1	// SPDX-License-Identifier: GPL-2.0
2	/*
3	* Pressure stall information for CPU, memory and IO
4	*
5	* Copyright (c) 2018 Facebook, Inc.
6	* Author: Johannes Weiner <hannes@cmpxchg.org>
7	*
8	* Polling support by Suren Baghdasaryan <surenb@google.com>
9	* Copyright (c) 2018 Google, Inc.
10	*
11	* When CPU, memory and IO are contended, tasks experience delays that
12	* reduce throughput and introduce latencies into the workload. Memory
13	* and IO contention, in addition, can cause a full loss of forward
14	* progress in which the CPU goes idle.
15	*
16	* This code aggregates individual task delays into resource pressure
17	* metrics that indicate problems with both workload health and
18	* resource utilization.
19	*
20	* Model
21	*
22	* The time in which a task can execute on a CPU is our baseline for
23	* productivity. Pressure expresses the amount of time in which this
24	* potential cannot be realized due to resource contention.
25	*
26	* This concept of productivity has two components: the workload and
27	* the CPU. To measure the impact of pressure on both, we define two
28	* contention states for a resource: SOME and FULL.
29	*
30	* In the SOME state of a given resource, one or more tasks are
31	* delayed on that resource. This affects the workload's ability to
32	* perform work, but the CPU may still be executing other tasks.
33	*
34	* In the FULL state of a given resource, all non-idle tasks are
35	* delayed on that resource such that nobody is advancing and the CPU
36	* goes idle. This leaves both workload and CPU unproductive.
37	*
38	* SOME = nr_delayed_tasks != 0
39	* FULL = nr_delayed_tasks != 0 && nr_productive_tasks == 0
40	*
41	* What it means for a task to be productive is defined differently
42	* for each resource. For IO, productive means a running task. For
43	* memory, productive means a running task that isn't a reclaimer. For
44	* CPU, productive means an on-CPU task.
45	*
46	* Naturally, the FULL state doesn't exist for the CPU resource at the
47	* system level, but exist at the cgroup level. At the cgroup level,
48	* FULL means all non-idle tasks in the cgroup are delayed on the CPU
49	* resource which is being used by others outside of the cgroup or
50	* throttled by the cgroup cpu.max configuration.
51	*
52	* The percentage of wall clock time spent in those compound stall
53	* states gives pressure numbers between 0 and 100 for each resource,
54	* where the SOME percentage indicates workload slowdowns and the FULL
55	* percentage indicates reduced CPU utilization:
56	*
57	* %SOME = time(SOME) / period
58	* %FULL = time(FULL) / period
59	*
60	* Multiple CPUs
61	*
62	* The more tasks and available CPUs there are, the more work can be
63	* performed concurrently. This means that the potential that can go
64	* unrealized due to resource contention *also* scales with non-idle
65	* tasks and CPUs.
66	*
67	* Consider a scenario where 257 number crunching tasks are trying to
68	* run concurrently on 256 CPUs. If we simply aggregated the task
69	* states, we would have to conclude a CPU SOME pressure number of
70	* 100%, since *somebody* is waiting on a runqueue at all
71	* times. However, that is clearly not the amount of contention the
72	* workload is experiencing: only one out of 256 possible execution
73	* threads will be contended at any given time, or about 0.4%.
74	*
75	* Conversely, consider a scenario of 4 tasks and 4 CPUs where at any
76	* given time *one* of the tasks is delayed due to a lack of memory.
77	* Again, looking purely at the task state would yield a memory FULL
78	* pressure number of 0%, since *somebody* is always making forward
79	* progress. But again this wouldn't capture the amount of execution
80	* potential lost, which is 1 out of 4 CPUs, or 25%.
81	*
82	* To calculate wasted potential (pressure) with multiple processors,
83	* we have to base our calculation on the number of non-idle tasks in
84	* conjunction with the number of available CPUs, which is the number
85	* of potential execution threads. SOME becomes then the proportion of
86	* delayed tasks to possible threads, and FULL is the share of possible
87	* threads that are unproductive due to delays:
88	*
89	* threads = min(nr_nonidle_tasks, nr_cpus)
90	* SOME = min(nr_delayed_tasks / threads, 1)
91	* FULL = (threads - min(nr_productive_tasks, threads)) / threads
92	*
93	* For the 257 number crunchers on 256 CPUs, this yields:
94	*
95	* threads = min(257, 256)
96	* SOME = min(1 / 256, 1) = 0.4%
97	* FULL = (256 - min(256, 256)) / 256 = 0%
98	*
99	* For the 1 out of 4 memory-delayed tasks, this yields:
100	*
101	* threads = min(4, 4)
102	* SOME = min(1 / 4, 1) = 25%
103	* FULL = (4 - min(3, 4)) / 4 = 25%
104	*
105	* [ Substitute nr_cpus with 1, and you can see that it's a natural
106	* extension of the single-CPU model. ]
107	*
108	* Implementation
109	*
110	* To assess the precise time spent in each such state, we would have
111	* to freeze the system on task changes and start/stop the state
112	* clocks accordingly. Obviously that doesn't scale in practice.
113	*
114	* Because the scheduler aims to distribute the compute load evenly
115	* among the available CPUs, we can track task state locally to each
116	* CPU and, at much lower frequency, extrapolate the global state for
117	* the cumulative stall times and the running averages.
118	*
119	* For each runqueue, we track:
120	*
121	* tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0)
122	* tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_productive_tasks[cpu])
123	* tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0)
124	*
125	* and then periodically aggregate:
126	*
127	* tNONIDLE = sum(tNONIDLE[i])
128	*
129	* tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE
130	* tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE
131	*
132	* %SOME = tSOME / period
133	* %FULL = tFULL / period
134	*
135	* This gives us an approximation of pressure that is practical
136	* cost-wise, yet way more sensitive and accurate than periodic
137	* sampling of the aggregate task states would be.
138	*/
139
140	static int psi_bug __read_mostly;
141
142	DEFINE_STATIC_KEY_FALSE(psi_disabled);
143	static DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled);
144
145	#ifdef CONFIG_PSI_DEFAULT_DISABLED
146	static bool psi_enable;
147	#else
148	static bool psi_enable = true;
149	#endif
150	static int __init setup_psi(char *str)
151	{
152	return kstrtobool(s: str, res: &psi_enable) == 0;
153	}
154	__setup("psi=", setup_psi);
155
156	/* Running averages - we need to be higher-res than loadavg */
157	#define PSI_FREQ (2*HZ+1) /* 2 sec intervals */
158	#define EXP_10s 1677 /* 1/exp(2s/10s) as fixed-point */
159	#define EXP_60s 1981 /* 1/exp(2s/60s) */
160	#define EXP_300s 2034 /* 1/exp(2s/300s) */
161
162	/* PSI trigger definitions */
163	#define WINDOW_MAX_US 10000000 /* Max window size is 10s */
164	#define UPDATES_PER_WINDOW 10 /* 10 updates per window */
165
166	/* Sampling frequency in nanoseconds */
167	static u64 psi_period __read_mostly;
168
169	/* System-level pressure and stall tracking */
170	static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
171	struct psi_group psi_system = {
172	.pcpu = &system_group_pcpu,
173	};
174
175	static void psi_avgs_work(struct work_struct *work);
176
177	static void poll_timer_fn(struct timer_list *t);
178
179	static void group_init(struct psi_group *group)
180	{
181	int cpu;
182
183	group->enabled = true;
184	for_each_possible_cpu(cpu)
185	seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);
186	group->avg_last_update = sched_clock();
187	group->avg_next_update = group->avg_last_update + psi_period;
188	mutex_init(&group->avgs_lock);
189
190	/* Init avg trigger-related members */
191	INIT_LIST_HEAD(list: &group->avg_triggers);
192	memset(group->avg_nr_triggers, 0, sizeof(group->avg_nr_triggers));
193	INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
194
195	/* Init rtpoll trigger-related members */
196	atomic_set(v: &group->rtpoll_scheduled, i: 0);
197	mutex_init(&group->rtpoll_trigger_lock);
198	INIT_LIST_HEAD(list: &group->rtpoll_triggers);
199	group->rtpoll_min_period = U32_MAX;
200	group->rtpoll_next_update = ULLONG_MAX;
201	init_waitqueue_head(&group->rtpoll_wait);
202	timer_setup(&group->rtpoll_timer, poll_timer_fn, 0);
203	rcu_assign_pointer(group->rtpoll_task, NULL);
204	}
205
206	void __init psi_init(void)
207	{
208	if (!psi_enable) {
209	static_branch_enable(&psi_disabled);
210	static_branch_disable(&psi_cgroups_enabled);
211	return;
212	}
213
214	if (!cgroup_psi_enabled())
215	static_branch_disable(&psi_cgroups_enabled);
216
217	psi_period = jiffies_to_nsecs(PSI_FREQ);
218	group_init(group: &psi_system);
219	}
220
221	static u32 test_states(unsigned int *tasks, u32 state_mask)
222	{
223	const bool oncpu = state_mask & PSI_ONCPU;
224
225	if (tasks[NR_IOWAIT]) {
226	state_mask |= BIT(PSI_IO_SOME);
227	if (!tasks[NR_RUNNING])
228	state_mask |= BIT(PSI_IO_FULL);
229	}
230
231	if (tasks[NR_MEMSTALL]) {
232	state_mask |= BIT(PSI_MEM_SOME);
233	if (tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING])
234	state_mask |= BIT(PSI_MEM_FULL);
235	}
236
237	if (tasks[NR_RUNNING] > oncpu)
238	state_mask |= BIT(PSI_CPU_SOME);
239
240	if (tasks[NR_RUNNING] && !oncpu)
241	state_mask |= BIT(PSI_CPU_FULL);
242
243	if (tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] || tasks[NR_RUNNING])
244	state_mask |= BIT(PSI_NONIDLE);
245
246	return state_mask;
247	}
248
249	static void get_recent_times(struct psi_group *group, int cpu,
250	enum psi_aggregators aggregator, u32 *times,
251	u32 *pchanged_states)
252	{
253	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
254	int current_cpu = raw_smp_processor_id();
255	unsigned int tasks[NR_PSI_TASK_COUNTS];
256	u64 now, state_start;
257	enum psi_states s;
258	unsigned int seq;
259	u32 state_mask;
260
261	*pchanged_states = 0;
262
263	/* Snapshot a coherent view of the CPU state */
264	do {
265	seq = read_seqcount_begin(&groupc->seq);
266	now = cpu_clock(cpu);
267	memcpy(times, groupc->times, sizeof(groupc->times));
268	state_mask = groupc->state_mask;
269	state_start = groupc->state_start;
270	if (cpu == current_cpu)
271	memcpy(tasks, groupc->tasks, sizeof(groupc->tasks));
272	} while (read_seqcount_retry(&groupc->seq, seq));
273
274	/* Calculate state time deltas against the previous snapshot */
275	for (s = 0; s < NR_PSI_STATES; s++) {
276	u32 delta;
277	/*
278	* In addition to already concluded states, we also
279	* incorporate currently active states on the CPU,
280	* since states may last for many sampling periods.
281	*
282	* This way we keep our delta sampling buckets small
283	* (u32) and our reported pressure close to what's
284	* actually happening.
285	*/
286	if (state_mask & (1 << s))
287	times[s] += now - state_start;
288
289	delta = times[s] - groupc->times_prev[aggregator][s];
290	groupc->times_prev[aggregator][s] = times[s];
291
292	times[s] = delta;
293	if (delta)
294	*pchanged_states |= (1 << s);
295	}
296
297	/*
298	* When collect_percpu_times() from the avgs_work, we don't want to
299	* re-arm avgs_work when all CPUs are IDLE. But the current CPU running
300	* this avgs_work is never IDLE, cause avgs_work can't be shut off.
301	* So for the current CPU, we need to re-arm avgs_work only when
302	* (NR_RUNNING > 1 || NR_IOWAIT > 0 || NR_MEMSTALL > 0), for other CPUs
303	* we can just check PSI_NONIDLE delta.
304	*/
305	if (current_work() == &group->avgs_work.work) {
306	bool reschedule;
307
308	if (cpu == current_cpu)
309	reschedule = tasks[NR_RUNNING] +
310	tasks[NR_IOWAIT] +
311	tasks[NR_MEMSTALL] > 1;
312	else
313	reschedule = *pchanged_states & (1 << PSI_NONIDLE);
314
315	if (reschedule)
316	*pchanged_states |= PSI_STATE_RESCHEDULE;
317	}
318	}
319
320	static void calc_avgs(unsigned long avg[3], int missed_periods,
321	u64 time, u64 period)
322	{
323	unsigned long pct;
324
325	/* Fill in zeroes for periods of no activity */
326	if (missed_periods) {
327	avg[0] = calc_load_n(load: avg[0], EXP_10s, active: 0, n: missed_periods);
328	avg[1] = calc_load_n(load: avg[1], EXP_60s, active: 0, n: missed_periods);
329	avg[2] = calc_load_n(load: avg[2], EXP_300s, active: 0, n: missed_periods);
330	}
331
332	/* Sample the most recent active period */
333	pct = div_u64(dividend: time * 100, divisor: period);
334	pct *= FIXED_1;
335	avg[0] = calc_load(load: avg[0], EXP_10s, active: pct);
336	avg[1] = calc_load(load: avg[1], EXP_60s, active: pct);
337	avg[2] = calc_load(load: avg[2], EXP_300s, active: pct);
338	}
339
340	static void collect_percpu_times(struct psi_group *group,
341	enum psi_aggregators aggregator,
342	u32 *pchanged_states)
343	{
344	u64 deltas[NR_PSI_STATES - 1] = { 0, };
345	unsigned long nonidle_total = 0;
346	u32 changed_states = 0;
347	int cpu;
348	int s;
349
350	/*
351	* Collect the per-cpu time buckets and average them into a
352	* single time sample that is normalized to wall clock time.
353	*
354	* For averaging, each CPU is weighted by its non-idle time in
355	* the sampling period. This eliminates artifacts from uneven
356	* loading, or even entirely idle CPUs.
357	*/
358	for_each_possible_cpu(cpu) {
359	u32 times[NR_PSI_STATES];
360	u32 nonidle;
361	u32 cpu_changed_states;
362
363	get_recent_times(group, cpu, aggregator, times,
364	pchanged_states: &cpu_changed_states);
365	changed_states |= cpu_changed_states;
366
367	nonidle = nsecs_to_jiffies(n: times[PSI_NONIDLE]);
368	nonidle_total += nonidle;
369
370	for (s = 0; s < PSI_NONIDLE; s++)
371	deltas[s] += (u64)times[s] * nonidle;
372	}
373
374	/*
375	* Integrate the sample into the running statistics that are
376	* reported to userspace: the cumulative stall times and the
377	* decaying averages.
378	*
379	* Pressure percentages are sampled at PSI_FREQ. We might be
380	* called more often when the user polls more frequently than
381	* that; we might be called less often when there is no task
382	* activity, thus no data, and clock ticks are sporadic. The
383	* below handles both.
384	*/
385
386	/* total= */
387	for (s = 0; s < NR_PSI_STATES - 1; s++)
388	group->total[aggregator][s] +=
389	div_u64(dividend: deltas[s], max(nonidle_total, 1UL));
390
391	if (pchanged_states)
392	*pchanged_states = changed_states;
393	}
394
395	/* Trigger tracking window manipulations */
396	static void window_reset(struct psi_window *win, u64 now, u64 value,
397	u64 prev_growth)
398	{
399	win->start_time = now;
400	win->start_value = value;
401	win->prev_growth = prev_growth;
402	}
403
404	/*
405	* PSI growth tracking window update and growth calculation routine.
406	*
407	* This approximates a sliding tracking window by interpolating
408	* partially elapsed windows using historical growth data from the
409	* previous intervals. This minimizes memory requirements (by not storing
410	* all the intermediate values in the previous window) and simplifies
411	* the calculations. It works well because PSI signal changes only in
412	* positive direction and over relatively small window sizes the growth
413	* is close to linear.
414	*/
415	static u64 window_update(struct psi_window *win, u64 now, u64 value)
416	{
417	u64 elapsed;
418	u64 growth;
419
420	elapsed = now - win->start_time;
421	growth = value - win->start_value;
422	/*
423	* After each tracking window passes win->start_value and
424	* win->start_time get reset and win->prev_growth stores
425	* the average per-window growth of the previous window.
426	* win->prev_growth is then used to interpolate additional
427	* growth from the previous window assuming it was linear.
428	*/
429	if (elapsed > win->size)
430	window_reset(win, now, value, prev_growth: growth);
431	else {
432	u32 remaining;
433
434	remaining = win->size - elapsed;
435	growth += div64_u64(dividend: win->prev_growth * remaining, divisor: win->size);
436	}
437
438	return growth;
439	}
440
441	static void update_triggers(struct psi_group *group, u64 now,
442	enum psi_aggregators aggregator)
443	{
444	struct psi_trigger *t;
445	u64 *total = group->total[aggregator];
446	struct list_head *triggers;
447	u64 *aggregator_total;
448
449	if (aggregator == PSI_AVGS) {
450	triggers = &group->avg_triggers;
451	aggregator_total = group->avg_total;
452	} else {
453	triggers = &group->rtpoll_triggers;
454	aggregator_total = group->rtpoll_total;
455	}
456
457	/*
458	* On subsequent updates, calculate growth deltas and let
459	* watchers know when their specified thresholds are exceeded.
460	*/
461	list_for_each_entry(t, triggers, node) {
462	u64 growth;
463	bool new_stall;
464
465	new_stall = aggregator_total[t->state] != total[t->state];
466
467	/* Check for stall activity or a previous threshold breach */
468	if (!new_stall && !t->pending_event)
469	continue;
470	/*
471	* Check for new stall activity, as well as deferred
472	* events that occurred in the last window after the
473	* trigger had already fired (we want to ratelimit
474	* events without dropping any).
475	*/
476	if (new_stall) {
477	/* Calculate growth since last update */
478	growth = window_update(win: &t->win, now, value: total[t->state]);
479	if (!t->pending_event) {
480	if (growth < t->threshold)
481	continue;
482
483	t->pending_event = true;
484	}
485	}
486	/* Limit event signaling to once per window */
487	if (now < t->last_event_time + t->win.size)
488	continue;
489
490	/* Generate an event */
491	if (cmpxchg(&t->event, 0, 1) == 0) {
492	if (t->of)
493	kernfs_notify(kn: t->of->kn);
494	else
495	wake_up_interruptible(&t->event_wait);
496	}
497	t->last_event_time = now;
498	/* Reset threshold breach flag once event got generated */
499	t->pending_event = false;
500	}
501	}
502
503	static u64 update_averages(struct psi_group *group, u64 now)
504	{
505	unsigned long missed_periods = 0;
506	u64 expires, period;
507	u64 avg_next_update;
508	int s;
509
510	/* avgX= */
511	expires = group->avg_next_update;
512	if (now - expires >= psi_period)
513	missed_periods = div_u64(dividend: now - expires, divisor: psi_period);
514
515	/*
516	* The periodic clock tick can get delayed for various
517	* reasons, especially on loaded systems. To avoid clock
518	* drift, we schedule the clock in fixed psi_period intervals.
519	* But the deltas we sample out of the per-cpu buckets above
520	* are based on the actual time elapsing between clock ticks.
521	*/
522	avg_next_update = expires + ((1 + missed_periods) * psi_period);
523	period = now - (group->avg_last_update + (missed_periods * psi_period));
524	group->avg_last_update = now;
525
526	for (s = 0; s < NR_PSI_STATES - 1; s++) {
527	u32 sample;
528
529	sample = group->total[PSI_AVGS][s] - group->avg_total[s];
530	/*
531	* Due to the lockless sampling of the time buckets,
532	* recorded time deltas can slip into the next period,
533	* which under full pressure can result in samples in
534	* excess of the period length.
535	*
536	* We don't want to report non-sensical pressures in
537	* excess of 100%, nor do we want to drop such events
538	* on the floor. Instead we punt any overage into the
539	* future until pressure subsides. By doing this we
540	* don't underreport the occurring pressure curve, we
541	* just report it delayed by one period length.
542	*
543	* The error isn't cumulative. As soon as another
544	* delta slips from a period P to P+1, by definition
545	* it frees up its time T in P.
546	*/
547	if (sample > period)
548	sample = period;
549	group->avg_total[s] += sample;
550	calc_avgs(avg: group->avg[s], missed_periods, time: sample, period);
551	}
552
553	return avg_next_update;
554	}
555
556	static void psi_avgs_work(struct work_struct *work)
557	{
558	struct delayed_work *dwork;
559	struct psi_group *group;
560	u32 changed_states;
561	u64 now;
562
563	dwork = to_delayed_work(work);
564	group = container_of(dwork, struct psi_group, avgs_work);
565
566	mutex_lock(&group->avgs_lock);
567
568	now = sched_clock();
569
570	collect_percpu_times(group, aggregator: PSI_AVGS, pchanged_states: &changed_states);
571	/*
572	* If there is task activity, periodically fold the per-cpu
573	* times and feed samples into the running averages. If things
574	* are idle and there is no data to process, stop the clock.
575	* Once restarted, we'll catch up the running averages in one
576	* go - see calc_avgs() and missed_periods.
577	*/
578	if (now >= group->avg_next_update) {
579	update_triggers(group, now, aggregator: PSI_AVGS);
580	group->avg_next_update = update_averages(group, now);
581	}
582
583	if (changed_states & PSI_STATE_RESCHEDULE) {
584	schedule_delayed_work(dwork, delay: nsecs_to_jiffies(
585	n: group->avg_next_update - now) + 1);
586	}
587
588	mutex_unlock(lock: &group->avgs_lock);
589	}
590
591	static void init_rtpoll_triggers(struct psi_group *group, u64 now)
592	{
593	struct psi_trigger *t;
594
595	list_for_each_entry(t, &group->rtpoll_triggers, node)
596	window_reset(win: &t->win, now,
597	value: group->total[PSI_POLL][t->state], prev_growth: 0);
598	memcpy(group->rtpoll_total, group->total[PSI_POLL],
599	sizeof(group->rtpoll_total));
600	group->rtpoll_next_update = now + group->rtpoll_min_period;
601	}
602
603	/* Schedule rtpolling if it's not already scheduled or forced. */
604	static void psi_schedule_rtpoll_work(struct psi_group *group, unsigned long delay,
605	bool force)
606	{
607	struct task_struct *task;
608
609	/*
610	* atomic_xchg should be called even when !force to provide a
611	* full memory barrier (see the comment inside psi_rtpoll_work).
612	*/
613	if (atomic_xchg(v: &group->rtpoll_scheduled, new: 1) && !force)
614	return;
615
616	rcu_read_lock();
617
618	task = rcu_dereference(group->rtpoll_task);
619	/*
620	* kworker might be NULL in case psi_trigger_destroy races with
621	* psi_task_change (hotpath) which can't use locks
622	*/
623	if (likely(task))
624	mod_timer(timer: &group->rtpoll_timer, expires: jiffies + delay);
625	else
626	atomic_set(v: &group->rtpoll_scheduled, i: 0);
627
628	rcu_read_unlock();
629	}
630
631	static void psi_rtpoll_work(struct psi_group *group)
632	{
633	bool force_reschedule = false;
634	u32 changed_states;
635	u64 now;
636
637	mutex_lock(&group->rtpoll_trigger_lock);
638
639	now = sched_clock();
640
641	if (now > group->rtpoll_until) {
642	/*
643	* We are either about to start or might stop rtpolling if no
644	* state change was recorded. Resetting rtpoll_scheduled leaves
645	* a small window for psi_group_change to sneak in and schedule
646	* an immediate rtpoll_work before we get to rescheduling. One
647	* potential extra wakeup at the end of the rtpolling window
648	* should be negligible and rtpoll_next_update still keeps
649	* updates correctly on schedule.
650	*/
651	atomic_set(v: &group->rtpoll_scheduled, i: 0);
652	/*
653	* A task change can race with the rtpoll worker that is supposed to
654	* report on it. To avoid missing events, ensure ordering between
655	* rtpoll_scheduled and the task state accesses, such that if the
656	* rtpoll worker misses the state update, the task change is
657	* guaranteed to reschedule the rtpoll worker:
658	*
659	* rtpoll worker:
660	* atomic_set(rtpoll_scheduled, 0)
661	* smp_mb()
662	* LOAD states
663	*
664	* task change:
665	* STORE states
666	* if atomic_xchg(rtpoll_scheduled, 1) == 0:
667	* schedule rtpoll worker
668	*
669	* The atomic_xchg() implies a full barrier.
670	*/
671	smp_mb();
672	} else {
673	/* The rtpolling window is not over, keep rescheduling */
674	force_reschedule = true;
675	}
676
677
678	collect_percpu_times(group, aggregator: PSI_POLL, pchanged_states: &changed_states);
679
680	if (changed_states & group->rtpoll_states) {
681	/* Initialize trigger windows when entering rtpolling mode */
682	if (now > group->rtpoll_until)
683	init_rtpoll_triggers(group, now);
684
685	/*
686	* Keep the monitor active for at least the duration of the
687	* minimum tracking window as long as monitor states are
688	* changing.
689	*/
690	group->rtpoll_until = now +
691	group->rtpoll_min_period * UPDATES_PER_WINDOW;
692	}
693
694	if (now > group->rtpoll_until) {
695	group->rtpoll_next_update = ULLONG_MAX;
696	goto out;
697	}
698
699	if (now >= group->rtpoll_next_update) {
700	if (changed_states & group->rtpoll_states) {
701	update_triggers(group, now, aggregator: PSI_POLL);
702	memcpy(group->rtpoll_total, group->total[PSI_POLL],
703	sizeof(group->rtpoll_total));
704	}
705	group->rtpoll_next_update = now + group->rtpoll_min_period;
706	}
707
708	psi_schedule_rtpoll_work(group,
709	delay: nsecs_to_jiffies(n: group->rtpoll_next_update - now) + 1,
710	force: force_reschedule);
711
712	out:
713	mutex_unlock(lock: &group->rtpoll_trigger_lock);
714	}
715
716	static int psi_rtpoll_worker(void *data)
717	{
718	struct psi_group *group = (struct psi_group *)data;
719
720	sched_set_fifo_low(current);
721
722	while (true) {
723	wait_event_interruptible(group->rtpoll_wait,
724	atomic_cmpxchg(&group->rtpoll_wakeup, 1, 0) ||
725	kthread_should_stop());
726	if (kthread_should_stop())
727	break;
728
729	psi_rtpoll_work(group);
730	}
731	return 0;
732	}
733
734	static void poll_timer_fn(struct timer_list *t)
735	{
736	struct psi_group *group = timer_container_of(group, t, rtpoll_timer);
737
738	atomic_set(v: &group->rtpoll_wakeup, i: 1);
739	wake_up_interruptible(&group->rtpoll_wait);
740	}
741
742	static void record_times(struct psi_group_cpu *groupc, u64 now)
743	{
744	u32 delta;
745
746	delta = now - groupc->state_start;
747	groupc->state_start = now;
748
749	if (groupc->state_mask & (1 << PSI_IO_SOME)) {
750	groupc->times[PSI_IO_SOME] += delta;
751	if (groupc->state_mask & (1 << PSI_IO_FULL))
752	groupc->times[PSI_IO_FULL] += delta;
753	}
754
755	if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
756	groupc->times[PSI_MEM_SOME] += delta;
757	if (groupc->state_mask & (1 << PSI_MEM_FULL))
758	groupc->times[PSI_MEM_FULL] += delta;
759	}
760
761	if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
762	groupc->times[PSI_CPU_SOME] += delta;
763	if (groupc->state_mask & (1 << PSI_CPU_FULL))
764	groupc->times[PSI_CPU_FULL] += delta;
765	}
766
767	if (groupc->state_mask & (1 << PSI_NONIDLE))
768	groupc->times[PSI_NONIDLE] += delta;
769	}
770
771	static void psi_group_change(struct psi_group *group, int cpu,
772	unsigned int clear, unsigned int set,
773	bool wake_clock)
774	{
775	struct psi_group_cpu *groupc;
776	unsigned int t, m;
777	u32 state_mask;
778	u64 now;
779
780	lockdep_assert_rq_held(cpu_rq(cpu));
781	groupc = per_cpu_ptr(group->pcpu, cpu);
782
783	/*
784	* First we update the task counts according to the state
785	* change requested through the @clear and @set bits.
786	*
787	* Then if the cgroup PSI stats accounting enabled, we
788	* assess the aggregate resource states this CPU's tasks
789	* have been in since the last change, and account any
790	* SOME and FULL time these may have resulted in.
791	*/
792	write_seqcount_begin(&groupc->seq);
793	now = cpu_clock(cpu);
794
795	/*
796	* Start with TSK_ONCPU, which doesn't have a corresponding
797	* task count - it's just a boolean flag directly encoded in
798	* the state mask. Clear, set, or carry the current state if
799	* no changes are requested.
800	*/
801	if (unlikely(clear & TSK_ONCPU)) {
802	state_mask = 0;
803	clear &= ~TSK_ONCPU;
804	} else if (unlikely(set & TSK_ONCPU)) {
805	state_mask = PSI_ONCPU;
806	set &= ~TSK_ONCPU;
807	} else {
808	state_mask = groupc->state_mask & PSI_ONCPU;
809	}
810
811	/*
812	* The rest of the state mask is calculated based on the task
813	* counts. Update those first, then construct the mask.
814	*/
815	for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
816	if (!(m & (1 << t)))
817	continue;
818	if (groupc->tasks[t]) {
819	groupc->tasks[t]--;
820	} else if (!psi_bug) {
821	printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n",
822	cpu, t, groupc->tasks[0],
823	groupc->tasks[1], groupc->tasks[2],
824	groupc->tasks[3], clear, set);
825	psi_bug = 1;
826	}
827	}
828
829	for (t = 0; set; set &= ~(1 << t), t++)
830	if (set & (1 << t))
831	groupc->tasks[t]++;
832
833	if (!group->enabled) {
834	/*
835	* On the first group change after disabling PSI, conclude
836	* the current state and flush its time. This is unlikely
837	* to matter to the user, but aggregation (get_recent_times)
838	* may have already incorporated the live state into times_prev;
839	* avoid a delta sample underflow when PSI is later re-enabled.
840	*/
841	if (unlikely(groupc->state_mask & (1 << PSI_NONIDLE)))
842	record_times(groupc, now);
843
844	groupc->state_mask = state_mask;
845
846	write_seqcount_end(&groupc->seq);
847	return;
848	}
849
850	state_mask = test_states(tasks: groupc->tasks, state_mask);
851
852	/*
853	* Since we care about lost potential, a memstall is FULL
854	* when there are no other working tasks, but also when
855	* the CPU is actively reclaiming and nothing productive
856	* could run even if it were runnable. So when the current
857	* task in a cgroup is in_memstall, the corresponding groupc
858	* on that cpu is in PSI_MEM_FULL state.
859	*/
860	if (unlikely((state_mask & PSI_ONCPU) && cpu_curr(cpu)->in_memstall))
861	state_mask |= (1 << PSI_MEM_FULL);
862
863	record_times(groupc, now);
864
865	groupc->state_mask = state_mask;
866
867	write_seqcount_end(&groupc->seq);
868
869	if (state_mask & group->rtpoll_states)
870	psi_schedule_rtpoll_work(group, delay: 1, force: false);
871
872	if (wake_clock && !delayed_work_pending(&group->avgs_work))
873	schedule_delayed_work(dwork: &group->avgs_work, PSI_FREQ);
874	}
875
876	static inline struct psi_group *task_psi_group(struct task_struct *task)
877	{
878	#ifdef CONFIG_CGROUPS
879	if (static_branch_likely(&psi_cgroups_enabled))
880	return cgroup_psi(cgrp: task_dfl_cgroup(task));
881	#endif
882	return &psi_system;
883	}
884
885	static void psi_flags_change(struct task_struct *task, int clear, int set)
886	{
887	if (((task->psi_flags & set) ||
888	(task->psi_flags & clear) != clear) &&
889	!psi_bug) {
890	printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
891	task->pid, task->comm, task_cpu(task),
892	task->psi_flags, clear, set);
893	psi_bug = 1;
894	}
895
896	task->psi_flags &= ~clear;
897	task->psi_flags |= set;
898	}
899
900	void psi_task_change(struct task_struct *task, int clear, int set)
901	{
902	int cpu = task_cpu(p: task);
903	struct psi_group *group;
904
905	if (!task->pid)
906	return;
907
908	psi_flags_change(task, clear, set);
909
910	group = task_psi_group(task);
911	do {
912	psi_group_change(group, cpu, clear, set, wake_clock: true);
913	} while ((group = group->parent));
914	}
915
916	void psi_task_switch(struct task_struct *prev, struct task_struct *next,
917	bool sleep)
918	{
919	struct psi_group *group, *common = NULL;
920	int cpu = task_cpu(p: prev);
921
922	if (next->pid) {
923	psi_flags_change(task: next, clear: 0, TSK_ONCPU);
924	/*
925	* Set TSK_ONCPU on @next's cgroups. If @next shares any
926	* ancestors with @prev, those will already have @prev's
927	* TSK_ONCPU bit set, and we can stop the iteration there.
928	*/
929	group = task_psi_group(task: next);
930	do {
931	if (per_cpu_ptr(group->pcpu, cpu)->state_mask &
932	PSI_ONCPU) {
933	common = group;
934	break;
935	}
936
937	psi_group_change(group, cpu, clear: 0, TSK_ONCPU, wake_clock: true);
938	} while ((group = group->parent));
939	}
940
941	if (prev->pid) {
942	int clear = TSK_ONCPU, set = 0;
943	bool wake_clock = true;
944
945	/*
946	* When we're going to sleep, psi_dequeue() lets us
947	* handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
948	* TSK_IOWAIT here, where we can combine it with
949	* TSK_ONCPU and save walking common ancestors twice.
950	*/
951	if (sleep) {
952	clear |= TSK_RUNNING;
953	if (prev->in_memstall)
954	clear |= TSK_MEMSTALL_RUNNING;
955	if (prev->in_iowait)
956	set |= TSK_IOWAIT;
957
958	/*
959	* Periodic aggregation shuts off if there is a period of no
960	* task changes, so we wake it back up if necessary. However,
961	* don't do this if the task change is the aggregation worker
962	* itself going to sleep, or we'll ping-pong forever.
963	*/
964	if (unlikely((prev->flags & PF_WQ_WORKER) &&
965	wq_worker_last_func(prev) == psi_avgs_work))
966	wake_clock = false;
967	}
968
969	psi_flags_change(task: prev, clear, set);
970
971	group = task_psi_group(task: prev);
972	do {
973	if (group == common)
974	break;
975	psi_group_change(group, cpu, clear, set, wake_clock);
976	} while ((group = group->parent));
977
978	/*
979	* TSK_ONCPU is handled up to the common ancestor. If there are
980	* any other differences between the two tasks (e.g. prev goes
981	* to sleep, or only one task is memstall), finish propagating
982	* those differences all the way up to the root.
983	*/
984	if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) {
985	clear &= ~TSK_ONCPU;
986	for (; group; group = group->parent)
987	psi_group_change(group, cpu, clear, set, wake_clock);
988	}
989	}
990	}
991
992	#ifdef CONFIG_IRQ_TIME_ACCOUNTING
993	void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_struct *prev)
994	{
995	int cpu = task_cpu(p: curr);
996	struct psi_group *group;
997	struct psi_group_cpu *groupc;
998	s64 delta;
999	u64 irq;
1000
1001	if (static_branch_likely(&psi_disabled) || !irqtime_enabled())
1002	return;
1003
1004	if (!curr->pid)
1005	return;
1006
1007	lockdep_assert_rq_held(rq);
1008	group = task_psi_group(task: curr);
1009	if (prev && task_psi_group(task: prev) == group)
1010	return;
1011
1012	irq = irq_time_read(cpu);
1013	delta = (s64)(irq - rq->psi_irq_time);
1014	if (delta < 0)
1015	return;
1016	rq->psi_irq_time = irq;
1017
1018	do {
1019	u64 now;
1020
1021	if (!group->enabled)
1022	continue;
1023
1024	groupc = per_cpu_ptr(group->pcpu, cpu);
1025
1026	write_seqcount_begin(&groupc->seq);
1027	now = cpu_clock(cpu);
1028
1029	record_times(groupc, now);
1030	groupc->times[PSI_IRQ_FULL] += delta;
1031
1032	write_seqcount_end(&groupc->seq);
1033
1034	if (group->rtpoll_states & (1 << PSI_IRQ_FULL))
1035	psi_schedule_rtpoll_work(group, delay: 1, force: false);
1036	} while ((group = group->parent));
1037	}
1038	#endif
1039
1040	/**
1041	* psi_memstall_enter - mark the beginning of a memory stall section
1042	* @flags: flags to handle nested sections
1043	*
1044	* Marks the calling task as being stalled due to a lack of memory,
1045	* such as waiting for a refault or performing reclaim.
1046	*/
1047	void psi_memstall_enter(unsigned long *flags)
1048	{
1049	struct rq_flags rf;
1050	struct rq *rq;
1051
1052	if (static_branch_likely(&psi_disabled))
1053	return;
1054
1055	*flags = current->in_memstall;
1056	if (*flags)
1057	return;
1058	/*
1059	* in_memstall setting & accounting needs to be atomic wrt
1060	* changes to the task's scheduling state, otherwise we can
1061	* race with CPU migration.
1062	*/
1063	rq = this_rq_lock_irq(rf: &rf);
1064
1065	current->in_memstall = 1;
1066	psi_task_change(current, clear: 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
1067
1068	rq_unlock_irq(rq, rf: &rf);
1069	}
1070	EXPORT_SYMBOL_GPL(psi_memstall_enter);
1071
1072	/**
1073	* psi_memstall_leave - mark the end of an memory stall section
1074	* @flags: flags to handle nested memdelay sections
1075	*
1076	* Marks the calling task as no longer stalled due to lack of memory.
1077	*/
1078	void psi_memstall_leave(unsigned long *flags)
1079	{
1080	struct rq_flags rf;
1081	struct rq *rq;
1082
1083	if (static_branch_likely(&psi_disabled))
1084	return;
1085
1086	if (*flags)
1087	return;
1088	/*
1089	* in_memstall clearing & accounting needs to be atomic wrt
1090	* changes to the task's scheduling state, otherwise we could
1091	* race with CPU migration.
1092	*/
1093	rq = this_rq_lock_irq(rf: &rf);
1094
1095	current->in_memstall = 0;
1096	psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, set: 0);
1097
1098	rq_unlock_irq(rq, rf: &rf);
1099	}
1100	EXPORT_SYMBOL_GPL(psi_memstall_leave);
1101
1102	#ifdef CONFIG_CGROUPS
1103	int psi_cgroup_alloc(struct cgroup *cgroup)
1104	{
1105	if (!static_branch_likely(&psi_cgroups_enabled))
1106	return 0;
1107
1108	cgroup->psi = kzalloc(sizeof(struct psi_group), GFP_KERNEL);
1109	if (!cgroup->psi)
1110	return -ENOMEM;
1111
1112	cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
1113	if (!cgroup->psi->pcpu) {
1114	kfree(objp: cgroup->psi);
1115	return -ENOMEM;
1116	}
1117	group_init(group: cgroup->psi);
1118	cgroup->psi->parent = cgroup_psi(cgrp: cgroup_parent(cgrp: cgroup));
1119	return 0;
1120	}
1121
1122	void psi_cgroup_free(struct cgroup *cgroup)
1123	{
1124	if (!static_branch_likely(&psi_cgroups_enabled))
1125	return;
1126
1127	cancel_delayed_work_sync(dwork: &cgroup->psi->avgs_work);
1128	free_percpu(pdata: cgroup->psi->pcpu);
1129	/* All triggers must be removed by now */
1130	WARN_ONCE(cgroup->psi->rtpoll_states, "psi: trigger leak\n");
1131	kfree(objp: cgroup->psi);
1132	}
1133
1134	/**
1135	* cgroup_move_task - move task to a different cgroup
1136	* @task: the task
1137	* @to: the target css_set
1138	*
1139	* Move task to a new cgroup and safely migrate its associated stall
1140	* state between the different groups.
1141	*
1142	* This function acquires the task's rq lock to lock out concurrent
1143	* changes to the task's scheduling state and - in case the task is
1144	* running - concurrent changes to its stall state.
1145	*/
1146	void cgroup_move_task(struct task_struct *task, struct css_set *to)
1147	{
1148	unsigned int task_flags;
1149	struct rq_flags rf;
1150	struct rq *rq;
1151
1152	if (!static_branch_likely(&psi_cgroups_enabled)) {
1153	/*
1154	* Lame to do this here, but the scheduler cannot be locked
1155	* from the outside, so we move cgroups from inside sched/.
1156	*/
1157	rcu_assign_pointer(task->cgroups, to);
1158	return;
1159	}
1160
1161	rq = task_rq_lock(p: task, rf: &rf);
1162
1163	/*
1164	* We may race with schedule() dropping the rq lock between
1165	* deactivating prev and switching to next. Because the psi
1166	* updates from the deactivation are deferred to the switch
1167	* callback to save cgroup tree updates, the task's scheduling
1168	* state here is not coherent with its psi state:
1169	*
1170	* schedule() cgroup_move_task()
1171	* rq_lock()
1172	* deactivate_task()
1173	* p->on_rq = 0
1174	* psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1175	* pick_next_task()
1176	* rq_unlock()
1177	* rq_lock()
1178	* psi_task_change() // old cgroup
1179	* task->cgroups = to
1180	* psi_task_change() // new cgroup
1181	* rq_unlock()
1182	* rq_lock()
1183	* psi_sched_switch() // does deferred updates in new cgroup
1184	*
1185	* Don't rely on the scheduling state. Use psi_flags instead.
1186	*/
1187	task_flags = task->psi_flags;
1188
1189	if (task_flags)
1190	psi_task_change(task, clear: task_flags, set: 0);
1191
1192	/* See comment above */
1193	rcu_assign_pointer(task->cgroups, to);
1194
1195	if (task_flags)
1196	psi_task_change(task, clear: 0, set: task_flags);
1197
1198	task_rq_unlock(rq, p: task, rf: &rf);
1199	}
1200
1201	void psi_cgroup_restart(struct psi_group *group)
1202	{
1203	int cpu;
1204
1205	/*
1206	* After we disable psi_group->enabled, we don't actually
1207	* stop percpu tasks accounting in each psi_group_cpu,
1208	* instead only stop test_states() loop, record_times()
1209	* and averaging worker, see psi_group_change() for details.
1210	*
1211	* When disable cgroup PSI, this function has nothing to sync
1212	* since cgroup pressure files are hidden and percpu psi_group_cpu
1213	* would see !psi_group->enabled and only do task accounting.
1214	*
1215	* When re-enable cgroup PSI, this function use psi_group_change()
1216	* to get correct state mask from test_states() loop on tasks[],
1217	* and restart groupc->state_start from now, use .clear = .set = 0
1218	* here since no task status really changed.
1219	*/
1220	if (!group->enabled)
1221	return;
1222
1223	for_each_possible_cpu(cpu) {
1224	struct rq *rq = cpu_rq(cpu);
1225	struct rq_flags rf;
1226
1227	rq_lock_irq(rq, rf: &rf);
1228	psi_group_change(group, cpu, clear: 0, set: 0, wake_clock: true);
1229	rq_unlock_irq(rq, rf: &rf);
1230	}
1231	}
1232	#endif /* CONFIG_CGROUPS */
1233
1234	int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1235	{
1236	bool only_full = false;
1237	int full;
1238	u64 now;
1239
1240	if (static_branch_likely(&psi_disabled))
1241	return -EOPNOTSUPP;
1242
1243	#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1244	if (!irqtime_enabled() && res == PSI_IRQ)
1245	return -EOPNOTSUPP;
1246	#endif
1247
1248	/* Update averages before reporting them */
1249	mutex_lock(&group->avgs_lock);
1250	now = sched_clock();
1251	collect_percpu_times(group, aggregator: PSI_AVGS, NULL);
1252	if (now >= group->avg_next_update)
1253	group->avg_next_update = update_averages(group, now);
1254	mutex_unlock(lock: &group->avgs_lock);
1255
1256	#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1257	only_full = res == PSI_IRQ;
1258	#endif
1259
1260	for (full = 0; full < 2 - only_full; full++) {
1261	unsigned long avg[3] = { 0, };
1262	u64 total = 0;
1263	int w;
1264
1265	/* CPU FULL is undefined at the system level */
1266	if (!(group == &psi_system && res == PSI_CPU && full)) {
1267	for (w = 0; w < 3; w++)
1268	avg[w] = group->avg[res * 2 + full][w];
1269	total = div_u64(dividend: group->total[PSI_AVGS][res * 2 + full],
1270	NSEC_PER_USEC);
1271	}
1272
1273	seq_printf(m, fmt: "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1274	full || only_full ? "full" : "some",
1275	LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1276	LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1277	LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1278	total);
1279	}
1280
1281	return 0;
1282	}
1283
1284	struct psi_trigger *psi_trigger_create(struct psi_group *group, char *buf,
1285	enum psi_res res, struct file *file,
1286	struct kernfs_open_file *of)
1287	{
1288	struct psi_trigger *t;
1289	enum psi_states state;
1290	u32 threshold_us;
1291	bool privileged;
1292	u32 window_us;
1293
1294	if (static_branch_likely(&psi_disabled))
1295	return ERR_PTR(error: -EOPNOTSUPP);
1296
1297	/*
1298	* Checking the privilege here on file->f_cred implies that a privileged user
1299	* could open the file and delegate the write to an unprivileged one.
1300	*/
1301	privileged = cap_raised(file->f_cred->cap_effective, CAP_SYS_RESOURCE);
1302
1303	if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1304	state = PSI_IO_SOME + res * 2;
1305	else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1306	state = PSI_IO_FULL + res * 2;
1307	else
1308	return ERR_PTR(error: -EINVAL);
1309
1310	#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1311	if (res == PSI_IRQ && --state != PSI_IRQ_FULL)
1312	return ERR_PTR(error: -EINVAL);
1313	#endif
1314
1315	if (state >= PSI_NONIDLE)
1316	return ERR_PTR(error: -EINVAL);
1317
1318	if (window_us == 0 || window_us > WINDOW_MAX_US)
1319	return ERR_PTR(error: -EINVAL);
1320
1321	/*
1322	* Unprivileged users can only use 2s windows so that averages aggregation
1323	* work is used, and no RT threads need to be spawned.
1324	*/
1325	if (!privileged && window_us % 2000000)
1326	return ERR_PTR(error: -EINVAL);
1327
1328	/* Check threshold */
1329	if (threshold_us == 0 || threshold_us > window_us)
1330	return ERR_PTR(error: -EINVAL);
1331
1332	t = kmalloc(sizeof(*t), GFP_KERNEL);
1333	if (!t)
1334	return ERR_PTR(error: -ENOMEM);
1335
1336	t->group = group;
1337	t->state = state;
1338	t->threshold = threshold_us * NSEC_PER_USEC;
1339	t->win.size = window_us * NSEC_PER_USEC;
1340	window_reset(win: &t->win, now: sched_clock(),
1341	value: group->total[PSI_POLL][t->state], prev_growth: 0);
1342
1343	t->event = 0;
1344	t->last_event_time = 0;
1345	t->of = of;
1346	if (!of)
1347	init_waitqueue_head(&t->event_wait);
1348	t->pending_event = false;
1349	t->aggregator = privileged ? PSI_POLL : PSI_AVGS;
1350
1351	if (privileged) {
1352	mutex_lock(&group->rtpoll_trigger_lock);
1353
1354	if (!rcu_access_pointer(group->rtpoll_task)) {
1355	struct task_struct *task;
1356
1357	task = kthread_create(psi_rtpoll_worker, group, "psimon");
1358	if (IS_ERR(ptr: task)) {
1359	kfree(objp: t);
1360	mutex_unlock(lock: &group->rtpoll_trigger_lock);
1361	return ERR_CAST(ptr: task);
1362	}
1363	atomic_set(v: &group->rtpoll_wakeup, i: 0);
1364	wake_up_process(tsk: task);
1365	rcu_assign_pointer(group->rtpoll_task, task);
1366	}
1367
1368	list_add(new: &t->node, head: &group->rtpoll_triggers);
1369	group->rtpoll_min_period = min(group->rtpoll_min_period,
1370	div_u64(t->win.size, UPDATES_PER_WINDOW));
1371	group->rtpoll_nr_triggers[t->state]++;
1372	group->rtpoll_states |= (1 << t->state);
1373
1374	mutex_unlock(lock: &group->rtpoll_trigger_lock);
1375	} else {
1376	mutex_lock(&group->avgs_lock);
1377
1378	list_add(new: &t->node, head: &group->avg_triggers);
1379	group->avg_nr_triggers[t->state]++;
1380
1381	mutex_unlock(lock: &group->avgs_lock);
1382	}
1383	return t;
1384	}
1385
1386	void psi_trigger_destroy(struct psi_trigger *t)
1387	{
1388	struct psi_group *group;
1389	struct task_struct *task_to_destroy = NULL;
1390
1391	/*
1392	* We do not check psi_disabled since it might have been disabled after
1393	* the trigger got created.
1394	*/
1395	if (!t)
1396	return;
1397
1398	group = t->group;
1399	/*
1400	* Wakeup waiters to stop polling and clear the queue to prevent it from
1401	* being accessed later. Can happen if cgroup is deleted from under a
1402	* polling process.
1403	*/
1404	if (t->of)
1405	kernfs_notify(kn: t->of->kn);
1406	else
1407	wake_up_interruptible(&t->event_wait);
1408
1409	if (t->aggregator == PSI_AVGS) {
1410	mutex_lock(&group->avgs_lock);
1411	if (!list_empty(head: &t->node)) {
1412	list_del(entry: &t->node);
1413	group->avg_nr_triggers[t->state]--;
1414	}
1415	mutex_unlock(lock: &group->avgs_lock);
1416	} else {
1417	mutex_lock(&group->rtpoll_trigger_lock);
1418	if (!list_empty(head: &t->node)) {
1419	struct psi_trigger *tmp;
1420	u64 period = ULLONG_MAX;
1421
1422	list_del(entry: &t->node);
1423	group->rtpoll_nr_triggers[t->state]--;
1424	if (!group->rtpoll_nr_triggers[t->state])
1425	group->rtpoll_states &= ~(1 << t->state);
1426	/*
1427	* Reset min update period for the remaining triggers
1428	* iff the destroying trigger had the min window size.
1429	*/
1430	if (group->rtpoll_min_period == div_u64(dividend: t->win.size, UPDATES_PER_WINDOW)) {
1431	list_for_each_entry(tmp, &group->rtpoll_triggers, node)
1432	period = min(period, div_u64(tmp->win.size,
1433	UPDATES_PER_WINDOW));
1434	group->rtpoll_min_period = period;
1435	}
1436	/* Destroy rtpoll_task when the last trigger is destroyed */
1437	if (group->rtpoll_states == 0) {
1438	group->rtpoll_until = 0;
1439	task_to_destroy = rcu_dereference_protected(
1440	group->rtpoll_task,
1441	lockdep_is_held(&group->rtpoll_trigger_lock));
1442	rcu_assign_pointer(group->rtpoll_task, NULL);
1443	timer_delete(timer: &group->rtpoll_timer);
1444	}
1445	}
1446	mutex_unlock(lock: &group->rtpoll_trigger_lock);
1447	}
1448
1449	/*
1450	* Wait for psi_schedule_rtpoll_work RCU to complete its read-side
1451	* critical section before destroying the trigger and optionally the
1452	* rtpoll_task.
1453	*/
1454	synchronize_rcu();
1455	/*
1456	* Stop kthread 'psimon' after releasing rtpoll_trigger_lock to prevent
1457	* a deadlock while waiting for psi_rtpoll_work to acquire
1458	* rtpoll_trigger_lock
1459	*/
1460	if (task_to_destroy) {
1461	/*
1462	* After the RCU grace period has expired, the worker
1463	* can no longer be found through group->rtpoll_task.
1464	*/
1465	kthread_stop(k: task_to_destroy);
1466	atomic_set(v: &group->rtpoll_scheduled, i: 0);
1467	}
1468	kfree(objp: t);
1469	}
1470
1471	__poll_t psi_trigger_poll(void **trigger_ptr,
1472	struct file *file, poll_table *wait)
1473	{
1474	__poll_t ret = DEFAULT_POLLMASK;
1475	struct psi_trigger *t;
1476
1477	if (static_branch_likely(&psi_disabled))
1478	return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1479
1480	t = smp_load_acquire(trigger_ptr);
1481	if (!t)
1482	return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1483
1484	if (t->of)
1485	kernfs_generic_poll(of: t->of, pt: wait);
1486	else
1487	poll_wait(filp: file, wait_address: &t->event_wait, p: wait);
1488
1489	if (cmpxchg(&t->event, 1, 0) == 1)
1490	ret |= EPOLLPRI;
1491
1492	return ret;
1493	}
1494
1495	#ifdef CONFIG_PROC_FS
1496	static int psi_io_show(struct seq_file *m, void *v)
1497	{
1498	return psi_show(m, group: &psi_system, res: PSI_IO);
1499	}
1500
1501	static int psi_memory_show(struct seq_file *m, void *v)
1502	{
1503	return psi_show(m, group: &psi_system, res: PSI_MEM);
1504	}
1505
1506	static int psi_cpu_show(struct seq_file *m, void *v)
1507	{
1508	return psi_show(m, group: &psi_system, res: PSI_CPU);
1509	}
1510
1511	static int psi_io_open(struct inode *inode, struct file *file)
1512	{
1513	return single_open(file, psi_io_show, NULL);
1514	}
1515
1516	static int psi_memory_open(struct inode *inode, struct file *file)
1517	{
1518	return single_open(file, psi_memory_show, NULL);
1519	}
1520
1521	static int psi_cpu_open(struct inode *inode, struct file *file)
1522	{
1523	return single_open(file, psi_cpu_show, NULL);
1524	}
1525
1526	static ssize_t psi_write(struct file *file, const char __user *user_buf,
1527	size_t nbytes, enum psi_res res)
1528	{
1529	char buf[32];
1530	size_t buf_size;
1531	struct seq_file *seq;
1532	struct psi_trigger *new;
1533
1534	if (static_branch_likely(&psi_disabled))
1535	return -EOPNOTSUPP;
1536
1537	if (!nbytes)
1538	return -EINVAL;
1539
1540	buf_size = min(nbytes, sizeof(buf));
1541	if (copy_from_user(to: buf, from: user_buf, n: buf_size))
1542	return -EFAULT;
1543
1544	buf[buf_size - 1] = '\0';
1545
1546	seq = file->private_data;
1547
1548	/* Take seq->lock to protect seq->private from concurrent writes */
1549	mutex_lock(&seq->lock);
1550
1551	/* Allow only one trigger per file descriptor */
1552	if (seq->private) {
1553	mutex_unlock(lock: &seq->lock);
1554	return -EBUSY;
1555	}
1556
1557	new = psi_trigger_create(group: &psi_system, buf, res, file, NULL);
1558	if (IS_ERR(ptr: new)) {
1559	mutex_unlock(lock: &seq->lock);
1560	return PTR_ERR(ptr: new);
1561	}
1562
1563	smp_store_release(&seq->private, new);
1564	mutex_unlock(lock: &seq->lock);
1565
1566	return nbytes;
1567	}
1568
1569	static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1570	size_t nbytes, loff_t *ppos)
1571	{
1572	return psi_write(file, user_buf, nbytes, res: PSI_IO);
1573	}
1574
1575	static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1576	size_t nbytes, loff_t *ppos)
1577	{
1578	return psi_write(file, user_buf, nbytes, res: PSI_MEM);
1579	}
1580
1581	static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1582	size_t nbytes, loff_t *ppos)
1583	{
1584	return psi_write(file, user_buf, nbytes, res: PSI_CPU);
1585	}
1586
1587	static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1588	{
1589	struct seq_file *seq = file->private_data;
1590
1591	return psi_trigger_poll(trigger_ptr: &seq->private, file, wait);
1592	}
1593
1594	static int psi_fop_release(struct inode *inode, struct file *file)
1595	{
1596	struct seq_file *seq = file->private_data;
1597
1598	psi_trigger_destroy(t: seq->private);
1599	return single_release(inode, file);
1600	}
1601
1602	static const struct proc_ops psi_io_proc_ops = {
1603	.proc_open = psi_io_open,
1604	.proc_read = seq_read,
1605	.proc_lseek = seq_lseek,
1606	.proc_write = psi_io_write,
1607	.proc_poll = psi_fop_poll,
1608	.proc_release = psi_fop_release,
1609	};
1610
1611	static const struct proc_ops psi_memory_proc_ops = {
1612	.proc_open = psi_memory_open,
1613	.proc_read = seq_read,
1614	.proc_lseek = seq_lseek,
1615	.proc_write = psi_memory_write,
1616	.proc_poll = psi_fop_poll,
1617	.proc_release = psi_fop_release,
1618	};
1619
1620	static const struct proc_ops psi_cpu_proc_ops = {
1621	.proc_open = psi_cpu_open,
1622	.proc_read = seq_read,
1623	.proc_lseek = seq_lseek,
1624	.proc_write = psi_cpu_write,
1625	.proc_poll = psi_fop_poll,
1626	.proc_release = psi_fop_release,
1627	};
1628
1629	#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1630	static int psi_irq_show(struct seq_file *m, void *v)
1631	{
1632	return psi_show(m, group: &psi_system, res: PSI_IRQ);
1633	}
1634
1635	static int psi_irq_open(struct inode *inode, struct file *file)
1636	{
1637	return single_open(file, psi_irq_show, NULL);
1638	}
1639
1640	static ssize_t psi_irq_write(struct file *file, const char __user *user_buf,
1641	size_t nbytes, loff_t *ppos)
1642	{
1643	return psi_write(file, user_buf, nbytes, res: PSI_IRQ);
1644	}
1645
1646	static const struct proc_ops psi_irq_proc_ops = {
1647	.proc_open = psi_irq_open,
1648	.proc_read = seq_read,
1649	.proc_lseek = seq_lseek,
1650	.proc_write = psi_irq_write,
1651	.proc_poll = psi_fop_poll,
1652	.proc_release = psi_fop_release,
1653	};
1654	#endif
1655
1656	static int __init psi_proc_init(void)
1657	{
1658	if (psi_enable) {
1659	proc_mkdir("pressure", NULL);
1660	proc_create(name: "pressure/io", mode: 0666, NULL, proc_ops: &psi_io_proc_ops);
1661	proc_create(name: "pressure/memory", mode: 0666, NULL, proc_ops: &psi_memory_proc_ops);
1662	proc_create(name: "pressure/cpu", mode: 0666, NULL, proc_ops: &psi_cpu_proc_ops);
1663	#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1664	proc_create(name: "pressure/irq", mode: 0666, NULL, proc_ops: &psi_irq_proc_ops);
1665	#endif
1666	}
1667	return 0;
1668	}
1669	module_init(psi_proc_init);
1670
1671	#endif /* CONFIG_PROC_FS */
1672