2 * Pressure stall information for CPU, memory and IO
4 * Copyright (c) 2018 Facebook, Inc.
5 * Author: Johannes Weiner <hannes@cmpxchg.org>
7 * Polling support by Suren Baghdasaryan <surenb@google.com>
8 * Copyright (c) 2018 Google, Inc.
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.
15 * This code aggregates individual task delays into resource pressure
16 * metrics that indicate problems with both workload health and
17 * resource utilization.
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.
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.
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.
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.
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.
42 * SOME = nr_delayed_tasks != 0
43 * FULL = nr_delayed_tasks != 0 && nr_running_tasks == 0
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:
50 * %SOME = time(SOME) / period
51 * %FULL = time(FULL) / period
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
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%.
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%.
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:
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
86 * For the 257 number crunchers on 256 CPUs, this yields:
88 * threads = min(257, 256)
89 * SOME = min(1 / 256, 1) = 0.4%
90 * FULL = (256 - min(257, 256)) / 256 = 0%
92 * For the 1 out of 4 memory-delayed tasks, this yields:
95 * SOME = min(1 / 4, 1) = 25%
96 * FULL = (4 - min(3, 4)) / 4 = 25%
98 * [ Substitute nr_cpus with 1, and you can see that it's a natural
99 * extension of the single-CPU model. ]
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.
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.
112 * For each runqueue, we track:
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)
118 * and then periodically aggregate:
120 * tNONIDLE = sum(tNONIDLE[i])
122 * tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE
123 * tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE
125 * %SOME = tSOME / period
126 * %FULL = tFULL / period
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.
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>
148 static int psi_bug __read_mostly;
150 DEFINE_STATIC_KEY_FALSE(psi_disabled);
151 DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled);
153 #ifdef CONFIG_PSI_DEFAULT_DISABLED
154 static bool psi_enable;
156 static bool psi_enable = true;
158 static int __init setup_psi(char *str)
160 return kstrtobool(str, &psi_enable) == 0;
162 __setup("psi=", setup_psi);
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) */
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 */
175 /* Sampling frequency in nanoseconds */
176 static u64 psi_period __read_mostly;
178 /* System-level pressure and stall tracking */
179 static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
180 struct psi_group psi_system = {
181 .pcpu = &system_group_pcpu,
184 static void psi_avgs_work(struct work_struct *work);
186 static void poll_timer_fn(struct timer_list *t);
188 static void group_init(struct psi_group *group)
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);
212 void __init psi_init(void)
215 static_branch_enable(&psi_disabled);
219 if (!cgroup_psi_enabled())
220 static_branch_disable(&psi_cgroups_enabled);
222 psi_period = jiffies_to_nsecs(PSI_FREQ);
223 group_init(&psi_system);
226 static bool test_state(unsigned int *tasks, enum psi_states state)
230 return unlikely(tasks[NR_IOWAIT]);
232 return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]);
234 return unlikely(tasks[NR_MEMSTALL]);
236 return unlikely(tasks[NR_MEMSTALL] && !tasks[NR_RUNNING]);
238 return unlikely(tasks[NR_RUNNING] > tasks[NR_ONCPU]);
240 return unlikely(tasks[NR_RUNNING] && !tasks[NR_ONCPU]);
242 return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
249 static void get_recent_times(struct psi_group *group, int cpu,
250 enum psi_aggregators aggregator, u32 *times,
251 u32 *pchanged_states)
253 struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
254 u64 now, state_start;
259 *pchanged_states = 0;
261 /* Snapshot a coherent view of the CPU state */
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));
270 /* Calculate state time deltas against the previous snapshot */
271 for (s = 0; s < NR_PSI_STATES; s++) {
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.
278 * This way we keep our delta sampling buckets small
279 * (u32) and our reported pressure close to what's
280 * actually happening.
282 if (state_mask & (1 << s))
283 times[s] += now - state_start;
285 delta = times[s] - groupc->times_prev[aggregator][s];
286 groupc->times_prev[aggregator][s] = times[s];
290 *pchanged_states |= (1 << s);
294 static void calc_avgs(unsigned long avg[3], int missed_periods,
295 u64 time, u64 period)
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);
306 /* Sample the most recent active period */
307 pct = div_u64(time * 100, period);
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);
314 static void collect_percpu_times(struct psi_group *group,
315 enum psi_aggregators aggregator,
316 u32 *pchanged_states)
318 u64 deltas[NR_PSI_STATES - 1] = { 0, };
319 unsigned long nonidle_total = 0;
320 u32 changed_states = 0;
325 * Collect the per-cpu time buckets and average them into a
326 * single time sample that is normalized to wallclock time.
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.
332 for_each_possible_cpu(cpu) {
333 u32 times[NR_PSI_STATES];
335 u32 cpu_changed_states;
337 get_recent_times(group, cpu, aggregator, times,
338 &cpu_changed_states);
339 changed_states |= cpu_changed_states;
341 nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
342 nonidle_total += nonidle;
344 for (s = 0; s < PSI_NONIDLE; s++)
345 deltas[s] += (u64)times[s] * nonidle;
349 * Integrate the sample into the running statistics that are
350 * reported to userspace: the cumulative stall times and the
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.
361 for (s = 0; s < NR_PSI_STATES - 1; s++)
362 group->total[aggregator][s] +=
363 div_u64(deltas[s], max(nonidle_total, 1UL));
366 *pchanged_states = changed_states;
369 static u64 update_averages(struct psi_group *group, u64 now)
371 unsigned long missed_periods = 0;
377 expires = group->avg_next_update;
378 if (now - expires >= psi_period)
379 missed_periods = div_u64(now - expires, psi_period);
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.
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;
392 for (s = 0; s < NR_PSI_STATES - 1; s++) {
395 sample = group->total[PSI_AVGS][s] - group->avg_total[s];
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.
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.
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.
415 group->avg_total[s] += sample;
416 calc_avgs(group->avg[s], missed_periods, sample, period);
419 return avg_next_update;
422 static void psi_avgs_work(struct work_struct *work)
424 struct delayed_work *dwork;
425 struct psi_group *group;
430 dwork = to_delayed_work(work);
431 group = container_of(dwork, struct psi_group, avgs_work);
433 mutex_lock(&group->avgs_lock);
437 collect_percpu_times(group, PSI_AVGS, &changed_states);
438 nonidle = changed_states & (1 << PSI_NONIDLE);
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.
446 if (now >= group->avg_next_update)
447 group->avg_next_update = update_averages(group, now);
450 schedule_delayed_work(dwork, nsecs_to_jiffies(
451 group->avg_next_update - now) + 1);
454 mutex_unlock(&group->avgs_lock);
457 /* Trigger tracking window manipulations */
458 static void window_reset(struct psi_window *win, u64 now, u64 value,
461 win->start_time = now;
462 win->start_value = value;
463 win->prev_growth = prev_growth;
467 * PSI growth tracking window update and growth calculation routine.
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.
477 static u64 window_update(struct psi_window *win, u64 now, u64 value)
482 elapsed = now - win->start_time;
483 growth = value - win->start_value;
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.
491 if (elapsed > win->size)
492 window_reset(win, now, value, growth);
496 remaining = win->size - elapsed;
497 growth += div64_u64(win->prev_growth * remaining, win->size);
503 static void init_triggers(struct psi_group *group, u64 now)
505 struct psi_trigger *t;
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;
515 static u64 update_triggers(struct psi_group *group, u64 now)
517 struct psi_trigger *t;
518 bool new_stall = false;
519 u64 *total = group->total[PSI_POLL];
522 * On subsequent updates, calculate growth deltas and let
523 * watchers know when their specified thresholds are exceeded.
525 list_for_each_entry(t, &group->triggers, node) {
528 /* Check for stall activity */
529 if (group->polling_total[t->state] == total[t->state])
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.
540 /* Calculate growth since last update */
541 growth = window_update(&t->win, now, total[t->state]);
542 if (growth < t->threshold)
545 /* Limit event signaling to once per window */
546 if (now < t->last_event_time + t->win.size)
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;
556 memcpy(group->polling_total, total,
557 sizeof(group->polling_total));
559 return now + group->poll_min_period;
562 /* Schedule polling if it's not already scheduled. */
563 static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay)
565 struct task_struct *task;
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.
573 if (timer_pending(&group->poll_timer))
578 task = rcu_dereference(group->poll_task);
580 * kworker might be NULL in case psi_trigger_destroy races with
581 * psi_task_change (hotpath) which can't use locks
584 mod_timer(&group->poll_timer, jiffies + delay);
589 static void psi_poll_work(struct psi_group *group)
594 mutex_lock(&group->trigger_lock);
598 collect_percpu_times(group, PSI_POLL, &changed_states);
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);
606 * Keep the monitor active for at least the duration of the
607 * minimum tracking window as long as monitor states are
610 group->polling_until = now +
611 group->poll_min_period * UPDATES_PER_WINDOW;
614 if (now > group->polling_until) {
615 group->polling_next_update = ULLONG_MAX;
619 if (now >= group->polling_next_update)
620 group->polling_next_update = update_triggers(group, now);
622 psi_schedule_poll_work(group,
623 nsecs_to_jiffies(group->polling_next_update - now) + 1);
626 mutex_unlock(&group->trigger_lock);
629 static int psi_poll_worker(void *data)
631 struct psi_group *group = (struct psi_group *)data;
633 sched_set_fifo_low(current);
636 wait_event_interruptible(group->poll_wait,
637 atomic_cmpxchg(&group->poll_wakeup, 1, 0) ||
638 kthread_should_stop());
639 if (kthread_should_stop())
642 psi_poll_work(group);
647 static void poll_timer_fn(struct timer_list *t)
649 struct psi_group *group = from_timer(group, t, poll_timer);
651 atomic_set(&group->poll_wakeup, 1);
652 wake_up_interruptible(&group->poll_wait);
655 static void record_times(struct psi_group_cpu *groupc, u64 now)
659 delta = now - groupc->state_start;
660 groupc->state_start = now;
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;
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;
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;
680 if (groupc->state_mask & (1 << PSI_NONIDLE))
681 groupc->times[PSI_NONIDLE] += delta;
684 static void psi_group_change(struct psi_group *group, int cpu,
685 unsigned int clear, unsigned int set, u64 now,
688 struct psi_group_cpu *groupc;
693 groupc = per_cpu_ptr(group->pcpu, cpu);
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.
700 * Then we update the task counts according to the state
701 * change requested through the @clear and @set bits.
703 write_seqcount_begin(&groupc->seq);
705 record_times(groupc, now);
707 for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
710 if (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);
721 for (t = 0; set; set &= ~(1 << t), t++)
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);
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.
739 if (unlikely(groupc->tasks[NR_ONCPU] && cpu_curr(cpu)->in_memstall))
740 state_mask |= (1 << PSI_MEM_FULL);
742 groupc->state_mask = state_mask;
744 write_seqcount_end(&groupc->seq);
746 if (state_mask & group->poll_states)
747 psi_schedule_poll_work(group, 1);
749 if (wake_clock && !delayed_work_pending(&group->avgs_work))
750 schedule_delayed_work(&group->avgs_work, PSI_FREQ);
753 static struct psi_group *iterate_groups(struct task_struct *task, void **iter)
755 if (*iter == &psi_system)
758 #ifdef CONFIG_CGROUPS
759 if (static_branch_likely(&psi_cgroups_enabled)) {
760 struct cgroup *cgroup = NULL;
763 cgroup = task->cgroups->dfl_cgrp;
765 cgroup = cgroup_parent(*iter);
767 if (cgroup && cgroup_parent(cgroup)) {
769 return cgroup_psi(cgroup);
777 static void psi_flags_change(struct task_struct *task, int clear, int set)
779 if (((task->psi_flags & set) ||
780 (task->psi_flags & clear) != clear) &&
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);
788 task->psi_flags &= ~clear;
789 task->psi_flags |= set;
792 void psi_task_change(struct task_struct *task, int clear, int set)
794 int cpu = task_cpu(task);
795 struct psi_group *group;
796 bool wake_clock = true;
803 psi_flags_change(task, clear, set);
805 now = cpu_clock(cpu);
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.
812 if (unlikely((clear & TSK_RUNNING) &&
813 (task->flags & PF_WQ_WORKER) &&
814 wq_worker_last_func(task) == psi_avgs_work))
817 while ((group = iterate_groups(task, &iter)))
818 psi_group_change(group, cpu, clear, set, now, wake_clock);
821 void psi_task_switch(struct task_struct *prev, struct task_struct *next,
824 struct psi_group *group, *common = NULL;
825 int cpu = task_cpu(prev);
827 u64 now = cpu_clock(cpu);
830 bool identical_state;
832 psi_flags_change(next, 0, TSK_ONCPU);
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.
840 identical_state = prev->psi_flags == next->psi_flags;
842 while ((group = iterate_groups(next, &iter))) {
843 if (identical_state &&
844 per_cpu_ptr(group->pcpu, cpu)->tasks[NR_ONCPU]) {
849 psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
854 int clear = TSK_ONCPU, set = 0;
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.
862 clear |= TSK_RUNNING;
867 psi_flags_change(prev, clear, set);
870 while ((group = iterate_groups(prev, &iter)) && group != common)
871 psi_group_change(group, cpu, clear, set, now, true);
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.
879 for (; group; group = iterate_groups(prev, &iter))
880 psi_group_change(group, cpu, clear, set, now, true);
886 * psi_memstall_enter - mark the beginning of a memory stall section
887 * @flags: flags to handle nested sections
889 * Marks the calling task as being stalled due to a lack of memory,
890 * such as waiting for a refault or performing reclaim.
892 void psi_memstall_enter(unsigned long *flags)
897 if (static_branch_likely(&psi_disabled))
900 *flags = current->in_memstall;
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.
908 rq = this_rq_lock_irq(&rf);
910 current->in_memstall = 1;
911 psi_task_change(current, 0, TSK_MEMSTALL);
913 rq_unlock_irq(rq, &rf);
917 * psi_memstall_leave - mark the end of an memory stall section
918 * @flags: flags to handle nested memdelay sections
920 * Marks the calling task as no longer stalled due to lack of memory.
922 void psi_memstall_leave(unsigned long *flags)
927 if (static_branch_likely(&psi_disabled))
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.
937 rq = this_rq_lock_irq(&rf);
939 current->in_memstall = 0;
940 psi_task_change(current, TSK_MEMSTALL, 0);
942 rq_unlock_irq(rq, &rf);
945 #ifdef CONFIG_CGROUPS
946 int psi_cgroup_alloc(struct cgroup *cgroup)
948 if (static_branch_likely(&psi_disabled))
951 cgroup->psi.pcpu = alloc_percpu(struct psi_group_cpu);
952 if (!cgroup->psi.pcpu)
954 group_init(&cgroup->psi);
958 void psi_cgroup_free(struct cgroup *cgroup)
960 if (static_branch_likely(&psi_disabled))
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");
970 * cgroup_move_task - move task to a different cgroup
972 * @to: the target css_set
974 * Move task to a new cgroup and safely migrate its associated stall
975 * state between the different groups.
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.
981 void cgroup_move_task(struct task_struct *task, struct css_set *to)
983 unsigned int task_flags;
987 if (static_branch_likely(&psi_disabled)) {
989 * Lame to do this here, but the scheduler cannot be locked
990 * from the outside, so we move cgroups from inside sched/.
992 rcu_assign_pointer(task->cgroups, to);
996 rq = task_rq_lock(task, &rf);
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:
1005 * schedule() cgroup_move_task()
1009 * psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1013 * psi_task_change() // old cgroup
1014 * task->cgroups = to
1015 * psi_task_change() // new cgroup
1018 * psi_sched_switch() // does deferred updates in new cgroup
1020 * Don't rely on the scheduling state. Use psi_flags instead.
1022 task_flags = task->psi_flags;
1025 psi_task_change(task, task_flags, 0);
1027 /* See comment above */
1028 rcu_assign_pointer(task->cgroups, to);
1031 psi_task_change(task, 0, task_flags);
1033 task_rq_unlock(rq, task, &rf);
1035 #endif /* CONFIG_CGROUPS */
1037 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1042 if (static_branch_likely(&psi_disabled))
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);
1053 for (full = 0; full < 2; full++) {
1054 unsigned long avg[3];
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],
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]),
1074 static int psi_io_show(struct seq_file *m, void *v)
1076 return psi_show(m, &psi_system, PSI_IO);
1079 static int psi_memory_show(struct seq_file *m, void *v)
1081 return psi_show(m, &psi_system, PSI_MEM);
1084 static int psi_cpu_show(struct seq_file *m, void *v)
1086 return psi_show(m, &psi_system, PSI_CPU);
1089 static int psi_open(struct file *file, int (*psi_show)(struct seq_file *, void *))
1091 if (file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE))
1094 return single_open(file, psi_show, NULL);
1097 static int psi_io_open(struct inode *inode, struct file *file)
1099 return psi_open(file, psi_io_show);
1102 static int psi_memory_open(struct inode *inode, struct file *file)
1104 return psi_open(file, psi_memory_show);
1107 static int psi_cpu_open(struct inode *inode, struct file *file)
1109 return psi_open(file, psi_cpu_show);
1112 struct psi_trigger *psi_trigger_create(struct psi_group *group,
1113 char *buf, size_t nbytes, enum psi_res res)
1115 struct psi_trigger *t;
1116 enum psi_states state;
1120 if (static_branch_likely(&psi_disabled))
1121 return ERR_PTR(-EOPNOTSUPP);
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;
1128 return ERR_PTR(-EINVAL);
1130 if (state >= PSI_NONIDLE)
1131 return ERR_PTR(-EINVAL);
1133 if (window_us < WINDOW_MIN_US ||
1134 window_us > WINDOW_MAX_US)
1135 return ERR_PTR(-EINVAL);
1137 /* Check threshold */
1138 if (threshold_us == 0 || threshold_us > window_us)
1139 return ERR_PTR(-EINVAL);
1141 t = kmalloc(sizeof(*t), GFP_KERNEL);
1143 return ERR_PTR(-ENOMEM);
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);
1152 t->last_event_time = 0;
1153 init_waitqueue_head(&t->event_wait);
1154 kref_init(&t->refcount);
1156 mutex_lock(&group->trigger_lock);
1158 if (!rcu_access_pointer(group->poll_task)) {
1159 struct task_struct *task;
1161 task = kthread_create(psi_poll_worker, group, "psimon");
1164 mutex_unlock(&group->trigger_lock);
1165 return ERR_CAST(task);
1167 atomic_set(&group->poll_wakeup, 0);
1168 wake_up_process(task);
1169 rcu_assign_pointer(group->poll_task, task);
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);
1178 mutex_unlock(&group->trigger_lock);
1183 static void psi_trigger_destroy(struct kref *ref)
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;
1189 if (static_branch_likely(&psi_disabled))
1193 * Wakeup waiters to stop polling. Can happen if cgroup is deleted
1194 * from under a polling process.
1196 wake_up_interruptible(&t->event_wait);
1198 mutex_lock(&group->trigger_lock);
1200 if (!list_empty(&t->node)) {
1201 struct psi_trigger *tmp;
1202 u64 period = ULLONG_MAX;
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(
1218 lockdep_is_held(&group->trigger_lock));
1219 rcu_assign_pointer(group->poll_task, NULL);
1220 del_timer(&group->poll_timer);
1224 mutex_unlock(&group->trigger_lock);
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
1233 * Stop kthread 'psimon' after releasing trigger_lock to prevent a
1234 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1236 if (task_to_destroy) {
1238 * After the RCU grace period has expired, the worker
1239 * can no longer be found through group->poll_task.
1241 kthread_stop(task_to_destroy);
1246 void psi_trigger_replace(void **trigger_ptr, struct psi_trigger *new)
1248 struct psi_trigger *old = *trigger_ptr;
1250 if (static_branch_likely(&psi_disabled))
1253 rcu_assign_pointer(*trigger_ptr, new);
1255 kref_put(&old->refcount, psi_trigger_destroy);
1258 __poll_t psi_trigger_poll(void **trigger_ptr,
1259 struct file *file, poll_table *wait)
1261 __poll_t ret = DEFAULT_POLLMASK;
1262 struct psi_trigger *t;
1264 if (static_branch_likely(&psi_disabled))
1265 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1269 t = rcu_dereference(*(void __rcu __force **)trigger_ptr);
1272 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1274 kref_get(&t->refcount);
1278 poll_wait(file, &t->event_wait, wait);
1280 if (cmpxchg(&t->event, 1, 0) == 1)
1283 kref_put(&t->refcount, psi_trigger_destroy);
1288 static ssize_t psi_write(struct file *file, const char __user *user_buf,
1289 size_t nbytes, enum psi_res res)
1293 struct seq_file *seq;
1294 struct psi_trigger *new;
1296 if (static_branch_likely(&psi_disabled))
1302 buf_size = min(nbytes, sizeof(buf));
1303 if (copy_from_user(buf, user_buf, buf_size))
1306 buf[buf_size - 1] = '\0';
1308 new = psi_trigger_create(&psi_system, buf, nbytes, res);
1310 return PTR_ERR(new);
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);
1321 static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1322 size_t nbytes, loff_t *ppos)
1324 return psi_write(file, user_buf, nbytes, PSI_IO);
1327 static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1328 size_t nbytes, loff_t *ppos)
1330 return psi_write(file, user_buf, nbytes, PSI_MEM);
1333 static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1334 size_t nbytes, loff_t *ppos)
1336 return psi_write(file, user_buf, nbytes, PSI_CPU);
1339 static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1341 struct seq_file *seq = file->private_data;
1343 return psi_trigger_poll(&seq->private, file, wait);
1346 static int psi_fop_release(struct inode *inode, struct file *file)
1348 struct seq_file *seq = file->private_data;
1350 psi_trigger_replace(&seq->private, NULL);
1351 return single_release(inode, file);
1354 static 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,
1363 static 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,
1372 static 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,
1381 static int __init psi_proc_init(void)
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);
1391 module_init(psi_proc_init);