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 group->poll_min_period = U32_MAX;
194 group->polling_next_update = ULLONG_MAX;
195 init_waitqueue_head(&group->poll_wait);
196 timer_setup(&group->poll_timer, poll_timer_fn, 0);
197 rcu_assign_pointer(group->poll_task, NULL);
200 void __init psi_init(void)
203 static_branch_enable(&psi_disabled);
204 static_branch_disable(&psi_cgroups_enabled);
208 if (!cgroup_psi_enabled())
209 static_branch_disable(&psi_cgroups_enabled);
211 psi_period = jiffies_to_nsecs(PSI_FREQ);
212 group_init(&psi_system);
215 static bool test_state(unsigned int *tasks, enum psi_states state, bool oncpu)
219 return unlikely(tasks[NR_IOWAIT]);
221 return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]);
223 return unlikely(tasks[NR_MEMSTALL]);
225 return unlikely(tasks[NR_MEMSTALL] &&
226 tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]);
228 return unlikely(tasks[NR_RUNNING] > oncpu);
230 return unlikely(tasks[NR_RUNNING] && !oncpu);
232 return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
239 static void get_recent_times(struct psi_group *group, int cpu,
240 enum psi_aggregators aggregator, u32 *times,
241 u32 *pchanged_states)
243 struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
244 u64 now, state_start;
249 *pchanged_states = 0;
251 /* Snapshot a coherent view of the CPU state */
253 seq = read_seqcount_begin(&groupc->seq);
254 now = cpu_clock(cpu);
255 memcpy(times, groupc->times, sizeof(groupc->times));
256 state_mask = groupc->state_mask;
257 state_start = groupc->state_start;
258 } while (read_seqcount_retry(&groupc->seq, seq));
260 /* Calculate state time deltas against the previous snapshot */
261 for (s = 0; s < NR_PSI_STATES; s++) {
264 * In addition to already concluded states, we also
265 * incorporate currently active states on the CPU,
266 * since states may last for many sampling periods.
268 * This way we keep our delta sampling buckets small
269 * (u32) and our reported pressure close to what's
270 * actually happening.
272 if (state_mask & (1 << s))
273 times[s] += now - state_start;
275 delta = times[s] - groupc->times_prev[aggregator][s];
276 groupc->times_prev[aggregator][s] = times[s];
280 *pchanged_states |= (1 << s);
284 static void calc_avgs(unsigned long avg[3], int missed_periods,
285 u64 time, u64 period)
289 /* Fill in zeroes for periods of no activity */
290 if (missed_periods) {
291 avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
292 avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
293 avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
296 /* Sample the most recent active period */
297 pct = div_u64(time * 100, period);
299 avg[0] = calc_load(avg[0], EXP_10s, pct);
300 avg[1] = calc_load(avg[1], EXP_60s, pct);
301 avg[2] = calc_load(avg[2], EXP_300s, pct);
304 static void collect_percpu_times(struct psi_group *group,
305 enum psi_aggregators aggregator,
306 u32 *pchanged_states)
308 u64 deltas[NR_PSI_STATES - 1] = { 0, };
309 unsigned long nonidle_total = 0;
310 u32 changed_states = 0;
315 * Collect the per-cpu time buckets and average them into a
316 * single time sample that is normalized to wallclock time.
318 * For averaging, each CPU is weighted by its non-idle time in
319 * the sampling period. This eliminates artifacts from uneven
320 * loading, or even entirely idle CPUs.
322 for_each_possible_cpu(cpu) {
323 u32 times[NR_PSI_STATES];
325 u32 cpu_changed_states;
327 get_recent_times(group, cpu, aggregator, times,
328 &cpu_changed_states);
329 changed_states |= cpu_changed_states;
331 nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
332 nonidle_total += nonidle;
334 for (s = 0; s < PSI_NONIDLE; s++)
335 deltas[s] += (u64)times[s] * nonidle;
339 * Integrate the sample into the running statistics that are
340 * reported to userspace: the cumulative stall times and the
343 * Pressure percentages are sampled at PSI_FREQ. We might be
344 * called more often when the user polls more frequently than
345 * that; we might be called less often when there is no task
346 * activity, thus no data, and clock ticks are sporadic. The
347 * below handles both.
351 for (s = 0; s < NR_PSI_STATES - 1; s++)
352 group->total[aggregator][s] +=
353 div_u64(deltas[s], max(nonidle_total, 1UL));
356 *pchanged_states = changed_states;
359 static u64 update_averages(struct psi_group *group, u64 now)
361 unsigned long missed_periods = 0;
367 expires = group->avg_next_update;
368 if (now - expires >= psi_period)
369 missed_periods = div_u64(now - expires, psi_period);
372 * The periodic clock tick can get delayed for various
373 * reasons, especially on loaded systems. To avoid clock
374 * drift, we schedule the clock in fixed psi_period intervals.
375 * But the deltas we sample out of the per-cpu buckets above
376 * are based on the actual time elapsing between clock ticks.
378 avg_next_update = expires + ((1 + missed_periods) * psi_period);
379 period = now - (group->avg_last_update + (missed_periods * psi_period));
380 group->avg_last_update = now;
382 for (s = 0; s < NR_PSI_STATES - 1; s++) {
385 sample = group->total[PSI_AVGS][s] - group->avg_total[s];
387 * Due to the lockless sampling of the time buckets,
388 * recorded time deltas can slip into the next period,
389 * which under full pressure can result in samples in
390 * excess of the period length.
392 * We don't want to report non-sensical pressures in
393 * excess of 100%, nor do we want to drop such events
394 * on the floor. Instead we punt any overage into the
395 * future until pressure subsides. By doing this we
396 * don't underreport the occurring pressure curve, we
397 * just report it delayed by one period length.
399 * The error isn't cumulative. As soon as another
400 * delta slips from a period P to P+1, by definition
401 * it frees up its time T in P.
405 group->avg_total[s] += sample;
406 calc_avgs(group->avg[s], missed_periods, sample, period);
409 return avg_next_update;
412 static void psi_avgs_work(struct work_struct *work)
414 struct delayed_work *dwork;
415 struct psi_group *group;
420 dwork = to_delayed_work(work);
421 group = container_of(dwork, struct psi_group, avgs_work);
423 mutex_lock(&group->avgs_lock);
427 collect_percpu_times(group, PSI_AVGS, &changed_states);
428 nonidle = changed_states & (1 << PSI_NONIDLE);
430 * If there is task activity, periodically fold the per-cpu
431 * times and feed samples into the running averages. If things
432 * are idle and there is no data to process, stop the clock.
433 * Once restarted, we'll catch up the running averages in one
434 * go - see calc_avgs() and missed_periods.
436 if (now >= group->avg_next_update)
437 group->avg_next_update = update_averages(group, now);
440 schedule_delayed_work(dwork, nsecs_to_jiffies(
441 group->avg_next_update - now) + 1);
444 mutex_unlock(&group->avgs_lock);
447 /* Trigger tracking window manipulations */
448 static void window_reset(struct psi_window *win, u64 now, u64 value,
451 win->start_time = now;
452 win->start_value = value;
453 win->prev_growth = prev_growth;
457 * PSI growth tracking window update and growth calculation routine.
459 * This approximates a sliding tracking window by interpolating
460 * partially elapsed windows using historical growth data from the
461 * previous intervals. This minimizes memory requirements (by not storing
462 * all the intermediate values in the previous window) and simplifies
463 * the calculations. It works well because PSI signal changes only in
464 * positive direction and over relatively small window sizes the growth
465 * is close to linear.
467 static u64 window_update(struct psi_window *win, u64 now, u64 value)
472 elapsed = now - win->start_time;
473 growth = value - win->start_value;
475 * After each tracking window passes win->start_value and
476 * win->start_time get reset and win->prev_growth stores
477 * the average per-window growth of the previous window.
478 * win->prev_growth is then used to interpolate additional
479 * growth from the previous window assuming it was linear.
481 if (elapsed > win->size)
482 window_reset(win, now, value, growth);
486 remaining = win->size - elapsed;
487 growth += div64_u64(win->prev_growth * remaining, win->size);
493 static void init_triggers(struct psi_group *group, u64 now)
495 struct psi_trigger *t;
497 list_for_each_entry(t, &group->triggers, node)
498 window_reset(&t->win, now,
499 group->total[PSI_POLL][t->state], 0);
500 memcpy(group->polling_total, group->total[PSI_POLL],
501 sizeof(group->polling_total));
502 group->polling_next_update = now + group->poll_min_period;
505 static u64 update_triggers(struct psi_group *group, u64 now)
507 struct psi_trigger *t;
508 bool update_total = false;
509 u64 *total = group->total[PSI_POLL];
512 * On subsequent updates, calculate growth deltas and let
513 * watchers know when their specified thresholds are exceeded.
515 list_for_each_entry(t, &group->triggers, node) {
519 new_stall = group->polling_total[t->state] != total[t->state];
521 /* Check for stall activity or a previous threshold breach */
522 if (!new_stall && !t->pending_event)
525 * Check for new stall activity, as well as deferred
526 * events that occurred in the last window after the
527 * trigger had already fired (we want to ratelimit
528 * events without dropping any).
532 * Multiple triggers might be looking at the same state,
533 * remember to update group->polling_total[] once we've
534 * been through all of them. Also remember to extend the
535 * polling time if we see new stall activity.
539 /* Calculate growth since last update */
540 growth = window_update(&t->win, now, total[t->state]);
541 if (growth < t->threshold)
544 t->pending_event = true;
546 /* Limit event signaling to once per window */
547 if (now < t->last_event_time + t->win.size)
550 /* Generate an event */
551 if (cmpxchg(&t->event, 0, 1) == 0)
552 wake_up_interruptible(&t->event_wait);
553 t->last_event_time = now;
554 /* Reset threshold breach flag once event got generated */
555 t->pending_event = false;
559 memcpy(group->polling_total, total,
560 sizeof(group->polling_total));
562 return now + group->poll_min_period;
565 /* Schedule polling if it's not already scheduled. */
566 static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay)
568 struct task_struct *task;
571 * Do not reschedule if already scheduled.
572 * Possible race with a timer scheduled after this check but before
573 * mod_timer below can be tolerated because group->polling_next_update
574 * will keep updates on schedule.
576 if (timer_pending(&group->poll_timer))
581 task = rcu_dereference(group->poll_task);
583 * kworker might be NULL in case psi_trigger_destroy races with
584 * psi_task_change (hotpath) which can't use locks
587 mod_timer(&group->poll_timer, jiffies + delay);
592 static void psi_poll_work(struct psi_group *group)
597 mutex_lock(&group->trigger_lock);
601 collect_percpu_times(group, PSI_POLL, &changed_states);
603 if (changed_states & group->poll_states) {
604 /* Initialize trigger windows when entering polling mode */
605 if (now > group->polling_until)
606 init_triggers(group, now);
609 * Keep the monitor active for at least the duration of the
610 * minimum tracking window as long as monitor states are
613 group->polling_until = now +
614 group->poll_min_period * UPDATES_PER_WINDOW;
617 if (now > group->polling_until) {
618 group->polling_next_update = ULLONG_MAX;
622 if (now >= group->polling_next_update)
623 group->polling_next_update = update_triggers(group, now);
625 psi_schedule_poll_work(group,
626 nsecs_to_jiffies(group->polling_next_update - now) + 1);
629 mutex_unlock(&group->trigger_lock);
632 static int psi_poll_worker(void *data)
634 struct psi_group *group = (struct psi_group *)data;
636 sched_set_fifo_low(current);
639 wait_event_interruptible(group->poll_wait,
640 atomic_cmpxchg(&group->poll_wakeup, 1, 0) ||
641 kthread_should_stop());
642 if (kthread_should_stop())
645 psi_poll_work(group);
650 static void poll_timer_fn(struct timer_list *t)
652 struct psi_group *group = from_timer(group, t, poll_timer);
654 atomic_set(&group->poll_wakeup, 1);
655 wake_up_interruptible(&group->poll_wait);
658 static void record_times(struct psi_group_cpu *groupc, u64 now)
662 delta = now - groupc->state_start;
663 groupc->state_start = now;
665 if (groupc->state_mask & (1 << PSI_IO_SOME)) {
666 groupc->times[PSI_IO_SOME] += delta;
667 if (groupc->state_mask & (1 << PSI_IO_FULL))
668 groupc->times[PSI_IO_FULL] += delta;
671 if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
672 groupc->times[PSI_MEM_SOME] += delta;
673 if (groupc->state_mask & (1 << PSI_MEM_FULL))
674 groupc->times[PSI_MEM_FULL] += delta;
677 if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
678 groupc->times[PSI_CPU_SOME] += delta;
679 if (groupc->state_mask & (1 << PSI_CPU_FULL))
680 groupc->times[PSI_CPU_FULL] += delta;
683 if (groupc->state_mask & (1 << PSI_NONIDLE))
684 groupc->times[PSI_NONIDLE] += delta;
687 static void psi_group_change(struct psi_group *group, int cpu,
688 unsigned int clear, unsigned int set, u64 now,
691 struct psi_group_cpu *groupc;
696 groupc = per_cpu_ptr(group->pcpu, cpu);
699 * First we assess the aggregate resource states this CPU's
700 * tasks have been in since the last change, and account any
701 * SOME and FULL time these may have resulted in.
703 * Then we update the task counts according to the state
704 * change requested through the @clear and @set bits.
706 write_seqcount_begin(&groupc->seq);
708 record_times(groupc, now);
711 * Start with TSK_ONCPU, which doesn't have a corresponding
712 * task count - it's just a boolean flag directly encoded in
713 * the state mask. Clear, set, or carry the current state if
714 * no changes are requested.
716 if (unlikely(clear & TSK_ONCPU)) {
719 } else if (unlikely(set & TSK_ONCPU)) {
720 state_mask = PSI_ONCPU;
723 state_mask = groupc->state_mask & PSI_ONCPU;
727 * The rest of the state mask is calculated based on the task
728 * counts. Update those first, then construct the mask.
730 for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
733 if (groupc->tasks[t]) {
735 } else if (!psi_bug) {
736 printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n",
737 cpu, t, groupc->tasks[0],
738 groupc->tasks[1], groupc->tasks[2],
739 groupc->tasks[3], clear, set);
744 for (t = 0; set; set &= ~(1 << t), t++)
748 for (s = 0; s < NR_PSI_STATES; s++) {
749 if (test_state(groupc->tasks, s, state_mask & PSI_ONCPU))
750 state_mask |= (1 << s);
754 * Since we care about lost potential, a memstall is FULL
755 * when there are no other working tasks, but also when
756 * the CPU is actively reclaiming and nothing productive
757 * could run even if it were runnable. So when the current
758 * task in a cgroup is in_memstall, the corresponding groupc
759 * on that cpu is in PSI_MEM_FULL state.
761 if (unlikely((state_mask & PSI_ONCPU) && cpu_curr(cpu)->in_memstall))
762 state_mask |= (1 << PSI_MEM_FULL);
764 groupc->state_mask = state_mask;
766 write_seqcount_end(&groupc->seq);
768 if (state_mask & group->poll_states)
769 psi_schedule_poll_work(group, 1);
771 if (wake_clock && !delayed_work_pending(&group->avgs_work))
772 schedule_delayed_work(&group->avgs_work, PSI_FREQ);
775 static inline struct psi_group *task_psi_group(struct task_struct *task)
777 #ifdef CONFIG_CGROUPS
778 if (static_branch_likely(&psi_cgroups_enabled))
779 return cgroup_psi(task_dfl_cgroup(task));
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;
808 psi_flags_change(task, clear, set);
810 now = cpu_clock(cpu);
812 group = task_psi_group(task);
814 psi_group_change(group, cpu, clear, set, now, true);
815 } while ((group = group->parent));
818 void psi_task_switch(struct task_struct *prev, struct task_struct *next,
821 struct psi_group *group, *common = NULL;
822 int cpu = task_cpu(prev);
823 u64 now = cpu_clock(cpu);
826 psi_flags_change(next, 0, TSK_ONCPU);
828 * Set TSK_ONCPU on @next's cgroups. If @next shares any
829 * ancestors with @prev, those will already have @prev's
830 * TSK_ONCPU bit set, and we can stop the iteration there.
832 group = task_psi_group(next);
834 if (per_cpu_ptr(group->pcpu, cpu)->state_mask &
840 psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
841 } while ((group = group->parent));
845 int clear = TSK_ONCPU, set = 0;
846 bool wake_clock = true;
849 * When we're going to sleep, psi_dequeue() lets us
850 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
851 * TSK_IOWAIT here, where we can combine it with
852 * TSK_ONCPU and save walking common ancestors twice.
855 clear |= TSK_RUNNING;
856 if (prev->in_memstall)
857 clear |= TSK_MEMSTALL_RUNNING;
862 * Periodic aggregation shuts off if there is a period of no
863 * task changes, so we wake it back up if necessary. However,
864 * don't do this if the task change is the aggregation worker
865 * itself going to sleep, or we'll ping-pong forever.
867 if (unlikely((prev->flags & PF_WQ_WORKER) &&
868 wq_worker_last_func(prev) == psi_avgs_work))
872 psi_flags_change(prev, clear, set);
874 group = task_psi_group(prev);
878 psi_group_change(group, cpu, clear, set, now, wake_clock);
879 } while ((group = group->parent));
882 * TSK_ONCPU is handled up to the common ancestor. If there are
883 * any other differences between the two tasks (e.g. prev goes
884 * to sleep, or only one task is memstall), finish propagating
885 * those differences all the way up to the root.
887 if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) {
889 for (; group; group = group->parent)
890 psi_group_change(group, cpu, clear, set, now, wake_clock);
895 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
896 void psi_account_irqtime(struct task_struct *task, u32 delta)
898 int cpu = task_cpu(task);
899 struct psi_group *group;
900 struct psi_group_cpu *groupc;
906 now = cpu_clock(cpu);
908 group = task_psi_group(task);
910 groupc = per_cpu_ptr(group->pcpu, cpu);
912 write_seqcount_begin(&groupc->seq);
914 record_times(groupc, now);
915 groupc->times[PSI_IRQ_FULL] += delta;
917 write_seqcount_end(&groupc->seq);
919 if (group->poll_states & (1 << PSI_IRQ_FULL))
920 psi_schedule_poll_work(group, 1);
921 } while ((group = group->parent));
926 * psi_memstall_enter - mark the beginning of a memory stall section
927 * @flags: flags to handle nested sections
929 * Marks the calling task as being stalled due to a lack of memory,
930 * such as waiting for a refault or performing reclaim.
932 void psi_memstall_enter(unsigned long *flags)
937 if (static_branch_likely(&psi_disabled))
940 *flags = current->in_memstall;
944 * in_memstall setting & accounting needs to be atomic wrt
945 * changes to the task's scheduling state, otherwise we can
946 * race with CPU migration.
948 rq = this_rq_lock_irq(&rf);
950 current->in_memstall = 1;
951 psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
953 rq_unlock_irq(rq, &rf);
957 * psi_memstall_leave - mark the end of an memory stall section
958 * @flags: flags to handle nested memdelay sections
960 * Marks the calling task as no longer stalled due to lack of memory.
962 void psi_memstall_leave(unsigned long *flags)
967 if (static_branch_likely(&psi_disabled))
973 * in_memstall clearing & accounting needs to be atomic wrt
974 * changes to the task's scheduling state, otherwise we could
975 * race with CPU migration.
977 rq = this_rq_lock_irq(&rf);
979 current->in_memstall = 0;
980 psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0);
982 rq_unlock_irq(rq, &rf);
985 #ifdef CONFIG_CGROUPS
986 int psi_cgroup_alloc(struct cgroup *cgroup)
988 if (!static_branch_likely(&psi_cgroups_enabled))
991 cgroup->psi = kzalloc(sizeof(struct psi_group), GFP_KERNEL);
995 cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
996 if (!cgroup->psi->pcpu) {
1000 group_init(cgroup->psi);
1001 cgroup->psi->parent = cgroup_psi(cgroup_parent(cgroup));
1005 void psi_cgroup_free(struct cgroup *cgroup)
1007 if (!static_branch_likely(&psi_cgroups_enabled))
1010 cancel_delayed_work_sync(&cgroup->psi->avgs_work);
1011 free_percpu(cgroup->psi->pcpu);
1012 /* All triggers must be removed by now */
1013 WARN_ONCE(cgroup->psi->poll_states, "psi: trigger leak\n");
1018 * cgroup_move_task - move task to a different cgroup
1020 * @to: the target css_set
1022 * Move task to a new cgroup and safely migrate its associated stall
1023 * state between the different groups.
1025 * This function acquires the task's rq lock to lock out concurrent
1026 * changes to the task's scheduling state and - in case the task is
1027 * running - concurrent changes to its stall state.
1029 void cgroup_move_task(struct task_struct *task, struct css_set *to)
1031 unsigned int task_flags;
1035 if (!static_branch_likely(&psi_cgroups_enabled)) {
1037 * Lame to do this here, but the scheduler cannot be locked
1038 * from the outside, so we move cgroups from inside sched/.
1040 rcu_assign_pointer(task->cgroups, to);
1044 rq = task_rq_lock(task, &rf);
1047 * We may race with schedule() dropping the rq lock between
1048 * deactivating prev and switching to next. Because the psi
1049 * updates from the deactivation are deferred to the switch
1050 * callback to save cgroup tree updates, the task's scheduling
1051 * state here is not coherent with its psi state:
1053 * schedule() cgroup_move_task()
1057 * psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1061 * psi_task_change() // old cgroup
1062 * task->cgroups = to
1063 * psi_task_change() // new cgroup
1066 * psi_sched_switch() // does deferred updates in new cgroup
1068 * Don't rely on the scheduling state. Use psi_flags instead.
1070 task_flags = task->psi_flags;
1073 psi_task_change(task, task_flags, 0);
1075 /* See comment above */
1076 rcu_assign_pointer(task->cgroups, to);
1079 psi_task_change(task, 0, task_flags);
1081 task_rq_unlock(rq, task, &rf);
1083 #endif /* CONFIG_CGROUPS */
1085 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1087 bool only_full = false;
1091 if (static_branch_likely(&psi_disabled))
1094 /* Update averages before reporting them */
1095 mutex_lock(&group->avgs_lock);
1096 now = sched_clock();
1097 collect_percpu_times(group, PSI_AVGS, NULL);
1098 if (now >= group->avg_next_update)
1099 group->avg_next_update = update_averages(group, now);
1100 mutex_unlock(&group->avgs_lock);
1102 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1103 only_full = res == PSI_IRQ;
1106 for (full = 0; full < 2 - only_full; full++) {
1107 unsigned long avg[3] = { 0, };
1111 /* CPU FULL is undefined at the system level */
1112 if (!(group == &psi_system && res == PSI_CPU && full)) {
1113 for (w = 0; w < 3; w++)
1114 avg[w] = group->avg[res * 2 + full][w];
1115 total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1119 seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1120 full || only_full ? "full" : "some",
1121 LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1122 LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1123 LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1130 struct psi_trigger *psi_trigger_create(struct psi_group *group,
1131 char *buf, enum psi_res res)
1133 struct psi_trigger *t;
1134 enum psi_states state;
1138 if (static_branch_likely(&psi_disabled))
1139 return ERR_PTR(-EOPNOTSUPP);
1141 if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1142 state = PSI_IO_SOME + res * 2;
1143 else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1144 state = PSI_IO_FULL + res * 2;
1146 return ERR_PTR(-EINVAL);
1148 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1149 if (res == PSI_IRQ && --state != PSI_IRQ_FULL)
1150 return ERR_PTR(-EINVAL);
1153 if (state >= PSI_NONIDLE)
1154 return ERR_PTR(-EINVAL);
1156 if (window_us < WINDOW_MIN_US ||
1157 window_us > WINDOW_MAX_US)
1158 return ERR_PTR(-EINVAL);
1160 /* Check threshold */
1161 if (threshold_us == 0 || threshold_us > window_us)
1162 return ERR_PTR(-EINVAL);
1164 t = kmalloc(sizeof(*t), GFP_KERNEL);
1166 return ERR_PTR(-ENOMEM);
1170 t->threshold = threshold_us * NSEC_PER_USEC;
1171 t->win.size = window_us * NSEC_PER_USEC;
1172 window_reset(&t->win, sched_clock(),
1173 group->total[PSI_POLL][t->state], 0);
1176 t->last_event_time = 0;
1177 init_waitqueue_head(&t->event_wait);
1178 t->pending_event = false;
1180 mutex_lock(&group->trigger_lock);
1182 if (!rcu_access_pointer(group->poll_task)) {
1183 struct task_struct *task;
1185 task = kthread_create(psi_poll_worker, group, "psimon");
1188 mutex_unlock(&group->trigger_lock);
1189 return ERR_CAST(task);
1191 atomic_set(&group->poll_wakeup, 0);
1192 wake_up_process(task);
1193 rcu_assign_pointer(group->poll_task, task);
1196 list_add(&t->node, &group->triggers);
1197 group->poll_min_period = min(group->poll_min_period,
1198 div_u64(t->win.size, UPDATES_PER_WINDOW));
1199 group->nr_triggers[t->state]++;
1200 group->poll_states |= (1 << t->state);
1202 mutex_unlock(&group->trigger_lock);
1207 void psi_trigger_destroy(struct psi_trigger *t)
1209 struct psi_group *group;
1210 struct task_struct *task_to_destroy = NULL;
1213 * We do not check psi_disabled since it might have been disabled after
1214 * the trigger got created.
1221 * Wakeup waiters to stop polling. Can happen if cgroup is deleted
1222 * from under a polling process.
1224 wake_up_interruptible(&t->event_wait);
1226 mutex_lock(&group->trigger_lock);
1228 if (!list_empty(&t->node)) {
1229 struct psi_trigger *tmp;
1230 u64 period = ULLONG_MAX;
1233 group->nr_triggers[t->state]--;
1234 if (!group->nr_triggers[t->state])
1235 group->poll_states &= ~(1 << t->state);
1236 /* reset min update period for the remaining triggers */
1237 list_for_each_entry(tmp, &group->triggers, node)
1238 period = min(period, div_u64(tmp->win.size,
1239 UPDATES_PER_WINDOW));
1240 group->poll_min_period = period;
1241 /* Destroy poll_task when the last trigger is destroyed */
1242 if (group->poll_states == 0) {
1243 group->polling_until = 0;
1244 task_to_destroy = rcu_dereference_protected(
1246 lockdep_is_held(&group->trigger_lock));
1247 rcu_assign_pointer(group->poll_task, NULL);
1248 del_timer(&group->poll_timer);
1252 mutex_unlock(&group->trigger_lock);
1255 * Wait for psi_schedule_poll_work RCU to complete its read-side
1256 * critical section before destroying the trigger and optionally the
1261 * Stop kthread 'psimon' after releasing trigger_lock to prevent a
1262 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1264 if (task_to_destroy) {
1266 * After the RCU grace period has expired, the worker
1267 * can no longer be found through group->poll_task.
1269 kthread_stop(task_to_destroy);
1274 __poll_t psi_trigger_poll(void **trigger_ptr,
1275 struct file *file, poll_table *wait)
1277 __poll_t ret = DEFAULT_POLLMASK;
1278 struct psi_trigger *t;
1280 if (static_branch_likely(&psi_disabled))
1281 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1283 t = smp_load_acquire(trigger_ptr);
1285 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1287 poll_wait(file, &t->event_wait, wait);
1289 if (cmpxchg(&t->event, 1, 0) == 1)
1295 #ifdef CONFIG_PROC_FS
1296 static int psi_io_show(struct seq_file *m, void *v)
1298 return psi_show(m, &psi_system, PSI_IO);
1301 static int psi_memory_show(struct seq_file *m, void *v)
1303 return psi_show(m, &psi_system, PSI_MEM);
1306 static int psi_cpu_show(struct seq_file *m, void *v)
1308 return psi_show(m, &psi_system, PSI_CPU);
1311 static int psi_open(struct file *file, int (*psi_show)(struct seq_file *, void *))
1313 if (file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE))
1316 return single_open(file, psi_show, NULL);
1319 static int psi_io_open(struct inode *inode, struct file *file)
1321 return psi_open(file, psi_io_show);
1324 static int psi_memory_open(struct inode *inode, struct file *file)
1326 return psi_open(file, psi_memory_show);
1329 static int psi_cpu_open(struct inode *inode, struct file *file)
1331 return psi_open(file, psi_cpu_show);
1334 static ssize_t psi_write(struct file *file, const char __user *user_buf,
1335 size_t nbytes, enum psi_res res)
1339 struct seq_file *seq;
1340 struct psi_trigger *new;
1342 if (static_branch_likely(&psi_disabled))
1348 buf_size = min(nbytes, sizeof(buf));
1349 if (copy_from_user(buf, user_buf, buf_size))
1352 buf[buf_size - 1] = '\0';
1354 seq = file->private_data;
1356 /* Take seq->lock to protect seq->private from concurrent writes */
1357 mutex_lock(&seq->lock);
1359 /* Allow only one trigger per file descriptor */
1361 mutex_unlock(&seq->lock);
1365 new = psi_trigger_create(&psi_system, buf, res);
1367 mutex_unlock(&seq->lock);
1368 return PTR_ERR(new);
1371 smp_store_release(&seq->private, new);
1372 mutex_unlock(&seq->lock);
1377 static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1378 size_t nbytes, loff_t *ppos)
1380 return psi_write(file, user_buf, nbytes, PSI_IO);
1383 static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1384 size_t nbytes, loff_t *ppos)
1386 return psi_write(file, user_buf, nbytes, PSI_MEM);
1389 static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1390 size_t nbytes, loff_t *ppos)
1392 return psi_write(file, user_buf, nbytes, PSI_CPU);
1395 static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1397 struct seq_file *seq = file->private_data;
1399 return psi_trigger_poll(&seq->private, file, wait);
1402 static int psi_fop_release(struct inode *inode, struct file *file)
1404 struct seq_file *seq = file->private_data;
1406 psi_trigger_destroy(seq->private);
1407 return single_release(inode, file);
1410 static const struct proc_ops psi_io_proc_ops = {
1411 .proc_open = psi_io_open,
1412 .proc_read = seq_read,
1413 .proc_lseek = seq_lseek,
1414 .proc_write = psi_io_write,
1415 .proc_poll = psi_fop_poll,
1416 .proc_release = psi_fop_release,
1419 static const struct proc_ops psi_memory_proc_ops = {
1420 .proc_open = psi_memory_open,
1421 .proc_read = seq_read,
1422 .proc_lseek = seq_lseek,
1423 .proc_write = psi_memory_write,
1424 .proc_poll = psi_fop_poll,
1425 .proc_release = psi_fop_release,
1428 static const struct proc_ops psi_cpu_proc_ops = {
1429 .proc_open = psi_cpu_open,
1430 .proc_read = seq_read,
1431 .proc_lseek = seq_lseek,
1432 .proc_write = psi_cpu_write,
1433 .proc_poll = psi_fop_poll,
1434 .proc_release = psi_fop_release,
1437 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1438 static int psi_irq_show(struct seq_file *m, void *v)
1440 return psi_show(m, &psi_system, PSI_IRQ);
1443 static int psi_irq_open(struct inode *inode, struct file *file)
1445 return psi_open(file, psi_irq_show);
1448 static ssize_t psi_irq_write(struct file *file, const char __user *user_buf,
1449 size_t nbytes, loff_t *ppos)
1451 return psi_write(file, user_buf, nbytes, PSI_IRQ);
1454 static const struct proc_ops psi_irq_proc_ops = {
1455 .proc_open = psi_irq_open,
1456 .proc_read = seq_read,
1457 .proc_lseek = seq_lseek,
1458 .proc_write = psi_irq_write,
1459 .proc_poll = psi_fop_poll,
1460 .proc_release = psi_fop_release,
1464 static int __init psi_proc_init(void)
1467 proc_mkdir("pressure", NULL);
1468 proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops);
1469 proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops);
1470 proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops);
1471 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1472 proc_create("pressure/irq", 0666, NULL, &psi_irq_proc_ops);
1477 module_init(psi_proc_init);
1479 #endif /* CONFIG_PROC_FS */