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