projects/linux·blob master · refs

1710 lines · 47.5 KiB

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 <[email protected]>
7 *
8 * Polling support by Suren Baghdasaryan <[email protected]>
9 * Copyright (c) 2018 Google, Inc.
10 *
11 * When CPU, memory and IO are contended, tasks experience delays that
12 * reduce throughput and introduce latencies into the workload. Memory
13 * and IO contention, in addition, can cause a full loss of forward
14 * progress in which the CPU goes idle.
15 *
16 * This code aggregates individual task delays into resource pressure
17 * metrics that indicate problems with both workload health and
18 * resource utilization.
19 *
20 * Model
21 *
22 * The time in which a task can execute on a CPU is our baseline for
23 * productivity. Pressure expresses the amount of time in which this
24 * potential cannot be realized due to resource contention.
25 *
26 * This concept of productivity has two components: the workload and
27 * the CPU. To measure the impact of pressure on both, we define two
28 * contention states for a resource: SOME and FULL.
29 *
30 * In the SOME state of a given resource, one or more tasks are
31 * delayed on that resource. This affects the workload's ability to
32 * perform work, but the CPU may still be executing other tasks.
33 *
34 * In the FULL state of a given resource, all non-idle tasks are
35 * delayed on that resource such that nobody is advancing and the CPU
36 * goes idle. This leaves both workload and CPU unproductive.
37 *
38 * SOME = nr_delayed_tasks != 0
39 * FULL = nr_delayed_tasks != 0 && nr_productive_tasks == 0
40 *
41 * What it means for a task to be productive is defined differently
42 * for each resource. For IO, productive means a running task. For
43 * memory, productive means a running task that isn't a reclaimer. For
44 * CPU, productive means an on-CPU task.
45 *
46 * Naturally, the FULL state doesn't exist for the CPU resource at the
47 * system level, but exist at the cgroup level. At the cgroup level,
48 * FULL means all non-idle tasks in the cgroup are delayed on the CPU
49 * resource which is being used by others outside of the cgroup or
50 * throttled by the cgroup cpu.max configuration.
51 *
52 * The percentage of wall clock time spent in those compound stall
53 * states gives pressure numbers between 0 and 100 for each resource,
54 * where the SOME percentage indicates workload slowdowns and the FULL
55 * percentage indicates reduced CPU utilization:
56 *
57 * %SOME = time(SOME) / period
58 * %FULL = time(FULL) / period
59 *
60 * Multiple CPUs
61 *
62 * The more tasks and available CPUs there are, the more work can be
63 * performed concurrently. This means that the potential that can go
64 * unrealized due to resource contention *also* scales with non-idle
65 * tasks and CPUs.
66 *
67 * Consider a scenario where 257 number crunching tasks are trying to
68 * run concurrently on 256 CPUs. If we simply aggregated the task
69 * states, we would have to conclude a CPU SOME pressure number of
70 * 100%, since *somebody* is waiting on a runqueue at all
71 * times. However, that is clearly not the amount of contention the
72 * workload is experiencing: only one out of 256 possible execution
73 * threads will be contended at any given time, or about 0.4%.
74 *
75 * Conversely, consider a scenario of 4 tasks and 4 CPUs where at any
76 * given time *one* of the tasks is delayed due to a lack of memory.
77 * Again, looking purely at the task state would yield a memory FULL
78 * pressure number of 0%, since *somebody* is always making forward
79 * progress. But again this wouldn't capture the amount of execution
80 * potential lost, which is 1 out of 4 CPUs, or 25%.
81 *
82 * To calculate wasted potential (pressure) with multiple processors,
83 * we have to base our calculation on the number of non-idle tasks in
84 * conjunction with the number of available CPUs, which is the number
85 * of potential execution threads. SOME becomes then the proportion of
86 * delayed tasks to possible threads, and FULL is the share of possible
87 * threads that are unproductive due to delays:
88 *
89 * threads = min(nr_nonidle_tasks, nr_cpus)
90 * SOME = min(nr_delayed_tasks / threads, 1)
91 * FULL = (threads - min(nr_productive_tasks, threads)) / threads
92 *
93 * For the 257 number crunchers on 256 CPUs, this yields:
94 *
95 * threads = min(257, 256)
96 * SOME = min(1 / 256, 1) = 0.4%
97 * FULL = (256 - min(256, 256)) / 256 = 0%
98 *
99 * For the 1 out of 4 memory-delayed tasks, this yields:
100 *
101 * threads = min(4, 4)
102 * SOME = min(1 / 4, 1) = 25%
103 * FULL = (4 - min(3, 4)) / 4 = 25%
104 *
105 * [ Substitute nr_cpus with 1, and you can see that it's a natural
106 * extension of the single-CPU model. ]
107 *
108 * Implementation
109 *
110 * To assess the precise time spent in each such state, we would have
111 * to freeze the system on task changes and start/stop the state
112 * clocks accordingly. Obviously that doesn't scale in practice.
113 *
114 * Because the scheduler aims to distribute the compute load evenly
115 * among the available CPUs, we can track task state locally to each
116 * CPU and, at much lower frequency, extrapolate the global state for
117 * the cumulative stall times and the running averages.
118 *
119 * For each runqueue, we track:
120 *
121 * tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0)
122 * tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_productive_tasks[cpu])
123 * tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0)
124 *
125 * and then periodically aggregate:
126 *
127 * tNONIDLE = sum(tNONIDLE[i])
128 *
129 * tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE
130 * tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE
131 *
132 * %SOME = tSOME / period
133 * %FULL = tFULL / period
134 *
135 * This gives us an approximation of pressure that is practical
136 * cost-wise, yet way more sensitive and accurate than periodic
137 * sampling of the aggregate task states would be.
138 */
139#include <linux/sched/clock.h>
140#include <linux/workqueue.h>
141#include <linux/psi.h>
142#include "sched.h"
143
144static int psi_bug __read_mostly;
145
146DEFINE_STATIC_KEY_FALSE(psi_disabled);
147static DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled);
148
149#ifdef CONFIG_PSI_DEFAULT_DISABLED
150static bool psi_enable;
151#else
152static bool psi_enable = true;
153#endif
154static int __init setup_psi(char *str)
155{
156 return kstrtobool(str, &psi_enable) == 0;
157}
158__setup("psi=", setup_psi);
159
160/* Running averages - we need to be higher-res than loadavg */
161#define PSI_FREQ (2*HZ+1) /* 2 sec intervals */
162#define EXP_10s 1677 /* 1/exp(2s/10s) as fixed-point */
163#define EXP_60s 1981 /* 1/exp(2s/60s) */
164#define EXP_300s 2034 /* 1/exp(2s/300s) */
165
166/* PSI trigger definitions */
167#define WINDOW_MAX_US 10000000 /* Max window size is 10s */
168#define UPDATES_PER_WINDOW 10 /* 10 updates per window */
169
170/* Sampling frequency in nanoseconds */
171static u64 psi_period __read_mostly;
172
173/* System-level pressure and stall tracking */
174static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
175struct psi_group psi_system = {
176 .pcpu = &system_group_pcpu,
177};
178
179static DEFINE_PER_CPU(seqcount_t, psi_seq) = SEQCNT_ZERO(psi_seq);
180
181static inline void psi_write_begin(int cpu)
182{
183 write_seqcount_begin(per_cpu_ptr(&psi_seq, cpu));
184}
185
186static inline void psi_write_end(int cpu)
187{
188 write_seqcount_end(per_cpu_ptr(&psi_seq, cpu));
189}
190
191static inline u32 psi_read_begin(int cpu)
192{
193 return read_seqcount_begin(per_cpu_ptr(&psi_seq, cpu));
194}
195
196static inline bool psi_read_retry(int cpu, u32 seq)
197{
198 return read_seqcount_retry(per_cpu_ptr(&psi_seq, cpu), seq);
199}
200
201static void psi_avgs_work(struct work_struct *work);
202
203static void poll_timer_fn(struct timer_list *t);
204
205static void group_init(struct psi_group *group)
206{
207 group->enabled = true;
208 group->avg_last_update = sched_clock();
209 group->avg_next_update = group->avg_last_update + psi_period;
210 mutex_init(&group->avgs_lock);
211
212 /* Init avg trigger-related members */
213 INIT_LIST_HEAD(&group->avg_triggers);
214 memset(group->avg_nr_triggers, 0, sizeof(group->avg_nr_triggers));
215 INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
216
217 /* Init rtpoll trigger-related members */
218 atomic_set(&group->rtpoll_scheduled, 0);
219 mutex_init(&group->rtpoll_trigger_lock);
220 INIT_LIST_HEAD(&group->rtpoll_triggers);
221 group->rtpoll_min_period = U32_MAX;
222 group->rtpoll_next_update = ULLONG_MAX;
223 init_waitqueue_head(&group->rtpoll_wait);
224 timer_setup(&group->rtpoll_timer, poll_timer_fn, 0);
225 rcu_assign_pointer(group->rtpoll_task, NULL);
226}
227
228void __init psi_init(void)
229{
230 if (!psi_enable) {
231 static_branch_enable(&psi_disabled);
232 static_branch_disable(&psi_cgroups_enabled);
233 return;
234 }
235
236 if (!cgroup_psi_enabled())
237 static_branch_disable(&psi_cgroups_enabled);
238
239 psi_period = jiffies_to_nsecs(PSI_FREQ);
240 group_init(&psi_system);
241}
242
243static u32 test_states(unsigned int *tasks, u32 state_mask)
244{
245 const bool oncpu = state_mask & PSI_ONCPU;
246
247 if (tasks[NR_IOWAIT]) {
248 state_mask |= BIT(PSI_IO_SOME);
249 if (!tasks[NR_RUNNING])
250 state_mask |= BIT(PSI_IO_FULL);
251 }
252
253 if (tasks[NR_MEMSTALL]) {
254 state_mask |= BIT(PSI_MEM_SOME);
255 if (tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING])
256 state_mask |= BIT(PSI_MEM_FULL);
257 }
258
259 if (tasks[NR_RUNNING] > oncpu)
260 state_mask |= BIT(PSI_CPU_SOME);
261
262 if (tasks[NR_RUNNING] && !oncpu)
263 state_mask |= BIT(PSI_CPU_FULL);
264
265 if (tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] || tasks[NR_RUNNING])
266 state_mask |= BIT(PSI_NONIDLE);
267
268 return state_mask;
269}
270
271static void get_recent_times(struct psi_group *group, int cpu,
272 enum psi_aggregators aggregator, u32 *times,
273 u32 *pchanged_states)
274{
275 struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
276 int current_cpu = raw_smp_processor_id();
277 unsigned int tasks[NR_PSI_TASK_COUNTS];
278 u64 now, state_start;
279 enum psi_states s;
280 unsigned int seq;
281 u32 state_mask;
282
283 *pchanged_states = 0;
284
285 /* Snapshot a coherent view of the CPU state */
286 do {
287 seq = psi_read_begin(cpu);
288 now = cpu_clock(cpu);
289 memcpy(times, groupc->times, sizeof(groupc->times));
290 state_mask = groupc->state_mask;
291 state_start = groupc->state_start;
292 if (cpu == current_cpu)
293 memcpy(tasks, groupc->tasks, sizeof(groupc->tasks));
294 } while (psi_read_retry(cpu, seq));
295
296 /* Calculate state time deltas against the previous snapshot */
297 for (s = 0; s < NR_PSI_STATES; s++) {
298 u32 delta;
299 /*
300 * In addition to already concluded states, we also
301 * incorporate currently active states on the CPU,
302 * since states may last for many sampling periods.
303 *
304 * This way we keep our delta sampling buckets small
305 * (u32) and our reported pressure close to what's
306 * actually happening.
307 */
308 if (state_mask & (1 << s))
309 times[s] += now - state_start;
310
311 delta = times[s] - groupc->times_prev[aggregator][s];
312 groupc->times_prev[aggregator][s] = times[s];
313
314 times[s] = delta;
315 if (delta)
316 *pchanged_states |= (1 << s);
317 }
318
319 /*
320 * When collect_percpu_times() from the avgs_work, we don't want to
321 * re-arm avgs_work when all CPUs are IDLE. But the current CPU running
322 * this avgs_work is never IDLE, cause avgs_work can't be shut off.
323 * So for the current CPU, we need to re-arm avgs_work only when
324 * (NR_RUNNING > 1 || NR_IOWAIT > 0 || NR_MEMSTALL > 0), for other CPUs
325 * we can just check PSI_NONIDLE delta.
326 */
327 if (current_work() == &group->avgs_work.work) {
328 bool reschedule;
329
330 if (cpu == current_cpu)
331 reschedule = tasks[NR_RUNNING] +
332 tasks[NR_IOWAIT] +
333 tasks[NR_MEMSTALL] > 1;
334 else
335 reschedule = *pchanged_states & (1 << PSI_NONIDLE);
336
337 if (reschedule)
338 *pchanged_states |= PSI_STATE_RESCHEDULE;
339 }
340}
341
342static void calc_avgs(unsigned long avg[3], int missed_periods,
343 u64 time, u64 period)
344{
345 unsigned long pct;
346
347 /* Fill in zeroes for periods of no activity */
348 if (missed_periods) {
349 avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
350 avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
351 avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
352 }
353
354 /* Sample the most recent active period */
355 pct = div_u64(time * 100, period);
356 pct *= FIXED_1;
357 avg[0] = calc_load(avg[0], EXP_10s, pct);
358 avg[1] = calc_load(avg[1], EXP_60s, pct);
359 avg[2] = calc_load(avg[2], EXP_300s, pct);
360}
361
362static void collect_percpu_times(struct psi_group *group,
363 enum psi_aggregators aggregator,
364 u32 *pchanged_states)
365{
366 u64 deltas[NR_PSI_STATES - 1] = { 0, };
367 unsigned long nonidle_total = 0;
368 u32 changed_states = 0;
369 int cpu;
370 int s;
371
372 /*
373 * Collect the per-cpu time buckets and average them into a
374 * single time sample that is normalized to wall clock time.
375 *
376 * For averaging, each CPU is weighted by its non-idle time in
377 * the sampling period. This eliminates artifacts from uneven
378 * loading, or even entirely idle CPUs.
379 */
380 for_each_possible_cpu(cpu) {
381 u32 times[NR_PSI_STATES];
382 u32 nonidle;
383 u32 cpu_changed_states;
384
385 get_recent_times(group, cpu, aggregator, times,
386 &cpu_changed_states);
387 changed_states |= cpu_changed_states;
388
389 nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
390 nonidle_total += nonidle;
391
392 for (s = 0; s < PSI_NONIDLE; s++)
393 deltas[s] += (u64)times[s] * nonidle;
394 }
395
396 /*
397 * Integrate the sample into the running statistics that are
398 * reported to userspace: the cumulative stall times and the
399 * decaying averages.
400 *
401 * Pressure percentages are sampled at PSI_FREQ. We might be
402 * called more often when the user polls more frequently than
403 * that; we might be called less often when there is no task
404 * activity, thus no data, and clock ticks are sporadic. The
405 * below handles both.
406 */
407
408 /* total= */
409 for (s = 0; s < NR_PSI_STATES - 1; s++)
410 group->total[aggregator][s] +=
411 div_u64(deltas[s], max(nonidle_total, 1UL));
412
413 if (pchanged_states)
414 *pchanged_states = changed_states;
415}
416
417/* Trigger tracking window manipulations */
418static void window_reset(struct psi_window *win, u64 now, u64 value,
419 u64 prev_growth)
420{
421 win->start_time = now;
422 win->start_value = value;
423 win->prev_growth = prev_growth;
424}
425
426/*
427 * PSI growth tracking window update and growth calculation routine.
428 *
429 * This approximates a sliding tracking window by interpolating
430 * partially elapsed windows using historical growth data from the
431 * previous intervals. This minimizes memory requirements (by not storing
432 * all the intermediate values in the previous window) and simplifies
433 * the calculations. It works well because PSI signal changes only in
434 * positive direction and over relatively small window sizes the growth
435 * is close to linear.
436 */
437static u64 window_update(struct psi_window *win, u64 now, u64 value)
438{
439 u64 elapsed;
440 u64 growth;
441
442 elapsed = now - win->start_time;
443 growth = value - win->start_value;
444 /*
445 * After each tracking window passes win->start_value and
446 * win->start_time get reset and win->prev_growth stores
447 * the average per-window growth of the previous window.
448 * win->prev_growth is then used to interpolate additional
449 * growth from the previous window assuming it was linear.
450 */
451 if (elapsed > win->size)
452 window_reset(win, now, value, growth);
453 else {
454 u32 remaining;
455
456 remaining = win->size - elapsed;
457 growth += div64_u64(win->prev_growth * remaining, win->size);
458 }
459
460 return growth;
461}
462
463static void update_triggers(struct psi_group *group, u64 now,
464 enum psi_aggregators aggregator)
465{
466 struct psi_trigger *t;
467 u64 *total = group->total[aggregator];
468 struct list_head *triggers;
469 u64 *aggregator_total;
470
471 if (aggregator == PSI_AVGS) {
472 triggers = &group->avg_triggers;
473 aggregator_total = group->avg_total;
474 } else {
475 triggers = &group->rtpoll_triggers;
476 aggregator_total = group->rtpoll_total;
477 }
478
479 /*
480 * On subsequent updates, calculate growth deltas and let
481 * watchers know when their specified thresholds are exceeded.
482 */
483 list_for_each_entry(t, triggers, node) {
484 u64 growth;
485 bool new_stall;
486
487 new_stall = aggregator_total[t->state] != total[t->state];
488
489 /* Check for stall activity or a previous threshold breach */
490 if (!new_stall && !t->pending_event)
491 continue;
492 /*
493 * Check for new stall activity, as well as deferred
494 * events that occurred in the last window after the
495 * trigger had already fired (we want to ratelimit
496 * events without dropping any).
497 */
498 if (new_stall) {
499 /* Calculate growth since last update */
500 growth = window_update(&t->win, now, total[t->state]);
501 if (!t->pending_event) {
502 if (growth < t->threshold)
503 continue;
504
505 t->pending_event = true;
506 }
507 }
508 /* Limit event signaling to once per window */
509 if (now < t->last_event_time + t->win.size)
510 continue;
511
512 /* Generate an event */
513 if (cmpxchg(&t->event, 0, 1) == 0) {
514 if (t->of)
515 kernfs_notify(t->of->kn);
516 else
517 wake_up_interruptible(&t->event_wait);
518 }
519 t->last_event_time = now;
520 /* Reset threshold breach flag once event got generated */
521 t->pending_event = false;
522 }
523}
524
525static u64 update_averages(struct psi_group *group, u64 now)
526{
527 unsigned long missed_periods = 0;
528 u64 expires, period;
529 u64 avg_next_update;
530 int s;
531
532 /* avgX= */
533 expires = group->avg_next_update;
534 if (now - expires >= psi_period)
535 missed_periods = div_u64(now - expires, psi_period);
536
537 /*
538 * The periodic clock tick can get delayed for various
539 * reasons, especially on loaded systems. To avoid clock
540 * drift, we schedule the clock in fixed psi_period intervals.
541 * But the deltas we sample out of the per-cpu buckets above
542 * are based on the actual time elapsing between clock ticks.
543 */
544 avg_next_update = expires + ((1 + missed_periods) * psi_period);
545 period = now - (group->avg_last_update + (missed_periods * psi_period));
546 group->avg_last_update = now;
547
548 for (s = 0; s < NR_PSI_STATES - 1; s++) {
549 u32 sample;
550
551 sample = group->total[PSI_AVGS][s] - group->avg_total[s];
552 /*
553 * Due to the lockless sampling of the time buckets,
554 * recorded time deltas can slip into the next period,
555 * which under full pressure can result in samples in
556 * excess of the period length.
557 *
558 * We don't want to report non-sensical pressures in
559 * excess of 100%, nor do we want to drop such events
560 * on the floor. Instead we punt any overage into the
561 * future until pressure subsides. By doing this we
562 * don't underreport the occurring pressure curve, we
563 * just report it delayed by one period length.
564 *
565 * The error isn't cumulative. As soon as another
566 * delta slips from a period P to P+1, by definition
567 * it frees up its time T in P.
568 */
569 if (sample > period)
570 sample = period;
571 group->avg_total[s] += sample;
572 calc_avgs(group->avg[s], missed_periods, sample, period);
573 }
574
575 return avg_next_update;
576}
577
578static void psi_avgs_work(struct work_struct *work)
579{
580 struct delayed_work *dwork;
581 struct psi_group *group;
582 u32 changed_states;
583 u64 now;
584
585 dwork = to_delayed_work(work);
586 group = container_of(dwork, struct psi_group, avgs_work);
587
588 mutex_lock(&group->avgs_lock);
589
590 now = sched_clock();
591
592 collect_percpu_times(group, PSI_AVGS, &changed_states);
593 /*
594 * If there is task activity, periodically fold the per-cpu
595 * times and feed samples into the running averages. If things
596 * are idle and there is no data to process, stop the clock.
597 * Once restarted, we'll catch up the running averages in one
598 * go - see calc_avgs() and missed_periods.
599 */
600 if (now >= group->avg_next_update) {
601 update_triggers(group, now, PSI_AVGS);
602 group->avg_next_update = update_averages(group, now);
603 }
604
605 if (changed_states & PSI_STATE_RESCHEDULE) {
606 schedule_delayed_work(dwork, nsecs_to_jiffies(
607 group->avg_next_update - now) + 1);
608 }
609
610 mutex_unlock(&group->avgs_lock);
611}
612
613static void init_rtpoll_triggers(struct psi_group *group, u64 now)
614{
615 struct psi_trigger *t;
616
617 list_for_each_entry(t, &group->rtpoll_triggers, node)
618 window_reset(&t->win, now,
619 group->total[PSI_POLL][t->state], 0);
620 memcpy(group->rtpoll_total, group->total[PSI_POLL],
621 sizeof(group->rtpoll_total));
622 group->rtpoll_next_update = now + group->rtpoll_min_period;
623}
624
625/* Schedule rtpolling if it's not already scheduled or forced. */
626static void psi_schedule_rtpoll_work(struct psi_group *group, unsigned long delay,
627 bool force)
628{
629 /*
630 * atomic_xchg should be called even when !force to provide a
631 * full memory barrier (see the comment inside psi_rtpoll_work).
632 */
633 if (atomic_xchg(&group->rtpoll_scheduled, 1) && !force)
634 return;
635
636 /*
637 * Only test whether a worker is installed; do not dereference the task.
638 * A racing trigger destruction may leave the timer armed.
639 * psi_cgroup_free() shuts it down before freeing the group.
640 * The system PSI group is static.
641 */
642 if (likely(rcu_access_pointer(group->rtpoll_task)))
643 mod_timer(&group->rtpoll_timer, jiffies + delay);
644 else
645 atomic_set(&group->rtpoll_scheduled, 0);
646}
647
648static void psi_rtpoll_work(struct psi_group *group)
649{
650 bool force_reschedule = false;
651 u32 changed_states;
652 u64 now;
653
654 mutex_lock(&group->rtpoll_trigger_lock);
655
656 now = sched_clock();
657
658 if (now > group->rtpoll_until) {
659 /*
660 * We are either about to start or might stop rtpolling if no
661 * state change was recorded. Resetting rtpoll_scheduled leaves
662 * a small window for psi_group_change to sneak in and schedule
663 * an immediate rtpoll_work before we get to rescheduling. One
664 * potential extra wakeup at the end of the rtpolling window
665 * should be negligible and rtpoll_next_update still keeps
666 * updates correctly on schedule.
667 */
668 atomic_set(&group->rtpoll_scheduled, 0);
669 /*
670 * A task change can race with the rtpoll worker that is supposed to
671 * report on it. To avoid missing events, ensure ordering between
672 * rtpoll_scheduled and the task state accesses, such that if the
673 * rtpoll worker misses the state update, the task change is
674 * guaranteed to reschedule the rtpoll worker:
675 *
676 * rtpoll worker:
677 * atomic_set(rtpoll_scheduled, 0)
678 * smp_mb()
679 * LOAD states
680 *
681 * task change:
682 * STORE states
683 * if atomic_xchg(rtpoll_scheduled, 1) == 0:
684 * schedule rtpoll worker
685 *
686 * The atomic_xchg() implies a full barrier.
687 */
688 smp_mb();
689 } else {
690 /* The rtpolling window is not over, keep rescheduling */
691 force_reschedule = true;
692 }
693
694
695 collect_percpu_times(group, PSI_POLL, &changed_states);
696
697 if (changed_states & group->rtpoll_states) {
698 /* Initialize trigger windows when entering rtpolling mode */
699 if (now > group->rtpoll_until)
700 init_rtpoll_triggers(group, now);
701
702 /*
703 * Keep the monitor active for at least the duration of the
704 * minimum tracking window as long as monitor states are
705 * changing.
706 */
707 group->rtpoll_until = now +
708 group->rtpoll_min_period * UPDATES_PER_WINDOW;
709 }
710
711 if (now > group->rtpoll_until) {
712 group->rtpoll_next_update = ULLONG_MAX;
713 goto out;
714 }
715
716 if (now >= group->rtpoll_next_update) {
717 if (changed_states & group->rtpoll_states) {
718 update_triggers(group, now, PSI_POLL);
719 memcpy(group->rtpoll_total, group->total[PSI_POLL],
720 sizeof(group->rtpoll_total));
721 }
722 group->rtpoll_next_update = now + group->rtpoll_min_period;
723 }
724
725 psi_schedule_rtpoll_work(group,
726 nsecs_to_jiffies(group->rtpoll_next_update - now) + 1,
727 force_reschedule);
728
729out:
730 mutex_unlock(&group->rtpoll_trigger_lock);
731}
732
733static int psi_rtpoll_worker(void *data)
734{
735 struct psi_group *group = (struct psi_group *)data;
736
737 sched_set_fifo_low(current);
738
739 while (true) {
740 wait_event_interruptible(group->rtpoll_wait,
741 atomic_cmpxchg(&group->rtpoll_wakeup, 1, 0) ||
742 kthread_should_stop());
743 if (kthread_should_stop())
744 break;
745
746 psi_rtpoll_work(group);
747 }
748 return 0;
749}
750
751static void poll_timer_fn(struct timer_list *t)
752{
753 struct psi_group *group = timer_container_of(group, t, rtpoll_timer);
754
755 atomic_set(&group->rtpoll_wakeup, 1);
756 wake_up_interruptible(&group->rtpoll_wait);
757}
758
759static void record_times(struct psi_group_cpu *groupc, u64 now)
760{
761 u32 delta;
762
763 delta = now - groupc->state_start;
764 groupc->state_start = now;
765
766 if (groupc->state_mask & (1 << PSI_IO_SOME)) {
767 groupc->times[PSI_IO_SOME] += delta;
768 if (groupc->state_mask & (1 << PSI_IO_FULL))
769 groupc->times[PSI_IO_FULL] += delta;
770 }
771
772 if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
773 groupc->times[PSI_MEM_SOME] += delta;
774 if (groupc->state_mask & (1 << PSI_MEM_FULL))
775 groupc->times[PSI_MEM_FULL] += delta;
776 }
777
778 if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
779 groupc->times[PSI_CPU_SOME] += delta;
780 if (groupc->state_mask & (1 << PSI_CPU_FULL))
781 groupc->times[PSI_CPU_FULL] += delta;
782 }
783
784 if (groupc->state_mask & (1 << PSI_NONIDLE))
785 groupc->times[PSI_NONIDLE] += delta;
786}
787
788#define for_each_group(iter, group) \
789 for (typeof(group) iter = group; iter; iter = iter->parent)
790
791static void psi_group_change(struct psi_group *group, int cpu,
792 unsigned int clear, unsigned int set,
793 u64 now, bool wake_clock)
794{
795 struct psi_group_cpu *groupc;
796 unsigned int t, m;
797 u32 state_mask;
798
799 lockdep_assert_rq_held(cpu_rq(cpu));
800 groupc = per_cpu_ptr(group->pcpu, cpu);
801
802 /*
803 * Start with TSK_ONCPU, which doesn't have a corresponding
804 * task count - it's just a boolean flag directly encoded in
805 * the state mask. Clear, set, or carry the current state if
806 * no changes are requested.
807 */
808 if (unlikely(clear & TSK_ONCPU)) {
809 state_mask = 0;
810 clear &= ~TSK_ONCPU;
811 } else if (unlikely(set & TSK_ONCPU)) {
812 state_mask = PSI_ONCPU;
813 set &= ~TSK_ONCPU;
814 } else {
815 state_mask = groupc->state_mask & PSI_ONCPU;
816 }
817
818 /*
819 * The rest of the state mask is calculated based on the task
820 * counts. Update those first, then construct the mask.
821 */
822 for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
823 if (!(m & (1 << t)))
824 continue;
825 if (groupc->tasks[t]) {
826 groupc->tasks[t]--;
827 } else if (!psi_bug) {
828 printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n",
829 cpu, t, groupc->tasks[0],
830 groupc->tasks[1], groupc->tasks[2],
831 groupc->tasks[3], clear, set);
832 psi_bug = 1;
833 }
834 }
835
836 for (t = 0; set; set &= ~(1 << t), t++)
837 if (set & (1 << t))
838 groupc->tasks[t]++;
839
840 if (!group->enabled) {
841 /*
842 * On the first group change after disabling PSI, conclude
843 * the current state and flush its time. This is unlikely
844 * to matter to the user, but aggregation (get_recent_times)
845 * may have already incorporated the live state into times_prev;
846 * avoid a delta sample underflow when PSI is later re-enabled.
847 */
848 if (unlikely(groupc->state_mask & (1 << PSI_NONIDLE)))
849 record_times(groupc, now);
850
851 groupc->state_mask = state_mask;
852
853 return;
854 }
855
856 state_mask = test_states(groupc->tasks, state_mask);
857
858 /*
859 * Since we care about lost potential, a memstall is FULL
860 * when there are no other working tasks, but also when
861 * the CPU is actively reclaiming and nothing productive
862 * could run even if it were runnable. So when the current
863 * task in a cgroup is in_memstall, the corresponding groupc
864 * on that cpu is in PSI_MEM_FULL state.
865 */
866 if (unlikely((state_mask & PSI_ONCPU) && cpu_curr(cpu)->in_memstall))
867 state_mask |= (1 << PSI_MEM_FULL);
868
869 record_times(groupc, now);
870
871 groupc->state_mask = state_mask;
872
873 if (state_mask & group->rtpoll_states)
874 psi_schedule_rtpoll_work(group, 1, false);
875
876 if (wake_clock && !delayed_work_pending(&group->avgs_work))
877 schedule_delayed_work(&group->avgs_work, PSI_FREQ);
878}
879
880static inline struct psi_group *task_psi_group(struct task_struct *task)
881{
882#ifdef CONFIG_CGROUPS
883 if (static_branch_likely(&psi_cgroups_enabled))
884 return cgroup_psi(task_dfl_cgroup(task));
885#endif
886 return &psi_system;
887}
888
889static void psi_flags_change(struct task_struct *task, int clear, int set)
890{
891 if (((task->psi_flags & set) ||
892 (task->psi_flags & clear) != clear) &&
893 !psi_bug) {
894 printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
895 task->pid, task->comm, task_cpu(task),
896 task->psi_flags, clear, set);
897 psi_bug = 1;
898 }
899
900 task->psi_flags &= ~clear;
901 task->psi_flags |= set;
902}
903
904void psi_task_change(struct task_struct *task, int clear, int set)
905{
906 int cpu = task_cpu(task);
907 u64 now;
908
909 if (!task->pid)
910 return;
911
912 psi_flags_change(task, clear, set);
913
914 psi_write_begin(cpu);
915 now = cpu_clock(cpu);
916 for_each_group(group, task_psi_group(task))
917 psi_group_change(group, cpu, clear, set, now, true);
918 psi_write_end(cpu);
919}
920
921void psi_task_switch(struct task_struct *prev, struct task_struct *next,
922 bool sleep)
923{
924 struct psi_group *common = NULL;
925 int cpu = task_cpu(prev);
926 u64 now;
927
928 psi_write_begin(cpu);
929 now = cpu_clock(cpu);
930
931 if (next->pid) {
932 psi_flags_change(next, 0, TSK_ONCPU);
933 /*
934 * Set TSK_ONCPU on @next's cgroups. If @next shares any
935 * ancestors with @prev, those will already have @prev's
936 * TSK_ONCPU bit set, and we can stop the iteration there.
937 */
938 for_each_group(group, task_psi_group(next)) {
939 struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
940
941 if (groupc->state_mask & PSI_ONCPU) {
942 common = group;
943 break;
944 }
945 psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
946 }
947 }
948
949 if (prev->pid) {
950 int clear = TSK_ONCPU, set = 0;
951 bool wake_clock = true;
952
953 /*
954 * When we're going to sleep, psi_dequeue() lets us
955 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
956 * TSK_IOWAIT here, where we can combine it with
957 * TSK_ONCPU and save walking common ancestors twice.
958 */
959 if (sleep) {
960 clear |= TSK_RUNNING;
961 if (prev->in_memstall)
962 clear |= TSK_MEMSTALL_RUNNING;
963 if (prev->in_iowait)
964 set |= TSK_IOWAIT;
965
966 /*
967 * Periodic aggregation shuts off if there is a period of no
968 * task changes, so we wake it back up if necessary. However,
969 * don't do this if the task change is the aggregation worker
970 * itself going to sleep, or we'll ping-pong forever.
971 */
972 if (unlikely((prev->flags & PF_WQ_WORKER) &&
973 wq_worker_last_func(prev) == psi_avgs_work))
974 wake_clock = false;
975 }
976
977 psi_flags_change(prev, clear, set);
978
979 for_each_group(group, task_psi_group(prev)) {
980 if (group == common)
981 break;
982 psi_group_change(group, cpu, clear, set, now, wake_clock);
983 }
984
985 /*
986 * TSK_ONCPU is handled up to the common ancestor. If there are
987 * any other differences between the two tasks (e.g. prev goes
988 * to sleep, or only one task is memstall), finish propagating
989 * those differences all the way up to the root.
990 */
991 if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) {
992 clear &= ~TSK_ONCPU;
993 for_each_group(group, common)
994 psi_group_change(group, cpu, clear, set, now, wake_clock);
995 }
996 }
997 psi_write_end(cpu);
998}
999
1000#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1001void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_struct *prev)
1002{
1003 int cpu = task_cpu(curr);
1004 struct psi_group_cpu *groupc;
1005 s64 delta;
1006 u64 irq;
1007 u64 now;
1008
1009 if (static_branch_likely(&psi_disabled) || !irqtime_enabled())
1010 return;
1011
1012 if (!curr->pid)
1013 return;
1014
1015 lockdep_assert_rq_held(rq);
1016 if (prev && task_psi_group(prev) == task_psi_group(curr))
1017 return;
1018
1019 irq = irq_time_read(cpu);
1020 delta = (s64)(irq - rq->psi_irq_time);
1021 if (delta <= 0)
1022 return;
1023 rq->psi_irq_time = irq;
1024
1025 psi_write_begin(cpu);
1026 now = cpu_clock(cpu);
1027
1028 for_each_group(group, task_psi_group(curr)) {
1029 if (!group->enabled)
1030 continue;
1031
1032 groupc = per_cpu_ptr(group->pcpu, cpu);
1033
1034 record_times(groupc, now);
1035 groupc->times[PSI_IRQ_FULL] += delta;
1036
1037 if (group->rtpoll_states & (1 << PSI_IRQ_FULL))
1038 psi_schedule_rtpoll_work(group, 1, false);
1039 }
1040 psi_write_end(cpu);
1041}
1042#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1043
1044/**
1045 * psi_memstall_enter - mark the beginning of a memory stall section
1046 * @flags: flags to handle nested sections
1047 *
1048 * Marks the calling task as being stalled due to a lack of memory,
1049 * such as waiting for a refault or performing reclaim.
1050 */
1051void psi_memstall_enter(unsigned long *flags)
1052{
1053 struct rq_flags rf;
1054 struct rq *rq;
1055
1056 if (static_branch_likely(&psi_disabled))
1057 return;
1058
1059 *flags = current->in_memstall;
1060 if (*flags)
1061 return;
1062 /*
1063 * in_memstall setting & accounting needs to be atomic wrt
1064 * changes to the task's scheduling state, otherwise we can
1065 * race with CPU migration.
1066 */
1067 rq = this_rq_lock_irq(&rf);
1068
1069 current->in_memstall = 1;
1070 psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
1071
1072 rq_unlock_irq(rq, &rf);
1073}
1074EXPORT_SYMBOL_GPL(psi_memstall_enter);
1075
1076/**
1077 * psi_memstall_leave - mark the end of an memory stall section
1078 * @flags: flags to handle nested memdelay sections
1079 *
1080 * Marks the calling task as no longer stalled due to lack of memory.
1081 */
1082void psi_memstall_leave(unsigned long *flags)
1083{
1084 struct rq_flags rf;
1085 struct rq *rq;
1086
1087 if (static_branch_likely(&psi_disabled))
1088 return;
1089
1090 if (*flags)
1091 return;
1092 /*
1093 * in_memstall clearing & accounting needs to be atomic wrt
1094 * changes to the task's scheduling state, otherwise we could
1095 * race with CPU migration.
1096 */
1097 rq = this_rq_lock_irq(&rf);
1098
1099 current->in_memstall = 0;
1100 psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0);
1101
1102 rq_unlock_irq(rq, &rf);
1103}
1104EXPORT_SYMBOL_GPL(psi_memstall_leave);
1105
1106#ifdef CONFIG_CGROUPS
1107int psi_cgroup_alloc(struct cgroup *cgroup)
1108{
1109 if (!static_branch_likely(&psi_cgroups_enabled))
1110 return 0;
1111
1112 cgroup->psi = kzalloc_obj(struct psi_group);
1113 if (!cgroup->psi)
1114 return -ENOMEM;
1115
1116 cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
1117 if (!cgroup->psi->pcpu) {
1118 kfree(cgroup->psi);
1119 return -ENOMEM;
1120 }
1121 group_init(cgroup->psi);
1122 cgroup->psi->parent = cgroup_psi(cgroup_parent(cgroup));
1123 return 0;
1124}
1125
1126void psi_cgroup_free(struct cgroup *cgroup)
1127{
1128 if (!static_branch_likely(&psi_cgroups_enabled))
1129 return;
1130
1131 cancel_delayed_work_sync(&cgroup->psi->avgs_work);
1132 /*
1133 * A psi_schedule_rtpoll_work() call racing the last trigger's
1134 * destruction may have re-armed the timer after psi_trigger_destroy()
1135 * deleted it. Spurious firing while the group is alive is harmless.
1136 */
1137 timer_shutdown_sync(&cgroup->psi->rtpoll_timer);
1138 free_percpu(cgroup->psi->pcpu);
1139 /* All triggers must be removed by now */
1140 WARN_ONCE(cgroup->psi->rtpoll_states, "psi: trigger leak\n");
1141 kfree(cgroup->psi);
1142}
1143
1144/**
1145 * cgroup_move_task - move task to a different cgroup
1146 * @task: the task
1147 * @to: the target css_set
1148 *
1149 * Move task to a new cgroup and safely migrate its associated stall
1150 * state between the different groups.
1151 *
1152 * This function acquires the task's rq lock to lock out concurrent
1153 * changes to the task's scheduling state and - in case the task is
1154 * running - concurrent changes to its stall state.
1155 */
1156void cgroup_move_task(struct task_struct *task, struct css_set *to)
1157{
1158 unsigned int task_flags;
1159 struct rq_flags rf;
1160 struct rq *rq;
1161
1162 if (!static_branch_likely(&psi_cgroups_enabled)) {
1163 /*
1164 * Lame to do this here, but the scheduler cannot be locked
1165 * from the outside, so we move cgroups from inside sched/.
1166 */
1167 rcu_assign_pointer(task->cgroups, to);
1168 return;
1169 }
1170
1171 rq = task_rq_lock(task, &rf);
1172
1173 /*
1174 * We may race with schedule() dropping the rq lock between
1175 * deactivating prev and switching to next. Because the psi
1176 * updates from the deactivation are deferred to the switch
1177 * callback to save cgroup tree updates, the task's scheduling
1178 * state here is not coherent with its psi state:
1179 *
1180 * schedule() cgroup_move_task()
1181 * rq_lock()
1182 * deactivate_task()
1183 * p->on_rq = 0
1184 * psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1185 * pick_next_task()
1186 * rq_unlock()
1187 * rq_lock()
1188 * psi_task_change() // old cgroup
1189 * task->cgroups = to
1190 * psi_task_change() // new cgroup
1191 * rq_unlock()
1192 * rq_lock()
1193 * psi_sched_switch() // does deferred updates in new cgroup
1194 *
1195 * Don't rely on the scheduling state. Use psi_flags instead.
1196 */
1197 task_flags = task->psi_flags;
1198
1199 if (task_flags)
1200 psi_task_change(task, task_flags, 0);
1201
1202 /* See comment above */
1203 rcu_assign_pointer(task->cgroups, to);
1204
1205 if (task_flags)
1206 psi_task_change(task, 0, task_flags);
1207
1208 task_rq_unlock(rq, task, &rf);
1209}
1210
1211void psi_cgroup_restart(struct psi_group *group)
1212{
1213 int cpu;
1214
1215 /*
1216 * After we disable psi_group->enabled, we don't actually
1217 * stop percpu tasks accounting in each psi_group_cpu,
1218 * instead only stop test_states() loop, record_times()
1219 * and averaging worker, see psi_group_change() for details.
1220 *
1221 * When disable cgroup PSI, this function has nothing to sync
1222 * since cgroup pressure files are hidden and percpu psi_group_cpu
1223 * would see !psi_group->enabled and only do task accounting.
1224 *
1225 * When re-enable cgroup PSI, this function use psi_group_change()
1226 * to get correct state mask from test_states() loop on tasks[],
1227 * and restart groupc->state_start from now, use .clear = .set = 0
1228 * here since no task status really changed.
1229 */
1230 if (!group->enabled)
1231 return;
1232
1233 for_each_possible_cpu(cpu) {
1234 u64 now;
1235
1236 guard(rq_lock_irq)(cpu_rq(cpu));
1237
1238 psi_write_begin(cpu);
1239 now = cpu_clock(cpu);
1240 psi_group_change(group, cpu, 0, 0, now, true);
1241 psi_write_end(cpu);
1242 }
1243}
1244#endif /* CONFIG_CGROUPS */
1245
1246int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1247{
1248 bool only_full = false;
1249 int full;
1250 u64 now;
1251
1252 if (static_branch_likely(&psi_disabled))
1253 return -EOPNOTSUPP;
1254
1255#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1256 if (!irqtime_enabled() && res == PSI_IRQ)
1257 return -EOPNOTSUPP;
1258#endif
1259
1260 /* Update averages before reporting them */
1261 mutex_lock(&group->avgs_lock);
1262 now = sched_clock();
1263 collect_percpu_times(group, PSI_AVGS, NULL);
1264 if (now >= group->avg_next_update)
1265 group->avg_next_update = update_averages(group, now);
1266 mutex_unlock(&group->avgs_lock);
1267
1268#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1269 only_full = res == PSI_IRQ;
1270#endif
1271
1272 for (full = 0; full < 2 - only_full; full++) {
1273 unsigned long avg[3] = { 0, };
1274 u64 total = 0;
1275 int w;
1276
1277 /* CPU FULL is undefined at the system level */
1278 if (!(group == &psi_system && res == PSI_CPU && full)) {
1279 for (w = 0; w < 3; w++)
1280 avg[w] = group->avg[res * 2 + full][w];
1281 total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1282 NSEC_PER_USEC);
1283 }
1284
1285 seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1286 full || only_full ? "full" : "some",
1287 LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1288 LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1289 LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1290 total);
1291 }
1292
1293 return 0;
1294}
1295
1296/*
1297 * Create @group's rtpoll worker after psi_trigger_create() reported the need
1298 * for one. kthread creation depends on the whole fork path and we don't want
1299 * all of that nested inside cgroup_mutex, so the caller must drop it and any
1300 * other lock that forks can wait behind. If two callers race, the loser stops
1301 * its never-woken kthread.
1302 */
1303int psi_trigger_create_rtpoll_worker(struct psi_group *group)
1304{
1305 struct task_struct *task;
1306
1307 task = kthread_create(psi_rtpoll_worker, group, "psimon");
1308 if (IS_ERR(task))
1309 return PTR_ERR(task);
1310
1311 scoped_guard(mutex, &group->rtpoll_trigger_lock) {
1312 if (!rcu_access_pointer(group->rtpoll_task)) {
1313 atomic_set(&group->rtpoll_wakeup, 0);
1314 wake_up_process(task);
1315 rcu_assign_pointer(group->rtpoll_task, task);
1316
1317 /*
1318 * Poll once to catch up on scheduling attempts dropped
1319 * while there was no rtpoll worker.
1320 */
1321 psi_schedule_rtpoll_work(group, 1, true);
1322 return 0;
1323 }
1324 }
1325
1326 kthread_stop(task);
1327 return 0;
1328}
1329
1330struct psi_trigger *psi_trigger_create(struct psi_group *group, char *buf,
1331 enum psi_res res, struct file *file,
1332 struct kernfs_open_file *of,
1333 bool *need_rtpoll_worker)
1334{
1335 struct psi_trigger *t;
1336 enum psi_states state;
1337 u32 threshold_us;
1338 bool privileged;
1339 u32 window_us;
1340
1341 *need_rtpoll_worker = false;
1342
1343 if (static_branch_likely(&psi_disabled))
1344 return ERR_PTR(-EOPNOTSUPP);
1345
1346 /*
1347 * Checking the privilege here on file->f_cred implies that a privileged user
1348 * could open the file and delegate the write to an unprivileged one.
1349 */
1350 privileged = cap_raised(file->f_cred->cap_effective, CAP_SYS_RESOURCE);
1351
1352 if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1353 state = PSI_IO_SOME + res * 2;
1354 else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1355 state = PSI_IO_FULL + res * 2;
1356 else
1357 return ERR_PTR(-EINVAL);
1358
1359#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1360 if (res == PSI_IRQ && --state != PSI_IRQ_FULL)
1361 return ERR_PTR(-EINVAL);
1362#endif
1363
1364 if (state >= PSI_NONIDLE)
1365 return ERR_PTR(-EINVAL);
1366
1367 if (window_us == 0 || window_us > WINDOW_MAX_US)
1368 return ERR_PTR(-EINVAL);
1369
1370 /*
1371 * Unprivileged users can only use 2s windows so that averages aggregation
1372 * work is used, and no RT threads need to be spawned.
1373 */
1374 if (!privileged && window_us % 2000000)
1375 return ERR_PTR(-EINVAL);
1376
1377 /* Check threshold */
1378 if (threshold_us == 0 || threshold_us > window_us)
1379 return ERR_PTR(-EINVAL);
1380
1381 t = kmalloc_obj(*t);
1382 if (!t)
1383 return ERR_PTR(-ENOMEM);
1384
1385 t->group = group;
1386 t->state = state;
1387 t->threshold = threshold_us * NSEC_PER_USEC;
1388 t->win.size = window_us * NSEC_PER_USEC;
1389 window_reset(&t->win, sched_clock(),
1390 group->total[PSI_POLL][t->state], 0);
1391
1392 t->event = 0;
1393 t->last_event_time = 0;
1394 t->of = of;
1395 if (!of)
1396 init_waitqueue_head(&t->event_wait);
1397 t->pending_event = false;
1398 t->aggregator = privileged ? PSI_POLL : PSI_AVGS;
1399
1400 if (privileged) {
1401 mutex_lock(&group->rtpoll_trigger_lock);
1402
1403 list_add(&t->node, &group->rtpoll_triggers);
1404 group->rtpoll_min_period = min(group->rtpoll_min_period,
1405 div_u64(t->win.size, UPDATES_PER_WINDOW));
1406 group->rtpoll_nr_triggers[t->state]++;
1407 group->rtpoll_states |= (1 << t->state);
1408
1409 *need_rtpoll_worker = !rcu_access_pointer(group->rtpoll_task);
1410
1411 mutex_unlock(&group->rtpoll_trigger_lock);
1412 } else {
1413 mutex_lock(&group->avgs_lock);
1414
1415 list_add(&t->node, &group->avg_triggers);
1416 group->avg_nr_triggers[t->state]++;
1417
1418 mutex_unlock(&group->avgs_lock);
1419 }
1420 return t;
1421}
1422
1423void psi_trigger_destroy(struct psi_trigger *t)
1424{
1425 struct psi_group *group;
1426 struct task_struct *task_to_destroy = NULL;
1427
1428 /*
1429 * We do not check psi_disabled since it might have been disabled after
1430 * the trigger got created.
1431 */
1432 if (!t)
1433 return;
1434
1435 group = t->group;
1436 /*
1437 * Wakeup waiters to stop polling and clear the queue to prevent it from
1438 * being accessed later. Can happen if cgroup is deleted from under a
1439 * polling process.
1440 */
1441 if (t->of)
1442 kernfs_notify(t->of->kn);
1443 else
1444 wake_up_interruptible(&t->event_wait);
1445
1446 if (t->aggregator == PSI_AVGS) {
1447 mutex_lock(&group->avgs_lock);
1448 if (!list_empty(&t->node)) {
1449 list_del(&t->node);
1450 group->avg_nr_triggers[t->state]--;
1451 }
1452 mutex_unlock(&group->avgs_lock);
1453 } else {
1454 mutex_lock(&group->rtpoll_trigger_lock);
1455 if (!list_empty(&t->node)) {
1456 struct psi_trigger *tmp;
1457 u64 period = ULLONG_MAX;
1458
1459 list_del(&t->node);
1460 group->rtpoll_nr_triggers[t->state]--;
1461 if (!group->rtpoll_nr_triggers[t->state])
1462 group->rtpoll_states &= ~(1 << t->state);
1463 /*
1464 * Reset min update period for the remaining triggers
1465 * iff the destroying trigger had the min window size.
1466 */
1467 if (group->rtpoll_min_period == div_u64(t->win.size, UPDATES_PER_WINDOW)) {
1468 list_for_each_entry(tmp, &group->rtpoll_triggers, node)
1469 period = min(period, div_u64(tmp->win.size,
1470 UPDATES_PER_WINDOW));
1471 group->rtpoll_min_period = period;
1472 }
1473 /* Destroy rtpoll_task when the last trigger is destroyed */
1474 if (group->rtpoll_states == 0) {
1475 group->rtpoll_until = 0;
1476 task_to_destroy = rcu_dereference_protected(
1477 group->rtpoll_task,
1478 lockdep_is_held(&group->rtpoll_trigger_lock));
1479 rcu_assign_pointer(group->rtpoll_task, NULL);
1480 timer_delete(&group->rtpoll_timer);
1481 }
1482 }
1483 mutex_unlock(&group->rtpoll_trigger_lock);
1484 }
1485
1486 /*
1487 * Stop kthread 'psimon' after releasing rtpoll_trigger_lock to prevent
1488 * a deadlock while waiting for psi_rtpoll_work to acquire
1489 * rtpoll_trigger_lock
1490 */
1491 if (task_to_destroy) {
1492 kthread_stop(task_to_destroy);
1493 atomic_set(&group->rtpoll_scheduled, 0);
1494 }
1495 kfree(t);
1496}
1497
1498__poll_t psi_trigger_poll(void **trigger_ptr,
1499 struct file *file, poll_table *wait)
1500{
1501 __poll_t ret = DEFAULT_POLLMASK;
1502 struct psi_trigger *t;
1503
1504 if (static_branch_likely(&psi_disabled))
1505 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1506
1507 t = smp_load_acquire(trigger_ptr);
1508 if (!t)
1509 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1510
1511 if (t->of)
1512 kernfs_generic_poll(t->of, wait);
1513 else
1514 poll_wait(file, &t->event_wait, wait);
1515
1516 if (cmpxchg(&t->event, 1, 0) == 1)
1517 ret |= EPOLLPRI;
1518
1519 return ret;
1520}
1521
1522#ifdef CONFIG_PROC_FS
1523static int psi_io_show(struct seq_file *m, void *v)
1524{
1525 return psi_show(m, &psi_system, PSI_IO);
1526}
1527
1528static int psi_memory_show(struct seq_file *m, void *v)
1529{
1530 return psi_show(m, &psi_system, PSI_MEM);
1531}
1532
1533static int psi_cpu_show(struct seq_file *m, void *v)
1534{
1535 return psi_show(m, &psi_system, PSI_CPU);
1536}
1537
1538static int psi_io_open(struct inode *inode, struct file *file)
1539{
1540 return single_open(file, psi_io_show, NULL);
1541}
1542
1543static int psi_memory_open(struct inode *inode, struct file *file)
1544{
1545 return single_open(file, psi_memory_show, NULL);
1546}
1547
1548static int psi_cpu_open(struct inode *inode, struct file *file)
1549{
1550 return single_open(file, psi_cpu_show, NULL);
1551}
1552
1553static ssize_t psi_write(struct file *file, const char __user *user_buf,
1554 size_t nbytes, enum psi_res res)
1555{
1556 char buf[32];
1557 size_t buf_size;
1558 struct seq_file *seq;
1559 struct psi_trigger *new;
1560 bool need_rtpoll_worker;
1561 int ret;
1562
1563 if (static_branch_likely(&psi_disabled))
1564 return -EOPNOTSUPP;
1565
1566 if (!nbytes)
1567 return -EINVAL;
1568
1569 buf_size = min(nbytes, sizeof(buf));
1570 if (copy_from_user(buf, user_buf, buf_size))
1571 return -EFAULT;
1572
1573 buf[buf_size - 1] = '\0';
1574
1575 seq = file->private_data;
1576
1577 /* Take seq->lock to protect seq->private from concurrent writes */
1578 mutex_lock(&seq->lock);
1579
1580 /* Allow only one trigger per file descriptor */
1581 if (seq->private) {
1582 mutex_unlock(&seq->lock);
1583 return -EBUSY;
1584 }
1585
1586 new = psi_trigger_create(&psi_system, buf, res, file, NULL,
1587 &need_rtpoll_worker);
1588 if (IS_ERR(new)) {
1589 mutex_unlock(&seq->lock);
1590 return PTR_ERR(new);
1591 }
1592
1593 if (need_rtpoll_worker) {
1594 ret = psi_trigger_create_rtpoll_worker(&psi_system);
1595 if (ret) {
1596 psi_trigger_destroy(new);
1597 mutex_unlock(&seq->lock);
1598 return ret;
1599 }
1600 }
1601
1602 smp_store_release(&seq->private, new);
1603 mutex_unlock(&seq->lock);
1604
1605 return nbytes;
1606}
1607
1608static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1609 size_t nbytes, loff_t *ppos)
1610{
1611 return psi_write(file, user_buf, nbytes, PSI_IO);
1612}
1613
1614static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1615 size_t nbytes, loff_t *ppos)
1616{
1617 return psi_write(file, user_buf, nbytes, PSI_MEM);
1618}
1619
1620static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1621 size_t nbytes, loff_t *ppos)
1622{
1623 return psi_write(file, user_buf, nbytes, PSI_CPU);
1624}
1625
1626static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1627{
1628 struct seq_file *seq = file->private_data;
1629
1630 return psi_trigger_poll(&seq->private, file, wait);
1631}
1632
1633static int psi_fop_release(struct inode *inode, struct file *file)
1634{
1635 struct seq_file *seq = file->private_data;
1636
1637 psi_trigger_destroy(seq->private);
1638 return single_release(inode, file);
1639}
1640
1641static const struct proc_ops psi_io_proc_ops = {
1642 .proc_open = psi_io_open,
1643 .proc_read = seq_read,
1644 .proc_lseek = seq_lseek,
1645 .proc_write = psi_io_write,
1646 .proc_poll = psi_fop_poll,
1647 .proc_release = psi_fop_release,
1648};
1649
1650static const struct proc_ops psi_memory_proc_ops = {
1651 .proc_open = psi_memory_open,
1652 .proc_read = seq_read,
1653 .proc_lseek = seq_lseek,
1654 .proc_write = psi_memory_write,
1655 .proc_poll = psi_fop_poll,
1656 .proc_release = psi_fop_release,
1657};
1658
1659static const struct proc_ops psi_cpu_proc_ops = {
1660 .proc_open = psi_cpu_open,
1661 .proc_read = seq_read,
1662 .proc_lseek = seq_lseek,
1663 .proc_write = psi_cpu_write,
1664 .proc_poll = psi_fop_poll,
1665 .proc_release = psi_fop_release,
1666};
1667
1668#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1669static int psi_irq_show(struct seq_file *m, void *v)
1670{
1671 return psi_show(m, &psi_system, PSI_IRQ);
1672}
1673
1674static int psi_irq_open(struct inode *inode, struct file *file)
1675{
1676 return single_open(file, psi_irq_show, NULL);
1677}
1678
1679static ssize_t psi_irq_write(struct file *file, const char __user *user_buf,
1680 size_t nbytes, loff_t *ppos)
1681{
1682 return psi_write(file, user_buf, nbytes, PSI_IRQ);
1683}
1684
1685static const struct proc_ops psi_irq_proc_ops = {
1686 .proc_open = psi_irq_open,
1687 .proc_read = seq_read,
1688 .proc_lseek = seq_lseek,
1689 .proc_write = psi_irq_write,
1690 .proc_poll = psi_fop_poll,
1691 .proc_release = psi_fop_release,
1692};
1693#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1694
1695static int __init psi_proc_init(void)
1696{
1697 if (psi_enable) {
1698 proc_mkdir("pressure", NULL);
1699 proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops);
1700 proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops);
1701 proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops);
1702#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1703 proc_create("pressure/irq", 0666, NULL, &psi_irq_proc_ops);
1704#endif
1705 }
1706 return 0;
1707}
1708module_init(psi_proc_init);
1709
1710#endif /* CONFIG_PROC_FS */