Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Pressure stall information for CPU, memory and IO
4 *
5 * Copyright (c) 2018 Facebook, Inc.
6 * Author: Johannes Weiner <hannes@cmpxchg.org>
7 *
8 * Polling support by Suren Baghdasaryan <surenb@google.com>
9 * Copyright (c) 2018 Google, Inc.
10 *
11 * When CPU, memory and IO are contended, tasks experience delays that
12 * reduce throughput and introduce latencies into the workload. Memory
13 * and IO contention, in addition, can cause a full loss of forward
14 * progress in which the CPU goes idle.
15 *
16 * This code aggregates individual task delays into resource pressure
17 * metrics that indicate problems with both workload health and
18 * resource utilization.
19 *
20 * Model
21 *
22 * The time in which a task can execute on a CPU is our baseline for
23 * productivity. Pressure expresses the amount of time in which this
24 * potential cannot be realized due to resource contention.
25 *
26 * This concept of productivity has two components: the workload and
27 * the CPU. To measure the impact of pressure on both, we define two
28 * contention states for a resource: SOME and FULL.
29 *
30 * In the SOME state of a given resource, one or more tasks are
31 * delayed on that resource. This affects the workload's ability to
32 * perform work, but the CPU may still be executing other tasks.
33 *
34 * In the FULL state of a given resource, all non-idle tasks are
35 * delayed on that resource such that nobody is advancing and the CPU
36 * goes idle. This leaves both workload and CPU unproductive.
37 *
38 * SOME = nr_delayed_tasks != 0
39 * FULL = nr_delayed_tasks != 0 && nr_productive_tasks == 0
40 *
41 * What it means for a task to be productive is defined differently
42 * for each resource. For IO, productive means a running task. For
43 * memory, productive means a running task that isn't a reclaimer. For
44 * CPU, productive means an oncpu task.
45 *
46 * Naturally, the FULL state doesn't exist for the CPU resource at the
47 * system level, but exist at the cgroup level. At the cgroup level,
48 * FULL means all non-idle tasks in the cgroup are delayed on the CPU
49 * resource which is being used by others outside of the cgroup or
50 * throttled by the cgroup cpu.max configuration.
51 *
52 * The percentage of wallclock time spent in those compound stall
53 * states gives pressure numbers between 0 and 100 for each resource,
54 * where the SOME percentage indicates workload slowdowns and the FULL
55 * percentage indicates reduced CPU utilization:
56 *
57 * %SOME = time(SOME) / period
58 * %FULL = time(FULL) / period
59 *
60 * Multiple CPUs
61 *
62 * The more tasks and available CPUs there are, the more work can be
63 * performed concurrently. This means that the potential that can go
64 * unrealized due to resource contention *also* scales with non-idle
65 * tasks and CPUs.
66 *
67 * Consider a scenario where 257 number crunching tasks are trying to
68 * run concurrently on 256 CPUs. If we simply aggregated the task
69 * states, we would have to conclude a CPU SOME pressure number of
70 * 100%, since *somebody* is waiting on a runqueue at all
71 * times. However, that is clearly not the amount of contention the
72 * workload is experiencing: only one out of 256 possible execution
73 * threads will be contended at any given time, or about 0.4%.
74 *
75 * Conversely, consider a scenario of 4 tasks and 4 CPUs where at any
76 * given time *one* of the tasks is delayed due to a lack of memory.
77 * Again, looking purely at the task state would yield a memory FULL
78 * pressure number of 0%, since *somebody* is always making forward
79 * progress. But again this wouldn't capture the amount of execution
80 * potential lost, which is 1 out of 4 CPUs, or 25%.
81 *
82 * To calculate wasted potential (pressure) with multiple processors,
83 * we have to base our calculation on the number of non-idle tasks in
84 * conjunction with the number of available CPUs, which is the number
85 * of potential execution threads. SOME becomes then the proportion of
86 * delayed tasks to possible threads, and FULL is the share of possible
87 * threads that are unproductive due to delays:
88 *
89 * threads = min(nr_nonidle_tasks, nr_cpus)
90 * SOME = min(nr_delayed_tasks / threads, 1)
91 * FULL = (threads - min(nr_productive_tasks, threads)) / threads
92 *
93 * For the 257 number crunchers on 256 CPUs, this yields:
94 *
95 * threads = min(257, 256)
96 * SOME = min(1 / 256, 1) = 0.4%
97 * FULL = (256 - min(256, 256)) / 256 = 0%
98 *
99 * For the 1 out of 4 memory-delayed tasks, this yields:
100 *
101 * threads = min(4, 4)
102 * SOME = min(1 / 4, 1) = 25%
103 * FULL = (4 - min(3, 4)) / 4 = 25%
104 *
105 * [ Substitute nr_cpus with 1, and you can see that it's a natural
106 * extension of the single-CPU model. ]
107 *
108 * Implementation
109 *
110 * To assess the precise time spent in each such state, we would have
111 * to freeze the system on task changes and start/stop the state
112 * clocks accordingly. Obviously that doesn't scale in practice.
113 *
114 * Because the scheduler aims to distribute the compute load evenly
115 * among the available CPUs, we can track task state locally to each
116 * CPU and, at much lower frequency, extrapolate the global state for
117 * the cumulative stall times and the running averages.
118 *
119 * For each runqueue, we track:
120 *
121 * tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0)
122 * tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_productive_tasks[cpu])
123 * tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0)
124 *
125 * and then periodically aggregate:
126 *
127 * tNONIDLE = sum(tNONIDLE[i])
128 *
129 * tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE
130 * tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE
131 *
132 * %SOME = tSOME / period
133 * %FULL = tFULL / period
134 *
135 * This gives us an approximation of pressure that is practical
136 * cost-wise, yet way more sensitive and accurate than periodic
137 * sampling of the aggregate task states would be.
138 */
139
140static int psi_bug __read_mostly;
141
142DEFINE_STATIC_KEY_FALSE(psi_disabled);
143static DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled);
144
145#ifdef CONFIG_PSI_DEFAULT_DISABLED
146static bool psi_enable;
147#else
148static bool psi_enable = true;
149#endif
150static int __init setup_psi(char *str)
151{
152 return kstrtobool(str, &psi_enable) == 0;
153}
154__setup("psi=", setup_psi);
155
156/* Running averages - we need to be higher-res than loadavg */
157#define PSI_FREQ (2*HZ+1) /* 2 sec intervals */
158#define EXP_10s 1677 /* 1/exp(2s/10s) as fixed-point */
159#define EXP_60s 1981 /* 1/exp(2s/60s) */
160#define EXP_300s 2034 /* 1/exp(2s/300s) */
161
162/* PSI trigger definitions */
163#define WINDOW_MAX_US 10000000 /* Max window size is 10s */
164#define UPDATES_PER_WINDOW 10 /* 10 updates per window */
165
166/* Sampling frequency in nanoseconds */
167static u64 psi_period __read_mostly;
168
169/* System-level pressure and stall tracking */
170static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
171struct psi_group psi_system = {
172 .pcpu = &system_group_pcpu,
173};
174
175static void psi_avgs_work(struct work_struct *work);
176
177static void poll_timer_fn(struct timer_list *t);
178
179static void group_init(struct psi_group *group)
180{
181 int cpu;
182
183 group->enabled = true;
184 for_each_possible_cpu(cpu)
185 seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);
186 group->avg_last_update = sched_clock();
187 group->avg_next_update = group->avg_last_update + psi_period;
188 mutex_init(&group->avgs_lock);
189
190 /* Init avg trigger-related members */
191 INIT_LIST_HEAD(&group->avg_triggers);
192 memset(group->avg_nr_triggers, 0, sizeof(group->avg_nr_triggers));
193 INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
194
195 /* Init rtpoll trigger-related members */
196 atomic_set(&group->rtpoll_scheduled, 0);
197 mutex_init(&group->rtpoll_trigger_lock);
198 INIT_LIST_HEAD(&group->rtpoll_triggers);
199 group->rtpoll_min_period = U32_MAX;
200 group->rtpoll_next_update = ULLONG_MAX;
201 init_waitqueue_head(&group->rtpoll_wait);
202 timer_setup(&group->rtpoll_timer, poll_timer_fn, 0);
203 rcu_assign_pointer(group->rtpoll_task, NULL);
204}
205
206void __init psi_init(void)
207{
208 if (!psi_enable) {
209 static_branch_enable(&psi_disabled);
210 static_branch_disable(&psi_cgroups_enabled);
211 return;
212 }
213
214 if (!cgroup_psi_enabled())
215 static_branch_disable(&psi_cgroups_enabled);
216
217 psi_period = jiffies_to_nsecs(PSI_FREQ);
218 group_init(&psi_system);
219}
220
221static bool test_state(unsigned int *tasks, enum psi_states state, bool oncpu)
222{
223 switch (state) {
224 case PSI_IO_SOME:
225 return unlikely(tasks[NR_IOWAIT]);
226 case PSI_IO_FULL:
227 return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]);
228 case PSI_MEM_SOME:
229 return unlikely(tasks[NR_MEMSTALL]);
230 case PSI_MEM_FULL:
231 return unlikely(tasks[NR_MEMSTALL] &&
232 tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]);
233 case PSI_CPU_SOME:
234 return unlikely(tasks[NR_RUNNING] > oncpu);
235 case PSI_CPU_FULL:
236 return unlikely(tasks[NR_RUNNING] && !oncpu);
237 case PSI_NONIDLE:
238 return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
239 tasks[NR_RUNNING];
240 default:
241 return false;
242 }
243}
244
245static void get_recent_times(struct psi_group *group, int cpu,
246 enum psi_aggregators aggregator, u32 *times,
247 u32 *pchanged_states)
248{
249 struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
250 int current_cpu = raw_smp_processor_id();
251 unsigned int tasks[NR_PSI_TASK_COUNTS];
252 u64 now, state_start;
253 enum psi_states s;
254 unsigned int seq;
255 u32 state_mask;
256
257 *pchanged_states = 0;
258
259 /* Snapshot a coherent view of the CPU state */
260 do {
261 seq = read_seqcount_begin(&groupc->seq);
262 now = cpu_clock(cpu);
263 memcpy(times, groupc->times, sizeof(groupc->times));
264 state_mask = groupc->state_mask;
265 state_start = groupc->state_start;
266 if (cpu == current_cpu)
267 memcpy(tasks, groupc->tasks, sizeof(groupc->tasks));
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 /*
294 * When collect_percpu_times() from the avgs_work, we don't want to
295 * re-arm avgs_work when all CPUs are IDLE. But the current CPU running
296 * this avgs_work is never IDLE, cause avgs_work can't be shut off.
297 * So for the current CPU, we need to re-arm avgs_work only when
298 * (NR_RUNNING > 1 || NR_IOWAIT > 0 || NR_MEMSTALL > 0), for other CPUs
299 * we can just check PSI_NONIDLE delta.
300 */
301 if (current_work() == &group->avgs_work.work) {
302 bool reschedule;
303
304 if (cpu == current_cpu)
305 reschedule = tasks[NR_RUNNING] +
306 tasks[NR_IOWAIT] +
307 tasks[NR_MEMSTALL] > 1;
308 else
309 reschedule = *pchanged_states & (1 << PSI_NONIDLE);
310
311 if (reschedule)
312 *pchanged_states |= PSI_STATE_RESCHEDULE;
313 }
314}
315
316static void calc_avgs(unsigned long avg[3], int missed_periods,
317 u64 time, u64 period)
318{
319 unsigned long pct;
320
321 /* Fill in zeroes for periods of no activity */
322 if (missed_periods) {
323 avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
324 avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
325 avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
326 }
327
328 /* Sample the most recent active period */
329 pct = div_u64(time * 100, period);
330 pct *= FIXED_1;
331 avg[0] = calc_load(avg[0], EXP_10s, pct);
332 avg[1] = calc_load(avg[1], EXP_60s, pct);
333 avg[2] = calc_load(avg[2], EXP_300s, pct);
334}
335
336static void collect_percpu_times(struct psi_group *group,
337 enum psi_aggregators aggregator,
338 u32 *pchanged_states)
339{
340 u64 deltas[NR_PSI_STATES - 1] = { 0, };
341 unsigned long nonidle_total = 0;
342 u32 changed_states = 0;
343 int cpu;
344 int s;
345
346 /*
347 * Collect the per-cpu time buckets and average them into a
348 * single time sample that is normalized to wallclock time.
349 *
350 * For averaging, each CPU is weighted by its non-idle time in
351 * the sampling period. This eliminates artifacts from uneven
352 * loading, or even entirely idle CPUs.
353 */
354 for_each_possible_cpu(cpu) {
355 u32 times[NR_PSI_STATES];
356 u32 nonidle;
357 u32 cpu_changed_states;
358
359 get_recent_times(group, cpu, aggregator, times,
360 &cpu_changed_states);
361 changed_states |= cpu_changed_states;
362
363 nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
364 nonidle_total += nonidle;
365
366 for (s = 0; s < PSI_NONIDLE; s++)
367 deltas[s] += (u64)times[s] * nonidle;
368 }
369
370 /*
371 * Integrate the sample into the running statistics that are
372 * reported to userspace: the cumulative stall times and the
373 * decaying averages.
374 *
375 * Pressure percentages are sampled at PSI_FREQ. We might be
376 * called more often when the user polls more frequently than
377 * that; we might be called less often when there is no task
378 * activity, thus no data, and clock ticks are sporadic. The
379 * below handles both.
380 */
381
382 /* total= */
383 for (s = 0; s < NR_PSI_STATES - 1; s++)
384 group->total[aggregator][s] +=
385 div_u64(deltas[s], max(nonidle_total, 1UL));
386
387 if (pchanged_states)
388 *pchanged_states = changed_states;
389}
390
391/* Trigger tracking window manipulations */
392static void window_reset(struct psi_window *win, u64 now, u64 value,
393 u64 prev_growth)
394{
395 win->start_time = now;
396 win->start_value = value;
397 win->prev_growth = prev_growth;
398}
399
400/*
401 * PSI growth tracking window update and growth calculation routine.
402 *
403 * This approximates a sliding tracking window by interpolating
404 * partially elapsed windows using historical growth data from the
405 * previous intervals. This minimizes memory requirements (by not storing
406 * all the intermediate values in the previous window) and simplifies
407 * the calculations. It works well because PSI signal changes only in
408 * positive direction and over relatively small window sizes the growth
409 * is close to linear.
410 */
411static u64 window_update(struct psi_window *win, u64 now, u64 value)
412{
413 u64 elapsed;
414 u64 growth;
415
416 elapsed = now - win->start_time;
417 growth = value - win->start_value;
418 /*
419 * After each tracking window passes win->start_value and
420 * win->start_time get reset and win->prev_growth stores
421 * the average per-window growth of the previous window.
422 * win->prev_growth is then used to interpolate additional
423 * growth from the previous window assuming it was linear.
424 */
425 if (elapsed > win->size)
426 window_reset(win, now, value, growth);
427 else {
428 u32 remaining;
429
430 remaining = win->size - elapsed;
431 growth += div64_u64(win->prev_growth * remaining, win->size);
432 }
433
434 return growth;
435}
436
437static void update_triggers(struct psi_group *group, u64 now,
438 enum psi_aggregators aggregator)
439{
440 struct psi_trigger *t;
441 u64 *total = group->total[aggregator];
442 struct list_head *triggers;
443 u64 *aggregator_total;
444
445 if (aggregator == PSI_AVGS) {
446 triggers = &group->avg_triggers;
447 aggregator_total = group->avg_total;
448 } else {
449 triggers = &group->rtpoll_triggers;
450 aggregator_total = group->rtpoll_total;
451 }
452
453 /*
454 * On subsequent updates, calculate growth deltas and let
455 * watchers know when their specified thresholds are exceeded.
456 */
457 list_for_each_entry(t, triggers, node) {
458 u64 growth;
459 bool new_stall;
460
461 new_stall = aggregator_total[t->state] != total[t->state];
462
463 /* Check for stall activity or a previous threshold breach */
464 if (!new_stall && !t->pending_event)
465 continue;
466 /*
467 * Check for new stall activity, as well as deferred
468 * events that occurred in the last window after the
469 * trigger had already fired (we want to ratelimit
470 * events without dropping any).
471 */
472 if (new_stall) {
473 /* Calculate growth since last update */
474 growth = window_update(&t->win, now, total[t->state]);
475 if (!t->pending_event) {
476 if (growth < t->threshold)
477 continue;
478
479 t->pending_event = true;
480 }
481 }
482 /* Limit event signaling to once per window */
483 if (now < t->last_event_time + t->win.size)
484 continue;
485
486 /* Generate an event */
487 if (cmpxchg(&t->event, 0, 1) == 0) {
488 if (t->of)
489 kernfs_notify(t->of->kn);
490 else
491 wake_up_interruptible(&t->event_wait);
492 }
493 t->last_event_time = now;
494 /* Reset threshold breach flag once event got generated */
495 t->pending_event = false;
496 }
497}
498
499static u64 update_averages(struct psi_group *group, u64 now)
500{
501 unsigned long missed_periods = 0;
502 u64 expires, period;
503 u64 avg_next_update;
504 int s;
505
506 /* avgX= */
507 expires = group->avg_next_update;
508 if (now - expires >= psi_period)
509 missed_periods = div_u64(now - expires, psi_period);
510
511 /*
512 * The periodic clock tick can get delayed for various
513 * reasons, especially on loaded systems. To avoid clock
514 * drift, we schedule the clock in fixed psi_period intervals.
515 * But the deltas we sample out of the per-cpu buckets above
516 * are based on the actual time elapsing between clock ticks.
517 */
518 avg_next_update = expires + ((1 + missed_periods) * psi_period);
519 period = now - (group->avg_last_update + (missed_periods * psi_period));
520 group->avg_last_update = now;
521
522 for (s = 0; s < NR_PSI_STATES - 1; s++) {
523 u32 sample;
524
525 sample = group->total[PSI_AVGS][s] - group->avg_total[s];
526 /*
527 * Due to the lockless sampling of the time buckets,
528 * recorded time deltas can slip into the next period,
529 * which under full pressure can result in samples in
530 * excess of the period length.
531 *
532 * We don't want to report non-sensical pressures in
533 * excess of 100%, nor do we want to drop such events
534 * on the floor. Instead we punt any overage into the
535 * future until pressure subsides. By doing this we
536 * don't underreport the occurring pressure curve, we
537 * just report it delayed by one period length.
538 *
539 * The error isn't cumulative. As soon as another
540 * delta slips from a period P to P+1, by definition
541 * it frees up its time T in P.
542 */
543 if (sample > period)
544 sample = period;
545 group->avg_total[s] += sample;
546 calc_avgs(group->avg[s], missed_periods, sample, period);
547 }
548
549 return avg_next_update;
550}
551
552static void psi_avgs_work(struct work_struct *work)
553{
554 struct delayed_work *dwork;
555 struct psi_group *group;
556 u32 changed_states;
557 u64 now;
558
559 dwork = to_delayed_work(work);
560 group = container_of(dwork, struct psi_group, avgs_work);
561
562 mutex_lock(&group->avgs_lock);
563
564 now = sched_clock();
565
566 collect_percpu_times(group, PSI_AVGS, &changed_states);
567 /*
568 * If there is task activity, periodically fold the per-cpu
569 * times and feed samples into the running averages. If things
570 * are idle and there is no data to process, stop the clock.
571 * Once restarted, we'll catch up the running averages in one
572 * go - see calc_avgs() and missed_periods.
573 */
574 if (now >= group->avg_next_update) {
575 update_triggers(group, now, PSI_AVGS);
576 group->avg_next_update = update_averages(group, now);
577 }
578
579 if (changed_states & PSI_STATE_RESCHEDULE) {
580 schedule_delayed_work(dwork, nsecs_to_jiffies(
581 group->avg_next_update - now) + 1);
582 }
583
584 mutex_unlock(&group->avgs_lock);
585}
586
587static void init_rtpoll_triggers(struct psi_group *group, u64 now)
588{
589 struct psi_trigger *t;
590
591 list_for_each_entry(t, &group->rtpoll_triggers, node)
592 window_reset(&t->win, now,
593 group->total[PSI_POLL][t->state], 0);
594 memcpy(group->rtpoll_total, group->total[PSI_POLL],
595 sizeof(group->rtpoll_total));
596 group->rtpoll_next_update = now + group->rtpoll_min_period;
597}
598
599/* Schedule rtpolling if it's not already scheduled or forced. */
600static void psi_schedule_rtpoll_work(struct psi_group *group, unsigned long delay,
601 bool force)
602{
603 struct task_struct *task;
604
605 /*
606 * atomic_xchg should be called even when !force to provide a
607 * full memory barrier (see the comment inside psi_rtpoll_work).
608 */
609 if (atomic_xchg(&group->rtpoll_scheduled, 1) && !force)
610 return;
611
612 rcu_read_lock();
613
614 task = rcu_dereference(group->rtpoll_task);
615 /*
616 * kworker might be NULL in case psi_trigger_destroy races with
617 * psi_task_change (hotpath) which can't use locks
618 */
619 if (likely(task))
620 mod_timer(&group->rtpoll_timer, jiffies + delay);
621 else
622 atomic_set(&group->rtpoll_scheduled, 0);
623
624 rcu_read_unlock();
625}
626
627static void psi_rtpoll_work(struct psi_group *group)
628{
629 bool force_reschedule = false;
630 u32 changed_states;
631 u64 now;
632
633 mutex_lock(&group->rtpoll_trigger_lock);
634
635 now = sched_clock();
636
637 if (now > group->rtpoll_until) {
638 /*
639 * We are either about to start or might stop rtpolling if no
640 * state change was recorded. Resetting rtpoll_scheduled leaves
641 * a small window for psi_group_change to sneak in and schedule
642 * an immediate rtpoll_work before we get to rescheduling. One
643 * potential extra wakeup at the end of the rtpolling window
644 * should be negligible and rtpoll_next_update still keeps
645 * updates correctly on schedule.
646 */
647 atomic_set(&group->rtpoll_scheduled, 0);
648 /*
649 * A task change can race with the rtpoll worker that is supposed to
650 * report on it. To avoid missing events, ensure ordering between
651 * rtpoll_scheduled and the task state accesses, such that if the
652 * rtpoll worker misses the state update, the task change is
653 * guaranteed to reschedule the rtpoll worker:
654 *
655 * rtpoll worker:
656 * atomic_set(rtpoll_scheduled, 0)
657 * smp_mb()
658 * LOAD states
659 *
660 * task change:
661 * STORE states
662 * if atomic_xchg(rtpoll_scheduled, 1) == 0:
663 * schedule rtpoll worker
664 *
665 * The atomic_xchg() implies a full barrier.
666 */
667 smp_mb();
668 } else {
669 /* The rtpolling window is not over, keep rescheduling */
670 force_reschedule = true;
671 }
672
673
674 collect_percpu_times(group, PSI_POLL, &changed_states);
675
676 if (changed_states & group->rtpoll_states) {
677 /* Initialize trigger windows when entering rtpolling mode */
678 if (now > group->rtpoll_until)
679 init_rtpoll_triggers(group, now);
680
681 /*
682 * Keep the monitor active for at least the duration of the
683 * minimum tracking window as long as monitor states are
684 * changing.
685 */
686 group->rtpoll_until = now +
687 group->rtpoll_min_period * UPDATES_PER_WINDOW;
688 }
689
690 if (now > group->rtpoll_until) {
691 group->rtpoll_next_update = ULLONG_MAX;
692 goto out;
693 }
694
695 if (now >= group->rtpoll_next_update) {
696 if (changed_states & group->rtpoll_states) {
697 update_triggers(group, now, PSI_POLL);
698 memcpy(group->rtpoll_total, group->total[PSI_POLL],
699 sizeof(group->rtpoll_total));
700 }
701 group->rtpoll_next_update = now + group->rtpoll_min_period;
702 }
703
704 psi_schedule_rtpoll_work(group,
705 nsecs_to_jiffies(group->rtpoll_next_update - now) + 1,
706 force_reschedule);
707
708out:
709 mutex_unlock(&group->rtpoll_trigger_lock);
710}
711
712static int psi_rtpoll_worker(void *data)
713{
714 struct psi_group *group = (struct psi_group *)data;
715
716 sched_set_fifo_low(current);
717
718 while (true) {
719 wait_event_interruptible(group->rtpoll_wait,
720 atomic_cmpxchg(&group->rtpoll_wakeup, 1, 0) ||
721 kthread_should_stop());
722 if (kthread_should_stop())
723 break;
724
725 psi_rtpoll_work(group);
726 }
727 return 0;
728}
729
730static void poll_timer_fn(struct timer_list *t)
731{
732 struct psi_group *group = from_timer(group, t, rtpoll_timer);
733
734 atomic_set(&group->rtpoll_wakeup, 1);
735 wake_up_interruptible(&group->rtpoll_wait);
736}
737
738static void record_times(struct psi_group_cpu *groupc, u64 now)
739{
740 u32 delta;
741
742 delta = now - groupc->state_start;
743 groupc->state_start = now;
744
745 if (groupc->state_mask & (1 << PSI_IO_SOME)) {
746 groupc->times[PSI_IO_SOME] += delta;
747 if (groupc->state_mask & (1 << PSI_IO_FULL))
748 groupc->times[PSI_IO_FULL] += delta;
749 }
750
751 if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
752 groupc->times[PSI_MEM_SOME] += delta;
753 if (groupc->state_mask & (1 << PSI_MEM_FULL))
754 groupc->times[PSI_MEM_FULL] += delta;
755 }
756
757 if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
758 groupc->times[PSI_CPU_SOME] += delta;
759 if (groupc->state_mask & (1 << PSI_CPU_FULL))
760 groupc->times[PSI_CPU_FULL] += delta;
761 }
762
763 if (groupc->state_mask & (1 << PSI_NONIDLE))
764 groupc->times[PSI_NONIDLE] += delta;
765}
766
767static void psi_group_change(struct psi_group *group, int cpu,
768 unsigned int clear, unsigned int set, u64 now,
769 bool wake_clock)
770{
771 struct psi_group_cpu *groupc;
772 unsigned int t, m;
773 enum psi_states s;
774 u32 state_mask;
775
776 groupc = per_cpu_ptr(group->pcpu, cpu);
777
778 /*
779 * First we update the task counts according to the state
780 * change requested through the @clear and @set bits.
781 *
782 * Then if the cgroup PSI stats accounting enabled, we
783 * assess the aggregate resource states this CPU's tasks
784 * have been in since the last change, and account any
785 * SOME and FULL time these may have resulted in.
786 */
787 write_seqcount_begin(&groupc->seq);
788
789 /*
790 * Start with TSK_ONCPU, which doesn't have a corresponding
791 * task count - it's just a boolean flag directly encoded in
792 * the state mask. Clear, set, or carry the current state if
793 * no changes are requested.
794 */
795 if (unlikely(clear & TSK_ONCPU)) {
796 state_mask = 0;
797 clear &= ~TSK_ONCPU;
798 } else if (unlikely(set & TSK_ONCPU)) {
799 state_mask = PSI_ONCPU;
800 set &= ~TSK_ONCPU;
801 } else {
802 state_mask = groupc->state_mask & PSI_ONCPU;
803 }
804
805 /*
806 * The rest of the state mask is calculated based on the task
807 * counts. Update those first, then construct the mask.
808 */
809 for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
810 if (!(m & (1 << t)))
811 continue;
812 if (groupc->tasks[t]) {
813 groupc->tasks[t]--;
814 } else if (!psi_bug) {
815 printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n",
816 cpu, t, groupc->tasks[0],
817 groupc->tasks[1], groupc->tasks[2],
818 groupc->tasks[3], clear, set);
819 psi_bug = 1;
820 }
821 }
822
823 for (t = 0; set; set &= ~(1 << t), t++)
824 if (set & (1 << t))
825 groupc->tasks[t]++;
826
827 if (!group->enabled) {
828 /*
829 * On the first group change after disabling PSI, conclude
830 * the current state and flush its time. This is unlikely
831 * to matter to the user, but aggregation (get_recent_times)
832 * may have already incorporated the live state into times_prev;
833 * avoid a delta sample underflow when PSI is later re-enabled.
834 */
835 if (unlikely(groupc->state_mask & (1 << PSI_NONIDLE)))
836 record_times(groupc, now);
837
838 groupc->state_mask = state_mask;
839
840 write_seqcount_end(&groupc->seq);
841 return;
842 }
843
844 for (s = 0; s < NR_PSI_STATES; s++) {
845 if (test_state(groupc->tasks, s, state_mask & PSI_ONCPU))
846 state_mask |= (1 << s);
847 }
848
849 /*
850 * Since we care about lost potential, a memstall is FULL
851 * when there are no other working tasks, but also when
852 * the CPU is actively reclaiming and nothing productive
853 * could run even if it were runnable. So when the current
854 * task in a cgroup is in_memstall, the corresponding groupc
855 * on that cpu is in PSI_MEM_FULL state.
856 */
857 if (unlikely((state_mask & PSI_ONCPU) && cpu_curr(cpu)->in_memstall))
858 state_mask |= (1 << PSI_MEM_FULL);
859
860 record_times(groupc, now);
861
862 groupc->state_mask = state_mask;
863
864 write_seqcount_end(&groupc->seq);
865
866 if (state_mask & group->rtpoll_states)
867 psi_schedule_rtpoll_work(group, 1, false);
868
869 if (wake_clock && !delayed_work_pending(&group->avgs_work))
870 schedule_delayed_work(&group->avgs_work, PSI_FREQ);
871}
872
873static inline struct psi_group *task_psi_group(struct task_struct *task)
874{
875#ifdef CONFIG_CGROUPS
876 if (static_branch_likely(&psi_cgroups_enabled))
877 return cgroup_psi(task_dfl_cgroup(task));
878#endif
879 return &psi_system;
880}
881
882static void psi_flags_change(struct task_struct *task, int clear, int set)
883{
884 if (((task->psi_flags & set) ||
885 (task->psi_flags & clear) != clear) &&
886 !psi_bug) {
887 printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
888 task->pid, task->comm, task_cpu(task),
889 task->psi_flags, clear, set);
890 psi_bug = 1;
891 }
892
893 task->psi_flags &= ~clear;
894 task->psi_flags |= set;
895}
896
897void psi_task_change(struct task_struct *task, int clear, int set)
898{
899 int cpu = task_cpu(task);
900 struct psi_group *group;
901 u64 now;
902
903 if (!task->pid)
904 return;
905
906 psi_flags_change(task, clear, set);
907
908 now = cpu_clock(cpu);
909
910 group = task_psi_group(task);
911 do {
912 psi_group_change(group, cpu, clear, set, now, true);
913 } while ((group = group->parent));
914}
915
916void psi_task_switch(struct task_struct *prev, struct task_struct *next,
917 bool sleep)
918{
919 struct psi_group *group, *common = NULL;
920 int cpu = task_cpu(prev);
921 u64 now = cpu_clock(cpu);
922
923 if (next->pid) {
924 psi_flags_change(next, 0, TSK_ONCPU);
925 /*
926 * Set TSK_ONCPU on @next's cgroups. If @next shares any
927 * ancestors with @prev, those will already have @prev's
928 * TSK_ONCPU bit set, and we can stop the iteration there.
929 */
930 group = task_psi_group(next);
931 do {
932 if (per_cpu_ptr(group->pcpu, cpu)->state_mask &
933 PSI_ONCPU) {
934 common = group;
935 break;
936 }
937
938 psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
939 } while ((group = group->parent));
940 }
941
942 if (prev->pid) {
943 int clear = TSK_ONCPU, set = 0;
944 bool wake_clock = true;
945
946 /*
947 * When we're going to sleep, psi_dequeue() lets us
948 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
949 * TSK_IOWAIT here, where we can combine it with
950 * TSK_ONCPU and save walking common ancestors twice.
951 */
952 if (sleep) {
953 clear |= TSK_RUNNING;
954 if (prev->in_memstall)
955 clear |= TSK_MEMSTALL_RUNNING;
956 if (prev->in_iowait)
957 set |= TSK_IOWAIT;
958
959 /*
960 * Periodic aggregation shuts off if there is a period of no
961 * task changes, so we wake it back up if necessary. However,
962 * don't do this if the task change is the aggregation worker
963 * itself going to sleep, or we'll ping-pong forever.
964 */
965 if (unlikely((prev->flags & PF_WQ_WORKER) &&
966 wq_worker_last_func(prev) == psi_avgs_work))
967 wake_clock = false;
968 }
969
970 psi_flags_change(prev, clear, set);
971
972 group = task_psi_group(prev);
973 do {
974 if (group == common)
975 break;
976 psi_group_change(group, cpu, clear, set, now, wake_clock);
977 } while ((group = group->parent));
978
979 /*
980 * TSK_ONCPU is handled up to the common ancestor. If there are
981 * any other differences between the two tasks (e.g. prev goes
982 * to sleep, or only one task is memstall), finish propagating
983 * those differences all the way up to the root.
984 */
985 if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) {
986 clear &= ~TSK_ONCPU;
987 for (; group; group = group->parent)
988 psi_group_change(group, cpu, clear, set, now, wake_clock);
989 }
990 }
991}
992
993#ifdef CONFIG_IRQ_TIME_ACCOUNTING
994void psi_account_irqtime(struct task_struct *task, u32 delta)
995{
996 int cpu = task_cpu(task);
997 struct psi_group *group;
998 struct psi_group_cpu *groupc;
999 u64 now;
1000
1001 if (static_branch_likely(&psi_disabled))
1002 return;
1003
1004 if (!task->pid)
1005 return;
1006
1007 now = cpu_clock(cpu);
1008
1009 group = task_psi_group(task);
1010 do {
1011 if (!group->enabled)
1012 continue;
1013
1014 groupc = per_cpu_ptr(group->pcpu, cpu);
1015
1016 write_seqcount_begin(&groupc->seq);
1017
1018 record_times(groupc, now);
1019 groupc->times[PSI_IRQ_FULL] += delta;
1020
1021 write_seqcount_end(&groupc->seq);
1022
1023 if (group->rtpoll_states & (1 << PSI_IRQ_FULL))
1024 psi_schedule_rtpoll_work(group, 1, false);
1025 } while ((group = group->parent));
1026}
1027#endif
1028
1029/**
1030 * psi_memstall_enter - mark the beginning of a memory stall section
1031 * @flags: flags to handle nested sections
1032 *
1033 * Marks the calling task as being stalled due to a lack of memory,
1034 * such as waiting for a refault or performing reclaim.
1035 */
1036void psi_memstall_enter(unsigned long *flags)
1037{
1038 struct rq_flags rf;
1039 struct rq *rq;
1040
1041 if (static_branch_likely(&psi_disabled))
1042 return;
1043
1044 *flags = current->in_memstall;
1045 if (*flags)
1046 return;
1047 /*
1048 * in_memstall setting & accounting needs to be atomic wrt
1049 * changes to the task's scheduling state, otherwise we can
1050 * race with CPU migration.
1051 */
1052 rq = this_rq_lock_irq(&rf);
1053
1054 current->in_memstall = 1;
1055 psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
1056
1057 rq_unlock_irq(rq, &rf);
1058}
1059EXPORT_SYMBOL_GPL(psi_memstall_enter);
1060
1061/**
1062 * psi_memstall_leave - mark the end of an memory stall section
1063 * @flags: flags to handle nested memdelay sections
1064 *
1065 * Marks the calling task as no longer stalled due to lack of memory.
1066 */
1067void psi_memstall_leave(unsigned long *flags)
1068{
1069 struct rq_flags rf;
1070 struct rq *rq;
1071
1072 if (static_branch_likely(&psi_disabled))
1073 return;
1074
1075 if (*flags)
1076 return;
1077 /*
1078 * in_memstall clearing & accounting needs to be atomic wrt
1079 * changes to the task's scheduling state, otherwise we could
1080 * race with CPU migration.
1081 */
1082 rq = this_rq_lock_irq(&rf);
1083
1084 current->in_memstall = 0;
1085 psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0);
1086
1087 rq_unlock_irq(rq, &rf);
1088}
1089EXPORT_SYMBOL_GPL(psi_memstall_leave);
1090
1091#ifdef CONFIG_CGROUPS
1092int psi_cgroup_alloc(struct cgroup *cgroup)
1093{
1094 if (!static_branch_likely(&psi_cgroups_enabled))
1095 return 0;
1096
1097 cgroup->psi = kzalloc(sizeof(struct psi_group), GFP_KERNEL);
1098 if (!cgroup->psi)
1099 return -ENOMEM;
1100
1101 cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
1102 if (!cgroup->psi->pcpu) {
1103 kfree(cgroup->psi);
1104 return -ENOMEM;
1105 }
1106 group_init(cgroup->psi);
1107 cgroup->psi->parent = cgroup_psi(cgroup_parent(cgroup));
1108 return 0;
1109}
1110
1111void psi_cgroup_free(struct cgroup *cgroup)
1112{
1113 if (!static_branch_likely(&psi_cgroups_enabled))
1114 return;
1115
1116 cancel_delayed_work_sync(&cgroup->psi->avgs_work);
1117 free_percpu(cgroup->psi->pcpu);
1118 /* All triggers must be removed by now */
1119 WARN_ONCE(cgroup->psi->rtpoll_states, "psi: trigger leak\n");
1120 kfree(cgroup->psi);
1121}
1122
1123/**
1124 * cgroup_move_task - move task to a different cgroup
1125 * @task: the task
1126 * @to: the target css_set
1127 *
1128 * Move task to a new cgroup and safely migrate its associated stall
1129 * state between the different groups.
1130 *
1131 * This function acquires the task's rq lock to lock out concurrent
1132 * changes to the task's scheduling state and - in case the task is
1133 * running - concurrent changes to its stall state.
1134 */
1135void cgroup_move_task(struct task_struct *task, struct css_set *to)
1136{
1137 unsigned int task_flags;
1138 struct rq_flags rf;
1139 struct rq *rq;
1140
1141 if (!static_branch_likely(&psi_cgroups_enabled)) {
1142 /*
1143 * Lame to do this here, but the scheduler cannot be locked
1144 * from the outside, so we move cgroups from inside sched/.
1145 */
1146 rcu_assign_pointer(task->cgroups, to);
1147 return;
1148 }
1149
1150 rq = task_rq_lock(task, &rf);
1151
1152 /*
1153 * We may race with schedule() dropping the rq lock between
1154 * deactivating prev and switching to next. Because the psi
1155 * updates from the deactivation are deferred to the switch
1156 * callback to save cgroup tree updates, the task's scheduling
1157 * state here is not coherent with its psi state:
1158 *
1159 * schedule() cgroup_move_task()
1160 * rq_lock()
1161 * deactivate_task()
1162 * p->on_rq = 0
1163 * psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1164 * pick_next_task()
1165 * rq_unlock()
1166 * rq_lock()
1167 * psi_task_change() // old cgroup
1168 * task->cgroups = to
1169 * psi_task_change() // new cgroup
1170 * rq_unlock()
1171 * rq_lock()
1172 * psi_sched_switch() // does deferred updates in new cgroup
1173 *
1174 * Don't rely on the scheduling state. Use psi_flags instead.
1175 */
1176 task_flags = task->psi_flags;
1177
1178 if (task_flags)
1179 psi_task_change(task, task_flags, 0);
1180
1181 /* See comment above */
1182 rcu_assign_pointer(task->cgroups, to);
1183
1184 if (task_flags)
1185 psi_task_change(task, 0, task_flags);
1186
1187 task_rq_unlock(rq, task, &rf);
1188}
1189
1190void psi_cgroup_restart(struct psi_group *group)
1191{
1192 int cpu;
1193
1194 /*
1195 * After we disable psi_group->enabled, we don't actually
1196 * stop percpu tasks accounting in each psi_group_cpu,
1197 * instead only stop test_state() loop, record_times()
1198 * and averaging worker, see psi_group_change() for details.
1199 *
1200 * When disable cgroup PSI, this function has nothing to sync
1201 * since cgroup pressure files are hidden and percpu psi_group_cpu
1202 * would see !psi_group->enabled and only do task accounting.
1203 *
1204 * When re-enable cgroup PSI, this function use psi_group_change()
1205 * to get correct state mask from test_state() loop on tasks[],
1206 * and restart groupc->state_start from now, use .clear = .set = 0
1207 * here since no task status really changed.
1208 */
1209 if (!group->enabled)
1210 return;
1211
1212 for_each_possible_cpu(cpu) {
1213 struct rq *rq = cpu_rq(cpu);
1214 struct rq_flags rf;
1215 u64 now;
1216
1217 rq_lock_irq(rq, &rf);
1218 now = cpu_clock(cpu);
1219 psi_group_change(group, cpu, 0, 0, now, true);
1220 rq_unlock_irq(rq, &rf);
1221 }
1222}
1223#endif /* CONFIG_CGROUPS */
1224
1225int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1226{
1227 bool only_full = false;
1228 int full;
1229 u64 now;
1230
1231 if (static_branch_likely(&psi_disabled))
1232 return -EOPNOTSUPP;
1233
1234 /* Update averages before reporting them */
1235 mutex_lock(&group->avgs_lock);
1236 now = sched_clock();
1237 collect_percpu_times(group, PSI_AVGS, NULL);
1238 if (now >= group->avg_next_update)
1239 group->avg_next_update = update_averages(group, now);
1240 mutex_unlock(&group->avgs_lock);
1241
1242#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1243 only_full = res == PSI_IRQ;
1244#endif
1245
1246 for (full = 0; full < 2 - only_full; full++) {
1247 unsigned long avg[3] = { 0, };
1248 u64 total = 0;
1249 int w;
1250
1251 /* CPU FULL is undefined at the system level */
1252 if (!(group == &psi_system && res == PSI_CPU && full)) {
1253 for (w = 0; w < 3; w++)
1254 avg[w] = group->avg[res * 2 + full][w];
1255 total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1256 NSEC_PER_USEC);
1257 }
1258
1259 seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1260 full || only_full ? "full" : "some",
1261 LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1262 LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1263 LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1264 total);
1265 }
1266
1267 return 0;
1268}
1269
1270struct psi_trigger *psi_trigger_create(struct psi_group *group, char *buf,
1271 enum psi_res res, struct file *file,
1272 struct kernfs_open_file *of)
1273{
1274 struct psi_trigger *t;
1275 enum psi_states state;
1276 u32 threshold_us;
1277 bool privileged;
1278 u32 window_us;
1279
1280 if (static_branch_likely(&psi_disabled))
1281 return ERR_PTR(-EOPNOTSUPP);
1282
1283 /*
1284 * Checking the privilege here on file->f_cred implies that a privileged user
1285 * could open the file and delegate the write to an unprivileged one.
1286 */
1287 privileged = cap_raised(file->f_cred->cap_effective, CAP_SYS_RESOURCE);
1288
1289 if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1290 state = PSI_IO_SOME + res * 2;
1291 else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1292 state = PSI_IO_FULL + res * 2;
1293 else
1294 return ERR_PTR(-EINVAL);
1295
1296#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1297 if (res == PSI_IRQ && --state != PSI_IRQ_FULL)
1298 return ERR_PTR(-EINVAL);
1299#endif
1300
1301 if (state >= PSI_NONIDLE)
1302 return ERR_PTR(-EINVAL);
1303
1304 if (window_us == 0 || window_us > WINDOW_MAX_US)
1305 return ERR_PTR(-EINVAL);
1306
1307 /*
1308 * Unprivileged users can only use 2s windows so that averages aggregation
1309 * work is used, and no RT threads need to be spawned.
1310 */
1311 if (!privileged && window_us % 2000000)
1312 return ERR_PTR(-EINVAL);
1313
1314 /* Check threshold */
1315 if (threshold_us == 0 || threshold_us > window_us)
1316 return ERR_PTR(-EINVAL);
1317
1318 t = kmalloc(sizeof(*t), GFP_KERNEL);
1319 if (!t)
1320 return ERR_PTR(-ENOMEM);
1321
1322 t->group = group;
1323 t->state = state;
1324 t->threshold = threshold_us * NSEC_PER_USEC;
1325 t->win.size = window_us * NSEC_PER_USEC;
1326 window_reset(&t->win, sched_clock(),
1327 group->total[PSI_POLL][t->state], 0);
1328
1329 t->event = 0;
1330 t->last_event_time = 0;
1331 t->of = of;
1332 if (!of)
1333 init_waitqueue_head(&t->event_wait);
1334 t->pending_event = false;
1335 t->aggregator = privileged ? PSI_POLL : PSI_AVGS;
1336
1337 if (privileged) {
1338 mutex_lock(&group->rtpoll_trigger_lock);
1339
1340 if (!rcu_access_pointer(group->rtpoll_task)) {
1341 struct task_struct *task;
1342
1343 task = kthread_create(psi_rtpoll_worker, group, "psimon");
1344 if (IS_ERR(task)) {
1345 kfree(t);
1346 mutex_unlock(&group->rtpoll_trigger_lock);
1347 return ERR_CAST(task);
1348 }
1349 atomic_set(&group->rtpoll_wakeup, 0);
1350 wake_up_process(task);
1351 rcu_assign_pointer(group->rtpoll_task, task);
1352 }
1353
1354 list_add(&t->node, &group->rtpoll_triggers);
1355 group->rtpoll_min_period = min(group->rtpoll_min_period,
1356 div_u64(t->win.size, UPDATES_PER_WINDOW));
1357 group->rtpoll_nr_triggers[t->state]++;
1358 group->rtpoll_states |= (1 << t->state);
1359
1360 mutex_unlock(&group->rtpoll_trigger_lock);
1361 } else {
1362 mutex_lock(&group->avgs_lock);
1363
1364 list_add(&t->node, &group->avg_triggers);
1365 group->avg_nr_triggers[t->state]++;
1366
1367 mutex_unlock(&group->avgs_lock);
1368 }
1369 return t;
1370}
1371
1372void psi_trigger_destroy(struct psi_trigger *t)
1373{
1374 struct psi_group *group;
1375 struct task_struct *task_to_destroy = NULL;
1376
1377 /*
1378 * We do not check psi_disabled since it might have been disabled after
1379 * the trigger got created.
1380 */
1381 if (!t)
1382 return;
1383
1384 group = t->group;
1385 /*
1386 * Wakeup waiters to stop polling and clear the queue to prevent it from
1387 * being accessed later. Can happen if cgroup is deleted from under a
1388 * polling process.
1389 */
1390 if (t->of)
1391 kernfs_notify(t->of->kn);
1392 else
1393 wake_up_interruptible(&t->event_wait);
1394
1395 if (t->aggregator == PSI_AVGS) {
1396 mutex_lock(&group->avgs_lock);
1397 if (!list_empty(&t->node)) {
1398 list_del(&t->node);
1399 group->avg_nr_triggers[t->state]--;
1400 }
1401 mutex_unlock(&group->avgs_lock);
1402 } else {
1403 mutex_lock(&group->rtpoll_trigger_lock);
1404 if (!list_empty(&t->node)) {
1405 struct psi_trigger *tmp;
1406 u64 period = ULLONG_MAX;
1407
1408 list_del(&t->node);
1409 group->rtpoll_nr_triggers[t->state]--;
1410 if (!group->rtpoll_nr_triggers[t->state])
1411 group->rtpoll_states &= ~(1 << t->state);
1412 /*
1413 * Reset min update period for the remaining triggers
1414 * iff the destroying trigger had the min window size.
1415 */
1416 if (group->rtpoll_min_period == div_u64(t->win.size, UPDATES_PER_WINDOW)) {
1417 list_for_each_entry(tmp, &group->rtpoll_triggers, node)
1418 period = min(period, div_u64(tmp->win.size,
1419 UPDATES_PER_WINDOW));
1420 group->rtpoll_min_period = period;
1421 }
1422 /* Destroy rtpoll_task when the last trigger is destroyed */
1423 if (group->rtpoll_states == 0) {
1424 group->rtpoll_until = 0;
1425 task_to_destroy = rcu_dereference_protected(
1426 group->rtpoll_task,
1427 lockdep_is_held(&group->rtpoll_trigger_lock));
1428 rcu_assign_pointer(group->rtpoll_task, NULL);
1429 del_timer(&group->rtpoll_timer);
1430 }
1431 }
1432 mutex_unlock(&group->rtpoll_trigger_lock);
1433 }
1434
1435 /*
1436 * Wait for psi_schedule_rtpoll_work RCU to complete its read-side
1437 * critical section before destroying the trigger and optionally the
1438 * rtpoll_task.
1439 */
1440 synchronize_rcu();
1441 /*
1442 * Stop kthread 'psimon' after releasing rtpoll_trigger_lock to prevent
1443 * a deadlock while waiting for psi_rtpoll_work to acquire
1444 * rtpoll_trigger_lock
1445 */
1446 if (task_to_destroy) {
1447 /*
1448 * After the RCU grace period has expired, the worker
1449 * can no longer be found through group->rtpoll_task.
1450 */
1451 kthread_stop(task_to_destroy);
1452 atomic_set(&group->rtpoll_scheduled, 0);
1453 }
1454 kfree(t);
1455}
1456
1457__poll_t psi_trigger_poll(void **trigger_ptr,
1458 struct file *file, poll_table *wait)
1459{
1460 __poll_t ret = DEFAULT_POLLMASK;
1461 struct psi_trigger *t;
1462
1463 if (static_branch_likely(&psi_disabled))
1464 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1465
1466 t = smp_load_acquire(trigger_ptr);
1467 if (!t)
1468 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1469
1470 if (t->of)
1471 kernfs_generic_poll(t->of, wait);
1472 else
1473 poll_wait(file, &t->event_wait, wait);
1474
1475 if (cmpxchg(&t->event, 1, 0) == 1)
1476 ret |= EPOLLPRI;
1477
1478 return ret;
1479}
1480
1481#ifdef CONFIG_PROC_FS
1482static int psi_io_show(struct seq_file *m, void *v)
1483{
1484 return psi_show(m, &psi_system, PSI_IO);
1485}
1486
1487static int psi_memory_show(struct seq_file *m, void *v)
1488{
1489 return psi_show(m, &psi_system, PSI_MEM);
1490}
1491
1492static int psi_cpu_show(struct seq_file *m, void *v)
1493{
1494 return psi_show(m, &psi_system, PSI_CPU);
1495}
1496
1497static int psi_io_open(struct inode *inode, struct file *file)
1498{
1499 return single_open(file, psi_io_show, NULL);
1500}
1501
1502static int psi_memory_open(struct inode *inode, struct file *file)
1503{
1504 return single_open(file, psi_memory_show, NULL);
1505}
1506
1507static int psi_cpu_open(struct inode *inode, struct file *file)
1508{
1509 return single_open(file, psi_cpu_show, NULL);
1510}
1511
1512static ssize_t psi_write(struct file *file, const char __user *user_buf,
1513 size_t nbytes, enum psi_res res)
1514{
1515 char buf[32];
1516 size_t buf_size;
1517 struct seq_file *seq;
1518 struct psi_trigger *new;
1519
1520 if (static_branch_likely(&psi_disabled))
1521 return -EOPNOTSUPP;
1522
1523 if (!nbytes)
1524 return -EINVAL;
1525
1526 buf_size = min(nbytes, sizeof(buf));
1527 if (copy_from_user(buf, user_buf, buf_size))
1528 return -EFAULT;
1529
1530 buf[buf_size - 1] = '\0';
1531
1532 seq = file->private_data;
1533
1534 /* Take seq->lock to protect seq->private from concurrent writes */
1535 mutex_lock(&seq->lock);
1536
1537 /* Allow only one trigger per file descriptor */
1538 if (seq->private) {
1539 mutex_unlock(&seq->lock);
1540 return -EBUSY;
1541 }
1542
1543 new = psi_trigger_create(&psi_system, buf, res, file, NULL);
1544 if (IS_ERR(new)) {
1545 mutex_unlock(&seq->lock);
1546 return PTR_ERR(new);
1547 }
1548
1549 smp_store_release(&seq->private, new);
1550 mutex_unlock(&seq->lock);
1551
1552 return nbytes;
1553}
1554
1555static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1556 size_t nbytes, loff_t *ppos)
1557{
1558 return psi_write(file, user_buf, nbytes, PSI_IO);
1559}
1560
1561static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1562 size_t nbytes, loff_t *ppos)
1563{
1564 return psi_write(file, user_buf, nbytes, PSI_MEM);
1565}
1566
1567static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1568 size_t nbytes, loff_t *ppos)
1569{
1570 return psi_write(file, user_buf, nbytes, PSI_CPU);
1571}
1572
1573static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1574{
1575 struct seq_file *seq = file->private_data;
1576
1577 return psi_trigger_poll(&seq->private, file, wait);
1578}
1579
1580static int psi_fop_release(struct inode *inode, struct file *file)
1581{
1582 struct seq_file *seq = file->private_data;
1583
1584 psi_trigger_destroy(seq->private);
1585 return single_release(inode, file);
1586}
1587
1588static const struct proc_ops psi_io_proc_ops = {
1589 .proc_open = psi_io_open,
1590 .proc_read = seq_read,
1591 .proc_lseek = seq_lseek,
1592 .proc_write = psi_io_write,
1593 .proc_poll = psi_fop_poll,
1594 .proc_release = psi_fop_release,
1595};
1596
1597static const struct proc_ops psi_memory_proc_ops = {
1598 .proc_open = psi_memory_open,
1599 .proc_read = seq_read,
1600 .proc_lseek = seq_lseek,
1601 .proc_write = psi_memory_write,
1602 .proc_poll = psi_fop_poll,
1603 .proc_release = psi_fop_release,
1604};
1605
1606static const struct proc_ops psi_cpu_proc_ops = {
1607 .proc_open = psi_cpu_open,
1608 .proc_read = seq_read,
1609 .proc_lseek = seq_lseek,
1610 .proc_write = psi_cpu_write,
1611 .proc_poll = psi_fop_poll,
1612 .proc_release = psi_fop_release,
1613};
1614
1615#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1616static int psi_irq_show(struct seq_file *m, void *v)
1617{
1618 return psi_show(m, &psi_system, PSI_IRQ);
1619}
1620
1621static int psi_irq_open(struct inode *inode, struct file *file)
1622{
1623 return single_open(file, psi_irq_show, NULL);
1624}
1625
1626static ssize_t psi_irq_write(struct file *file, const char __user *user_buf,
1627 size_t nbytes, loff_t *ppos)
1628{
1629 return psi_write(file, user_buf, nbytes, PSI_IRQ);
1630}
1631
1632static const struct proc_ops psi_irq_proc_ops = {
1633 .proc_open = psi_irq_open,
1634 .proc_read = seq_read,
1635 .proc_lseek = seq_lseek,
1636 .proc_write = psi_irq_write,
1637 .proc_poll = psi_fop_poll,
1638 .proc_release = psi_fop_release,
1639};
1640#endif
1641
1642static int __init psi_proc_init(void)
1643{
1644 if (psi_enable) {
1645 proc_mkdir("pressure", NULL);
1646 proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops);
1647 proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops);
1648 proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops);
1649#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1650 proc_create("pressure/irq", 0666, NULL, &psi_irq_proc_ops);
1651#endif
1652 }
1653 return 0;
1654}
1655module_init(psi_proc_init);
1656
1657#endif /* CONFIG_PROC_FS */
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);