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 oncpu 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 wallclock time spent in those compound stall
  53 * states gives pressure numbers between 0 and 100 for each resource,
  54 * where the SOME percentage indicates workload slowdowns and the FULL
  55 * percentage indicates reduced CPU utilization:
  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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 140static int psi_bug __read_mostly;
 141
 142DEFINE_STATIC_KEY_FALSE(psi_disabled);
 143DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled);
 144
 145#ifdef CONFIG_PSI_DEFAULT_DISABLED
 146static bool psi_enable;
 147#else
 148static bool psi_enable = true;
 149#endif
 150static int __init setup_psi(char *str)
 151{
 152	return kstrtobool(str, &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_MIN_US 500000	/* Min window size is 500ms */
 164#define WINDOW_MAX_US 10000000	/* Max window size is 10s */
 165#define UPDATES_PER_WINDOW 10	/* 10 updates per window */
 166
 167/* Sampling frequency in nanoseconds */
 168static u64 psi_period __read_mostly;
 169
 170/* System-level pressure and stall tracking */
 171static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu);
 172struct psi_group psi_system = {
 173	.pcpu = &system_group_pcpu,
 174};
 175
 176static void psi_avgs_work(struct work_struct *work);
 177
 178static void poll_timer_fn(struct timer_list *t);
 179
 180static void group_init(struct psi_group *group)
 181{
 182	int cpu;
 183
 184	group->enabled = true;
 185	for_each_possible_cpu(cpu)
 186		seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);
 187	group->avg_last_update = sched_clock();
 188	group->avg_next_update = group->avg_last_update + psi_period;
 189	INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work);
 190	mutex_init(&group->avgs_lock);
 191	/* Init trigger-related members */
 192	atomic_set(&group->poll_scheduled, 0);
 193	mutex_init(&group->trigger_lock);
 194	INIT_LIST_HEAD(&group->triggers);
 
 
 195	group->poll_min_period = U32_MAX;
 
 196	group->polling_next_update = ULLONG_MAX;
 
 197	init_waitqueue_head(&group->poll_wait);
 198	timer_setup(&group->poll_timer, poll_timer_fn, 0);
 199	rcu_assign_pointer(group->poll_task, NULL);
 200}
 201
 202void __init psi_init(void)
 203{
 204	if (!psi_enable) {
 205		static_branch_enable(&psi_disabled);
 206		static_branch_disable(&psi_cgroups_enabled);
 207		return;
 208	}
 209
 210	if (!cgroup_psi_enabled())
 211		static_branch_disable(&psi_cgroups_enabled);
 212
 213	psi_period = jiffies_to_nsecs(PSI_FREQ);
 214	group_init(&psi_system);
 215}
 216
 217static bool test_state(unsigned int *tasks, enum psi_states state, bool oncpu)
 218{
 219	switch (state) {
 220	case PSI_IO_SOME:
 221		return unlikely(tasks[NR_IOWAIT]);
 222	case PSI_IO_FULL:
 223		return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]);
 224	case PSI_MEM_SOME:
 225		return unlikely(tasks[NR_MEMSTALL]);
 226	case PSI_MEM_FULL:
 227		return unlikely(tasks[NR_MEMSTALL] &&
 228			tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]);
 229	case PSI_CPU_SOME:
 230		return unlikely(tasks[NR_RUNNING] > oncpu);
 231	case PSI_CPU_FULL:
 232		return unlikely(tasks[NR_RUNNING] && !oncpu);
 233	case PSI_NONIDLE:
 234		return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] ||
 235			tasks[NR_RUNNING];
 236	default:
 237		return false;
 238	}
 239}
 240
 241static void get_recent_times(struct psi_group *group, int cpu,
 242			     enum psi_aggregators aggregator, u32 *times,
 243			     u32 *pchanged_states)
 244{
 245	struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
 246	int current_cpu = raw_smp_processor_id();
 247	unsigned int tasks[NR_PSI_TASK_COUNTS];
 248	u64 now, state_start;
 249	enum psi_states s;
 250	unsigned int seq;
 251	u32 state_mask;
 252
 253	*pchanged_states = 0;
 254
 255	/* Snapshot a coherent view of the CPU state */
 256	do {
 257		seq = read_seqcount_begin(&groupc->seq);
 258		now = cpu_clock(cpu);
 259		memcpy(times, groupc->times, sizeof(groupc->times));
 260		state_mask = groupc->state_mask;
 261		state_start = groupc->state_start;
 262		if (cpu == current_cpu)
 263			memcpy(tasks, groupc->tasks, sizeof(groupc->tasks));
 264	} while (read_seqcount_retry(&groupc->seq, seq));
 265
 266	/* Calculate state time deltas against the previous snapshot */
 267	for (s = 0; s < NR_PSI_STATES; s++) {
 268		u32 delta;
 269		/*
 270		 * In addition to already concluded states, we also
 271		 * incorporate currently active states on the CPU,
 272		 * since states may last for many sampling periods.
 273		 *
 274		 * This way we keep our delta sampling buckets small
 275		 * (u32) and our reported pressure close to what's
 276		 * actually happening.
 277		 */
 278		if (state_mask & (1 << s))
 279			times[s] += now - state_start;
 280
 281		delta = times[s] - groupc->times_prev[aggregator][s];
 282		groupc->times_prev[aggregator][s] = times[s];
 283
 284		times[s] = delta;
 285		if (delta)
 286			*pchanged_states |= (1 << s);
 287	}
 288
 289	/*
 290	 * When collect_percpu_times() from the avgs_work, we don't want to
 291	 * re-arm avgs_work when all CPUs are IDLE. But the current CPU running
 292	 * this avgs_work is never IDLE, cause avgs_work can't be shut off.
 293	 * So for the current CPU, we need to re-arm avgs_work only when
 294	 * (NR_RUNNING > 1 || NR_IOWAIT > 0 || NR_MEMSTALL > 0), for other CPUs
 295	 * we can just check PSI_NONIDLE delta.
 296	 */
 297	if (current_work() == &group->avgs_work.work) {
 298		bool reschedule;
 299
 300		if (cpu == current_cpu)
 301			reschedule = tasks[NR_RUNNING] +
 302				     tasks[NR_IOWAIT] +
 303				     tasks[NR_MEMSTALL] > 1;
 304		else
 305			reschedule = *pchanged_states & (1 << PSI_NONIDLE);
 306
 307		if (reschedule)
 308			*pchanged_states |= PSI_STATE_RESCHEDULE;
 309	}
 310}
 311
 312static void calc_avgs(unsigned long avg[3], int missed_periods,
 313		      u64 time, u64 period)
 314{
 315	unsigned long pct;
 316
 317	/* Fill in zeroes for periods of no activity */
 318	if (missed_periods) {
 319		avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
 320		avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
 321		avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
 322	}
 323
 324	/* Sample the most recent active period */
 325	pct = div_u64(time * 100, period);
 326	pct *= FIXED_1;
 327	avg[0] = calc_load(avg[0], EXP_10s, pct);
 328	avg[1] = calc_load(avg[1], EXP_60s, pct);
 329	avg[2] = calc_load(avg[2], EXP_300s, pct);
 330}
 331
 332static void collect_percpu_times(struct psi_group *group,
 333				 enum psi_aggregators aggregator,
 334				 u32 *pchanged_states)
 335{
 336	u64 deltas[NR_PSI_STATES - 1] = { 0, };
 337	unsigned long nonidle_total = 0;
 338	u32 changed_states = 0;
 339	int cpu;
 340	int s;
 341
 342	/*
 343	 * Collect the per-cpu time buckets and average them into a
 344	 * single time sample that is normalized to wallclock time.
 345	 *
 346	 * For averaging, each CPU is weighted by its non-idle time in
 347	 * the sampling period. This eliminates artifacts from uneven
 348	 * loading, or even entirely idle CPUs.
 349	 */
 350	for_each_possible_cpu(cpu) {
 351		u32 times[NR_PSI_STATES];
 352		u32 nonidle;
 353		u32 cpu_changed_states;
 354
 355		get_recent_times(group, cpu, aggregator, times,
 356				&cpu_changed_states);
 357		changed_states |= cpu_changed_states;
 358
 359		nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
 360		nonidle_total += nonidle;
 361
 362		for (s = 0; s < PSI_NONIDLE; s++)
 363			deltas[s] += (u64)times[s] * nonidle;
 364	}
 365
 366	/*
 367	 * Integrate the sample into the running statistics that are
 368	 * reported to userspace: the cumulative stall times and the
 369	 * decaying averages.
 370	 *
 371	 * Pressure percentages are sampled at PSI_FREQ. We might be
 372	 * called more often when the user polls more frequently than
 373	 * that; we might be called less often when there is no task
 374	 * activity, thus no data, and clock ticks are sporadic. The
 375	 * below handles both.
 376	 */
 377
 378	/* total= */
 379	for (s = 0; s < NR_PSI_STATES - 1; s++)
 380		group->total[aggregator][s] +=
 381				div_u64(deltas[s], max(nonidle_total, 1UL));
 382
 383	if (pchanged_states)
 384		*pchanged_states = changed_states;
 385}
 386
 387static u64 update_averages(struct psi_group *group, u64 now)
 388{
 389	unsigned long missed_periods = 0;
 390	u64 expires, period;
 391	u64 avg_next_update;
 392	int s;
 393
 394	/* avgX= */
 395	expires = group->avg_next_update;
 396	if (now - expires >= psi_period)
 397		missed_periods = div_u64(now - expires, psi_period);
 398
 399	/*
 400	 * The periodic clock tick can get delayed for various
 401	 * reasons, especially on loaded systems. To avoid clock
 402	 * drift, we schedule the clock in fixed psi_period intervals.
 403	 * But the deltas we sample out of the per-cpu buckets above
 404	 * are based on the actual time elapsing between clock ticks.
 405	 */
 406	avg_next_update = expires + ((1 + missed_periods) * psi_period);
 407	period = now - (group->avg_last_update + (missed_periods * psi_period));
 408	group->avg_last_update = now;
 409
 410	for (s = 0; s < NR_PSI_STATES - 1; s++) {
 411		u32 sample;
 412
 413		sample = group->total[PSI_AVGS][s] - group->avg_total[s];
 414		/*
 415		 * Due to the lockless sampling of the time buckets,
 416		 * recorded time deltas can slip into the next period,
 417		 * which under full pressure can result in samples in
 418		 * excess of the period length.
 419		 *
 420		 * We don't want to report non-sensical pressures in
 421		 * excess of 100%, nor do we want to drop such events
 422		 * on the floor. Instead we punt any overage into the
 423		 * future until pressure subsides. By doing this we
 424		 * don't underreport the occurring pressure curve, we
 425		 * just report it delayed by one period length.
 426		 *
 427		 * The error isn't cumulative. As soon as another
 428		 * delta slips from a period P to P+1, by definition
 429		 * it frees up its time T in P.
 430		 */
 431		if (sample > period)
 432			sample = period;
 433		group->avg_total[s] += sample;
 434		calc_avgs(group->avg[s], missed_periods, sample, period);
 435	}
 436
 437	return avg_next_update;
 438}
 439
 440static void psi_avgs_work(struct work_struct *work)
 441{
 442	struct delayed_work *dwork;
 443	struct psi_group *group;
 444	u32 changed_states;
 
 445	u64 now;
 446
 447	dwork = to_delayed_work(work);
 448	group = container_of(dwork, struct psi_group, avgs_work);
 449
 450	mutex_lock(&group->avgs_lock);
 451
 452	now = sched_clock();
 453
 454	collect_percpu_times(group, PSI_AVGS, &changed_states);
 
 455	/*
 456	 * If there is task activity, periodically fold the per-cpu
 457	 * times and feed samples into the running averages. If things
 458	 * are idle and there is no data to process, stop the clock.
 459	 * Once restarted, we'll catch up the running averages in one
 460	 * go - see calc_avgs() and missed_periods.
 461	 */
 462	if (now >= group->avg_next_update)
 463		group->avg_next_update = update_averages(group, now);
 464
 465	if (changed_states & PSI_STATE_RESCHEDULE) {
 466		schedule_delayed_work(dwork, nsecs_to_jiffies(
 467				group->avg_next_update - now) + 1);
 468	}
 469
 470	mutex_unlock(&group->avgs_lock);
 471}
 472
 473/* Trigger tracking window manipulations */
 474static void window_reset(struct psi_window *win, u64 now, u64 value,
 475			 u64 prev_growth)
 476{
 477	win->start_time = now;
 478	win->start_value = value;
 479	win->prev_growth = prev_growth;
 480}
 481
 482/*
 483 * PSI growth tracking window update and growth calculation routine.
 484 *
 485 * This approximates a sliding tracking window by interpolating
 486 * partially elapsed windows using historical growth data from the
 487 * previous intervals. This minimizes memory requirements (by not storing
 488 * all the intermediate values in the previous window) and simplifies
 489 * the calculations. It works well because PSI signal changes only in
 490 * positive direction and over relatively small window sizes the growth
 491 * is close to linear.
 492 */
 493static u64 window_update(struct psi_window *win, u64 now, u64 value)
 494{
 495	u64 elapsed;
 496	u64 growth;
 497
 498	elapsed = now - win->start_time;
 499	growth = value - win->start_value;
 500	/*
 501	 * After each tracking window passes win->start_value and
 502	 * win->start_time get reset and win->prev_growth stores
 503	 * the average per-window growth of the previous window.
 504	 * win->prev_growth is then used to interpolate additional
 505	 * growth from the previous window assuming it was linear.
 506	 */
 507	if (elapsed > win->size)
 508		window_reset(win, now, value, growth);
 509	else {
 510		u32 remaining;
 511
 512		remaining = win->size - elapsed;
 513		growth += div64_u64(win->prev_growth * remaining, win->size);
 514	}
 515
 516	return growth;
 517}
 518
 519static void init_triggers(struct psi_group *group, u64 now)
 520{
 521	struct psi_trigger *t;
 522
 523	list_for_each_entry(t, &group->triggers, node)
 524		window_reset(&t->win, now,
 525				group->total[PSI_POLL][t->state], 0);
 526	memcpy(group->polling_total, group->total[PSI_POLL],
 527		   sizeof(group->polling_total));
 528	group->polling_next_update = now + group->poll_min_period;
 529}
 530
 531static u64 update_triggers(struct psi_group *group, u64 now)
 532{
 533	struct psi_trigger *t;
 534	bool update_total = false;
 535	u64 *total = group->total[PSI_POLL];
 536
 537	/*
 538	 * On subsequent updates, calculate growth deltas and let
 539	 * watchers know when their specified thresholds are exceeded.
 540	 */
 541	list_for_each_entry(t, &group->triggers, node) {
 542		u64 growth;
 543		bool new_stall;
 544
 545		new_stall = group->polling_total[t->state] != total[t->state];
 546
 547		/* Check for stall activity or a previous threshold breach */
 548		if (!new_stall && !t->pending_event)
 549			continue;
 
 550		/*
 551		 * Check for new stall activity, as well as deferred
 552		 * events that occurred in the last window after the
 553		 * trigger had already fired (we want to ratelimit
 554		 * events without dropping any).
 555		 */
 556		if (new_stall) {
 557			/*
 558			 * Multiple triggers might be looking at the same state,
 559			 * remember to update group->polling_total[] once we've
 560			 * been through all of them. Also remember to extend the
 561			 * polling time if we see new stall activity.
 562			 */
 563			update_total = true;
 564
 565			/* Calculate growth since last update */
 566			growth = window_update(&t->win, now, total[t->state]);
 567			if (!t->pending_event) {
 568				if (growth < t->threshold)
 569					continue;
 570
 571				t->pending_event = true;
 572			}
 573		}
 574		/* Limit event signaling to once per window */
 575		if (now < t->last_event_time + t->win.size)
 576			continue;
 577
 578		/* Generate an event */
 579		if (cmpxchg(&t->event, 0, 1) == 0)
 580			wake_up_interruptible(&t->event_wait);
 581		t->last_event_time = now;
 582		/* Reset threshold breach flag once event got generated */
 583		t->pending_event = false;
 584	}
 585
 586	if (update_total)
 587		memcpy(group->polling_total, total,
 588				sizeof(group->polling_total));
 589
 590	return now + group->poll_min_period;
 591}
 592
 593/* Schedule polling if it's not already scheduled or forced. */
 594static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay,
 595				   bool force)
 596{
 597	struct task_struct *task;
 598
 599	/*
 600	 * atomic_xchg should be called even when !force to provide a
 601	 * full memory barrier (see the comment inside psi_poll_work).
 
 
 602	 */
 603	if (atomic_xchg(&group->poll_scheduled, 1) && !force)
 604		return;
 605
 606	rcu_read_lock();
 607
 608	task = rcu_dereference(group->poll_task);
 609	/*
 610	 * kworker might be NULL in case psi_trigger_destroy races with
 611	 * psi_task_change (hotpath) which can't use locks
 612	 */
 613	if (likely(task))
 614		mod_timer(&group->poll_timer, jiffies + delay);
 615	else
 616		atomic_set(&group->poll_scheduled, 0);
 617
 618	rcu_read_unlock();
 619}
 620
 621static void psi_poll_work(struct psi_group *group)
 622{
 623	bool force_reschedule = false;
 624	u32 changed_states;
 625	u64 now;
 626
 627	mutex_lock(&group->trigger_lock);
 628
 629	now = sched_clock();
 630
 631	if (now > group->polling_until) {
 632		/*
 633		 * We are either about to start or might stop polling if no
 634		 * state change was recorded. Resetting poll_scheduled leaves
 635		 * a small window for psi_group_change to sneak in and schedule
 636		 * an immediate poll_work before we get to rescheduling. One
 637		 * potential extra wakeup at the end of the polling window
 638		 * should be negligible and polling_next_update still keeps
 639		 * updates correctly on schedule.
 640		 */
 641		atomic_set(&group->poll_scheduled, 0);
 642		/*
 643		 * A task change can race with the poll worker that is supposed to
 644		 * report on it. To avoid missing events, ensure ordering between
 645		 * poll_scheduled and the task state accesses, such that if the poll
 646		 * worker misses the state update, the task change is guaranteed to
 647		 * reschedule the poll worker:
 648		 *
 649		 * poll worker:
 650		 *   atomic_set(poll_scheduled, 0)
 651		 *   smp_mb()
 652		 *   LOAD states
 653		 *
 654		 * task change:
 655		 *   STORE states
 656		 *   if atomic_xchg(poll_scheduled, 1) == 0:
 657		 *     schedule poll worker
 658		 *
 659		 * The atomic_xchg() implies a full barrier.
 660		 */
 661		smp_mb();
 662	} else {
 663		/* Polling window is not over, keep rescheduling */
 664		force_reschedule = true;
 665	}
 666
 667
 668	collect_percpu_times(group, PSI_POLL, &changed_states);
 669
 670	if (changed_states & group->poll_states) {
 671		/* Initialize trigger windows when entering polling mode */
 672		if (now > group->polling_until)
 673			init_triggers(group, now);
 674
 675		/*
 676		 * Keep the monitor active for at least the duration of the
 677		 * minimum tracking window as long as monitor states are
 678		 * changing.
 679		 */
 680		group->polling_until = now +
 681			group->poll_min_period * UPDATES_PER_WINDOW;
 682	}
 683
 684	if (now > group->polling_until) {
 685		group->polling_next_update = ULLONG_MAX;
 686		goto out;
 687	}
 688
 689	if (now >= group->polling_next_update)
 690		group->polling_next_update = update_triggers(group, now);
 691
 692	psi_schedule_poll_work(group,
 693		nsecs_to_jiffies(group->polling_next_update - now) + 1,
 694		force_reschedule);
 695
 696out:
 697	mutex_unlock(&group->trigger_lock);
 698}
 699
 700static int psi_poll_worker(void *data)
 701{
 702	struct psi_group *group = (struct psi_group *)data;
 703
 704	sched_set_fifo_low(current);
 705
 706	while (true) {
 707		wait_event_interruptible(group->poll_wait,
 708				atomic_cmpxchg(&group->poll_wakeup, 1, 0) ||
 709				kthread_should_stop());
 710		if (kthread_should_stop())
 711			break;
 712
 713		psi_poll_work(group);
 714	}
 715	return 0;
 716}
 717
 718static void poll_timer_fn(struct timer_list *t)
 719{
 720	struct psi_group *group = from_timer(group, t, poll_timer);
 721
 722	atomic_set(&group->poll_wakeup, 1);
 723	wake_up_interruptible(&group->poll_wait);
 724}
 725
 726static void record_times(struct psi_group_cpu *groupc, u64 now)
 727{
 728	u32 delta;
 729
 730	delta = now - groupc->state_start;
 731	groupc->state_start = now;
 732
 733	if (groupc->state_mask & (1 << PSI_IO_SOME)) {
 734		groupc->times[PSI_IO_SOME] += delta;
 735		if (groupc->state_mask & (1 << PSI_IO_FULL))
 736			groupc->times[PSI_IO_FULL] += delta;
 737	}
 738
 739	if (groupc->state_mask & (1 << PSI_MEM_SOME)) {
 740		groupc->times[PSI_MEM_SOME] += delta;
 741		if (groupc->state_mask & (1 << PSI_MEM_FULL))
 742			groupc->times[PSI_MEM_FULL] += delta;
 743	}
 744
 745	if (groupc->state_mask & (1 << PSI_CPU_SOME)) {
 746		groupc->times[PSI_CPU_SOME] += delta;
 747		if (groupc->state_mask & (1 << PSI_CPU_FULL))
 748			groupc->times[PSI_CPU_FULL] += delta;
 749	}
 750
 751	if (groupc->state_mask & (1 << PSI_NONIDLE))
 752		groupc->times[PSI_NONIDLE] += delta;
 753}
 754
 755static void psi_group_change(struct psi_group *group, int cpu,
 756			     unsigned int clear, unsigned int set, u64 now,
 757			     bool wake_clock)
 758{
 759	struct psi_group_cpu *groupc;
 
 760	unsigned int t, m;
 761	enum psi_states s;
 762	u32 state_mask;
 763
 764	groupc = per_cpu_ptr(group->pcpu, cpu);
 765
 766	/*
 767	 * First we update the task counts according to the state
 768	 * change requested through the @clear and @set bits.
 769	 *
 770	 * Then if the cgroup PSI stats accounting enabled, we
 771	 * assess the aggregate resource states this CPU's tasks
 772	 * have been in since the last change, and account any
 773	 * SOME and FULL time these may have resulted in.
 
 
 
 774	 */
 775	write_seqcount_begin(&groupc->seq);
 776
 777	/*
 778	 * Start with TSK_ONCPU, which doesn't have a corresponding
 779	 * task count - it's just a boolean flag directly encoded in
 780	 * the state mask. Clear, set, or carry the current state if
 781	 * no changes are requested.
 782	 */
 783	if (unlikely(clear & TSK_ONCPU)) {
 784		state_mask = 0;
 785		clear &= ~TSK_ONCPU;
 786	} else if (unlikely(set & TSK_ONCPU)) {
 787		state_mask = PSI_ONCPU;
 788		set &= ~TSK_ONCPU;
 789	} else {
 790		state_mask = groupc->state_mask & PSI_ONCPU;
 791	}
 792
 793	/*
 794	 * The rest of the state mask is calculated based on the task
 795	 * counts. Update those first, then construct the mask.
 796	 */
 797	for (t = 0, m = clear; m; m &= ~(1 << t), t++) {
 798		if (!(m & (1 << t)))
 799			continue;
 800		if (groupc->tasks[t]) {
 801			groupc->tasks[t]--;
 802		} else if (!psi_bug) {
 803			printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n",
 804					cpu, t, groupc->tasks[0],
 805					groupc->tasks[1], groupc->tasks[2],
 806					groupc->tasks[3], clear, set);
 807			psi_bug = 1;
 808		}
 809	}
 810
 811	for (t = 0; set; set &= ~(1 << t), t++)
 812		if (set & (1 << t))
 813			groupc->tasks[t]++;
 814
 815	if (!group->enabled) {
 816		/*
 817		 * On the first group change after disabling PSI, conclude
 818		 * the current state and flush its time. This is unlikely
 819		 * to matter to the user, but aggregation (get_recent_times)
 820		 * may have already incorporated the live state into times_prev;
 821		 * avoid a delta sample underflow when PSI is later re-enabled.
 822		 */
 823		if (unlikely(groupc->state_mask & (1 << PSI_NONIDLE)))
 824			record_times(groupc, now);
 825
 826		groupc->state_mask = state_mask;
 827
 828		write_seqcount_end(&groupc->seq);
 829		return;
 830	}
 831
 832	for (s = 0; s < NR_PSI_STATES; s++) {
 833		if (test_state(groupc->tasks, s, state_mask & PSI_ONCPU))
 834			state_mask |= (1 << s);
 835	}
 836
 837	/*
 838	 * Since we care about lost potential, a memstall is FULL
 839	 * when there are no other working tasks, but also when
 840	 * the CPU is actively reclaiming and nothing productive
 841	 * could run even if it were runnable. So when the current
 842	 * task in a cgroup is in_memstall, the corresponding groupc
 843	 * on that cpu is in PSI_MEM_FULL state.
 844	 */
 845	if (unlikely((state_mask & PSI_ONCPU) && cpu_curr(cpu)->in_memstall))
 846		state_mask |= (1 << PSI_MEM_FULL);
 847
 848	record_times(groupc, now);
 849
 850	groupc->state_mask = state_mask;
 851
 852	write_seqcount_end(&groupc->seq);
 853
 854	if (state_mask & group->poll_states)
 855		psi_schedule_poll_work(group, 1, false);
 856
 857	if (wake_clock && !delayed_work_pending(&group->avgs_work))
 858		schedule_delayed_work(&group->avgs_work, PSI_FREQ);
 859}
 860
 861static inline struct psi_group *task_psi_group(struct task_struct *task)
 862{
 
 
 
 863#ifdef CONFIG_CGROUPS
 864	if (static_branch_likely(&psi_cgroups_enabled))
 865		return cgroup_psi(task_dfl_cgroup(task));
 
 
 
 
 
 
 
 
 
 
 
 866#endif
 
 867	return &psi_system;
 868}
 869
 870static void psi_flags_change(struct task_struct *task, int clear, int set)
 871{
 872	if (((task->psi_flags & set) ||
 873	     (task->psi_flags & clear) != clear) &&
 874	    !psi_bug) {
 875		printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n",
 876				task->pid, task->comm, task_cpu(task),
 877				task->psi_flags, clear, set);
 878		psi_bug = 1;
 879	}
 880
 881	task->psi_flags &= ~clear;
 882	task->psi_flags |= set;
 883}
 884
 885void psi_task_change(struct task_struct *task, int clear, int set)
 886{
 887	int cpu = task_cpu(task);
 888	struct psi_group *group;
 
 
 889	u64 now;
 890
 891	if (!task->pid)
 892		return;
 893
 894	psi_flags_change(task, clear, set);
 895
 896	now = cpu_clock(cpu);
 
 
 
 
 
 
 
 
 
 
 897
 898	group = task_psi_group(task);
 899	do {
 900		psi_group_change(group, cpu, clear, set, now, true);
 901	} while ((group = group->parent));
 902}
 903
 904void psi_task_switch(struct task_struct *prev, struct task_struct *next,
 905		     bool sleep)
 906{
 907	struct psi_group *group, *common = NULL;
 908	int cpu = task_cpu(prev);
 
 909	u64 now = cpu_clock(cpu);
 910
 911	if (next->pid) {
 
 
 912		psi_flags_change(next, 0, TSK_ONCPU);
 913		/*
 914		 * Set TSK_ONCPU on @next's cgroups. If @next shares any
 915		 * ancestors with @prev, those will already have @prev's
 916		 * TSK_ONCPU bit set, and we can stop the iteration there.
 
 
 917		 */
 918		group = task_psi_group(next);
 919		do {
 920			if (per_cpu_ptr(group->pcpu, cpu)->state_mask &
 921			    PSI_ONCPU) {
 
 922				common = group;
 923				break;
 924			}
 925
 926			psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
 927		} while ((group = group->parent));
 928	}
 929
 930	if (prev->pid) {
 931		int clear = TSK_ONCPU, set = 0;
 932		bool wake_clock = true;
 933
 934		/*
 935		 * When we're going to sleep, psi_dequeue() lets us
 936		 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and
 937		 * TSK_IOWAIT here, where we can combine it with
 938		 * TSK_ONCPU and save walking common ancestors twice.
 939		 */
 940		if (sleep) {
 941			clear |= TSK_RUNNING;
 942			if (prev->in_memstall)
 943				clear |= TSK_MEMSTALL_RUNNING;
 944			if (prev->in_iowait)
 945				set |= TSK_IOWAIT;
 946
 947			/*
 948			 * Periodic aggregation shuts off if there is a period of no
 949			 * task changes, so we wake it back up if necessary. However,
 950			 * don't do this if the task change is the aggregation worker
 951			 * itself going to sleep, or we'll ping-pong forever.
 952			 */
 953			if (unlikely((prev->flags & PF_WQ_WORKER) &&
 954				     wq_worker_last_func(prev) == psi_avgs_work))
 955				wake_clock = false;
 956		}
 957
 958		psi_flags_change(prev, clear, set);
 959
 960		group = task_psi_group(prev);
 961		do {
 962			if (group == common)
 963				break;
 964			psi_group_change(group, cpu, clear, set, now, wake_clock);
 965		} while ((group = group->parent));
 966
 967		/*
 968		 * TSK_ONCPU is handled up to the common ancestor. If there are
 969		 * any other differences between the two tasks (e.g. prev goes
 970		 * to sleep, or only one task is memstall), finish propagating
 971		 * those differences all the way up to the root.
 972		 */
 973		if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) {
 974			clear &= ~TSK_ONCPU;
 975			for (; group; group = group->parent)
 976				psi_group_change(group, cpu, clear, set, now, wake_clock);
 977		}
 978	}
 979}
 980
 981#ifdef CONFIG_IRQ_TIME_ACCOUNTING
 982void psi_account_irqtime(struct task_struct *task, u32 delta)
 983{
 984	int cpu = task_cpu(task);
 985	struct psi_group *group;
 986	struct psi_group_cpu *groupc;
 987	u64 now;
 988
 989	if (!task->pid)
 990		return;
 991
 992	now = cpu_clock(cpu);
 993
 994	group = task_psi_group(task);
 995	do {
 996		if (!group->enabled)
 997			continue;
 998
 999		groupc = per_cpu_ptr(group->pcpu, cpu);
1000
1001		write_seqcount_begin(&groupc->seq);
1002
1003		record_times(groupc, now);
1004		groupc->times[PSI_IRQ_FULL] += delta;
1005
1006		write_seqcount_end(&groupc->seq);
1007
1008		if (group->poll_states & (1 << PSI_IRQ_FULL))
1009			psi_schedule_poll_work(group, 1, false);
1010	} while ((group = group->parent));
1011}
1012#endif
1013
1014/**
1015 * psi_memstall_enter - mark the beginning of a memory stall section
1016 * @flags: flags to handle nested sections
1017 *
1018 * Marks the calling task as being stalled due to a lack of memory,
1019 * such as waiting for a refault or performing reclaim.
1020 */
1021void psi_memstall_enter(unsigned long *flags)
1022{
1023	struct rq_flags rf;
1024	struct rq *rq;
1025
1026	if (static_branch_likely(&psi_disabled))
1027		return;
1028
1029	*flags = current->in_memstall;
1030	if (*flags)
1031		return;
1032	/*
1033	 * in_memstall setting & accounting needs to be atomic wrt
1034	 * changes to the task's scheduling state, otherwise we can
1035	 * race with CPU migration.
1036	 */
1037	rq = this_rq_lock_irq(&rf);
1038
1039	current->in_memstall = 1;
1040	psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING);
1041
1042	rq_unlock_irq(rq, &rf);
1043}
1044EXPORT_SYMBOL_GPL(psi_memstall_enter);
1045
1046/**
1047 * psi_memstall_leave - mark the end of an memory stall section
1048 * @flags: flags to handle nested memdelay sections
1049 *
1050 * Marks the calling task as no longer stalled due to lack of memory.
1051 */
1052void psi_memstall_leave(unsigned long *flags)
1053{
1054	struct rq_flags rf;
1055	struct rq *rq;
1056
1057	if (static_branch_likely(&psi_disabled))
1058		return;
1059
1060	if (*flags)
1061		return;
1062	/*
1063	 * in_memstall clearing & accounting needs to be atomic wrt
1064	 * changes to the task's scheduling state, otherwise we could
1065	 * race with CPU migration.
1066	 */
1067	rq = this_rq_lock_irq(&rf);
1068
1069	current->in_memstall = 0;
1070	psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0);
1071
1072	rq_unlock_irq(rq, &rf);
1073}
1074EXPORT_SYMBOL_GPL(psi_memstall_leave);
1075
1076#ifdef CONFIG_CGROUPS
1077int psi_cgroup_alloc(struct cgroup *cgroup)
1078{
1079	if (!static_branch_likely(&psi_cgroups_enabled))
1080		return 0;
1081
1082	cgroup->psi = kzalloc(sizeof(struct psi_group), GFP_KERNEL);
1083	if (!cgroup->psi)
1084		return -ENOMEM;
1085
1086	cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
1087	if (!cgroup->psi->pcpu) {
1088		kfree(cgroup->psi);
1089		return -ENOMEM;
1090	}
1091	group_init(cgroup->psi);
1092	cgroup->psi->parent = cgroup_psi(cgroup_parent(cgroup));
1093	return 0;
1094}
1095
1096void psi_cgroup_free(struct cgroup *cgroup)
1097{
1098	if (!static_branch_likely(&psi_cgroups_enabled))
1099		return;
1100
1101	cancel_delayed_work_sync(&cgroup->psi->avgs_work);
1102	free_percpu(cgroup->psi->pcpu);
1103	/* All triggers must be removed by now */
1104	WARN_ONCE(cgroup->psi->poll_states, "psi: trigger leak\n");
1105	kfree(cgroup->psi);
1106}
1107
1108/**
1109 * cgroup_move_task - move task to a different cgroup
1110 * @task: the task
1111 * @to: the target css_set
1112 *
1113 * Move task to a new cgroup and safely migrate its associated stall
1114 * state between the different groups.
1115 *
1116 * This function acquires the task's rq lock to lock out concurrent
1117 * changes to the task's scheduling state and - in case the task is
1118 * running - concurrent changes to its stall state.
1119 */
1120void cgroup_move_task(struct task_struct *task, struct css_set *to)
1121{
1122	unsigned int task_flags;
1123	struct rq_flags rf;
1124	struct rq *rq;
1125
1126	if (!static_branch_likely(&psi_cgroups_enabled)) {
1127		/*
1128		 * Lame to do this here, but the scheduler cannot be locked
1129		 * from the outside, so we move cgroups from inside sched/.
1130		 */
1131		rcu_assign_pointer(task->cgroups, to);
1132		return;
1133	}
1134
1135	rq = task_rq_lock(task, &rf);
1136
1137	/*
1138	 * We may race with schedule() dropping the rq lock between
1139	 * deactivating prev and switching to next. Because the psi
1140	 * updates from the deactivation are deferred to the switch
1141	 * callback to save cgroup tree updates, the task's scheduling
1142	 * state here is not coherent with its psi state:
1143	 *
1144	 * schedule()                   cgroup_move_task()
1145	 *   rq_lock()
1146	 *   deactivate_task()
1147	 *     p->on_rq = 0
1148	 *     psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates
1149	 *   pick_next_task()
1150	 *     rq_unlock()
1151	 *                                rq_lock()
1152	 *                                psi_task_change() // old cgroup
1153	 *                                task->cgroups = to
1154	 *                                psi_task_change() // new cgroup
1155	 *                                rq_unlock()
1156	 *     rq_lock()
1157	 *   psi_sched_switch() // does deferred updates in new cgroup
1158	 *
1159	 * Don't rely on the scheduling state. Use psi_flags instead.
1160	 */
1161	task_flags = task->psi_flags;
1162
1163	if (task_flags)
1164		psi_task_change(task, task_flags, 0);
1165
1166	/* See comment above */
1167	rcu_assign_pointer(task->cgroups, to);
1168
1169	if (task_flags)
1170		psi_task_change(task, 0, task_flags);
1171
1172	task_rq_unlock(rq, task, &rf);
1173}
1174
1175void psi_cgroup_restart(struct psi_group *group)
1176{
1177	int cpu;
1178
1179	/*
1180	 * After we disable psi_group->enabled, we don't actually
1181	 * stop percpu tasks accounting in each psi_group_cpu,
1182	 * instead only stop test_state() loop, record_times()
1183	 * and averaging worker, see psi_group_change() for details.
1184	 *
1185	 * When disable cgroup PSI, this function has nothing to sync
1186	 * since cgroup pressure files are hidden and percpu psi_group_cpu
1187	 * would see !psi_group->enabled and only do task accounting.
1188	 *
1189	 * When re-enable cgroup PSI, this function use psi_group_change()
1190	 * to get correct state mask from test_state() loop on tasks[],
1191	 * and restart groupc->state_start from now, use .clear = .set = 0
1192	 * here since no task status really changed.
1193	 */
1194	if (!group->enabled)
1195		return;
1196
1197	for_each_possible_cpu(cpu) {
1198		struct rq *rq = cpu_rq(cpu);
1199		struct rq_flags rf;
1200		u64 now;
1201
1202		rq_lock_irq(rq, &rf);
1203		now = cpu_clock(cpu);
1204		psi_group_change(group, cpu, 0, 0, now, true);
1205		rq_unlock_irq(rq, &rf);
1206	}
1207}
1208#endif /* CONFIG_CGROUPS */
1209
1210int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
1211{
1212	bool only_full = false;
1213	int full;
1214	u64 now;
1215
1216	if (static_branch_likely(&psi_disabled))
1217		return -EOPNOTSUPP;
1218
1219	/* Update averages before reporting them */
1220	mutex_lock(&group->avgs_lock);
1221	now = sched_clock();
1222	collect_percpu_times(group, PSI_AVGS, NULL);
1223	if (now >= group->avg_next_update)
1224		group->avg_next_update = update_averages(group, now);
1225	mutex_unlock(&group->avgs_lock);
1226
1227#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1228	only_full = res == PSI_IRQ;
1229#endif
1230
1231	for (full = 0; full < 2 - only_full; full++) {
1232		unsigned long avg[3] = { 0, };
1233		u64 total = 0;
1234		int w;
1235
1236		/* CPU FULL is undefined at the system level */
1237		if (!(group == &psi_system && res == PSI_CPU && full)) {
1238			for (w = 0; w < 3; w++)
1239				avg[w] = group->avg[res * 2 + full][w];
1240			total = div_u64(group->total[PSI_AVGS][res * 2 + full],
1241					NSEC_PER_USEC);
1242		}
1243
1244		seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n",
1245			   full || only_full ? "full" : "some",
1246			   LOAD_INT(avg[0]), LOAD_FRAC(avg[0]),
1247			   LOAD_INT(avg[1]), LOAD_FRAC(avg[1]),
1248			   LOAD_INT(avg[2]), LOAD_FRAC(avg[2]),
1249			   total);
1250	}
1251
1252	return 0;
1253}
1254
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1255struct psi_trigger *psi_trigger_create(struct psi_group *group,
1256			char *buf, enum psi_res res)
1257{
1258	struct psi_trigger *t;
1259	enum psi_states state;
1260	u32 threshold_us;
1261	u32 window_us;
1262
1263	if (static_branch_likely(&psi_disabled))
1264		return ERR_PTR(-EOPNOTSUPP);
1265
1266	if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2)
1267		state = PSI_IO_SOME + res * 2;
1268	else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2)
1269		state = PSI_IO_FULL + res * 2;
1270	else
1271		return ERR_PTR(-EINVAL);
1272
1273#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1274	if (res == PSI_IRQ && --state != PSI_IRQ_FULL)
1275		return ERR_PTR(-EINVAL);
1276#endif
1277
1278	if (state >= PSI_NONIDLE)
1279		return ERR_PTR(-EINVAL);
1280
1281	if (window_us < WINDOW_MIN_US ||
1282		window_us > WINDOW_MAX_US)
1283		return ERR_PTR(-EINVAL);
1284
1285	/* Check threshold */
1286	if (threshold_us == 0 || threshold_us > window_us)
1287		return ERR_PTR(-EINVAL);
1288
1289	t = kmalloc(sizeof(*t), GFP_KERNEL);
1290	if (!t)
1291		return ERR_PTR(-ENOMEM);
1292
1293	t->group = group;
1294	t->state = state;
1295	t->threshold = threshold_us * NSEC_PER_USEC;
1296	t->win.size = window_us * NSEC_PER_USEC;
1297	window_reset(&t->win, sched_clock(),
1298			group->total[PSI_POLL][t->state], 0);
1299
1300	t->event = 0;
1301	t->last_event_time = 0;
1302	init_waitqueue_head(&t->event_wait);
1303	t->pending_event = false;
1304
1305	mutex_lock(&group->trigger_lock);
1306
1307	if (!rcu_access_pointer(group->poll_task)) {
1308		struct task_struct *task;
1309
1310		task = kthread_create(psi_poll_worker, group, "psimon");
1311		if (IS_ERR(task)) {
1312			kfree(t);
1313			mutex_unlock(&group->trigger_lock);
1314			return ERR_CAST(task);
1315		}
1316		atomic_set(&group->poll_wakeup, 0);
1317		wake_up_process(task);
1318		rcu_assign_pointer(group->poll_task, task);
1319	}
1320
1321	list_add(&t->node, &group->triggers);
1322	group->poll_min_period = min(group->poll_min_period,
1323		div_u64(t->win.size, UPDATES_PER_WINDOW));
1324	group->nr_triggers[t->state]++;
1325	group->poll_states |= (1 << t->state);
1326
1327	mutex_unlock(&group->trigger_lock);
1328
1329	return t;
1330}
1331
1332void psi_trigger_destroy(struct psi_trigger *t)
1333{
1334	struct psi_group *group;
 
1335	struct task_struct *task_to_destroy = NULL;
1336
1337	/*
1338	 * We do not check psi_disabled since it might have been disabled after
1339	 * the trigger got created.
1340	 */
1341	if (!t)
1342		return;
1343
1344	group = t->group;
1345	/*
1346	 * Wakeup waiters to stop polling and clear the queue to prevent it from
1347	 * being accessed later. Can happen if cgroup is deleted from under a
1348	 * polling process.
1349	 */
1350	wake_up_pollfree(&t->event_wait);
1351
1352	mutex_lock(&group->trigger_lock);
1353
1354	if (!list_empty(&t->node)) {
1355		struct psi_trigger *tmp;
1356		u64 period = ULLONG_MAX;
1357
1358		list_del(&t->node);
1359		group->nr_triggers[t->state]--;
1360		if (!group->nr_triggers[t->state])
1361			group->poll_states &= ~(1 << t->state);
1362		/* reset min update period for the remaining triggers */
1363		list_for_each_entry(tmp, &group->triggers, node)
1364			period = min(period, div_u64(tmp->win.size,
1365					UPDATES_PER_WINDOW));
1366		group->poll_min_period = period;
1367		/* Destroy poll_task when the last trigger is destroyed */
1368		if (group->poll_states == 0) {
1369			group->polling_until = 0;
1370			task_to_destroy = rcu_dereference_protected(
1371					group->poll_task,
1372					lockdep_is_held(&group->trigger_lock));
1373			rcu_assign_pointer(group->poll_task, NULL);
1374			del_timer(&group->poll_timer);
1375		}
1376	}
1377
1378	mutex_unlock(&group->trigger_lock);
1379
1380	/*
1381	 * Wait for psi_schedule_poll_work RCU to complete its read-side
1382	 * critical section before destroying the trigger and optionally the
1383	 * poll_task.
1384	 */
1385	synchronize_rcu();
1386	/*
1387	 * Stop kthread 'psimon' after releasing trigger_lock to prevent a
1388	 * deadlock while waiting for psi_poll_work to acquire trigger_lock
1389	 */
1390	if (task_to_destroy) {
1391		/*
1392		 * After the RCU grace period has expired, the worker
1393		 * can no longer be found through group->poll_task.
1394		 */
1395		kthread_stop(task_to_destroy);
1396		atomic_set(&group->poll_scheduled, 0);
1397	}
1398	kfree(t);
1399}
1400
 
 
 
 
 
 
 
 
 
 
 
 
1401__poll_t psi_trigger_poll(void **trigger_ptr,
1402				struct file *file, poll_table *wait)
1403{
1404	__poll_t ret = DEFAULT_POLLMASK;
1405	struct psi_trigger *t;
1406
1407	if (static_branch_likely(&psi_disabled))
1408		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
1409
1410	t = smp_load_acquire(trigger_ptr);
1411	if (!t)
 
 
 
1412		return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI;
 
 
 
 
1413
1414	poll_wait(file, &t->event_wait, wait);
1415
1416	if (cmpxchg(&t->event, 1, 0) == 1)
1417		ret |= EPOLLPRI;
1418
1419	return ret;
1420}
1421
1422#ifdef CONFIG_PROC_FS
1423static int psi_io_show(struct seq_file *m, void *v)
1424{
1425	return psi_show(m, &psi_system, PSI_IO);
1426}
1427
1428static int psi_memory_show(struct seq_file *m, void *v)
1429{
1430	return psi_show(m, &psi_system, PSI_MEM);
1431}
1432
1433static int psi_cpu_show(struct seq_file *m, void *v)
1434{
1435	return psi_show(m, &psi_system, PSI_CPU);
1436}
1437
1438static int psi_open(struct file *file, int (*psi_show)(struct seq_file *, void *))
1439{
1440	if (file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE))
1441		return -EPERM;
1442
1443	return single_open(file, psi_show, NULL);
1444}
1445
1446static int psi_io_open(struct inode *inode, struct file *file)
1447{
1448	return psi_open(file, psi_io_show);
1449}
1450
1451static int psi_memory_open(struct inode *inode, struct file *file)
1452{
1453	return psi_open(file, psi_memory_show);
1454}
1455
1456static int psi_cpu_open(struct inode *inode, struct file *file)
1457{
1458	return psi_open(file, psi_cpu_show);
1459}
1460
1461static ssize_t psi_write(struct file *file, const char __user *user_buf,
1462			 size_t nbytes, enum psi_res res)
1463{
1464	char buf[32];
1465	size_t buf_size;
1466	struct seq_file *seq;
1467	struct psi_trigger *new;
1468
1469	if (static_branch_likely(&psi_disabled))
1470		return -EOPNOTSUPP;
1471
1472	if (!nbytes)
1473		return -EINVAL;
1474
1475	buf_size = min(nbytes, sizeof(buf));
1476	if (copy_from_user(buf, user_buf, buf_size))
1477		return -EFAULT;
1478
1479	buf[buf_size - 1] = '\0';
1480
1481	seq = file->private_data;
 
 
1482
 
1483	/* Take seq->lock to protect seq->private from concurrent writes */
1484	mutex_lock(&seq->lock);
1485
1486	/* Allow only one trigger per file descriptor */
1487	if (seq->private) {
1488		mutex_unlock(&seq->lock);
1489		return -EBUSY;
1490	}
1491
1492	new = psi_trigger_create(&psi_system, buf, res);
1493	if (IS_ERR(new)) {
1494		mutex_unlock(&seq->lock);
1495		return PTR_ERR(new);
1496	}
1497
1498	smp_store_release(&seq->private, new);
1499	mutex_unlock(&seq->lock);
1500
1501	return nbytes;
1502}
1503
1504static ssize_t psi_io_write(struct file *file, const char __user *user_buf,
1505			    size_t nbytes, loff_t *ppos)
1506{
1507	return psi_write(file, user_buf, nbytes, PSI_IO);
1508}
1509
1510static ssize_t psi_memory_write(struct file *file, const char __user *user_buf,
1511				size_t nbytes, loff_t *ppos)
1512{
1513	return psi_write(file, user_buf, nbytes, PSI_MEM);
1514}
1515
1516static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf,
1517			     size_t nbytes, loff_t *ppos)
1518{
1519	return psi_write(file, user_buf, nbytes, PSI_CPU);
1520}
1521
1522static __poll_t psi_fop_poll(struct file *file, poll_table *wait)
1523{
1524	struct seq_file *seq = file->private_data;
1525
1526	return psi_trigger_poll(&seq->private, file, wait);
1527}
1528
1529static int psi_fop_release(struct inode *inode, struct file *file)
1530{
1531	struct seq_file *seq = file->private_data;
1532
1533	psi_trigger_destroy(seq->private);
1534	return single_release(inode, file);
1535}
1536
1537static const struct proc_ops psi_io_proc_ops = {
1538	.proc_open	= psi_io_open,
1539	.proc_read	= seq_read,
1540	.proc_lseek	= seq_lseek,
1541	.proc_write	= psi_io_write,
1542	.proc_poll	= psi_fop_poll,
1543	.proc_release	= psi_fop_release,
1544};
1545
1546static const struct proc_ops psi_memory_proc_ops = {
1547	.proc_open	= psi_memory_open,
1548	.proc_read	= seq_read,
1549	.proc_lseek	= seq_lseek,
1550	.proc_write	= psi_memory_write,
1551	.proc_poll	= psi_fop_poll,
1552	.proc_release	= psi_fop_release,
1553};
1554
1555static const struct proc_ops psi_cpu_proc_ops = {
1556	.proc_open	= psi_cpu_open,
1557	.proc_read	= seq_read,
1558	.proc_lseek	= seq_lseek,
1559	.proc_write	= psi_cpu_write,
1560	.proc_poll	= psi_fop_poll,
1561	.proc_release	= psi_fop_release,
1562};
1563
1564#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1565static int psi_irq_show(struct seq_file *m, void *v)
1566{
1567	return psi_show(m, &psi_system, PSI_IRQ);
1568}
1569
1570static int psi_irq_open(struct inode *inode, struct file *file)
1571{
1572	return psi_open(file, psi_irq_show);
1573}
1574
1575static ssize_t psi_irq_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, PSI_IRQ);
1579}
1580
1581static const struct proc_ops psi_irq_proc_ops = {
1582	.proc_open	= psi_irq_open,
1583	.proc_read	= seq_read,
1584	.proc_lseek	= seq_lseek,
1585	.proc_write	= psi_irq_write,
1586	.proc_poll	= psi_fop_poll,
1587	.proc_release	= psi_fop_release,
1588};
1589#endif
1590
1591static int __init psi_proc_init(void)
1592{
1593	if (psi_enable) {
1594		proc_mkdir("pressure", NULL);
1595		proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops);
1596		proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops);
1597		proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops);
1598#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1599		proc_create("pressure/irq", 0666, NULL, &psi_irq_proc_ops);
1600#endif
1601	}
1602	return 0;
1603}
1604module_init(psi_proc_init);
1605
1606#endif /* CONFIG_PROC_FS */