Loading...
1/*
2 * Pressure stall information for CPU, memory and IO
3 *
4 * Copyright (c) 2018 Facebook, Inc.
5 * Author: Johannes Weiner <hannes@cmpxchg.org>
6 *
7 * Polling support by Suren Baghdasaryan <surenb@google.com>
8 * Copyright (c) 2018 Google, Inc.
9 *
10 * When CPU, memory and IO are contended, tasks experience delays that
11 * reduce throughput and introduce latencies into the workload. Memory
12 * and IO contention, in addition, can cause a full loss of forward
13 * progress in which the CPU goes idle.
14 *
15 * This code aggregates individual task delays into resource pressure
16 * metrics that indicate problems with both workload health and
17 * resource utilization.
18 *
19 * Model
20 *
21 * The time in which a task can execute on a CPU is our baseline for
22 * productivity. Pressure expresses the amount of time in which this
23 * potential cannot be realized due to resource contention.
24 *
25 * This concept of productivity has two components: the workload and
26 * the CPU. To measure the impact of pressure on both, we define two
27 * contention states for a resource: SOME and FULL.
28 *
29 * In the SOME state of a given resource, one or more tasks are
30 * delayed on that resource. This affects the workload's ability to
31 * perform work, but the CPU may still be executing other tasks.
32 *
33 * In the FULL state of a given resource, all non-idle tasks are
34 * delayed on that resource such that nobody is advancing and the CPU
35 * goes idle. This leaves both workload and CPU unproductive.
36 *
37 * (Naturally, the FULL state doesn't exist for the CPU resource.)
38 *
39 * SOME = nr_delayed_tasks != 0
40 * FULL = nr_delayed_tasks != 0 && nr_running_tasks == 0
41 *
42 * The percentage of wallclock time spent in those compound stall
43 * states gives pressure numbers between 0 and 100 for each resource,
44 * where the SOME percentage indicates workload slowdowns and the FULL
45 * percentage indicates reduced CPU utilization:
46 *
47 * %SOME = time(SOME) / period
48 * %FULL = time(FULL) / period
49 *
50 * Multiple CPUs
51 *
52 * The more tasks and available CPUs there are, the more work can be
53 * performed concurrently. This means that the potential that can go
54 * unrealized due to resource contention *also* scales with non-idle
55 * tasks and CPUs.
56 *
57 * Consider a scenario where 257 number crunching tasks are trying to
58 * run concurrently on 256 CPUs. If we simply aggregated the task
59 * states, we would have to conclude a CPU SOME pressure number of
60 * 100%, since *somebody* is waiting on a runqueue at all
61 * times. However, that is clearly not the amount of contention the
62 * workload is experiencing: only one out of 256 possible exceution
63 * threads will be contended at any given time, or about 0.4%.
64 *
65 * Conversely, consider a scenario of 4 tasks and 4 CPUs where at any
66 * given time *one* of the tasks is delayed due to a lack of memory.
67 * Again, looking purely at the task state would yield a memory FULL
68 * pressure number of 0%, since *somebody* is always making forward
69 * progress. But again this wouldn't capture the amount of execution
70 * potential lost, which is 1 out of 4 CPUs, or 25%.
71 *
72 * To calculate wasted potential (pressure) with multiple processors,
73 * we have to base our calculation on the number of non-idle tasks in
74 * conjunction with the number of available CPUs, which is the number
75 * of potential execution threads. SOME becomes then the proportion of
76 * delayed tasks to possibe threads, and FULL is the share of possible
77 * threads that are unproductive due to delays:
78 *
79 * threads = min(nr_nonidle_tasks, nr_cpus)
80 * SOME = min(nr_delayed_tasks / threads, 1)
81 * FULL = (threads - min(nr_running_tasks, threads)) / threads
82 *
83 * For the 257 number crunchers on 256 CPUs, this yields:
84 *
85 * threads = min(257, 256)
86 * SOME = min(1 / 256, 1) = 0.4%
87 * FULL = (256 - min(257, 256)) / 256 = 0%
88 *
89 * For the 1 out of 4 memory-delayed tasks, this yields:
90 *
91 * threads = min(4, 4)
92 * SOME = min(1 / 4, 1) = 25%
93 * FULL = (4 - min(3, 4)) / 4 = 25%
94 *
95 * [ Substitute nr_cpus with 1, and you can see that it's a natural
96 * extension of the single-CPU model. ]
97 *
98 * Implementation
99 *
100 * To assess the precise time spent in each such state, we would have
101 * to freeze the system on task changes and start/stop the state
102 * clocks accordingly. Obviously that doesn't scale in practice.
103 *
104 * Because the scheduler aims to distribute the compute load evenly
105 * among the available CPUs, we can track task state locally to each
106 * CPU and, at much lower frequency, extrapolate the global state for
107 * the cumulative stall times and the running averages.
108 *
109 * For each runqueue, we track:
110 *
111 * tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0)
112 * tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_running_tasks[cpu])
113 * tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0)
114 *
115 * and then periodically aggregate:
116 *
117 * tNONIDLE = sum(tNONIDLE[i])
118 *
119 * tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE
120 * tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE
121 *
122 * %SOME = tSOME / period
123 * %FULL = tFULL / period
124 *
125 * This gives us an approximation of pressure that is practical
126 * cost-wise, yet way more sensitive and accurate than periodic
127 * sampling of the aggregate task states would be.
128 */
129
130#include "../workqueue_internal.h"
131#include <linux/sched/loadavg.h>
132#include <linux/seq_file.h>
133#include <linux/proc_fs.h>
134#include <linux/seqlock.h>
135#include <linux/uaccess.h>
136#include <linux/cgroup.h>
137#include <linux/module.h>
138#include <linux/sched.h>
139#include <linux/ctype.h>
140#include <linux/file.h>
141#include <linux/poll.h>
142#include <linux/psi.h>
143#include "sched.h"
144
145static int psi_bug __read_mostly;
146
147DEFINE_STATIC_KEY_FALSE(psi_disabled);
148
149#ifdef CONFIG_PSI_DEFAULT_DISABLED
150static bool psi_enable;
151#else
152static bool psi_enable = true;
153#endif
154static int __init setup_psi(char *str)
155{
156 return kstrtobool(str, &psi_enable) == 0;
157}
158__setup("psi=", setup_psi);
159
160/* Running averages - we need to be higher-res than loadavg */
161#define PSI_FREQ (2*HZ+1) /* 2 sec intervals */
162#define EXP_10s 1677 /* 1/exp(2s/10s) as fixed-point */
163#define EXP_60s 1981 /* 1/exp(2s/60s) */
164#define EXP_300s 2034 /* 1/exp(2s/300s) */
165
166/* PSI trigger definitions */
167#define WINDOW_MIN_US 500000 /* Min window size is 500ms */
168#define WINDOW_MAX_US 10000000 /* Max window size is 10s */
169#define UPDATES_PER_WINDOW 10 /* 10 updates per window */
170
171/* Sampling frequency in nanoseconds */
172static u64 psi_period __read_mostly;
173
174/* System-level pressure and stall tracking */
175static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
176struct psi_group psi_system = {
177 .pcpu = &system_group_pcpu,
178};
179
180static void psi_avgs_work(struct work_struct *work);
181
182static void group_init(struct psi_group *group)
183{
184 int cpu;
185
186 for_each_possible_cpu(cpu)
187 seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);
188 group->avg_next_update = sched_clock() + psi_period;
189 INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
190 mutex_init(&group->avgs_lock);
191 /* Init trigger-related members */
192 atomic_set(&group->poll_scheduled, 0);
193 mutex_init(&group->trigger_lock);
194 INIT_LIST_HEAD(&group->triggers);
195 memset(group->nr_triggers, 0, sizeof(group->nr_triggers));
196 group->poll_states = 0;
197 group->poll_min_period = U32_MAX;
198 memset(group->polling_total, 0, sizeof(group->polling_total));
199 group->polling_next_update = ULLONG_MAX;
200 group->polling_until = 0;
201 rcu_assign_pointer(group->poll_kworker, NULL);
202}
203
204void __init psi_init(void)
205{
206 if (!psi_enable) {
207 static_branch_enable(&psi_disabled);
208 return;
209 }
210
211 psi_period = jiffies_to_nsecs(PSI_FREQ);
212 group_init(&psi_system);
213}
214
215static bool test_state(unsigned int *tasks, enum psi_states state)
216{
217 switch (state) {
218 case PSI_IO_SOME:
219 return tasks[NR_IOWAIT];
220 case PSI_IO_FULL:
221 return tasks[NR_IOWAIT] && !tasks[NR_RUNNING];
222 case PSI_MEM_SOME:
223 return tasks[NR_MEMSTALL];
224 case PSI_MEM_FULL:
225 return tasks[NR_MEMSTALL] && !tasks[NR_RUNNING];
226 case PSI_CPU_SOME:
227 return tasks[NR_RUNNING] > 1;
228 case PSI_NONIDLE:
229 return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
230 tasks[NR_RUNNING];
231 default:
232 return false;
233 }
234}
235
236static void get_recent_times(struct psi_group *group, int cpu,
237 enum psi_aggregators aggregator, u32 *times,
238 u32 *pchanged_states)
239{
240 struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
241 u64 now, state_start;
242 enum psi_states s;
243 unsigned int seq;
244 u32 state_mask;
245
246 *pchanged_states = 0;
247
248 /* Snapshot a coherent view of the CPU state */
249 do {
250 seq = read_seqcount_begin(&groupc->seq);
251 now = cpu_clock(cpu);
252 memcpy(times, groupc->times, sizeof(groupc->times));
253 state_mask = groupc->state_mask;
254 state_start = groupc->state_start;
255 } while (read_seqcount_retry(&groupc->seq, seq));
256
257 /* Calculate state time deltas against the previous snapshot */
258 for (s = 0; s < NR_PSI_STATES; s++) {
259 u32 delta;
260 /*
261 * In addition to already concluded states, we also
262 * incorporate currently active states on the CPU,
263 * since states may last for many sampling periods.
264 *
265 * This way we keep our delta sampling buckets small
266 * (u32) and our reported pressure close to what's
267 * actually happening.
268 */
269 if (state_mask & (1 << s))
270 times[s] += now - state_start;
271
272 delta = times[s] - groupc->times_prev[aggregator][s];
273 groupc->times_prev[aggregator][s] = times[s];
274
275 times[s] = delta;
276 if (delta)
277 *pchanged_states |= (1 << s);
278 }
279}
280
281static void calc_avgs(unsigned long avg[3], int missed_periods,
282 u64 time, u64 period)
283{
284 unsigned long pct;
285
286 /* Fill in zeroes for periods of no activity */
287 if (missed_periods) {
288 avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
289 avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
290 avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
291 }
292
293 /* Sample the most recent active period */
294 pct = div_u64(time * 100, period);
295 pct *= FIXED_1;
296 avg[0] = calc_load(avg[0], EXP_10s, pct);
297 avg[1] = calc_load(avg[1], EXP_60s, pct);
298 avg[2] = calc_load(avg[2], EXP_300s, pct);
299}
300
301static void collect_percpu_times(struct psi_group *group,
302 enum psi_aggregators aggregator,
303 u32 *pchanged_states)
304{
305 u64 deltas[NR_PSI_STATES - 1] = { 0, };
306 unsigned long nonidle_total = 0;
307 u32 changed_states = 0;
308 int cpu;
309 int s;
310
311 /*
312 * Collect the per-cpu time buckets and average them into a
313 * single time sample that is normalized to wallclock time.
314 *
315 * For averaging, each CPU is weighted by its non-idle time in
316 * the sampling period. This eliminates artifacts from uneven
317 * loading, or even entirely idle CPUs.
318 */
319 for_each_possible_cpu(cpu) {
320 u32 times[NR_PSI_STATES];
321 u32 nonidle;
322 u32 cpu_changed_states;
323
324 get_recent_times(group, cpu, aggregator, times,
325 &cpu_changed_states);
326 changed_states |= cpu_changed_states;
327
328 nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
329 nonidle_total += nonidle;
330
331 for (s = 0; s < PSI_NONIDLE; s++)
332 deltas[s] += (u64)times[s] * nonidle;
333 }
334
335 /*
336 * Integrate the sample into the running statistics that are
337 * reported to userspace: the cumulative stall times and the
338 * decaying averages.
339 *
340 * Pressure percentages are sampled at PSI_FREQ. We might be
341 * called more often when the user polls more frequently than
342 * that; we might be called less often when there is no task
343 * activity, thus no data, and clock ticks are sporadic. The
344 * below handles both.
345 */
346
347 /* total= */
348 for (s = 0; s < NR_PSI_STATES - 1; s++)
349 group->total[aggregator][s] +=
350 div_u64(deltas[s], max(nonidle_total, 1UL));
351
352 if (pchanged_states)
353 *pchanged_states = changed_states;
354}
355
356static u64 update_averages(struct psi_group *group, u64 now)
357{
358 unsigned long missed_periods = 0;
359 u64 expires, period;
360 u64 avg_next_update;
361 int s;
362
363 /* avgX= */
364 expires = group->avg_next_update;
365 if (now - expires >= psi_period)
366 missed_periods = div_u64(now - expires, psi_period);
367
368 /*
369 * The periodic clock tick can get delayed for various
370 * reasons, especially on loaded systems. To avoid clock
371 * drift, we schedule the clock in fixed psi_period intervals.
372 * But the deltas we sample out of the per-cpu buckets above
373 * are based on the actual time elapsing between clock ticks.
374 */
375 avg_next_update = expires + ((1 + missed_periods) * psi_period);
376 period = now - (group->avg_last_update + (missed_periods * psi_period));
377 group->avg_last_update = now;
378
379 for (s = 0; s < NR_PSI_STATES - 1; s++) {
380 u32 sample;
381
382 sample = group->total[PSI_AVGS][s] - group->avg_total[s];
383 /*
384 * Due to the lockless sampling of the time buckets,
385 * recorded time deltas can slip into the next period,
386 * which under full pressure can result in samples in
387 * excess of the period length.
388 *
389 * We don't want to report non-sensical pressures in
390 * excess of 100%, nor do we want to drop such events
391 * on the floor. Instead we punt any overage into the
392 * future until pressure subsides. By doing this we
393 * don't underreport the occurring pressure curve, we
394 * just report it delayed by one period length.
395 *
396 * The error isn't cumulative. As soon as another
397 * delta slips from a period P to P+1, by definition
398 * it frees up its time T in P.
399 */
400 if (sample > period)
401 sample = period;
402 group->avg_total[s] += sample;
403 calc_avgs(group->avg[s], missed_periods, sample, period);
404 }
405
406 return avg_next_update;
407}
408
409static void psi_avgs_work(struct work_struct *work)
410{
411 struct delayed_work *dwork;
412 struct psi_group *group;
413 u32 changed_states;
414 bool nonidle;
415 u64 now;
416
417 dwork = to_delayed_work(work);
418 group = container_of(dwork, struct psi_group, avgs_work);
419
420 mutex_lock(&group->avgs_lock);
421
422 now = sched_clock();
423
424 collect_percpu_times(group, PSI_AVGS, &changed_states);
425 nonidle = changed_states & (1 << PSI_NONIDLE);
426 /*
427 * If there is task activity, periodically fold the per-cpu
428 * times and feed samples into the running averages. If things
429 * are idle and there is no data to process, stop the clock.
430 * Once restarted, we'll catch up the running averages in one
431 * go - see calc_avgs() and missed_periods.
432 */
433 if (now >= group->avg_next_update)
434 group->avg_next_update = update_averages(group, now);
435
436 if (nonidle) {
437 schedule_delayed_work(dwork, nsecs_to_jiffies(
438 group->avg_next_update - now) + 1);
439 }
440
441 mutex_unlock(&group->avgs_lock);
442}
443
444/* Trigger tracking window manupulations */
445static void window_reset(struct psi_window *win, u64 now, u64 value,
446 u64 prev_growth)
447{
448 win->start_time = now;
449 win->start_value = value;
450 win->prev_growth = prev_growth;
451}
452
453/*
454 * PSI growth tracking window update and growth calculation routine.
455 *
456 * This approximates a sliding tracking window by interpolating
457 * partially elapsed windows using historical growth data from the
458 * previous intervals. This minimizes memory requirements (by not storing
459 * all the intermediate values in the previous window) and simplifies
460 * the calculations. It works well because PSI signal changes only in
461 * positive direction and over relatively small window sizes the growth
462 * is close to linear.
463 */
464static u64 window_update(struct psi_window *win, u64 now, u64 value)
465{
466 u64 elapsed;
467 u64 growth;
468
469 elapsed = now - win->start_time;
470 growth = value - win->start_value;
471 /*
472 * After each tracking window passes win->start_value and
473 * win->start_time get reset and win->prev_growth stores
474 * the average per-window growth of the previous window.
475 * win->prev_growth is then used to interpolate additional
476 * growth from the previous window assuming it was linear.
477 */
478 if (elapsed > win->size)
479 window_reset(win, now, value, growth);
480 else {
481 u32 remaining;
482
483 remaining = win->size - elapsed;
484 growth += div_u64(win->prev_growth * remaining, win->size);
485 }
486
487 return growth;
488}
489
490static void init_triggers(struct psi_group *group, u64 now)
491{
492 struct psi_trigger *t;
493
494 list_for_each_entry(t, &group->triggers, node)
495 window_reset(&t->win, now,
496 group->total[PSI_POLL][t->state], 0);
497 memcpy(group->polling_total, group->total[PSI_POLL],
498 sizeof(group->polling_total));
499 group->polling_next_update = now + group->poll_min_period;
500}
501
502static u64 update_triggers(struct psi_group *group, u64 now)
503{
504 struct psi_trigger *t;
505 bool new_stall = false;
506 u64 *total = group->total[PSI_POLL];
507
508 /*
509 * On subsequent updates, calculate growth deltas and let
510 * watchers know when their specified thresholds are exceeded.
511 */
512 list_for_each_entry(t, &group->triggers, node) {
513 u64 growth;
514
515 /* Check for stall activity */
516 if (group->polling_total[t->state] == total[t->state])
517 continue;
518
519 /*
520 * Multiple triggers might be looking at the same state,
521 * remember to update group->polling_total[] once we've
522 * been through all of them. Also remember to extend the
523 * polling time if we see new stall activity.
524 */
525 new_stall = true;
526
527 /* Calculate growth since last update */
528 growth = window_update(&t->win, now, total[t->state]);
529 if (growth < t->threshold)
530 continue;
531
532 /* Limit event signaling to once per window */
533 if (now < t->last_event_time + t->win.size)
534 continue;
535
536 /* Generate an event */
537 if (cmpxchg(&t->event, 0, 1) == 0)
538 wake_up_interruptible(&t->event_wait);
539 t->last_event_time = now;
540 }
541
542 if (new_stall)
543 memcpy(group->polling_total, total,
544 sizeof(group->polling_total));
545
546 return now + group->poll_min_period;
547}
548
549/*
550 * Schedule polling if it's not already scheduled. It's safe to call even from
551 * hotpath because even though kthread_queue_delayed_work takes worker->lock
552 * spinlock that spinlock is never contended due to poll_scheduled atomic
553 * preventing such competition.
554 */
555static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay)
556{
557 struct kthread_worker *kworker;
558
559 /* Do not reschedule if already scheduled */
560 if (atomic_cmpxchg(&group->poll_scheduled, 0, 1) != 0)
561 return;
562
563 rcu_read_lock();
564
565 kworker = rcu_dereference(group->poll_kworker);
566 /*
567 * kworker might be NULL in case psi_trigger_destroy races with
568 * psi_task_change (hotpath) which can't use locks
569 */
570 if (likely(kworker))
571 kthread_queue_delayed_work(kworker, &group->poll_work, delay);
572 else
573 atomic_set(&group->poll_scheduled, 0);
574
575 rcu_read_unlock();
576}
577
578static void psi_poll_work(struct kthread_work *work)
579{
580 struct kthread_delayed_work *dwork;
581 struct psi_group *group;
582 u32 changed_states;
583 u64 now;
584
585 dwork = container_of(work, struct kthread_delayed_work, work);
586 group = container_of(dwork, struct psi_group, poll_work);
587
588 atomic_set(&group->poll_scheduled, 0);
589
590 mutex_lock(&group->trigger_lock);
591
592 now = sched_clock();
593
594 collect_percpu_times(group, PSI_POLL, &changed_states);
595
596 if (changed_states & group->poll_states) {
597 /* Initialize trigger windows when entering polling mode */
598 if (now > group->polling_until)
599 init_triggers(group, now);
600
601 /*
602 * Keep the monitor active for at least the duration of the
603 * minimum tracking window as long as monitor states are
604 * changing.
605 */
606 group->polling_until = now +
607 group->poll_min_period * UPDATES_PER_WINDOW;
608 }
609
610 if (now > group->polling_until) {
611 group->polling_next_update = ULLONG_MAX;
612 goto out;
613 }
614
615 if (now >= group->polling_next_update)
616 group->polling_next_update = update_triggers(group, now);
617
618 psi_schedule_poll_work(group,
619 nsecs_to_jiffies(group->polling_next_update - now) + 1);
620
621out:
622 mutex_unlock(&group->trigger_lock);
623}
624
625static void record_times(struct psi_group_cpu *groupc, int cpu,
626 bool memstall_tick)
627{
628 u32 delta;
629 u64 now;
630
631 now = cpu_clock(cpu);
632 delta = now - groupc->state_start;
633 groupc->state_start = now;
634
635 if (groupc->state_mask & (1 << PSI_IO_SOME)) {
636 groupc->times[PSI_IO_SOME] += delta;
637 if (groupc->state_mask & (1 << PSI_IO_FULL))
638 groupc->times[PSI_IO_FULL] += delta;
639 }
640
641 if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
642 groupc->times[PSI_MEM_SOME] += delta;
643 if (groupc->state_mask & (1 << PSI_MEM_FULL))
644 groupc->times[PSI_MEM_FULL] += delta;
645 else if (memstall_tick) {
646 u32 sample;
647 /*
648 * Since we care about lost potential, a
649 * memstall is FULL when there are no other
650 * working tasks, but also when the CPU is
651 * actively reclaiming and nothing productive
652 * could run even if it were runnable.
653 *
654 * When the timer tick sees a reclaiming CPU,
655 * regardless of runnable tasks, sample a FULL
656 * tick (or less if it hasn't been a full tick
657 * since the last state change).
658 */
659 sample = min(delta, (u32)jiffies_to_nsecs(1));
660 groupc->times[PSI_MEM_FULL] += sample;
661 }
662 }
663
664 if (groupc->state_mask & (1 << PSI_CPU_SOME))
665 groupc->times[PSI_CPU_SOME] += delta;
666
667 if (groupc->state_mask & (1 << PSI_NONIDLE))
668 groupc->times[PSI_NONIDLE] += delta;
669}
670
671static u32 psi_group_change(struct psi_group *group, int cpu,
672 unsigned int clear, unsigned int set)
673{
674 struct psi_group_cpu *groupc;
675 unsigned int t, m;
676 enum psi_states s;
677 u32 state_mask = 0;
678
679 groupc = per_cpu_ptr(group->pcpu, cpu);
680
681 /*
682 * First we assess the aggregate resource states this CPU's
683 * tasks have been in since the last change, and account any
684 * SOME and FULL time these may have resulted in.
685 *
686 * Then we update the task counts according to the state
687 * change requested through the @clear and @set bits.
688 */
689 write_seqcount_begin(&groupc->seq);
690
691 record_times(groupc, cpu, false);
692
693 for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
694 if (!(m & (1 << t)))
695 continue;
696 if (groupc->tasks[t] == 0 && !psi_bug) {
697 printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u] clear=%x set=%x\n",
698 cpu, t, groupc->tasks[0],
699 groupc->tasks[1], groupc->tasks[2],
700 clear, set);
701 psi_bug = 1;
702 }
703 groupc->tasks[t]--;
704 }
705
706 for (t = 0; set; set &= ~(1 << t), t++)
707 if (set & (1 << t))
708 groupc->tasks[t]++;
709
710 /* Calculate state mask representing active states */
711 for (s = 0; s < NR_PSI_STATES; s++) {
712 if (test_state(groupc->tasks, s))
713 state_mask |= (1 << s);
714 }
715 groupc->state_mask = state_mask;
716
717 write_seqcount_end(&groupc->seq);
718
719 return state_mask;
720}
721
722static struct psi_group *iterate_groups(struct task_struct *task, void **iter)
723{
724#ifdef CONFIG_CGROUPS
725 struct cgroup *cgroup = NULL;
726
727 if (!*iter)
728 cgroup = task->cgroups->dfl_cgrp;
729 else if (*iter == &psi_system)
730 return NULL;
731 else
732 cgroup = cgroup_parent(*iter);
733
734 if (cgroup && cgroup_parent(cgroup)) {
735 *iter = cgroup;
736 return cgroup_psi(cgroup);
737 }
738#else
739 if (*iter)
740 return NULL;
741#endif
742 *iter = &psi_system;
743 return &psi_system;
744}
745
746void psi_task_change(struct task_struct *task, int clear, int set)
747{
748 int cpu = task_cpu(task);
749 struct psi_group *group;
750 bool wake_clock = true;
751 void *iter = NULL;
752
753 if (!task->pid)
754 return;
755
756 if (((task->psi_flags & set) ||
757 (task->psi_flags & clear) != clear) &&
758 !psi_bug) {
759 printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
760 task->pid, task->comm, cpu,
761 task->psi_flags, clear, set);
762 psi_bug = 1;
763 }
764
765 task->psi_flags &= ~clear;
766 task->psi_flags |= set;
767
768 /*
769 * Periodic aggregation shuts off if there is a period of no
770 * task changes, so we wake it back up if necessary. However,
771 * don't do this if the task change is the aggregation worker
772 * itself going to sleep, or we'll ping-pong forever.
773 */
774 if (unlikely((clear & TSK_RUNNING) &&
775 (task->flags & PF_WQ_WORKER) &&
776 wq_worker_last_func(task) == psi_avgs_work))
777 wake_clock = false;
778
779 while ((group = iterate_groups(task, &iter))) {
780 u32 state_mask = psi_group_change(group, cpu, clear, set);
781
782 if (state_mask & group->poll_states)
783 psi_schedule_poll_work(group, 1);
784
785 if (wake_clock && !delayed_work_pending(&group->avgs_work))
786 schedule_delayed_work(&group->avgs_work, PSI_FREQ);
787 }
788}
789
790void psi_memstall_tick(struct task_struct *task, int cpu)
791{
792 struct psi_group *group;
793 void *iter = NULL;
794
795 while ((group = iterate_groups(task, &iter))) {
796 struct psi_group_cpu *groupc;
797
798 groupc = per_cpu_ptr(group->pcpu, cpu);
799 write_seqcount_begin(&groupc->seq);
800 record_times(groupc, cpu, true);
801 write_seqcount_end(&groupc->seq);
802 }
803}
804
805/**
806 * psi_memstall_enter - mark the beginning of a memory stall section
807 * @flags: flags to handle nested sections
808 *
809 * Marks the calling task as being stalled due to a lack of memory,
810 * such as waiting for a refault or performing reclaim.
811 */
812void psi_memstall_enter(unsigned long *flags)
813{
814 struct rq_flags rf;
815 struct rq *rq;
816
817 if (static_branch_likely(&psi_disabled))
818 return;
819
820 *flags = current->flags & PF_MEMSTALL;
821 if (*flags)
822 return;
823 /*
824 * PF_MEMSTALL setting & accounting needs to be atomic wrt
825 * changes to the task's scheduling state, otherwise we can
826 * race with CPU migration.
827 */
828 rq = this_rq_lock_irq(&rf);
829
830 current->flags |= PF_MEMSTALL;
831 psi_task_change(current, 0, TSK_MEMSTALL);
832
833 rq_unlock_irq(rq, &rf);
834}
835
836/**
837 * psi_memstall_leave - mark the end of an memory stall section
838 * @flags: flags to handle nested memdelay sections
839 *
840 * Marks the calling task as no longer stalled due to lack of memory.
841 */
842void psi_memstall_leave(unsigned long *flags)
843{
844 struct rq_flags rf;
845 struct rq *rq;
846
847 if (static_branch_likely(&psi_disabled))
848 return;
849
850 if (*flags)
851 return;
852 /*
853 * PF_MEMSTALL clearing & accounting needs to be atomic wrt
854 * changes to the task's scheduling state, otherwise we could
855 * race with CPU migration.
856 */
857 rq = this_rq_lock_irq(&rf);
858
859 current->flags &= ~PF_MEMSTALL;
860 psi_task_change(current, TSK_MEMSTALL, 0);
861
862 rq_unlock_irq(rq, &rf);
863}
864
865#ifdef CONFIG_CGROUPS
866int psi_cgroup_alloc(struct cgroup *cgroup)
867{
868 if (static_branch_likely(&psi_disabled))
869 return 0;
870
871 cgroup->psi.pcpu = alloc_percpu(struct psi_group_cpu);
872 if (!cgroup->psi.pcpu)
873 return -ENOMEM;
874 group_init(&cgroup->psi);
875 return 0;
876}
877
878void psi_cgroup_free(struct cgroup *cgroup)
879{
880 if (static_branch_likely(&psi_disabled))
881 return;
882
883 cancel_delayed_work_sync(&cgroup->psi.avgs_work);
884 free_percpu(cgroup->psi.pcpu);
885 /* All triggers must be removed by now */
886 WARN_ONCE(cgroup->psi.poll_states, "psi: trigger leak\n");
887}
888
889/**
890 * cgroup_move_task - move task to a different cgroup
891 * @task: the task
892 * @to: the target css_set
893 *
894 * Move task to a new cgroup and safely migrate its associated stall
895 * state between the different groups.
896 *
897 * This function acquires the task's rq lock to lock out concurrent
898 * changes to the task's scheduling state and - in case the task is
899 * running - concurrent changes to its stall state.
900 */
901void cgroup_move_task(struct task_struct *task, struct css_set *to)
902{
903 unsigned int task_flags = 0;
904 struct rq_flags rf;
905 struct rq *rq;
906
907 if (static_branch_likely(&psi_disabled)) {
908 /*
909 * Lame to do this here, but the scheduler cannot be locked
910 * from the outside, so we move cgroups from inside sched/.
911 */
912 rcu_assign_pointer(task->cgroups, to);
913 return;
914 }
915
916 rq = task_rq_lock(task, &rf);
917
918 if (task_on_rq_queued(task))
919 task_flags = TSK_RUNNING;
920 else if (task->in_iowait)
921 task_flags = TSK_IOWAIT;
922
923 if (task->flags & PF_MEMSTALL)
924 task_flags |= TSK_MEMSTALL;
925
926 if (task_flags)
927 psi_task_change(task, task_flags, 0);
928
929 /* See comment above */
930 rcu_assign_pointer(task->cgroups, to);
931
932 if (task_flags)
933 psi_task_change(task, 0, task_flags);
934
935 task_rq_unlock(rq, task, &rf);
936}
937#endif /* CONFIG_CGROUPS */
938
939int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
940{
941 int full;
942 u64 now;
943
944 if (static_branch_likely(&psi_disabled))
945 return -EOPNOTSUPP;
946
947 /* Update averages before reporting them */
948 mutex_lock(&group->avgs_lock);
949 now = sched_clock();
950 collect_percpu_times(group, PSI_AVGS, NULL);
951 if (now >= group->avg_next_update)
952 group->avg_next_update = update_averages(group, now);
953 mutex_unlock(&group->avgs_lock);
954
955 for (full = 0; full < 2 - (res == PSI_CPU); full++) {
956 unsigned long avg[3];
957 u64 total;
958 int w;
959
960 for (w = 0; w < 3; w++)
961 avg[w] = group->avg[res * 2 + full][w];
962 total = div_u64(group->total[PSI_AVGS][res * 2 + full],
963 NSEC_PER_USEC);
964
965 seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
966 full ? "full" : "some",
967 LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
968 LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
969 LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
970 total);
971 }
972
973 return 0;
974}
975
976static int psi_io_show(struct seq_file *m, void *v)
977{
978 return psi_show(m, &psi_system, PSI_IO);
979}
980
981static int psi_memory_show(struct seq_file *m, void *v)
982{
983 return psi_show(m, &psi_system, PSI_MEM);
984}
985
986static int psi_cpu_show(struct seq_file *m, void *v)
987{
988 return psi_show(m, &psi_system, PSI_CPU);
989}
990
991static int psi_io_open(struct inode *inode, struct file *file)
992{
993 return single_open(file, psi_io_show, NULL);
994}
995
996static int psi_memory_open(struct inode *inode, struct file *file)
997{
998 return single_open(file, psi_memory_show, NULL);
999}
1000
1001static int psi_cpu_open(struct inode *inode, struct file *file)
1002{
1003 return single_open(file, psi_cpu_show, NULL);
1004}
1005
1006struct psi_trigger *psi_trigger_create(struct psi_group *group,
1007 char *buf, size_t nbytes, enum psi_res res)
1008{
1009 struct psi_trigger *t;
1010 enum psi_states state;
1011 u32 threshold_us;
1012 u32 window_us;
1013
1014 if (static_branch_likely(&psi_disabled))
1015 return ERR_PTR(-EOPNOTSUPP);
1016
1017 if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1018 state = PSI_IO_SOME + res * 2;
1019 else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1020 state = PSI_IO_FULL + res * 2;
1021 else
1022 return ERR_PTR(-EINVAL);
1023
1024 if (state >= PSI_NONIDLE)
1025 return ERR_PTR(-EINVAL);
1026
1027 if (window_us < WINDOW_MIN_US ||
1028 window_us > WINDOW_MAX_US)
1029 return ERR_PTR(-EINVAL);
1030
1031 /* Check threshold */
1032 if (threshold_us == 0 || threshold_us > window_us)
1033 return ERR_PTR(-EINVAL);
1034
1035 t = kmalloc(sizeof(*t), GFP_KERNEL);
1036 if (!t)
1037 return ERR_PTR(-ENOMEM);
1038
1039 t->group = group;
1040 t->state = state;
1041 t->threshold = threshold_us * NSEC_PER_USEC;
1042 t->win.size = window_us * NSEC_PER_USEC;
1043 window_reset(&t->win, 0, 0, 0);
1044
1045 t->event = 0;
1046 t->last_event_time = 0;
1047 init_waitqueue_head(&t->event_wait);
1048 kref_init(&t->refcount);
1049
1050 mutex_lock(&group->trigger_lock);
1051
1052 if (!rcu_access_pointer(group->poll_kworker)) {
1053 struct sched_param param = {
1054 .sched_priority = 1,
1055 };
1056 struct kthread_worker *kworker;
1057
1058 kworker = kthread_create_worker(0, "psimon");
1059 if (IS_ERR(kworker)) {
1060 kfree(t);
1061 mutex_unlock(&group->trigger_lock);
1062 return ERR_CAST(kworker);
1063 }
1064 sched_setscheduler_nocheck(kworker->task, SCHED_FIFO, ¶m);
1065 kthread_init_delayed_work(&group->poll_work,
1066 psi_poll_work);
1067 rcu_assign_pointer(group->poll_kworker, kworker);
1068 }
1069
1070 list_add(&t->node, &group->triggers);
1071 group->poll_min_period = min(group->poll_min_period,
1072 div_u64(t->win.size, UPDATES_PER_WINDOW));
1073 group->nr_triggers[t->state]++;
1074 group->poll_states |= (1 << t->state);
1075
1076 mutex_unlock(&group->trigger_lock);
1077
1078 return t;
1079}
1080
1081static void psi_trigger_destroy(struct kref *ref)
1082{
1083 struct psi_trigger *t = container_of(ref, struct psi_trigger, refcount);
1084 struct psi_group *group = t->group;
1085 struct kthread_worker *kworker_to_destroy = NULL;
1086
1087 if (static_branch_likely(&psi_disabled))
1088 return;
1089
1090 /*
1091 * Wakeup waiters to stop polling. Can happen if cgroup is deleted
1092 * from under a polling process.
1093 */
1094 wake_up_interruptible(&t->event_wait);
1095
1096 mutex_lock(&group->trigger_lock);
1097
1098 if (!list_empty(&t->node)) {
1099 struct psi_trigger *tmp;
1100 u64 period = ULLONG_MAX;
1101
1102 list_del(&t->node);
1103 group->nr_triggers[t->state]--;
1104 if (!group->nr_triggers[t->state])
1105 group->poll_states &= ~(1 << t->state);
1106 /* reset min update period for the remaining triggers */
1107 list_for_each_entry(tmp, &group->triggers, node)
1108 period = min(period, div_u64(tmp->win.size,
1109 UPDATES_PER_WINDOW));
1110 group->poll_min_period = period;
1111 /* Destroy poll_kworker when the last trigger is destroyed */
1112 if (group->poll_states == 0) {
1113 group->polling_until = 0;
1114 kworker_to_destroy = rcu_dereference_protected(
1115 group->poll_kworker,
1116 lockdep_is_held(&group->trigger_lock));
1117 rcu_assign_pointer(group->poll_kworker, NULL);
1118 }
1119 }
1120
1121 mutex_unlock(&group->trigger_lock);
1122
1123 /*
1124 * Wait for both *trigger_ptr from psi_trigger_replace and
1125 * poll_kworker RCUs to complete their read-side critical sections
1126 * before destroying the trigger and optionally the poll_kworker
1127 */
1128 synchronize_rcu();
1129 /*
1130 * Destroy the kworker after releasing trigger_lock to prevent a
1131 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1132 */
1133 if (kworker_to_destroy) {
1134 /*
1135 * After the RCU grace period has expired, the worker
1136 * can no longer be found through group->poll_kworker.
1137 * But it might have been already scheduled before
1138 * that - deschedule it cleanly before destroying it.
1139 */
1140 kthread_cancel_delayed_work_sync(&group->poll_work);
1141 atomic_set(&group->poll_scheduled, 0);
1142
1143 kthread_destroy_worker(kworker_to_destroy);
1144 }
1145 kfree(t);
1146}
1147
1148void psi_trigger_replace(void **trigger_ptr, struct psi_trigger *new)
1149{
1150 struct psi_trigger *old = *trigger_ptr;
1151
1152 if (static_branch_likely(&psi_disabled))
1153 return;
1154
1155 rcu_assign_pointer(*trigger_ptr, new);
1156 if (old)
1157 kref_put(&old->refcount, psi_trigger_destroy);
1158}
1159
1160__poll_t psi_trigger_poll(void **trigger_ptr,
1161 struct file *file, poll_table *wait)
1162{
1163 __poll_t ret = DEFAULT_POLLMASK;
1164 struct psi_trigger *t;
1165
1166 if (static_branch_likely(&psi_disabled))
1167 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1168
1169 rcu_read_lock();
1170
1171 t = rcu_dereference(*(void __rcu __force **)trigger_ptr);
1172 if (!t) {
1173 rcu_read_unlock();
1174 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1175 }
1176 kref_get(&t->refcount);
1177
1178 rcu_read_unlock();
1179
1180 poll_wait(file, &t->event_wait, wait);
1181
1182 if (cmpxchg(&t->event, 1, 0) == 1)
1183 ret |= EPOLLPRI;
1184
1185 kref_put(&t->refcount, psi_trigger_destroy);
1186
1187 return ret;
1188}
1189
1190static ssize_t psi_write(struct file *file, const char __user *user_buf,
1191 size_t nbytes, enum psi_res res)
1192{
1193 char buf[32];
1194 size_t buf_size;
1195 struct seq_file *seq;
1196 struct psi_trigger *new;
1197
1198 if (static_branch_likely(&psi_disabled))
1199 return -EOPNOTSUPP;
1200
1201 buf_size = min(nbytes, sizeof(buf));
1202 if (copy_from_user(buf, user_buf, buf_size))
1203 return -EFAULT;
1204
1205 buf[buf_size - 1] = '\0';
1206
1207 new = psi_trigger_create(&psi_system, buf, nbytes, res);
1208 if (IS_ERR(new))
1209 return PTR_ERR(new);
1210
1211 seq = file->private_data;
1212 /* Take seq->lock to protect seq->private from concurrent writes */
1213 mutex_lock(&seq->lock);
1214 psi_trigger_replace(&seq->private, new);
1215 mutex_unlock(&seq->lock);
1216
1217 return nbytes;
1218}
1219
1220static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1221 size_t nbytes, loff_t *ppos)
1222{
1223 return psi_write(file, user_buf, nbytes, PSI_IO);
1224}
1225
1226static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1227 size_t nbytes, loff_t *ppos)
1228{
1229 return psi_write(file, user_buf, nbytes, PSI_MEM);
1230}
1231
1232static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1233 size_t nbytes, loff_t *ppos)
1234{
1235 return psi_write(file, user_buf, nbytes, PSI_CPU);
1236}
1237
1238static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1239{
1240 struct seq_file *seq = file->private_data;
1241
1242 return psi_trigger_poll(&seq->private, file, wait);
1243}
1244
1245static int psi_fop_release(struct inode *inode, struct file *file)
1246{
1247 struct seq_file *seq = file->private_data;
1248
1249 psi_trigger_replace(&seq->private, NULL);
1250 return single_release(inode, file);
1251}
1252
1253static const struct file_operations psi_io_fops = {
1254 .open = psi_io_open,
1255 .read = seq_read,
1256 .llseek = seq_lseek,
1257 .write = psi_io_write,
1258 .poll = psi_fop_poll,
1259 .release = psi_fop_release,
1260};
1261
1262static const struct file_operations psi_memory_fops = {
1263 .open = psi_memory_open,
1264 .read = seq_read,
1265 .llseek = seq_lseek,
1266 .write = psi_memory_write,
1267 .poll = psi_fop_poll,
1268 .release = psi_fop_release,
1269};
1270
1271static const struct file_operations psi_cpu_fops = {
1272 .open = psi_cpu_open,
1273 .read = seq_read,
1274 .llseek = seq_lseek,
1275 .write = psi_cpu_write,
1276 .poll = psi_fop_poll,
1277 .release = psi_fop_release,
1278};
1279
1280static int __init psi_proc_init(void)
1281{
1282 proc_mkdir("pressure", NULL);
1283 proc_create("pressure/io", 0, NULL, &psi_io_fops);
1284 proc_create("pressure/memory", 0, NULL, &psi_memory_fops);
1285 proc_create("pressure/cpu", 0, NULL, &psi_cpu_fops);
1286 return 0;
1287}
1288module_init(psi_proc_init);
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 on-CPU 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 wall clock 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
140static int psi_bug __read_mostly;
141
142DEFINE_STATIC_KEY_FALSE(psi_disabled);
143static DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled);
144
145#ifdef CONFIG_PSI_DEFAULT_DISABLED
146static bool psi_enable;
147#else
148static bool psi_enable = true;
149#endif
150static 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_MAX_US 10000000 /* Max window size is 10s */
164#define UPDATES_PER_WINDOW 10 /* 10 updates per window */
165
166/* Sampling frequency in nanoseconds */
167static u64 psi_period __read_mostly;
168
169/* System-level pressure and stall tracking */
170static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
171struct psi_group psi_system = {
172 .pcpu = &system_group_pcpu,
173};
174
175static void psi_avgs_work(struct work_struct *work);
176
177static void poll_timer_fn(struct timer_list *t);
178
179static void group_init(struct psi_group *group)
180{
181 int cpu;
182
183 group->enabled = true;
184 for_each_possible_cpu(cpu)
185 seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);
186 group->avg_last_update = sched_clock();
187 group->avg_next_update = group->avg_last_update + psi_period;
188 mutex_init(&group->avgs_lock);
189
190 /* Init avg trigger-related members */
191 INIT_LIST_HEAD(&group->avg_triggers);
192 memset(group->avg_nr_triggers, 0, sizeof(group->avg_nr_triggers));
193 INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
194
195 /* Init rtpoll trigger-related members */
196 atomic_set(&group->rtpoll_scheduled, 0);
197 mutex_init(&group->rtpoll_trigger_lock);
198 INIT_LIST_HEAD(&group->rtpoll_triggers);
199 group->rtpoll_min_period = U32_MAX;
200 group->rtpoll_next_update = ULLONG_MAX;
201 init_waitqueue_head(&group->rtpoll_wait);
202 timer_setup(&group->rtpoll_timer, poll_timer_fn, 0);
203 rcu_assign_pointer(group->rtpoll_task, NULL);
204}
205
206void __init psi_init(void)
207{
208 if (!psi_enable) {
209 static_branch_enable(&psi_disabled);
210 static_branch_disable(&psi_cgroups_enabled);
211 return;
212 }
213
214 if (!cgroup_psi_enabled())
215 static_branch_disable(&psi_cgroups_enabled);
216
217 psi_period = jiffies_to_nsecs(PSI_FREQ);
218 group_init(&psi_system);
219}
220
221static u32 test_states(unsigned int *tasks, u32 state_mask)
222{
223 const bool oncpu = state_mask & PSI_ONCPU;
224
225 if (tasks[NR_IOWAIT]) {
226 state_mask |= BIT(PSI_IO_SOME);
227 if (!tasks[NR_RUNNING])
228 state_mask |= BIT(PSI_IO_FULL);
229 }
230
231 if (tasks[NR_MEMSTALL]) {
232 state_mask |= BIT(PSI_MEM_SOME);
233 if (tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING])
234 state_mask |= BIT(PSI_MEM_FULL);
235 }
236
237 if (tasks[NR_RUNNING] > oncpu)
238 state_mask |= BIT(PSI_CPU_SOME);
239
240 if (tasks[NR_RUNNING] && !oncpu)
241 state_mask |= BIT(PSI_CPU_FULL);
242
243 if (tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] || tasks[NR_RUNNING])
244 state_mask |= BIT(PSI_NONIDLE);
245
246 return state_mask;
247}
248
249static void get_recent_times(struct psi_group *group, int cpu,
250 enum psi_aggregators aggregator, u32 *times,
251 u32 *pchanged_states)
252{
253 struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
254 int current_cpu = raw_smp_processor_id();
255 unsigned int tasks[NR_PSI_TASK_COUNTS];
256 u64 now, state_start;
257 enum psi_states s;
258 unsigned int seq;
259 u32 state_mask;
260
261 *pchanged_states = 0;
262
263 /* Snapshot a coherent view of the CPU state */
264 do {
265 seq = read_seqcount_begin(&groupc->seq);
266 now = cpu_clock(cpu);
267 memcpy(times, groupc->times, sizeof(groupc->times));
268 state_mask = groupc->state_mask;
269 state_start = groupc->state_start;
270 if (cpu == current_cpu)
271 memcpy(tasks, groupc->tasks, sizeof(groupc->tasks));
272 } while (read_seqcount_retry(&groupc->seq, seq));
273
274 /* Calculate state time deltas against the previous snapshot */
275 for (s = 0; s < NR_PSI_STATES; s++) {
276 u32 delta;
277 /*
278 * In addition to already concluded states, we also
279 * incorporate currently active states on the CPU,
280 * since states may last for many sampling periods.
281 *
282 * This way we keep our delta sampling buckets small
283 * (u32) and our reported pressure close to what's
284 * actually happening.
285 */
286 if (state_mask & (1 << s))
287 times[s] += now - state_start;
288
289 delta = times[s] - groupc->times_prev[aggregator][s];
290 groupc->times_prev[aggregator][s] = times[s];
291
292 times[s] = delta;
293 if (delta)
294 *pchanged_states |= (1 << s);
295 }
296
297 /*
298 * When collect_percpu_times() from the avgs_work, we don't want to
299 * re-arm avgs_work when all CPUs are IDLE. But the current CPU running
300 * this avgs_work is never IDLE, cause avgs_work can't be shut off.
301 * So for the current CPU, we need to re-arm avgs_work only when
302 * (NR_RUNNING > 1 || NR_IOWAIT > 0 || NR_MEMSTALL > 0), for other CPUs
303 * we can just check PSI_NONIDLE delta.
304 */
305 if (current_work() == &group->avgs_work.work) {
306 bool reschedule;
307
308 if (cpu == current_cpu)
309 reschedule = tasks[NR_RUNNING] +
310 tasks[NR_IOWAIT] +
311 tasks[NR_MEMSTALL] > 1;
312 else
313 reschedule = *pchanged_states & (1 << PSI_NONIDLE);
314
315 if (reschedule)
316 *pchanged_states |= PSI_STATE_RESCHEDULE;
317 }
318}
319
320static void calc_avgs(unsigned long avg[3], int missed_periods,
321 u64 time, u64 period)
322{
323 unsigned long pct;
324
325 /* Fill in zeroes for periods of no activity */
326 if (missed_periods) {
327 avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
328 avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
329 avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
330 }
331
332 /* Sample the most recent active period */
333 pct = div_u64(time * 100, period);
334 pct *= FIXED_1;
335 avg[0] = calc_load(avg[0], EXP_10s, pct);
336 avg[1] = calc_load(avg[1], EXP_60s, pct);
337 avg[2] = calc_load(avg[2], EXP_300s, pct);
338}
339
340static void collect_percpu_times(struct psi_group *group,
341 enum psi_aggregators aggregator,
342 u32 *pchanged_states)
343{
344 u64 deltas[NR_PSI_STATES - 1] = { 0, };
345 unsigned long nonidle_total = 0;
346 u32 changed_states = 0;
347 int cpu;
348 int s;
349
350 /*
351 * Collect the per-cpu time buckets and average them into a
352 * single time sample that is normalized to wall clock time.
353 *
354 * For averaging, each CPU is weighted by its non-idle time in
355 * the sampling period. This eliminates artifacts from uneven
356 * loading, or even entirely idle CPUs.
357 */
358 for_each_possible_cpu(cpu) {
359 u32 times[NR_PSI_STATES];
360 u32 nonidle;
361 u32 cpu_changed_states;
362
363 get_recent_times(group, cpu, aggregator, times,
364 &cpu_changed_states);
365 changed_states |= cpu_changed_states;
366
367 nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
368 nonidle_total += nonidle;
369
370 for (s = 0; s < PSI_NONIDLE; s++)
371 deltas[s] += (u64)times[s] * nonidle;
372 }
373
374 /*
375 * Integrate the sample into the running statistics that are
376 * reported to userspace: the cumulative stall times and the
377 * decaying averages.
378 *
379 * Pressure percentages are sampled at PSI_FREQ. We might be
380 * called more often when the user polls more frequently than
381 * that; we might be called less often when there is no task
382 * activity, thus no data, and clock ticks are sporadic. The
383 * below handles both.
384 */
385
386 /* total= */
387 for (s = 0; s < NR_PSI_STATES - 1; s++)
388 group->total[aggregator][s] +=
389 div_u64(deltas[s], max(nonidle_total, 1UL));
390
391 if (pchanged_states)
392 *pchanged_states = changed_states;
393}
394
395/* Trigger tracking window manipulations */
396static void window_reset(struct psi_window *win, u64 now, u64 value,
397 u64 prev_growth)
398{
399 win->start_time = now;
400 win->start_value = value;
401 win->prev_growth = prev_growth;
402}
403
404/*
405 * PSI growth tracking window update and growth calculation routine.
406 *
407 * This approximates a sliding tracking window by interpolating
408 * partially elapsed windows using historical growth data from the
409 * previous intervals. This minimizes memory requirements (by not storing
410 * all the intermediate values in the previous window) and simplifies
411 * the calculations. It works well because PSI signal changes only in
412 * positive direction and over relatively small window sizes the growth
413 * is close to linear.
414 */
415static u64 window_update(struct psi_window *win, u64 now, u64 value)
416{
417 u64 elapsed;
418 u64 growth;
419
420 elapsed = now - win->start_time;
421 growth = value - win->start_value;
422 /*
423 * After each tracking window passes win->start_value and
424 * win->start_time get reset and win->prev_growth stores
425 * the average per-window growth of the previous window.
426 * win->prev_growth is then used to interpolate additional
427 * growth from the previous window assuming it was linear.
428 */
429 if (elapsed > win->size)
430 window_reset(win, now, value, growth);
431 else {
432 u32 remaining;
433
434 remaining = win->size - elapsed;
435 growth += div64_u64(win->prev_growth * remaining, win->size);
436 }
437
438 return growth;
439}
440
441static void update_triggers(struct psi_group *group, u64 now,
442 enum psi_aggregators aggregator)
443{
444 struct psi_trigger *t;
445 u64 *total = group->total[aggregator];
446 struct list_head *triggers;
447 u64 *aggregator_total;
448
449 if (aggregator == PSI_AVGS) {
450 triggers = &group->avg_triggers;
451 aggregator_total = group->avg_total;
452 } else {
453 triggers = &group->rtpoll_triggers;
454 aggregator_total = group->rtpoll_total;
455 }
456
457 /*
458 * On subsequent updates, calculate growth deltas and let
459 * watchers know when their specified thresholds are exceeded.
460 */
461 list_for_each_entry(t, triggers, node) {
462 u64 growth;
463 bool new_stall;
464
465 new_stall = aggregator_total[t->state] != total[t->state];
466
467 /* Check for stall activity or a previous threshold breach */
468 if (!new_stall && !t->pending_event)
469 continue;
470 /*
471 * Check for new stall activity, as well as deferred
472 * events that occurred in the last window after the
473 * trigger had already fired (we want to ratelimit
474 * events without dropping any).
475 */
476 if (new_stall) {
477 /* Calculate growth since last update */
478 growth = window_update(&t->win, now, total[t->state]);
479 if (!t->pending_event) {
480 if (growth < t->threshold)
481 continue;
482
483 t->pending_event = true;
484 }
485 }
486 /* Limit event signaling to once per window */
487 if (now < t->last_event_time + t->win.size)
488 continue;
489
490 /* Generate an event */
491 if (cmpxchg(&t->event, 0, 1) == 0) {
492 if (t->of)
493 kernfs_notify(t->of->kn);
494 else
495 wake_up_interruptible(&t->event_wait);
496 }
497 t->last_event_time = now;
498 /* Reset threshold breach flag once event got generated */
499 t->pending_event = false;
500 }
501}
502
503static u64 update_averages(struct psi_group *group, u64 now)
504{
505 unsigned long missed_periods = 0;
506 u64 expires, period;
507 u64 avg_next_update;
508 int s;
509
510 /* avgX= */
511 expires = group->avg_next_update;
512 if (now - expires >= psi_period)
513 missed_periods = div_u64(now - expires, psi_period);
514
515 /*
516 * The periodic clock tick can get delayed for various
517 * reasons, especially on loaded systems. To avoid clock
518 * drift, we schedule the clock in fixed psi_period intervals.
519 * But the deltas we sample out of the per-cpu buckets above
520 * are based on the actual time elapsing between clock ticks.
521 */
522 avg_next_update = expires + ((1 + missed_periods) * psi_period);
523 period = now - (group->avg_last_update + (missed_periods * psi_period));
524 group->avg_last_update = now;
525
526 for (s = 0; s < NR_PSI_STATES - 1; s++) {
527 u32 sample;
528
529 sample = group->total[PSI_AVGS][s] - group->avg_total[s];
530 /*
531 * Due to the lockless sampling of the time buckets,
532 * recorded time deltas can slip into the next period,
533 * which under full pressure can result in samples in
534 * excess of the period length.
535 *
536 * We don't want to report non-sensical pressures in
537 * excess of 100%, nor do we want to drop such events
538 * on the floor. Instead we punt any overage into the
539 * future until pressure subsides. By doing this we
540 * don't underreport the occurring pressure curve, we
541 * just report it delayed by one period length.
542 *
543 * The error isn't cumulative. As soon as another
544 * delta slips from a period P to P+1, by definition
545 * it frees up its time T in P.
546 */
547 if (sample > period)
548 sample = period;
549 group->avg_total[s] += sample;
550 calc_avgs(group->avg[s], missed_periods, sample, period);
551 }
552
553 return avg_next_update;
554}
555
556static void psi_avgs_work(struct work_struct *work)
557{
558 struct delayed_work *dwork;
559 struct psi_group *group;
560 u32 changed_states;
561 u64 now;
562
563 dwork = to_delayed_work(work);
564 group = container_of(dwork, struct psi_group, avgs_work);
565
566 mutex_lock(&group->avgs_lock);
567
568 now = sched_clock();
569
570 collect_percpu_times(group, PSI_AVGS, &changed_states);
571 /*
572 * If there is task activity, periodically fold the per-cpu
573 * times and feed samples into the running averages. If things
574 * are idle and there is no data to process, stop the clock.
575 * Once restarted, we'll catch up the running averages in one
576 * go - see calc_avgs() and missed_periods.
577 */
578 if (now >= group->avg_next_update) {
579 update_triggers(group, now, PSI_AVGS);
580 group->avg_next_update = update_averages(group, now);
581 }
582
583 if (changed_states & PSI_STATE_RESCHEDULE) {
584 schedule_delayed_work(dwork, nsecs_to_jiffies(
585 group->avg_next_update - now) + 1);
586 }
587
588 mutex_unlock(&group->avgs_lock);
589}
590
591static void init_rtpoll_triggers(struct psi_group *group, u64 now)
592{
593 struct psi_trigger *t;
594
595 list_for_each_entry(t, &group->rtpoll_triggers, node)
596 window_reset(&t->win, now,
597 group->total[PSI_POLL][t->state], 0);
598 memcpy(group->rtpoll_total, group->total[PSI_POLL],
599 sizeof(group->rtpoll_total));
600 group->rtpoll_next_update = now + group->rtpoll_min_period;
601}
602
603/* Schedule rtpolling if it's not already scheduled or forced. */
604static void psi_schedule_rtpoll_work(struct psi_group *group, unsigned long delay,
605 bool force)
606{
607 struct task_struct *task;
608
609 /*
610 * atomic_xchg should be called even when !force to provide a
611 * full memory barrier (see the comment inside psi_rtpoll_work).
612 */
613 if (atomic_xchg(&group->rtpoll_scheduled, 1) && !force)
614 return;
615
616 rcu_read_lock();
617
618 task = rcu_dereference(group->rtpoll_task);
619 /*
620 * kworker might be NULL in case psi_trigger_destroy races with
621 * psi_task_change (hotpath) which can't use locks
622 */
623 if (likely(task))
624 mod_timer(&group->rtpoll_timer, jiffies + delay);
625 else
626 atomic_set(&group->rtpoll_scheduled, 0);
627
628 rcu_read_unlock();
629}
630
631static void psi_rtpoll_work(struct psi_group *group)
632{
633 bool force_reschedule = false;
634 u32 changed_states;
635 u64 now;
636
637 mutex_lock(&group->rtpoll_trigger_lock);
638
639 now = sched_clock();
640
641 if (now > group->rtpoll_until) {
642 /*
643 * We are either about to start or might stop rtpolling if no
644 * state change was recorded. Resetting rtpoll_scheduled leaves
645 * a small window for psi_group_change to sneak in and schedule
646 * an immediate rtpoll_work before we get to rescheduling. One
647 * potential extra wakeup at the end of the rtpolling window
648 * should be negligible and rtpoll_next_update still keeps
649 * updates correctly on schedule.
650 */
651 atomic_set(&group->rtpoll_scheduled, 0);
652 /*
653 * A task change can race with the rtpoll worker that is supposed to
654 * report on it. To avoid missing events, ensure ordering between
655 * rtpoll_scheduled and the task state accesses, such that if the
656 * rtpoll worker misses the state update, the task change is
657 * guaranteed to reschedule the rtpoll worker:
658 *
659 * rtpoll worker:
660 * atomic_set(rtpoll_scheduled, 0)
661 * smp_mb()
662 * LOAD states
663 *
664 * task change:
665 * STORE states
666 * if atomic_xchg(rtpoll_scheduled, 1) == 0:
667 * schedule rtpoll worker
668 *
669 * The atomic_xchg() implies a full barrier.
670 */
671 smp_mb();
672 } else {
673 /* The rtpolling window is not over, keep rescheduling */
674 force_reschedule = true;
675 }
676
677
678 collect_percpu_times(group, PSI_POLL, &changed_states);
679
680 if (changed_states & group->rtpoll_states) {
681 /* Initialize trigger windows when entering rtpolling mode */
682 if (now > group->rtpoll_until)
683 init_rtpoll_triggers(group, now);
684
685 /*
686 * Keep the monitor active for at least the duration of the
687 * minimum tracking window as long as monitor states are
688 * changing.
689 */
690 group->rtpoll_until = now +
691 group->rtpoll_min_period * UPDATES_PER_WINDOW;
692 }
693
694 if (now > group->rtpoll_until) {
695 group->rtpoll_next_update = ULLONG_MAX;
696 goto out;
697 }
698
699 if (now >= group->rtpoll_next_update) {
700 if (changed_states & group->rtpoll_states) {
701 update_triggers(group, now, PSI_POLL);
702 memcpy(group->rtpoll_total, group->total[PSI_POLL],
703 sizeof(group->rtpoll_total));
704 }
705 group->rtpoll_next_update = now + group->rtpoll_min_period;
706 }
707
708 psi_schedule_rtpoll_work(group,
709 nsecs_to_jiffies(group->rtpoll_next_update - now) + 1,
710 force_reschedule);
711
712out:
713 mutex_unlock(&group->rtpoll_trigger_lock);
714}
715
716static int psi_rtpoll_worker(void *data)
717{
718 struct psi_group *group = (struct psi_group *)data;
719
720 sched_set_fifo_low(current);
721
722 while (true) {
723 wait_event_interruptible(group->rtpoll_wait,
724 atomic_cmpxchg(&group->rtpoll_wakeup, 1, 0) ||
725 kthread_should_stop());
726 if (kthread_should_stop())
727 break;
728
729 psi_rtpoll_work(group);
730 }
731 return 0;
732}
733
734static void poll_timer_fn(struct timer_list *t)
735{
736 struct psi_group *group = from_timer(group, t, rtpoll_timer);
737
738 atomic_set(&group->rtpoll_wakeup, 1);
739 wake_up_interruptible(&group->rtpoll_wait);
740}
741
742static void record_times(struct psi_group_cpu *groupc, u64 now)
743{
744 u32 delta;
745
746 delta = now - groupc->state_start;
747 groupc->state_start = now;
748
749 if (groupc->state_mask & (1 << PSI_IO_SOME)) {
750 groupc->times[PSI_IO_SOME] += delta;
751 if (groupc->state_mask & (1 << PSI_IO_FULL))
752 groupc->times[PSI_IO_FULL] += delta;
753 }
754
755 if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
756 groupc->times[PSI_MEM_SOME] += delta;
757 if (groupc->state_mask & (1 << PSI_MEM_FULL))
758 groupc->times[PSI_MEM_FULL] += delta;
759 }
760
761 if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
762 groupc->times[PSI_CPU_SOME] += delta;
763 if (groupc->state_mask & (1 << PSI_CPU_FULL))
764 groupc->times[PSI_CPU_FULL] += delta;
765 }
766
767 if (groupc->state_mask & (1 << PSI_NONIDLE))
768 groupc->times[PSI_NONIDLE] += delta;
769}
770
771static void psi_group_change(struct psi_group *group, int cpu,
772 unsigned int clear, unsigned int set,
773 bool wake_clock)
774{
775 struct psi_group_cpu *groupc;
776 unsigned int t, m;
777 u32 state_mask;
778 u64 now;
779
780 lockdep_assert_rq_held(cpu_rq(cpu));
781 groupc = per_cpu_ptr(group->pcpu, cpu);
782
783 /*
784 * First we update the task counts according to the state
785 * change requested through the @clear and @set bits.
786 *
787 * Then if the cgroup PSI stats accounting enabled, we
788 * assess the aggregate resource states this CPU's tasks
789 * have been in since the last change, and account any
790 * SOME and FULL time these may have resulted in.
791 */
792 write_seqcount_begin(&groupc->seq);
793 now = cpu_clock(cpu);
794
795 /*
796 * Start with TSK_ONCPU, which doesn't have a corresponding
797 * task count - it's just a boolean flag directly encoded in
798 * the state mask. Clear, set, or carry the current state if
799 * no changes are requested.
800 */
801 if (unlikely(clear & TSK_ONCPU)) {
802 state_mask = 0;
803 clear &= ~TSK_ONCPU;
804 } else if (unlikely(set & TSK_ONCPU)) {
805 state_mask = PSI_ONCPU;
806 set &= ~TSK_ONCPU;
807 } else {
808 state_mask = groupc->state_mask & PSI_ONCPU;
809 }
810
811 /*
812 * The rest of the state mask is calculated based on the task
813 * counts. Update those first, then construct the mask.
814 */
815 for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
816 if (!(m & (1 << t)))
817 continue;
818 if (groupc->tasks[t]) {
819 groupc->tasks[t]--;
820 } else if (!psi_bug) {
821 printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n",
822 cpu, t, groupc->tasks[0],
823 groupc->tasks[1], groupc->tasks[2],
824 groupc->tasks[3], clear, set);
825 psi_bug = 1;
826 }
827 }
828
829 for (t = 0; set; set &= ~(1 << t), t++)
830 if (set & (1 << t))
831 groupc->tasks[t]++;
832
833 if (!group->enabled) {
834 /*
835 * On the first group change after disabling PSI, conclude
836 * the current state and flush its time. This is unlikely
837 * to matter to the user, but aggregation (get_recent_times)
838 * may have already incorporated the live state into times_prev;
839 * avoid a delta sample underflow when PSI is later re-enabled.
840 */
841 if (unlikely(groupc->state_mask & (1 << PSI_NONIDLE)))
842 record_times(groupc, now);
843
844 groupc->state_mask = state_mask;
845
846 write_seqcount_end(&groupc->seq);
847 return;
848 }
849
850 state_mask = test_states(groupc->tasks, state_mask);
851
852 /*
853 * Since we care about lost potential, a memstall is FULL
854 * when there are no other working tasks, but also when
855 * the CPU is actively reclaiming and nothing productive
856 * could run even if it were runnable. So when the current
857 * task in a cgroup is in_memstall, the corresponding groupc
858 * on that cpu is in PSI_MEM_FULL state.
859 */
860 if (unlikely((state_mask & PSI_ONCPU) && cpu_curr(cpu)->in_memstall))
861 state_mask |= (1 << PSI_MEM_FULL);
862
863 record_times(groupc, now);
864
865 groupc->state_mask = state_mask;
866
867 write_seqcount_end(&groupc->seq);
868
869 if (state_mask & group->rtpoll_states)
870 psi_schedule_rtpoll_work(group, 1, false);
871
872 if (wake_clock && !delayed_work_pending(&group->avgs_work))
873 schedule_delayed_work(&group->avgs_work, PSI_FREQ);
874}
875
876static inline struct psi_group *task_psi_group(struct task_struct *task)
877{
878#ifdef CONFIG_CGROUPS
879 if (static_branch_likely(&psi_cgroups_enabled))
880 return cgroup_psi(task_dfl_cgroup(task));
881#endif
882 return &psi_system;
883}
884
885static void psi_flags_change(struct task_struct *task, int clear, int set)
886{
887 if (((task->psi_flags & set) ||
888 (task->psi_flags & clear) != clear) &&
889 !psi_bug) {
890 printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
891 task->pid, task->comm, task_cpu(task),
892 task->psi_flags, clear, set);
893 psi_bug = 1;
894 }
895
896 task->psi_flags &= ~clear;
897 task->psi_flags |= set;
898}
899
900void psi_task_change(struct task_struct *task, int clear, int set)
901{
902 int cpu = task_cpu(task);
903 struct psi_group *group;
904
905 if (!task->pid)
906 return;
907
908 psi_flags_change(task, clear, set);
909
910 group = task_psi_group(task);
911 do {
912 psi_group_change(group, cpu, clear, set, true);
913 } while ((group = group->parent));
914}
915
916void psi_task_switch(struct task_struct *prev, struct task_struct *next,
917 bool sleep)
918{
919 struct psi_group *group, *common = NULL;
920 int cpu = task_cpu(prev);
921
922 if (next->pid) {
923 psi_flags_change(next, 0, TSK_ONCPU);
924 /*
925 * Set TSK_ONCPU on @next's cgroups. If @next shares any
926 * ancestors with @prev, those will already have @prev's
927 * TSK_ONCPU bit set, and we can stop the iteration there.
928 */
929 group = task_psi_group(next);
930 do {
931 if (per_cpu_ptr(group->pcpu, cpu)->state_mask &
932 PSI_ONCPU) {
933 common = group;
934 break;
935 }
936
937 psi_group_change(group, cpu, 0, TSK_ONCPU, true);
938 } while ((group = group->parent));
939 }
940
941 if (prev->pid) {
942 int clear = TSK_ONCPU, set = 0;
943 bool wake_clock = true;
944
945 /*
946 * When we're going to sleep, psi_dequeue() lets us
947 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
948 * TSK_IOWAIT here, where we can combine it with
949 * TSK_ONCPU and save walking common ancestors twice.
950 */
951 if (sleep) {
952 clear |= TSK_RUNNING;
953 if (prev->in_memstall)
954 clear |= TSK_MEMSTALL_RUNNING;
955 if (prev->in_iowait)
956 set |= TSK_IOWAIT;
957
958 /*
959 * Periodic aggregation shuts off if there is a period of no
960 * task changes, so we wake it back up if necessary. However,
961 * don't do this if the task change is the aggregation worker
962 * itself going to sleep, or we'll ping-pong forever.
963 */
964 if (unlikely((prev->flags & PF_WQ_WORKER) &&
965 wq_worker_last_func(prev) == psi_avgs_work))
966 wake_clock = false;
967 }
968
969 psi_flags_change(prev, clear, set);
970
971 group = task_psi_group(prev);
972 do {
973 if (group == common)
974 break;
975 psi_group_change(group, cpu, clear, set, wake_clock);
976 } while ((group = group->parent));
977
978 /*
979 * TSK_ONCPU is handled up to the common ancestor. If there are
980 * any other differences between the two tasks (e.g. prev goes
981 * to sleep, or only one task is memstall), finish propagating
982 * those differences all the way up to the root.
983 */
984 if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) {
985 clear &= ~TSK_ONCPU;
986 for (; group; group = group->parent)
987 psi_group_change(group, cpu, clear, set, wake_clock);
988 }
989 }
990}
991
992#ifdef CONFIG_IRQ_TIME_ACCOUNTING
993void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_struct *prev)
994{
995 int cpu = task_cpu(curr);
996 struct psi_group *group;
997 struct psi_group_cpu *groupc;
998 s64 delta;
999 u64 irq;
1000
1001 if (static_branch_likely(&psi_disabled))
1002 return;
1003
1004 if (!curr->pid)
1005 return;
1006
1007 lockdep_assert_rq_held(rq);
1008 group = task_psi_group(curr);
1009 if (prev && task_psi_group(prev) == group)
1010 return;
1011
1012 irq = irq_time_read(cpu);
1013 delta = (s64)(irq - rq->psi_irq_time);
1014 if (delta < 0)
1015 return;
1016 rq->psi_irq_time = irq;
1017
1018 do {
1019 u64 now;
1020
1021 if (!group->enabled)
1022 continue;
1023
1024 groupc = per_cpu_ptr(group->pcpu, cpu);
1025
1026 write_seqcount_begin(&groupc->seq);
1027 now = cpu_clock(cpu);
1028
1029 record_times(groupc, now);
1030 groupc->times[PSI_IRQ_FULL] += delta;
1031
1032 write_seqcount_end(&groupc->seq);
1033
1034 if (group->rtpoll_states & (1 << PSI_IRQ_FULL))
1035 psi_schedule_rtpoll_work(group, 1, false);
1036 } while ((group = group->parent));
1037}
1038#endif
1039
1040/**
1041 * psi_memstall_enter - mark the beginning of a memory stall section
1042 * @flags: flags to handle nested sections
1043 *
1044 * Marks the calling task as being stalled due to a lack of memory,
1045 * such as waiting for a refault or performing reclaim.
1046 */
1047void psi_memstall_enter(unsigned long *flags)
1048{
1049 struct rq_flags rf;
1050 struct rq *rq;
1051
1052 if (static_branch_likely(&psi_disabled))
1053 return;
1054
1055 *flags = current->in_memstall;
1056 if (*flags)
1057 return;
1058 /*
1059 * in_memstall setting & accounting needs to be atomic wrt
1060 * changes to the task's scheduling state, otherwise we can
1061 * race with CPU migration.
1062 */
1063 rq = this_rq_lock_irq(&rf);
1064
1065 current->in_memstall = 1;
1066 psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
1067
1068 rq_unlock_irq(rq, &rf);
1069}
1070EXPORT_SYMBOL_GPL(psi_memstall_enter);
1071
1072/**
1073 * psi_memstall_leave - mark the end of an memory stall section
1074 * @flags: flags to handle nested memdelay sections
1075 *
1076 * Marks the calling task as no longer stalled due to lack of memory.
1077 */
1078void psi_memstall_leave(unsigned long *flags)
1079{
1080 struct rq_flags rf;
1081 struct rq *rq;
1082
1083 if (static_branch_likely(&psi_disabled))
1084 return;
1085
1086 if (*flags)
1087 return;
1088 /*
1089 * in_memstall clearing & accounting needs to be atomic wrt
1090 * changes to the task's scheduling state, otherwise we could
1091 * race with CPU migration.
1092 */
1093 rq = this_rq_lock_irq(&rf);
1094
1095 current->in_memstall = 0;
1096 psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0);
1097
1098 rq_unlock_irq(rq, &rf);
1099}
1100EXPORT_SYMBOL_GPL(psi_memstall_leave);
1101
1102#ifdef CONFIG_CGROUPS
1103int psi_cgroup_alloc(struct cgroup *cgroup)
1104{
1105 if (!static_branch_likely(&psi_cgroups_enabled))
1106 return 0;
1107
1108 cgroup->psi = kzalloc(sizeof(struct psi_group), GFP_KERNEL);
1109 if (!cgroup->psi)
1110 return -ENOMEM;
1111
1112 cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
1113 if (!cgroup->psi->pcpu) {
1114 kfree(cgroup->psi);
1115 return -ENOMEM;
1116 }
1117 group_init(cgroup->psi);
1118 cgroup->psi->parent = cgroup_psi(cgroup_parent(cgroup));
1119 return 0;
1120}
1121
1122void psi_cgroup_free(struct cgroup *cgroup)
1123{
1124 if (!static_branch_likely(&psi_cgroups_enabled))
1125 return;
1126
1127 cancel_delayed_work_sync(&cgroup->psi->avgs_work);
1128 free_percpu(cgroup->psi->pcpu);
1129 /* All triggers must be removed by now */
1130 WARN_ONCE(cgroup->psi->rtpoll_states, "psi: trigger leak\n");
1131 kfree(cgroup->psi);
1132}
1133
1134/**
1135 * cgroup_move_task - move task to a different cgroup
1136 * @task: the task
1137 * @to: the target css_set
1138 *
1139 * Move task to a new cgroup and safely migrate its associated stall
1140 * state between the different groups.
1141 *
1142 * This function acquires the task's rq lock to lock out concurrent
1143 * changes to the task's scheduling state and - in case the task is
1144 * running - concurrent changes to its stall state.
1145 */
1146void cgroup_move_task(struct task_struct *task, struct css_set *to)
1147{
1148 unsigned int task_flags;
1149 struct rq_flags rf;
1150 struct rq *rq;
1151
1152 if (!static_branch_likely(&psi_cgroups_enabled)) {
1153 /*
1154 * Lame to do this here, but the scheduler cannot be locked
1155 * from the outside, so we move cgroups from inside sched/.
1156 */
1157 rcu_assign_pointer(task->cgroups, to);
1158 return;
1159 }
1160
1161 rq = task_rq_lock(task, &rf);
1162
1163 /*
1164 * We may race with schedule() dropping the rq lock between
1165 * deactivating prev and switching to next. Because the psi
1166 * updates from the deactivation are deferred to the switch
1167 * callback to save cgroup tree updates, the task's scheduling
1168 * state here is not coherent with its psi state:
1169 *
1170 * schedule() cgroup_move_task()
1171 * rq_lock()
1172 * deactivate_task()
1173 * p->on_rq = 0
1174 * psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1175 * pick_next_task()
1176 * rq_unlock()
1177 * rq_lock()
1178 * psi_task_change() // old cgroup
1179 * task->cgroups = to
1180 * psi_task_change() // new cgroup
1181 * rq_unlock()
1182 * rq_lock()
1183 * psi_sched_switch() // does deferred updates in new cgroup
1184 *
1185 * Don't rely on the scheduling state. Use psi_flags instead.
1186 */
1187 task_flags = task->psi_flags;
1188
1189 if (task_flags)
1190 psi_task_change(task, task_flags, 0);
1191
1192 /* See comment above */
1193 rcu_assign_pointer(task->cgroups, to);
1194
1195 if (task_flags)
1196 psi_task_change(task, 0, task_flags);
1197
1198 task_rq_unlock(rq, task, &rf);
1199}
1200
1201void psi_cgroup_restart(struct psi_group *group)
1202{
1203 int cpu;
1204
1205 /*
1206 * After we disable psi_group->enabled, we don't actually
1207 * stop percpu tasks accounting in each psi_group_cpu,
1208 * instead only stop test_states() loop, record_times()
1209 * and averaging worker, see psi_group_change() for details.
1210 *
1211 * When disable cgroup PSI, this function has nothing to sync
1212 * since cgroup pressure files are hidden and percpu psi_group_cpu
1213 * would see !psi_group->enabled and only do task accounting.
1214 *
1215 * When re-enable cgroup PSI, this function use psi_group_change()
1216 * to get correct state mask from test_states() loop on tasks[],
1217 * and restart groupc->state_start from now, use .clear = .set = 0
1218 * here since no task status really changed.
1219 */
1220 if (!group->enabled)
1221 return;
1222
1223 for_each_possible_cpu(cpu) {
1224 struct rq *rq = cpu_rq(cpu);
1225 struct rq_flags rf;
1226
1227 rq_lock_irq(rq, &rf);
1228 psi_group_change(group, cpu, 0, 0, true);
1229 rq_unlock_irq(rq, &rf);
1230 }
1231}
1232#endif /* CONFIG_CGROUPS */
1233
1234int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1235{
1236 bool only_full = false;
1237 int full;
1238 u64 now;
1239
1240 if (static_branch_likely(&psi_disabled))
1241 return -EOPNOTSUPP;
1242
1243 /* Update averages before reporting them */
1244 mutex_lock(&group->avgs_lock);
1245 now = sched_clock();
1246 collect_percpu_times(group, PSI_AVGS, NULL);
1247 if (now >= group->avg_next_update)
1248 group->avg_next_update = update_averages(group, now);
1249 mutex_unlock(&group->avgs_lock);
1250
1251#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1252 only_full = res == PSI_IRQ;
1253#endif
1254
1255 for (full = 0; full < 2 - only_full; full++) {
1256 unsigned long avg[3] = { 0, };
1257 u64 total = 0;
1258 int w;
1259
1260 /* CPU FULL is undefined at the system level */
1261 if (!(group == &psi_system && res == PSI_CPU && full)) {
1262 for (w = 0; w < 3; w++)
1263 avg[w] = group->avg[res * 2 + full][w];
1264 total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1265 NSEC_PER_USEC);
1266 }
1267
1268 seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1269 full || only_full ? "full" : "some",
1270 LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1271 LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1272 LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1273 total);
1274 }
1275
1276 return 0;
1277}
1278
1279struct psi_trigger *psi_trigger_create(struct psi_group *group, char *buf,
1280 enum psi_res res, struct file *file,
1281 struct kernfs_open_file *of)
1282{
1283 struct psi_trigger *t;
1284 enum psi_states state;
1285 u32 threshold_us;
1286 bool privileged;
1287 u32 window_us;
1288
1289 if (static_branch_likely(&psi_disabled))
1290 return ERR_PTR(-EOPNOTSUPP);
1291
1292 /*
1293 * Checking the privilege here on file->f_cred implies that a privileged user
1294 * could open the file and delegate the write to an unprivileged one.
1295 */
1296 privileged = cap_raised(file->f_cred->cap_effective, CAP_SYS_RESOURCE);
1297
1298 if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1299 state = PSI_IO_SOME + res * 2;
1300 else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1301 state = PSI_IO_FULL + res * 2;
1302 else
1303 return ERR_PTR(-EINVAL);
1304
1305#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1306 if (res == PSI_IRQ && --state != PSI_IRQ_FULL)
1307 return ERR_PTR(-EINVAL);
1308#endif
1309
1310 if (state >= PSI_NONIDLE)
1311 return ERR_PTR(-EINVAL);
1312
1313 if (window_us == 0 || window_us > WINDOW_MAX_US)
1314 return ERR_PTR(-EINVAL);
1315
1316 /*
1317 * Unprivileged users can only use 2s windows so that averages aggregation
1318 * work is used, and no RT threads need to be spawned.
1319 */
1320 if (!privileged && window_us % 2000000)
1321 return ERR_PTR(-EINVAL);
1322
1323 /* Check threshold */
1324 if (threshold_us == 0 || threshold_us > window_us)
1325 return ERR_PTR(-EINVAL);
1326
1327 t = kmalloc(sizeof(*t), GFP_KERNEL);
1328 if (!t)
1329 return ERR_PTR(-ENOMEM);
1330
1331 t->group = group;
1332 t->state = state;
1333 t->threshold = threshold_us * NSEC_PER_USEC;
1334 t->win.size = window_us * NSEC_PER_USEC;
1335 window_reset(&t->win, sched_clock(),
1336 group->total[PSI_POLL][t->state], 0);
1337
1338 t->event = 0;
1339 t->last_event_time = 0;
1340 t->of = of;
1341 if (!of)
1342 init_waitqueue_head(&t->event_wait);
1343 t->pending_event = false;
1344 t->aggregator = privileged ? PSI_POLL : PSI_AVGS;
1345
1346 if (privileged) {
1347 mutex_lock(&group->rtpoll_trigger_lock);
1348
1349 if (!rcu_access_pointer(group->rtpoll_task)) {
1350 struct task_struct *task;
1351
1352 task = kthread_create(psi_rtpoll_worker, group, "psimon");
1353 if (IS_ERR(task)) {
1354 kfree(t);
1355 mutex_unlock(&group->rtpoll_trigger_lock);
1356 return ERR_CAST(task);
1357 }
1358 atomic_set(&group->rtpoll_wakeup, 0);
1359 wake_up_process(task);
1360 rcu_assign_pointer(group->rtpoll_task, task);
1361 }
1362
1363 list_add(&t->node, &group->rtpoll_triggers);
1364 group->rtpoll_min_period = min(group->rtpoll_min_period,
1365 div_u64(t->win.size, UPDATES_PER_WINDOW));
1366 group->rtpoll_nr_triggers[t->state]++;
1367 group->rtpoll_states |= (1 << t->state);
1368
1369 mutex_unlock(&group->rtpoll_trigger_lock);
1370 } else {
1371 mutex_lock(&group->avgs_lock);
1372
1373 list_add(&t->node, &group->avg_triggers);
1374 group->avg_nr_triggers[t->state]++;
1375
1376 mutex_unlock(&group->avgs_lock);
1377 }
1378 return t;
1379}
1380
1381void psi_trigger_destroy(struct psi_trigger *t)
1382{
1383 struct psi_group *group;
1384 struct task_struct *task_to_destroy = NULL;
1385
1386 /*
1387 * We do not check psi_disabled since it might have been disabled after
1388 * the trigger got created.
1389 */
1390 if (!t)
1391 return;
1392
1393 group = t->group;
1394 /*
1395 * Wakeup waiters to stop polling and clear the queue to prevent it from
1396 * being accessed later. Can happen if cgroup is deleted from under a
1397 * polling process.
1398 */
1399 if (t->of)
1400 kernfs_notify(t->of->kn);
1401 else
1402 wake_up_interruptible(&t->event_wait);
1403
1404 if (t->aggregator == PSI_AVGS) {
1405 mutex_lock(&group->avgs_lock);
1406 if (!list_empty(&t->node)) {
1407 list_del(&t->node);
1408 group->avg_nr_triggers[t->state]--;
1409 }
1410 mutex_unlock(&group->avgs_lock);
1411 } else {
1412 mutex_lock(&group->rtpoll_trigger_lock);
1413 if (!list_empty(&t->node)) {
1414 struct psi_trigger *tmp;
1415 u64 period = ULLONG_MAX;
1416
1417 list_del(&t->node);
1418 group->rtpoll_nr_triggers[t->state]--;
1419 if (!group->rtpoll_nr_triggers[t->state])
1420 group->rtpoll_states &= ~(1 << t->state);
1421 /*
1422 * Reset min update period for the remaining triggers
1423 * iff the destroying trigger had the min window size.
1424 */
1425 if (group->rtpoll_min_period == div_u64(t->win.size, UPDATES_PER_WINDOW)) {
1426 list_for_each_entry(tmp, &group->rtpoll_triggers, node)
1427 period = min(period, div_u64(tmp->win.size,
1428 UPDATES_PER_WINDOW));
1429 group->rtpoll_min_period = period;
1430 }
1431 /* Destroy rtpoll_task when the last trigger is destroyed */
1432 if (group->rtpoll_states == 0) {
1433 group->rtpoll_until = 0;
1434 task_to_destroy = rcu_dereference_protected(
1435 group->rtpoll_task,
1436 lockdep_is_held(&group->rtpoll_trigger_lock));
1437 rcu_assign_pointer(group->rtpoll_task, NULL);
1438 del_timer(&group->rtpoll_timer);
1439 }
1440 }
1441 mutex_unlock(&group->rtpoll_trigger_lock);
1442 }
1443
1444 /*
1445 * Wait for psi_schedule_rtpoll_work RCU to complete its read-side
1446 * critical section before destroying the trigger and optionally the
1447 * rtpoll_task.
1448 */
1449 synchronize_rcu();
1450 /*
1451 * Stop kthread 'psimon' after releasing rtpoll_trigger_lock to prevent
1452 * a deadlock while waiting for psi_rtpoll_work to acquire
1453 * rtpoll_trigger_lock
1454 */
1455 if (task_to_destroy) {
1456 /*
1457 * After the RCU grace period has expired, the worker
1458 * can no longer be found through group->rtpoll_task.
1459 */
1460 kthread_stop(task_to_destroy);
1461 atomic_set(&group->rtpoll_scheduled, 0);
1462 }
1463 kfree(t);
1464}
1465
1466__poll_t psi_trigger_poll(void **trigger_ptr,
1467 struct file *file, poll_table *wait)
1468{
1469 __poll_t ret = DEFAULT_POLLMASK;
1470 struct psi_trigger *t;
1471
1472 if (static_branch_likely(&psi_disabled))
1473 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1474
1475 t = smp_load_acquire(trigger_ptr);
1476 if (!t)
1477 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1478
1479 if (t->of)
1480 kernfs_generic_poll(t->of, wait);
1481 else
1482 poll_wait(file, &t->event_wait, wait);
1483
1484 if (cmpxchg(&t->event, 1, 0) == 1)
1485 ret |= EPOLLPRI;
1486
1487 return ret;
1488}
1489
1490#ifdef CONFIG_PROC_FS
1491static int psi_io_show(struct seq_file *m, void *v)
1492{
1493 return psi_show(m, &psi_system, PSI_IO);
1494}
1495
1496static int psi_memory_show(struct seq_file *m, void *v)
1497{
1498 return psi_show(m, &psi_system, PSI_MEM);
1499}
1500
1501static int psi_cpu_show(struct seq_file *m, void *v)
1502{
1503 return psi_show(m, &psi_system, PSI_CPU);
1504}
1505
1506static int psi_io_open(struct inode *inode, struct file *file)
1507{
1508 return single_open(file, psi_io_show, NULL);
1509}
1510
1511static int psi_memory_open(struct inode *inode, struct file *file)
1512{
1513 return single_open(file, psi_memory_show, NULL);
1514}
1515
1516static int psi_cpu_open(struct inode *inode, struct file *file)
1517{
1518 return single_open(file, psi_cpu_show, NULL);
1519}
1520
1521static ssize_t psi_write(struct file *file, const char __user *user_buf,
1522 size_t nbytes, enum psi_res res)
1523{
1524 char buf[32];
1525 size_t buf_size;
1526 struct seq_file *seq;
1527 struct psi_trigger *new;
1528
1529 if (static_branch_likely(&psi_disabled))
1530 return -EOPNOTSUPP;
1531
1532 if (!nbytes)
1533 return -EINVAL;
1534
1535 buf_size = min(nbytes, sizeof(buf));
1536 if (copy_from_user(buf, user_buf, buf_size))
1537 return -EFAULT;
1538
1539 buf[buf_size - 1] = '\0';
1540
1541 seq = file->private_data;
1542
1543 /* Take seq->lock to protect seq->private from concurrent writes */
1544 mutex_lock(&seq->lock);
1545
1546 /* Allow only one trigger per file descriptor */
1547 if (seq->private) {
1548 mutex_unlock(&seq->lock);
1549 return -EBUSY;
1550 }
1551
1552 new = psi_trigger_create(&psi_system, buf, res, file, NULL);
1553 if (IS_ERR(new)) {
1554 mutex_unlock(&seq->lock);
1555 return PTR_ERR(new);
1556 }
1557
1558 smp_store_release(&seq->private, new);
1559 mutex_unlock(&seq->lock);
1560
1561 return nbytes;
1562}
1563
1564static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1565 size_t nbytes, loff_t *ppos)
1566{
1567 return psi_write(file, user_buf, nbytes, PSI_IO);
1568}
1569
1570static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1571 size_t nbytes, loff_t *ppos)
1572{
1573 return psi_write(file, user_buf, nbytes, PSI_MEM);
1574}
1575
1576static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1577 size_t nbytes, loff_t *ppos)
1578{
1579 return psi_write(file, user_buf, nbytes, PSI_CPU);
1580}
1581
1582static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1583{
1584 struct seq_file *seq = file->private_data;
1585
1586 return psi_trigger_poll(&seq->private, file, wait);
1587}
1588
1589static int psi_fop_release(struct inode *inode, struct file *file)
1590{
1591 struct seq_file *seq = file->private_data;
1592
1593 psi_trigger_destroy(seq->private);
1594 return single_release(inode, file);
1595}
1596
1597static const struct proc_ops psi_io_proc_ops = {
1598 .proc_open = psi_io_open,
1599 .proc_read = seq_read,
1600 .proc_lseek = seq_lseek,
1601 .proc_write = psi_io_write,
1602 .proc_poll = psi_fop_poll,
1603 .proc_release = psi_fop_release,
1604};
1605
1606static const struct proc_ops psi_memory_proc_ops = {
1607 .proc_open = psi_memory_open,
1608 .proc_read = seq_read,
1609 .proc_lseek = seq_lseek,
1610 .proc_write = psi_memory_write,
1611 .proc_poll = psi_fop_poll,
1612 .proc_release = psi_fop_release,
1613};
1614
1615static const struct proc_ops psi_cpu_proc_ops = {
1616 .proc_open = psi_cpu_open,
1617 .proc_read = seq_read,
1618 .proc_lseek = seq_lseek,
1619 .proc_write = psi_cpu_write,
1620 .proc_poll = psi_fop_poll,
1621 .proc_release = psi_fop_release,
1622};
1623
1624#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1625static int psi_irq_show(struct seq_file *m, void *v)
1626{
1627 return psi_show(m, &psi_system, PSI_IRQ);
1628}
1629
1630static int psi_irq_open(struct inode *inode, struct file *file)
1631{
1632 return single_open(file, psi_irq_show, NULL);
1633}
1634
1635static ssize_t psi_irq_write(struct file *file, const char __user *user_buf,
1636 size_t nbytes, loff_t *ppos)
1637{
1638 return psi_write(file, user_buf, nbytes, PSI_IRQ);
1639}
1640
1641static const struct proc_ops psi_irq_proc_ops = {
1642 .proc_open = psi_irq_open,
1643 .proc_read = seq_read,
1644 .proc_lseek = seq_lseek,
1645 .proc_write = psi_irq_write,
1646 .proc_poll = psi_fop_poll,
1647 .proc_release = psi_fop_release,
1648};
1649#endif
1650
1651static int __init psi_proc_init(void)
1652{
1653 if (psi_enable) {
1654 proc_mkdir("pressure", NULL);
1655 proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops);
1656 proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops);
1657 proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops);
1658#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1659 proc_create("pressure/irq", 0666, NULL, &psi_irq_proc_ops);
1660#endif
1661 }
1662 return 0;
1663}
1664module_init(psi_proc_init);
1665
1666#endif /* CONFIG_PROC_FS */