1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Pressure stall information for CPU, memory and IO
4  *
5  * Copyright (c) 2018 Facebook, Inc.
6  * Author: Johannes Weiner <hannes@cmpxchg.org>
7  *
8  * Polling support by Suren Baghdasaryan <surenb@google.com>
9  * Copyright (c) 2018 Google, Inc.
10  *
11  * When CPU, memory and IO are contended, tasks experience delays that
12  * reduce throughput and introduce latencies into the workload. Memory
13  * and IO contention, in addition, can cause a full loss of forward
14  * progress in which the CPU goes idle.
15  *
16  * This code aggregates individual task delays into resource pressure
17  * metrics that indicate problems with both workload health and
18  * resource utilization.
19  *
20  *			Model
21  *
22  * The time in which a task can execute on a CPU is our baseline for
23  * productivity. Pressure expresses the amount of time in which this
24  * potential cannot be realized due to resource contention.
25  *
26  * This concept of productivity has two components: the workload and
27  * the CPU. To measure the impact of pressure on both, we define two
28  * contention states for a resource: SOME and FULL.
29  *
30  * In the SOME state of a given resource, one or more tasks are
31  * delayed on that resource. This affects the workload's ability to
32  * perform work, but the CPU may still be executing other tasks.
33  *
34  * In the FULL state of a given resource, all non-idle tasks are
35  * delayed on that resource such that nobody is advancing and the CPU
36  * goes idle. This leaves both workload and CPU unproductive.
37  *
38  *	SOME = nr_delayed_tasks != 0
39  *	FULL = nr_delayed_tasks != 0 && nr_productive_tasks == 0
40  *
41  * What it means for a task to be productive is defined differently
42  * for each resource. For IO, productive means a running task. For
43  * memory, productive means a running task that isn't a reclaimer. For
44  * CPU, productive means an oncpu task.
45  *
46  * Naturally, the FULL state doesn't exist for the CPU resource at the
47  * system level, but exist at the cgroup level. At the cgroup level,
48  * FULL means all non-idle tasks in the cgroup are delayed on the CPU
49  * resource which is being used by others outside of the cgroup or
50  * throttled by the cgroup cpu.max configuration.
51  *
52  * The percentage of wallclock time spent in those compound stall
53  * states gives pressure numbers between 0 and 100 for each resource,
54  * where the SOME percentage indicates workload slowdowns and the FULL
55  * percentage indicates reduced CPU utilization:
56  *
57  *	%SOME = time(SOME) / period
58  *	%FULL = time(FULL) / period
59  *
60  *			Multiple CPUs
61  *
62  * The more tasks and available CPUs there are, the more work can be
63  * performed concurrently. This means that the potential that can go
64  * unrealized due to resource contention *also* scales with non-idle
65  * tasks and CPUs.
66  *
67  * Consider a scenario where 257 number crunching tasks are trying to
68  * run concurrently on 256 CPUs. If we simply aggregated the task
69  * states, we would have to conclude a CPU SOME pressure number of
70  * 100%, since *somebody* is waiting on a runqueue at all
71  * times. However, that is clearly not the amount of contention the
72  * workload is experiencing: only one out of 256 possible execution
73  * threads will be contended at any given time, or about 0.4%.
74  *
75  * Conversely, consider a scenario of 4 tasks and 4 CPUs where at any
76  * given time *one* of the tasks is delayed due to a lack of memory.
77  * Again, looking purely at the task state would yield a memory FULL
78  * pressure number of 0%, since *somebody* is always making forward
79  * progress. But again this wouldn't capture the amount of execution
80  * potential lost, which is 1 out of 4 CPUs, or 25%.
81  *
82  * To calculate wasted potential (pressure) with multiple processors,
83  * we have to base our calculation on the number of non-idle tasks in
84  * conjunction with the number of available CPUs, which is the number
85  * of potential execution threads. SOME becomes then the proportion of
86  * delayed tasks to possible threads, and FULL is the share of possible
87  * threads that are unproductive due to delays:
88  *
89  *	threads = min(nr_nonidle_tasks, nr_cpus)
90  *	   SOME = min(nr_delayed_tasks / threads, 1)
91  *	   FULL = (threads - min(nr_productive_tasks, threads)) / threads
92  *
93  * For the 257 number crunchers on 256 CPUs, this yields:
94  *
95  *	threads = min(257, 256)
96  *	   SOME = min(1 / 256, 1)             = 0.4%
97  *	   FULL = (256 - min(256, 256)) / 256 = 0%
98  *
99  * For the 1 out of 4 memory-delayed tasks, this yields:
100  *
101  *	threads = min(4, 4)
102  *	   SOME = min(1 / 4, 1)               = 25%
103  *	   FULL = (4 - min(3, 4)) / 4         = 25%
104  *
105  * [ Substitute nr_cpus with 1, and you can see that it's a natural
106  *   extension of the single-CPU model. ]
107  *
108  *			Implementation
109  *
110  * To assess the precise time spent in each such state, we would have
111  * to freeze the system on task changes and start/stop the state
112  * clocks accordingly. Obviously that doesn't scale in practice.
113  *
114  * Because the scheduler aims to distribute the compute load evenly
115  * among the available CPUs, we can track task state locally to each
116  * CPU and, at much lower frequency, extrapolate the global state for
117  * the cumulative stall times and the running averages.
118  *
119  * For each runqueue, we track:
120  *
121  *	   tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0)
122  *	   tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_productive_tasks[cpu])
123  *	tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0)
124  *
125  * and then periodically aggregate:
126  *
127  *	tNONIDLE = sum(tNONIDLE[i])
128  *
129  *	   tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE
130  *	   tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE
131  *
132  *	   %SOME = tSOME / period
133  *	   %FULL = tFULL / period
134  *
135  * This gives us an approximation of pressure that is practical
136  * cost-wise, yet way more sensitive and accurate than periodic
137  * sampling of the aggregate task states would be.
138  */
139 
140 static int psi_bug __read_mostly;
141 
142 DEFINE_STATIC_KEY_FALSE(psi_disabled);
143 DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled);
144 
145 #ifdef CONFIG_PSI_DEFAULT_DISABLED
146 static bool psi_enable;
147 #else
148 static bool psi_enable = true;
149 #endif
setup_psi(char * str)150 static int __init setup_psi(char *str)
151 {
152 	return kstrtobool(str, &psi_enable) == 0;
153 }
154 __setup("psi=", setup_psi);
155 
156 /* Running averages - we need to be higher-res than loadavg */
157 #define PSI_FREQ	(2*HZ+1)	/* 2 sec intervals */
158 #define EXP_10s		1677		/* 1/exp(2s/10s) as fixed-point */
159 #define EXP_60s		1981		/* 1/exp(2s/60s) */
160 #define EXP_300s	2034		/* 1/exp(2s/300s) */
161 
162 /* PSI trigger definitions */
163 #define WINDOW_MIN_US 500000	/* Min window size is 500ms */
164 #define WINDOW_MAX_US 10000000	/* Max window size is 10s */
165 #define UPDATES_PER_WINDOW 10	/* 10 updates per window */
166 
167 /* Sampling frequency in nanoseconds */
168 static u64 psi_period __read_mostly;
169 
170 /* System-level pressure and stall tracking */
171 static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
172 struct psi_group psi_system = {
173 	.pcpu = &system_group_pcpu,
174 };
175 
176 static void psi_avgs_work(struct work_struct *work);
177 
178 static void poll_timer_fn(struct timer_list *t);
179 
group_init(struct psi_group * group)180 static void group_init(struct psi_group *group)
181 {
182 	int cpu;
183 
184 	group->enabled = true;
185 	for_each_possible_cpu(cpu)
186 		seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);
187 	group->avg_last_update = sched_clock();
188 	group->avg_next_update = group->avg_last_update + psi_period;
189 	INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
190 	mutex_init(&group->avgs_lock);
191 	/* Init trigger-related members */
192 	mutex_init(&group->trigger_lock);
193 	INIT_LIST_HEAD(&group->triggers);
194 	group->poll_min_period = U32_MAX;
195 	group->polling_next_update = ULLONG_MAX;
196 	init_waitqueue_head(&group->poll_wait);
197 	timer_setup(&group->poll_timer, poll_timer_fn, 0);
198 	rcu_assign_pointer(group->poll_task, NULL);
199 }
200 
psi_init(void)201 void __init psi_init(void)
202 {
203 	if (!psi_enable) {
204 		static_branch_enable(&psi_disabled);
205 		static_branch_disable(&psi_cgroups_enabled);
206 		return;
207 	}
208 
209 	if (!cgroup_psi_enabled())
210 		static_branch_disable(&psi_cgroups_enabled);
211 
212 	psi_period = jiffies_to_nsecs(PSI_FREQ);
213 	group_init(&psi_system);
214 }
215 
test_state(unsigned int * tasks,enum psi_states state,bool oncpu)216 static bool test_state(unsigned int *tasks, enum psi_states state, bool oncpu)
217 {
218 	switch (state) {
219 	case PSI_IO_SOME:
220 		return unlikely(tasks[NR_IOWAIT]);
221 	case PSI_IO_FULL:
222 		return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]);
223 	case PSI_MEM_SOME:
224 		return unlikely(tasks[NR_MEMSTALL]);
225 	case PSI_MEM_FULL:
226 		return unlikely(tasks[NR_MEMSTALL] &&
227 			tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]);
228 	case PSI_CPU_SOME:
229 		return unlikely(tasks[NR_RUNNING] > oncpu);
230 	case PSI_CPU_FULL:
231 		return unlikely(tasks[NR_RUNNING] && !oncpu);
232 	case PSI_NONIDLE:
233 		return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
234 			tasks[NR_RUNNING];
235 	default:
236 		return false;
237 	}
238 }
239 
get_recent_times(struct psi_group * group,int cpu,enum psi_aggregators aggregator,u32 * times,u32 * pchanged_states)240 static void get_recent_times(struct psi_group *group, int cpu,
241 			     enum psi_aggregators aggregator, u32 *times,
242 			     u32 *pchanged_states)
243 {
244 	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
245 	u64 now, state_start;
246 	enum psi_states s;
247 	unsigned int seq;
248 	u32 state_mask;
249 
250 	*pchanged_states = 0;
251 
252 	/* Snapshot a coherent view of the CPU state */
253 	do {
254 		seq = read_seqcount_begin(&groupc->seq);
255 		now = cpu_clock(cpu);
256 		memcpy(times, groupc->times, sizeof(groupc->times));
257 		state_mask = groupc->state_mask;
258 		state_start = groupc->state_start;
259 	} while (read_seqcount_retry(&groupc->seq, seq));
260 
261 	/* Calculate state time deltas against the previous snapshot */
262 	for (s = 0; s < NR_PSI_STATES; s++) {
263 		u32 delta;
264 		/*
265 		 * In addition to already concluded states, we also
266 		 * incorporate currently active states on the CPU,
267 		 * since states may last for many sampling periods.
268 		 *
269 		 * This way we keep our delta sampling buckets small
270 		 * (u32) and our reported pressure close to what's
271 		 * actually happening.
272 		 */
273 		if (state_mask & (1 << s))
274 			times[s] += now - state_start;
275 
276 		delta = times[s] - groupc->times_prev[aggregator][s];
277 		groupc->times_prev[aggregator][s] = times[s];
278 
279 		times[s] = delta;
280 		if (delta)
281 			*pchanged_states |= (1 << s);
282 	}
283 }
284 
calc_avgs(unsigned long avg[3],int missed_periods,u64 time,u64 period)285 static void calc_avgs(unsigned long avg[3], int missed_periods,
286 		      u64 time, u64 period)
287 {
288 	unsigned long pct;
289 
290 	/* Fill in zeroes for periods of no activity */
291 	if (missed_periods) {
292 		avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
293 		avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
294 		avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
295 	}
296 
297 	/* Sample the most recent active period */
298 	pct = div_u64(time * 100, period);
299 	pct *= FIXED_1;
300 	avg[0] = calc_load(avg[0], EXP_10s, pct);
301 	avg[1] = calc_load(avg[1], EXP_60s, pct);
302 	avg[2] = calc_load(avg[2], EXP_300s, pct);
303 }
304 
collect_percpu_times(struct psi_group * group,enum psi_aggregators aggregator,u32 * pchanged_states)305 static void collect_percpu_times(struct psi_group *group,
306 				 enum psi_aggregators aggregator,
307 				 u32 *pchanged_states)
308 {
309 	u64 deltas[NR_PSI_STATES - 1] = { 0, };
310 	unsigned long nonidle_total = 0;
311 	u32 changed_states = 0;
312 	int cpu;
313 	int s;
314 
315 	/*
316 	 * Collect the per-cpu time buckets and average them into a
317 	 * single time sample that is normalized to wallclock time.
318 	 *
319 	 * For averaging, each CPU is weighted by its non-idle time in
320 	 * the sampling period. This eliminates artifacts from uneven
321 	 * loading, or even entirely idle CPUs.
322 	 */
323 	for_each_possible_cpu(cpu) {
324 		u32 times[NR_PSI_STATES];
325 		u32 nonidle;
326 		u32 cpu_changed_states;
327 
328 		get_recent_times(group, cpu, aggregator, times,
329 				&cpu_changed_states);
330 		changed_states |= cpu_changed_states;
331 
332 		nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
333 		nonidle_total += nonidle;
334 
335 		for (s = 0; s < PSI_NONIDLE; s++)
336 			deltas[s] += (u64)times[s] * nonidle;
337 	}
338 
339 	/*
340 	 * Integrate the sample into the running statistics that are
341 	 * reported to userspace: the cumulative stall times and the
342 	 * decaying averages.
343 	 *
344 	 * Pressure percentages are sampled at PSI_FREQ. We might be
345 	 * called more often when the user polls more frequently than
346 	 * that; we might be called less often when there is no task
347 	 * activity, thus no data, and clock ticks are sporadic. The
348 	 * below handles both.
349 	 */
350 
351 	/* total= */
352 	for (s = 0; s < NR_PSI_STATES - 1; s++)
353 		group->total[aggregator][s] +=
354 				div_u64(deltas[s], max(nonidle_total, 1UL));
355 
356 	if (pchanged_states)
357 		*pchanged_states = changed_states;
358 }
359 
update_averages(struct psi_group * group,u64 now)360 static u64 update_averages(struct psi_group *group, u64 now)
361 {
362 	unsigned long missed_periods = 0;
363 	u64 expires, period;
364 	u64 avg_next_update;
365 	int s;
366 
367 	/* avgX= */
368 	expires = group->avg_next_update;
369 	if (now - expires >= psi_period)
370 		missed_periods = div_u64(now - expires, psi_period);
371 
372 	/*
373 	 * The periodic clock tick can get delayed for various
374 	 * reasons, especially on loaded systems. To avoid clock
375 	 * drift, we schedule the clock in fixed psi_period intervals.
376 	 * But the deltas we sample out of the per-cpu buckets above
377 	 * are based on the actual time elapsing between clock ticks.
378 	 */
379 	avg_next_update = expires + ((1 + missed_periods) * psi_period);
380 	period = now - (group->avg_last_update + (missed_periods * psi_period));
381 	group->avg_last_update = now;
382 
383 	for (s = 0; s < NR_PSI_STATES - 1; s++) {
384 		u32 sample;
385 
386 		sample = group->total[PSI_AVGS][s] - group->avg_total[s];
387 		/*
388 		 * Due to the lockless sampling of the time buckets,
389 		 * recorded time deltas can slip into the next period,
390 		 * which under full pressure can result in samples in
391 		 * excess of the period length.
392 		 *
393 		 * We don't want to report non-sensical pressures in
394 		 * excess of 100%, nor do we want to drop such events
395 		 * on the floor. Instead we punt any overage into the
396 		 * future until pressure subsides. By doing this we
397 		 * don't underreport the occurring pressure curve, we
398 		 * just report it delayed by one period length.
399 		 *
400 		 * The error isn't cumulative. As soon as another
401 		 * delta slips from a period P to P+1, by definition
402 		 * it frees up its time T in P.
403 		 */
404 		if (sample > period)
405 			sample = period;
406 		group->avg_total[s] += sample;
407 		calc_avgs(group->avg[s], missed_periods, sample, period);
408 	}
409 
410 	return avg_next_update;
411 }
412 
psi_avgs_work(struct work_struct * work)413 static void psi_avgs_work(struct work_struct *work)
414 {
415 	struct delayed_work *dwork;
416 	struct psi_group *group;
417 	u32 changed_states;
418 	bool nonidle;
419 	u64 now;
420 
421 	dwork = to_delayed_work(work);
422 	group = container_of(dwork, struct psi_group, avgs_work);
423 
424 	mutex_lock(&group->avgs_lock);
425 
426 	now = sched_clock();
427 
428 	collect_percpu_times(group, PSI_AVGS, &changed_states);
429 	nonidle = changed_states & (1 << PSI_NONIDLE);
430 	/*
431 	 * If there is task activity, periodically fold the per-cpu
432 	 * times and feed samples into the running averages. If things
433 	 * are idle and there is no data to process, stop the clock.
434 	 * Once restarted, we'll catch up the running averages in one
435 	 * go - see calc_avgs() and missed_periods.
436 	 */
437 	if (now >= group->avg_next_update)
438 		group->avg_next_update = update_averages(group, now);
439 
440 	if (nonidle) {
441 		schedule_delayed_work(dwork, nsecs_to_jiffies(
442 				group->avg_next_update - now) + 1);
443 	}
444 
445 	mutex_unlock(&group->avgs_lock);
446 }
447 
448 /* Trigger tracking window manipulations */
window_reset(struct psi_window * win,u64 now,u64 value,u64 prev_growth)449 static void window_reset(struct psi_window *win, u64 now, u64 value,
450 			 u64 prev_growth)
451 {
452 	win->start_time = now;
453 	win->start_value = value;
454 	win->prev_growth = prev_growth;
455 }
456 
457 /*
458  * PSI growth tracking window update and growth calculation routine.
459  *
460  * This approximates a sliding tracking window by interpolating
461  * partially elapsed windows using historical growth data from the
462  * previous intervals. This minimizes memory requirements (by not storing
463  * all the intermediate values in the previous window) and simplifies
464  * the calculations. It works well because PSI signal changes only in
465  * positive direction and over relatively small window sizes the growth
466  * is close to linear.
467  */
window_update(struct psi_window * win,u64 now,u64 value)468 static u64 window_update(struct psi_window *win, u64 now, u64 value)
469 {
470 	u64 elapsed;
471 	u64 growth;
472 
473 	elapsed = now - win->start_time;
474 	growth = value - win->start_value;
475 	/*
476 	 * After each tracking window passes win->start_value and
477 	 * win->start_time get reset and win->prev_growth stores
478 	 * the average per-window growth of the previous window.
479 	 * win->prev_growth is then used to interpolate additional
480 	 * growth from the previous window assuming it was linear.
481 	 */
482 	if (elapsed > win->size)
483 		window_reset(win, now, value, growth);
484 	else {
485 		u32 remaining;
486 
487 		remaining = win->size - elapsed;
488 		growth += div64_u64(win->prev_growth * remaining, win->size);
489 	}
490 
491 	return growth;
492 }
493 
init_triggers(struct psi_group * group,u64 now)494 static void init_triggers(struct psi_group *group, u64 now)
495 {
496 	struct psi_trigger *t;
497 
498 	list_for_each_entry(t, &group->triggers, node)
499 		window_reset(&t->win, now,
500 				group->total[PSI_POLL][t->state], 0);
501 	memcpy(group->polling_total, group->total[PSI_POLL],
502 		   sizeof(group->polling_total));
503 	group->polling_next_update = now + group->poll_min_period;
504 }
505 
update_triggers(struct psi_group * group,u64 now)506 static u64 update_triggers(struct psi_group *group, u64 now)
507 {
508 	struct psi_trigger *t;
509 	bool update_total = false;
510 	u64 *total = group->total[PSI_POLL];
511 
512 	/*
513 	 * On subsequent updates, calculate growth deltas and let
514 	 * watchers know when their specified thresholds are exceeded.
515 	 */
516 	list_for_each_entry(t, &group->triggers, node) {
517 		u64 growth;
518 		bool new_stall;
519 
520 		new_stall = group->polling_total[t->state] != total[t->state];
521 
522 		/* Check for stall activity or a previous threshold breach */
523 		if (!new_stall && !t->pending_event)
524 			continue;
525 		/*
526 		 * Check for new stall activity, as well as deferred
527 		 * events that occurred in the last window after the
528 		 * trigger had already fired (we want to ratelimit
529 		 * events without dropping any).
530 		 */
531 		if (new_stall) {
532 			/*
533 			 * Multiple triggers might be looking at the same state,
534 			 * remember to update group->polling_total[] once we've
535 			 * been through all of them. Also remember to extend the
536 			 * polling time if we see new stall activity.
537 			 */
538 			update_total = true;
539 
540 			/* Calculate growth since last update */
541 			growth = window_update(&t->win, now, total[t->state]);
542 			if (!t->pending_event) {
543 				if (growth < t->threshold)
544 					continue;
545 
546 				t->pending_event = true;
547 			}
548 		}
549 		/* Limit event signaling to once per window */
550 		if (now < t->last_event_time + t->win.size)
551 			continue;
552 
553 		/* Generate an event */
554 		if (cmpxchg(&t->event, 0, 1) == 0)
555 			wake_up_interruptible(&t->event_wait);
556 		t->last_event_time = now;
557 		/* Reset threshold breach flag once event got generated */
558 		t->pending_event = false;
559 	}
560 
561 	if (update_total)
562 		memcpy(group->polling_total, total,
563 				sizeof(group->polling_total));
564 
565 	return now + group->poll_min_period;
566 }
567 
568 /* Schedule polling if it's not already scheduled. */
psi_schedule_poll_work(struct psi_group * group,unsigned long delay)569 static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay)
570 {
571 	struct task_struct *task;
572 
573 	/*
574 	 * Do not reschedule if already scheduled.
575 	 * Possible race with a timer scheduled after this check but before
576 	 * mod_timer below can be tolerated because group->polling_next_update
577 	 * will keep updates on schedule.
578 	 */
579 	if (timer_pending(&group->poll_timer))
580 		return;
581 
582 	rcu_read_lock();
583 
584 	task = rcu_dereference(group->poll_task);
585 	/*
586 	 * kworker might be NULL in case psi_trigger_destroy races with
587 	 * psi_task_change (hotpath) which can't use locks
588 	 */
589 	if (likely(task))
590 		mod_timer(&group->poll_timer, jiffies + delay);
591 
592 	rcu_read_unlock();
593 }
594 
psi_poll_work(struct psi_group * group)595 static void psi_poll_work(struct psi_group *group)
596 {
597 	u32 changed_states;
598 	u64 now;
599 
600 	mutex_lock(&group->trigger_lock);
601 
602 	now = sched_clock();
603 
604 	collect_percpu_times(group, PSI_POLL, &changed_states);
605 
606 	if (changed_states & group->poll_states) {
607 		/* Initialize trigger windows when entering polling mode */
608 		if (now > group->polling_until)
609 			init_triggers(group, now);
610 
611 		/*
612 		 * Keep the monitor active for at least the duration of the
613 		 * minimum tracking window as long as monitor states are
614 		 * changing.
615 		 */
616 		group->polling_until = now +
617 			group->poll_min_period * UPDATES_PER_WINDOW;
618 	}
619 
620 	if (now > group->polling_until) {
621 		group->polling_next_update = ULLONG_MAX;
622 		goto out;
623 	}
624 
625 	if (now >= group->polling_next_update)
626 		group->polling_next_update = update_triggers(group, now);
627 
628 	psi_schedule_poll_work(group,
629 		nsecs_to_jiffies(group->polling_next_update - now) + 1);
630 
631 out:
632 	mutex_unlock(&group->trigger_lock);
633 }
634 
psi_poll_worker(void * data)635 static int psi_poll_worker(void *data)
636 {
637 	struct psi_group *group = (struct psi_group *)data;
638 
639 	sched_set_fifo_low(current);
640 
641 	while (true) {
642 		wait_event_interruptible(group->poll_wait,
643 				atomic_cmpxchg(&group->poll_wakeup, 1, 0) ||
644 				kthread_should_stop());
645 		if (kthread_should_stop())
646 			break;
647 
648 		psi_poll_work(group);
649 	}
650 	return 0;
651 }
652 
poll_timer_fn(struct timer_list * t)653 static void poll_timer_fn(struct timer_list *t)
654 {
655 	struct psi_group *group = from_timer(group, t, poll_timer);
656 
657 	atomic_set(&group->poll_wakeup, 1);
658 	wake_up_interruptible(&group->poll_wait);
659 }
660 
record_times(struct psi_group_cpu * groupc,u64 now)661 static void record_times(struct psi_group_cpu *groupc, u64 now)
662 {
663 	u32 delta;
664 
665 	delta = now - groupc->state_start;
666 	groupc->state_start = now;
667 
668 	if (groupc->state_mask & (1 << PSI_IO_SOME)) {
669 		groupc->times[PSI_IO_SOME] += delta;
670 		if (groupc->state_mask & (1 << PSI_IO_FULL))
671 			groupc->times[PSI_IO_FULL] += delta;
672 	}
673 
674 	if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
675 		groupc->times[PSI_MEM_SOME] += delta;
676 		if (groupc->state_mask & (1 << PSI_MEM_FULL))
677 			groupc->times[PSI_MEM_FULL] += delta;
678 	}
679 
680 	if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
681 		groupc->times[PSI_CPU_SOME] += delta;
682 		if (groupc->state_mask & (1 << PSI_CPU_FULL))
683 			groupc->times[PSI_CPU_FULL] += delta;
684 	}
685 
686 	if (groupc->state_mask & (1 << PSI_NONIDLE))
687 		groupc->times[PSI_NONIDLE] += delta;
688 }
689 
psi_group_change(struct psi_group * group,int cpu,unsigned int clear,unsigned int set,u64 now,bool wake_clock)690 static void psi_group_change(struct psi_group *group, int cpu,
691 			     unsigned int clear, unsigned int set, u64 now,
692 			     bool wake_clock)
693 {
694 	struct psi_group_cpu *groupc;
695 	unsigned int t, m;
696 	enum psi_states s;
697 	u32 state_mask;
698 
699 	groupc = per_cpu_ptr(group->pcpu, cpu);
700 
701 	/*
702 	 * First we update the task counts according to the state
703 	 * change requested through the @clear and @set bits.
704 	 *
705 	 * Then if the cgroup PSI stats accounting enabled, we
706 	 * assess the aggregate resource states this CPU's tasks
707 	 * have been in since the last change, and account any
708 	 * SOME and FULL time these may have resulted in.
709 	 */
710 	write_seqcount_begin(&groupc->seq);
711 
712 	/*
713 	 * Start with TSK_ONCPU, which doesn't have a corresponding
714 	 * task count - it's just a boolean flag directly encoded in
715 	 * the state mask. Clear, set, or carry the current state if
716 	 * no changes are requested.
717 	 */
718 	if (unlikely(clear & TSK_ONCPU)) {
719 		state_mask = 0;
720 		clear &= ~TSK_ONCPU;
721 	} else if (unlikely(set & TSK_ONCPU)) {
722 		state_mask = PSI_ONCPU;
723 		set &= ~TSK_ONCPU;
724 	} else {
725 		state_mask = groupc->state_mask & PSI_ONCPU;
726 	}
727 
728 	/*
729 	 * The rest of the state mask is calculated based on the task
730 	 * counts. Update those first, then construct the mask.
731 	 */
732 	for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
733 		if (!(m & (1 << t)))
734 			continue;
735 		if (groupc->tasks[t]) {
736 			groupc->tasks[t]--;
737 		} else if (!psi_bug) {
738 			printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n",
739 					cpu, t, groupc->tasks[0],
740 					groupc->tasks[1], groupc->tasks[2],
741 					groupc->tasks[3], clear, set);
742 			psi_bug = 1;
743 		}
744 	}
745 
746 	for (t = 0; set; set &= ~(1 << t), t++)
747 		if (set & (1 << t))
748 			groupc->tasks[t]++;
749 
750 	if (!group->enabled) {
751 		/*
752 		 * On the first group change after disabling PSI, conclude
753 		 * the current state and flush its time. This is unlikely
754 		 * to matter to the user, but aggregation (get_recent_times)
755 		 * may have already incorporated the live state into times_prev;
756 		 * avoid a delta sample underflow when PSI is later re-enabled.
757 		 */
758 		if (unlikely(groupc->state_mask & (1 << PSI_NONIDLE)))
759 			record_times(groupc, now);
760 
761 		groupc->state_mask = state_mask;
762 
763 		write_seqcount_end(&groupc->seq);
764 		return;
765 	}
766 
767 	for (s = 0; s < NR_PSI_STATES; s++) {
768 		if (test_state(groupc->tasks, s, state_mask & PSI_ONCPU))
769 			state_mask |= (1 << s);
770 	}
771 
772 	/*
773 	 * Since we care about lost potential, a memstall is FULL
774 	 * when there are no other working tasks, but also when
775 	 * the CPU is actively reclaiming and nothing productive
776 	 * could run even if it were runnable. So when the current
777 	 * task in a cgroup is in_memstall, the corresponding groupc
778 	 * on that cpu is in PSI_MEM_FULL state.
779 	 */
780 	if (unlikely((state_mask & PSI_ONCPU) && cpu_curr(cpu)->in_memstall))
781 		state_mask |= (1 << PSI_MEM_FULL);
782 
783 	record_times(groupc, now);
784 
785 	groupc->state_mask = state_mask;
786 
787 	write_seqcount_end(&groupc->seq);
788 
789 	if (state_mask & group->poll_states)
790 		psi_schedule_poll_work(group, 1);
791 
792 	if (wake_clock && !delayed_work_pending(&group->avgs_work))
793 		schedule_delayed_work(&group->avgs_work, PSI_FREQ);
794 }
795 
task_psi_group(struct task_struct * task)796 static inline struct psi_group *task_psi_group(struct task_struct *task)
797 {
798 #ifdef CONFIG_CGROUPS
799 	if (static_branch_likely(&psi_cgroups_enabled))
800 		return cgroup_psi(task_dfl_cgroup(task));
801 #endif
802 	return &psi_system;
803 }
804 
psi_flags_change(struct task_struct * task,int clear,int set)805 static void psi_flags_change(struct task_struct *task, int clear, int set)
806 {
807 	if (((task->psi_flags & set) ||
808 	     (task->psi_flags & clear) != clear) &&
809 	    !psi_bug) {
810 		printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
811 				task->pid, task->comm, task_cpu(task),
812 				task->psi_flags, clear, set);
813 		psi_bug = 1;
814 	}
815 
816 	task->psi_flags &= ~clear;
817 	task->psi_flags |= set;
818 }
819 
psi_task_change(struct task_struct * task,int clear,int set)820 void psi_task_change(struct task_struct *task, int clear, int set)
821 {
822 	int cpu = task_cpu(task);
823 	struct psi_group *group;
824 	u64 now;
825 
826 	if (!task->pid)
827 		return;
828 
829 	psi_flags_change(task, clear, set);
830 
831 	now = cpu_clock(cpu);
832 
833 	group = task_psi_group(task);
834 	do {
835 		psi_group_change(group, cpu, clear, set, now, true);
836 	} while ((group = group->parent));
837 }
838 
psi_task_switch(struct task_struct * prev,struct task_struct * next,bool sleep)839 void psi_task_switch(struct task_struct *prev, struct task_struct *next,
840 		     bool sleep)
841 {
842 	struct psi_group *group, *common = NULL;
843 	int cpu = task_cpu(prev);
844 	u64 now = cpu_clock(cpu);
845 
846 	if (next->pid) {
847 		psi_flags_change(next, 0, TSK_ONCPU);
848 		/*
849 		 * Set TSK_ONCPU on @next's cgroups. If @next shares any
850 		 * ancestors with @prev, those will already have @prev's
851 		 * TSK_ONCPU bit set, and we can stop the iteration there.
852 		 */
853 		group = task_psi_group(next);
854 		do {
855 			if (per_cpu_ptr(group->pcpu, cpu)->state_mask &
856 			    PSI_ONCPU) {
857 				common = group;
858 				break;
859 			}
860 
861 			psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
862 		} while ((group = group->parent));
863 	}
864 
865 	if (prev->pid) {
866 		int clear = TSK_ONCPU, set = 0;
867 		bool wake_clock = true;
868 
869 		/*
870 		 * When we're going to sleep, psi_dequeue() lets us
871 		 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
872 		 * TSK_IOWAIT here, where we can combine it with
873 		 * TSK_ONCPU and save walking common ancestors twice.
874 		 */
875 		if (sleep) {
876 			clear |= TSK_RUNNING;
877 			if (prev->in_memstall)
878 				clear |= TSK_MEMSTALL_RUNNING;
879 			if (prev->in_iowait)
880 				set |= TSK_IOWAIT;
881 
882 			/*
883 			 * Periodic aggregation shuts off if there is a period of no
884 			 * task changes, so we wake it back up if necessary. However,
885 			 * don't do this if the task change is the aggregation worker
886 			 * itself going to sleep, or we'll ping-pong forever.
887 			 */
888 			if (unlikely((prev->flags & PF_WQ_WORKER) &&
889 				     wq_worker_last_func(prev) == psi_avgs_work))
890 				wake_clock = false;
891 		}
892 
893 		psi_flags_change(prev, clear, set);
894 
895 		group = task_psi_group(prev);
896 		do {
897 			if (group == common)
898 				break;
899 			psi_group_change(group, cpu, clear, set, now, wake_clock);
900 		} while ((group = group->parent));
901 
902 		/*
903 		 * TSK_ONCPU is handled up to the common ancestor. If there are
904 		 * any other differences between the two tasks (e.g. prev goes
905 		 * to sleep, or only one task is memstall), finish propagating
906 		 * those differences all the way up to the root.
907 		 */
908 		if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) {
909 			clear &= ~TSK_ONCPU;
910 			for (; group; group = group->parent)
911 				psi_group_change(group, cpu, clear, set, now, wake_clock);
912 		}
913 	}
914 }
915 
916 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
psi_account_irqtime(struct task_struct * task,u32 delta)917 void psi_account_irqtime(struct task_struct *task, u32 delta)
918 {
919 	int cpu = task_cpu(task);
920 	struct psi_group *group;
921 	struct psi_group_cpu *groupc;
922 	u64 now;
923 
924 	if (!task->pid)
925 		return;
926 
927 	now = cpu_clock(cpu);
928 
929 	group = task_psi_group(task);
930 	do {
931 		if (!group->enabled)
932 			continue;
933 
934 		groupc = per_cpu_ptr(group->pcpu, cpu);
935 
936 		write_seqcount_begin(&groupc->seq);
937 
938 		record_times(groupc, now);
939 		groupc->times[PSI_IRQ_FULL] += delta;
940 
941 		write_seqcount_end(&groupc->seq);
942 
943 		if (group->poll_states & (1 << PSI_IRQ_FULL))
944 			psi_schedule_poll_work(group, 1);
945 	} while ((group = group->parent));
946 }
947 #endif
948 
949 /**
950  * psi_memstall_enter - mark the beginning of a memory stall section
951  * @flags: flags to handle nested sections
952  *
953  * Marks the calling task as being stalled due to a lack of memory,
954  * such as waiting for a refault or performing reclaim.
955  */
psi_memstall_enter(unsigned long * flags)956 void psi_memstall_enter(unsigned long *flags)
957 {
958 	struct rq_flags rf;
959 	struct rq *rq;
960 
961 	if (static_branch_likely(&psi_disabled))
962 		return;
963 
964 	*flags = current->in_memstall;
965 	if (*flags)
966 		return;
967 	/*
968 	 * in_memstall setting & accounting needs to be atomic wrt
969 	 * changes to the task's scheduling state, otherwise we can
970 	 * race with CPU migration.
971 	 */
972 	rq = this_rq_lock_irq(&rf);
973 
974 	current->in_memstall = 1;
975 	psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
976 
977 	rq_unlock_irq(rq, &rf);
978 }
979 EXPORT_SYMBOL_GPL(psi_memstall_enter);
980 
981 /**
982  * psi_memstall_leave - mark the end of an memory stall section
983  * @flags: flags to handle nested memdelay sections
984  *
985  * Marks the calling task as no longer stalled due to lack of memory.
986  */
psi_memstall_leave(unsigned long * flags)987 void psi_memstall_leave(unsigned long *flags)
988 {
989 	struct rq_flags rf;
990 	struct rq *rq;
991 
992 	if (static_branch_likely(&psi_disabled))
993 		return;
994 
995 	if (*flags)
996 		return;
997 	/*
998 	 * in_memstall clearing & accounting needs to be atomic wrt
999 	 * changes to the task's scheduling state, otherwise we could
1000 	 * race with CPU migration.
1001 	 */
1002 	rq = this_rq_lock_irq(&rf);
1003 
1004 	current->in_memstall = 0;
1005 	psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0);
1006 
1007 	rq_unlock_irq(rq, &rf);
1008 }
1009 EXPORT_SYMBOL_GPL(psi_memstall_leave);
1010 
1011 #ifdef CONFIG_CGROUPS
psi_cgroup_alloc(struct cgroup * cgroup)1012 int psi_cgroup_alloc(struct cgroup *cgroup)
1013 {
1014 	if (!static_branch_likely(&psi_cgroups_enabled))
1015 		return 0;
1016 
1017 	cgroup->psi = kzalloc(sizeof(struct psi_group), GFP_KERNEL);
1018 	if (!cgroup->psi)
1019 		return -ENOMEM;
1020 
1021 	cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
1022 	if (!cgroup->psi->pcpu) {
1023 		kfree(cgroup->psi);
1024 		return -ENOMEM;
1025 	}
1026 	group_init(cgroup->psi);
1027 	cgroup->psi->parent = cgroup_psi(cgroup_parent(cgroup));
1028 	return 0;
1029 }
1030 
psi_cgroup_free(struct cgroup * cgroup)1031 void psi_cgroup_free(struct cgroup *cgroup)
1032 {
1033 	if (!static_branch_likely(&psi_cgroups_enabled))
1034 		return;
1035 
1036 	cancel_delayed_work_sync(&cgroup->psi->avgs_work);
1037 	free_percpu(cgroup->psi->pcpu);
1038 	/* All triggers must be removed by now */
1039 	WARN_ONCE(cgroup->psi->poll_states, "psi: trigger leak\n");
1040 	kfree(cgroup->psi);
1041 }
1042 
1043 /**
1044  * cgroup_move_task - move task to a different cgroup
1045  * @task: the task
1046  * @to: the target css_set
1047  *
1048  * Move task to a new cgroup and safely migrate its associated stall
1049  * state between the different groups.
1050  *
1051  * This function acquires the task's rq lock to lock out concurrent
1052  * changes to the task's scheduling state and - in case the task is
1053  * running - concurrent changes to its stall state.
1054  */
cgroup_move_task(struct task_struct * task,struct css_set * to)1055 void cgroup_move_task(struct task_struct *task, struct css_set *to)
1056 {
1057 	unsigned int task_flags;
1058 	struct rq_flags rf;
1059 	struct rq *rq;
1060 
1061 	if (!static_branch_likely(&psi_cgroups_enabled)) {
1062 		/*
1063 		 * Lame to do this here, but the scheduler cannot be locked
1064 		 * from the outside, so we move cgroups from inside sched/.
1065 		 */
1066 		rcu_assign_pointer(task->cgroups, to);
1067 		return;
1068 	}
1069 
1070 	rq = task_rq_lock(task, &rf);
1071 
1072 	/*
1073 	 * We may race with schedule() dropping the rq lock between
1074 	 * deactivating prev and switching to next. Because the psi
1075 	 * updates from the deactivation are deferred to the switch
1076 	 * callback to save cgroup tree updates, the task's scheduling
1077 	 * state here is not coherent with its psi state:
1078 	 *
1079 	 * schedule()                   cgroup_move_task()
1080 	 *   rq_lock()
1081 	 *   deactivate_task()
1082 	 *     p->on_rq = 0
1083 	 *     psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1084 	 *   pick_next_task()
1085 	 *     rq_unlock()
1086 	 *                                rq_lock()
1087 	 *                                psi_task_change() // old cgroup
1088 	 *                                task->cgroups = to
1089 	 *                                psi_task_change() // new cgroup
1090 	 *                                rq_unlock()
1091 	 *     rq_lock()
1092 	 *   psi_sched_switch() // does deferred updates in new cgroup
1093 	 *
1094 	 * Don't rely on the scheduling state. Use psi_flags instead.
1095 	 */
1096 	task_flags = task->psi_flags;
1097 
1098 	if (task_flags)
1099 		psi_task_change(task, task_flags, 0);
1100 
1101 	/* See comment above */
1102 	rcu_assign_pointer(task->cgroups, to);
1103 
1104 	if (task_flags)
1105 		psi_task_change(task, 0, task_flags);
1106 
1107 	task_rq_unlock(rq, task, &rf);
1108 }
1109 
psi_cgroup_restart(struct psi_group * group)1110 void psi_cgroup_restart(struct psi_group *group)
1111 {
1112 	int cpu;
1113 
1114 	/*
1115 	 * After we disable psi_group->enabled, we don't actually
1116 	 * stop percpu tasks accounting in each psi_group_cpu,
1117 	 * instead only stop test_state() loop, record_times()
1118 	 * and averaging worker, see psi_group_change() for details.
1119 	 *
1120 	 * When disable cgroup PSI, this function has nothing to sync
1121 	 * since cgroup pressure files are hidden and percpu psi_group_cpu
1122 	 * would see !psi_group->enabled and only do task accounting.
1123 	 *
1124 	 * When re-enable cgroup PSI, this function use psi_group_change()
1125 	 * to get correct state mask from test_state() loop on tasks[],
1126 	 * and restart groupc->state_start from now, use .clear = .set = 0
1127 	 * here since no task status really changed.
1128 	 */
1129 	if (!group->enabled)
1130 		return;
1131 
1132 	for_each_possible_cpu(cpu) {
1133 		struct rq *rq = cpu_rq(cpu);
1134 		struct rq_flags rf;
1135 		u64 now;
1136 
1137 		rq_lock_irq(rq, &rf);
1138 		now = cpu_clock(cpu);
1139 		psi_group_change(group, cpu, 0, 0, now, true);
1140 		rq_unlock_irq(rq, &rf);
1141 	}
1142 }
1143 #endif /* CONFIG_CGROUPS */
1144 
psi_show(struct seq_file * m,struct psi_group * group,enum psi_res res)1145 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1146 {
1147 	bool only_full = false;
1148 	int full;
1149 	u64 now;
1150 
1151 	if (static_branch_likely(&psi_disabled))
1152 		return -EOPNOTSUPP;
1153 
1154 	/* Update averages before reporting them */
1155 	mutex_lock(&group->avgs_lock);
1156 	now = sched_clock();
1157 	collect_percpu_times(group, PSI_AVGS, NULL);
1158 	if (now >= group->avg_next_update)
1159 		group->avg_next_update = update_averages(group, now);
1160 	mutex_unlock(&group->avgs_lock);
1161 
1162 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1163 	only_full = res == PSI_IRQ;
1164 #endif
1165 
1166 	for (full = 0; full < 2 - only_full; full++) {
1167 		unsigned long avg[3] = { 0, };
1168 		u64 total = 0;
1169 		int w;
1170 
1171 		/* CPU FULL is undefined at the system level */
1172 		if (!(group == &psi_system && res == PSI_CPU && full)) {
1173 			for (w = 0; w < 3; w++)
1174 				avg[w] = group->avg[res * 2 + full][w];
1175 			total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1176 					NSEC_PER_USEC);
1177 		}
1178 
1179 		seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1180 			   full || only_full ? "full" : "some",
1181 			   LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1182 			   LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1183 			   LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1184 			   total);
1185 	}
1186 
1187 	return 0;
1188 }
1189 
psi_trigger_create(struct psi_group * group,char * buf,enum psi_res res)1190 struct psi_trigger *psi_trigger_create(struct psi_group *group,
1191 			char *buf, enum psi_res res)
1192 {
1193 	struct psi_trigger *t;
1194 	enum psi_states state;
1195 	u32 threshold_us;
1196 	u32 window_us;
1197 
1198 	if (static_branch_likely(&psi_disabled))
1199 		return ERR_PTR(-EOPNOTSUPP);
1200 
1201 	if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1202 		state = PSI_IO_SOME + res * 2;
1203 	else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1204 		state = PSI_IO_FULL + res * 2;
1205 	else
1206 		return ERR_PTR(-EINVAL);
1207 
1208 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1209 	if (res == PSI_IRQ && --state != PSI_IRQ_FULL)
1210 		return ERR_PTR(-EINVAL);
1211 #endif
1212 
1213 	if (state >= PSI_NONIDLE)
1214 		return ERR_PTR(-EINVAL);
1215 
1216 	if (window_us < WINDOW_MIN_US ||
1217 		window_us > WINDOW_MAX_US)
1218 		return ERR_PTR(-EINVAL);
1219 
1220 	/* Check threshold */
1221 	if (threshold_us == 0 || threshold_us > window_us)
1222 		return ERR_PTR(-EINVAL);
1223 
1224 	t = kmalloc(sizeof(*t), GFP_KERNEL);
1225 	if (!t)
1226 		return ERR_PTR(-ENOMEM);
1227 
1228 	t->group = group;
1229 	t->state = state;
1230 	t->threshold = threshold_us * NSEC_PER_USEC;
1231 	t->win.size = window_us * NSEC_PER_USEC;
1232 	window_reset(&t->win, sched_clock(),
1233 			group->total[PSI_POLL][t->state], 0);
1234 
1235 	t->event = 0;
1236 	t->last_event_time = 0;
1237 	init_waitqueue_head(&t->event_wait);
1238 	t->pending_event = false;
1239 
1240 	mutex_lock(&group->trigger_lock);
1241 
1242 	if (!rcu_access_pointer(group->poll_task)) {
1243 		struct task_struct *task;
1244 
1245 		task = kthread_create(psi_poll_worker, group, "psimon");
1246 		if (IS_ERR(task)) {
1247 			kfree(t);
1248 			mutex_unlock(&group->trigger_lock);
1249 			return ERR_CAST(task);
1250 		}
1251 		atomic_set(&group->poll_wakeup, 0);
1252 		wake_up_process(task);
1253 		rcu_assign_pointer(group->poll_task, task);
1254 	}
1255 
1256 	list_add(&t->node, &group->triggers);
1257 	group->poll_min_period = min(group->poll_min_period,
1258 		div_u64(t->win.size, UPDATES_PER_WINDOW));
1259 	group->nr_triggers[t->state]++;
1260 	group->poll_states |= (1 << t->state);
1261 
1262 	mutex_unlock(&group->trigger_lock);
1263 
1264 	return t;
1265 }
1266 
psi_trigger_destroy(struct psi_trigger * t)1267 void psi_trigger_destroy(struct psi_trigger *t)
1268 {
1269 	struct psi_group *group;
1270 	struct task_struct *task_to_destroy = NULL;
1271 
1272 	/*
1273 	 * We do not check psi_disabled since it might have been disabled after
1274 	 * the trigger got created.
1275 	 */
1276 	if (!t)
1277 		return;
1278 
1279 	group = t->group;
1280 	/*
1281 	 * Wakeup waiters to stop polling. Can happen if cgroup is deleted
1282 	 * from under a polling process.
1283 	 */
1284 	wake_up_interruptible(&t->event_wait);
1285 
1286 	mutex_lock(&group->trigger_lock);
1287 
1288 	if (!list_empty(&t->node)) {
1289 		struct psi_trigger *tmp;
1290 		u64 period = ULLONG_MAX;
1291 
1292 		list_del(&t->node);
1293 		group->nr_triggers[t->state]--;
1294 		if (!group->nr_triggers[t->state])
1295 			group->poll_states &= ~(1 << t->state);
1296 		/* reset min update period for the remaining triggers */
1297 		list_for_each_entry(tmp, &group->triggers, node)
1298 			period = min(period, div_u64(tmp->win.size,
1299 					UPDATES_PER_WINDOW));
1300 		group->poll_min_period = period;
1301 		/* Destroy poll_task when the last trigger is destroyed */
1302 		if (group->poll_states == 0) {
1303 			group->polling_until = 0;
1304 			task_to_destroy = rcu_dereference_protected(
1305 					group->poll_task,
1306 					lockdep_is_held(&group->trigger_lock));
1307 			rcu_assign_pointer(group->poll_task, NULL);
1308 			del_timer(&group->poll_timer);
1309 		}
1310 	}
1311 
1312 	mutex_unlock(&group->trigger_lock);
1313 
1314 	/*
1315 	 * Wait for psi_schedule_poll_work RCU to complete its read-side
1316 	 * critical section before destroying the trigger and optionally the
1317 	 * poll_task.
1318 	 */
1319 	synchronize_rcu();
1320 	/*
1321 	 * Stop kthread 'psimon' after releasing trigger_lock to prevent a
1322 	 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1323 	 */
1324 	if (task_to_destroy) {
1325 		/*
1326 		 * After the RCU grace period has expired, the worker
1327 		 * can no longer be found through group->poll_task.
1328 		 */
1329 		kthread_stop(task_to_destroy);
1330 	}
1331 	kfree(t);
1332 }
1333 
psi_trigger_poll(void ** trigger_ptr,struct file * file,poll_table * wait)1334 __poll_t psi_trigger_poll(void **trigger_ptr,
1335 				struct file *file, poll_table *wait)
1336 {
1337 	__poll_t ret = DEFAULT_POLLMASK;
1338 	struct psi_trigger *t;
1339 
1340 	if (static_branch_likely(&psi_disabled))
1341 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1342 
1343 	t = smp_load_acquire(trigger_ptr);
1344 	if (!t)
1345 		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1346 
1347 	poll_wait(file, &t->event_wait, wait);
1348 
1349 	if (cmpxchg(&t->event, 1, 0) == 1)
1350 		ret |= EPOLLPRI;
1351 
1352 	return ret;
1353 }
1354 
1355 #ifdef CONFIG_PROC_FS
psi_io_show(struct seq_file * m,void * v)1356 static int psi_io_show(struct seq_file *m, void *v)
1357 {
1358 	return psi_show(m, &psi_system, PSI_IO);
1359 }
1360 
psi_memory_show(struct seq_file * m,void * v)1361 static int psi_memory_show(struct seq_file *m, void *v)
1362 {
1363 	return psi_show(m, &psi_system, PSI_MEM);
1364 }
1365 
psi_cpu_show(struct seq_file * m,void * v)1366 static int psi_cpu_show(struct seq_file *m, void *v)
1367 {
1368 	return psi_show(m, &psi_system, PSI_CPU);
1369 }
1370 
psi_open(struct file * file,int (* psi_show)(struct seq_file *,void *))1371 static int psi_open(struct file *file, int (*psi_show)(struct seq_file *, void *))
1372 {
1373 	if (file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE))
1374 		return -EPERM;
1375 
1376 	return single_open(file, psi_show, NULL);
1377 }
1378 
psi_io_open(struct inode * inode,struct file * file)1379 static int psi_io_open(struct inode *inode, struct file *file)
1380 {
1381 	return psi_open(file, psi_io_show);
1382 }
1383 
psi_memory_open(struct inode * inode,struct file * file)1384 static int psi_memory_open(struct inode *inode, struct file *file)
1385 {
1386 	return psi_open(file, psi_memory_show);
1387 }
1388 
psi_cpu_open(struct inode * inode,struct file * file)1389 static int psi_cpu_open(struct inode *inode, struct file *file)
1390 {
1391 	return psi_open(file, psi_cpu_show);
1392 }
1393 
psi_write(struct file * file,const char __user * user_buf,size_t nbytes,enum psi_res res)1394 static ssize_t psi_write(struct file *file, const char __user *user_buf,
1395 			 size_t nbytes, enum psi_res res)
1396 {
1397 	char buf[32];
1398 	size_t buf_size;
1399 	struct seq_file *seq;
1400 	struct psi_trigger *new;
1401 
1402 	if (static_branch_likely(&psi_disabled))
1403 		return -EOPNOTSUPP;
1404 
1405 	if (!nbytes)
1406 		return -EINVAL;
1407 
1408 	buf_size = min(nbytes, sizeof(buf));
1409 	if (copy_from_user(buf, user_buf, buf_size))
1410 		return -EFAULT;
1411 
1412 	buf[buf_size - 1] = '\0';
1413 
1414 	seq = file->private_data;
1415 
1416 	/* Take seq->lock to protect seq->private from concurrent writes */
1417 	mutex_lock(&seq->lock);
1418 
1419 	/* Allow only one trigger per file descriptor */
1420 	if (seq->private) {
1421 		mutex_unlock(&seq->lock);
1422 		return -EBUSY;
1423 	}
1424 
1425 	new = psi_trigger_create(&psi_system, buf, res);
1426 	if (IS_ERR(new)) {
1427 		mutex_unlock(&seq->lock);
1428 		return PTR_ERR(new);
1429 	}
1430 
1431 	smp_store_release(&seq->private, new);
1432 	mutex_unlock(&seq->lock);
1433 
1434 	return nbytes;
1435 }
1436 
psi_io_write(struct file * file,const char __user * user_buf,size_t nbytes,loff_t * ppos)1437 static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1438 			    size_t nbytes, loff_t *ppos)
1439 {
1440 	return psi_write(file, user_buf, nbytes, PSI_IO);
1441 }
1442 
psi_memory_write(struct file * file,const char __user * user_buf,size_t nbytes,loff_t * ppos)1443 static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1444 				size_t nbytes, loff_t *ppos)
1445 {
1446 	return psi_write(file, user_buf, nbytes, PSI_MEM);
1447 }
1448 
psi_cpu_write(struct file * file,const char __user * user_buf,size_t nbytes,loff_t * ppos)1449 static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1450 			     size_t nbytes, loff_t *ppos)
1451 {
1452 	return psi_write(file, user_buf, nbytes, PSI_CPU);
1453 }
1454 
psi_fop_poll(struct file * file,poll_table * wait)1455 static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1456 {
1457 	struct seq_file *seq = file->private_data;
1458 
1459 	return psi_trigger_poll(&seq->private, file, wait);
1460 }
1461 
psi_fop_release(struct inode * inode,struct file * file)1462 static int psi_fop_release(struct inode *inode, struct file *file)
1463 {
1464 	struct seq_file *seq = file->private_data;
1465 
1466 	psi_trigger_destroy(seq->private);
1467 	return single_release(inode, file);
1468 }
1469 
1470 static const struct proc_ops psi_io_proc_ops = {
1471 	.proc_open	= psi_io_open,
1472 	.proc_read	= seq_read,
1473 	.proc_lseek	= seq_lseek,
1474 	.proc_write	= psi_io_write,
1475 	.proc_poll	= psi_fop_poll,
1476 	.proc_release	= psi_fop_release,
1477 };
1478 
1479 static const struct proc_ops psi_memory_proc_ops = {
1480 	.proc_open	= psi_memory_open,
1481 	.proc_read	= seq_read,
1482 	.proc_lseek	= seq_lseek,
1483 	.proc_write	= psi_memory_write,
1484 	.proc_poll	= psi_fop_poll,
1485 	.proc_release	= psi_fop_release,
1486 };
1487 
1488 static const struct proc_ops psi_cpu_proc_ops = {
1489 	.proc_open	= psi_cpu_open,
1490 	.proc_read	= seq_read,
1491 	.proc_lseek	= seq_lseek,
1492 	.proc_write	= psi_cpu_write,
1493 	.proc_poll	= psi_fop_poll,
1494 	.proc_release	= psi_fop_release,
1495 };
1496 
1497 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
psi_irq_show(struct seq_file * m,void * v)1498 static int psi_irq_show(struct seq_file *m, void *v)
1499 {
1500 	return psi_show(m, &psi_system, PSI_IRQ);
1501 }
1502 
psi_irq_open(struct inode * inode,struct file * file)1503 static int psi_irq_open(struct inode *inode, struct file *file)
1504 {
1505 	return psi_open(file, psi_irq_show);
1506 }
1507 
psi_irq_write(struct file * file,const char __user * user_buf,size_t nbytes,loff_t * ppos)1508 static ssize_t psi_irq_write(struct file *file, const char __user *user_buf,
1509 			     size_t nbytes, loff_t *ppos)
1510 {
1511 	return psi_write(file, user_buf, nbytes, PSI_IRQ);
1512 }
1513 
1514 static const struct proc_ops psi_irq_proc_ops = {
1515 	.proc_open	= psi_irq_open,
1516 	.proc_read	= seq_read,
1517 	.proc_lseek	= seq_lseek,
1518 	.proc_write	= psi_irq_write,
1519 	.proc_poll	= psi_fop_poll,
1520 	.proc_release	= psi_fop_release,
1521 };
1522 #endif
1523 
psi_proc_init(void)1524 static int __init psi_proc_init(void)
1525 {
1526 	if (psi_enable) {
1527 		proc_mkdir("pressure", NULL);
1528 		proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops);
1529 		proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops);
1530 		proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops);
1531 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1532 		proc_create("pressure/irq", 0666, NULL, &psi_irq_proc_ops);
1533 #endif
1534 	}
1535 	return 0;
1536 }
1537 module_init(psi_proc_init);
1538 
1539 #endif /* CONFIG_PROC_FS */
1540