xref: /openbmc/linux/kernel/sched/psi.c (revision 227a0655)
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 #include "../workqueue_internal.h"
141 #include <linux/sched/loadavg.h>
142 #include <linux/seq_file.h>
143 #include <linux/proc_fs.h>
144 #include <linux/seqlock.h>
145 #include <linux/uaccess.h>
146 #include <linux/cgroup.h>
147 #include <linux/module.h>
148 #include <linux/sched.h>
149 #include <linux/ctype.h>
150 #include <linux/file.h>
151 #include <linux/poll.h>
152 #include <linux/psi.h>
153 #include "sched.h"
154 
155 static int psi_bug __read_mostly;
156 
157 DEFINE_STATIC_KEY_FALSE(psi_disabled);
158 DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled);
159 
160 #ifdef CONFIG_PSI_DEFAULT_DISABLED
161 static bool psi_enable;
162 #else
163 static bool psi_enable = true;
164 #endif
165 static int __init setup_psi(char *str)
166 {
167 	return kstrtobool(str, &psi_enable) == 0;
168 }
169 __setup("psi=", setup_psi);
170 
171 /* Running averages - we need to be higher-res than loadavg */
172 #define PSI_FREQ	(2*HZ+1)	/* 2 sec intervals */
173 #define EXP_10s		1677		/* 1/exp(2s/10s) as fixed-point */
174 #define EXP_60s		1981		/* 1/exp(2s/60s) */
175 #define EXP_300s	2034		/* 1/exp(2s/300s) */
176 
177 /* PSI trigger definitions */
178 #define WINDOW_MIN_US 500000	/* Min window size is 500ms */
179 #define WINDOW_MAX_US 10000000	/* Max window size is 10s */
180 #define UPDATES_PER_WINDOW 10	/* 10 updates per window */
181 
182 /* Sampling frequency in nanoseconds */
183 static u64 psi_period __read_mostly;
184 
185 /* System-level pressure and stall tracking */
186 static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
187 struct psi_group psi_system = {
188 	.pcpu = &system_group_pcpu,
189 };
190 
191 static void psi_avgs_work(struct work_struct *work);
192 
193 static void poll_timer_fn(struct timer_list *t);
194 
195 static void group_init(struct psi_group *group)
196 {
197 	int cpu;
198 
199 	for_each_possible_cpu(cpu)
200 		seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);
201 	group->avg_last_update = sched_clock();
202 	group->avg_next_update = group->avg_last_update + psi_period;
203 	INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
204 	mutex_init(&group->avgs_lock);
205 	/* Init trigger-related members */
206 	mutex_init(&group->trigger_lock);
207 	INIT_LIST_HEAD(&group->triggers);
208 	memset(group->nr_triggers, 0, sizeof(group->nr_triggers));
209 	group->poll_states = 0;
210 	group->poll_min_period = U32_MAX;
211 	memset(group->polling_total, 0, sizeof(group->polling_total));
212 	group->polling_next_update = ULLONG_MAX;
213 	group->polling_until = 0;
214 	init_waitqueue_head(&group->poll_wait);
215 	timer_setup(&group->poll_timer, poll_timer_fn, 0);
216 	rcu_assign_pointer(group->poll_task, NULL);
217 }
218 
219 void __init psi_init(void)
220 {
221 	if (!psi_enable) {
222 		static_branch_enable(&psi_disabled);
223 		return;
224 	}
225 
226 	if (!cgroup_psi_enabled())
227 		static_branch_disable(&psi_cgroups_enabled);
228 
229 	psi_period = jiffies_to_nsecs(PSI_FREQ);
230 	group_init(&psi_system);
231 }
232 
233 static bool test_state(unsigned int *tasks, enum psi_states state)
234 {
235 	switch (state) {
236 	case PSI_IO_SOME:
237 		return unlikely(tasks[NR_IOWAIT]);
238 	case PSI_IO_FULL:
239 		return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]);
240 	case PSI_MEM_SOME:
241 		return unlikely(tasks[NR_MEMSTALL]);
242 	case PSI_MEM_FULL:
243 		return unlikely(tasks[NR_MEMSTALL] &&
244 			tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]);
245 	case PSI_CPU_SOME:
246 		return unlikely(tasks[NR_RUNNING] > tasks[NR_ONCPU]);
247 	case PSI_CPU_FULL:
248 		return unlikely(tasks[NR_RUNNING] && !tasks[NR_ONCPU]);
249 	case PSI_NONIDLE:
250 		return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
251 			tasks[NR_RUNNING];
252 	default:
253 		return false;
254 	}
255 }
256 
257 static void get_recent_times(struct psi_group *group, int cpu,
258 			     enum psi_aggregators aggregator, u32 *times,
259 			     u32 *pchanged_states)
260 {
261 	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
262 	u64 now, state_start;
263 	enum psi_states s;
264 	unsigned int seq;
265 	u32 state_mask;
266 
267 	*pchanged_states = 0;
268 
269 	/* Snapshot a coherent view of the CPU state */
270 	do {
271 		seq = read_seqcount_begin(&groupc->seq);
272 		now = cpu_clock(cpu);
273 		memcpy(times, groupc->times, sizeof(groupc->times));
274 		state_mask = groupc->state_mask;
275 		state_start = groupc->state_start;
276 	} while (read_seqcount_retry(&groupc->seq, seq));
277 
278 	/* Calculate state time deltas against the previous snapshot */
279 	for (s = 0; s < NR_PSI_STATES; s++) {
280 		u32 delta;
281 		/*
282 		 * In addition to already concluded states, we also
283 		 * incorporate currently active states on the CPU,
284 		 * since states may last for many sampling periods.
285 		 *
286 		 * This way we keep our delta sampling buckets small
287 		 * (u32) and our reported pressure close to what's
288 		 * actually happening.
289 		 */
290 		if (state_mask & (1 << s))
291 			times[s] += now - state_start;
292 
293 		delta = times[s] - groupc->times_prev[aggregator][s];
294 		groupc->times_prev[aggregator][s] = times[s];
295 
296 		times[s] = delta;
297 		if (delta)
298 			*pchanged_states |= (1 << s);
299 	}
300 }
301 
302 static void calc_avgs(unsigned long avg[3], int missed_periods,
303 		      u64 time, u64 period)
304 {
305 	unsigned long pct;
306 
307 	/* Fill in zeroes for periods of no activity */
308 	if (missed_periods) {
309 		avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
310 		avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
311 		avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
312 	}
313 
314 	/* Sample the most recent active period */
315 	pct = div_u64(time * 100, period);
316 	pct *= FIXED_1;
317 	avg[0] = calc_load(avg[0], EXP_10s, pct);
318 	avg[1] = calc_load(avg[1], EXP_60s, pct);
319 	avg[2] = calc_load(avg[2], EXP_300s, pct);
320 }
321 
322 static void collect_percpu_times(struct psi_group *group,
323 				 enum psi_aggregators aggregator,
324 				 u32 *pchanged_states)
325 {
326 	u64 deltas[NR_PSI_STATES - 1] = { 0, };
327 	unsigned long nonidle_total = 0;
328 	u32 changed_states = 0;
329 	int cpu;
330 	int s;
331 
332 	/*
333 	 * Collect the per-cpu time buckets and average them into a
334 	 * single time sample that is normalized to wallclock time.
335 	 *
336 	 * For averaging, each CPU is weighted by its non-idle time in
337 	 * the sampling period. This eliminates artifacts from uneven
338 	 * loading, or even entirely idle CPUs.
339 	 */
340 	for_each_possible_cpu(cpu) {
341 		u32 times[NR_PSI_STATES];
342 		u32 nonidle;
343 		u32 cpu_changed_states;
344 
345 		get_recent_times(group, cpu, aggregator, times,
346 				&cpu_changed_states);
347 		changed_states |= cpu_changed_states;
348 
349 		nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
350 		nonidle_total += nonidle;
351 
352 		for (s = 0; s < PSI_NONIDLE; s++)
353 			deltas[s] += (u64)times[s] * nonidle;
354 	}
355 
356 	/*
357 	 * Integrate the sample into the running statistics that are
358 	 * reported to userspace: the cumulative stall times and the
359 	 * decaying averages.
360 	 *
361 	 * Pressure percentages are sampled at PSI_FREQ. We might be
362 	 * called more often when the user polls more frequently than
363 	 * that; we might be called less often when there is no task
364 	 * activity, thus no data, and clock ticks are sporadic. The
365 	 * below handles both.
366 	 */
367 
368 	/* total= */
369 	for (s = 0; s < NR_PSI_STATES - 1; s++)
370 		group->total[aggregator][s] +=
371 				div_u64(deltas[s], max(nonidle_total, 1UL));
372 
373 	if (pchanged_states)
374 		*pchanged_states = changed_states;
375 }
376 
377 static u64 update_averages(struct psi_group *group, u64 now)
378 {
379 	unsigned long missed_periods = 0;
380 	u64 expires, period;
381 	u64 avg_next_update;
382 	int s;
383 
384 	/* avgX= */
385 	expires = group->avg_next_update;
386 	if (now - expires >= psi_period)
387 		missed_periods = div_u64(now - expires, psi_period);
388 
389 	/*
390 	 * The periodic clock tick can get delayed for various
391 	 * reasons, especially on loaded systems. To avoid clock
392 	 * drift, we schedule the clock in fixed psi_period intervals.
393 	 * But the deltas we sample out of the per-cpu buckets above
394 	 * are based on the actual time elapsing between clock ticks.
395 	 */
396 	avg_next_update = expires + ((1 + missed_periods) * psi_period);
397 	period = now - (group->avg_last_update + (missed_periods * psi_period));
398 	group->avg_last_update = now;
399 
400 	for (s = 0; s < NR_PSI_STATES - 1; s++) {
401 		u32 sample;
402 
403 		sample = group->total[PSI_AVGS][s] - group->avg_total[s];
404 		/*
405 		 * Due to the lockless sampling of the time buckets,
406 		 * recorded time deltas can slip into the next period,
407 		 * which under full pressure can result in samples in
408 		 * excess of the period length.
409 		 *
410 		 * We don't want to report non-sensical pressures in
411 		 * excess of 100%, nor do we want to drop such events
412 		 * on the floor. Instead we punt any overage into the
413 		 * future until pressure subsides. By doing this we
414 		 * don't underreport the occurring pressure curve, we
415 		 * just report it delayed by one period length.
416 		 *
417 		 * The error isn't cumulative. As soon as another
418 		 * delta slips from a period P to P+1, by definition
419 		 * it frees up its time T in P.
420 		 */
421 		if (sample > period)
422 			sample = period;
423 		group->avg_total[s] += sample;
424 		calc_avgs(group->avg[s], missed_periods, sample, period);
425 	}
426 
427 	return avg_next_update;
428 }
429 
430 static void psi_avgs_work(struct work_struct *work)
431 {
432 	struct delayed_work *dwork;
433 	struct psi_group *group;
434 	u32 changed_states;
435 	bool nonidle;
436 	u64 now;
437 
438 	dwork = to_delayed_work(work);
439 	group = container_of(dwork, struct psi_group, avgs_work);
440 
441 	mutex_lock(&group->avgs_lock);
442 
443 	now = sched_clock();
444 
445 	collect_percpu_times(group, PSI_AVGS, &changed_states);
446 	nonidle = changed_states & (1 << PSI_NONIDLE);
447 	/*
448 	 * If there is task activity, periodically fold the per-cpu
449 	 * times and feed samples into the running averages. If things
450 	 * are idle and there is no data to process, stop the clock.
451 	 * Once restarted, we'll catch up the running averages in one
452 	 * go - see calc_avgs() and missed_periods.
453 	 */
454 	if (now >= group->avg_next_update)
455 		group->avg_next_update = update_averages(group, now);
456 
457 	if (nonidle) {
458 		schedule_delayed_work(dwork, nsecs_to_jiffies(
459 				group->avg_next_update - now) + 1);
460 	}
461 
462 	mutex_unlock(&group->avgs_lock);
463 }
464 
465 /* Trigger tracking window manipulations */
466 static void window_reset(struct psi_window *win, u64 now, u64 value,
467 			 u64 prev_growth)
468 {
469 	win->start_time = now;
470 	win->start_value = value;
471 	win->prev_growth = prev_growth;
472 }
473 
474 /*
475  * PSI growth tracking window update and growth calculation routine.
476  *
477  * This approximates a sliding tracking window by interpolating
478  * partially elapsed windows using historical growth data from the
479  * previous intervals. This minimizes memory requirements (by not storing
480  * all the intermediate values in the previous window) and simplifies
481  * the calculations. It works well because PSI signal changes only in
482  * positive direction and over relatively small window sizes the growth
483  * is close to linear.
484  */
485 static u64 window_update(struct psi_window *win, u64 now, u64 value)
486 {
487 	u64 elapsed;
488 	u64 growth;
489 
490 	elapsed = now - win->start_time;
491 	growth = value - win->start_value;
492 	/*
493 	 * After each tracking window passes win->start_value and
494 	 * win->start_time get reset and win->prev_growth stores
495 	 * the average per-window growth of the previous window.
496 	 * win->prev_growth is then used to interpolate additional
497 	 * growth from the previous window assuming it was linear.
498 	 */
499 	if (elapsed > win->size)
500 		window_reset(win, now, value, growth);
501 	else {
502 		u32 remaining;
503 
504 		remaining = win->size - elapsed;
505 		growth += div64_u64(win->prev_growth * remaining, win->size);
506 	}
507 
508 	return growth;
509 }
510 
511 static void init_triggers(struct psi_group *group, u64 now)
512 {
513 	struct psi_trigger *t;
514 
515 	list_for_each_entry(t, &group->triggers, node)
516 		window_reset(&t->win, now,
517 				group->total[PSI_POLL][t->state], 0);
518 	memcpy(group->polling_total, group->total[PSI_POLL],
519 		   sizeof(group->polling_total));
520 	group->polling_next_update = now + group->poll_min_period;
521 }
522 
523 static u64 update_triggers(struct psi_group *group, u64 now)
524 {
525 	struct psi_trigger *t;
526 	bool new_stall = false;
527 	u64 *total = group->total[PSI_POLL];
528 
529 	/*
530 	 * On subsequent updates, calculate growth deltas and let
531 	 * watchers know when their specified thresholds are exceeded.
532 	 */
533 	list_for_each_entry(t, &group->triggers, node) {
534 		u64 growth;
535 
536 		/* Check for stall activity */
537 		if (group->polling_total[t->state] == total[t->state])
538 			continue;
539 
540 		/*
541 		 * Multiple triggers might be looking at the same state,
542 		 * remember to update group->polling_total[] once we've
543 		 * been through all of them. Also remember to extend the
544 		 * polling time if we see new stall activity.
545 		 */
546 		new_stall = true;
547 
548 		/* Calculate growth since last update */
549 		growth = window_update(&t->win, now, total[t->state]);
550 		if (growth < t->threshold)
551 			continue;
552 
553 		/* Limit event signaling to once per window */
554 		if (now < t->last_event_time + t->win.size)
555 			continue;
556 
557 		/* Generate an event */
558 		if (cmpxchg(&t->event, 0, 1) == 0)
559 			wake_up_interruptible(&t->event_wait);
560 		t->last_event_time = now;
561 	}
562 
563 	if (new_stall)
564 		memcpy(group->polling_total, total,
565 				sizeof(group->polling_total));
566 
567 	return now + group->poll_min_period;
568 }
569 
570 /* Schedule polling if it's not already scheduled. */
571 static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay)
572 {
573 	struct task_struct *task;
574 
575 	/*
576 	 * Do not reschedule if already scheduled.
577 	 * Possible race with a timer scheduled after this check but before
578 	 * mod_timer below can be tolerated because group->polling_next_update
579 	 * will keep updates on schedule.
580 	 */
581 	if (timer_pending(&group->poll_timer))
582 		return;
583 
584 	rcu_read_lock();
585 
586 	task = rcu_dereference(group->poll_task);
587 	/*
588 	 * kworker might be NULL in case psi_trigger_destroy races with
589 	 * psi_task_change (hotpath) which can't use locks
590 	 */
591 	if (likely(task))
592 		mod_timer(&group->poll_timer, jiffies + delay);
593 
594 	rcu_read_unlock();
595 }
596 
597 static void psi_poll_work(struct psi_group *group)
598 {
599 	u32 changed_states;
600 	u64 now;
601 
602 	mutex_lock(&group->trigger_lock);
603 
604 	now = sched_clock();
605 
606 	collect_percpu_times(group, PSI_POLL, &changed_states);
607 
608 	if (changed_states & group->poll_states) {
609 		/* Initialize trigger windows when entering polling mode */
610 		if (now > group->polling_until)
611 			init_triggers(group, now);
612 
613 		/*
614 		 * Keep the monitor active for at least the duration of the
615 		 * minimum tracking window as long as monitor states are
616 		 * changing.
617 		 */
618 		group->polling_until = now +
619 			group->poll_min_period * UPDATES_PER_WINDOW;
620 	}
621 
622 	if (now > group->polling_until) {
623 		group->polling_next_update = ULLONG_MAX;
624 		goto out;
625 	}
626 
627 	if (now >= group->polling_next_update)
628 		group->polling_next_update = update_triggers(group, now);
629 
630 	psi_schedule_poll_work(group,
631 		nsecs_to_jiffies(group->polling_next_update - now) + 1);
632 
633 out:
634 	mutex_unlock(&group->trigger_lock);
635 }
636 
637 static int psi_poll_worker(void *data)
638 {
639 	struct psi_group *group = (struct psi_group *)data;
640 
641 	sched_set_fifo_low(current);
642 
643 	while (true) {
644 		wait_event_interruptible(group->poll_wait,
645 				atomic_cmpxchg(&group->poll_wakeup, 1, 0) ||
646 				kthread_should_stop());
647 		if (kthread_should_stop())
648 			break;
649 
650 		psi_poll_work(group);
651 	}
652 	return 0;
653 }
654 
655 static void poll_timer_fn(struct timer_list *t)
656 {
657 	struct psi_group *group = from_timer(group, t, poll_timer);
658 
659 	atomic_set(&group->poll_wakeup, 1);
660 	wake_up_interruptible(&group->poll_wait);
661 }
662 
663 static void record_times(struct psi_group_cpu *groupc, u64 now)
664 {
665 	u32 delta;
666 
667 	delta = now - groupc->state_start;
668 	groupc->state_start = now;
669 
670 	if (groupc->state_mask & (1 << PSI_IO_SOME)) {
671 		groupc->times[PSI_IO_SOME] += delta;
672 		if (groupc->state_mask & (1 << PSI_IO_FULL))
673 			groupc->times[PSI_IO_FULL] += delta;
674 	}
675 
676 	if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
677 		groupc->times[PSI_MEM_SOME] += delta;
678 		if (groupc->state_mask & (1 << PSI_MEM_FULL))
679 			groupc->times[PSI_MEM_FULL] += delta;
680 	}
681 
682 	if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
683 		groupc->times[PSI_CPU_SOME] += delta;
684 		if (groupc->state_mask & (1 << PSI_CPU_FULL))
685 			groupc->times[PSI_CPU_FULL] += delta;
686 	}
687 
688 	if (groupc->state_mask & (1 << PSI_NONIDLE))
689 		groupc->times[PSI_NONIDLE] += delta;
690 }
691 
692 static void psi_group_change(struct psi_group *group, int cpu,
693 			     unsigned int clear, unsigned int set, u64 now,
694 			     bool wake_clock)
695 {
696 	struct psi_group_cpu *groupc;
697 	u32 state_mask = 0;
698 	unsigned int t, m;
699 	enum psi_states s;
700 
701 	groupc = per_cpu_ptr(group->pcpu, cpu);
702 
703 	/*
704 	 * First we assess the aggregate resource states this CPU's
705 	 * tasks have been in since the last change, and account any
706 	 * SOME and FULL time these may have resulted in.
707 	 *
708 	 * Then we update the task counts according to the state
709 	 * change requested through the @clear and @set bits.
710 	 */
711 	write_seqcount_begin(&groupc->seq);
712 
713 	record_times(groupc, now);
714 
715 	for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
716 		if (!(m & (1 << t)))
717 			continue;
718 		if (groupc->tasks[t]) {
719 			groupc->tasks[t]--;
720 		} else if (!psi_bug) {
721 			printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u %u] clear=%x set=%x\n",
722 					cpu, t, groupc->tasks[0],
723 					groupc->tasks[1], groupc->tasks[2],
724 					groupc->tasks[3], groupc->tasks[4],
725 					clear, set);
726 			psi_bug = 1;
727 		}
728 	}
729 
730 	for (t = 0; set; set &= ~(1 << t), t++)
731 		if (set & (1 << t))
732 			groupc->tasks[t]++;
733 
734 	/* Calculate state mask representing active states */
735 	for (s = 0; s < NR_PSI_STATES; s++) {
736 		if (test_state(groupc->tasks, s))
737 			state_mask |= (1 << s);
738 	}
739 
740 	/*
741 	 * Since we care about lost potential, a memstall is FULL
742 	 * when there are no other working tasks, but also when
743 	 * the CPU is actively reclaiming and nothing productive
744 	 * could run even if it were runnable. So when the current
745 	 * task in a cgroup is in_memstall, the corresponding groupc
746 	 * on that cpu is in PSI_MEM_FULL state.
747 	 */
748 	if (unlikely(groupc->tasks[NR_ONCPU] && cpu_curr(cpu)->in_memstall))
749 		state_mask |= (1 << PSI_MEM_FULL);
750 
751 	groupc->state_mask = state_mask;
752 
753 	write_seqcount_end(&groupc->seq);
754 
755 	if (state_mask & group->poll_states)
756 		psi_schedule_poll_work(group, 1);
757 
758 	if (wake_clock && !delayed_work_pending(&group->avgs_work))
759 		schedule_delayed_work(&group->avgs_work, PSI_FREQ);
760 }
761 
762 static struct psi_group *iterate_groups(struct task_struct *task, void **iter)
763 {
764 	if (*iter == &psi_system)
765 		return NULL;
766 
767 #ifdef CONFIG_CGROUPS
768 	if (static_branch_likely(&psi_cgroups_enabled)) {
769 		struct cgroup *cgroup = NULL;
770 
771 		if (!*iter)
772 			cgroup = task->cgroups->dfl_cgrp;
773 		else
774 			cgroup = cgroup_parent(*iter);
775 
776 		if (cgroup && cgroup_parent(cgroup)) {
777 			*iter = cgroup;
778 			return cgroup_psi(cgroup);
779 		}
780 	}
781 #endif
782 	*iter = &psi_system;
783 	return &psi_system;
784 }
785 
786 static void psi_flags_change(struct task_struct *task, int clear, int set)
787 {
788 	if (((task->psi_flags & set) ||
789 	     (task->psi_flags & clear) != clear) &&
790 	    !psi_bug) {
791 		printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
792 				task->pid, task->comm, task_cpu(task),
793 				task->psi_flags, clear, set);
794 		psi_bug = 1;
795 	}
796 
797 	task->psi_flags &= ~clear;
798 	task->psi_flags |= set;
799 }
800 
801 void psi_task_change(struct task_struct *task, int clear, int set)
802 {
803 	int cpu = task_cpu(task);
804 	struct psi_group *group;
805 	bool wake_clock = true;
806 	void *iter = NULL;
807 	u64 now;
808 
809 	if (!task->pid)
810 		return;
811 
812 	psi_flags_change(task, clear, set);
813 
814 	now = cpu_clock(cpu);
815 	/*
816 	 * Periodic aggregation shuts off if there is a period of no
817 	 * task changes, so we wake it back up if necessary. However,
818 	 * don't do this if the task change is the aggregation worker
819 	 * itself going to sleep, or we'll ping-pong forever.
820 	 */
821 	if (unlikely((clear & TSK_RUNNING) &&
822 		     (task->flags & PF_WQ_WORKER) &&
823 		     wq_worker_last_func(task) == psi_avgs_work))
824 		wake_clock = false;
825 
826 	while ((group = iterate_groups(task, &iter)))
827 		psi_group_change(group, cpu, clear, set, now, wake_clock);
828 }
829 
830 void psi_task_switch(struct task_struct *prev, struct task_struct *next,
831 		     bool sleep)
832 {
833 	struct psi_group *group, *common = NULL;
834 	int cpu = task_cpu(prev);
835 	void *iter;
836 	u64 now = cpu_clock(cpu);
837 
838 	if (next->pid) {
839 		bool identical_state;
840 
841 		psi_flags_change(next, 0, TSK_ONCPU);
842 		/*
843 		 * When switching between tasks that have an identical
844 		 * runtime state, the cgroup that contains both tasks
845 		 * we reach the first common ancestor. Iterate @next's
846 		 * ancestors only until we encounter @prev's ONCPU.
847 		 */
848 		identical_state = prev->psi_flags == next->psi_flags;
849 		iter = NULL;
850 		while ((group = iterate_groups(next, &iter))) {
851 			if (identical_state &&
852 			    per_cpu_ptr(group->pcpu, cpu)->tasks[NR_ONCPU]) {
853 				common = group;
854 				break;
855 			}
856 
857 			psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
858 		}
859 	}
860 
861 	if (prev->pid) {
862 		int clear = TSK_ONCPU, set = 0;
863 
864 		/*
865 		 * When we're going to sleep, psi_dequeue() lets us
866 		 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
867 		 * TSK_IOWAIT here, where we can combine it with
868 		 * TSK_ONCPU and save walking common ancestors twice.
869 		 */
870 		if (sleep) {
871 			clear |= TSK_RUNNING;
872 			if (prev->in_memstall)
873 				clear |= TSK_MEMSTALL_RUNNING;
874 			if (prev->in_iowait)
875 				set |= TSK_IOWAIT;
876 		}
877 
878 		psi_flags_change(prev, clear, set);
879 
880 		iter = NULL;
881 		while ((group = iterate_groups(prev, &iter)) && group != common)
882 			psi_group_change(group, cpu, clear, set, now, true);
883 
884 		/*
885 		 * TSK_ONCPU is handled up to the common ancestor. If we're tasked
886 		 * with dequeuing too, finish that for the rest of the hierarchy.
887 		 */
888 		if (sleep) {
889 			clear &= ~TSK_ONCPU;
890 			for (; group; group = iterate_groups(prev, &iter))
891 				psi_group_change(group, cpu, clear, set, now, true);
892 		}
893 	}
894 }
895 
896 /**
897  * psi_memstall_enter - mark the beginning of a memory stall section
898  * @flags: flags to handle nested sections
899  *
900  * Marks the calling task as being stalled due to a lack of memory,
901  * such as waiting for a refault or performing reclaim.
902  */
903 void psi_memstall_enter(unsigned long *flags)
904 {
905 	struct rq_flags rf;
906 	struct rq *rq;
907 
908 	if (static_branch_likely(&psi_disabled))
909 		return;
910 
911 	*flags = current->in_memstall;
912 	if (*flags)
913 		return;
914 	/*
915 	 * in_memstall setting & accounting needs to be atomic wrt
916 	 * changes to the task's scheduling state, otherwise we can
917 	 * race with CPU migration.
918 	 */
919 	rq = this_rq_lock_irq(&rf);
920 
921 	current->in_memstall = 1;
922 	psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
923 
924 	rq_unlock_irq(rq, &rf);
925 }
926 
927 /**
928  * psi_memstall_leave - mark the end of an memory stall section
929  * @flags: flags to handle nested memdelay sections
930  *
931  * Marks the calling task as no longer stalled due to lack of memory.
932  */
933 void psi_memstall_leave(unsigned long *flags)
934 {
935 	struct rq_flags rf;
936 	struct rq *rq;
937 
938 	if (static_branch_likely(&psi_disabled))
939 		return;
940 
941 	if (*flags)
942 		return;
943 	/*
944 	 * in_memstall clearing & accounting needs to be atomic wrt
945 	 * changes to the task's scheduling state, otherwise we could
946 	 * race with CPU migration.
947 	 */
948 	rq = this_rq_lock_irq(&rf);
949 
950 	current->in_memstall = 0;
951 	psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0);
952 
953 	rq_unlock_irq(rq, &rf);
954 }
955 
956 #ifdef CONFIG_CGROUPS
957 int psi_cgroup_alloc(struct cgroup *cgroup)
958 {
959 	if (static_branch_likely(&psi_disabled))
960 		return 0;
961 
962 	cgroup->psi.pcpu = alloc_percpu(struct psi_group_cpu);
963 	if (!cgroup->psi.pcpu)
964 		return -ENOMEM;
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 }
979 
980 /**
981  * cgroup_move_task - move task to a different cgroup
982  * @task: the task
983  * @to: the target css_set
984  *
985  * Move task to a new cgroup and safely migrate its associated stall
986  * state between the different groups.
987  *
988  * This function acquires the task's rq lock to lock out concurrent
989  * changes to the task's scheduling state and - in case the task is
990  * running - concurrent changes to its stall state.
991  */
992 void cgroup_move_task(struct task_struct *task, struct css_set *to)
993 {
994 	unsigned int task_flags;
995 	struct rq_flags rf;
996 	struct rq *rq;
997 
998 	if (static_branch_likely(&psi_disabled)) {
999 		/*
1000 		 * Lame to do this here, but the scheduler cannot be locked
1001 		 * from the outside, so we move cgroups from inside sched/.
1002 		 */
1003 		rcu_assign_pointer(task->cgroups, to);
1004 		return;
1005 	}
1006 
1007 	rq = task_rq_lock(task, &rf);
1008 
1009 	/*
1010 	 * We may race with schedule() dropping the rq lock between
1011 	 * deactivating prev and switching to next. Because the psi
1012 	 * updates from the deactivation are deferred to the switch
1013 	 * callback to save cgroup tree updates, the task's scheduling
1014 	 * state here is not coherent with its psi state:
1015 	 *
1016 	 * schedule()                   cgroup_move_task()
1017 	 *   rq_lock()
1018 	 *   deactivate_task()
1019 	 *     p->on_rq = 0
1020 	 *     psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1021 	 *   pick_next_task()
1022 	 *     rq_unlock()
1023 	 *                                rq_lock()
1024 	 *                                psi_task_change() // old cgroup
1025 	 *                                task->cgroups = to
1026 	 *                                psi_task_change() // new cgroup
1027 	 *                                rq_unlock()
1028 	 *     rq_lock()
1029 	 *   psi_sched_switch() // does deferred updates in new cgroup
1030 	 *
1031 	 * Don't rely on the scheduling state. Use psi_flags instead.
1032 	 */
1033 	task_flags = task->psi_flags;
1034 
1035 	if (task_flags)
1036 		psi_task_change(task, task_flags, 0);
1037 
1038 	/* See comment above */
1039 	rcu_assign_pointer(task->cgroups, to);
1040 
1041 	if (task_flags)
1042 		psi_task_change(task, 0, task_flags);
1043 
1044 	task_rq_unlock(rq, task, &rf);
1045 }
1046 #endif /* CONFIG_CGROUPS */
1047 
1048 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1049 {
1050 	int full;
1051 	u64 now;
1052 
1053 	if (static_branch_likely(&psi_disabled))
1054 		return -EOPNOTSUPP;
1055 
1056 	/* Update averages before reporting them */
1057 	mutex_lock(&group->avgs_lock);
1058 	now = sched_clock();
1059 	collect_percpu_times(group, PSI_AVGS, NULL);
1060 	if (now >= group->avg_next_update)
1061 		group->avg_next_update = update_averages(group, now);
1062 	mutex_unlock(&group->avgs_lock);
1063 
1064 	for (full = 0; full < 2; full++) {
1065 		unsigned long avg[3];
1066 		u64 total;
1067 		int w;
1068 
1069 		for (w = 0; w < 3; w++)
1070 			avg[w] = group->avg[res * 2 + full][w];
1071 		total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1072 				NSEC_PER_USEC);
1073 
1074 		seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1075 			   full ? "full" : "some",
1076 			   LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1077 			   LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1078 			   LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1079 			   total);
1080 	}
1081 
1082 	return 0;
1083 }
1084 
1085 struct psi_trigger *psi_trigger_create(struct psi_group *group,
1086 			char *buf, size_t nbytes, enum psi_res res)
1087 {
1088 	struct psi_trigger *t;
1089 	enum psi_states state;
1090 	u32 threshold_us;
1091 	u32 window_us;
1092 
1093 	if (static_branch_likely(&psi_disabled))
1094 		return ERR_PTR(-EOPNOTSUPP);
1095 
1096 	if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1097 		state = PSI_IO_SOME + res * 2;
1098 	else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1099 		state = PSI_IO_FULL + res * 2;
1100 	else
1101 		return ERR_PTR(-EINVAL);
1102 
1103 	if (state >= PSI_NONIDLE)
1104 		return ERR_PTR(-EINVAL);
1105 
1106 	if (window_us < WINDOW_MIN_US ||
1107 		window_us > WINDOW_MAX_US)
1108 		return ERR_PTR(-EINVAL);
1109 
1110 	/* Check threshold */
1111 	if (threshold_us == 0 || threshold_us > window_us)
1112 		return ERR_PTR(-EINVAL);
1113 
1114 	t = kmalloc(sizeof(*t), GFP_KERNEL);
1115 	if (!t)
1116 		return ERR_PTR(-ENOMEM);
1117 
1118 	t->group = group;
1119 	t->state = state;
1120 	t->threshold = threshold_us * NSEC_PER_USEC;
1121 	t->win.size = window_us * NSEC_PER_USEC;
1122 	window_reset(&t->win, 0, 0, 0);
1123 
1124 	t->event = 0;
1125 	t->last_event_time = 0;
1126 	init_waitqueue_head(&t->event_wait);
1127 
1128 	mutex_lock(&group->trigger_lock);
1129 
1130 	if (!rcu_access_pointer(group->poll_task)) {
1131 		struct task_struct *task;
1132 
1133 		task = kthread_create(psi_poll_worker, group, "psimon");
1134 		if (IS_ERR(task)) {
1135 			kfree(t);
1136 			mutex_unlock(&group->trigger_lock);
1137 			return ERR_CAST(task);
1138 		}
1139 		atomic_set(&group->poll_wakeup, 0);
1140 		wake_up_process(task);
1141 		rcu_assign_pointer(group->poll_task, task);
1142 	}
1143 
1144 	list_add(&t->node, &group->triggers);
1145 	group->poll_min_period = min(group->poll_min_period,
1146 		div_u64(t->win.size, UPDATES_PER_WINDOW));
1147 	group->nr_triggers[t->state]++;
1148 	group->poll_states |= (1 << t->state);
1149 
1150 	mutex_unlock(&group->trigger_lock);
1151 
1152 	return t;
1153 }
1154 
1155 void psi_trigger_destroy(struct psi_trigger *t)
1156 {
1157 	struct psi_group *group;
1158 	struct task_struct *task_to_destroy = NULL;
1159 
1160 	/*
1161 	 * We do not check psi_disabled since it might have been disabled after
1162 	 * the trigger got created.
1163 	 */
1164 	if (!t)
1165 		return;
1166 
1167 	group = t->group;
1168 	/*
1169 	 * Wakeup waiters to stop polling. Can happen if cgroup is deleted
1170 	 * from under a polling process.
1171 	 */
1172 	wake_up_interruptible(&t->event_wait);
1173 
1174 	mutex_lock(&group->trigger_lock);
1175 
1176 	if (!list_empty(&t->node)) {
1177 		struct psi_trigger *tmp;
1178 		u64 period = ULLONG_MAX;
1179 
1180 		list_del(&t->node);
1181 		group->nr_triggers[t->state]--;
1182 		if (!group->nr_triggers[t->state])
1183 			group->poll_states &= ~(1 << t->state);
1184 		/* reset min update period for the remaining triggers */
1185 		list_for_each_entry(tmp, &group->triggers, node)
1186 			period = min(period, div_u64(tmp->win.size,
1187 					UPDATES_PER_WINDOW));
1188 		group->poll_min_period = period;
1189 		/* Destroy poll_task when the last trigger is destroyed */
1190 		if (group->poll_states == 0) {
1191 			group->polling_until = 0;
1192 			task_to_destroy = rcu_dereference_protected(
1193 					group->poll_task,
1194 					lockdep_is_held(&group->trigger_lock));
1195 			rcu_assign_pointer(group->poll_task, NULL);
1196 			del_timer(&group->poll_timer);
1197 		}
1198 	}
1199 
1200 	mutex_unlock(&group->trigger_lock);
1201 
1202 	/*
1203 	 * Wait for psi_schedule_poll_work RCU to complete its read-side
1204 	 * critical section before destroying the trigger and optionally the
1205 	 * poll_task.
1206 	 */
1207 	synchronize_rcu();
1208 	/*
1209 	 * Stop kthread 'psimon' after releasing trigger_lock to prevent a
1210 	 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1211 	 */
1212 	if (task_to_destroy) {
1213 		/*
1214 		 * After the RCU grace period has expired, the worker
1215 		 * can no longer be found through group->poll_task.
1216 		 */
1217 		kthread_stop(task_to_destroy);
1218 	}
1219 	kfree(t);
1220 }
1221 
1222 __poll_t psi_trigger_poll(void **trigger_ptr,
1223 				struct file *file, poll_table *wait)
1224 {
1225 	__poll_t ret = DEFAULT_POLLMASK;
1226 	struct psi_trigger *t;
1227 
1228 	if (static_branch_likely(&psi_disabled))
1229 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1230 
1231 	t = smp_load_acquire(trigger_ptr);
1232 	if (!t)
1233 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1234 
1235 	poll_wait(file, &t->event_wait, wait);
1236 
1237 	if (cmpxchg(&t->event, 1, 0) == 1)
1238 		ret |= EPOLLPRI;
1239 
1240 	return ret;
1241 }
1242 
1243 #ifdef CONFIG_PROC_FS
1244 static int psi_io_show(struct seq_file *m, void *v)
1245 {
1246 	return psi_show(m, &psi_system, PSI_IO);
1247 }
1248 
1249 static int psi_memory_show(struct seq_file *m, void *v)
1250 {
1251 	return psi_show(m, &psi_system, PSI_MEM);
1252 }
1253 
1254 static int psi_cpu_show(struct seq_file *m, void *v)
1255 {
1256 	return psi_show(m, &psi_system, PSI_CPU);
1257 }
1258 
1259 static int psi_open(struct file *file, int (*psi_show)(struct seq_file *, void *))
1260 {
1261 	if (file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE))
1262 		return -EPERM;
1263 
1264 	return single_open(file, psi_show, NULL);
1265 }
1266 
1267 static int psi_io_open(struct inode *inode, struct file *file)
1268 {
1269 	return psi_open(file, psi_io_show);
1270 }
1271 
1272 static int psi_memory_open(struct inode *inode, struct file *file)
1273 {
1274 	return psi_open(file, psi_memory_show);
1275 }
1276 
1277 static int psi_cpu_open(struct inode *inode, struct file *file)
1278 {
1279 	return psi_open(file, psi_cpu_show);
1280 }
1281 
1282 static ssize_t psi_write(struct file *file, const char __user *user_buf,
1283 			 size_t nbytes, enum psi_res res)
1284 {
1285 	char buf[32];
1286 	size_t buf_size;
1287 	struct seq_file *seq;
1288 	struct psi_trigger *new;
1289 
1290 	if (static_branch_likely(&psi_disabled))
1291 		return -EOPNOTSUPP;
1292 
1293 	if (!nbytes)
1294 		return -EINVAL;
1295 
1296 	buf_size = min(nbytes, sizeof(buf));
1297 	if (copy_from_user(buf, user_buf, buf_size))
1298 		return -EFAULT;
1299 
1300 	buf[buf_size - 1] = '\0';
1301 
1302 	seq = file->private_data;
1303 
1304 	/* Take seq->lock to protect seq->private from concurrent writes */
1305 	mutex_lock(&seq->lock);
1306 
1307 	/* Allow only one trigger per file descriptor */
1308 	if (seq->private) {
1309 		mutex_unlock(&seq->lock);
1310 		return -EBUSY;
1311 	}
1312 
1313 	new = psi_trigger_create(&psi_system, buf, nbytes, res);
1314 	if (IS_ERR(new)) {
1315 		mutex_unlock(&seq->lock);
1316 		return PTR_ERR(new);
1317 	}
1318 
1319 	smp_store_release(&seq->private, new);
1320 	mutex_unlock(&seq->lock);
1321 
1322 	return nbytes;
1323 }
1324 
1325 static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1326 			    size_t nbytes, loff_t *ppos)
1327 {
1328 	return psi_write(file, user_buf, nbytes, PSI_IO);
1329 }
1330 
1331 static ssize_t psi_memory_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_MEM);
1335 }
1336 
1337 static ssize_t psi_cpu_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_CPU);
1341 }
1342 
1343 static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1344 {
1345 	struct seq_file *seq = file->private_data;
1346 
1347 	return psi_trigger_poll(&seq->private, file, wait);
1348 }
1349 
1350 static int psi_fop_release(struct inode *inode, struct file *file)
1351 {
1352 	struct seq_file *seq = file->private_data;
1353 
1354 	psi_trigger_destroy(seq->private);
1355 	return single_release(inode, file);
1356 }
1357 
1358 static const struct proc_ops psi_io_proc_ops = {
1359 	.proc_open	= psi_io_open,
1360 	.proc_read	= seq_read,
1361 	.proc_lseek	= seq_lseek,
1362 	.proc_write	= psi_io_write,
1363 	.proc_poll	= psi_fop_poll,
1364 	.proc_release	= psi_fop_release,
1365 };
1366 
1367 static const struct proc_ops psi_memory_proc_ops = {
1368 	.proc_open	= psi_memory_open,
1369 	.proc_read	= seq_read,
1370 	.proc_lseek	= seq_lseek,
1371 	.proc_write	= psi_memory_write,
1372 	.proc_poll	= psi_fop_poll,
1373 	.proc_release	= psi_fop_release,
1374 };
1375 
1376 static const struct proc_ops psi_cpu_proc_ops = {
1377 	.proc_open	= psi_cpu_open,
1378 	.proc_read	= seq_read,
1379 	.proc_lseek	= seq_lseek,
1380 	.proc_write	= psi_cpu_write,
1381 	.proc_poll	= psi_fop_poll,
1382 	.proc_release	= psi_fop_release,
1383 };
1384 
1385 static int __init psi_proc_init(void)
1386 {
1387 	if (psi_enable) {
1388 		proc_mkdir("pressure", NULL);
1389 		proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops);
1390 		proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops);
1391 		proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops);
1392 	}
1393 	return 0;
1394 }
1395 module_init(psi_proc_init);
1396 
1397 #endif /* CONFIG_PROC_FS */
1398