
690 	struct psi_group_cpu *groupc;
691 	u32 state_mask = 0;
692 	unsigned int t, m;
693 	enum psi_states s;
694 
695 	groupc = per_cpu_ptr(group->pcpu, cpu);
696 
697 	/*
698 	 * First we assess the aggregate resource states this CPU's
699 	 * tasks have been in since the last change, and account any
700 	 * SOME and FULL time these may have resulted in.
701 	 *
702 	 * Then we update the task counts according to the state
703 	 * change requested through the @clear and @set bits.
704 	 */
705 	write_seqcount_begin(&groupc->seq);
706 
707 	record_times(groupc, now);
708 
709 	for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
710 		if (!(m & (1 << t)))
711 			continue;
712 		if (groupc->tasks[t]) {
713 			groupc->tasks[t]--;
714 		} else if (!psi_bug) {
715 			printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u %u] clear=%x set=%x\n",
716 					cpu, t, groupc->tasks[0],
717 					groupc->tasks[1], groupc->tasks[2],
718 					groupc->tasks[3], groupc->tasks[4],
719 					clear, set);
720 			psi_bug = 1;
721 		}
722 	}
723 
724 	for (t = 0; set; set &= ~(1 << t), t++)
725 		if (set & (1 << t))
726 			groupc->tasks[t]++;
727 
728 	/* Calculate state mask representing active states */
729 	for (s = 0; s < NR_PSI_STATES; s++) {
730 		if (test_state(groupc->tasks, s))
731 			state_mask |= (1 << s);
732 	}
733 
734 	/*
735 	 * Since we care about lost potential, a memstall is FULL
736 	 * when there are no other working tasks, but also when
737 	 * the CPU is actively reclaiming and nothing productive
738 	 * could run even if it were runnable. So when the current
739 	 * task in a cgroup is in_memstall, the corresponding groupc
740 	 * on that cpu is in PSI_MEM_FULL state.
741 	 */
742 	if (unlikely(groupc->tasks[NR_ONCPU] && cpu_curr(cpu)->in_memstall))
743 		state_mask |= (1 << PSI_MEM_FULL);
744 
745 	groupc->state_mask = state_mask;
746 
747 	write_seqcount_end(&groupc->seq);
748 
749 	if (state_mask & group->poll_states)
750 		psi_schedule_poll_work(group, 1);
751 
752 	if (wake_clock && !delayed_work_pending(&group->avgs_work))
753 		schedule_delayed_work(&group->avgs_work, PSI_FREQ);
754 }
755 
756 static struct psi_group *iterate_groups(struct task_struct *task, void **iter)
757 {
758 	if (*iter == &psi_system)
759 		return NULL;
760 
761 #ifdef CONFIG_CGROUPS
762 	if (static_branch_likely(&psi_cgroups_enabled)) {
763 		struct cgroup *cgroup = NULL;
764 
765 		if (!*iter)
766 			cgroup = task->cgroups->dfl_cgrp;
767 		else
768 			cgroup = cgroup_parent(*iter);
769 
770 		if (cgroup && cgroup_parent(cgroup)) {
771 			*iter = cgroup;
772 			return cgroup_psi(cgroup);
773 		}
774 	}
775 #endif
776 	*iter = &psi_system;
777 	return &psi_system;
778 }
779 
780 static void psi_flags_change(struct task_struct *task, int clear, int set)
781 {
782 	if (((task->psi_flags & set) ||
783 	     (task->psi_flags & clear) != clear) &&
784 	    !psi_bug) {
785 		printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
786 				task->pid, task->comm, task_cpu(task),
787 				task->psi_flags, clear, set);
788 		psi_bug = 1;
789 	}
790 
791 	task->psi_flags &= ~clear;
792 	task->psi_flags |= set;
793 }
794 
795 void psi_task_change(struct task_struct *task, int clear, int set)
796 {
797 	int cpu = task_cpu(task);
798 	struct psi_group *group;
799 	bool wake_clock = true;
800 	void *iter = NULL;
801 	u64 now;
802 
803 	if (!task->pid)
804 		return;
805 
806 	psi_flags_change(task, clear, set);
807 
808 	now = cpu_clock(cpu);
809 	/*
810 	 * Periodic aggregation shuts off if there is a period of no
811 	 * task changes, so we wake it back up if necessary. However,
812 	 * don't do this if the task change is the aggregation worker
813 	 * itself going to sleep, or we'll ping-pong forever.
814 	 */
815 	if (unlikely((clear & TSK_RUNNING) &&
816 		     (task->flags & PF_WQ_WORKER) &&
817 		     wq_worker_last_func(task) == psi_avgs_work))
818 		wake_clock = false;
819 
820 	while ((group = iterate_groups(task, &iter)))
821 		psi_group_change(group, cpu, clear, set, now, wake_clock);
822 }
823 
824 void psi_task_switch(struct task_struct *prev, struct task_struct *next,
825 		     bool sleep)
826 {
827 	struct psi_group *group, *common = NULL;
828 	int cpu = task_cpu(prev);
829 	void *iter;
830 	u64 now = cpu_clock(cpu);
831 
832 	if (next->pid) {
833 		bool identical_state;
834 
835 		psi_flags_change(next, 0, TSK_ONCPU);
836 		/*
837 		 * When switching between tasks that have an identical
838 		 * runtime state, the cgroup that contains both tasks
839 		 * we reach the first common ancestor. Iterate @next's
840 		 * ancestors only until we encounter @prev's ONCPU.
841 		 */
842 		identical_state = prev->psi_flags == next->psi_flags;
843 		iter = NULL;
844 		while ((group = iterate_groups(next, &iter))) {
845 			if (identical_state &&
846 			    per_cpu_ptr(group->pcpu, cpu)->tasks[NR_ONCPU]) {
847 				common = group;
848 				break;
849 			}
850 
851 			psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
852 		}
853 	}
854 
855 	if (prev->pid) {
856 		int clear = TSK_ONCPU, set = 0;
857 
858 		/*
859 		 * When we're going to sleep, psi_dequeue() lets us
860 		 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
861 		 * TSK_IOWAIT here, where we can combine it with
862 		 * TSK_ONCPU and save walking common ancestors twice.
863 		 */
864 		if (sleep) {
865 			clear |= TSK_RUNNING;
866 			if (prev->in_memstall)
867 				clear |= TSK_MEMSTALL_RUNNING;
868 			if (prev->in_iowait)
869 				set |= TSK_IOWAIT;
870 		}
871 
872 		psi_flags_change(prev, clear, set);
873 
874 		iter = NULL;
875 		while ((group = iterate_groups(prev, &iter)) && group != common)
876 			psi_group_change(group, cpu, clear, set, now, true);
877 
878 		/*
879 		 * TSK_ONCPU is handled up to the common ancestor. If we're tasked
880 		 * with dequeuing too, finish that for the rest of the hierarchy.
881 		 */
882 		if (sleep) {
883 			clear &= ~TSK_ONCPU;
884 			for (; group; group = iterate_groups(prev, &iter))
885 				psi_group_change(group, cpu, clear, set, now, true);
886 		}
887 	}
888 }
889 
890 /**
891  * psi_memstall_enter - mark the beginning of a memory stall section
892  * @flags: flags to handle nested sections
893  *
894  * Marks the calling task as being stalled due to a lack of memory,
895  * such as waiting for a refault or performing reclaim.
896  */
897 void psi_memstall_enter(unsigned long *flags)
898 {
899 	struct rq_flags rf;
900 	struct rq *rq;
901 
902 	if (static_branch_likely(&psi_disabled))
903 		return;
904 
905 	*flags = current->in_memstall;
906 	if (*flags)
907 		return;
908 	/*
909 	 * in_memstall setting & accounting needs to be atomic wrt
910 	 * changes to the task's scheduling state, otherwise we can
911 	 * race with CPU migration.
912 	 */
913 	rq = this_rq_lock_irq(&rf);
914 
915 	current->in_memstall = 1;
916 	psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
917 
918 	rq_unlock_irq(rq, &rf);
919 }
920 
921 /**
922  * psi_memstall_leave - mark the end of an memory stall section
923  * @flags: flags to handle nested memdelay sections
924  *
925  * Marks the calling task as no longer stalled due to lack of memory.
926  */
927 void psi_memstall_leave(unsigned long *flags)
928 {
929 	struct rq_flags rf;
930 	struct rq *rq;
931 
932 	if (static_branch_likely(&psi_disabled))
933 		return;
934 
935 	if (*flags)
936 		return;
937 	/*
938 	 * in_memstall clearing & accounting needs to be atomic wrt
939 	 * changes to the task's scheduling state, otherwise we could
940 	 * race with CPU migration.
941 	 */
942 	rq = this_rq_lock_irq(&rf);
943 
944 	current->in_memstall = 0;
945 	psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0);
946 
947 	rq_unlock_irq(rq, &rf);
948 }
949 
950 #ifdef CONFIG_CGROUPS
951 int psi_cgroup_alloc(struct cgroup *cgroup)
952 {
953 	if (static_branch_likely(&psi_disabled))
954 		return 0;
955 
956 	cgroup->psi = kzalloc(sizeof(struct psi_group), GFP_KERNEL);
957 	if (!cgroup->psi)
958 		return -ENOMEM;
959 
960 	cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
961 	if (!cgroup->psi->pcpu) {
962 		kfree(cgroup->psi);
963 		return -ENOMEM;
964 	}
965 	group_init(cgroup->psi);
966 	return 0;
967 }
968 
969 void psi_cgroup_free(struct cgroup *cgroup)
970 {
971 	if (static_branch_likely(&psi_disabled))
972 		return;
973 
974 	cancel_delayed_work_sync(&cgroup->psi->avgs_work);
975 	free_percpu(cgroup->psi->pcpu);
976 	/* All triggers must be removed by now */
977 	WARN_ONCE(cgroup->psi->poll_states, "psi: trigger leak\n");
978 	kfree(cgroup->psi);
979 }
980 
981 /**
982  * cgroup_move_task - move task to a different cgroup
983  * @task: the task
984  * @to: the target css_set
985  *
986  * Move task to a new cgroup and safely migrate its associated stall
987  * state between the different groups.
988  *
989  * This function acquires the task's rq lock to lock out concurrent
990  * changes to the task's scheduling state and - in case the task is
991  * running - concurrent changes to its stall state.
992  */
993 void cgroup_move_task(struct task_struct *task, struct css_set *to)
994 {
995 	unsigned int task_flags;
996 	struct rq_flags rf;
997 	struct rq *rq;
998 
999 	if (static_branch_likely(&psi_disabled)) {
1000 		/*
1001 		 * Lame to do this here, but the scheduler cannot be locked
1002 		 * from the outside, so we move cgroups from inside sched/.
1003 		 */
1004 		rcu_assign_pointer(task->cgroups, to);
1005 		return;
1006 	}
1007 
1008 	rq = task_rq_lock(task, &rf);
1009 
1010 	/*
1011 	 * We may race with schedule() dropping the rq lock between
1012 	 * deactivating prev and switching to next. Because the psi
1013 	 * updates from the deactivation are deferred to the switch
1014 	 * callback to save cgroup tree updates, the task's scheduling
1015 	 * state here is not coherent with its psi state:
1016 	 *
1017 	 * schedule()                   cgroup_move_task()
1018 	 *   rq_lock()
1019 	 *   deactivate_task()
1020 	 *     p->on_rq = 0
1021 	 *     psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1022 	 *   pick_next_task()
1023 	 *     rq_unlock()
1024 	 *                                rq_lock()
1025 	 *                                psi_task_change() // old cgroup
1026 	 *                                task->cgroups = to
1027 	 *                                psi_task_change() // new cgroup
1028 	 *                                rq_unlock()
1029 	 *     rq_lock()
1030 	 *   psi_sched_switch() // does deferred updates in new cgroup
1031 	 *
1032 	 * Don't rely on the scheduling state. Use psi_flags instead.
1033 	 */
1034 	task_flags = task->psi_flags;
1035 
1036 	if (task_flags)
1037 		psi_task_change(task, task_flags, 0);
1038 
1039 	/* See comment above */
1040 	rcu_assign_pointer(task->cgroups, to);
1041 
1042 	if (task_flags)
1043 		psi_task_change(task, 0, task_flags);
1044 
1045 	task_rq_unlock(rq, task, &rf);
1046 }
1047 #endif /* CONFIG_CGROUPS */
1048 
1049 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1050 {
1051 	int full;
1052 	u64 now;
1053 
1054 	if (static_branch_likely(&psi_disabled))
1055 		return -EOPNOTSUPP;
1056 
1057 	/* Update averages before reporting them */
1058 	mutex_lock(&group->avgs_lock);
1059 	now = sched_clock();
1060 	collect_percpu_times(group, PSI_AVGS, NULL);
1061 	if (now >= group->avg_next_update)
1062 		group->avg_next_update = update_averages(group, now);
1063 	mutex_unlock(&group->avgs_lock);
1064 
1065 	for (full = 0; full < 2; full++) {
1066 		unsigned long avg[3] = { 0, };
1067 		u64 total = 0;
1068 		int w;
1069 
1070 		/* CPU FULL is undefined at the system level */
1071 		if (!(group == &psi_system && res == PSI_CPU && full)) {
1072 			for (w = 0; w < 3; w++)
1073 				avg[w] = group->avg[res * 2 + full][w];
1074 			total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1075 					NSEC_PER_USEC);
1076 		}
1077 
1078 		seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1079 			   full ? "full" : "some",
1080 			   LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1081 			   LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1082 			   LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1083 			   total);
1084 	}
1085 
1086 	return 0;
1087 }
1088 
1089 struct psi_trigger *psi_trigger_create(struct psi_group *group,
1090 			char *buf, enum psi_res res)
1091 {
1092 	struct psi_trigger *t;
1093 	enum psi_states state;
1094 	u32 threshold_us;
1095 	u32 window_us;
1096 
1097 	if (static_branch_likely(&psi_disabled))
1098 		return ERR_PTR(-EOPNOTSUPP);
1099 
1100 	if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1101 		state = PSI_IO_SOME + res * 2;
1102 	else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1103 		state = PSI_IO_FULL + res * 2;
1104 	else
1105 		return ERR_PTR(-EINVAL);
1106 
1107 	if (state >= PSI_NONIDLE)
1108 		return ERR_PTR(-EINVAL);
1109 
1110 	if (window_us < WINDOW_MIN_US ||
1111 		window_us > WINDOW_MAX_US)
1112 		return ERR_PTR(-EINVAL);
1113 
1114 	/* Check threshold */
1115 	if (threshold_us == 0 || threshold_us > window_us)
1116 		return ERR_PTR(-EINVAL);
1117 
1118 	t = kmalloc(sizeof(*t), GFP_KERNEL);
1119 	if (!t)
1120 		return ERR_PTR(-ENOMEM);
1121 
1122 	t->group = group;
1123 	t->state = state;
1124 	t->threshold = threshold_us * NSEC_PER_USEC;
1125 	t->win.size = window_us * NSEC_PER_USEC;
1126 	window_reset(&t->win, sched_clock(),
1127 			group->total[PSI_POLL][t->state], 0);
1128 
1129 	t->event = 0;
1130 	t->last_event_time = 0;
1131 	init_waitqueue_head(&t->event_wait);
1132 	t->pending_event = false;
1133 
1134 	mutex_lock(&group->trigger_lock);
1135 
1136 	if (!rcu_access_pointer(group->poll_task)) {
1137 		struct task_struct *task;
1138 
1139 		task = kthread_create(psi_poll_worker, group, "psimon");
1140 		if (IS_ERR(task)) {
1141 			kfree(t);
1142 			mutex_unlock(&group->trigger_lock);
1143 			return ERR_CAST(task);
1144 		}
1145 		atomic_set(&group->poll_wakeup, 0);
1146 		wake_up_process(task);
1147 		rcu_assign_pointer(group->poll_task, task);
1148 	}
1149 
1150 	list_add(&t->node, &group->triggers);
1151 	group->poll_min_period = min(group->poll_min_period,
1152 		div_u64(t->win.size, UPDATES_PER_WINDOW));
1153 	group->nr_triggers[t->state]++;
1154 	group->poll_states |= (1 << t->state);
1155 
1156 	mutex_unlock(&group->trigger_lock);
1157 
1158 	return t;
1159 }
1160 
1161 void psi_trigger_destroy(struct psi_trigger *t)
1162 {
1163 	struct psi_group *group;
1164 	struct task_struct *task_to_destroy = NULL;
1165 
1166 	/*
1167 	 * We do not check psi_disabled since it might have been disabled after
1168 	 * the trigger got created.
1169 	 */
1170 	if (!t)
1171 		return;
1172 
1173 	group = t->group;
1174 	/*
1175 	 * Wakeup waiters to stop polling. Can happen if cgroup is deleted
1176 	 * from under a polling process.
1177 	 */
1178 	wake_up_interruptible(&t->event_wait);
1179 
1180 	mutex_lock(&group->trigger_lock);
1181 
1182 	if (!list_empty(&t->node)) {
1183 		struct psi_trigger *tmp;
1184 		u64 period = ULLONG_MAX;
1185 
1186 		list_del(&t->node);
1187 		group->nr_triggers[t->state]--;
1188 		if (!group->nr_triggers[t->state])
1189 			group->poll_states &= ~(1 << t->state);
1190 		/* reset min update period for the remaining triggers */
1191 		list_for_each_entry(tmp, &group->triggers, node)
1192 			period = min(period, div_u64(tmp->win.size,
1193 					UPDATES_PER_WINDOW));
1194 		group->poll_min_period = period;
1195 		/* Destroy poll_task when the last trigger is destroyed */
1196 		if (group->poll_states == 0) {
1197 			group->polling_until = 0;
1198 			task_to_destroy = rcu_dereference_protected(
1199 					group->poll_task,
1200 					lockdep_is_held(&group->trigger_lock));
1201 			rcu_assign_pointer(group->poll_task, NULL);
1202 			del_timer(&group->poll_timer);
1203 		}
1204 	}
1205 
1206 	mutex_unlock(&group->trigger_lock);
1207 
1208 	/*
1209 	 * Wait for psi_schedule_poll_work RCU to complete its read-side
1210 	 * critical section before destroying the trigger and optionally the
1211 	 * poll_task.
1212 	 */
1213 	synchronize_rcu();
1214 	/*
1215 	 * Stop kthread 'psimon' after releasing trigger_lock to prevent a
1216 	 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1217 	 */
1218 	if (task_to_destroy) {
1219 		/*
1220 		 * After the RCU grace period has expired, the worker
1221 		 * can no longer be found through group->poll_task.
1222 		 */
1223 		kthread_stop(task_to_destroy);
1224 	}
1225 	kfree(t);
1226 }
1227 
1228 __poll_t psi_trigger_poll(void **trigger_ptr,
1229 				struct file *file, poll_table *wait)
1230 {
1231 	__poll_t ret = DEFAULT_POLLMASK;
1232 	struct psi_trigger *t;
1233 
1234 	if (static_branch_likely(&psi_disabled))
1235 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1236 
1237 	t = smp_load_acquire(trigger_ptr);
1238 	if (!t)
1239 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1240 
1241 	poll_wait(file, &t->event_wait, wait);
1242 
1243 	if (cmpxchg(&t->event, 1, 0) == 1)
1244 		ret |= EPOLLPRI;
1245 
1246 	return ret;
1247 }
1248 
1249 #ifdef CONFIG_PROC_FS
1250 static int psi_io_show(struct seq_file *m, void *v)
1251 {
1252 	return psi_show(m, &psi_system, PSI_IO);
1253 }
1254 
1255 static int psi_memory_show(struct seq_file *m, void *v)
1256 {
1257 	return psi_show(m, &psi_system, PSI_MEM);
1258 }
1259 
1260 static int psi_cpu_show(struct seq_file *m, void *v)
1261 {
1262 	return psi_show(m, &psi_system, PSI_CPU);
1263 }
1264 
1265 static int psi_open(struct file *file, int (*psi_show)(struct seq_file *, void *))
1266 {
1267 	if (file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE))
1268 		return -EPERM;
1269 
1270 	return single_open(file, psi_show, NULL);
1271 }
1272 
1273 static int psi_io_open(struct inode *inode, struct file *file)
1274 {
1275 	return psi_open(file, psi_io_show);
1276 }
1277 
1278 static int psi_memory_open(struct inode *inode, struct file *file)
1279 {
1280 	return psi_open(file, psi_memory_show);
1281 }
1282 
1283 static int psi_cpu_open(struct inode *inode, struct file *file)
1284 {
1285 	return psi_open(file, psi_cpu_show);
1286 }
1287 
1288 static ssize_t psi_write(struct file *file, const char __user *user_buf,
1289 			 size_t nbytes, enum psi_res res)
1290 {
1291 	char buf[32];
1292 	size_t buf_size;
1293 	struct seq_file *seq;
1294 	struct psi_trigger *new;
1295 
1296 	if (static_branch_likely(&psi_disabled))
1297 		return -EOPNOTSUPP;
1298 
1299 	if (!nbytes)
1300 		return -EINVAL;
1301 
1302 	buf_size = min(nbytes, sizeof(buf));
1303 	if (copy_from_user(buf, user_buf, buf_size))
1304 		return -EFAULT;
1305 
1306 	buf[buf_size - 1] = '\0';
1307 
1308 	seq = file->private_data;
1309 
1310 	/* Take seq->lock to protect seq->private from concurrent writes */
1311 	mutex_lock(&seq->lock);
1312 
1313 	/* Allow only one trigger per file descriptor */
1314 	if (seq->private) {
1315 		mutex_unlock(&seq->lock);
1316 		return -EBUSY;
1317 	}
1318 
1319 	new = psi_trigger_create(&psi_system, buf, res);
1320 	if (IS_ERR(new)) {
1321 		mutex_unlock(&seq->lock);
1322 		return PTR_ERR(new);
1323 	}
1324 
1325 	smp_store_release(&seq->private, new);
1326 	mutex_unlock(&seq->lock);
1327 
1328 	return nbytes;
1329 }
1330 
1331 static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1332 			    size_t nbytes, loff_t *ppos)
1333 {
1334 	return psi_write(file, user_buf, nbytes, PSI_IO);
1335 }
1336 
1337 static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1338 				size_t nbytes, loff_t *ppos)
1339 {
1340 	return psi_write(file, user_buf, nbytes, PSI_MEM);
1341 }
1342 
1343 static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1344 			     size_t nbytes, loff_t *ppos)
1345 {
1346 	return psi_write(file, user_buf, nbytes, PSI_CPU);
1347 }
1348 
1349 static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1350 {
1351 	struct seq_file *seq = file->private_data;
1352 
1353 	return psi_trigger_poll(&seq->private, file, wait);
1354 }
1355 
1356 static int psi_fop_release(struct inode *inode, struct file *file)
1357 {
1358 	struct seq_file *seq = file->private_data;
1359 
1360 	psi_trigger_destroy(seq->private);
1361 	return single_release(inode, file);
1362 }
1363 
1364 static const struct proc_ops psi_io_proc_ops = {
1365 	.proc_open	= psi_io_open,
1366 	.proc_read	= seq_read,
1367 	.proc_lseek	= seq_lseek,
1368 	.proc_write	= psi_io_write,
1369 	.proc_poll	= psi_fop_poll,
1370 	.proc_release	= psi_fop_release,
1371 };
1372 
1373 static const struct proc_ops psi_memory_proc_ops = {
1374 	.proc_open	= psi_memory_open,
1375 	.proc_read	= seq_read,
1376 	.proc_lseek	= seq_lseek,
1377 	.proc_write	= psi_memory_write,
1378 	.proc_poll	= psi_fop_poll,
1379 	.proc_release	= psi_fop_release,
1380 };
1381 
1382 static const struct proc_ops psi_cpu_proc_ops = {
1383 	.proc_open	= psi_cpu_open,
1384 	.proc_read	= seq_read,
1385 	.proc_lseek	= seq_lseek,
1386 	.proc_write	= psi_cpu_write,
1387 	.proc_poll	= psi_fop_poll,
1388 	.proc_release	= psi_fop_release,
1389 };
1390 
1391 static int __init psi_proc_init(void)
1392 {
1393 	if (psi_enable) {
1394 		proc_mkdir("pressure", NULL);
1395 		proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops);
1396 		proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops);
1397 		proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops);
1398 	}
1399 	return 0;
1400 }
1401 module_init(psi_proc_init);
1402 
1403 #endif /* CONFIG_PROC_FS */
1404