1 // SPDX-License-Identifier: GPL-2.0
3 * Pressure stall information for CPU, memory and IO
5 * Copyright (c) 2018 Facebook, Inc.
6 * Author: Johannes Weiner <hannes@cmpxchg.org>
8 * Polling support by Suren Baghdasaryan <surenb@google.com>
9 * Copyright (c) 2018 Google, Inc.
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.
16 * This code aggregates individual task delays into resource pressure
17 * metrics that indicate problems with both workload health and
18 * resource utilization.
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.
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.
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.
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.
38 * SOME = nr_delayed_tasks != 0
39 * FULL = nr_delayed_tasks != 0 && nr_productive_tasks == 0
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.
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.
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:
57 * %SOME = time(SOME) / period
58 * %FULL = time(FULL) / period
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
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%.
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%.
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:
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
93 * For the 257 number crunchers on 256 CPUs, this yields:
95 * threads = min(257, 256)
96 * SOME = min(1 / 256, 1) = 0.4%
97 * FULL = (256 - min(256, 256)) / 256 = 0%
99 * For the 1 out of 4 memory-delayed tasks, this yields:
101 * threads = min(4, 4)
102 * SOME = min(1 / 4, 1) = 25%
103 * FULL = (4 - min(3, 4)) / 4 = 25%
105 * [ Substitute nr_cpus with 1, and you can see that it's a natural
106 * extension of the single-CPU model. ]
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.
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.
119 * For each runqueue, we track:
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)
125 * and then periodically aggregate:
127 * tNONIDLE = sum(tNONIDLE[i])
129 * tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE
130 * tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE
132 * %SOME = tSOME / period
133 * %FULL = tFULL / period
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.
140 static int psi_bug __read_mostly;
142 DEFINE_STATIC_KEY_FALSE(psi_disabled);
143 DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled);
145 #ifdef CONFIG_PSI_DEFAULT_DISABLED
146 static bool psi_enable;
148 static bool psi_enable = true;
150 static int __init setup_psi(char *str)
152 return kstrtobool(str, &psi_enable) == 0;
154 __setup("psi=", setup_psi);
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) */
162 /* PSI trigger definitions */
163 #define WINDOW_MIN_US 500000 /* Min window size is 500ms */
164 #define WINDOW_MAX_US 10000000 /* Max window size is 10s */
165 #define UPDATES_PER_WINDOW 10 /* 10 updates per window */
167 /* Sampling frequency in nanoseconds */
168 static u64 psi_period __read_mostly;
170 /* System-level pressure and stall tracking */
171 static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
172 struct psi_group psi_system = {
173 .pcpu = &system_group_pcpu,
176 static void psi_avgs_work(struct work_struct *work);
178 static void poll_timer_fn(struct timer_list *t);
180 static void group_init(struct psi_group *group)
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 INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
189 mutex_init(&group->avgs_lock);
190 /* Init trigger-related members */
191 mutex_init(&group->trigger_lock);
192 INIT_LIST_HEAD(&group->triggers);
193 memset(group->nr_triggers, 0, sizeof(group->nr_triggers));
194 group->poll_states = 0;
195 group->poll_min_period = U32_MAX;
196 memset(group->polling_total, 0, sizeof(group->polling_total));
197 group->polling_next_update = ULLONG_MAX;
198 group->polling_until = 0;
199 init_waitqueue_head(&group->poll_wait);
200 timer_setup(&group->poll_timer, poll_timer_fn, 0);
201 rcu_assign_pointer(group->poll_task, NULL);
204 void __init psi_init(void)
207 static_branch_enable(&psi_disabled);
211 if (!cgroup_psi_enabled())
212 static_branch_disable(&psi_cgroups_enabled);
214 psi_period = jiffies_to_nsecs(PSI_FREQ);
215 group_init(&psi_system);
218 static bool test_state(unsigned int *tasks, enum psi_states state)
222 return unlikely(tasks[NR_IOWAIT]);
224 return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]);
226 return unlikely(tasks[NR_MEMSTALL]);
228 return unlikely(tasks[NR_MEMSTALL] &&
229 tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]);
231 return unlikely(tasks[NR_RUNNING] > tasks[NR_ONCPU]);
233 return unlikely(tasks[NR_RUNNING] && !tasks[NR_ONCPU]);
235 return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
242 static void get_recent_times(struct psi_group *group, int cpu,
243 enum psi_aggregators aggregator, u32 *times,
244 u32 *pchanged_states)
246 struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
247 u64 now, state_start;
252 *pchanged_states = 0;
254 /* Snapshot a coherent view of the CPU state */
256 seq = read_seqcount_begin(&groupc->seq);
257 now = cpu_clock(cpu);
258 memcpy(times, groupc->times, sizeof(groupc->times));
259 state_mask = groupc->state_mask;
260 state_start = groupc->state_start;
261 } while (read_seqcount_retry(&groupc->seq, seq));
263 /* Calculate state time deltas against the previous snapshot */
264 for (s = 0; s < NR_PSI_STATES; s++) {
267 * In addition to already concluded states, we also
268 * incorporate currently active states on the CPU,
269 * since states may last for many sampling periods.
271 * This way we keep our delta sampling buckets small
272 * (u32) and our reported pressure close to what's
273 * actually happening.
275 if (state_mask & (1 << s))
276 times[s] += now - state_start;
278 delta = times[s] - groupc->times_prev[aggregator][s];
279 groupc->times_prev[aggregator][s] = times[s];
283 *pchanged_states |= (1 << s);
287 static void calc_avgs(unsigned long avg[3], int missed_periods,
288 u64 time, u64 period)
292 /* Fill in zeroes for periods of no activity */
293 if (missed_periods) {
294 avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
295 avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
296 avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
299 /* Sample the most recent active period */
300 pct = div_u64(time * 100, period);
302 avg[0] = calc_load(avg[0], EXP_10s, pct);
303 avg[1] = calc_load(avg[1], EXP_60s, pct);
304 avg[2] = calc_load(avg[2], EXP_300s, pct);
307 static void collect_percpu_times(struct psi_group *group,
308 enum psi_aggregators aggregator,
309 u32 *pchanged_states)
311 u64 deltas[NR_PSI_STATES - 1] = { 0, };
312 unsigned long nonidle_total = 0;
313 u32 changed_states = 0;
318 * Collect the per-cpu time buckets and average them into a
319 * single time sample that is normalized to wallclock time.
321 * For averaging, each CPU is weighted by its non-idle time in
322 * the sampling period. This eliminates artifacts from uneven
323 * loading, or even entirely idle CPUs.
325 for_each_possible_cpu(cpu) {
326 u32 times[NR_PSI_STATES];
328 u32 cpu_changed_states;
330 get_recent_times(group, cpu, aggregator, times,
331 &cpu_changed_states);
332 changed_states |= cpu_changed_states;
334 nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
335 nonidle_total += nonidle;
337 for (s = 0; s < PSI_NONIDLE; s++)
338 deltas[s] += (u64)times[s] * nonidle;
342 * Integrate the sample into the running statistics that are
343 * reported to userspace: the cumulative stall times and the
346 * Pressure percentages are sampled at PSI_FREQ. We might be
347 * called more often when the user polls more frequently than
348 * that; we might be called less often when there is no task
349 * activity, thus no data, and clock ticks are sporadic. The
350 * below handles both.
354 for (s = 0; s < NR_PSI_STATES - 1; s++)
355 group->total[aggregator][s] +=
356 div_u64(deltas[s], max(nonidle_total, 1UL));
359 *pchanged_states = changed_states;
362 static u64 update_averages(struct psi_group *group, u64 now)
364 unsigned long missed_periods = 0;
370 expires = group->avg_next_update;
371 if (now - expires >= psi_period)
372 missed_periods = div_u64(now - expires, psi_period);
375 * The periodic clock tick can get delayed for various
376 * reasons, especially on loaded systems. To avoid clock
377 * drift, we schedule the clock in fixed psi_period intervals.
378 * But the deltas we sample out of the per-cpu buckets above
379 * are based on the actual time elapsing between clock ticks.
381 avg_next_update = expires + ((1 + missed_periods) * psi_period);
382 period = now - (group->avg_last_update + (missed_periods * psi_period));
383 group->avg_last_update = now;
385 for (s = 0; s < NR_PSI_STATES - 1; s++) {
388 sample = group->total[PSI_AVGS][s] - group->avg_total[s];
390 * Due to the lockless sampling of the time buckets,
391 * recorded time deltas can slip into the next period,
392 * which under full pressure can result in samples in
393 * excess of the period length.
395 * We don't want to report non-sensical pressures in
396 * excess of 100%, nor do we want to drop such events
397 * on the floor. Instead we punt any overage into the
398 * future until pressure subsides. By doing this we
399 * don't underreport the occurring pressure curve, we
400 * just report it delayed by one period length.
402 * The error isn't cumulative. As soon as another
403 * delta slips from a period P to P+1, by definition
404 * it frees up its time T in P.
408 group->avg_total[s] += sample;
409 calc_avgs(group->avg[s], missed_periods, sample, period);
412 return avg_next_update;
415 static void psi_avgs_work(struct work_struct *work)
417 struct delayed_work *dwork;
418 struct psi_group *group;
423 dwork = to_delayed_work(work);
424 group = container_of(dwork, struct psi_group, avgs_work);
426 mutex_lock(&group->avgs_lock);
430 collect_percpu_times(group, PSI_AVGS, &changed_states);
431 nonidle = changed_states & (1 << PSI_NONIDLE);
433 * If there is task activity, periodically fold the per-cpu
434 * times and feed samples into the running averages. If things
435 * are idle and there is no data to process, stop the clock.
436 * Once restarted, we'll catch up the running averages in one
437 * go - see calc_avgs() and missed_periods.
439 if (now >= group->avg_next_update)
440 group->avg_next_update = update_averages(group, now);
443 schedule_delayed_work(dwork, nsecs_to_jiffies(
444 group->avg_next_update - now) + 1);
447 mutex_unlock(&group->avgs_lock);
450 /* Trigger tracking window manipulations */
451 static void window_reset(struct psi_window *win, u64 now, u64 value,
454 win->start_time = now;
455 win->start_value = value;
456 win->prev_growth = prev_growth;
460 * PSI growth tracking window update and growth calculation routine.
462 * This approximates a sliding tracking window by interpolating
463 * partially elapsed windows using historical growth data from the
464 * previous intervals. This minimizes memory requirements (by not storing
465 * all the intermediate values in the previous window) and simplifies
466 * the calculations. It works well because PSI signal changes only in
467 * positive direction and over relatively small window sizes the growth
468 * is close to linear.
470 static u64 window_update(struct psi_window *win, u64 now, u64 value)
475 elapsed = now - win->start_time;
476 growth = value - win->start_value;
478 * After each tracking window passes win->start_value and
479 * win->start_time get reset and win->prev_growth stores
480 * the average per-window growth of the previous window.
481 * win->prev_growth is then used to interpolate additional
482 * growth from the previous window assuming it was linear.
484 if (elapsed > win->size)
485 window_reset(win, now, value, growth);
489 remaining = win->size - elapsed;
490 growth += div64_u64(win->prev_growth * remaining, win->size);
496 static void init_triggers(struct psi_group *group, u64 now)
498 struct psi_trigger *t;
500 list_for_each_entry(t, &group->triggers, node)
501 window_reset(&t->win, now,
502 group->total[PSI_POLL][t->state], 0);
503 memcpy(group->polling_total, group->total[PSI_POLL],
504 sizeof(group->polling_total));
505 group->polling_next_update = now + group->poll_min_period;
508 static u64 update_triggers(struct psi_group *group, u64 now)
510 struct psi_trigger *t;
511 bool update_total = false;
512 u64 *total = group->total[PSI_POLL];
515 * On subsequent updates, calculate growth deltas and let
516 * watchers know when their specified thresholds are exceeded.
518 list_for_each_entry(t, &group->triggers, node) {
522 new_stall = group->polling_total[t->state] != total[t->state];
524 /* Check for stall activity or a previous threshold breach */
525 if (!new_stall && !t->pending_event)
528 * Check for new stall activity, as well as deferred
529 * events that occurred in the last window after the
530 * trigger had already fired (we want to ratelimit
531 * events without dropping any).
535 * Multiple triggers might be looking at the same state,
536 * remember to update group->polling_total[] once we've
537 * been through all of them. Also remember to extend the
538 * polling time if we see new stall activity.
542 /* Calculate growth since last update */
543 growth = window_update(&t->win, now, total[t->state]);
544 if (growth < t->threshold)
547 t->pending_event = true;
549 /* Limit event signaling to once per window */
550 if (now < t->last_event_time + t->win.size)
553 /* Generate an event */
554 if (cmpxchg(&t->event, 0, 1) == 0)
555 wake_up_interruptible(&t->event_wait);
556 t->last_event_time = now;
557 /* Reset threshold breach flag once event got generated */
558 t->pending_event = false;
562 memcpy(group->polling_total, total,
563 sizeof(group->polling_total));
565 return now + group->poll_min_period;
568 /* Schedule polling if it's not already scheduled. */
569 static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay)
571 struct task_struct *task;
574 * Do not reschedule if already scheduled.
575 * Possible race with a timer scheduled after this check but before
576 * mod_timer below can be tolerated because group->polling_next_update
577 * will keep updates on schedule.
579 if (timer_pending(&group->poll_timer))
584 task = rcu_dereference(group->poll_task);
586 * kworker might be NULL in case psi_trigger_destroy races with
587 * psi_task_change (hotpath) which can't use locks
590 mod_timer(&group->poll_timer, jiffies + delay);
595 static void psi_poll_work(struct psi_group *group)
600 mutex_lock(&group->trigger_lock);
604 collect_percpu_times(group, PSI_POLL, &changed_states);
606 if (changed_states & group->poll_states) {
607 /* Initialize trigger windows when entering polling mode */
608 if (now > group->polling_until)
609 init_triggers(group, now);
612 * Keep the monitor active for at least the duration of the
613 * minimum tracking window as long as monitor states are
616 group->polling_until = now +
617 group->poll_min_period * UPDATES_PER_WINDOW;
620 if (now > group->polling_until) {
621 group->polling_next_update = ULLONG_MAX;
625 if (now >= group->polling_next_update)
626 group->polling_next_update = update_triggers(group, now);
628 psi_schedule_poll_work(group,
629 nsecs_to_jiffies(group->polling_next_update - now) + 1);
632 mutex_unlock(&group->trigger_lock);
635 static int psi_poll_worker(void *data)
637 struct psi_group *group = (struct psi_group *)data;
639 sched_set_fifo_low(current);
642 wait_event_interruptible(group->poll_wait,
643 atomic_cmpxchg(&group->poll_wakeup, 1, 0) ||
644 kthread_should_stop());
645 if (kthread_should_stop())
648 psi_poll_work(group);
653 static void poll_timer_fn(struct timer_list *t)
655 struct psi_group *group = from_timer(group, t, poll_timer);
657 atomic_set(&group->poll_wakeup, 1);
658 wake_up_interruptible(&group->poll_wait);
661 static void record_times(struct psi_group_cpu *groupc, u64 now)
665 delta = now - groupc->state_start;
666 groupc->state_start = now;
668 if (groupc->state_mask & (1 << PSI_IO_SOME)) {
669 groupc->times[PSI_IO_SOME] += delta;
670 if (groupc->state_mask & (1 << PSI_IO_FULL))
671 groupc->times[PSI_IO_FULL] += delta;
674 if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
675 groupc->times[PSI_MEM_SOME] += delta;
676 if (groupc->state_mask & (1 << PSI_MEM_FULL))
677 groupc->times[PSI_MEM_FULL] += delta;
680 if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
681 groupc->times[PSI_CPU_SOME] += delta;
682 if (groupc->state_mask & (1 << PSI_CPU_FULL))
683 groupc->times[PSI_CPU_FULL] += delta;
686 if (groupc->state_mask & (1 << PSI_NONIDLE))
687 groupc->times[PSI_NONIDLE] += delta;
690 static void psi_group_change(struct psi_group *group, int cpu,
691 unsigned int clear, unsigned int set, u64 now,
694 struct psi_group_cpu *groupc;
699 groupc = per_cpu_ptr(group->pcpu, cpu);
702 * First we assess the aggregate resource states this CPU's
703 * tasks have been in since the last change, and account any
704 * SOME and FULL time these may have resulted in.
706 * Then we update the task counts according to the state
707 * change requested through the @clear and @set bits.
709 write_seqcount_begin(&groupc->seq);
711 record_times(groupc, now);
713 for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
716 if (groupc->tasks[t]) {
718 } else if (!psi_bug) {
719 printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u %u] clear=%x set=%x\n",
720 cpu, t, groupc->tasks[0],
721 groupc->tasks[1], groupc->tasks[2],
722 groupc->tasks[3], groupc->tasks[4],
728 for (t = 0; set; set &= ~(1 << t), t++)
732 /* Calculate state mask representing active states */
733 for (s = 0; s < NR_PSI_STATES; s++) {
734 if (test_state(groupc->tasks, s))
735 state_mask |= (1 << s);
739 * Since we care about lost potential, a memstall is FULL
740 * when there are no other working tasks, but also when
741 * the CPU is actively reclaiming and nothing productive
742 * could run even if it were runnable. So when the current
743 * task in a cgroup is in_memstall, the corresponding groupc
744 * on that cpu is in PSI_MEM_FULL state.
746 if (unlikely(groupc->tasks[NR_ONCPU] && cpu_curr(cpu)->in_memstall))
747 state_mask |= (1 << PSI_MEM_FULL);
749 groupc->state_mask = state_mask;
751 write_seqcount_end(&groupc->seq);
753 if (state_mask & group->poll_states)
754 psi_schedule_poll_work(group, 1);
756 if (wake_clock && !delayed_work_pending(&group->avgs_work))
757 schedule_delayed_work(&group->avgs_work, PSI_FREQ);
760 static struct psi_group *iterate_groups(struct task_struct *task, void **iter)
762 if (*iter == &psi_system)
765 #ifdef CONFIG_CGROUPS
766 if (static_branch_likely(&psi_cgroups_enabled)) {
767 struct cgroup *cgroup = NULL;
770 cgroup = task->cgroups->dfl_cgrp;
772 cgroup = cgroup_parent(*iter);
774 if (cgroup && cgroup_parent(cgroup)) {
776 return cgroup_psi(cgroup);
784 static void psi_flags_change(struct task_struct *task, int clear, int set)
786 if (((task->psi_flags & set) ||
787 (task->psi_flags & clear) != clear) &&
789 printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
790 task->pid, task->comm, task_cpu(task),
791 task->psi_flags, clear, set);
795 task->psi_flags &= ~clear;
796 task->psi_flags |= set;
799 void psi_task_change(struct task_struct *task, int clear, int set)
801 int cpu = task_cpu(task);
802 struct psi_group *group;
803 bool wake_clock = true;
810 psi_flags_change(task, clear, set);
812 now = cpu_clock(cpu);
814 * Periodic aggregation shuts off if there is a period of no
815 * task changes, so we wake it back up if necessary. However,
816 * don't do this if the task change is the aggregation worker
817 * itself going to sleep, or we'll ping-pong forever.
819 if (unlikely((clear & TSK_RUNNING) &&
820 (task->flags & PF_WQ_WORKER) &&
821 wq_worker_last_func(task) == psi_avgs_work))
824 while ((group = iterate_groups(task, &iter)))
825 psi_group_change(group, cpu, clear, set, now, wake_clock);
828 void psi_task_switch(struct task_struct *prev, struct task_struct *next,
831 struct psi_group *group, *common = NULL;
832 int cpu = task_cpu(prev);
834 u64 now = cpu_clock(cpu);
837 bool identical_state;
839 psi_flags_change(next, 0, TSK_ONCPU);
841 * When switching between tasks that have an identical
842 * runtime state, the cgroup that contains both tasks
843 * we reach the first common ancestor. Iterate @next's
844 * ancestors only until we encounter @prev's ONCPU.
846 identical_state = prev->psi_flags == next->psi_flags;
848 while ((group = iterate_groups(next, &iter))) {
849 if (identical_state &&
850 per_cpu_ptr(group->pcpu, cpu)->tasks[NR_ONCPU]) {
855 psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
860 int clear = TSK_ONCPU, set = 0;
863 * When we're going to sleep, psi_dequeue() lets us
864 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
865 * TSK_IOWAIT here, where we can combine it with
866 * TSK_ONCPU and save walking common ancestors twice.
869 clear |= TSK_RUNNING;
870 if (prev->in_memstall)
871 clear |= TSK_MEMSTALL_RUNNING;
876 psi_flags_change(prev, clear, set);
879 while ((group = iterate_groups(prev, &iter)) && group != common)
880 psi_group_change(group, cpu, clear, set, now, true);
883 * TSK_ONCPU is handled up to the common ancestor. If we're tasked
884 * with dequeuing too, finish that for the rest of the hierarchy.
888 for (; group; group = iterate_groups(prev, &iter))
889 psi_group_change(group, cpu, clear, set, now, true);
895 * psi_memstall_enter - mark the beginning of a memory stall section
896 * @flags: flags to handle nested sections
898 * Marks the calling task as being stalled due to a lack of memory,
899 * such as waiting for a refault or performing reclaim.
901 void psi_memstall_enter(unsigned long *flags)
906 if (static_branch_likely(&psi_disabled))
909 *flags = current->in_memstall;
913 * in_memstall setting & accounting needs to be atomic wrt
914 * changes to the task's scheduling state, otherwise we can
915 * race with CPU migration.
917 rq = this_rq_lock_irq(&rf);
919 current->in_memstall = 1;
920 psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
922 rq_unlock_irq(rq, &rf);
926 * psi_memstall_leave - mark the end of an memory stall section
927 * @flags: flags to handle nested memdelay sections
929 * Marks the calling task as no longer stalled due to lack of memory.
931 void psi_memstall_leave(unsigned long *flags)
936 if (static_branch_likely(&psi_disabled))
942 * in_memstall clearing & accounting needs to be atomic wrt
943 * changes to the task's scheduling state, otherwise we could
944 * race with CPU migration.
946 rq = this_rq_lock_irq(&rf);
948 current->in_memstall = 0;
949 psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0);
951 rq_unlock_irq(rq, &rf);
954 #ifdef CONFIG_CGROUPS
955 int psi_cgroup_alloc(struct cgroup *cgroup)
957 if (static_branch_likely(&psi_disabled))
960 cgroup->psi = kmalloc(sizeof(struct psi_group), GFP_KERNEL);
964 cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
965 if (!cgroup->psi->pcpu) {
969 group_init(cgroup->psi);
973 void psi_cgroup_free(struct cgroup *cgroup)
975 if (static_branch_likely(&psi_disabled))
978 cancel_delayed_work_sync(&cgroup->psi->avgs_work);
979 free_percpu(cgroup->psi->pcpu);
980 /* All triggers must be removed by now */
981 WARN_ONCE(cgroup->psi->poll_states, "psi: trigger leak\n");
986 * cgroup_move_task - move task to a different cgroup
988 * @to: the target css_set
990 * Move task to a new cgroup and safely migrate its associated stall
991 * state between the different groups.
993 * This function acquires the task's rq lock to lock out concurrent
994 * changes to the task's scheduling state and - in case the task is
995 * running - concurrent changes to its stall state.
997 void cgroup_move_task(struct task_struct *task, struct css_set *to)
999 unsigned int task_flags;
1003 if (static_branch_likely(&psi_disabled)) {
1005 * Lame to do this here, but the scheduler cannot be locked
1006 * from the outside, so we move cgroups from inside sched/.
1008 rcu_assign_pointer(task->cgroups, to);
1012 rq = task_rq_lock(task, &rf);
1015 * We may race with schedule() dropping the rq lock between
1016 * deactivating prev and switching to next. Because the psi
1017 * updates from the deactivation are deferred to the switch
1018 * callback to save cgroup tree updates, the task's scheduling
1019 * state here is not coherent with its psi state:
1021 * schedule() cgroup_move_task()
1025 * psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1029 * psi_task_change() // old cgroup
1030 * task->cgroups = to
1031 * psi_task_change() // new cgroup
1034 * psi_sched_switch() // does deferred updates in new cgroup
1036 * Don't rely on the scheduling state. Use psi_flags instead.
1038 task_flags = task->psi_flags;
1041 psi_task_change(task, task_flags, 0);
1043 /* See comment above */
1044 rcu_assign_pointer(task->cgroups, to);
1047 psi_task_change(task, 0, task_flags);
1049 task_rq_unlock(rq, task, &rf);
1051 #endif /* CONFIG_CGROUPS */
1053 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1058 if (static_branch_likely(&psi_disabled))
1061 /* Update averages before reporting them */
1062 mutex_lock(&group->avgs_lock);
1063 now = sched_clock();
1064 collect_percpu_times(group, PSI_AVGS, NULL);
1065 if (now >= group->avg_next_update)
1066 group->avg_next_update = update_averages(group, now);
1067 mutex_unlock(&group->avgs_lock);
1069 for (full = 0; full < 2; full++) {
1070 unsigned long avg[3] = { 0, };
1074 /* CPU FULL is undefined at the system level */
1075 if (!(group == &psi_system && res == PSI_CPU && full)) {
1076 for (w = 0; w < 3; w++)
1077 avg[w] = group->avg[res * 2 + full][w];
1078 total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1082 seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1083 full ? "full" : "some",
1084 LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1085 LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1086 LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1093 struct psi_trigger *psi_trigger_create(struct psi_group *group,
1094 char *buf, size_t nbytes, enum psi_res res)
1096 struct psi_trigger *t;
1097 enum psi_states state;
1101 if (static_branch_likely(&psi_disabled))
1102 return ERR_PTR(-EOPNOTSUPP);
1104 if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1105 state = PSI_IO_SOME + res * 2;
1106 else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1107 state = PSI_IO_FULL + res * 2;
1109 return ERR_PTR(-EINVAL);
1111 if (state >= PSI_NONIDLE)
1112 return ERR_PTR(-EINVAL);
1114 if (window_us < WINDOW_MIN_US ||
1115 window_us > WINDOW_MAX_US)
1116 return ERR_PTR(-EINVAL);
1118 /* Check threshold */
1119 if (threshold_us == 0 || threshold_us > window_us)
1120 return ERR_PTR(-EINVAL);
1122 t = kmalloc(sizeof(*t), GFP_KERNEL);
1124 return ERR_PTR(-ENOMEM);
1128 t->threshold = threshold_us * NSEC_PER_USEC;
1129 t->win.size = window_us * NSEC_PER_USEC;
1130 window_reset(&t->win, sched_clock(),
1131 group->total[PSI_POLL][t->state], 0);
1134 t->last_event_time = 0;
1135 init_waitqueue_head(&t->event_wait);
1136 t->pending_event = false;
1138 mutex_lock(&group->trigger_lock);
1140 if (!rcu_access_pointer(group->poll_task)) {
1141 struct task_struct *task;
1143 task = kthread_create(psi_poll_worker, group, "psimon");
1146 mutex_unlock(&group->trigger_lock);
1147 return ERR_CAST(task);
1149 atomic_set(&group->poll_wakeup, 0);
1150 wake_up_process(task);
1151 rcu_assign_pointer(group->poll_task, task);
1154 list_add(&t->node, &group->triggers);
1155 group->poll_min_period = min(group->poll_min_period,
1156 div_u64(t->win.size, UPDATES_PER_WINDOW));
1157 group->nr_triggers[t->state]++;
1158 group->poll_states |= (1 << t->state);
1160 mutex_unlock(&group->trigger_lock);
1165 void psi_trigger_destroy(struct psi_trigger *t)
1167 struct psi_group *group;
1168 struct task_struct *task_to_destroy = NULL;
1171 * We do not check psi_disabled since it might have been disabled after
1172 * the trigger got created.
1179 * Wakeup waiters to stop polling. Can happen if cgroup is deleted
1180 * from under a polling process.
1182 wake_up_interruptible(&t->event_wait);
1184 mutex_lock(&group->trigger_lock);
1186 if (!list_empty(&t->node)) {
1187 struct psi_trigger *tmp;
1188 u64 period = ULLONG_MAX;
1191 group->nr_triggers[t->state]--;
1192 if (!group->nr_triggers[t->state])
1193 group->poll_states &= ~(1 << t->state);
1194 /* reset min update period for the remaining triggers */
1195 list_for_each_entry(tmp, &group->triggers, node)
1196 period = min(period, div_u64(tmp->win.size,
1197 UPDATES_PER_WINDOW));
1198 group->poll_min_period = period;
1199 /* Destroy poll_task when the last trigger is destroyed */
1200 if (group->poll_states == 0) {
1201 group->polling_until = 0;
1202 task_to_destroy = rcu_dereference_protected(
1204 lockdep_is_held(&group->trigger_lock));
1205 rcu_assign_pointer(group->poll_task, NULL);
1206 del_timer(&group->poll_timer);
1210 mutex_unlock(&group->trigger_lock);
1213 * Wait for psi_schedule_poll_work RCU to complete its read-side
1214 * critical section before destroying the trigger and optionally the
1219 * Stop kthread 'psimon' after releasing trigger_lock to prevent a
1220 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1222 if (task_to_destroy) {
1224 * After the RCU grace period has expired, the worker
1225 * can no longer be found through group->poll_task.
1227 kthread_stop(task_to_destroy);
1232 __poll_t psi_trigger_poll(void **trigger_ptr,
1233 struct file *file, poll_table *wait)
1235 __poll_t ret = DEFAULT_POLLMASK;
1236 struct psi_trigger *t;
1238 if (static_branch_likely(&psi_disabled))
1239 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1241 t = smp_load_acquire(trigger_ptr);
1243 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1245 poll_wait(file, &t->event_wait, wait);
1247 if (cmpxchg(&t->event, 1, 0) == 1)
1253 #ifdef CONFIG_PROC_FS
1254 static int psi_io_show(struct seq_file *m, void *v)
1256 return psi_show(m, &psi_system, PSI_IO);
1259 static int psi_memory_show(struct seq_file *m, void *v)
1261 return psi_show(m, &psi_system, PSI_MEM);
1264 static int psi_cpu_show(struct seq_file *m, void *v)
1266 return psi_show(m, &psi_system, PSI_CPU);
1269 static int psi_open(struct file *file, int (*psi_show)(struct seq_file *, void *))
1271 if (file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE))
1274 return single_open(file, psi_show, NULL);
1277 static int psi_io_open(struct inode *inode, struct file *file)
1279 return psi_open(file, psi_io_show);
1282 static int psi_memory_open(struct inode *inode, struct file *file)
1284 return psi_open(file, psi_memory_show);
1287 static int psi_cpu_open(struct inode *inode, struct file *file)
1289 return psi_open(file, psi_cpu_show);
1292 static ssize_t psi_write(struct file *file, const char __user *user_buf,
1293 size_t nbytes, enum psi_res res)
1297 struct seq_file *seq;
1298 struct psi_trigger *new;
1300 if (static_branch_likely(&psi_disabled))
1306 buf_size = min(nbytes, sizeof(buf));
1307 if (copy_from_user(buf, user_buf, buf_size))
1310 buf[buf_size - 1] = '\0';
1312 seq = file->private_data;
1314 /* Take seq->lock to protect seq->private from concurrent writes */
1315 mutex_lock(&seq->lock);
1317 /* Allow only one trigger per file descriptor */
1319 mutex_unlock(&seq->lock);
1323 new = psi_trigger_create(&psi_system, buf, nbytes, res);
1325 mutex_unlock(&seq->lock);
1326 return PTR_ERR(new);
1329 smp_store_release(&seq->private, new);
1330 mutex_unlock(&seq->lock);
1335 static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1336 size_t nbytes, loff_t *ppos)
1338 return psi_write(file, user_buf, nbytes, PSI_IO);
1341 static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1342 size_t nbytes, loff_t *ppos)
1344 return psi_write(file, user_buf, nbytes, PSI_MEM);
1347 static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1348 size_t nbytes, loff_t *ppos)
1350 return psi_write(file, user_buf, nbytes, PSI_CPU);
1353 static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1355 struct seq_file *seq = file->private_data;
1357 return psi_trigger_poll(&seq->private, file, wait);
1360 static int psi_fop_release(struct inode *inode, struct file *file)
1362 struct seq_file *seq = file->private_data;
1364 psi_trigger_destroy(seq->private);
1365 return single_release(inode, file);
1368 static const struct proc_ops psi_io_proc_ops = {
1369 .proc_open = psi_io_open,
1370 .proc_read = seq_read,
1371 .proc_lseek = seq_lseek,
1372 .proc_write = psi_io_write,
1373 .proc_poll = psi_fop_poll,
1374 .proc_release = psi_fop_release,
1377 static const struct proc_ops psi_memory_proc_ops = {
1378 .proc_open = psi_memory_open,
1379 .proc_read = seq_read,
1380 .proc_lseek = seq_lseek,
1381 .proc_write = psi_memory_write,
1382 .proc_poll = psi_fop_poll,
1383 .proc_release = psi_fop_release,
1386 static const struct proc_ops psi_cpu_proc_ops = {
1387 .proc_open = psi_cpu_open,
1388 .proc_read = seq_read,
1389 .proc_lseek = seq_lseek,
1390 .proc_write = psi_cpu_write,
1391 .proc_poll = psi_fop_poll,
1392 .proc_release = psi_fop_release,
1395 static int __init psi_proc_init(void)
1398 proc_mkdir("pressure", NULL);
1399 proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops);
1400 proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops);
1401 proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops);
1405 module_init(psi_proc_init);
1407 #endif /* CONFIG_PROC_FS */