xref: /openbmc/linux/kernel/sched/psi.c (revision f519cd13)
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_last_update = sched_clock();
189 	group->avg_next_update = group->avg_last_update + psi_period;
190 	INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
191 	mutex_init(&group->avgs_lock);
192 	/* Init trigger-related members */
193 	atomic_set(&group->poll_scheduled, 0);
194 	mutex_init(&group->trigger_lock);
195 	INIT_LIST_HEAD(&group->triggers);
196 	memset(group->nr_triggers, 0, sizeof(group->nr_triggers));
197 	group->poll_states = 0;
198 	group->poll_min_period = U32_MAX;
199 	memset(group->polling_total, 0, sizeof(group->polling_total));
200 	group->polling_next_update = ULLONG_MAX;
201 	group->polling_until = 0;
202 	rcu_assign_pointer(group->poll_kworker, NULL);
203 }
204 
205 void __init psi_init(void)
206 {
207 	if (!psi_enable) {
208 		static_branch_enable(&psi_disabled);
209 		return;
210 	}
211 
212 	psi_period = jiffies_to_nsecs(PSI_FREQ);
213 	group_init(&psi_system);
214 }
215 
216 static bool test_state(unsigned int *tasks, enum psi_states state)
217 {
218 	switch (state) {
219 	case PSI_IO_SOME:
220 		return tasks[NR_IOWAIT];
221 	case PSI_IO_FULL:
222 		return tasks[NR_IOWAIT] && !tasks[NR_RUNNING];
223 	case PSI_MEM_SOME:
224 		return tasks[NR_MEMSTALL];
225 	case PSI_MEM_FULL:
226 		return tasks[NR_MEMSTALL] && !tasks[NR_RUNNING];
227 	case PSI_CPU_SOME:
228 		return tasks[NR_RUNNING] > 1;
229 	case PSI_NONIDLE:
230 		return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
231 			tasks[NR_RUNNING];
232 	default:
233 		return false;
234 	}
235 }
236 
237 static void get_recent_times(struct psi_group *group, int cpu,
238 			     enum psi_aggregators aggregator, u32 *times,
239 			     u32 *pchanged_states)
240 {
241 	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
242 	u64 now, state_start;
243 	enum psi_states s;
244 	unsigned int seq;
245 	u32 state_mask;
246 
247 	*pchanged_states = 0;
248 
249 	/* Snapshot a coherent view of the CPU state */
250 	do {
251 		seq = read_seqcount_begin(&groupc->seq);
252 		now = cpu_clock(cpu);
253 		memcpy(times, groupc->times, sizeof(groupc->times));
254 		state_mask = groupc->state_mask;
255 		state_start = groupc->state_start;
256 	} while (read_seqcount_retry(&groupc->seq, seq));
257 
258 	/* Calculate state time deltas against the previous snapshot */
259 	for (s = 0; s < NR_PSI_STATES; s++) {
260 		u32 delta;
261 		/*
262 		 * In addition to already concluded states, we also
263 		 * incorporate currently active states on the CPU,
264 		 * since states may last for many sampling periods.
265 		 *
266 		 * This way we keep our delta sampling buckets small
267 		 * (u32) and our reported pressure close to what's
268 		 * actually happening.
269 		 */
270 		if (state_mask & (1 << s))
271 			times[s] += now - state_start;
272 
273 		delta = times[s] - groupc->times_prev[aggregator][s];
274 		groupc->times_prev[aggregator][s] = times[s];
275 
276 		times[s] = delta;
277 		if (delta)
278 			*pchanged_states |= (1 << s);
279 	}
280 }
281 
282 static void calc_avgs(unsigned long avg[3], int missed_periods,
283 		      u64 time, u64 period)
284 {
285 	unsigned long pct;
286 
287 	/* Fill in zeroes for periods of no activity */
288 	if (missed_periods) {
289 		avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
290 		avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
291 		avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
292 	}
293 
294 	/* Sample the most recent active period */
295 	pct = div_u64(time * 100, period);
296 	pct *= FIXED_1;
297 	avg[0] = calc_load(avg[0], EXP_10s, pct);
298 	avg[1] = calc_load(avg[1], EXP_60s, pct);
299 	avg[2] = calc_load(avg[2], EXP_300s, pct);
300 }
301 
302 static void collect_percpu_times(struct psi_group *group,
303 				 enum psi_aggregators aggregator,
304 				 u32 *pchanged_states)
305 {
306 	u64 deltas[NR_PSI_STATES - 1] = { 0, };
307 	unsigned long nonidle_total = 0;
308 	u32 changed_states = 0;
309 	int cpu;
310 	int s;
311 
312 	/*
313 	 * Collect the per-cpu time buckets and average them into a
314 	 * single time sample that is normalized to wallclock time.
315 	 *
316 	 * For averaging, each CPU is weighted by its non-idle time in
317 	 * the sampling period. This eliminates artifacts from uneven
318 	 * loading, or even entirely idle CPUs.
319 	 */
320 	for_each_possible_cpu(cpu) {
321 		u32 times[NR_PSI_STATES];
322 		u32 nonidle;
323 		u32 cpu_changed_states;
324 
325 		get_recent_times(group, cpu, aggregator, times,
326 				&cpu_changed_states);
327 		changed_states |= cpu_changed_states;
328 
329 		nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
330 		nonidle_total += nonidle;
331 
332 		for (s = 0; s < PSI_NONIDLE; s++)
333 			deltas[s] += (u64)times[s] * nonidle;
334 	}
335 
336 	/*
337 	 * Integrate the sample into the running statistics that are
338 	 * reported to userspace: the cumulative stall times and the
339 	 * decaying averages.
340 	 *
341 	 * Pressure percentages are sampled at PSI_FREQ. We might be
342 	 * called more often when the user polls more frequently than
343 	 * that; we might be called less often when there is no task
344 	 * activity, thus no data, and clock ticks are sporadic. The
345 	 * below handles both.
346 	 */
347 
348 	/* total= */
349 	for (s = 0; s < NR_PSI_STATES - 1; s++)
350 		group->total[aggregator][s] +=
351 				div_u64(deltas[s], max(nonidle_total, 1UL));
352 
353 	if (pchanged_states)
354 		*pchanged_states = changed_states;
355 }
356 
357 static u64 update_averages(struct psi_group *group, u64 now)
358 {
359 	unsigned long missed_periods = 0;
360 	u64 expires, period;
361 	u64 avg_next_update;
362 	int s;
363 
364 	/* avgX= */
365 	expires = group->avg_next_update;
366 	if (now - expires >= psi_period)
367 		missed_periods = div_u64(now - expires, psi_period);
368 
369 	/*
370 	 * The periodic clock tick can get delayed for various
371 	 * reasons, especially on loaded systems. To avoid clock
372 	 * drift, we schedule the clock in fixed psi_period intervals.
373 	 * But the deltas we sample out of the per-cpu buckets above
374 	 * are based on the actual time elapsing between clock ticks.
375 	 */
376 	avg_next_update = expires + ((1 + missed_periods) * psi_period);
377 	period = now - (group->avg_last_update + (missed_periods * psi_period));
378 	group->avg_last_update = now;
379 
380 	for (s = 0; s < NR_PSI_STATES - 1; s++) {
381 		u32 sample;
382 
383 		sample = group->total[PSI_AVGS][s] - group->avg_total[s];
384 		/*
385 		 * Due to the lockless sampling of the time buckets,
386 		 * recorded time deltas can slip into the next period,
387 		 * which under full pressure can result in samples in
388 		 * excess of the period length.
389 		 *
390 		 * We don't want to report non-sensical pressures in
391 		 * excess of 100%, nor do we want to drop such events
392 		 * on the floor. Instead we punt any overage into the
393 		 * future until pressure subsides. By doing this we
394 		 * don't underreport the occurring pressure curve, we
395 		 * just report it delayed by one period length.
396 		 *
397 		 * The error isn't cumulative. As soon as another
398 		 * delta slips from a period P to P+1, by definition
399 		 * it frees up its time T in P.
400 		 */
401 		if (sample > period)
402 			sample = period;
403 		group->avg_total[s] += sample;
404 		calc_avgs(group->avg[s], missed_periods, sample, period);
405 	}
406 
407 	return avg_next_update;
408 }
409 
410 static void psi_avgs_work(struct work_struct *work)
411 {
412 	struct delayed_work *dwork;
413 	struct psi_group *group;
414 	u32 changed_states;
415 	bool nonidle;
416 	u64 now;
417 
418 	dwork = to_delayed_work(work);
419 	group = container_of(dwork, struct psi_group, avgs_work);
420 
421 	mutex_lock(&group->avgs_lock);
422 
423 	now = sched_clock();
424 
425 	collect_percpu_times(group, PSI_AVGS, &changed_states);
426 	nonidle = changed_states & (1 << PSI_NONIDLE);
427 	/*
428 	 * If there is task activity, periodically fold the per-cpu
429 	 * times and feed samples into the running averages. If things
430 	 * are idle and there is no data to process, stop the clock.
431 	 * Once restarted, we'll catch up the running averages in one
432 	 * go - see calc_avgs() and missed_periods.
433 	 */
434 	if (now >= group->avg_next_update)
435 		group->avg_next_update = update_averages(group, now);
436 
437 	if (nonidle) {
438 		schedule_delayed_work(dwork, nsecs_to_jiffies(
439 				group->avg_next_update - now) + 1);
440 	}
441 
442 	mutex_unlock(&group->avgs_lock);
443 }
444 
445 /* Trigger tracking window manupulations */
446 static void window_reset(struct psi_window *win, u64 now, u64 value,
447 			 u64 prev_growth)
448 {
449 	win->start_time = now;
450 	win->start_value = value;
451 	win->prev_growth = prev_growth;
452 }
453 
454 /*
455  * PSI growth tracking window update and growth calculation routine.
456  *
457  * This approximates a sliding tracking window by interpolating
458  * partially elapsed windows using historical growth data from the
459  * previous intervals. This minimizes memory requirements (by not storing
460  * all the intermediate values in the previous window) and simplifies
461  * the calculations. It works well because PSI signal changes only in
462  * positive direction and over relatively small window sizes the growth
463  * is close to linear.
464  */
465 static u64 window_update(struct psi_window *win, u64 now, u64 value)
466 {
467 	u64 elapsed;
468 	u64 growth;
469 
470 	elapsed = now - win->start_time;
471 	growth = value - win->start_value;
472 	/*
473 	 * After each tracking window passes win->start_value and
474 	 * win->start_time get reset and win->prev_growth stores
475 	 * the average per-window growth of the previous window.
476 	 * win->prev_growth is then used to interpolate additional
477 	 * growth from the previous window assuming it was linear.
478 	 */
479 	if (elapsed > win->size)
480 		window_reset(win, now, value, growth);
481 	else {
482 		u32 remaining;
483 
484 		remaining = win->size - elapsed;
485 		growth += div64_u64(win->prev_growth * remaining, win->size);
486 	}
487 
488 	return growth;
489 }
490 
491 static void init_triggers(struct psi_group *group, u64 now)
492 {
493 	struct psi_trigger *t;
494 
495 	list_for_each_entry(t, &group->triggers, node)
496 		window_reset(&t->win, now,
497 				group->total[PSI_POLL][t->state], 0);
498 	memcpy(group->polling_total, group->total[PSI_POLL],
499 		   sizeof(group->polling_total));
500 	group->polling_next_update = now + group->poll_min_period;
501 }
502 
503 static u64 update_triggers(struct psi_group *group, u64 now)
504 {
505 	struct psi_trigger *t;
506 	bool new_stall = false;
507 	u64 *total = group->total[PSI_POLL];
508 
509 	/*
510 	 * On subsequent updates, calculate growth deltas and let
511 	 * watchers know when their specified thresholds are exceeded.
512 	 */
513 	list_for_each_entry(t, &group->triggers, node) {
514 		u64 growth;
515 
516 		/* Check for stall activity */
517 		if (group->polling_total[t->state] == total[t->state])
518 			continue;
519 
520 		/*
521 		 * Multiple triggers might be looking at the same state,
522 		 * remember to update group->polling_total[] once we've
523 		 * been through all of them. Also remember to extend the
524 		 * polling time if we see new stall activity.
525 		 */
526 		new_stall = true;
527 
528 		/* Calculate growth since last update */
529 		growth = window_update(&t->win, now, total[t->state]);
530 		if (growth < t->threshold)
531 			continue;
532 
533 		/* Limit event signaling to once per window */
534 		if (now < t->last_event_time + t->win.size)
535 			continue;
536 
537 		/* Generate an event */
538 		if (cmpxchg(&t->event, 0, 1) == 0)
539 			wake_up_interruptible(&t->event_wait);
540 		t->last_event_time = now;
541 	}
542 
543 	if (new_stall)
544 		memcpy(group->polling_total, total,
545 				sizeof(group->polling_total));
546 
547 	return now + group->poll_min_period;
548 }
549 
550 /*
551  * Schedule polling if it's not already scheduled. It's safe to call even from
552  * hotpath because even though kthread_queue_delayed_work takes worker->lock
553  * spinlock that spinlock is never contended due to poll_scheduled atomic
554  * preventing such competition.
555  */
556 static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay)
557 {
558 	struct kthread_worker *kworker;
559 
560 	/* Do not reschedule if already scheduled */
561 	if (atomic_cmpxchg(&group->poll_scheduled, 0, 1) != 0)
562 		return;
563 
564 	rcu_read_lock();
565 
566 	kworker = rcu_dereference(group->poll_kworker);
567 	/*
568 	 * kworker might be NULL in case psi_trigger_destroy races with
569 	 * psi_task_change (hotpath) which can't use locks
570 	 */
571 	if (likely(kworker))
572 		kthread_queue_delayed_work(kworker, &group->poll_work, delay);
573 	else
574 		atomic_set(&group->poll_scheduled, 0);
575 
576 	rcu_read_unlock();
577 }
578 
579 static void psi_poll_work(struct kthread_work *work)
580 {
581 	struct kthread_delayed_work *dwork;
582 	struct psi_group *group;
583 	u32 changed_states;
584 	u64 now;
585 
586 	dwork = container_of(work, struct kthread_delayed_work, work);
587 	group = container_of(dwork, struct psi_group, poll_work);
588 
589 	atomic_set(&group->poll_scheduled, 0);
590 
591 	mutex_lock(&group->trigger_lock);
592 
593 	now = sched_clock();
594 
595 	collect_percpu_times(group, PSI_POLL, &changed_states);
596 
597 	if (changed_states & group->poll_states) {
598 		/* Initialize trigger windows when entering polling mode */
599 		if (now > group->polling_until)
600 			init_triggers(group, now);
601 
602 		/*
603 		 * Keep the monitor active for at least the duration of the
604 		 * minimum tracking window as long as monitor states are
605 		 * changing.
606 		 */
607 		group->polling_until = now +
608 			group->poll_min_period * UPDATES_PER_WINDOW;
609 	}
610 
611 	if (now > group->polling_until) {
612 		group->polling_next_update = ULLONG_MAX;
613 		goto out;
614 	}
615 
616 	if (now >= group->polling_next_update)
617 		group->polling_next_update = update_triggers(group, now);
618 
619 	psi_schedule_poll_work(group,
620 		nsecs_to_jiffies(group->polling_next_update - now) + 1);
621 
622 out:
623 	mutex_unlock(&group->trigger_lock);
624 }
625 
626 static void record_times(struct psi_group_cpu *groupc, int cpu,
627 			 bool memstall_tick)
628 {
629 	u32 delta;
630 	u64 now;
631 
632 	now = cpu_clock(cpu);
633 	delta = now - groupc->state_start;
634 	groupc->state_start = now;
635 
636 	if (groupc->state_mask & (1 << PSI_IO_SOME)) {
637 		groupc->times[PSI_IO_SOME] += delta;
638 		if (groupc->state_mask & (1 << PSI_IO_FULL))
639 			groupc->times[PSI_IO_FULL] += delta;
640 	}
641 
642 	if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
643 		groupc->times[PSI_MEM_SOME] += delta;
644 		if (groupc->state_mask & (1 << PSI_MEM_FULL))
645 			groupc->times[PSI_MEM_FULL] += delta;
646 		else if (memstall_tick) {
647 			u32 sample;
648 			/*
649 			 * Since we care about lost potential, a
650 			 * memstall is FULL when there are no other
651 			 * working tasks, but also when the CPU is
652 			 * actively reclaiming and nothing productive
653 			 * could run even if it were runnable.
654 			 *
655 			 * When the timer tick sees a reclaiming CPU,
656 			 * regardless of runnable tasks, sample a FULL
657 			 * tick (or less if it hasn't been a full tick
658 			 * since the last state change).
659 			 */
660 			sample = min(delta, (u32)jiffies_to_nsecs(1));
661 			groupc->times[PSI_MEM_FULL] += sample;
662 		}
663 	}
664 
665 	if (groupc->state_mask & (1 << PSI_CPU_SOME))
666 		groupc->times[PSI_CPU_SOME] += delta;
667 
668 	if (groupc->state_mask & (1 << PSI_NONIDLE))
669 		groupc->times[PSI_NONIDLE] += delta;
670 }
671 
672 static u32 psi_group_change(struct psi_group *group, int cpu,
673 			    unsigned int clear, unsigned int set)
674 {
675 	struct psi_group_cpu *groupc;
676 	unsigned int t, m;
677 	enum psi_states s;
678 	u32 state_mask = 0;
679 
680 	groupc = per_cpu_ptr(group->pcpu, cpu);
681 
682 	/*
683 	 * First we assess the aggregate resource states this CPU's
684 	 * tasks have been in since the last change, and account any
685 	 * SOME and FULL time these may have resulted in.
686 	 *
687 	 * Then we update the task counts according to the state
688 	 * change requested through the @clear and @set bits.
689 	 */
690 	write_seqcount_begin(&groupc->seq);
691 
692 	record_times(groupc, cpu, false);
693 
694 	for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
695 		if (!(m & (1 << t)))
696 			continue;
697 		if (groupc->tasks[t] == 0 && !psi_bug) {
698 			printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u] clear=%x set=%x\n",
699 					cpu, t, groupc->tasks[0],
700 					groupc->tasks[1], groupc->tasks[2],
701 					clear, set);
702 			psi_bug = 1;
703 		}
704 		groupc->tasks[t]--;
705 	}
706 
707 	for (t = 0; set; set &= ~(1 << t), t++)
708 		if (set & (1 << t))
709 			groupc->tasks[t]++;
710 
711 	/* Calculate state mask representing active states */
712 	for (s = 0; s < NR_PSI_STATES; s++) {
713 		if (test_state(groupc->tasks, s))
714 			state_mask |= (1 << s);
715 	}
716 	groupc->state_mask = state_mask;
717 
718 	write_seqcount_end(&groupc->seq);
719 
720 	return state_mask;
721 }
722 
723 static struct psi_group *iterate_groups(struct task_struct *task, void **iter)
724 {
725 #ifdef CONFIG_CGROUPS
726 	struct cgroup *cgroup = NULL;
727 
728 	if (!*iter)
729 		cgroup = task->cgroups->dfl_cgrp;
730 	else if (*iter == &psi_system)
731 		return NULL;
732 	else
733 		cgroup = cgroup_parent(*iter);
734 
735 	if (cgroup && cgroup_parent(cgroup)) {
736 		*iter = cgroup;
737 		return cgroup_psi(cgroup);
738 	}
739 #else
740 	if (*iter)
741 		return NULL;
742 #endif
743 	*iter = &psi_system;
744 	return &psi_system;
745 }
746 
747 void psi_task_change(struct task_struct *task, int clear, int set)
748 {
749 	int cpu = task_cpu(task);
750 	struct psi_group *group;
751 	bool wake_clock = true;
752 	void *iter = NULL;
753 
754 	if (!task->pid)
755 		return;
756 
757 	if (((task->psi_flags & set) ||
758 	     (task->psi_flags & clear) != clear) &&
759 	    !psi_bug) {
760 		printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
761 				task->pid, task->comm, cpu,
762 				task->psi_flags, clear, set);
763 		psi_bug = 1;
764 	}
765 
766 	task->psi_flags &= ~clear;
767 	task->psi_flags |= set;
768 
769 	/*
770 	 * Periodic aggregation shuts off if there is a period of no
771 	 * task changes, so we wake it back up if necessary. However,
772 	 * don't do this if the task change is the aggregation worker
773 	 * itself going to sleep, or we'll ping-pong forever.
774 	 */
775 	if (unlikely((clear & TSK_RUNNING) &&
776 		     (task->flags & PF_WQ_WORKER) &&
777 		     wq_worker_last_func(task) == psi_avgs_work))
778 		wake_clock = false;
779 
780 	while ((group = iterate_groups(task, &iter))) {
781 		u32 state_mask = psi_group_change(group, cpu, clear, set);
782 
783 		if (state_mask & group->poll_states)
784 			psi_schedule_poll_work(group, 1);
785 
786 		if (wake_clock && !delayed_work_pending(&group->avgs_work))
787 			schedule_delayed_work(&group->avgs_work, PSI_FREQ);
788 	}
789 }
790 
791 void psi_memstall_tick(struct task_struct *task, int cpu)
792 {
793 	struct psi_group *group;
794 	void *iter = NULL;
795 
796 	while ((group = iterate_groups(task, &iter))) {
797 		struct psi_group_cpu *groupc;
798 
799 		groupc = per_cpu_ptr(group->pcpu, cpu);
800 		write_seqcount_begin(&groupc->seq);
801 		record_times(groupc, cpu, true);
802 		write_seqcount_end(&groupc->seq);
803 	}
804 }
805 
806 /**
807  * psi_memstall_enter - mark the beginning of a memory stall section
808  * @flags: flags to handle nested sections
809  *
810  * Marks the calling task as being stalled due to a lack of memory,
811  * such as waiting for a refault or performing reclaim.
812  */
813 void psi_memstall_enter(unsigned long *flags)
814 {
815 	struct rq_flags rf;
816 	struct rq *rq;
817 
818 	if (static_branch_likely(&psi_disabled))
819 		return;
820 
821 	*flags = current->flags & PF_MEMSTALL;
822 	if (*flags)
823 		return;
824 	/*
825 	 * PF_MEMSTALL setting & accounting needs to be atomic wrt
826 	 * changes to the task's scheduling state, otherwise we can
827 	 * race with CPU migration.
828 	 */
829 	rq = this_rq_lock_irq(&rf);
830 
831 	current->flags |= PF_MEMSTALL;
832 	psi_task_change(current, 0, TSK_MEMSTALL);
833 
834 	rq_unlock_irq(rq, &rf);
835 }
836 
837 /**
838  * psi_memstall_leave - mark the end of an memory stall section
839  * @flags: flags to handle nested memdelay sections
840  *
841  * Marks the calling task as no longer stalled due to lack of memory.
842  */
843 void psi_memstall_leave(unsigned long *flags)
844 {
845 	struct rq_flags rf;
846 	struct rq *rq;
847 
848 	if (static_branch_likely(&psi_disabled))
849 		return;
850 
851 	if (*flags)
852 		return;
853 	/*
854 	 * PF_MEMSTALL clearing & accounting needs to be atomic wrt
855 	 * changes to the task's scheduling state, otherwise we could
856 	 * race with CPU migration.
857 	 */
858 	rq = this_rq_lock_irq(&rf);
859 
860 	current->flags &= ~PF_MEMSTALL;
861 	psi_task_change(current, TSK_MEMSTALL, 0);
862 
863 	rq_unlock_irq(rq, &rf);
864 }
865 
866 #ifdef CONFIG_CGROUPS
867 int psi_cgroup_alloc(struct cgroup *cgroup)
868 {
869 	if (static_branch_likely(&psi_disabled))
870 		return 0;
871 
872 	cgroup->psi.pcpu = alloc_percpu(struct psi_group_cpu);
873 	if (!cgroup->psi.pcpu)
874 		return -ENOMEM;
875 	group_init(&cgroup->psi);
876 	return 0;
877 }
878 
879 void psi_cgroup_free(struct cgroup *cgroup)
880 {
881 	if (static_branch_likely(&psi_disabled))
882 		return;
883 
884 	cancel_delayed_work_sync(&cgroup->psi.avgs_work);
885 	free_percpu(cgroup->psi.pcpu);
886 	/* All triggers must be removed by now */
887 	WARN_ONCE(cgroup->psi.poll_states, "psi: trigger leak\n");
888 }
889 
890 /**
891  * cgroup_move_task - move task to a different cgroup
892  * @task: the task
893  * @to: the target css_set
894  *
895  * Move task to a new cgroup and safely migrate its associated stall
896  * state between the different groups.
897  *
898  * This function acquires the task's rq lock to lock out concurrent
899  * changes to the task's scheduling state and - in case the task is
900  * running - concurrent changes to its stall state.
901  */
902 void cgroup_move_task(struct task_struct *task, struct css_set *to)
903 {
904 	unsigned int task_flags = 0;
905 	struct rq_flags rf;
906 	struct rq *rq;
907 
908 	if (static_branch_likely(&psi_disabled)) {
909 		/*
910 		 * Lame to do this here, but the scheduler cannot be locked
911 		 * from the outside, so we move cgroups from inside sched/.
912 		 */
913 		rcu_assign_pointer(task->cgroups, to);
914 		return;
915 	}
916 
917 	rq = task_rq_lock(task, &rf);
918 
919 	if (task_on_rq_queued(task))
920 		task_flags = TSK_RUNNING;
921 	else if (task->in_iowait)
922 		task_flags = TSK_IOWAIT;
923 
924 	if (task->flags & PF_MEMSTALL)
925 		task_flags |= TSK_MEMSTALL;
926 
927 	if (task_flags)
928 		psi_task_change(task, task_flags, 0);
929 
930 	/* See comment above */
931 	rcu_assign_pointer(task->cgroups, to);
932 
933 	if (task_flags)
934 		psi_task_change(task, 0, task_flags);
935 
936 	task_rq_unlock(rq, task, &rf);
937 }
938 #endif /* CONFIG_CGROUPS */
939 
940 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
941 {
942 	int full;
943 	u64 now;
944 
945 	if (static_branch_likely(&psi_disabled))
946 		return -EOPNOTSUPP;
947 
948 	/* Update averages before reporting them */
949 	mutex_lock(&group->avgs_lock);
950 	now = sched_clock();
951 	collect_percpu_times(group, PSI_AVGS, NULL);
952 	if (now >= group->avg_next_update)
953 		group->avg_next_update = update_averages(group, now);
954 	mutex_unlock(&group->avgs_lock);
955 
956 	for (full = 0; full < 2 - (res == PSI_CPU); full++) {
957 		unsigned long avg[3];
958 		u64 total;
959 		int w;
960 
961 		for (w = 0; w < 3; w++)
962 			avg[w] = group->avg[res * 2 + full][w];
963 		total = div_u64(group->total[PSI_AVGS][res * 2 + full],
964 				NSEC_PER_USEC);
965 
966 		seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
967 			   full ? "full" : "some",
968 			   LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
969 			   LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
970 			   LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
971 			   total);
972 	}
973 
974 	return 0;
975 }
976 
977 static int psi_io_show(struct seq_file *m, void *v)
978 {
979 	return psi_show(m, &psi_system, PSI_IO);
980 }
981 
982 static int psi_memory_show(struct seq_file *m, void *v)
983 {
984 	return psi_show(m, &psi_system, PSI_MEM);
985 }
986 
987 static int psi_cpu_show(struct seq_file *m, void *v)
988 {
989 	return psi_show(m, &psi_system, PSI_CPU);
990 }
991 
992 static int psi_io_open(struct inode *inode, struct file *file)
993 {
994 	return single_open(file, psi_io_show, NULL);
995 }
996 
997 static int psi_memory_open(struct inode *inode, struct file *file)
998 {
999 	return single_open(file, psi_memory_show, NULL);
1000 }
1001 
1002 static int psi_cpu_open(struct inode *inode, struct file *file)
1003 {
1004 	return single_open(file, psi_cpu_show, NULL);
1005 }
1006 
1007 struct psi_trigger *psi_trigger_create(struct psi_group *group,
1008 			char *buf, size_t nbytes, enum psi_res res)
1009 {
1010 	struct psi_trigger *t;
1011 	enum psi_states state;
1012 	u32 threshold_us;
1013 	u32 window_us;
1014 
1015 	if (static_branch_likely(&psi_disabled))
1016 		return ERR_PTR(-EOPNOTSUPP);
1017 
1018 	if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1019 		state = PSI_IO_SOME + res * 2;
1020 	else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1021 		state = PSI_IO_FULL + res * 2;
1022 	else
1023 		return ERR_PTR(-EINVAL);
1024 
1025 	if (state >= PSI_NONIDLE)
1026 		return ERR_PTR(-EINVAL);
1027 
1028 	if (window_us < WINDOW_MIN_US ||
1029 		window_us > WINDOW_MAX_US)
1030 		return ERR_PTR(-EINVAL);
1031 
1032 	/* Check threshold */
1033 	if (threshold_us == 0 || threshold_us > window_us)
1034 		return ERR_PTR(-EINVAL);
1035 
1036 	t = kmalloc(sizeof(*t), GFP_KERNEL);
1037 	if (!t)
1038 		return ERR_PTR(-ENOMEM);
1039 
1040 	t->group = group;
1041 	t->state = state;
1042 	t->threshold = threshold_us * NSEC_PER_USEC;
1043 	t->win.size = window_us * NSEC_PER_USEC;
1044 	window_reset(&t->win, 0, 0, 0);
1045 
1046 	t->event = 0;
1047 	t->last_event_time = 0;
1048 	init_waitqueue_head(&t->event_wait);
1049 	kref_init(&t->refcount);
1050 
1051 	mutex_lock(&group->trigger_lock);
1052 
1053 	if (!rcu_access_pointer(group->poll_kworker)) {
1054 		struct sched_param param = {
1055 			.sched_priority = 1,
1056 		};
1057 		struct kthread_worker *kworker;
1058 
1059 		kworker = kthread_create_worker(0, "psimon");
1060 		if (IS_ERR(kworker)) {
1061 			kfree(t);
1062 			mutex_unlock(&group->trigger_lock);
1063 			return ERR_CAST(kworker);
1064 		}
1065 		sched_setscheduler_nocheck(kworker->task, SCHED_FIFO, &param);
1066 		kthread_init_delayed_work(&group->poll_work,
1067 				psi_poll_work);
1068 		rcu_assign_pointer(group->poll_kworker, kworker);
1069 	}
1070 
1071 	list_add(&t->node, &group->triggers);
1072 	group->poll_min_period = min(group->poll_min_period,
1073 		div_u64(t->win.size, UPDATES_PER_WINDOW));
1074 	group->nr_triggers[t->state]++;
1075 	group->poll_states |= (1 << t->state);
1076 
1077 	mutex_unlock(&group->trigger_lock);
1078 
1079 	return t;
1080 }
1081 
1082 static void psi_trigger_destroy(struct kref *ref)
1083 {
1084 	struct psi_trigger *t = container_of(ref, struct psi_trigger, refcount);
1085 	struct psi_group *group = t->group;
1086 	struct kthread_worker *kworker_to_destroy = NULL;
1087 
1088 	if (static_branch_likely(&psi_disabled))
1089 		return;
1090 
1091 	/*
1092 	 * Wakeup waiters to stop polling. Can happen if cgroup is deleted
1093 	 * from under a polling process.
1094 	 */
1095 	wake_up_interruptible(&t->event_wait);
1096 
1097 	mutex_lock(&group->trigger_lock);
1098 
1099 	if (!list_empty(&t->node)) {
1100 		struct psi_trigger *tmp;
1101 		u64 period = ULLONG_MAX;
1102 
1103 		list_del(&t->node);
1104 		group->nr_triggers[t->state]--;
1105 		if (!group->nr_triggers[t->state])
1106 			group->poll_states &= ~(1 << t->state);
1107 		/* reset min update period for the remaining triggers */
1108 		list_for_each_entry(tmp, &group->triggers, node)
1109 			period = min(period, div_u64(tmp->win.size,
1110 					UPDATES_PER_WINDOW));
1111 		group->poll_min_period = period;
1112 		/* Destroy poll_kworker when the last trigger is destroyed */
1113 		if (group->poll_states == 0) {
1114 			group->polling_until = 0;
1115 			kworker_to_destroy = rcu_dereference_protected(
1116 					group->poll_kworker,
1117 					lockdep_is_held(&group->trigger_lock));
1118 			rcu_assign_pointer(group->poll_kworker, NULL);
1119 		}
1120 	}
1121 
1122 	mutex_unlock(&group->trigger_lock);
1123 
1124 	/*
1125 	 * Wait for both *trigger_ptr from psi_trigger_replace and
1126 	 * poll_kworker RCUs to complete their read-side critical sections
1127 	 * before destroying the trigger and optionally the poll_kworker
1128 	 */
1129 	synchronize_rcu();
1130 	/*
1131 	 * Destroy the kworker after releasing trigger_lock to prevent a
1132 	 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1133 	 */
1134 	if (kworker_to_destroy) {
1135 		/*
1136 		 * After the RCU grace period has expired, the worker
1137 		 * can no longer be found through group->poll_kworker.
1138 		 * But it might have been already scheduled before
1139 		 * that - deschedule it cleanly before destroying it.
1140 		 */
1141 		kthread_cancel_delayed_work_sync(&group->poll_work);
1142 		atomic_set(&group->poll_scheduled, 0);
1143 
1144 		kthread_destroy_worker(kworker_to_destroy);
1145 	}
1146 	kfree(t);
1147 }
1148 
1149 void psi_trigger_replace(void **trigger_ptr, struct psi_trigger *new)
1150 {
1151 	struct psi_trigger *old = *trigger_ptr;
1152 
1153 	if (static_branch_likely(&psi_disabled))
1154 		return;
1155 
1156 	rcu_assign_pointer(*trigger_ptr, new);
1157 	if (old)
1158 		kref_put(&old->refcount, psi_trigger_destroy);
1159 }
1160 
1161 __poll_t psi_trigger_poll(void **trigger_ptr,
1162 				struct file *file, poll_table *wait)
1163 {
1164 	__poll_t ret = DEFAULT_POLLMASK;
1165 	struct psi_trigger *t;
1166 
1167 	if (static_branch_likely(&psi_disabled))
1168 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1169 
1170 	rcu_read_lock();
1171 
1172 	t = rcu_dereference(*(void __rcu __force **)trigger_ptr);
1173 	if (!t) {
1174 		rcu_read_unlock();
1175 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1176 	}
1177 	kref_get(&t->refcount);
1178 
1179 	rcu_read_unlock();
1180 
1181 	poll_wait(file, &t->event_wait, wait);
1182 
1183 	if (cmpxchg(&t->event, 1, 0) == 1)
1184 		ret |= EPOLLPRI;
1185 
1186 	kref_put(&t->refcount, psi_trigger_destroy);
1187 
1188 	return ret;
1189 }
1190 
1191 static ssize_t psi_write(struct file *file, const char __user *user_buf,
1192 			 size_t nbytes, enum psi_res res)
1193 {
1194 	char buf[32];
1195 	size_t buf_size;
1196 	struct seq_file *seq;
1197 	struct psi_trigger *new;
1198 
1199 	if (static_branch_likely(&psi_disabled))
1200 		return -EOPNOTSUPP;
1201 
1202 	buf_size = min(nbytes, sizeof(buf));
1203 	if (copy_from_user(buf, user_buf, buf_size))
1204 		return -EFAULT;
1205 
1206 	buf[buf_size - 1] = '\0';
1207 
1208 	new = psi_trigger_create(&psi_system, buf, nbytes, res);
1209 	if (IS_ERR(new))
1210 		return PTR_ERR(new);
1211 
1212 	seq = file->private_data;
1213 	/* Take seq->lock to protect seq->private from concurrent writes */
1214 	mutex_lock(&seq->lock);
1215 	psi_trigger_replace(&seq->private, new);
1216 	mutex_unlock(&seq->lock);
1217 
1218 	return nbytes;
1219 }
1220 
1221 static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1222 			    size_t nbytes, loff_t *ppos)
1223 {
1224 	return psi_write(file, user_buf, nbytes, PSI_IO);
1225 }
1226 
1227 static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1228 				size_t nbytes, loff_t *ppos)
1229 {
1230 	return psi_write(file, user_buf, nbytes, PSI_MEM);
1231 }
1232 
1233 static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1234 			     size_t nbytes, loff_t *ppos)
1235 {
1236 	return psi_write(file, user_buf, nbytes, PSI_CPU);
1237 }
1238 
1239 static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1240 {
1241 	struct seq_file *seq = file->private_data;
1242 
1243 	return psi_trigger_poll(&seq->private, file, wait);
1244 }
1245 
1246 static int psi_fop_release(struct inode *inode, struct file *file)
1247 {
1248 	struct seq_file *seq = file->private_data;
1249 
1250 	psi_trigger_replace(&seq->private, NULL);
1251 	return single_release(inode, file);
1252 }
1253 
1254 static const struct file_operations psi_io_fops = {
1255 	.open           = psi_io_open,
1256 	.read           = seq_read,
1257 	.llseek         = seq_lseek,
1258 	.write          = psi_io_write,
1259 	.poll           = psi_fop_poll,
1260 	.release        = psi_fop_release,
1261 };
1262 
1263 static const struct file_operations psi_memory_fops = {
1264 	.open           = psi_memory_open,
1265 	.read           = seq_read,
1266 	.llseek         = seq_lseek,
1267 	.write          = psi_memory_write,
1268 	.poll           = psi_fop_poll,
1269 	.release        = psi_fop_release,
1270 };
1271 
1272 static const struct file_operations psi_cpu_fops = {
1273 	.open           = psi_cpu_open,
1274 	.read           = seq_read,
1275 	.llseek         = seq_lseek,
1276 	.write          = psi_cpu_write,
1277 	.poll           = psi_fop_poll,
1278 	.release        = psi_fop_release,
1279 };
1280 
1281 static int __init psi_proc_init(void)
1282 {
1283 	proc_mkdir("pressure", NULL);
1284 	proc_create("pressure/io", 0, NULL, &psi_io_fops);
1285 	proc_create("pressure/memory", 0, NULL, &psi_memory_fops);
1286 	proc_create("pressure/cpu", 0, NULL, &psi_cpu_fops);
1287 	return 0;
1288 }
1289 module_init(psi_proc_init);
1290