1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Pressure stall information for CPU, memory and IO 4 * 5 * Copyright (c) 2018 Facebook, Inc. 6 * Author: Johannes Weiner <hannes@cmpxchg.org> 7 * 8 * Polling support by Suren Baghdasaryan <surenb@google.com> 9 * Copyright (c) 2018 Google, Inc. 10 * 11 * When CPU, memory and IO are contended, tasks experience delays that 12 * reduce throughput and introduce latencies into the workload. Memory 13 * and IO contention, in addition, can cause a full loss of forward 14 * progress in which the CPU goes idle. 15 * 16 * This code aggregates individual task delays into resource pressure 17 * metrics that indicate problems with both workload health and 18 * resource utilization. 19 * 20 * Model 21 * 22 * The time in which a task can execute on a CPU is our baseline for 23 * productivity. Pressure expresses the amount of time in which this 24 * potential cannot be realized due to resource contention. 25 * 26 * This concept of productivity has two components: the workload and 27 * the CPU. To measure the impact of pressure on both, we define two 28 * contention states for a resource: SOME and FULL. 29 * 30 * In the SOME state of a given resource, one or more tasks are 31 * delayed on that resource. This affects the workload's ability to 32 * perform work, but the CPU may still be executing other tasks. 33 * 34 * In the FULL state of a given resource, all non-idle tasks are 35 * delayed on that resource such that nobody is advancing and the CPU 36 * goes idle. This leaves both workload and CPU unproductive. 37 * 38 * SOME = nr_delayed_tasks != 0 39 * FULL = nr_delayed_tasks != 0 && nr_productive_tasks == 0 40 * 41 * What it means for a task to be productive is defined differently 42 * for each resource. For IO, productive means a running task. For 43 * memory, productive means a running task that isn't a reclaimer. For 44 * CPU, productive means an on-CPU task. 45 * 46 * Naturally, the FULL state doesn't exist for the CPU resource at the 47 * system level, but exist at the cgroup level. At the cgroup level, 48 * FULL means all non-idle tasks in the cgroup are delayed on the CPU 49 * resource which is being used by others outside of the cgroup or 50 * throttled by the cgroup cpu.max configuration. 51 * 52 * The percentage of wall clock time spent in those compound stall 53 * states gives pressure numbers between 0 and 100 for each resource, 54 * where the SOME percentage indicates workload slowdowns and the FULL 55 * percentage indicates reduced CPU utilization: 56 * 57 * %SOME = time(SOME) / period 58 * %FULL = time(FULL) / period 59 * 60 * Multiple CPUs 61 * 62 * The more tasks and available CPUs there are, the more work can be 63 * performed concurrently. This means that the potential that can go 64 * unrealized due to resource contention *also* scales with non-idle 65 * tasks and CPUs. 66 * 67 * Consider a scenario where 257 number crunching tasks are trying to 68 * run concurrently on 256 CPUs. If we simply aggregated the task 69 * states, we would have to conclude a CPU SOME pressure number of 70 * 100%, since *somebody* is waiting on a runqueue at all 71 * times. However, that is clearly not the amount of contention the 72 * workload is experiencing: only one out of 256 possible execution 73 * threads will be contended at any given time, or about 0.4%. 74 * 75 * Conversely, consider a scenario of 4 tasks and 4 CPUs where at any 76 * given time *one* of the tasks is delayed due to a lack of memory. 77 * Again, looking purely at the task state would yield a memory FULL 78 * pressure number of 0%, since *somebody* is always making forward 79 * progress. But again this wouldn't capture the amount of execution 80 * potential lost, which is 1 out of 4 CPUs, or 25%. 81 * 82 * To calculate wasted potential (pressure) with multiple processors, 83 * we have to base our calculation on the number of non-idle tasks in 84 * conjunction with the number of available CPUs, which is the number 85 * of potential execution threads. SOME becomes then the proportion of 86 * delayed tasks to possible threads, and FULL is the share of possible 87 * threads that are unproductive due to delays: 88 * 89 * threads = min(nr_nonidle_tasks, nr_cpus) 90 * SOME = min(nr_delayed_tasks / threads, 1) 91 * FULL = (threads - min(nr_productive_tasks, threads)) / threads 92 * 93 * For the 257 number crunchers on 256 CPUs, this yields: 94 * 95 * threads = min(257, 256) 96 * SOME = min(1 / 256, 1) = 0.4% 97 * FULL = (256 - min(256, 256)) / 256 = 0% 98 * 99 * For the 1 out of 4 memory-delayed tasks, this yields: 100 * 101 * threads = min(4, 4) 102 * SOME = min(1 / 4, 1) = 25% 103 * FULL = (4 - min(3, 4)) / 4 = 25% 104 * 105 * [ Substitute nr_cpus with 1, and you can see that it's a natural 106 * extension of the single-CPU model. ] 107 * 108 * Implementation 109 * 110 * To assess the precise time spent in each such state, we would have 111 * to freeze the system on task changes and start/stop the state 112 * clocks accordingly. Obviously that doesn't scale in practice. 113 * 114 * Because the scheduler aims to distribute the compute load evenly 115 * among the available CPUs, we can track task state locally to each 116 * CPU and, at much lower frequency, extrapolate the global state for 117 * the cumulative stall times and the running averages. 118 * 119 * For each runqueue, we track: 120 * 121 * tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0) 122 * tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_productive_tasks[cpu]) 123 * tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0) 124 * 125 * and then periodically aggregate: 126 * 127 * tNONIDLE = sum(tNONIDLE[i]) 128 * 129 * tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE 130 * tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE 131 * 132 * %SOME = tSOME / period 133 * %FULL = tFULL / period 134 * 135 * This gives us an approximation of pressure that is practical 136 * cost-wise, yet way more sensitive and accurate than periodic 137 * sampling of the aggregate task states would be. 138 */ 139 #include <linux/sched/clock.h> 140 #include <linux/workqueue.h> 141 #include <linux/psi.h> 142 #include "sched.h" 143 144 static int psi_bug __read_mostly; 145 146 DEFINE_STATIC_KEY_FALSE(psi_disabled); 147 static DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled); 148 149 #ifdef CONFIG_PSI_DEFAULT_DISABLED 150 static bool psi_enable; 151 #else 152 static bool psi_enable = true; 153 #endif 154 static int __init setup_psi(char *str) 155 { 156 return kstrtobool(str, &psi_enable) == 0; 157 } 158 __setup("psi=", setup_psi); 159 160 /* Running averages - we need to be higher-res than loadavg */ 161 #define PSI_FREQ (2*HZ+1) /* 2 sec intervals */ 162 #define EXP_10s 1677 /* 1/exp(2s/10s) as fixed-point */ 163 #define EXP_60s 1981 /* 1/exp(2s/60s) */ 164 #define EXP_300s 2034 /* 1/exp(2s/300s) */ 165 166 /* PSI trigger definitions */ 167 #define WINDOW_MAX_US 10000000 /* Max window size is 10s */ 168 #define UPDATES_PER_WINDOW 10 /* 10 updates per window */ 169 170 /* Sampling frequency in nanoseconds */ 171 static u64 psi_period __read_mostly; 172 173 /* System-level pressure and stall tracking */ 174 static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu); 175 struct psi_group psi_system = { 176 .pcpu = &system_group_pcpu, 177 }; 178 179 static void psi_avgs_work(struct work_struct *work); 180 181 static void poll_timer_fn(struct timer_list *t); 182 183 static void group_init(struct psi_group *group) 184 { 185 int cpu; 186 187 group->enabled = true; 188 for_each_possible_cpu(cpu) 189 seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq); 190 group->avg_last_update = sched_clock(); 191 group->avg_next_update = group->avg_last_update + psi_period; 192 mutex_init(&group->avgs_lock); 193 194 /* Init avg trigger-related members */ 195 INIT_LIST_HEAD(&group->avg_triggers); 196 memset(group->avg_nr_triggers, 0, sizeof(group->avg_nr_triggers)); 197 INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work); 198 199 /* Init rtpoll trigger-related members */ 200 atomic_set(&group->rtpoll_scheduled, 0); 201 mutex_init(&group->rtpoll_trigger_lock); 202 INIT_LIST_HEAD(&group->rtpoll_triggers); 203 group->rtpoll_min_period = U32_MAX; 204 group->rtpoll_next_update = ULLONG_MAX; 205 init_waitqueue_head(&group->rtpoll_wait); 206 timer_setup(&group->rtpoll_timer, poll_timer_fn, 0); 207 rcu_assign_pointer(group->rtpoll_task, NULL); 208 } 209 210 void __init psi_init(void) 211 { 212 if (!psi_enable) { 213 static_branch_enable(&psi_disabled); 214 static_branch_disable(&psi_cgroups_enabled); 215 return; 216 } 217 218 if (!cgroup_psi_enabled()) 219 static_branch_disable(&psi_cgroups_enabled); 220 221 psi_period = jiffies_to_nsecs(PSI_FREQ); 222 group_init(&psi_system); 223 } 224 225 static u32 test_states(unsigned int *tasks, u32 state_mask) 226 { 227 const bool oncpu = state_mask & PSI_ONCPU; 228 229 if (tasks[NR_IOWAIT]) { 230 state_mask |= BIT(PSI_IO_SOME); 231 if (!tasks[NR_RUNNING]) 232 state_mask |= BIT(PSI_IO_FULL); 233 } 234 235 if (tasks[NR_MEMSTALL]) { 236 state_mask |= BIT(PSI_MEM_SOME); 237 if (tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]) 238 state_mask |= BIT(PSI_MEM_FULL); 239 } 240 241 if (tasks[NR_RUNNING] > oncpu) 242 state_mask |= BIT(PSI_CPU_SOME); 243 244 if (tasks[NR_RUNNING] && !oncpu) 245 state_mask |= BIT(PSI_CPU_FULL); 246 247 if (tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] || tasks[NR_RUNNING]) 248 state_mask |= BIT(PSI_NONIDLE); 249 250 return state_mask; 251 } 252 253 static void get_recent_times(struct psi_group *group, int cpu, 254 enum psi_aggregators aggregator, u32 *times, 255 u32 *pchanged_states) 256 { 257 struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu); 258 int current_cpu = raw_smp_processor_id(); 259 unsigned int tasks[NR_PSI_TASK_COUNTS]; 260 u64 now, state_start; 261 enum psi_states s; 262 unsigned int seq; 263 u32 state_mask; 264 265 *pchanged_states = 0; 266 267 /* Snapshot a coherent view of the CPU state */ 268 do { 269 seq = read_seqcount_begin(&groupc->seq); 270 now = cpu_clock(cpu); 271 memcpy(times, groupc->times, sizeof(groupc->times)); 272 state_mask = groupc->state_mask; 273 state_start = groupc->state_start; 274 if (cpu == current_cpu) 275 memcpy(tasks, groupc->tasks, sizeof(groupc->tasks)); 276 } while (read_seqcount_retry(&groupc->seq, seq)); 277 278 /* Calculate state time deltas against the previous snapshot */ 279 for (s = 0; s < NR_PSI_STATES; s++) { 280 u32 delta; 281 /* 282 * In addition to already concluded states, we also 283 * incorporate currently active states on the CPU, 284 * since states may last for many sampling periods. 285 * 286 * This way we keep our delta sampling buckets small 287 * (u32) and our reported pressure close to what's 288 * actually happening. 289 */ 290 if (state_mask & (1 << s)) 291 times[s] += now - state_start; 292 293 delta = times[s] - groupc->times_prev[aggregator][s]; 294 groupc->times_prev[aggregator][s] = times[s]; 295 296 times[s] = delta; 297 if (delta) 298 *pchanged_states |= (1 << s); 299 } 300 301 /* 302 * When collect_percpu_times() from the avgs_work, we don't want to 303 * re-arm avgs_work when all CPUs are IDLE. But the current CPU running 304 * this avgs_work is never IDLE, cause avgs_work can't be shut off. 305 * So for the current CPU, we need to re-arm avgs_work only when 306 * (NR_RUNNING > 1 || NR_IOWAIT > 0 || NR_MEMSTALL > 0), for other CPUs 307 * we can just check PSI_NONIDLE delta. 308 */ 309 if (current_work() == &group->avgs_work.work) { 310 bool reschedule; 311 312 if (cpu == current_cpu) 313 reschedule = tasks[NR_RUNNING] + 314 tasks[NR_IOWAIT] + 315 tasks[NR_MEMSTALL] > 1; 316 else 317 reschedule = *pchanged_states & (1 << PSI_NONIDLE); 318 319 if (reschedule) 320 *pchanged_states |= PSI_STATE_RESCHEDULE; 321 } 322 } 323 324 static void calc_avgs(unsigned long avg[3], int missed_periods, 325 u64 time, u64 period) 326 { 327 unsigned long pct; 328 329 /* Fill in zeroes for periods of no activity */ 330 if (missed_periods) { 331 avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods); 332 avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods); 333 avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods); 334 } 335 336 /* Sample the most recent active period */ 337 pct = div_u64(time * 100, period); 338 pct *= FIXED_1; 339 avg[0] = calc_load(avg[0], EXP_10s, pct); 340 avg[1] = calc_load(avg[1], EXP_60s, pct); 341 avg[2] = calc_load(avg[2], EXP_300s, pct); 342 } 343 344 static void collect_percpu_times(struct psi_group *group, 345 enum psi_aggregators aggregator, 346 u32 *pchanged_states) 347 { 348 u64 deltas[NR_PSI_STATES - 1] = { 0, }; 349 unsigned long nonidle_total = 0; 350 u32 changed_states = 0; 351 int cpu; 352 int s; 353 354 /* 355 * Collect the per-cpu time buckets and average them into a 356 * single time sample that is normalized to wall clock time. 357 * 358 * For averaging, each CPU is weighted by its non-idle time in 359 * the sampling period. This eliminates artifacts from uneven 360 * loading, or even entirely idle CPUs. 361 */ 362 for_each_possible_cpu(cpu) { 363 u32 times[NR_PSI_STATES]; 364 u32 nonidle; 365 u32 cpu_changed_states; 366 367 get_recent_times(group, cpu, aggregator, times, 368 &cpu_changed_states); 369 changed_states |= cpu_changed_states; 370 371 nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]); 372 nonidle_total += nonidle; 373 374 for (s = 0; s < PSI_NONIDLE; s++) 375 deltas[s] += (u64)times[s] * nonidle; 376 } 377 378 /* 379 * Integrate the sample into the running statistics that are 380 * reported to userspace: the cumulative stall times and the 381 * decaying averages. 382 * 383 * Pressure percentages are sampled at PSI_FREQ. We might be 384 * called more often when the user polls more frequently than 385 * that; we might be called less often when there is no task 386 * activity, thus no data, and clock ticks are sporadic. The 387 * below handles both. 388 */ 389 390 /* total= */ 391 for (s = 0; s < NR_PSI_STATES - 1; s++) 392 group->total[aggregator][s] += 393 div_u64(deltas[s], max(nonidle_total, 1UL)); 394 395 if (pchanged_states) 396 *pchanged_states = changed_states; 397 } 398 399 /* Trigger tracking window manipulations */ 400 static void window_reset(struct psi_window *win, u64 now, u64 value, 401 u64 prev_growth) 402 { 403 win->start_time = now; 404 win->start_value = value; 405 win->prev_growth = prev_growth; 406 } 407 408 /* 409 * PSI growth tracking window update and growth calculation routine. 410 * 411 * This approximates a sliding tracking window by interpolating 412 * partially elapsed windows using historical growth data from the 413 * previous intervals. This minimizes memory requirements (by not storing 414 * all the intermediate values in the previous window) and simplifies 415 * the calculations. It works well because PSI signal changes only in 416 * positive direction and over relatively small window sizes the growth 417 * is close to linear. 418 */ 419 static u64 window_update(struct psi_window *win, u64 now, u64 value) 420 { 421 u64 elapsed; 422 u64 growth; 423 424 elapsed = now - win->start_time; 425 growth = value - win->start_value; 426 /* 427 * After each tracking window passes win->start_value and 428 * win->start_time get reset and win->prev_growth stores 429 * the average per-window growth of the previous window. 430 * win->prev_growth is then used to interpolate additional 431 * growth from the previous window assuming it was linear. 432 */ 433 if (elapsed > win->size) 434 window_reset(win, now, value, growth); 435 else { 436 u32 remaining; 437 438 remaining = win->size - elapsed; 439 growth += div64_u64(win->prev_growth * remaining, win->size); 440 } 441 442 return growth; 443 } 444 445 static void update_triggers(struct psi_group *group, u64 now, 446 enum psi_aggregators aggregator) 447 { 448 struct psi_trigger *t; 449 u64 *total = group->total[aggregator]; 450 struct list_head *triggers; 451 u64 *aggregator_total; 452 453 if (aggregator == PSI_AVGS) { 454 triggers = &group->avg_triggers; 455 aggregator_total = group->avg_total; 456 } else { 457 triggers = &group->rtpoll_triggers; 458 aggregator_total = group->rtpoll_total; 459 } 460 461 /* 462 * On subsequent updates, calculate growth deltas and let 463 * watchers know when their specified thresholds are exceeded. 464 */ 465 list_for_each_entry(t, triggers, node) { 466 u64 growth; 467 bool new_stall; 468 469 new_stall = aggregator_total[t->state] != total[t->state]; 470 471 /* Check for stall activity or a previous threshold breach */ 472 if (!new_stall && !t->pending_event) 473 continue; 474 /* 475 * Check for new stall activity, as well as deferred 476 * events that occurred in the last window after the 477 * trigger had already fired (we want to ratelimit 478 * events without dropping any). 479 */ 480 if (new_stall) { 481 /* Calculate growth since last update */ 482 growth = window_update(&t->win, now, total[t->state]); 483 if (!t->pending_event) { 484 if (growth < t->threshold) 485 continue; 486 487 t->pending_event = true; 488 } 489 } 490 /* Limit event signaling to once per window */ 491 if (now < t->last_event_time + t->win.size) 492 continue; 493 494 /* Generate an event */ 495 if (cmpxchg(&t->event, 0, 1) == 0) { 496 if (t->of) 497 kernfs_notify(t->of->kn); 498 else 499 wake_up_interruptible(&t->event_wait); 500 } 501 t->last_event_time = now; 502 /* Reset threshold breach flag once event got generated */ 503 t->pending_event = false; 504 } 505 } 506 507 static u64 update_averages(struct psi_group *group, u64 now) 508 { 509 unsigned long missed_periods = 0; 510 u64 expires, period; 511 u64 avg_next_update; 512 int s; 513 514 /* avgX= */ 515 expires = group->avg_next_update; 516 if (now - expires >= psi_period) 517 missed_periods = div_u64(now - expires, psi_period); 518 519 /* 520 * The periodic clock tick can get delayed for various 521 * reasons, especially on loaded systems. To avoid clock 522 * drift, we schedule the clock in fixed psi_period intervals. 523 * But the deltas we sample out of the per-cpu buckets above 524 * are based on the actual time elapsing between clock ticks. 525 */ 526 avg_next_update = expires + ((1 + missed_periods) * psi_period); 527 period = now - (group->avg_last_update + (missed_periods * psi_period)); 528 group->avg_last_update = now; 529 530 for (s = 0; s < NR_PSI_STATES - 1; s++) { 531 u32 sample; 532 533 sample = group->total[PSI_AVGS][s] - group->avg_total[s]; 534 /* 535 * Due to the lockless sampling of the time buckets, 536 * recorded time deltas can slip into the next period, 537 * which under full pressure can result in samples in 538 * excess of the period length. 539 * 540 * We don't want to report non-sensical pressures in 541 * excess of 100%, nor do we want to drop such events 542 * on the floor. Instead we punt any overage into the 543 * future until pressure subsides. By doing this we 544 * don't underreport the occurring pressure curve, we 545 * just report it delayed by one period length. 546 * 547 * The error isn't cumulative. As soon as another 548 * delta slips from a period P to P+1, by definition 549 * it frees up its time T in P. 550 */ 551 if (sample > period) 552 sample = period; 553 group->avg_total[s] += sample; 554 calc_avgs(group->avg[s], missed_periods, sample, period); 555 } 556 557 return avg_next_update; 558 } 559 560 static void psi_avgs_work(struct work_struct *work) 561 { 562 struct delayed_work *dwork; 563 struct psi_group *group; 564 u32 changed_states; 565 u64 now; 566 567 dwork = to_delayed_work(work); 568 group = container_of(dwork, struct psi_group, avgs_work); 569 570 mutex_lock(&group->avgs_lock); 571 572 now = sched_clock(); 573 574 collect_percpu_times(group, PSI_AVGS, &changed_states); 575 /* 576 * If there is task activity, periodically fold the per-cpu 577 * times and feed samples into the running averages. If things 578 * are idle and there is no data to process, stop the clock. 579 * Once restarted, we'll catch up the running averages in one 580 * go - see calc_avgs() and missed_periods. 581 */ 582 if (now >= group->avg_next_update) { 583 update_triggers(group, now, PSI_AVGS); 584 group->avg_next_update = update_averages(group, now); 585 } 586 587 if (changed_states & PSI_STATE_RESCHEDULE) { 588 schedule_delayed_work(dwork, nsecs_to_jiffies( 589 group->avg_next_update - now) + 1); 590 } 591 592 mutex_unlock(&group->avgs_lock); 593 } 594 595 static void init_rtpoll_triggers(struct psi_group *group, u64 now) 596 { 597 struct psi_trigger *t; 598 599 list_for_each_entry(t, &group->rtpoll_triggers, node) 600 window_reset(&t->win, now, 601 group->total[PSI_POLL][t->state], 0); 602 memcpy(group->rtpoll_total, group->total[PSI_POLL], 603 sizeof(group->rtpoll_total)); 604 group->rtpoll_next_update = now + group->rtpoll_min_period; 605 } 606 607 /* Schedule rtpolling if it's not already scheduled or forced. */ 608 static void psi_schedule_rtpoll_work(struct psi_group *group, unsigned long delay, 609 bool force) 610 { 611 struct task_struct *task; 612 613 /* 614 * atomic_xchg should be called even when !force to provide a 615 * full memory barrier (see the comment inside psi_rtpoll_work). 616 */ 617 if (atomic_xchg(&group->rtpoll_scheduled, 1) && !force) 618 return; 619 620 rcu_read_lock(); 621 622 task = rcu_dereference(group->rtpoll_task); 623 /* 624 * kworker might be NULL in case psi_trigger_destroy races with 625 * psi_task_change (hotpath) which can't use locks 626 */ 627 if (likely(task)) 628 mod_timer(&group->rtpoll_timer, jiffies + delay); 629 else 630 atomic_set(&group->rtpoll_scheduled, 0); 631 632 rcu_read_unlock(); 633 } 634 635 static void psi_rtpoll_work(struct psi_group *group) 636 { 637 bool force_reschedule = false; 638 u32 changed_states; 639 u64 now; 640 641 mutex_lock(&group->rtpoll_trigger_lock); 642 643 now = sched_clock(); 644 645 if (now > group->rtpoll_until) { 646 /* 647 * We are either about to start or might stop rtpolling if no 648 * state change was recorded. Resetting rtpoll_scheduled leaves 649 * a small window for psi_group_change to sneak in and schedule 650 * an immediate rtpoll_work before we get to rescheduling. One 651 * potential extra wakeup at the end of the rtpolling window 652 * should be negligible and rtpoll_next_update still keeps 653 * updates correctly on schedule. 654 */ 655 atomic_set(&group->rtpoll_scheduled, 0); 656 /* 657 * A task change can race with the rtpoll worker that is supposed to 658 * report on it. To avoid missing events, ensure ordering between 659 * rtpoll_scheduled and the task state accesses, such that if the 660 * rtpoll worker misses the state update, the task change is 661 * guaranteed to reschedule the rtpoll worker: 662 * 663 * rtpoll worker: 664 * atomic_set(rtpoll_scheduled, 0) 665 * smp_mb() 666 * LOAD states 667 * 668 * task change: 669 * STORE states 670 * if atomic_xchg(rtpoll_scheduled, 1) == 0: 671 * schedule rtpoll worker 672 * 673 * The atomic_xchg() implies a full barrier. 674 */ 675 smp_mb(); 676 } else { 677 /* The rtpolling window is not over, keep rescheduling */ 678 force_reschedule = true; 679 } 680 681 682 collect_percpu_times(group, PSI_POLL, &changed_states); 683 684 if (changed_states & group->rtpoll_states) { 685 /* Initialize trigger windows when entering rtpolling mode */ 686 if (now > group->rtpoll_until) 687 init_rtpoll_triggers(group, now); 688 689 /* 690 * Keep the monitor active for at least the duration of the 691 * minimum tracking window as long as monitor states are 692 * changing. 693 */ 694 group->rtpoll_until = now + 695 group->rtpoll_min_period * UPDATES_PER_WINDOW; 696 } 697 698 if (now > group->rtpoll_until) { 699 group->rtpoll_next_update = ULLONG_MAX; 700 goto out; 701 } 702 703 if (now >= group->rtpoll_next_update) { 704 if (changed_states & group->rtpoll_states) { 705 update_triggers(group, now, PSI_POLL); 706 memcpy(group->rtpoll_total, group->total[PSI_POLL], 707 sizeof(group->rtpoll_total)); 708 } 709 group->rtpoll_next_update = now + group->rtpoll_min_period; 710 } 711 712 psi_schedule_rtpoll_work(group, 713 nsecs_to_jiffies(group->rtpoll_next_update - now) + 1, 714 force_reschedule); 715 716 out: 717 mutex_unlock(&group->rtpoll_trigger_lock); 718 } 719 720 static int psi_rtpoll_worker(void *data) 721 { 722 struct psi_group *group = (struct psi_group *)data; 723 724 sched_set_fifo_low(current); 725 726 while (true) { 727 wait_event_interruptible(group->rtpoll_wait, 728 atomic_cmpxchg(&group->rtpoll_wakeup, 1, 0) || 729 kthread_should_stop()); 730 if (kthread_should_stop()) 731 break; 732 733 psi_rtpoll_work(group); 734 } 735 return 0; 736 } 737 738 static void poll_timer_fn(struct timer_list *t) 739 { 740 struct psi_group *group = timer_container_of(group, t, rtpoll_timer); 741 742 atomic_set(&group->rtpoll_wakeup, 1); 743 wake_up_interruptible(&group->rtpoll_wait); 744 } 745 746 static void record_times(struct psi_group_cpu *groupc, u64 now) 747 { 748 u32 delta; 749 750 delta = now - groupc->state_start; 751 groupc->state_start = now; 752 753 if (groupc->state_mask & (1 << PSI_IO_SOME)) { 754 groupc->times[PSI_IO_SOME] += delta; 755 if (groupc->state_mask & (1 << PSI_IO_FULL)) 756 groupc->times[PSI_IO_FULL] += delta; 757 } 758 759 if (groupc->state_mask & (1 << PSI_MEM_SOME)) { 760 groupc->times[PSI_MEM_SOME] += delta; 761 if (groupc->state_mask & (1 << PSI_MEM_FULL)) 762 groupc->times[PSI_MEM_FULL] += delta; 763 } 764 765 if (groupc->state_mask & (1 << PSI_CPU_SOME)) { 766 groupc->times[PSI_CPU_SOME] += delta; 767 if (groupc->state_mask & (1 << PSI_CPU_FULL)) 768 groupc->times[PSI_CPU_FULL] += delta; 769 } 770 771 if (groupc->state_mask & (1 << PSI_NONIDLE)) 772 groupc->times[PSI_NONIDLE] += delta; 773 } 774 775 static void psi_group_change(struct psi_group *group, int cpu, 776 unsigned int clear, unsigned int set, 777 bool wake_clock) 778 { 779 struct psi_group_cpu *groupc; 780 unsigned int t, m; 781 u32 state_mask; 782 u64 now; 783 784 lockdep_assert_rq_held(cpu_rq(cpu)); 785 groupc = per_cpu_ptr(group->pcpu, cpu); 786 787 /* 788 * First we update the task counts according to the state 789 * change requested through the @clear and @set bits. 790 * 791 * Then if the cgroup PSI stats accounting enabled, we 792 * assess the aggregate resource states this CPU's tasks 793 * have been in since the last change, and account any 794 * SOME and FULL time these may have resulted in. 795 */ 796 write_seqcount_begin(&groupc->seq); 797 now = cpu_clock(cpu); 798 799 /* 800 * Start with TSK_ONCPU, which doesn't have a corresponding 801 * task count - it's just a boolean flag directly encoded in 802 * the state mask. Clear, set, or carry the current state if 803 * no changes are requested. 804 */ 805 if (unlikely(clear & TSK_ONCPU)) { 806 state_mask = 0; 807 clear &= ~TSK_ONCPU; 808 } else if (unlikely(set & TSK_ONCPU)) { 809 state_mask = PSI_ONCPU; 810 set &= ~TSK_ONCPU; 811 } else { 812 state_mask = groupc->state_mask & PSI_ONCPU; 813 } 814 815 /* 816 * The rest of the state mask is calculated based on the task 817 * counts. Update those first, then construct the mask. 818 */ 819 for (t = 0, m = clear; m; m &= ~(1 << t), t++) { 820 if (!(m & (1 << t))) 821 continue; 822 if (groupc->tasks[t]) { 823 groupc->tasks[t]--; 824 } else if (!psi_bug) { 825 printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n", 826 cpu, t, groupc->tasks[0], 827 groupc->tasks[1], groupc->tasks[2], 828 groupc->tasks[3], clear, set); 829 psi_bug = 1; 830 } 831 } 832 833 for (t = 0; set; set &= ~(1 << t), t++) 834 if (set & (1 << t)) 835 groupc->tasks[t]++; 836 837 if (!group->enabled) { 838 /* 839 * On the first group change after disabling PSI, conclude 840 * the current state and flush its time. This is unlikely 841 * to matter to the user, but aggregation (get_recent_times) 842 * may have already incorporated the live state into times_prev; 843 * avoid a delta sample underflow when PSI is later re-enabled. 844 */ 845 if (unlikely(groupc->state_mask & (1 << PSI_NONIDLE))) 846 record_times(groupc, now); 847 848 groupc->state_mask = state_mask; 849 850 write_seqcount_end(&groupc->seq); 851 return; 852 } 853 854 state_mask = test_states(groupc->tasks, state_mask); 855 856 /* 857 * Since we care about lost potential, a memstall is FULL 858 * when there are no other working tasks, but also when 859 * the CPU is actively reclaiming and nothing productive 860 * could run even if it were runnable. So when the current 861 * task in a cgroup is in_memstall, the corresponding groupc 862 * on that cpu is in PSI_MEM_FULL state. 863 */ 864 if (unlikely((state_mask & PSI_ONCPU) && cpu_curr(cpu)->in_memstall)) 865 state_mask |= (1 << PSI_MEM_FULL); 866 867 record_times(groupc, now); 868 869 groupc->state_mask = state_mask; 870 871 write_seqcount_end(&groupc->seq); 872 873 if (state_mask & group->rtpoll_states) 874 psi_schedule_rtpoll_work(group, 1, false); 875 876 if (wake_clock && !delayed_work_pending(&group->avgs_work)) 877 schedule_delayed_work(&group->avgs_work, PSI_FREQ); 878 } 879 880 static inline struct psi_group *task_psi_group(struct task_struct *task) 881 { 882 #ifdef CONFIG_CGROUPS 883 if (static_branch_likely(&psi_cgroups_enabled)) 884 return cgroup_psi(task_dfl_cgroup(task)); 885 #endif 886 return &psi_system; 887 } 888 889 static void psi_flags_change(struct task_struct *task, int clear, int set) 890 { 891 if (((task->psi_flags & set) || 892 (task->psi_flags & clear) != clear) && 893 !psi_bug) { 894 printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n", 895 task->pid, task->comm, task_cpu(task), 896 task->psi_flags, clear, set); 897 psi_bug = 1; 898 } 899 900 task->psi_flags &= ~clear; 901 task->psi_flags |= set; 902 } 903 904 void psi_task_change(struct task_struct *task, int clear, int set) 905 { 906 int cpu = task_cpu(task); 907 struct psi_group *group; 908 909 if (!task->pid) 910 return; 911 912 psi_flags_change(task, clear, set); 913 914 group = task_psi_group(task); 915 do { 916 psi_group_change(group, cpu, clear, set, true); 917 } while ((group = group->parent)); 918 } 919 920 void psi_task_switch(struct task_struct *prev, struct task_struct *next, 921 bool sleep) 922 { 923 struct psi_group *group, *common = NULL; 924 int cpu = task_cpu(prev); 925 926 if (next->pid) { 927 psi_flags_change(next, 0, TSK_ONCPU); 928 /* 929 * Set TSK_ONCPU on @next's cgroups. If @next shares any 930 * ancestors with @prev, those will already have @prev's 931 * TSK_ONCPU bit set, and we can stop the iteration there. 932 */ 933 group = task_psi_group(next); 934 do { 935 if (per_cpu_ptr(group->pcpu, cpu)->state_mask & 936 PSI_ONCPU) { 937 common = group; 938 break; 939 } 940 941 psi_group_change(group, cpu, 0, TSK_ONCPU, true); 942 } while ((group = group->parent)); 943 } 944 945 if (prev->pid) { 946 int clear = TSK_ONCPU, set = 0; 947 bool wake_clock = true; 948 949 /* 950 * When we're going to sleep, psi_dequeue() lets us 951 * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and 952 * TSK_IOWAIT here, where we can combine it with 953 * TSK_ONCPU and save walking common ancestors twice. 954 */ 955 if (sleep) { 956 clear |= TSK_RUNNING; 957 if (prev->in_memstall) 958 clear |= TSK_MEMSTALL_RUNNING; 959 if (prev->in_iowait) 960 set |= TSK_IOWAIT; 961 962 /* 963 * Periodic aggregation shuts off if there is a period of no 964 * task changes, so we wake it back up if necessary. However, 965 * don't do this if the task change is the aggregation worker 966 * itself going to sleep, or we'll ping-pong forever. 967 */ 968 if (unlikely((prev->flags & PF_WQ_WORKER) && 969 wq_worker_last_func(prev) == psi_avgs_work)) 970 wake_clock = false; 971 } 972 973 psi_flags_change(prev, clear, set); 974 975 group = task_psi_group(prev); 976 do { 977 if (group == common) 978 break; 979 psi_group_change(group, cpu, clear, set, wake_clock); 980 } while ((group = group->parent)); 981 982 /* 983 * TSK_ONCPU is handled up to the common ancestor. If there are 984 * any other differences between the two tasks (e.g. prev goes 985 * to sleep, or only one task is memstall), finish propagating 986 * those differences all the way up to the root. 987 */ 988 if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) { 989 clear &= ~TSK_ONCPU; 990 for (; group; group = group->parent) 991 psi_group_change(group, cpu, clear, set, wake_clock); 992 } 993 } 994 } 995 996 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 997 void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_struct *prev) 998 { 999 int cpu = task_cpu(curr); 1000 struct psi_group *group; 1001 struct psi_group_cpu *groupc; 1002 s64 delta; 1003 u64 irq; 1004 1005 if (static_branch_likely(&psi_disabled) || !irqtime_enabled()) 1006 return; 1007 1008 if (!curr->pid) 1009 return; 1010 1011 lockdep_assert_rq_held(rq); 1012 group = task_psi_group(curr); 1013 if (prev && task_psi_group(prev) == group) 1014 return; 1015 1016 irq = irq_time_read(cpu); 1017 delta = (s64)(irq - rq->psi_irq_time); 1018 if (delta < 0) 1019 return; 1020 rq->psi_irq_time = irq; 1021 1022 do { 1023 u64 now; 1024 1025 if (!group->enabled) 1026 continue; 1027 1028 groupc = per_cpu_ptr(group->pcpu, cpu); 1029 1030 write_seqcount_begin(&groupc->seq); 1031 now = cpu_clock(cpu); 1032 1033 record_times(groupc, now); 1034 groupc->times[PSI_IRQ_FULL] += delta; 1035 1036 write_seqcount_end(&groupc->seq); 1037 1038 if (group->rtpoll_states & (1 << PSI_IRQ_FULL)) 1039 psi_schedule_rtpoll_work(group, 1, false); 1040 } while ((group = group->parent)); 1041 } 1042 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 1043 1044 /** 1045 * psi_memstall_enter - mark the beginning of a memory stall section 1046 * @flags: flags to handle nested sections 1047 * 1048 * Marks the calling task as being stalled due to a lack of memory, 1049 * such as waiting for a refault or performing reclaim. 1050 */ 1051 void psi_memstall_enter(unsigned long *flags) 1052 { 1053 struct rq_flags rf; 1054 struct rq *rq; 1055 1056 if (static_branch_likely(&psi_disabled)) 1057 return; 1058 1059 *flags = current->in_memstall; 1060 if (*flags) 1061 return; 1062 /* 1063 * in_memstall setting & accounting needs to be atomic wrt 1064 * changes to the task's scheduling state, otherwise we can 1065 * race with CPU migration. 1066 */ 1067 rq = this_rq_lock_irq(&rf); 1068 1069 current->in_memstall = 1; 1070 psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING); 1071 1072 rq_unlock_irq(rq, &rf); 1073 } 1074 EXPORT_SYMBOL_GPL(psi_memstall_enter); 1075 1076 /** 1077 * psi_memstall_leave - mark the end of an memory stall section 1078 * @flags: flags to handle nested memdelay sections 1079 * 1080 * Marks the calling task as no longer stalled due to lack of memory. 1081 */ 1082 void psi_memstall_leave(unsigned long *flags) 1083 { 1084 struct rq_flags rf; 1085 struct rq *rq; 1086 1087 if (static_branch_likely(&psi_disabled)) 1088 return; 1089 1090 if (*flags) 1091 return; 1092 /* 1093 * in_memstall clearing & accounting needs to be atomic wrt 1094 * changes to the task's scheduling state, otherwise we could 1095 * race with CPU migration. 1096 */ 1097 rq = this_rq_lock_irq(&rf); 1098 1099 current->in_memstall = 0; 1100 psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0); 1101 1102 rq_unlock_irq(rq, &rf); 1103 } 1104 EXPORT_SYMBOL_GPL(psi_memstall_leave); 1105 1106 #ifdef CONFIG_CGROUPS 1107 int psi_cgroup_alloc(struct cgroup *cgroup) 1108 { 1109 if (!static_branch_likely(&psi_cgroups_enabled)) 1110 return 0; 1111 1112 cgroup->psi = kzalloc(sizeof(struct psi_group), GFP_KERNEL); 1113 if (!cgroup->psi) 1114 return -ENOMEM; 1115 1116 cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu); 1117 if (!cgroup->psi->pcpu) { 1118 kfree(cgroup->psi); 1119 return -ENOMEM; 1120 } 1121 group_init(cgroup->psi); 1122 cgroup->psi->parent = cgroup_psi(cgroup_parent(cgroup)); 1123 return 0; 1124 } 1125 1126 void psi_cgroup_free(struct cgroup *cgroup) 1127 { 1128 if (!static_branch_likely(&psi_cgroups_enabled)) 1129 return; 1130 1131 cancel_delayed_work_sync(&cgroup->psi->avgs_work); 1132 free_percpu(cgroup->psi->pcpu); 1133 /* All triggers must be removed by now */ 1134 WARN_ONCE(cgroup->psi->rtpoll_states, "psi: trigger leak\n"); 1135 kfree(cgroup->psi); 1136 } 1137 1138 /** 1139 * cgroup_move_task - move task to a different cgroup 1140 * @task: the task 1141 * @to: the target css_set 1142 * 1143 * Move task to a new cgroup and safely migrate its associated stall 1144 * state between the different groups. 1145 * 1146 * This function acquires the task's rq lock to lock out concurrent 1147 * changes to the task's scheduling state and - in case the task is 1148 * running - concurrent changes to its stall state. 1149 */ 1150 void cgroup_move_task(struct task_struct *task, struct css_set *to) 1151 { 1152 unsigned int task_flags; 1153 struct rq_flags rf; 1154 struct rq *rq; 1155 1156 if (!static_branch_likely(&psi_cgroups_enabled)) { 1157 /* 1158 * Lame to do this here, but the scheduler cannot be locked 1159 * from the outside, so we move cgroups from inside sched/. 1160 */ 1161 rcu_assign_pointer(task->cgroups, to); 1162 return; 1163 } 1164 1165 rq = task_rq_lock(task, &rf); 1166 1167 /* 1168 * We may race with schedule() dropping the rq lock between 1169 * deactivating prev and switching to next. Because the psi 1170 * updates from the deactivation are deferred to the switch 1171 * callback to save cgroup tree updates, the task's scheduling 1172 * state here is not coherent with its psi state: 1173 * 1174 * schedule() cgroup_move_task() 1175 * rq_lock() 1176 * deactivate_task() 1177 * p->on_rq = 0 1178 * psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates 1179 * pick_next_task() 1180 * rq_unlock() 1181 * rq_lock() 1182 * psi_task_change() // old cgroup 1183 * task->cgroups = to 1184 * psi_task_change() // new cgroup 1185 * rq_unlock() 1186 * rq_lock() 1187 * psi_sched_switch() // does deferred updates in new cgroup 1188 * 1189 * Don't rely on the scheduling state. Use psi_flags instead. 1190 */ 1191 task_flags = task->psi_flags; 1192 1193 if (task_flags) 1194 psi_task_change(task, task_flags, 0); 1195 1196 /* See comment above */ 1197 rcu_assign_pointer(task->cgroups, to); 1198 1199 if (task_flags) 1200 psi_task_change(task, 0, task_flags); 1201 1202 task_rq_unlock(rq, task, &rf); 1203 } 1204 1205 void psi_cgroup_restart(struct psi_group *group) 1206 { 1207 int cpu; 1208 1209 /* 1210 * After we disable psi_group->enabled, we don't actually 1211 * stop percpu tasks accounting in each psi_group_cpu, 1212 * instead only stop test_states() loop, record_times() 1213 * and averaging worker, see psi_group_change() for details. 1214 * 1215 * When disable cgroup PSI, this function has nothing to sync 1216 * since cgroup pressure files are hidden and percpu psi_group_cpu 1217 * would see !psi_group->enabled and only do task accounting. 1218 * 1219 * When re-enable cgroup PSI, this function use psi_group_change() 1220 * to get correct state mask from test_states() loop on tasks[], 1221 * and restart groupc->state_start from now, use .clear = .set = 0 1222 * here since no task status really changed. 1223 */ 1224 if (!group->enabled) 1225 return; 1226 1227 for_each_possible_cpu(cpu) { 1228 struct rq *rq = cpu_rq(cpu); 1229 struct rq_flags rf; 1230 1231 rq_lock_irq(rq, &rf); 1232 psi_group_change(group, cpu, 0, 0, true); 1233 rq_unlock_irq(rq, &rf); 1234 } 1235 } 1236 #endif /* CONFIG_CGROUPS */ 1237 1238 int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res) 1239 { 1240 bool only_full = false; 1241 int full; 1242 u64 now; 1243 1244 if (static_branch_likely(&psi_disabled)) 1245 return -EOPNOTSUPP; 1246 1247 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1248 if (!irqtime_enabled() && res == PSI_IRQ) 1249 return -EOPNOTSUPP; 1250 #endif 1251 1252 /* Update averages before reporting them */ 1253 mutex_lock(&group->avgs_lock); 1254 now = sched_clock(); 1255 collect_percpu_times(group, PSI_AVGS, NULL); 1256 if (now >= group->avg_next_update) 1257 group->avg_next_update = update_averages(group, now); 1258 mutex_unlock(&group->avgs_lock); 1259 1260 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1261 only_full = res == PSI_IRQ; 1262 #endif 1263 1264 for (full = 0; full < 2 - only_full; full++) { 1265 unsigned long avg[3] = { 0, }; 1266 u64 total = 0; 1267 int w; 1268 1269 /* CPU FULL is undefined at the system level */ 1270 if (!(group == &psi_system && res == PSI_CPU && full)) { 1271 for (w = 0; w < 3; w++) 1272 avg[w] = group->avg[res * 2 + full][w]; 1273 total = div_u64(group->total[PSI_AVGS][res * 2 + full], 1274 NSEC_PER_USEC); 1275 } 1276 1277 seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n", 1278 full || only_full ? "full" : "some", 1279 LOAD_INT(avg[0]), LOAD_FRAC(avg[0]), 1280 LOAD_INT(avg[1]), LOAD_FRAC(avg[1]), 1281 LOAD_INT(avg[2]), LOAD_FRAC(avg[2]), 1282 total); 1283 } 1284 1285 return 0; 1286 } 1287 1288 struct psi_trigger *psi_trigger_create(struct psi_group *group, char *buf, 1289 enum psi_res res, struct file *file, 1290 struct kernfs_open_file *of) 1291 { 1292 struct psi_trigger *t; 1293 enum psi_states state; 1294 u32 threshold_us; 1295 bool privileged; 1296 u32 window_us; 1297 1298 if (static_branch_likely(&psi_disabled)) 1299 return ERR_PTR(-EOPNOTSUPP); 1300 1301 /* 1302 * Checking the privilege here on file->f_cred implies that a privileged user 1303 * could open the file and delegate the write to an unprivileged one. 1304 */ 1305 privileged = cap_raised(file->f_cred->cap_effective, CAP_SYS_RESOURCE); 1306 1307 if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2) 1308 state = PSI_IO_SOME + res * 2; 1309 else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2) 1310 state = PSI_IO_FULL + res * 2; 1311 else 1312 return ERR_PTR(-EINVAL); 1313 1314 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1315 if (res == PSI_IRQ && --state != PSI_IRQ_FULL) 1316 return ERR_PTR(-EINVAL); 1317 #endif 1318 1319 if (state >= PSI_NONIDLE) 1320 return ERR_PTR(-EINVAL); 1321 1322 if (window_us == 0 || window_us > WINDOW_MAX_US) 1323 return ERR_PTR(-EINVAL); 1324 1325 /* 1326 * Unprivileged users can only use 2s windows so that averages aggregation 1327 * work is used, and no RT threads need to be spawned. 1328 */ 1329 if (!privileged && window_us % 2000000) 1330 return ERR_PTR(-EINVAL); 1331 1332 /* Check threshold */ 1333 if (threshold_us == 0 || threshold_us > window_us) 1334 return ERR_PTR(-EINVAL); 1335 1336 t = kmalloc(sizeof(*t), GFP_KERNEL); 1337 if (!t) 1338 return ERR_PTR(-ENOMEM); 1339 1340 t->group = group; 1341 t->state = state; 1342 t->threshold = threshold_us * NSEC_PER_USEC; 1343 t->win.size = window_us * NSEC_PER_USEC; 1344 window_reset(&t->win, sched_clock(), 1345 group->total[PSI_POLL][t->state], 0); 1346 1347 t->event = 0; 1348 t->last_event_time = 0; 1349 t->of = of; 1350 if (!of) 1351 init_waitqueue_head(&t->event_wait); 1352 t->pending_event = false; 1353 t->aggregator = privileged ? PSI_POLL : PSI_AVGS; 1354 1355 if (privileged) { 1356 mutex_lock(&group->rtpoll_trigger_lock); 1357 1358 if (!rcu_access_pointer(group->rtpoll_task)) { 1359 struct task_struct *task; 1360 1361 task = kthread_create(psi_rtpoll_worker, group, "psimon"); 1362 if (IS_ERR(task)) { 1363 kfree(t); 1364 mutex_unlock(&group->rtpoll_trigger_lock); 1365 return ERR_CAST(task); 1366 } 1367 atomic_set(&group->rtpoll_wakeup, 0); 1368 wake_up_process(task); 1369 rcu_assign_pointer(group->rtpoll_task, task); 1370 } 1371 1372 list_add(&t->node, &group->rtpoll_triggers); 1373 group->rtpoll_min_period = min(group->rtpoll_min_period, 1374 div_u64(t->win.size, UPDATES_PER_WINDOW)); 1375 group->rtpoll_nr_triggers[t->state]++; 1376 group->rtpoll_states |= (1 << t->state); 1377 1378 mutex_unlock(&group->rtpoll_trigger_lock); 1379 } else { 1380 mutex_lock(&group->avgs_lock); 1381 1382 list_add(&t->node, &group->avg_triggers); 1383 group->avg_nr_triggers[t->state]++; 1384 1385 mutex_unlock(&group->avgs_lock); 1386 } 1387 return t; 1388 } 1389 1390 void psi_trigger_destroy(struct psi_trigger *t) 1391 { 1392 struct psi_group *group; 1393 struct task_struct *task_to_destroy = NULL; 1394 1395 /* 1396 * We do not check psi_disabled since it might have been disabled after 1397 * the trigger got created. 1398 */ 1399 if (!t) 1400 return; 1401 1402 group = t->group; 1403 /* 1404 * Wakeup waiters to stop polling and clear the queue to prevent it from 1405 * being accessed later. Can happen if cgroup is deleted from under a 1406 * polling process. 1407 */ 1408 if (t->of) 1409 kernfs_notify(t->of->kn); 1410 else 1411 wake_up_interruptible(&t->event_wait); 1412 1413 if (t->aggregator == PSI_AVGS) { 1414 mutex_lock(&group->avgs_lock); 1415 if (!list_empty(&t->node)) { 1416 list_del(&t->node); 1417 group->avg_nr_triggers[t->state]--; 1418 } 1419 mutex_unlock(&group->avgs_lock); 1420 } else { 1421 mutex_lock(&group->rtpoll_trigger_lock); 1422 if (!list_empty(&t->node)) { 1423 struct psi_trigger *tmp; 1424 u64 period = ULLONG_MAX; 1425 1426 list_del(&t->node); 1427 group->rtpoll_nr_triggers[t->state]--; 1428 if (!group->rtpoll_nr_triggers[t->state]) 1429 group->rtpoll_states &= ~(1 << t->state); 1430 /* 1431 * Reset min update period for the remaining triggers 1432 * iff the destroying trigger had the min window size. 1433 */ 1434 if (group->rtpoll_min_period == div_u64(t->win.size, UPDATES_PER_WINDOW)) { 1435 list_for_each_entry(tmp, &group->rtpoll_triggers, node) 1436 period = min(period, div_u64(tmp->win.size, 1437 UPDATES_PER_WINDOW)); 1438 group->rtpoll_min_period = period; 1439 } 1440 /* Destroy rtpoll_task when the last trigger is destroyed */ 1441 if (group->rtpoll_states == 0) { 1442 group->rtpoll_until = 0; 1443 task_to_destroy = rcu_dereference_protected( 1444 group->rtpoll_task, 1445 lockdep_is_held(&group->rtpoll_trigger_lock)); 1446 rcu_assign_pointer(group->rtpoll_task, NULL); 1447 timer_delete(&group->rtpoll_timer); 1448 } 1449 } 1450 mutex_unlock(&group->rtpoll_trigger_lock); 1451 } 1452 1453 /* 1454 * Wait for psi_schedule_rtpoll_work RCU to complete its read-side 1455 * critical section before destroying the trigger and optionally the 1456 * rtpoll_task. 1457 */ 1458 synchronize_rcu(); 1459 /* 1460 * Stop kthread 'psimon' after releasing rtpoll_trigger_lock to prevent 1461 * a deadlock while waiting for psi_rtpoll_work to acquire 1462 * rtpoll_trigger_lock 1463 */ 1464 if (task_to_destroy) { 1465 /* 1466 * After the RCU grace period has expired, the worker 1467 * can no longer be found through group->rtpoll_task. 1468 */ 1469 kthread_stop(task_to_destroy); 1470 atomic_set(&group->rtpoll_scheduled, 0); 1471 } 1472 kfree(t); 1473 } 1474 1475 __poll_t psi_trigger_poll(void **trigger_ptr, 1476 struct file *file, poll_table *wait) 1477 { 1478 __poll_t ret = DEFAULT_POLLMASK; 1479 struct psi_trigger *t; 1480 1481 if (static_branch_likely(&psi_disabled)) 1482 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI; 1483 1484 t = smp_load_acquire(trigger_ptr); 1485 if (!t) 1486 return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI; 1487 1488 if (t->of) 1489 kernfs_generic_poll(t->of, wait); 1490 else 1491 poll_wait(file, &t->event_wait, wait); 1492 1493 if (cmpxchg(&t->event, 1, 0) == 1) 1494 ret |= EPOLLPRI; 1495 1496 return ret; 1497 } 1498 1499 #ifdef CONFIG_PROC_FS 1500 static int psi_io_show(struct seq_file *m, void *v) 1501 { 1502 return psi_show(m, &psi_system, PSI_IO); 1503 } 1504 1505 static int psi_memory_show(struct seq_file *m, void *v) 1506 { 1507 return psi_show(m, &psi_system, PSI_MEM); 1508 } 1509 1510 static int psi_cpu_show(struct seq_file *m, void *v) 1511 { 1512 return psi_show(m, &psi_system, PSI_CPU); 1513 } 1514 1515 static int psi_io_open(struct inode *inode, struct file *file) 1516 { 1517 return single_open(file, psi_io_show, NULL); 1518 } 1519 1520 static int psi_memory_open(struct inode *inode, struct file *file) 1521 { 1522 return single_open(file, psi_memory_show, NULL); 1523 } 1524 1525 static int psi_cpu_open(struct inode *inode, struct file *file) 1526 { 1527 return single_open(file, psi_cpu_show, NULL); 1528 } 1529 1530 static ssize_t psi_write(struct file *file, const char __user *user_buf, 1531 size_t nbytes, enum psi_res res) 1532 { 1533 char buf[32]; 1534 size_t buf_size; 1535 struct seq_file *seq; 1536 struct psi_trigger *new; 1537 1538 if (static_branch_likely(&psi_disabled)) 1539 return -EOPNOTSUPP; 1540 1541 if (!nbytes) 1542 return -EINVAL; 1543 1544 buf_size = min(nbytes, sizeof(buf)); 1545 if (copy_from_user(buf, user_buf, buf_size)) 1546 return -EFAULT; 1547 1548 buf[buf_size - 1] = '\0'; 1549 1550 seq = file->private_data; 1551 1552 /* Take seq->lock to protect seq->private from concurrent writes */ 1553 mutex_lock(&seq->lock); 1554 1555 /* Allow only one trigger per file descriptor */ 1556 if (seq->private) { 1557 mutex_unlock(&seq->lock); 1558 return -EBUSY; 1559 } 1560 1561 new = psi_trigger_create(&psi_system, buf, res, file, NULL); 1562 if (IS_ERR(new)) { 1563 mutex_unlock(&seq->lock); 1564 return PTR_ERR(new); 1565 } 1566 1567 smp_store_release(&seq->private, new); 1568 mutex_unlock(&seq->lock); 1569 1570 return nbytes; 1571 } 1572 1573 static ssize_t psi_io_write(struct file *file, const char __user *user_buf, 1574 size_t nbytes, loff_t *ppos) 1575 { 1576 return psi_write(file, user_buf, nbytes, PSI_IO); 1577 } 1578 1579 static ssize_t psi_memory_write(struct file *file, const char __user *user_buf, 1580 size_t nbytes, loff_t *ppos) 1581 { 1582 return psi_write(file, user_buf, nbytes, PSI_MEM); 1583 } 1584 1585 static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf, 1586 size_t nbytes, loff_t *ppos) 1587 { 1588 return psi_write(file, user_buf, nbytes, PSI_CPU); 1589 } 1590 1591 static __poll_t psi_fop_poll(struct file *file, poll_table *wait) 1592 { 1593 struct seq_file *seq = file->private_data; 1594 1595 return psi_trigger_poll(&seq->private, file, wait); 1596 } 1597 1598 static int psi_fop_release(struct inode *inode, struct file *file) 1599 { 1600 struct seq_file *seq = file->private_data; 1601 1602 psi_trigger_destroy(seq->private); 1603 return single_release(inode, file); 1604 } 1605 1606 static const struct proc_ops psi_io_proc_ops = { 1607 .proc_open = psi_io_open, 1608 .proc_read = seq_read, 1609 .proc_lseek = seq_lseek, 1610 .proc_write = psi_io_write, 1611 .proc_poll = psi_fop_poll, 1612 .proc_release = psi_fop_release, 1613 }; 1614 1615 static const struct proc_ops psi_memory_proc_ops = { 1616 .proc_open = psi_memory_open, 1617 .proc_read = seq_read, 1618 .proc_lseek = seq_lseek, 1619 .proc_write = psi_memory_write, 1620 .proc_poll = psi_fop_poll, 1621 .proc_release = psi_fop_release, 1622 }; 1623 1624 static const struct proc_ops psi_cpu_proc_ops = { 1625 .proc_open = psi_cpu_open, 1626 .proc_read = seq_read, 1627 .proc_lseek = seq_lseek, 1628 .proc_write = psi_cpu_write, 1629 .proc_poll = psi_fop_poll, 1630 .proc_release = psi_fop_release, 1631 }; 1632 1633 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1634 static int psi_irq_show(struct seq_file *m, void *v) 1635 { 1636 return psi_show(m, &psi_system, PSI_IRQ); 1637 } 1638 1639 static int psi_irq_open(struct inode *inode, struct file *file) 1640 { 1641 return single_open(file, psi_irq_show, NULL); 1642 } 1643 1644 static ssize_t psi_irq_write(struct file *file, const char __user *user_buf, 1645 size_t nbytes, loff_t *ppos) 1646 { 1647 return psi_write(file, user_buf, nbytes, PSI_IRQ); 1648 } 1649 1650 static const struct proc_ops psi_irq_proc_ops = { 1651 .proc_open = psi_irq_open, 1652 .proc_read = seq_read, 1653 .proc_lseek = seq_lseek, 1654 .proc_write = psi_irq_write, 1655 .proc_poll = psi_fop_poll, 1656 .proc_release = psi_fop_release, 1657 }; 1658 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 1659 1660 static int __init psi_proc_init(void) 1661 { 1662 if (psi_enable) { 1663 proc_mkdir("pressure", NULL); 1664 proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops); 1665 proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops); 1666 proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops); 1667 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1668 proc_create("pressure/irq", 0666, NULL, &psi_irq_proc_ops); 1669 #endif 1670 } 1671 return 0; 1672 } 1673 module_init(psi_proc_init); 1674 1675 #endif /* CONFIG_PROC_FS */ 1676