xref: /openbmc/linux/kernel/sched/psi.c (revision e6f4c346)
1 /*
2  * Pressure stall information for CPU, memory and IO
3  *
4  * Copyright (c) 2018 Facebook, Inc.
5  * Author: Johannes Weiner <hannes@cmpxchg.org>
6  *
7  * Polling support by Suren Baghdasaryan <surenb@google.com>
8  * Copyright (c) 2018 Google, Inc.
9  *
10  * When CPU, memory and IO are contended, tasks experience delays that
11  * reduce throughput and introduce latencies into the workload. Memory
12  * and IO contention, in addition, can cause a full loss of forward
13  * progress in which the CPU goes idle.
14  *
15  * This code aggregates individual task delays into resource pressure
16  * metrics that indicate problems with both workload health and
17  * resource utilization.
18  *
19  *			Model
20  *
21  * The time in which a task can execute on a CPU is our baseline for
22  * productivity. Pressure expresses the amount of time in which this
23  * potential cannot be realized due to resource contention.
24  *
25  * This concept of productivity has two components: the workload and
26  * the CPU. To measure the impact of pressure on both, we define two
27  * contention states for a resource: SOME and FULL.
28  *
29  * In the SOME state of a given resource, one or more tasks are
30  * delayed on that resource. This affects the workload's ability to
31  * perform work, but the CPU may still be executing other tasks.
32  *
33  * In the FULL state of a given resource, all non-idle tasks are
34  * delayed on that resource such that nobody is advancing and the CPU
35  * goes idle. This leaves both workload and CPU unproductive.
36  *
37  * (Naturally, the FULL state doesn't exist for the CPU resource.)
38  *
39  *	SOME = nr_delayed_tasks != 0
40  *	FULL = nr_delayed_tasks != 0 && nr_running_tasks == 0
41  *
42  * The percentage of wallclock time spent in those compound stall
43  * states gives pressure numbers between 0 and 100 for each resource,
44  * where the SOME percentage indicates workload slowdowns and the FULL
45  * percentage indicates reduced CPU utilization:
46  *
47  *	%SOME = time(SOME) / period
48  *	%FULL = time(FULL) / period
49  *
50  *			Multiple CPUs
51  *
52  * The more tasks and available CPUs there are, the more work can be
53  * performed concurrently. This means that the potential that can go
54  * unrealized due to resource contention *also* scales with non-idle
55  * tasks and CPUs.
56  *
57  * Consider a scenario where 257 number crunching tasks are trying to
58  * run concurrently on 256 CPUs. If we simply aggregated the task
59  * states, we would have to conclude a CPU SOME pressure number of
60  * 100%, since *somebody* is waiting on a runqueue at all
61  * times. However, that is clearly not the amount of contention the
62  * workload is experiencing: only one out of 256 possible exceution
63  * threads will be contended at any given time, or about 0.4%.
64  *
65  * Conversely, consider a scenario of 4 tasks and 4 CPUs where at any
66  * given time *one* of the tasks is delayed due to a lack of memory.
67  * Again, looking purely at the task state would yield a memory FULL
68  * pressure number of 0%, since *somebody* is always making forward
69  * progress. But again this wouldn't capture the amount of execution
70  * potential lost, which is 1 out of 4 CPUs, or 25%.
71  *
72  * To calculate wasted potential (pressure) with multiple processors,
73  * we have to base our calculation on the number of non-idle tasks in
74  * conjunction with the number of available CPUs, which is the number
75  * of potential execution threads. SOME becomes then the proportion of
76  * delayed tasks to possibe threads, and FULL is the share of possible
77  * threads that are unproductive due to delays:
78  *
79  *	threads = min(nr_nonidle_tasks, nr_cpus)
80  *	   SOME = min(nr_delayed_tasks / threads, 1)
81  *	   FULL = (threads - min(nr_running_tasks, threads)) / threads
82  *
83  * For the 257 number crunchers on 256 CPUs, this yields:
84  *
85  *	threads = min(257, 256)
86  *	   SOME = min(1 / 256, 1)             = 0.4%
87  *	   FULL = (256 - min(257, 256)) / 256 = 0%
88  *
89  * For the 1 out of 4 memory-delayed tasks, this yields:
90  *
91  *	threads = min(4, 4)
92  *	   SOME = min(1 / 4, 1)               = 25%
93  *	   FULL = (4 - min(3, 4)) / 4         = 25%
94  *
95  * [ Substitute nr_cpus with 1, and you can see that it's a natural
96  *   extension of the single-CPU model. ]
97  *
98  *			Implementation
99  *
100  * To assess the precise time spent in each such state, we would have
101  * to freeze the system on task changes and start/stop the state
102  * clocks accordingly. Obviously that doesn't scale in practice.
103  *
104  * Because the scheduler aims to distribute the compute load evenly
105  * among the available CPUs, we can track task state locally to each
106  * CPU and, at much lower frequency, extrapolate the global state for
107  * the cumulative stall times and the running averages.
108  *
109  * For each runqueue, we track:
110  *
111  *	   tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0)
112  *	   tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_running_tasks[cpu])
113  *	tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0)
114  *
115  * and then periodically aggregate:
116  *
117  *	tNONIDLE = sum(tNONIDLE[i])
118  *
119  *	   tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE
120  *	   tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE
121  *
122  *	   %SOME = tSOME / period
123  *	   %FULL = tFULL / period
124  *
125  * This gives us an approximation of pressure that is practical
126  * cost-wise, yet way more sensitive and accurate than periodic
127  * sampling of the aggregate task states would be.
128  */
129 
130 #include "../workqueue_internal.h"
131 #include <linux/sched/loadavg.h>
132 #include <linux/seq_file.h>
133 #include <linux/proc_fs.h>
134 #include <linux/seqlock.h>
135 #include <linux/uaccess.h>
136 #include <linux/cgroup.h>
137 #include <linux/module.h>
138 #include <linux/sched.h>
139 #include <linux/ctype.h>
140 #include <linux/file.h>
141 #include <linux/poll.h>
142 #include <linux/psi.h>
143 #include "sched.h"
144 
145 static int psi_bug __read_mostly;
146 
147 DEFINE_STATIC_KEY_FALSE(psi_disabled);
148 
149 #ifdef CONFIG_PSI_DEFAULT_DISABLED
150 static bool psi_enable;
151 #else
152 static bool psi_enable = true;
153 #endif
154 static int __init setup_psi(char *str)
155 {
156 	return kstrtobool(str, &psi_enable) == 0;
157 }
158 __setup("psi=", setup_psi);
159 
160 /* Running averages - we need to be higher-res than loadavg */
161 #define PSI_FREQ	(2*HZ+1)	/* 2 sec intervals */
162 #define EXP_10s		1677		/* 1/exp(2s/10s) as fixed-point */
163 #define EXP_60s		1981		/* 1/exp(2s/60s) */
164 #define EXP_300s	2034		/* 1/exp(2s/300s) */
165 
166 /* PSI trigger definitions */
167 #define WINDOW_MIN_US 500000	/* Min window size is 500ms */
168 #define WINDOW_MAX_US 10000000	/* Max window size is 10s */
169 #define UPDATES_PER_WINDOW 10	/* 10 updates per window */
170 
171 /* Sampling frequency in nanoseconds */
172 static u64 psi_period __read_mostly;
173 
174 /* System-level pressure and stall tracking */
175 static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
176 struct psi_group psi_system = {
177 	.pcpu = &system_group_pcpu,
178 };
179 
180 static void psi_avgs_work(struct work_struct *work);
181 
182 static void group_init(struct psi_group *group)
183 {
184 	int cpu;
185 
186 	for_each_possible_cpu(cpu)
187 		seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);
188 	group->avg_next_update = sched_clock() + psi_period;
189 	INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
190 	mutex_init(&group->avgs_lock);
191 	/* Init trigger-related members */
192 	atomic_set(&group->poll_scheduled, 0);
193 	mutex_init(&group->trigger_lock);
194 	INIT_LIST_HEAD(&group->triggers);
195 	memset(group->nr_triggers, 0, sizeof(group->nr_triggers));
196 	group->poll_states = 0;
197 	group->poll_min_period = U32_MAX;
198 	memset(group->polling_total, 0, sizeof(group->polling_total));
199 	group->polling_next_update = ULLONG_MAX;
200 	group->polling_until = 0;
201 	rcu_assign_pointer(group->poll_kworker, NULL);
202 }
203 
204 void __init psi_init(void)
205 {
206 	if (!psi_enable) {
207 		static_branch_enable(&psi_disabled);
208 		return;
209 	}
210 
211 	psi_period = jiffies_to_nsecs(PSI_FREQ);
212 	group_init(&psi_system);
213 }
214 
215 static bool test_state(unsigned int *tasks, enum psi_states state)
216 {
217 	switch (state) {
218 	case PSI_IO_SOME:
219 		return tasks[NR_IOWAIT];
220 	case PSI_IO_FULL:
221 		return tasks[NR_IOWAIT] && !tasks[NR_RUNNING];
222 	case PSI_MEM_SOME:
223 		return tasks[NR_MEMSTALL];
224 	case PSI_MEM_FULL:
225 		return tasks[NR_MEMSTALL] && !tasks[NR_RUNNING];
226 	case PSI_CPU_SOME:
227 		return tasks[NR_RUNNING] > 1;
228 	case PSI_NONIDLE:
229 		return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
230 			tasks[NR_RUNNING];
231 	default:
232 		return false;
233 	}
234 }
235 
236 static void get_recent_times(struct psi_group *group, int cpu,
237 			     enum psi_aggregators aggregator, u32 *times,
238 			     u32 *pchanged_states)
239 {
240 	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
241 	u64 now, state_start;
242 	enum psi_states s;
243 	unsigned int seq;
244 	u32 state_mask;
245 
246 	*pchanged_states = 0;
247 
248 	/* Snapshot a coherent view of the CPU state */
249 	do {
250 		seq = read_seqcount_begin(&groupc->seq);
251 		now = cpu_clock(cpu);
252 		memcpy(times, groupc->times, sizeof(groupc->times));
253 		state_mask = groupc->state_mask;
254 		state_start = groupc->state_start;
255 	} while (read_seqcount_retry(&groupc->seq, seq));
256 
257 	/* Calculate state time deltas against the previous snapshot */
258 	for (s = 0; s < NR_PSI_STATES; s++) {
259 		u32 delta;
260 		/*
261 		 * In addition to already concluded states, we also
262 		 * incorporate currently active states on the CPU,
263 		 * since states may last for many sampling periods.
264 		 *
265 		 * This way we keep our delta sampling buckets small
266 		 * (u32) and our reported pressure close to what's
267 		 * actually happening.
268 		 */
269 		if (state_mask & (1 << s))
270 			times[s] += now - state_start;
271 
272 		delta = times[s] - groupc->times_prev[aggregator][s];
273 		groupc->times_prev[aggregator][s] = times[s];
274 
275 		times[s] = delta;
276 		if (delta)
277 			*pchanged_states |= (1 << s);
278 	}
279 }
280 
281 static void calc_avgs(unsigned long avg[3], int missed_periods,
282 		      u64 time, u64 period)
283 {
284 	unsigned long pct;
285 
286 	/* Fill in zeroes for periods of no activity */
287 	if (missed_periods) {
288 		avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
289 		avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
290 		avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
291 	}
292 
293 	/* Sample the most recent active period */
294 	pct = div_u64(time * 100, period);
295 	pct *= FIXED_1;
296 	avg[0] = calc_load(avg[0], EXP_10s, pct);
297 	avg[1] = calc_load(avg[1], EXP_60s, pct);
298 	avg[2] = calc_load(avg[2], EXP_300s, pct);
299 }
300 
301 static void collect_percpu_times(struct psi_group *group,
302 				 enum psi_aggregators aggregator,
303 				 u32 *pchanged_states)
304 {
305 	u64 deltas[NR_PSI_STATES - 1] = { 0, };
306 	unsigned long nonidle_total = 0;
307 	u32 changed_states = 0;
308 	int cpu;
309 	int s;
310 
311 	/*
312 	 * Collect the per-cpu time buckets and average them into a
313 	 * single time sample that is normalized to wallclock time.
314 	 *
315 	 * For averaging, each CPU is weighted by its non-idle time in
316 	 * the sampling period. This eliminates artifacts from uneven
317 	 * loading, or even entirely idle CPUs.
318 	 */
319 	for_each_possible_cpu(cpu) {
320 		u32 times[NR_PSI_STATES];
321 		u32 nonidle;
322 		u32 cpu_changed_states;
323 
324 		get_recent_times(group, cpu, aggregator, times,
325 				&cpu_changed_states);
326 		changed_states |= cpu_changed_states;
327 
328 		nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
329 		nonidle_total += nonidle;
330 
331 		for (s = 0; s < PSI_NONIDLE; s++)
332 			deltas[s] += (u64)times[s] * nonidle;
333 	}
334 
335 	/*
336 	 * Integrate the sample into the running statistics that are
337 	 * reported to userspace: the cumulative stall times and the
338 	 * decaying averages.
339 	 *
340 	 * Pressure percentages are sampled at PSI_FREQ. We might be
341 	 * called more often when the user polls more frequently than
342 	 * that; we might be called less often when there is no task
343 	 * activity, thus no data, and clock ticks are sporadic. The
344 	 * below handles both.
345 	 */
346 
347 	/* total= */
348 	for (s = 0; s < NR_PSI_STATES - 1; s++)
349 		group->total[aggregator][s] +=
350 				div_u64(deltas[s], max(nonidle_total, 1UL));
351 
352 	if (pchanged_states)
353 		*pchanged_states = changed_states;
354 }
355 
356 static u64 update_averages(struct psi_group *group, u64 now)
357 {
358 	unsigned long missed_periods = 0;
359 	u64 expires, period;
360 	u64 avg_next_update;
361 	int s;
362 
363 	/* avgX= */
364 	expires = group->avg_next_update;
365 	if (now - expires >= psi_period)
366 		missed_periods = div_u64(now - expires, psi_period);
367 
368 	/*
369 	 * The periodic clock tick can get delayed for various
370 	 * reasons, especially on loaded systems. To avoid clock
371 	 * drift, we schedule the clock in fixed psi_period intervals.
372 	 * But the deltas we sample out of the per-cpu buckets above
373 	 * are based on the actual time elapsing between clock ticks.
374 	 */
375 	avg_next_update = expires + ((1 + missed_periods) * psi_period);
376 	period = now - (group->avg_last_update + (missed_periods * psi_period));
377 	group->avg_last_update = now;
378 
379 	for (s = 0; s < NR_PSI_STATES - 1; s++) {
380 		u32 sample;
381 
382 		sample = group->total[PSI_AVGS][s] - group->avg_total[s];
383 		/*
384 		 * Due to the lockless sampling of the time buckets,
385 		 * recorded time deltas can slip into the next period,
386 		 * which under full pressure can result in samples in
387 		 * excess of the period length.
388 		 *
389 		 * We don't want to report non-sensical pressures in
390 		 * excess of 100%, nor do we want to drop such events
391 		 * on the floor. Instead we punt any overage into the
392 		 * future until pressure subsides. By doing this we
393 		 * don't underreport the occurring pressure curve, we
394 		 * just report it delayed by one period length.
395 		 *
396 		 * The error isn't cumulative. As soon as another
397 		 * delta slips from a period P to P+1, by definition
398 		 * it frees up its time T in P.
399 		 */
400 		if (sample > period)
401 			sample = period;
402 		group->avg_total[s] += sample;
403 		calc_avgs(group->avg[s], missed_periods, sample, period);
404 	}
405 
406 	return avg_next_update;
407 }
408 
409 static void psi_avgs_work(struct work_struct *work)
410 {
411 	struct delayed_work *dwork;
412 	struct psi_group *group;
413 	u32 changed_states;
414 	bool nonidle;
415 	u64 now;
416 
417 	dwork = to_delayed_work(work);
418 	group = container_of(dwork, struct psi_group, avgs_work);
419 
420 	mutex_lock(&group->avgs_lock);
421 
422 	now = sched_clock();
423 
424 	collect_percpu_times(group, PSI_AVGS, &changed_states);
425 	nonidle = changed_states & (1 << PSI_NONIDLE);
426 	/*
427 	 * If there is task activity, periodically fold the per-cpu
428 	 * times and feed samples into the running averages. If things
429 	 * are idle and there is no data to process, stop the clock.
430 	 * Once restarted, we'll catch up the running averages in one
431 	 * go - see calc_avgs() and missed_periods.
432 	 */
433 	if (now >= group->avg_next_update)
434 		group->avg_next_update = update_averages(group, now);
435 
436 	if (nonidle) {
437 		schedule_delayed_work(dwork, nsecs_to_jiffies(
438 				group->avg_next_update - now) + 1);
439 	}
440 
441 	mutex_unlock(&group->avgs_lock);
442 }
443 
444 /* Trigger tracking window manupulations */
445 static void window_reset(struct psi_window *win, u64 now, u64 value,
446 			 u64 prev_growth)
447 {
448 	win->start_time = now;
449 	win->start_value = value;
450 	win->prev_growth = prev_growth;
451 }
452 
453 /*
454  * PSI growth tracking window update and growth calculation routine.
455  *
456  * This approximates a sliding tracking window by interpolating
457  * partially elapsed windows using historical growth data from the
458  * previous intervals. This minimizes memory requirements (by not storing
459  * all the intermediate values in the previous window) and simplifies
460  * the calculations. It works well because PSI signal changes only in
461  * positive direction and over relatively small window sizes the growth
462  * is close to linear.
463  */
464 static u64 window_update(struct psi_window *win, u64 now, u64 value)
465 {
466 	u64 elapsed;
467 	u64 growth;
468 
469 	elapsed = now - win->start_time;
470 	growth = value - win->start_value;
471 	/*
472 	 * After each tracking window passes win->start_value and
473 	 * win->start_time get reset and win->prev_growth stores
474 	 * the average per-window growth of the previous window.
475 	 * win->prev_growth is then used to interpolate additional
476 	 * growth from the previous window assuming it was linear.
477 	 */
478 	if (elapsed > win->size)
479 		window_reset(win, now, value, growth);
480 	else {
481 		u32 remaining;
482 
483 		remaining = win->size - elapsed;
484 		growth += div_u64(win->prev_growth * remaining, win->size);
485 	}
486 
487 	return growth;
488 }
489 
490 static void init_triggers(struct psi_group *group, u64 now)
491 {
492 	struct psi_trigger *t;
493 
494 	list_for_each_entry(t, &group->triggers, node)
495 		window_reset(&t->win, now,
496 				group->total[PSI_POLL][t->state], 0);
497 	memcpy(group->polling_total, group->total[PSI_POLL],
498 		   sizeof(group->polling_total));
499 	group->polling_next_update = now + group->poll_min_period;
500 }
501 
502 static u64 update_triggers(struct psi_group *group, u64 now)
503 {
504 	struct psi_trigger *t;
505 	bool new_stall = false;
506 	u64 *total = group->total[PSI_POLL];
507 
508 	/*
509 	 * On subsequent updates, calculate growth deltas and let
510 	 * watchers know when their specified thresholds are exceeded.
511 	 */
512 	list_for_each_entry(t, &group->triggers, node) {
513 		u64 growth;
514 
515 		/* Check for stall activity */
516 		if (group->polling_total[t->state] == total[t->state])
517 			continue;
518 
519 		/*
520 		 * Multiple triggers might be looking at the same state,
521 		 * remember to update group->polling_total[] once we've
522 		 * been through all of them. Also remember to extend the
523 		 * polling time if we see new stall activity.
524 		 */
525 		new_stall = true;
526 
527 		/* Calculate growth since last update */
528 		growth = window_update(&t->win, now, total[t->state]);
529 		if (growth < t->threshold)
530 			continue;
531 
532 		/* Limit event signaling to once per window */
533 		if (now < t->last_event_time + t->win.size)
534 			continue;
535 
536 		/* Generate an event */
537 		if (cmpxchg(&t->event, 0, 1) == 0)
538 			wake_up_interruptible(&t->event_wait);
539 		t->last_event_time = now;
540 	}
541 
542 	if (new_stall)
543 		memcpy(group->polling_total, total,
544 				sizeof(group->polling_total));
545 
546 	return now + group->poll_min_period;
547 }
548 
549 /*
550  * Schedule polling if it's not already scheduled. It's safe to call even from
551  * hotpath because even though kthread_queue_delayed_work takes worker->lock
552  * spinlock that spinlock is never contended due to poll_scheduled atomic
553  * preventing such competition.
554  */
555 static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay)
556 {
557 	struct kthread_worker *kworker;
558 
559 	/* Do not reschedule if already scheduled */
560 	if (atomic_cmpxchg(&group->poll_scheduled, 0, 1) != 0)
561 		return;
562 
563 	rcu_read_lock();
564 
565 	kworker = rcu_dereference(group->poll_kworker);
566 	/*
567 	 * kworker might be NULL in case psi_trigger_destroy races with
568 	 * psi_task_change (hotpath) which can't use locks
569 	 */
570 	if (likely(kworker))
571 		kthread_queue_delayed_work(kworker, &group->poll_work, delay);
572 	else
573 		atomic_set(&group->poll_scheduled, 0);
574 
575 	rcu_read_unlock();
576 }
577 
578 static void psi_poll_work(struct kthread_work *work)
579 {
580 	struct kthread_delayed_work *dwork;
581 	struct psi_group *group;
582 	u32 changed_states;
583 	u64 now;
584 
585 	dwork = container_of(work, struct kthread_delayed_work, work);
586 	group = container_of(dwork, struct psi_group, poll_work);
587 
588 	atomic_set(&group->poll_scheduled, 0);
589 
590 	mutex_lock(&group->trigger_lock);
591 
592 	now = sched_clock();
593 
594 	collect_percpu_times(group, PSI_POLL, &changed_states);
595 
596 	if (changed_states & group->poll_states) {
597 		/* Initialize trigger windows when entering polling mode */
598 		if (now > group->polling_until)
599 			init_triggers(group, now);
600 
601 		/*
602 		 * Keep the monitor active for at least the duration of the
603 		 * minimum tracking window as long as monitor states are
604 		 * changing.
605 		 */
606 		group->polling_until = now +
607 			group->poll_min_period * UPDATES_PER_WINDOW;
608 	}
609 
610 	if (now > group->polling_until) {
611 		group->polling_next_update = ULLONG_MAX;
612 		goto out;
613 	}
614 
615 	if (now >= group->polling_next_update)
616 		group->polling_next_update = update_triggers(group, now);
617 
618 	psi_schedule_poll_work(group,
619 		nsecs_to_jiffies(group->polling_next_update - now) + 1);
620 
621 out:
622 	mutex_unlock(&group->trigger_lock);
623 }
624 
625 static void record_times(struct psi_group_cpu *groupc, int cpu,
626 			 bool memstall_tick)
627 {
628 	u32 delta;
629 	u64 now;
630 
631 	now = cpu_clock(cpu);
632 	delta = now - groupc->state_start;
633 	groupc->state_start = now;
634 
635 	if (groupc->state_mask & (1 << PSI_IO_SOME)) {
636 		groupc->times[PSI_IO_SOME] += delta;
637 		if (groupc->state_mask & (1 << PSI_IO_FULL))
638 			groupc->times[PSI_IO_FULL] += delta;
639 	}
640 
641 	if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
642 		groupc->times[PSI_MEM_SOME] += delta;
643 		if (groupc->state_mask & (1 << PSI_MEM_FULL))
644 			groupc->times[PSI_MEM_FULL] += delta;
645 		else if (memstall_tick) {
646 			u32 sample;
647 			/*
648 			 * Since we care about lost potential, a
649 			 * memstall is FULL when there are no other
650 			 * working tasks, but also when the CPU is
651 			 * actively reclaiming and nothing productive
652 			 * could run even if it were runnable.
653 			 *
654 			 * When the timer tick sees a reclaiming CPU,
655 			 * regardless of runnable tasks, sample a FULL
656 			 * tick (or less if it hasn't been a full tick
657 			 * since the last state change).
658 			 */
659 			sample = min(delta, (u32)jiffies_to_nsecs(1));
660 			groupc->times[PSI_MEM_FULL] += sample;
661 		}
662 	}
663 
664 	if (groupc->state_mask & (1 << PSI_CPU_SOME))
665 		groupc->times[PSI_CPU_SOME] += delta;
666 
667 	if (groupc->state_mask & (1 << PSI_NONIDLE))
668 		groupc->times[PSI_NONIDLE] += delta;
669 }
670 
671 static u32 psi_group_change(struct psi_group *group, int cpu,
672 			    unsigned int clear, unsigned int set)
673 {
674 	struct psi_group_cpu *groupc;
675 	unsigned int t, m;
676 	enum psi_states s;
677 	u32 state_mask = 0;
678 
679 	groupc = per_cpu_ptr(group->pcpu, cpu);
680 
681 	/*
682 	 * First we assess the aggregate resource states this CPU's
683 	 * tasks have been in since the last change, and account any
684 	 * SOME and FULL time these may have resulted in.
685 	 *
686 	 * Then we update the task counts according to the state
687 	 * change requested through the @clear and @set bits.
688 	 */
689 	write_seqcount_begin(&groupc->seq);
690 
691 	record_times(groupc, cpu, false);
692 
693 	for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
694 		if (!(m & (1 << t)))
695 			continue;
696 		if (groupc->tasks[t] == 0 && !psi_bug) {
697 			printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u] clear=%x set=%x\n",
698 					cpu, t, groupc->tasks[0],
699 					groupc->tasks[1], groupc->tasks[2],
700 					clear, set);
701 			psi_bug = 1;
702 		}
703 		groupc->tasks[t]--;
704 	}
705 
706 	for (t = 0; set; set &= ~(1 << t), t++)
707 		if (set & (1 << t))
708 			groupc->tasks[t]++;
709 
710 	/* Calculate state mask representing active states */
711 	for (s = 0; s < NR_PSI_STATES; s++) {
712 		if (test_state(groupc->tasks, s))
713 			state_mask |= (1 << s);
714 	}
715 	groupc->state_mask = state_mask;
716 
717 	write_seqcount_end(&groupc->seq);
718 
719 	return state_mask;
720 }
721 
722 static struct psi_group *iterate_groups(struct task_struct *task, void **iter)
723 {
724 #ifdef CONFIG_CGROUPS
725 	struct cgroup *cgroup = NULL;
726 
727 	if (!*iter)
728 		cgroup = task->cgroups->dfl_cgrp;
729 	else if (*iter == &psi_system)
730 		return NULL;
731 	else
732 		cgroup = cgroup_parent(*iter);
733 
734 	if (cgroup && cgroup_parent(cgroup)) {
735 		*iter = cgroup;
736 		return cgroup_psi(cgroup);
737 	}
738 #else
739 	if (*iter)
740 		return NULL;
741 #endif
742 	*iter = &psi_system;
743 	return &psi_system;
744 }
745 
746 void psi_task_change(struct task_struct *task, int clear, int set)
747 {
748 	int cpu = task_cpu(task);
749 	struct psi_group *group;
750 	bool wake_clock = true;
751 	void *iter = NULL;
752 
753 	if (!task->pid)
754 		return;
755 
756 	if (((task->psi_flags & set) ||
757 	     (task->psi_flags & clear) != clear) &&
758 	    !psi_bug) {
759 		printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
760 				task->pid, task->comm, cpu,
761 				task->psi_flags, clear, set);
762 		psi_bug = 1;
763 	}
764 
765 	task->psi_flags &= ~clear;
766 	task->psi_flags |= set;
767 
768 	/*
769 	 * Periodic aggregation shuts off if there is a period of no
770 	 * task changes, so we wake it back up if necessary. However,
771 	 * don't do this if the task change is the aggregation worker
772 	 * itself going to sleep, or we'll ping-pong forever.
773 	 */
774 	if (unlikely((clear & TSK_RUNNING) &&
775 		     (task->flags & PF_WQ_WORKER) &&
776 		     wq_worker_last_func(task) == psi_avgs_work))
777 		wake_clock = false;
778 
779 	while ((group = iterate_groups(task, &iter))) {
780 		u32 state_mask = psi_group_change(group, cpu, clear, set);
781 
782 		if (state_mask & group->poll_states)
783 			psi_schedule_poll_work(group, 1);
784 
785 		if (wake_clock && !delayed_work_pending(&group->avgs_work))
786 			schedule_delayed_work(&group->avgs_work, PSI_FREQ);
787 	}
788 }
789 
790 void psi_memstall_tick(struct task_struct *task, int cpu)
791 {
792 	struct psi_group *group;
793 	void *iter = NULL;
794 
795 	while ((group = iterate_groups(task, &iter))) {
796 		struct psi_group_cpu *groupc;
797 
798 		groupc = per_cpu_ptr(group->pcpu, cpu);
799 		write_seqcount_begin(&groupc->seq);
800 		record_times(groupc, cpu, true);
801 		write_seqcount_end(&groupc->seq);
802 	}
803 }
804 
805 /**
806  * psi_memstall_enter - mark the beginning of a memory stall section
807  * @flags: flags to handle nested sections
808  *
809  * Marks the calling task as being stalled due to a lack of memory,
810  * such as waiting for a refault or performing reclaim.
811  */
812 void psi_memstall_enter(unsigned long *flags)
813 {
814 	struct rq_flags rf;
815 	struct rq *rq;
816 
817 	if (static_branch_likely(&psi_disabled))
818 		return;
819 
820 	*flags = current->flags & PF_MEMSTALL;
821 	if (*flags)
822 		return;
823 	/*
824 	 * PF_MEMSTALL setting & accounting needs to be atomic wrt
825 	 * changes to the task's scheduling state, otherwise we can
826 	 * race with CPU migration.
827 	 */
828 	rq = this_rq_lock_irq(&rf);
829 
830 	current->flags |= PF_MEMSTALL;
831 	psi_task_change(current, 0, TSK_MEMSTALL);
832 
833 	rq_unlock_irq(rq, &rf);
834 }
835 
836 /**
837  * psi_memstall_leave - mark the end of an memory stall section
838  * @flags: flags to handle nested memdelay sections
839  *
840  * Marks the calling task as no longer stalled due to lack of memory.
841  */
842 void psi_memstall_leave(unsigned long *flags)
843 {
844 	struct rq_flags rf;
845 	struct rq *rq;
846 
847 	if (static_branch_likely(&psi_disabled))
848 		return;
849 
850 	if (*flags)
851 		return;
852 	/*
853 	 * PF_MEMSTALL clearing & accounting needs to be atomic wrt
854 	 * changes to the task's scheduling state, otherwise we could
855 	 * race with CPU migration.
856 	 */
857 	rq = this_rq_lock_irq(&rf);
858 
859 	current->flags &= ~PF_MEMSTALL;
860 	psi_task_change(current, TSK_MEMSTALL, 0);
861 
862 	rq_unlock_irq(rq, &rf);
863 }
864 
865 #ifdef CONFIG_CGROUPS
866 int psi_cgroup_alloc(struct cgroup *cgroup)
867 {
868 	if (static_branch_likely(&psi_disabled))
869 		return 0;
870 
871 	cgroup->psi.pcpu = alloc_percpu(struct psi_group_cpu);
872 	if (!cgroup->psi.pcpu)
873 		return -ENOMEM;
874 	group_init(&cgroup->psi);
875 	return 0;
876 }
877 
878 void psi_cgroup_free(struct cgroup *cgroup)
879 {
880 	if (static_branch_likely(&psi_disabled))
881 		return;
882 
883 	cancel_delayed_work_sync(&cgroup->psi.avgs_work);
884 	free_percpu(cgroup->psi.pcpu);
885 	/* All triggers must be removed by now */
886 	WARN_ONCE(cgroup->psi.poll_states, "psi: trigger leak\n");
887 }
888 
889 /**
890  * cgroup_move_task - move task to a different cgroup
891  * @task: the task
892  * @to: the target css_set
893  *
894  * Move task to a new cgroup and safely migrate its associated stall
895  * state between the different groups.
896  *
897  * This function acquires the task's rq lock to lock out concurrent
898  * changes to the task's scheduling state and - in case the task is
899  * running - concurrent changes to its stall state.
900  */
901 void cgroup_move_task(struct task_struct *task, struct css_set *to)
902 {
903 	unsigned int task_flags = 0;
904 	struct rq_flags rf;
905 	struct rq *rq;
906 
907 	if (static_branch_likely(&psi_disabled)) {
908 		/*
909 		 * Lame to do this here, but the scheduler cannot be locked
910 		 * from the outside, so we move cgroups from inside sched/.
911 		 */
912 		rcu_assign_pointer(task->cgroups, to);
913 		return;
914 	}
915 
916 	rq = task_rq_lock(task, &rf);
917 
918 	if (task_on_rq_queued(task))
919 		task_flags = TSK_RUNNING;
920 	else if (task->in_iowait)
921 		task_flags = TSK_IOWAIT;
922 
923 	if (task->flags & PF_MEMSTALL)
924 		task_flags |= TSK_MEMSTALL;
925 
926 	if (task_flags)
927 		psi_task_change(task, task_flags, 0);
928 
929 	/* See comment above */
930 	rcu_assign_pointer(task->cgroups, to);
931 
932 	if (task_flags)
933 		psi_task_change(task, 0, task_flags);
934 
935 	task_rq_unlock(rq, task, &rf);
936 }
937 #endif /* CONFIG_CGROUPS */
938 
939 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
940 {
941 	int full;
942 	u64 now;
943 
944 	if (static_branch_likely(&psi_disabled))
945 		return -EOPNOTSUPP;
946 
947 	/* Update averages before reporting them */
948 	mutex_lock(&group->avgs_lock);
949 	now = sched_clock();
950 	collect_percpu_times(group, PSI_AVGS, NULL);
951 	if (now >= group->avg_next_update)
952 		group->avg_next_update = update_averages(group, now);
953 	mutex_unlock(&group->avgs_lock);
954 
955 	for (full = 0; full < 2 - (res == PSI_CPU); full++) {
956 		unsigned long avg[3];
957 		u64 total;
958 		int w;
959 
960 		for (w = 0; w < 3; w++)
961 			avg[w] = group->avg[res * 2 + full][w];
962 		total = div_u64(group->total[PSI_AVGS][res * 2 + full],
963 				NSEC_PER_USEC);
964 
965 		seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
966 			   full ? "full" : "some",
967 			   LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
968 			   LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
969 			   LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
970 			   total);
971 	}
972 
973 	return 0;
974 }
975 
976 static int psi_io_show(struct seq_file *m, void *v)
977 {
978 	return psi_show(m, &psi_system, PSI_IO);
979 }
980 
981 static int psi_memory_show(struct seq_file *m, void *v)
982 {
983 	return psi_show(m, &psi_system, PSI_MEM);
984 }
985 
986 static int psi_cpu_show(struct seq_file *m, void *v)
987 {
988 	return psi_show(m, &psi_system, PSI_CPU);
989 }
990 
991 static int psi_io_open(struct inode *inode, struct file *file)
992 {
993 	return single_open(file, psi_io_show, NULL);
994 }
995 
996 static int psi_memory_open(struct inode *inode, struct file *file)
997 {
998 	return single_open(file, psi_memory_show, NULL);
999 }
1000 
1001 static int psi_cpu_open(struct inode *inode, struct file *file)
1002 {
1003 	return single_open(file, psi_cpu_show, NULL);
1004 }
1005 
1006 struct psi_trigger *psi_trigger_create(struct psi_group *group,
1007 			char *buf, size_t nbytes, enum psi_res res)
1008 {
1009 	struct psi_trigger *t;
1010 	enum psi_states state;
1011 	u32 threshold_us;
1012 	u32 window_us;
1013 
1014 	if (static_branch_likely(&psi_disabled))
1015 		return ERR_PTR(-EOPNOTSUPP);
1016 
1017 	if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1018 		state = PSI_IO_SOME + res * 2;
1019 	else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1020 		state = PSI_IO_FULL + res * 2;
1021 	else
1022 		return ERR_PTR(-EINVAL);
1023 
1024 	if (state >= PSI_NONIDLE)
1025 		return ERR_PTR(-EINVAL);
1026 
1027 	if (window_us < WINDOW_MIN_US ||
1028 		window_us > WINDOW_MAX_US)
1029 		return ERR_PTR(-EINVAL);
1030 
1031 	/* Check threshold */
1032 	if (threshold_us == 0 || threshold_us > window_us)
1033 		return ERR_PTR(-EINVAL);
1034 
1035 	t = kmalloc(sizeof(*t), GFP_KERNEL);
1036 	if (!t)
1037 		return ERR_PTR(-ENOMEM);
1038 
1039 	t->group = group;
1040 	t->state = state;
1041 	t->threshold = threshold_us * NSEC_PER_USEC;
1042 	t->win.size = window_us * NSEC_PER_USEC;
1043 	window_reset(&t->win, 0, 0, 0);
1044 
1045 	t->event = 0;
1046 	t->last_event_time = 0;
1047 	init_waitqueue_head(&t->event_wait);
1048 	kref_init(&t->refcount);
1049 
1050 	mutex_lock(&group->trigger_lock);
1051 
1052 	if (!rcu_access_pointer(group->poll_kworker)) {
1053 		struct sched_param param = {
1054 			.sched_priority = 1,
1055 		};
1056 		struct kthread_worker *kworker;
1057 
1058 		kworker = kthread_create_worker(0, "psimon");
1059 		if (IS_ERR(kworker)) {
1060 			kfree(t);
1061 			mutex_unlock(&group->trigger_lock);
1062 			return ERR_CAST(kworker);
1063 		}
1064 		sched_setscheduler_nocheck(kworker->task, SCHED_FIFO, &param);
1065 		kthread_init_delayed_work(&group->poll_work,
1066 				psi_poll_work);
1067 		rcu_assign_pointer(group->poll_kworker, kworker);
1068 	}
1069 
1070 	list_add(&t->node, &group->triggers);
1071 	group->poll_min_period = min(group->poll_min_period,
1072 		div_u64(t->win.size, UPDATES_PER_WINDOW));
1073 	group->nr_triggers[t->state]++;
1074 	group->poll_states |= (1 << t->state);
1075 
1076 	mutex_unlock(&group->trigger_lock);
1077 
1078 	return t;
1079 }
1080 
1081 static void psi_trigger_destroy(struct kref *ref)
1082 {
1083 	struct psi_trigger *t = container_of(ref, struct psi_trigger, refcount);
1084 	struct psi_group *group = t->group;
1085 	struct kthread_worker *kworker_to_destroy = NULL;
1086 
1087 	if (static_branch_likely(&psi_disabled))
1088 		return;
1089 
1090 	/*
1091 	 * Wakeup waiters to stop polling. Can happen if cgroup is deleted
1092 	 * from under a polling process.
1093 	 */
1094 	wake_up_interruptible(&t->event_wait);
1095 
1096 	mutex_lock(&group->trigger_lock);
1097 
1098 	if (!list_empty(&t->node)) {
1099 		struct psi_trigger *tmp;
1100 		u64 period = ULLONG_MAX;
1101 
1102 		list_del(&t->node);
1103 		group->nr_triggers[t->state]--;
1104 		if (!group->nr_triggers[t->state])
1105 			group->poll_states &= ~(1 << t->state);
1106 		/* reset min update period for the remaining triggers */
1107 		list_for_each_entry(tmp, &group->triggers, node)
1108 			period = min(period, div_u64(tmp->win.size,
1109 					UPDATES_PER_WINDOW));
1110 		group->poll_min_period = period;
1111 		/* Destroy poll_kworker when the last trigger is destroyed */
1112 		if (group->poll_states == 0) {
1113 			group->polling_until = 0;
1114 			kworker_to_destroy = rcu_dereference_protected(
1115 					group->poll_kworker,
1116 					lockdep_is_held(&group->trigger_lock));
1117 			rcu_assign_pointer(group->poll_kworker, NULL);
1118 		}
1119 	}
1120 
1121 	mutex_unlock(&group->trigger_lock);
1122 
1123 	/*
1124 	 * Wait for both *trigger_ptr from psi_trigger_replace and
1125 	 * poll_kworker RCUs to complete their read-side critical sections
1126 	 * before destroying the trigger and optionally the poll_kworker
1127 	 */
1128 	synchronize_rcu();
1129 	/*
1130 	 * Destroy the kworker after releasing trigger_lock to prevent a
1131 	 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1132 	 */
1133 	if (kworker_to_destroy) {
1134 		kthread_cancel_delayed_work_sync(&group->poll_work);
1135 		kthread_destroy_worker(kworker_to_destroy);
1136 	}
1137 	kfree(t);
1138 }
1139 
1140 void psi_trigger_replace(void **trigger_ptr, struct psi_trigger *new)
1141 {
1142 	struct psi_trigger *old = *trigger_ptr;
1143 
1144 	if (static_branch_likely(&psi_disabled))
1145 		return;
1146 
1147 	rcu_assign_pointer(*trigger_ptr, new);
1148 	if (old)
1149 		kref_put(&old->refcount, psi_trigger_destroy);
1150 }
1151 
1152 __poll_t psi_trigger_poll(void **trigger_ptr,
1153 				struct file *file, poll_table *wait)
1154 {
1155 	__poll_t ret = DEFAULT_POLLMASK;
1156 	struct psi_trigger *t;
1157 
1158 	if (static_branch_likely(&psi_disabled))
1159 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1160 
1161 	rcu_read_lock();
1162 
1163 	t = rcu_dereference(*(void __rcu __force **)trigger_ptr);
1164 	if (!t) {
1165 		rcu_read_unlock();
1166 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1167 	}
1168 	kref_get(&t->refcount);
1169 
1170 	rcu_read_unlock();
1171 
1172 	poll_wait(file, &t->event_wait, wait);
1173 
1174 	if (cmpxchg(&t->event, 1, 0) == 1)
1175 		ret |= EPOLLPRI;
1176 
1177 	kref_put(&t->refcount, psi_trigger_destroy);
1178 
1179 	return ret;
1180 }
1181 
1182 static ssize_t psi_write(struct file *file, const char __user *user_buf,
1183 			 size_t nbytes, enum psi_res res)
1184 {
1185 	char buf[32];
1186 	size_t buf_size;
1187 	struct seq_file *seq;
1188 	struct psi_trigger *new;
1189 
1190 	if (static_branch_likely(&psi_disabled))
1191 		return -EOPNOTSUPP;
1192 
1193 	buf_size = min(nbytes, (sizeof(buf) - 1));
1194 	if (copy_from_user(buf, user_buf, buf_size))
1195 		return -EFAULT;
1196 
1197 	buf[buf_size - 1] = '\0';
1198 
1199 	new = psi_trigger_create(&psi_system, buf, nbytes, res);
1200 	if (IS_ERR(new))
1201 		return PTR_ERR(new);
1202 
1203 	seq = file->private_data;
1204 	/* Take seq->lock to protect seq->private from concurrent writes */
1205 	mutex_lock(&seq->lock);
1206 	psi_trigger_replace(&seq->private, new);
1207 	mutex_unlock(&seq->lock);
1208 
1209 	return nbytes;
1210 }
1211 
1212 static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1213 			    size_t nbytes, loff_t *ppos)
1214 {
1215 	return psi_write(file, user_buf, nbytes, PSI_IO);
1216 }
1217 
1218 static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1219 				size_t nbytes, loff_t *ppos)
1220 {
1221 	return psi_write(file, user_buf, nbytes, PSI_MEM);
1222 }
1223 
1224 static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1225 			     size_t nbytes, loff_t *ppos)
1226 {
1227 	return psi_write(file, user_buf, nbytes, PSI_CPU);
1228 }
1229 
1230 static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1231 {
1232 	struct seq_file *seq = file->private_data;
1233 
1234 	return psi_trigger_poll(&seq->private, file, wait);
1235 }
1236 
1237 static int psi_fop_release(struct inode *inode, struct file *file)
1238 {
1239 	struct seq_file *seq = file->private_data;
1240 
1241 	psi_trigger_replace(&seq->private, NULL);
1242 	return single_release(inode, file);
1243 }
1244 
1245 static const struct file_operations psi_io_fops = {
1246 	.open           = psi_io_open,
1247 	.read           = seq_read,
1248 	.llseek         = seq_lseek,
1249 	.write          = psi_io_write,
1250 	.poll           = psi_fop_poll,
1251 	.release        = psi_fop_release,
1252 };
1253 
1254 static const struct file_operations psi_memory_fops = {
1255 	.open           = psi_memory_open,
1256 	.read           = seq_read,
1257 	.llseek         = seq_lseek,
1258 	.write          = psi_memory_write,
1259 	.poll           = psi_fop_poll,
1260 	.release        = psi_fop_release,
1261 };
1262 
1263 static const struct file_operations psi_cpu_fops = {
1264 	.open           = psi_cpu_open,
1265 	.read           = seq_read,
1266 	.llseek         = seq_lseek,
1267 	.write          = psi_cpu_write,
1268 	.poll           = psi_fop_poll,
1269 	.release        = psi_fop_release,
1270 };
1271 
1272 static int __init psi_proc_init(void)
1273 {
1274 	proc_mkdir("pressure", NULL);
1275 	proc_create("pressure/io", 0, NULL, &psi_io_fops);
1276 	proc_create("pressure/memory", 0, NULL, &psi_memory_fops);
1277 	proc_create("pressure/cpu", 0, NULL, &psi_cpu_fops);
1278 	return 0;
1279 }
1280 module_init(psi_proc_init);
1281