psi.c

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