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