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 */ 139 140 static int psi_bug __read_mostly; 141 142 DEFINE_STATIC_KEY_FALSE(psi_disabled); 143 static DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled); 144 145 #ifdef CONFIG_PSI_DEFAULT_DISABLED 146 static bool psi_enable; 147 #else 148 static bool psi_enable = true; 149 #endif 150 static int __init setup_psi(char *str) 151 { 152 return kstrtobool(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_MAX_US 10000000 /* Max window size is 10s */ 164 #define UPDATES_PER_WINDOW 10 /* 10 updates per window */ 165 166 /* Sampling frequency in nanoseconds */ 167 static u64 psi_period __read_mostly; 168 169 /* System-level pressure and stall tracking */ 170 static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu); 171 struct psi_group psi_system = { 172 .pcpu = &system_group_pcpu, 173 }; 174 175 static void psi_avgs_work(struct work_struct *work); 176 177 static void poll_timer_fn(struct timer_list *t); 178 179 static void group_init(struct psi_group *group) 180 { 181 int cpu; 182 183 group->enabled = true; 184 for_each_possible_cpu(cpu) 185 seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq); 186 group->avg_last_update = sched_clock(); 187 group->avg_next_update = group->avg_last_update + psi_period; 188 mutex_init(&group->avgs_lock); 189 190 /* Init avg trigger-related members */ 191 INIT_LIST_HEAD(&group->avg_triggers); 192 memset(group->avg_nr_triggers, 0, sizeof(group->avg_nr_triggers)); 193 INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work); 194 195 /* Init rtpoll trigger-related members */ 196 atomic_set(&group->rtpoll_scheduled, 0); 197 mutex_init(&group->rtpoll_trigger_lock); 198 INIT_LIST_HEAD(&group->rtpoll_triggers); 199 group->rtpoll_min_period = U32_MAX; 200 group->rtpoll_next_update = ULLONG_MAX; 201 init_waitqueue_head(&group->rtpoll_wait); 202 timer_setup(&group->rtpoll_timer, poll_timer_fn, 0); 203 rcu_assign_pointer(group->rtpoll_task, NULL); 204 } 205 206 void __init psi_init(void) 207 { 208 if (!psi_enable) { 209 static_branch_enable(&psi_disabled); 210 static_branch_disable(&psi_cgroups_enabled); 211 return; 212 } 213 214 if (!cgroup_psi_enabled()) 215 static_branch_disable(&psi_cgroups_enabled); 216 217 psi_period = jiffies_to_nsecs(PSI_FREQ); 218 group_init(&psi_system); 219 } 220 221 static bool test_state(unsigned int *tasks, enum psi_states state, bool oncpu) 222 { 223 switch (state) { 224 case PSI_IO_SOME: 225 return unlikely(tasks[NR_IOWAIT]); 226 case PSI_IO_FULL: 227 return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]); 228 case PSI_MEM_SOME: 229 return unlikely(tasks[NR_MEMSTALL]); 230 case PSI_MEM_FULL: 231 return unlikely(tasks[NR_MEMSTALL] && 232 tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]); 233 case PSI_CPU_SOME: 234 return unlikely(tasks[NR_RUNNING] > oncpu); 235 case PSI_CPU_FULL: 236 return unlikely(tasks[NR_RUNNING] && !oncpu); 237 case PSI_NONIDLE: 238 return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] || 239 tasks[NR_RUNNING]; 240 default: 241 return false; 242 } 243 } 244 245 static void get_recent_times(struct psi_group *group, int cpu, 246 enum psi_aggregators aggregator, u32 *times, 247 u32 *pchanged_states) 248 { 249 struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu); 250 int current_cpu = raw_smp_processor_id(); 251 unsigned int tasks[NR_PSI_TASK_COUNTS]; 252 u64 now, state_start; 253 enum psi_states s; 254 unsigned int seq; 255 u32 state_mask; 256 257 *pchanged_states = 0; 258 259 /* Snapshot a coherent view of the CPU state */ 260 do { 261 seq = read_seqcount_begin(&groupc->seq); 262 now = cpu_clock(cpu); 263 memcpy(times, groupc->times, sizeof(groupc->times)); 264 state_mask = groupc->state_mask; 265 state_start = groupc->state_start; 266 if (cpu == current_cpu) 267 memcpy(tasks, groupc->tasks, sizeof(groupc->tasks)); 268 } while (read_seqcount_retry(&groupc->seq, seq)); 269 270 /* Calculate state time deltas against the previous snapshot */ 271 for (s = 0; s < NR_PSI_STATES; s++) { 272 u32 delta; 273 /* 274 * In addition to already concluded states, we also 275 * incorporate currently active states on the CPU, 276 * since states may last for many sampling periods. 277 * 278 * This way we keep our delta sampling buckets small 279 * (u32) and our reported pressure close to what's 280 * actually happening. 281 */ 282 if (state_mask & (1 << s)) 283 times[s] += now - state_start; 284 285 delta = times[s] - groupc->times_prev[aggregator][s]; 286 groupc->times_prev[aggregator][s] = times[s]; 287 288 times[s] = delta; 289 if (delta) 290 *pchanged_states |= (1 << s); 291 } 292 293 /* 294 * When collect_percpu_times() from the avgs_work, we don't want to 295 * re-arm avgs_work when all CPUs are IDLE. But the current CPU running 296 * this avgs_work is never IDLE, cause avgs_work can't be shut off. 297 * So for the current CPU, we need to re-arm avgs_work only when 298 * (NR_RUNNING > 1 || NR_IOWAIT > 0 || NR_MEMSTALL > 0), for other CPUs 299 * we can just check PSI_NONIDLE delta. 300 */ 301 if (current_work() == &group->avgs_work.work) { 302 bool reschedule; 303 304 if (cpu == current_cpu) 305 reschedule = tasks[NR_RUNNING] + 306 tasks[NR_IOWAIT] + 307 tasks[NR_MEMSTALL] > 1; 308 else 309 reschedule = *pchanged_states & (1 << PSI_NONIDLE); 310 311 if (reschedule) 312 *pchanged_states |= PSI_STATE_RESCHEDULE; 313 } 314 } 315 316 static void calc_avgs(unsigned long avg[3], int missed_periods, 317 u64 time, u64 period) 318 { 319 unsigned long pct; 320 321 /* Fill in zeroes for periods of no activity */ 322 if (missed_periods) { 323 avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods); 324 avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods); 325 avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods); 326 } 327 328 /* Sample the most recent active period */ 329 pct = div_u64(time * 100, period); 330 pct *= FIXED_1; 331 avg[0] = calc_load(avg[0], EXP_10s, pct); 332 avg[1] = calc_load(avg[1], EXP_60s, pct); 333 avg[2] = calc_load(avg[2], EXP_300s, pct); 334 } 335 336 static void collect_percpu_times(struct psi_group *group, 337 enum psi_aggregators aggregator, 338 u32 *pchanged_states) 339 { 340 u64 deltas[NR_PSI_STATES - 1] = { 0, }; 341 unsigned long nonidle_total = 0; 342 u32 changed_states = 0; 343 int cpu; 344 int s; 345 346 /* 347 * Collect the per-cpu time buckets and average them into a 348 * single time sample that is normalized to wallclock time. 349 * 350 * For averaging, each CPU is weighted by its non-idle time in 351 * the sampling period. This eliminates artifacts from uneven 352 * loading, or even entirely idle CPUs. 353 */ 354 for_each_possible_cpu(cpu) { 355 u32 times[NR_PSI_STATES]; 356 u32 nonidle; 357 u32 cpu_changed_states; 358 359 get_recent_times(group, cpu, aggregator, times, 360 &cpu_changed_states); 361 changed_states |= cpu_changed_states; 362 363 nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]); 364 nonidle_total += nonidle; 365 366 for (s = 0; s < PSI_NONIDLE; s++) 367 deltas[s] += (u64)times[s] * nonidle; 368 } 369 370 /* 371 * Integrate the sample into the running statistics that are 372 * reported to userspace: the cumulative stall times and the 373 * decaying averages. 374 * 375 * Pressure percentages are sampled at PSI_FREQ. We might be 376 * called more often when the user polls more frequently than 377 * that; we might be called less often when there is no task 378 * activity, thus no data, and clock ticks are sporadic. The 379 * below handles both. 380 */ 381 382 /* total= */ 383 for (s = 0; s < NR_PSI_STATES - 1; s++) 384 group->total[aggregator][s] += 385 div_u64(deltas[s], max(nonidle_total, 1UL)); 386 387 if (pchanged_states) 388 *pchanged_states = changed_states; 389 } 390 391 /* Trigger tracking window manipulations */ 392 static void window_reset(struct psi_window *win, u64 now, u64 value, 393 u64 prev_growth) 394 { 395 win->start_time = now; 396 win->start_value = value; 397 win->prev_growth = prev_growth; 398 } 399 400 /* 401 * PSI growth tracking window update and growth calculation routine. 402 * 403 * This approximates a sliding tracking window by interpolating 404 * partially elapsed windows using historical growth data from the 405 * previous intervals. This minimizes memory requirements (by not storing 406 * all the intermediate values in the previous window) and simplifies 407 * the calculations. It works well because PSI signal changes only in 408 * positive direction and over relatively small window sizes the growth 409 * is close to linear. 410 */ 411 static u64 window_update(struct psi_window *win, u64 now, u64 value) 412 { 413 u64 elapsed; 414 u64 growth; 415 416 elapsed = now - win->start_time; 417 growth = value - win->start_value; 418 /* 419 * After each tracking window passes win->start_value and 420 * win->start_time get reset and win->prev_growth stores 421 * the average per-window growth of the previous window. 422 * win->prev_growth is then used to interpolate additional 423 * growth from the previous window assuming it was linear. 424 */ 425 if (elapsed > win->size) 426 window_reset(win, now, value, growth); 427 else { 428 u32 remaining; 429 430 remaining = win->size - elapsed; 431 growth += div64_u64(win->prev_growth * remaining, win->size); 432 } 433 434 return growth; 435 } 436 437 static u64 update_triggers(struct psi_group *group, u64 now, bool *update_total, 438 enum psi_aggregators aggregator) 439 { 440 struct psi_trigger *t; 441 u64 *total = group->total[aggregator]; 442 struct list_head *triggers; 443 u64 *aggregator_total; 444 *update_total = false; 445 446 if (aggregator == PSI_AVGS) { 447 triggers = &group->avg_triggers; 448 aggregator_total = group->avg_total; 449 } else { 450 triggers = &group->rtpoll_triggers; 451 aggregator_total = group->rtpoll_total; 452 } 453 454 /* 455 * On subsequent updates, calculate growth deltas and let 456 * watchers know when their specified thresholds are exceeded. 457 */ 458 list_for_each_entry(t, triggers, node) { 459 u64 growth; 460 bool new_stall; 461 462 new_stall = aggregator_total[t->state] != total[t->state]; 463 464 /* Check for stall activity or a previous threshold breach */ 465 if (!new_stall && !t->pending_event) 466 continue; 467 /* 468 * Check for new stall activity, as well as deferred 469 * events that occurred in the last window after the 470 * trigger had already fired (we want to ratelimit 471 * events without dropping any). 472 */ 473 if (new_stall) { 474 /* 475 * Multiple triggers might be looking at the same state, 476 * remember to update group->polling_total[] once we've 477 * been through all of them. Also remember to extend the 478 * polling time if we see new stall activity. 479 */ 480 *update_total = true; 481 482 /* Calculate growth since last update */ 483 growth = window_update(&t->win, now, total[t->state]); 484 if (!t->pending_event) { 485 if (growth < t->threshold) 486 continue; 487 488 t->pending_event = true; 489 } 490 } 491 /* Limit event signaling to once per window */ 492 if (now < t->last_event_time + t->win.size) 493 continue; 494 495 /* Generate an event */ 496 if (cmpxchg(&t->event, 0, 1) == 0) { 497 if (t->of) 498 kernfs_notify(t->of->kn); 499 else 500 wake_up_interruptible(&t->event_wait); 501 } 502 t->last_event_time = now; 503 /* Reset threshold breach flag once event got generated */ 504 t->pending_event = false; 505 } 506 507 return now + group->rtpoll_min_period; 508 } 509 510 static u64 update_averages(struct psi_group *group, u64 now) 511 { 512 unsigned long missed_periods = 0; 513 u64 expires, period; 514 u64 avg_next_update; 515 int s; 516 517 /* avgX= */ 518 expires = group->avg_next_update; 519 if (now - expires >= psi_period) 520 missed_periods = div_u64(now - expires, psi_period); 521 522 /* 523 * The periodic clock tick can get delayed for various 524 * reasons, especially on loaded systems. To avoid clock 525 * drift, we schedule the clock in fixed psi_period intervals. 526 * But the deltas we sample out of the per-cpu buckets above 527 * are based on the actual time elapsing between clock ticks. 528 */ 529 avg_next_update = expires + ((1 + missed_periods) * psi_period); 530 period = now - (group->avg_last_update + (missed_periods * psi_period)); 531 group->avg_last_update = now; 532 533 for (s = 0; s < NR_PSI_STATES - 1; s++) { 534 u32 sample; 535 536 sample = group->total[PSI_AVGS][s] - group->avg_total[s]; 537 /* 538 * Due to the lockless sampling of the time buckets, 539 * recorded time deltas can slip into the next period, 540 * which under full pressure can result in samples in 541 * excess of the period length. 542 * 543 * We don't want to report non-sensical pressures in 544 * excess of 100%, nor do we want to drop such events 545 * on the floor. Instead we punt any overage into the 546 * future until pressure subsides. By doing this we 547 * don't underreport the occurring pressure curve, we 548 * just report it delayed by one period length. 549 * 550 * The error isn't cumulative. As soon as another 551 * delta slips from a period P to P+1, by definition 552 * it frees up its time T in P. 553 */ 554 if (sample > period) 555 sample = period; 556 group->avg_total[s] += sample; 557 calc_avgs(group->avg[s], missed_periods, sample, period); 558 } 559 560 return avg_next_update; 561 } 562 563 static void psi_avgs_work(struct work_struct *work) 564 { 565 struct delayed_work *dwork; 566 struct psi_group *group; 567 u32 changed_states; 568 bool update_total; 569 u64 now; 570 571 dwork = to_delayed_work(work); 572 group = container_of(dwork, struct psi_group, avgs_work); 573 574 mutex_lock(&group->avgs_lock); 575 576 now = sched_clock(); 577 578 collect_percpu_times(group, PSI_AVGS, &changed_states); 579 /* 580 * If there is task activity, periodically fold the per-cpu 581 * times and feed samples into the running averages. If things 582 * are idle and there is no data to process, stop the clock. 583 * Once restarted, we'll catch up the running averages in one 584 * go - see calc_avgs() and missed_periods. 585 */ 586 if (now >= group->avg_next_update) { 587 update_triggers(group, now, &update_total, PSI_AVGS); 588 group->avg_next_update = update_averages(group, now); 589 } 590 591 if (changed_states & PSI_STATE_RESCHEDULE) { 592 schedule_delayed_work(dwork, nsecs_to_jiffies( 593 group->avg_next_update - now) + 1); 594 } 595 596 mutex_unlock(&group->avgs_lock); 597 } 598 599 static void init_rtpoll_triggers(struct psi_group *group, u64 now) 600 { 601 struct psi_trigger *t; 602 603 list_for_each_entry(t, &group->rtpoll_triggers, node) 604 window_reset(&t->win, now, 605 group->total[PSI_POLL][t->state], 0); 606 memcpy(group->rtpoll_total, group->total[PSI_POLL], 607 sizeof(group->rtpoll_total)); 608 group->rtpoll_next_update = now + group->rtpoll_min_period; 609 } 610 611 /* Schedule polling if it's not already scheduled or forced. */ 612 static void psi_schedule_rtpoll_work(struct psi_group *group, unsigned long delay, 613 bool force) 614 { 615 struct task_struct *task; 616 617 /* 618 * atomic_xchg should be called even when !force to provide a 619 * full memory barrier (see the comment inside psi_rtpoll_work). 620 */ 621 if (atomic_xchg(&group->rtpoll_scheduled, 1) && !force) 622 return; 623 624 rcu_read_lock(); 625 626 task = rcu_dereference(group->rtpoll_task); 627 /* 628 * kworker might be NULL in case psi_trigger_destroy races with 629 * psi_task_change (hotpath) which can't use locks 630 */ 631 if (likely(task)) 632 mod_timer(&group->rtpoll_timer, jiffies + delay); 633 else 634 atomic_set(&group->rtpoll_scheduled, 0); 635 636 rcu_read_unlock(); 637 } 638 639 static void psi_rtpoll_work(struct psi_group *group) 640 { 641 bool force_reschedule = false; 642 u32 changed_states; 643 bool update_total; 644 u64 now; 645 646 mutex_lock(&group->rtpoll_trigger_lock); 647 648 now = sched_clock(); 649 650 if (now > group->rtpoll_until) { 651 /* 652 * We are either about to start or might stop polling if no 653 * state change was recorded. Resetting poll_scheduled leaves 654 * a small window for psi_group_change to sneak in and schedule 655 * an immediate poll_work before we get to rescheduling. One 656 * potential extra wakeup at the end of the polling window 657 * should be negligible and polling_next_update still keeps 658 * updates correctly on schedule. 659 */ 660 atomic_set(&group->rtpoll_scheduled, 0); 661 /* 662 * A task change can race with the poll worker that is supposed to 663 * report on it. To avoid missing events, ensure ordering between 664 * poll_scheduled and the task state accesses, such that if the poll 665 * worker misses the state update, the task change is guaranteed to 666 * reschedule the poll worker: 667 * 668 * poll worker: 669 * atomic_set(poll_scheduled, 0) 670 * smp_mb() 671 * LOAD states 672 * 673 * task change: 674 * STORE states 675 * if atomic_xchg(poll_scheduled, 1) == 0: 676 * schedule poll worker 677 * 678 * The atomic_xchg() implies a full barrier. 679 */ 680 smp_mb(); 681 } else { 682 /* Polling window is not over, keep rescheduling */ 683 force_reschedule = true; 684 } 685 686 687 collect_percpu_times(group, PSI_POLL, &changed_states); 688 689 if (changed_states & group->rtpoll_states) { 690 /* Initialize trigger windows when entering polling mode */ 691 if (now > group->rtpoll_until) 692 init_rtpoll_triggers(group, now); 693 694 /* 695 * Keep the monitor active for at least the duration of the 696 * minimum tracking window as long as monitor states are 697 * changing. 698 */ 699 group->rtpoll_until = now + 700 group->rtpoll_min_period * UPDATES_PER_WINDOW; 701 } 702 703 if (now > group->rtpoll_until) { 704 group->rtpoll_next_update = ULLONG_MAX; 705 goto out; 706 } 707 708 if (now >= group->rtpoll_next_update) { 709 group->rtpoll_next_update = update_triggers(group, now, &update_total, PSI_POLL); 710 if (update_total) 711 memcpy(group->rtpoll_total, group->total[PSI_POLL], 712 sizeof(group->rtpoll_total)); 713 } 714 715 psi_schedule_rtpoll_work(group, 716 nsecs_to_jiffies(group->rtpoll_next_update - now) + 1, 717 force_reschedule); 718 719 out: 720 mutex_unlock(&group->rtpoll_trigger_lock); 721 } 722 723 static int psi_rtpoll_worker(void *data) 724 { 725 struct psi_group *group = (struct psi_group *)data; 726 727 sched_set_fifo_low(current); 728 729 while (true) { 730 wait_event_interruptible(group->rtpoll_wait, 731 atomic_cmpxchg(&group->rtpoll_wakeup, 1, 0) || 732 kthread_should_stop()); 733 if (kthread_should_stop()) 734 break; 735 736 psi_rtpoll_work(group); 737 } 738 return 0; 739 } 740 741 static void poll_timer_fn(struct timer_list *t) 742 { 743 struct psi_group *group = from_timer(group, t, rtpoll_timer); 744 745 atomic_set(&group->rtpoll_wakeup, 1); 746 wake_up_interruptible(&group->rtpoll_wait); 747 } 748 749 static void record_times(struct psi_group_cpu *groupc, u64 now) 750 { 751 u32 delta; 752 753 delta = now - groupc->state_start; 754 groupc->state_start = now; 755 756 if (groupc->state_mask & (1 << PSI_IO_SOME)) { 757 groupc->times[PSI_IO_SOME] += delta; 758 if (groupc->state_mask & (1 << PSI_IO_FULL)) 759 groupc->times[PSI_IO_FULL] += delta; 760 } 761 762 if (groupc->state_mask & (1 << PSI_MEM_SOME)) { 763 groupc->times[PSI_MEM_SOME] += delta; 764 if (groupc->state_mask & (1 << PSI_MEM_FULL)) 765 groupc->times[PSI_MEM_FULL] += delta; 766 } 767 768 if (groupc->state_mask & (1 << PSI_CPU_SOME)) { 769 groupc->times[PSI_CPU_SOME] += delta; 770 if (groupc->state_mask & (1 << PSI_CPU_FULL)) 771 groupc->times[PSI_CPU_FULL] += delta; 772 } 773 774 if (groupc->state_mask & (1 << PSI_NONIDLE)) 775 groupc->times[PSI_NONIDLE] += delta; 776 } 777 778 static void psi_group_change(struct psi_group *group, int cpu, 779 unsigned int clear, unsigned int set, u64 now, 780 bool wake_clock) 781 { 782 struct psi_group_cpu *groupc; 783 unsigned int t, m; 784 enum psi_states s; 785 u32 state_mask; 786 787 lockdep_assert_rq_held(cpu_rq(cpu)); 788 groupc = per_cpu_ptr(group->pcpu, cpu); 789 790 /* 791 * First we update the task counts according to the state 792 * change requested through the @clear and @set bits. 793 * 794 * Then if the cgroup PSI stats accounting enabled, we 795 * assess the aggregate resource states this CPU's tasks 796 * have been in since the last change, and account any 797 * SOME and FULL time these may have resulted in. 798 */ 799 write_seqcount_begin(&groupc->seq); 800 801 /* 802 * Start with TSK_ONCPU, which doesn't have a corresponding 803 * task count - it's just a boolean flag directly encoded in 804 * the state mask. Clear, set, or carry the current state if 805 * no changes are requested. 806 */ 807 if (unlikely(clear & TSK_ONCPU)) { 808 state_mask = 0; 809 clear &= ~TSK_ONCPU; 810 } else if (unlikely(set & TSK_ONCPU)) { 811 state_mask = PSI_ONCPU; 812 set &= ~TSK_ONCPU; 813 } else { 814 state_mask = groupc->state_mask & PSI_ONCPU; 815 } 816 817 /* 818 * The rest of the state mask is calculated based on the task 819 * counts. Update those first, then construct the mask. 820 */ 821 for (t = 0, m = clear; m; m &= ~(1 << t), t++) { 822 if (!(m & (1 << t))) 823 continue; 824 if (groupc->tasks[t]) { 825 groupc->tasks[t]--; 826 } else if (!psi_bug) { 827 printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n", 828 cpu, t, groupc->tasks[0], 829 groupc->tasks[1], groupc->tasks[2], 830 groupc->tasks[3], clear, set); 831 psi_bug = 1; 832 } 833 } 834 835 for (t = 0; set; set &= ~(1 << t), t++) 836 if (set & (1 << t)) 837 groupc->tasks[t]++; 838 839 if (!group->enabled) { 840 /* 841 * On the first group change after disabling PSI, conclude 842 * the current state and flush its time. This is unlikely 843 * to matter to the user, but aggregation (get_recent_times) 844 * may have already incorporated the live state into times_prev; 845 * avoid a delta sample underflow when PSI is later re-enabled. 846 */ 847 if (unlikely(groupc->state_mask & (1 << PSI_NONIDLE))) 848 record_times(groupc, now); 849 850 groupc->state_mask = state_mask; 851 852 write_seqcount_end(&groupc->seq); 853 return; 854 } 855 856 for (s = 0; s < NR_PSI_STATES; s++) { 857 if (test_state(groupc->tasks, s, state_mask & PSI_ONCPU)) 858 state_mask |= (1 << s); 859 } 860 861 /* 862 * Since we care about lost potential, a memstall is FULL 863 * when there are no other working tasks, but also when 864 * the CPU is actively reclaiming and nothing productive 865 * could run even if it were runnable. So when the current 866 * task in a cgroup is in_memstall, the corresponding groupc 867 * on that cpu is in PSI_MEM_FULL state. 868 */ 869 if (unlikely((state_mask & PSI_ONCPU) && cpu_curr(cpu)->in_memstall)) 870 state_mask |= (1 << PSI_MEM_FULL); 871 872 record_times(groupc, now); 873 874 groupc->state_mask = state_mask; 875 876 write_seqcount_end(&groupc->seq); 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 } 884 885 static 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 } 893 894 static 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 } 908 909 void psi_task_change(struct task_struct *task, int clear, int set) 910 { 911 int cpu = task_cpu(task); 912 struct psi_group *group; 913 u64 now; 914 915 if (!task->pid) 916 return; 917 918 psi_flags_change(task, clear, set); 919 920 now = cpu_clock(cpu); 921 922 group = task_psi_group(task); 923 do { 924 psi_group_change(group, cpu, clear, set, now, true); 925 } while ((group = group->parent)); 926 } 927 928 void psi_task_switch(struct task_struct *prev, struct task_struct *next, 929 bool sleep) 930 { 931 struct psi_group *group, *common = NULL; 932 int cpu = task_cpu(prev); 933 u64 now = cpu_clock(cpu); 934 935 if (next->pid) { 936 psi_flags_change(next, 0, TSK_ONCPU); 937 /* 938 * Set TSK_ONCPU on @next's cgroups. If @next shares any 939 * ancestors with @prev, those will already have @prev's 940 * TSK_ONCPU bit set, and we can stop the iteration there. 941 */ 942 group = task_psi_group(next); 943 do { 944 if (per_cpu_ptr(group->pcpu, cpu)->state_mask & 945 PSI_ONCPU) { 946 common = group; 947 break; 948 } 949 950 psi_group_change(group, cpu, 0, TSK_ONCPU, now, true); 951 } while ((group = group->parent)); 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 group = task_psi_group(prev); 985 do { 986 if (group == common) 987 break; 988 psi_group_change(group, cpu, clear, set, now, wake_clock); 989 } while ((group = group->parent)); 990 991 /* 992 * TSK_ONCPU is handled up to the common ancestor. If there are 993 * any other differences between the two tasks (e.g. prev goes 994 * to sleep, or only one task is memstall), finish propagating 995 * those differences all the way up to the root. 996 */ 997 if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) { 998 clear &= ~TSK_ONCPU; 999 for (; group; group = group->parent) 1000 psi_group_change(group, cpu, clear, set, now, wake_clock); 1001 } 1002 } 1003 } 1004 1005 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1006 void 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 *group; 1010 struct psi_group_cpu *groupc; 1011 u64 now, irq; 1012 s64 delta; 1013 1014 if (!curr->pid) 1015 return; 1016 1017 lockdep_assert_rq_held(rq); 1018 group = task_psi_group(curr); 1019 if (prev && task_psi_group(prev) == group) 1020 return; 1021 1022 now = cpu_clock(cpu); 1023 irq = irq_time_read(cpu); 1024 delta = (s64)(irq - rq->psi_irq_time); 1025 if (delta < 0) 1026 return; 1027 rq->psi_irq_time = irq; 1028 1029 do { 1030 if (!group->enabled) 1031 continue; 1032 1033 groupc = per_cpu_ptr(group->pcpu, cpu); 1034 1035 write_seqcount_begin(&groupc->seq); 1036 1037 record_times(groupc, now); 1038 groupc->times[PSI_IRQ_FULL] += delta; 1039 1040 write_seqcount_end(&groupc->seq); 1041 1042 if (group->rtpoll_states & (1 << PSI_IRQ_FULL)) 1043 psi_schedule_rtpoll_work(group, 1, false); 1044 } while ((group = group->parent)); 1045 } 1046 #endif 1047 1048 /** 1049 * psi_memstall_enter - mark the beginning of a memory stall section 1050 * @flags: flags to handle nested sections 1051 * 1052 * Marks the calling task as being stalled due to a lack of memory, 1053 * such as waiting for a refault or performing reclaim. 1054 */ 1055 void psi_memstall_enter(unsigned long *flags) 1056 { 1057 struct rq_flags rf; 1058 struct rq *rq; 1059 1060 if (static_branch_likely(&psi_disabled)) 1061 return; 1062 1063 *flags = current->in_memstall; 1064 if (*flags) 1065 return; 1066 /* 1067 * in_memstall setting & accounting needs to be atomic wrt 1068 * changes to the task's scheduling state, otherwise we can 1069 * race with CPU migration. 1070 */ 1071 rq = this_rq_lock_irq(&rf); 1072 1073 current->in_memstall = 1; 1074 psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING); 1075 1076 rq_unlock_irq(rq, &rf); 1077 } 1078 EXPORT_SYMBOL_GPL(psi_memstall_enter); 1079 1080 /** 1081 * psi_memstall_leave - mark the end of an memory stall section 1082 * @flags: flags to handle nested memdelay sections 1083 * 1084 * Marks the calling task as no longer stalled due to lack of memory. 1085 */ 1086 void psi_memstall_leave(unsigned long *flags) 1087 { 1088 struct rq_flags rf; 1089 struct rq *rq; 1090 1091 if (static_branch_likely(&psi_disabled)) 1092 return; 1093 1094 if (*flags) 1095 return; 1096 /* 1097 * in_memstall clearing & accounting needs to be atomic wrt 1098 * changes to the task's scheduling state, otherwise we could 1099 * race with CPU migration. 1100 */ 1101 rq = this_rq_lock_irq(&rf); 1102 1103 current->in_memstall = 0; 1104 psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0); 1105 1106 rq_unlock_irq(rq, &rf); 1107 } 1108 EXPORT_SYMBOL_GPL(psi_memstall_leave); 1109 1110 #ifdef CONFIG_CGROUPS 1111 int psi_cgroup_alloc(struct cgroup *cgroup) 1112 { 1113 if (!static_branch_likely(&psi_cgroups_enabled)) 1114 return 0; 1115 1116 cgroup->psi = kzalloc(sizeof(struct psi_group), GFP_KERNEL); 1117 if (!cgroup->psi) 1118 return -ENOMEM; 1119 1120 cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu); 1121 if (!cgroup->psi->pcpu) { 1122 kfree(cgroup->psi); 1123 return -ENOMEM; 1124 } 1125 group_init(cgroup->psi); 1126 cgroup->psi->parent = cgroup_psi(cgroup_parent(cgroup)); 1127 return 0; 1128 } 1129 1130 void psi_cgroup_free(struct cgroup *cgroup) 1131 { 1132 if (!static_branch_likely(&psi_cgroups_enabled)) 1133 return; 1134 1135 cancel_delayed_work_sync(&cgroup->psi->avgs_work); 1136 free_percpu(cgroup->psi->pcpu); 1137 /* All triggers must be removed by now */ 1138 WARN_ONCE(cgroup->psi->rtpoll_states, "psi: trigger leak\n"); 1139 kfree(cgroup->psi); 1140 } 1141 1142 /** 1143 * cgroup_move_task - move task to a different cgroup 1144 * @task: the task 1145 * @to: the target css_set 1146 * 1147 * Move task to a new cgroup and safely migrate its associated stall 1148 * state between the different groups. 1149 * 1150 * This function acquires the task's rq lock to lock out concurrent 1151 * changes to the task's scheduling state and - in case the task is 1152 * running - concurrent changes to its stall state. 1153 */ 1154 void cgroup_move_task(struct task_struct *task, struct css_set *to) 1155 { 1156 unsigned int task_flags; 1157 struct rq_flags rf; 1158 struct rq *rq; 1159 1160 if (!static_branch_likely(&psi_cgroups_enabled)) { 1161 /* 1162 * Lame to do this here, but the scheduler cannot be locked 1163 * from the outside, so we move cgroups from inside sched/. 1164 */ 1165 rcu_assign_pointer(task->cgroups, to); 1166 return; 1167 } 1168 1169 rq = task_rq_lock(task, &rf); 1170 1171 /* 1172 * We may race with schedule() dropping the rq lock between 1173 * deactivating prev and switching to next. Because the psi 1174 * updates from the deactivation are deferred to the switch 1175 * callback to save cgroup tree updates, the task's scheduling 1176 * state here is not coherent with its psi state: 1177 * 1178 * schedule() cgroup_move_task() 1179 * rq_lock() 1180 * deactivate_task() 1181 * p->on_rq = 0 1182 * psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates 1183 * pick_next_task() 1184 * rq_unlock() 1185 * rq_lock() 1186 * psi_task_change() // old cgroup 1187 * task->cgroups = to 1188 * psi_task_change() // new cgroup 1189 * rq_unlock() 1190 * rq_lock() 1191 * psi_sched_switch() // does deferred updates in new cgroup 1192 * 1193 * Don't rely on the scheduling state. Use psi_flags instead. 1194 */ 1195 task_flags = task->psi_flags; 1196 1197 if (task_flags) 1198 psi_task_change(task, task_flags, 0); 1199 1200 /* See comment above */ 1201 rcu_assign_pointer(task->cgroups, to); 1202 1203 if (task_flags) 1204 psi_task_change(task, 0, task_flags); 1205 1206 task_rq_unlock(rq, task, &rf); 1207 } 1208 1209 void psi_cgroup_restart(struct psi_group *group) 1210 { 1211 int cpu; 1212 1213 /* 1214 * After we disable psi_group->enabled, we don't actually 1215 * stop percpu tasks accounting in each psi_group_cpu, 1216 * instead only stop test_state() loop, record_times() 1217 * and averaging worker, see psi_group_change() for details. 1218 * 1219 * When disable cgroup PSI, this function has nothing to sync 1220 * since cgroup pressure files are hidden and percpu psi_group_cpu 1221 * would see !psi_group->enabled and only do task accounting. 1222 * 1223 * When re-enable cgroup PSI, this function use psi_group_change() 1224 * to get correct state mask from test_state() loop on tasks[], 1225 * and restart groupc->state_start from now, use .clear = .set = 0 1226 * here since no task status really changed. 1227 */ 1228 if (!group->enabled) 1229 return; 1230 1231 for_each_possible_cpu(cpu) { 1232 struct rq *rq = cpu_rq(cpu); 1233 struct rq_flags rf; 1234 u64 now; 1235 1236 rq_lock_irq(rq, &rf); 1237 now = cpu_clock(cpu); 1238 psi_group_change(group, cpu, 0, 0, now, true); 1239 rq_unlock_irq(rq, &rf); 1240 } 1241 } 1242 #endif /* CONFIG_CGROUPS */ 1243 1244 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res) 1245 { 1246 bool only_full = false; 1247 int full; 1248 u64 now; 1249 1250 if (static_branch_likely(&psi_disabled)) 1251 return -EOPNOTSUPP; 1252 1253 /* Update averages before reporting them */ 1254 mutex_lock(&group->avgs_lock); 1255 now = sched_clock(); 1256 collect_percpu_times(group, PSI_AVGS, NULL); 1257 if (now >= group->avg_next_update) 1258 group->avg_next_update = update_averages(group, now); 1259 mutex_unlock(&group->avgs_lock); 1260 1261 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1262 only_full = res == PSI_IRQ; 1263 #endif 1264 1265 for (full = 0; full < 2 - only_full; full++) { 1266 unsigned long avg[3] = { 0, }; 1267 u64 total = 0; 1268 int w; 1269 1270 /* CPU FULL is undefined at the system level */ 1271 if (!(group == &psi_system && res == PSI_CPU && full)) { 1272 for (w = 0; w < 3; w++) 1273 avg[w] = group->avg[res * 2 + full][w]; 1274 total = div_u64(group->total[PSI_AVGS][res * 2 + full], 1275 NSEC_PER_USEC); 1276 } 1277 1278 seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n", 1279 full || only_full ? "full" : "some", 1280 LOAD_INT(avg[0]), LOAD_FRAC(avg[0]), 1281 LOAD_INT(avg[1]), LOAD_FRAC(avg[1]), 1282 LOAD_INT(avg[2]), LOAD_FRAC(avg[2]), 1283 total); 1284 } 1285 1286 return 0; 1287 } 1288 1289 struct psi_trigger *psi_trigger_create(struct psi_group *group, char *buf, 1290 enum psi_res res, struct file *file, 1291 struct kernfs_open_file *of) 1292 { 1293 struct psi_trigger *t; 1294 enum psi_states state; 1295 u32 threshold_us; 1296 bool privileged; 1297 u32 window_us; 1298 1299 if (static_branch_likely(&psi_disabled)) 1300 return ERR_PTR(-EOPNOTSUPP); 1301 1302 /* 1303 * Checking the privilege here on file->f_cred implies that a privileged user 1304 * could open the file and delegate the write to an unprivileged one. 1305 */ 1306 privileged = cap_raised(file->f_cred->cap_effective, CAP_SYS_RESOURCE); 1307 1308 if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2) 1309 state = PSI_IO_SOME + res * 2; 1310 else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2) 1311 state = PSI_IO_FULL + res * 2; 1312 else 1313 return ERR_PTR(-EINVAL); 1314 1315 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1316 if (res == PSI_IRQ && --state != PSI_IRQ_FULL) 1317 return ERR_PTR(-EINVAL); 1318 #endif 1319 1320 if (state >= PSI_NONIDLE) 1321 return ERR_PTR(-EINVAL); 1322 1323 if (window_us == 0 || window_us > WINDOW_MAX_US) 1324 return ERR_PTR(-EINVAL); 1325 1326 /* 1327 * Unprivileged users can only use 2s windows so that averages aggregation 1328 * work is used, and no RT threads need to be spawned. 1329 */ 1330 if (!privileged && window_us % 2000000) 1331 return ERR_PTR(-EINVAL); 1332 1333 /* Check threshold */ 1334 if (threshold_us == 0 || threshold_us > window_us) 1335 return ERR_PTR(-EINVAL); 1336 1337 t = kmalloc(sizeof(*t), GFP_KERNEL); 1338 if (!t) 1339 return ERR_PTR(-ENOMEM); 1340 1341 t->group = group; 1342 t->state = state; 1343 t->threshold = threshold_us * NSEC_PER_USEC; 1344 t->win.size = window_us * NSEC_PER_USEC; 1345 window_reset(&t->win, sched_clock(), 1346 group->total[PSI_POLL][t->state], 0); 1347 1348 t->event = 0; 1349 t->last_event_time = 0; 1350 t->of = of; 1351 if (!of) 1352 init_waitqueue_head(&t->event_wait); 1353 t->pending_event = false; 1354 t->aggregator = privileged ? PSI_POLL : PSI_AVGS; 1355 1356 if (privileged) { 1357 mutex_lock(&group->rtpoll_trigger_lock); 1358 1359 if (!rcu_access_pointer(group->rtpoll_task)) { 1360 struct task_struct *task; 1361 1362 task = kthread_create(psi_rtpoll_worker, group, "psimon"); 1363 if (IS_ERR(task)) { 1364 kfree(t); 1365 mutex_unlock(&group->rtpoll_trigger_lock); 1366 return ERR_CAST(task); 1367 } 1368 atomic_set(&group->rtpoll_wakeup, 0); 1369 wake_up_process(task); 1370 rcu_assign_pointer(group->rtpoll_task, task); 1371 } 1372 1373 list_add(&t->node, &group->rtpoll_triggers); 1374 group->rtpoll_min_period = min(group->rtpoll_min_period, 1375 div_u64(t->win.size, UPDATES_PER_WINDOW)); 1376 group->rtpoll_nr_triggers[t->state]++; 1377 group->rtpoll_states |= (1 << t->state); 1378 1379 mutex_unlock(&group->rtpoll_trigger_lock); 1380 } else { 1381 mutex_lock(&group->avgs_lock); 1382 1383 list_add(&t->node, &group->avg_triggers); 1384 group->avg_nr_triggers[t->state]++; 1385 1386 mutex_unlock(&group->avgs_lock); 1387 } 1388 return t; 1389 } 1390 1391 void psi_trigger_destroy(struct psi_trigger *t) 1392 { 1393 struct psi_group *group; 1394 struct task_struct *task_to_destroy = NULL; 1395 1396 /* 1397 * We do not check psi_disabled since it might have been disabled after 1398 * the trigger got created. 1399 */ 1400 if (!t) 1401 return; 1402 1403 group = t->group; 1404 /* 1405 * Wakeup waiters to stop polling and clear the queue to prevent it from 1406 * being accessed later. Can happen if cgroup is deleted from under a 1407 * polling process. 1408 */ 1409 if (t->of) 1410 kernfs_notify(t->of->kn); 1411 else 1412 wake_up_interruptible(&t->event_wait); 1413 1414 if (t->aggregator == PSI_AVGS) { 1415 mutex_lock(&group->avgs_lock); 1416 if (!list_empty(&t->node)) { 1417 list_del(&t->node); 1418 group->avg_nr_triggers[t->state]--; 1419 } 1420 mutex_unlock(&group->avgs_lock); 1421 } else { 1422 mutex_lock(&group->rtpoll_trigger_lock); 1423 if (!list_empty(&t->node)) { 1424 struct psi_trigger *tmp; 1425 u64 period = ULLONG_MAX; 1426 1427 list_del(&t->node); 1428 group->rtpoll_nr_triggers[t->state]--; 1429 if (!group->rtpoll_nr_triggers[t->state]) 1430 group->rtpoll_states &= ~(1 << t->state); 1431 /* 1432 * Reset min update period for the remaining triggers 1433 * iff the destroying trigger had the min window size. 1434 */ 1435 if (group->rtpoll_min_period == div_u64(t->win.size, UPDATES_PER_WINDOW)) { 1436 list_for_each_entry(tmp, &group->rtpoll_triggers, node) 1437 period = min(period, div_u64(tmp->win.size, 1438 UPDATES_PER_WINDOW)); 1439 group->rtpoll_min_period = period; 1440 } 1441 /* Destroy rtpoll_task when the last trigger is destroyed */ 1442 if (group->rtpoll_states == 0) { 1443 group->rtpoll_until = 0; 1444 task_to_destroy = rcu_dereference_protected( 1445 group->rtpoll_task, 1446 lockdep_is_held(&group->rtpoll_trigger_lock)); 1447 rcu_assign_pointer(group->rtpoll_task, NULL); 1448 del_timer(&group->rtpoll_timer); 1449 } 1450 } 1451 mutex_unlock(&group->rtpoll_trigger_lock); 1452 } 1453 1454 /* 1455 * Wait for psi_schedule_rtpoll_work RCU to complete its read-side 1456 * critical section before destroying the trigger and optionally the 1457 * rtpoll_task. 1458 */ 1459 synchronize_rcu(); 1460 /* 1461 * Stop kthread 'psimon' after releasing rtpoll_trigger_lock to prevent 1462 * a deadlock while waiting for psi_rtpoll_work to acquire 1463 * rtpoll_trigger_lock 1464 */ 1465 if (task_to_destroy) { 1466 /* 1467 * After the RCU grace period has expired, the worker 1468 * can no longer be found through group->rtpoll_task. 1469 */ 1470 kthread_stop(task_to_destroy); 1471 atomic_set(&group->rtpoll_scheduled, 0); 1472 } 1473 kfree(t); 1474 } 1475 1476 __poll_t psi_trigger_poll(void **trigger_ptr, 1477 struct file *file, poll_table *wait) 1478 { 1479 __poll_t ret = DEFAULT_POLLMASK; 1480 struct psi_trigger *t; 1481 1482 if (static_branch_likely(&psi_disabled)) 1483 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI; 1484 1485 t = smp_load_acquire(trigger_ptr); 1486 if (!t) 1487 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI; 1488 1489 if (t->of) 1490 kernfs_generic_poll(t->of, wait); 1491 else 1492 poll_wait(file, &t->event_wait, wait); 1493 1494 if (cmpxchg(&t->event, 1, 0) == 1) 1495 ret |= EPOLLPRI; 1496 1497 return ret; 1498 } 1499 1500 #ifdef CONFIG_PROC_FS 1501 static int psi_io_show(struct seq_file *m, void *v) 1502 { 1503 return psi_show(m, &psi_system, PSI_IO); 1504 } 1505 1506 static int psi_memory_show(struct seq_file *m, void *v) 1507 { 1508 return psi_show(m, &psi_system, PSI_MEM); 1509 } 1510 1511 static int psi_cpu_show(struct seq_file *m, void *v) 1512 { 1513 return psi_show(m, &psi_system, PSI_CPU); 1514 } 1515 1516 static int psi_io_open(struct inode *inode, struct file *file) 1517 { 1518 return single_open(file, psi_io_show, NULL); 1519 } 1520 1521 static int psi_memory_open(struct inode *inode, struct file *file) 1522 { 1523 return single_open(file, psi_memory_show, NULL); 1524 } 1525 1526 static int psi_cpu_open(struct inode *inode, struct file *file) 1527 { 1528 return single_open(file, psi_cpu_show, NULL); 1529 } 1530 1531 static ssize_t psi_write(struct file *file, const char __user *user_buf, 1532 size_t nbytes, enum psi_res res) 1533 { 1534 char buf[32]; 1535 size_t buf_size; 1536 struct seq_file *seq; 1537 struct psi_trigger *new; 1538 1539 if (static_branch_likely(&psi_disabled)) 1540 return -EOPNOTSUPP; 1541 1542 if (!nbytes) 1543 return -EINVAL; 1544 1545 buf_size = min(nbytes, sizeof(buf)); 1546 if (copy_from_user(buf, user_buf, buf_size)) 1547 return -EFAULT; 1548 1549 buf[buf_size - 1] = '\0'; 1550 1551 seq = file->private_data; 1552 1553 /* Take seq->lock to protect seq->private from concurrent writes */ 1554 mutex_lock(&seq->lock); 1555 1556 /* Allow only one trigger per file descriptor */ 1557 if (seq->private) { 1558 mutex_unlock(&seq->lock); 1559 return -EBUSY; 1560 } 1561 1562 new = psi_trigger_create(&psi_system, buf, res, file, NULL); 1563 if (IS_ERR(new)) { 1564 mutex_unlock(&seq->lock); 1565 return PTR_ERR(new); 1566 } 1567 1568 smp_store_release(&seq->private, new); 1569 mutex_unlock(&seq->lock); 1570 1571 return nbytes; 1572 } 1573 1574 static ssize_t psi_io_write(struct file *file, const char __user *user_buf, 1575 size_t nbytes, loff_t *ppos) 1576 { 1577 return psi_write(file, user_buf, nbytes, PSI_IO); 1578 } 1579 1580 static ssize_t psi_memory_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_MEM); 1584 } 1585 1586 static ssize_t psi_cpu_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_CPU); 1590 } 1591 1592 static __poll_t psi_fop_poll(struct file *file, poll_table *wait) 1593 { 1594 struct seq_file *seq = file->private_data; 1595 1596 return psi_trigger_poll(&seq->private, file, wait); 1597 } 1598 1599 static int psi_fop_release(struct inode *inode, struct file *file) 1600 { 1601 struct seq_file *seq = file->private_data; 1602 1603 psi_trigger_destroy(seq->private); 1604 return single_release(inode, file); 1605 } 1606 1607 static const struct proc_ops psi_io_proc_ops = { 1608 .proc_open = psi_io_open, 1609 .proc_read = seq_read, 1610 .proc_lseek = seq_lseek, 1611 .proc_write = psi_io_write, 1612 .proc_poll = psi_fop_poll, 1613 .proc_release = psi_fop_release, 1614 }; 1615 1616 static const struct proc_ops psi_memory_proc_ops = { 1617 .proc_open = psi_memory_open, 1618 .proc_read = seq_read, 1619 .proc_lseek = seq_lseek, 1620 .proc_write = psi_memory_write, 1621 .proc_poll = psi_fop_poll, 1622 .proc_release = psi_fop_release, 1623 }; 1624 1625 static const struct proc_ops psi_cpu_proc_ops = { 1626 .proc_open = psi_cpu_open, 1627 .proc_read = seq_read, 1628 .proc_lseek = seq_lseek, 1629 .proc_write = psi_cpu_write, 1630 .proc_poll = psi_fop_poll, 1631 .proc_release = psi_fop_release, 1632 }; 1633 1634 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1635 static int psi_irq_show(struct seq_file *m, void *v) 1636 { 1637 return psi_show(m, &psi_system, PSI_IRQ); 1638 } 1639 1640 static int psi_irq_open(struct inode *inode, struct file *file) 1641 { 1642 return single_open(file, psi_irq_show, NULL); 1643 } 1644 1645 static ssize_t psi_irq_write(struct file *file, const char __user *user_buf, 1646 size_t nbytes, loff_t *ppos) 1647 { 1648 return psi_write(file, user_buf, nbytes, PSI_IRQ); 1649 } 1650 1651 static const struct proc_ops psi_irq_proc_ops = { 1652 .proc_open = psi_irq_open, 1653 .proc_read = seq_read, 1654 .proc_lseek = seq_lseek, 1655 .proc_write = psi_irq_write, 1656 .proc_poll = psi_fop_poll, 1657 .proc_release = psi_fop_release, 1658 }; 1659 #endif 1660 1661 static int __init psi_proc_init(void) 1662 { 1663 if (psi_enable) { 1664 proc_mkdir("pressure", NULL); 1665 proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops); 1666 proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops); 1667 proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops); 1668 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1669 proc_create("pressure/irq", 0666, NULL, &psi_irq_proc_ops); 1670 #endif 1671 } 1672 return 0; 1673 } 1674 module_init(psi_proc_init); 1675 1676 #endif /* CONFIG_PROC_FS */ 1677