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/*
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_last_update = sched_clock();
189 group->avg_next_update = group->avg_last_update + psi_period;
190 INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
191 mutex_init(&group->avgs_lock);
192 /* Init trigger-related members */
193 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_task, 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] > tasks[NR_ONCPU];
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 += div64_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/* Schedule polling if it's not already scheduled. */
550static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay)
551{
552 struct task_struct *task;
553
554 /*
555 * Do not reschedule if already scheduled.
556 * Possible race with a timer scheduled after this check but before
557 * mod_timer below can be tolerated because group->polling_next_update
558 * will keep updates on schedule.
559 */
560 if (timer_pending(&group->poll_timer))
561 return;
562
563 rcu_read_lock();
564
565 task = rcu_dereference(group->poll_task);
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(task))
571 mod_timer(&group->poll_timer, jiffies + delay);
572
573 rcu_read_unlock();
574}
575
576static void psi_poll_work(struct psi_group *group)
577{
578 u32 changed_states;
579 u64 now;
580
581 mutex_lock(&group->trigger_lock);
582
583 now = sched_clock();
584
585 collect_percpu_times(group, PSI_POLL, &changed_states);
586
587 if (changed_states & group->poll_states) {
588 /* Initialize trigger windows when entering polling mode */
589 if (now > group->polling_until)
590 init_triggers(group, now);
591
592 /*
593 * Keep the monitor active for at least the duration of the
594 * minimum tracking window as long as monitor states are
595 * changing.
596 */
597 group->polling_until = now +
598 group->poll_min_period * UPDATES_PER_WINDOW;
599 }
600
601 if (now > group->polling_until) {
602 group->polling_next_update = ULLONG_MAX;
603 goto out;
604 }
605
606 if (now >= group->polling_next_update)
607 group->polling_next_update = update_triggers(group, now);
608
609 psi_schedule_poll_work(group,
610 nsecs_to_jiffies(group->polling_next_update - now) + 1);
611
612out:
613 mutex_unlock(&group->trigger_lock);
614}
615
616static int psi_poll_worker(void *data)
617{
618 struct psi_group *group = (struct psi_group *)data;
619
620 sched_set_fifo_low(current);
621
622 while (true) {
623 wait_event_interruptible(group->poll_wait,
624 atomic_cmpxchg(&group->poll_wakeup, 1, 0) ||
625 kthread_should_stop());
626 if (kthread_should_stop())
627 break;
628
629 psi_poll_work(group);
630 }
631 return 0;
632}
633
634static void poll_timer_fn(struct timer_list *t)
635{
636 struct psi_group *group = from_timer(group, t, poll_timer);
637
638 atomic_set(&group->poll_wakeup, 1);
639 wake_up_interruptible(&group->poll_wait);
640}
641
642static void record_times(struct psi_group_cpu *groupc, int cpu,
643 bool memstall_tick)
644{
645 u32 delta;
646 u64 now;
647
648 now = cpu_clock(cpu);
649 delta = now - groupc->state_start;
650 groupc->state_start = now;
651
652 if (groupc->state_mask & (1 << PSI_IO_SOME)) {
653 groupc->times[PSI_IO_SOME] += delta;
654 if (groupc->state_mask & (1 << PSI_IO_FULL))
655 groupc->times[PSI_IO_FULL] += delta;
656 }
657
658 if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
659 groupc->times[PSI_MEM_SOME] += delta;
660 if (groupc->state_mask & (1 << PSI_MEM_FULL))
661 groupc->times[PSI_MEM_FULL] += delta;
662 else if (memstall_tick) {
663 u32 sample;
664 /*
665 * Since we care about lost potential, a
666 * memstall is FULL when there are no other
667 * working tasks, but also when the CPU is
668 * actively reclaiming and nothing productive
669 * could run even if it were runnable.
670 *
671 * When the timer tick sees a reclaiming CPU,
672 * regardless of runnable tasks, sample a FULL
673 * tick (or less if it hasn't been a full tick
674 * since the last state change).
675 */
676 sample = min(delta, (u32)jiffies_to_nsecs(1));
677 groupc->times[PSI_MEM_FULL] += sample;
678 }
679 }
680
681 if (groupc->state_mask & (1 << PSI_CPU_SOME))
682 groupc->times[PSI_CPU_SOME] += delta;
683
684 if (groupc->state_mask & (1 << PSI_NONIDLE))
685 groupc->times[PSI_NONIDLE] += delta;
686}
687
688static void psi_group_change(struct psi_group *group, int cpu,
689 unsigned int clear, unsigned int set,
690 bool wake_clock)
691{
692 struct psi_group_cpu *groupc;
693 u32 state_mask = 0;
694 unsigned int t, m;
695 enum psi_states s;
696
697 groupc = per_cpu_ptr(group->pcpu, cpu);
698
699 /*
700 * First we assess the aggregate resource states this CPU's
701 * tasks have been in since the last change, and account any
702 * SOME and FULL time these may have resulted in.
703 *
704 * Then we update the task counts according to the state
705 * change requested through the @clear and @set bits.
706 */
707 write_seqcount_begin(&groupc->seq);
708
709 record_times(groupc, cpu, false);
710
711 for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
712 if (!(m & (1 << t)))
713 continue;
714 if (groupc->tasks[t] == 0 && !psi_bug) {
715 printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n",
716 cpu, t, groupc->tasks[0],
717 groupc->tasks[1], groupc->tasks[2],
718 groupc->tasks[3], clear, set);
719 psi_bug = 1;
720 }
721 groupc->tasks[t]--;
722 }
723
724 for (t = 0; set; set &= ~(1 << t), t++)
725 if (set & (1 << t))
726 groupc->tasks[t]++;
727
728 /* Calculate state mask representing active states */
729 for (s = 0; s < NR_PSI_STATES; s++) {
730 if (test_state(groupc->tasks, s))
731 state_mask |= (1 << s);
732 }
733 groupc->state_mask = state_mask;
734
735 write_seqcount_end(&groupc->seq);
736
737 if (state_mask & group->poll_states)
738 psi_schedule_poll_work(group, 1);
739
740 if (wake_clock && !delayed_work_pending(&group->avgs_work))
741 schedule_delayed_work(&group->avgs_work, PSI_FREQ);
742}
743
744static struct psi_group *iterate_groups(struct task_struct *task, void **iter)
745{
746#ifdef CONFIG_CGROUPS
747 struct cgroup *cgroup = NULL;
748
749 if (!*iter)
750 cgroup = task->cgroups->dfl_cgrp;
751 else if (*iter == &psi_system)
752 return NULL;
753 else
754 cgroup = cgroup_parent(*iter);
755
756 if (cgroup && cgroup_parent(cgroup)) {
757 *iter = cgroup;
758 return cgroup_psi(cgroup);
759 }
760#else
761 if (*iter)
762 return NULL;
763#endif
764 *iter = &psi_system;
765 return &psi_system;
766}
767
768static void psi_flags_change(struct task_struct *task, int clear, int set)
769{
770 if (((task->psi_flags & set) ||
771 (task->psi_flags & clear) != clear) &&
772 !psi_bug) {
773 printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
774 task->pid, task->comm, task_cpu(task),
775 task->psi_flags, clear, set);
776 psi_bug = 1;
777 }
778
779 task->psi_flags &= ~clear;
780 task->psi_flags |= set;
781}
782
783void psi_task_change(struct task_struct *task, int clear, int set)
784{
785 int cpu = task_cpu(task);
786 struct psi_group *group;
787 bool wake_clock = true;
788 void *iter = NULL;
789
790 if (!task->pid)
791 return;
792
793 psi_flags_change(task, clear, set);
794
795 /*
796 * Periodic aggregation shuts off if there is a period of no
797 * task changes, so we wake it back up if necessary. However,
798 * don't do this if the task change is the aggregation worker
799 * itself going to sleep, or we'll ping-pong forever.
800 */
801 if (unlikely((clear & TSK_RUNNING) &&
802 (task->flags & PF_WQ_WORKER) &&
803 wq_worker_last_func(task) == psi_avgs_work))
804 wake_clock = false;
805
806 while ((group = iterate_groups(task, &iter)))
807 psi_group_change(group, cpu, clear, set, wake_clock);
808}
809
810void psi_task_switch(struct task_struct *prev, struct task_struct *next,
811 bool sleep)
812{
813 struct psi_group *group, *common = NULL;
814 int cpu = task_cpu(prev);
815 void *iter;
816
817 if (next->pid) {
818 psi_flags_change(next, 0, TSK_ONCPU);
819 /*
820 * When moving state between tasks, the group that
821 * contains them both does not change: we can stop
822 * updating the tree once we reach the first common
823 * ancestor. Iterate @next's ancestors until we
824 * encounter @prev's state.
825 */
826 iter = NULL;
827 while ((group = iterate_groups(next, &iter))) {
828 if (per_cpu_ptr(group->pcpu, cpu)->tasks[NR_ONCPU]) {
829 common = group;
830 break;
831 }
832
833 psi_group_change(group, cpu, 0, TSK_ONCPU, true);
834 }
835 }
836
837 /*
838 * If this is a voluntary sleep, dequeue will have taken care
839 * of the outgoing TSK_ONCPU alongside TSK_RUNNING already. We
840 * only need to deal with it during preemption.
841 */
842 if (sleep)
843 return;
844
845 if (prev->pid) {
846 psi_flags_change(prev, TSK_ONCPU, 0);
847
848 iter = NULL;
849 while ((group = iterate_groups(prev, &iter)) && group != common)
850 psi_group_change(group, cpu, TSK_ONCPU, 0, true);
851 }
852}
853
854void psi_memstall_tick(struct task_struct *task, int cpu)
855{
856 struct psi_group *group;
857 void *iter = NULL;
858
859 while ((group = iterate_groups(task, &iter))) {
860 struct psi_group_cpu *groupc;
861
862 groupc = per_cpu_ptr(group->pcpu, cpu);
863 write_seqcount_begin(&groupc->seq);
864 record_times(groupc, cpu, true);
865 write_seqcount_end(&groupc->seq);
866 }
867}
868
869/**
870 * psi_memstall_enter - mark the beginning of a memory stall section
871 * @flags: flags to handle nested sections
872 *
873 * Marks the calling task as being stalled due to a lack of memory,
874 * such as waiting for a refault or performing reclaim.
875 */
876void psi_memstall_enter(unsigned long *flags)
877{
878 struct rq_flags rf;
879 struct rq *rq;
880
881 if (static_branch_likely(&psi_disabled))
882 return;
883
884 *flags = current->in_memstall;
885 if (*flags)
886 return;
887 /*
888 * in_memstall setting & accounting needs to be atomic wrt
889 * changes to the task's scheduling state, otherwise we can
890 * race with CPU migration.
891 */
892 rq = this_rq_lock_irq(&rf);
893
894 current->in_memstall = 1;
895 psi_task_change(current, 0, TSK_MEMSTALL);
896
897 rq_unlock_irq(rq, &rf);
898}
899
900/**
901 * psi_memstall_leave - mark the end of an memory stall section
902 * @flags: flags to handle nested memdelay sections
903 *
904 * Marks the calling task as no longer stalled due to lack of memory.
905 */
906void psi_memstall_leave(unsigned long *flags)
907{
908 struct rq_flags rf;
909 struct rq *rq;
910
911 if (static_branch_likely(&psi_disabled))
912 return;
913
914 if (*flags)
915 return;
916 /*
917 * in_memstall clearing & accounting needs to be atomic wrt
918 * changes to the task's scheduling state, otherwise we could
919 * race with CPU migration.
920 */
921 rq = this_rq_lock_irq(&rf);
922
923 current->in_memstall = 0;
924 psi_task_change(current, TSK_MEMSTALL, 0);
925
926 rq_unlock_irq(rq, &rf);
927}
928
929#ifdef CONFIG_CGROUPS
930int psi_cgroup_alloc(struct cgroup *cgroup)
931{
932 if (static_branch_likely(&psi_disabled))
933 return 0;
934
935 cgroup->psi.pcpu = alloc_percpu(struct psi_group_cpu);
936 if (!cgroup->psi.pcpu)
937 return -ENOMEM;
938 group_init(&cgroup->psi);
939 return 0;
940}
941
942void psi_cgroup_free(struct cgroup *cgroup)
943{
944 if (static_branch_likely(&psi_disabled))
945 return;
946
947 cancel_delayed_work_sync(&cgroup->psi.avgs_work);
948 free_percpu(cgroup->psi.pcpu);
949 /* All triggers must be removed by now */
950 WARN_ONCE(cgroup->psi.poll_states, "psi: trigger leak\n");
951}
952
953/**
954 * cgroup_move_task - move task to a different cgroup
955 * @task: the task
956 * @to: the target css_set
957 *
958 * Move task to a new cgroup and safely migrate its associated stall
959 * state between the different groups.
960 *
961 * This function acquires the task's rq lock to lock out concurrent
962 * changes to the task's scheduling state and - in case the task is
963 * running - concurrent changes to its stall state.
964 */
965void cgroup_move_task(struct task_struct *task, struct css_set *to)
966{
967 unsigned int task_flags = 0;
968 struct rq_flags rf;
969 struct rq *rq;
970
971 if (static_branch_likely(&psi_disabled)) {
972 /*
973 * Lame to do this here, but the scheduler cannot be locked
974 * from the outside, so we move cgroups from inside sched/.
975 */
976 rcu_assign_pointer(task->cgroups, to);
977 return;
978 }
979
980 rq = task_rq_lock(task, &rf);
981
982 if (task_on_rq_queued(task)) {
983 task_flags = TSK_RUNNING;
984 if (task_current(rq, task))
985 task_flags |= TSK_ONCPU;
986 } else if (task->in_iowait)
987 task_flags = TSK_IOWAIT;
988
989 if (task->in_memstall)
990 task_flags |= TSK_MEMSTALL;
991
992 if (task_flags)
993 psi_task_change(task, task_flags, 0);
994
995 /* See comment above */
996 rcu_assign_pointer(task->cgroups, to);
997
998 if (task_flags)
999 psi_task_change(task, 0, task_flags);
1000
1001 task_rq_unlock(rq, task, &rf);
1002}
1003#endif /* CONFIG_CGROUPS */
1004
1005int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1006{
1007 int full;
1008 u64 now;
1009
1010 if (static_branch_likely(&psi_disabled))
1011 return -EOPNOTSUPP;
1012
1013 /* Update averages before reporting them */
1014 mutex_lock(&group->avgs_lock);
1015 now = sched_clock();
1016 collect_percpu_times(group, PSI_AVGS, NULL);
1017 if (now >= group->avg_next_update)
1018 group->avg_next_update = update_averages(group, now);
1019 mutex_unlock(&group->avgs_lock);
1020
1021 for (full = 0; full < 2 - (res == PSI_CPU); full++) {
1022 unsigned long avg[3];
1023 u64 total;
1024 int w;
1025
1026 for (w = 0; w < 3; w++)
1027 avg[w] = group->avg[res * 2 + full][w];
1028 total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1029 NSEC_PER_USEC);
1030
1031 seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1032 full ? "full" : "some",
1033 LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1034 LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1035 LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1036 total);
1037 }
1038
1039 return 0;
1040}
1041
1042static int psi_io_show(struct seq_file *m, void *v)
1043{
1044 return psi_show(m, &psi_system, PSI_IO);
1045}
1046
1047static int psi_memory_show(struct seq_file *m, void *v)
1048{
1049 return psi_show(m, &psi_system, PSI_MEM);
1050}
1051
1052static int psi_cpu_show(struct seq_file *m, void *v)
1053{
1054 return psi_show(m, &psi_system, PSI_CPU);
1055}
1056
1057static int psi_io_open(struct inode *inode, struct file *file)
1058{
1059 return single_open(file, psi_io_show, NULL);
1060}
1061
1062static int psi_memory_open(struct inode *inode, struct file *file)
1063{
1064 return single_open(file, psi_memory_show, NULL);
1065}
1066
1067static int psi_cpu_open(struct inode *inode, struct file *file)
1068{
1069 return single_open(file, psi_cpu_show, NULL);
1070}
1071
1072struct psi_trigger *psi_trigger_create(struct psi_group *group,
1073 char *buf, size_t nbytes, enum psi_res res)
1074{
1075 struct psi_trigger *t;
1076 enum psi_states state;
1077 u32 threshold_us;
1078 u32 window_us;
1079
1080 if (static_branch_likely(&psi_disabled))
1081 return ERR_PTR(-EOPNOTSUPP);
1082
1083 if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1084 state = PSI_IO_SOME + res * 2;
1085 else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1086 state = PSI_IO_FULL + res * 2;
1087 else
1088 return ERR_PTR(-EINVAL);
1089
1090 if (state >= PSI_NONIDLE)
1091 return ERR_PTR(-EINVAL);
1092
1093 if (window_us < WINDOW_MIN_US ||
1094 window_us > WINDOW_MAX_US)
1095 return ERR_PTR(-EINVAL);
1096
1097 /* Check threshold */
1098 if (threshold_us == 0 || threshold_us > window_us)
1099 return ERR_PTR(-EINVAL);
1100
1101 t = kmalloc(sizeof(*t), GFP_KERNEL);
1102 if (!t)
1103 return ERR_PTR(-ENOMEM);
1104
1105 t->group = group;
1106 t->state = state;
1107 t->threshold = threshold_us * NSEC_PER_USEC;
1108 t->win.size = window_us * NSEC_PER_USEC;
1109 window_reset(&t->win, 0, 0, 0);
1110
1111 t->event = 0;
1112 t->last_event_time = 0;
1113 init_waitqueue_head(&t->event_wait);
1114 kref_init(&t->refcount);
1115
1116 mutex_lock(&group->trigger_lock);
1117
1118 if (!rcu_access_pointer(group->poll_task)) {
1119 struct task_struct *task;
1120
1121 task = kthread_create(psi_poll_worker, group, "psimon");
1122 if (IS_ERR(task)) {
1123 kfree(t);
1124 mutex_unlock(&group->trigger_lock);
1125 return ERR_CAST(task);
1126 }
1127 atomic_set(&group->poll_wakeup, 0);
1128 init_waitqueue_head(&group->poll_wait);
1129 wake_up_process(task);
1130 timer_setup(&group->poll_timer, poll_timer_fn, 0);
1131 rcu_assign_pointer(group->poll_task, task);
1132 }
1133
1134 list_add(&t->node, &group->triggers);
1135 group->poll_min_period = min(group->poll_min_period,
1136 div_u64(t->win.size, UPDATES_PER_WINDOW));
1137 group->nr_triggers[t->state]++;
1138 group->poll_states |= (1 << t->state);
1139
1140 mutex_unlock(&group->trigger_lock);
1141
1142 return t;
1143}
1144
1145static void psi_trigger_destroy(struct kref *ref)
1146{
1147 struct psi_trigger *t = container_of(ref, struct psi_trigger, refcount);
1148 struct psi_group *group = t->group;
1149 struct task_struct *task_to_destroy = NULL;
1150
1151 if (static_branch_likely(&psi_disabled))
1152 return;
1153
1154 /*
1155 * Wakeup waiters to stop polling. Can happen if cgroup is deleted
1156 * from under a polling process.
1157 */
1158 wake_up_interruptible(&t->event_wait);
1159
1160 mutex_lock(&group->trigger_lock);
1161
1162 if (!list_empty(&t->node)) {
1163 struct psi_trigger *tmp;
1164 u64 period = ULLONG_MAX;
1165
1166 list_del(&t->node);
1167 group->nr_triggers[t->state]--;
1168 if (!group->nr_triggers[t->state])
1169 group->poll_states &= ~(1 << t->state);
1170 /* reset min update period for the remaining triggers */
1171 list_for_each_entry(tmp, &group->triggers, node)
1172 period = min(period, div_u64(tmp->win.size,
1173 UPDATES_PER_WINDOW));
1174 group->poll_min_period = period;
1175 /* Destroy poll_task when the last trigger is destroyed */
1176 if (group->poll_states == 0) {
1177 group->polling_until = 0;
1178 task_to_destroy = rcu_dereference_protected(
1179 group->poll_task,
1180 lockdep_is_held(&group->trigger_lock));
1181 rcu_assign_pointer(group->poll_task, NULL);
1182 }
1183 }
1184
1185 mutex_unlock(&group->trigger_lock);
1186
1187 /*
1188 * Wait for both *trigger_ptr from psi_trigger_replace and
1189 * poll_task RCUs to complete their read-side critical sections
1190 * before destroying the trigger and optionally the poll_task
1191 */
1192 synchronize_rcu();
1193 /*
1194 * Destroy the kworker after releasing trigger_lock to prevent a
1195 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1196 */
1197 if (task_to_destroy) {
1198 /*
1199 * After the RCU grace period has expired, the worker
1200 * can no longer be found through group->poll_task.
1201 * But it might have been already scheduled before
1202 * that - deschedule it cleanly before destroying it.
1203 */
1204 del_timer_sync(&group->poll_timer);
1205 kthread_stop(task_to_destroy);
1206 }
1207 kfree(t);
1208}
1209
1210void psi_trigger_replace(void **trigger_ptr, struct psi_trigger *new)
1211{
1212 struct psi_trigger *old = *trigger_ptr;
1213
1214 if (static_branch_likely(&psi_disabled))
1215 return;
1216
1217 rcu_assign_pointer(*trigger_ptr, new);
1218 if (old)
1219 kref_put(&old->refcount, psi_trigger_destroy);
1220}
1221
1222__poll_t psi_trigger_poll(void **trigger_ptr,
1223 struct file *file, poll_table *wait)
1224{
1225 __poll_t ret = DEFAULT_POLLMASK;
1226 struct psi_trigger *t;
1227
1228 if (static_branch_likely(&psi_disabled))
1229 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1230
1231 rcu_read_lock();
1232
1233 t = rcu_dereference(*(void __rcu __force **)trigger_ptr);
1234 if (!t) {
1235 rcu_read_unlock();
1236 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1237 }
1238 kref_get(&t->refcount);
1239
1240 rcu_read_unlock();
1241
1242 poll_wait(file, &t->event_wait, wait);
1243
1244 if (cmpxchg(&t->event, 1, 0) == 1)
1245 ret |= EPOLLPRI;
1246
1247 kref_put(&t->refcount, psi_trigger_destroy);
1248
1249 return ret;
1250}
1251
1252static ssize_t psi_write(struct file *file, const char __user *user_buf,
1253 size_t nbytes, enum psi_res res)
1254{
1255 char buf[32];
1256 size_t buf_size;
1257 struct seq_file *seq;
1258 struct psi_trigger *new;
1259
1260 if (static_branch_likely(&psi_disabled))
1261 return -EOPNOTSUPP;
1262
1263 if (!nbytes)
1264 return -EINVAL;
1265
1266 buf_size = min(nbytes, sizeof(buf));
1267 if (copy_from_user(buf, user_buf, buf_size))
1268 return -EFAULT;
1269
1270 buf[buf_size - 1] = '\0';
1271
1272 new = psi_trigger_create(&psi_system, buf, nbytes, res);
1273 if (IS_ERR(new))
1274 return PTR_ERR(new);
1275
1276 seq = file->private_data;
1277 /* Take seq->lock to protect seq->private from concurrent writes */
1278 mutex_lock(&seq->lock);
1279 psi_trigger_replace(&seq->private, new);
1280 mutex_unlock(&seq->lock);
1281
1282 return nbytes;
1283}
1284
1285static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1286 size_t nbytes, loff_t *ppos)
1287{
1288 return psi_write(file, user_buf, nbytes, PSI_IO);
1289}
1290
1291static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1292 size_t nbytes, loff_t *ppos)
1293{
1294 return psi_write(file, user_buf, nbytes, PSI_MEM);
1295}
1296
1297static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1298 size_t nbytes, loff_t *ppos)
1299{
1300 return psi_write(file, user_buf, nbytes, PSI_CPU);
1301}
1302
1303static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1304{
1305 struct seq_file *seq = file->private_data;
1306
1307 return psi_trigger_poll(&seq->private, file, wait);
1308}
1309
1310static int psi_fop_release(struct inode *inode, struct file *file)
1311{
1312 struct seq_file *seq = file->private_data;
1313
1314 psi_trigger_replace(&seq->private, NULL);
1315 return single_release(inode, file);
1316}
1317
1318static const struct proc_ops psi_io_proc_ops = {
1319 .proc_open = psi_io_open,
1320 .proc_read = seq_read,
1321 .proc_lseek = seq_lseek,
1322 .proc_write = psi_io_write,
1323 .proc_poll = psi_fop_poll,
1324 .proc_release = psi_fop_release,
1325};
1326
1327static const struct proc_ops psi_memory_proc_ops = {
1328 .proc_open = psi_memory_open,
1329 .proc_read = seq_read,
1330 .proc_lseek = seq_lseek,
1331 .proc_write = psi_memory_write,
1332 .proc_poll = psi_fop_poll,
1333 .proc_release = psi_fop_release,
1334};
1335
1336static const struct proc_ops psi_cpu_proc_ops = {
1337 .proc_open = psi_cpu_open,
1338 .proc_read = seq_read,
1339 .proc_lseek = seq_lseek,
1340 .proc_write = psi_cpu_write,
1341 .proc_poll = psi_fop_poll,
1342 .proc_release = psi_fop_release,
1343};
1344
1345static int __init psi_proc_init(void)
1346{
1347 if (psi_enable) {
1348 proc_mkdir("pressure", NULL);
1349 proc_create("pressure/io", 0, NULL, &psi_io_proc_ops);
1350 proc_create("pressure/memory", 0, NULL, &psi_memory_proc_ops);
1351 proc_create("pressure/cpu", 0, NULL, &psi_cpu_proc_ops);
1352 }
1353 return 0;
1354}
1355module_init(psi_proc_init);