1 /* 2 * buffered writeback throttling. loosely based on CoDel. We can't drop 3 * packets for IO scheduling, so the logic is something like this: 4 * 5 * - Monitor latencies in a defined window of time. 6 * - If the minimum latency in the above window exceeds some target, increment 7 * scaling step and scale down queue depth by a factor of 2x. The monitoring 8 * window is then shrunk to 100 / sqrt(scaling step + 1). 9 * - For any window where we don't have solid data on what the latencies 10 * look like, retain status quo. 11 * - If latencies look good, decrement scaling step. 12 * - If we're only doing writes, allow the scaling step to go negative. This 13 * will temporarily boost write performance, snapping back to a stable 14 * scaling step of 0 if reads show up or the heavy writers finish. Unlike 15 * positive scaling steps where we shrink the monitoring window, a negative 16 * scaling step retains the default step==0 window size. 17 * 18 * Copyright (C) 2016 Jens Axboe 19 * 20 */ 21 #include <linux/kernel.h> 22 #include <linux/blk_types.h> 23 #include <linux/slab.h> 24 #include <linux/backing-dev.h> 25 #include <linux/swap.h> 26 27 #include "blk-wbt.h" 28 29 #define CREATE_TRACE_POINTS 30 #include <trace/events/wbt.h> 31 32 enum { 33 /* 34 * Default setting, we'll scale up (to 75% of QD max) or down (min 1) 35 * from here depending on device stats 36 */ 37 RWB_DEF_DEPTH = 16, 38 39 /* 40 * 100msec window 41 */ 42 RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, 43 44 /* 45 * Disregard stats, if we don't meet this minimum 46 */ 47 RWB_MIN_WRITE_SAMPLES = 3, 48 49 /* 50 * If we have this number of consecutive windows with not enough 51 * information to scale up or down, scale up. 52 */ 53 RWB_UNKNOWN_BUMP = 5, 54 }; 55 56 static inline bool rwb_enabled(struct rq_wb *rwb) 57 { 58 return rwb && rwb->wb_normal != 0; 59 } 60 61 /* 62 * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded, 63 * false if 'v' + 1 would be bigger than 'below'. 64 */ 65 static bool atomic_inc_below(atomic_t *v, int below) 66 { 67 int cur = atomic_read(v); 68 69 for (;;) { 70 int old; 71 72 if (cur >= below) 73 return false; 74 old = atomic_cmpxchg(v, cur, cur + 1); 75 if (old == cur) 76 break; 77 cur = old; 78 } 79 80 return true; 81 } 82 83 static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) 84 { 85 if (rwb_enabled(rwb)) { 86 const unsigned long cur = jiffies; 87 88 if (cur != *var) 89 *var = cur; 90 } 91 } 92 93 /* 94 * If a task was rate throttled in balance_dirty_pages() within the last 95 * second or so, use that to indicate a higher cleaning rate. 96 */ 97 static bool wb_recent_wait(struct rq_wb *rwb) 98 { 99 struct bdi_writeback *wb = &rwb->queue->backing_dev_info.wb; 100 101 return time_before(jiffies, wb->dirty_sleep + HZ); 102 } 103 104 static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, bool is_kswapd) 105 { 106 return &rwb->rq_wait[is_kswapd]; 107 } 108 109 static void rwb_wake_all(struct rq_wb *rwb) 110 { 111 int i; 112 113 for (i = 0; i < WBT_NUM_RWQ; i++) { 114 struct rq_wait *rqw = &rwb->rq_wait[i]; 115 116 if (waitqueue_active(&rqw->wait)) 117 wake_up_all(&rqw->wait); 118 } 119 } 120 121 void __wbt_done(struct rq_wb *rwb, enum wbt_flags wb_acct) 122 { 123 struct rq_wait *rqw; 124 int inflight, limit; 125 126 if (!(wb_acct & WBT_TRACKED)) 127 return; 128 129 rqw = get_rq_wait(rwb, wb_acct & WBT_KSWAPD); 130 inflight = atomic_dec_return(&rqw->inflight); 131 132 /* 133 * wbt got disabled with IO in flight. Wake up any potential 134 * waiters, we don't have to do more than that. 135 */ 136 if (unlikely(!rwb_enabled(rwb))) { 137 rwb_wake_all(rwb); 138 return; 139 } 140 141 /* 142 * If the device does write back caching, drop further down 143 * before we wake people up. 144 */ 145 if (rwb->wc && !wb_recent_wait(rwb)) 146 limit = 0; 147 else 148 limit = rwb->wb_normal; 149 150 /* 151 * Don't wake anyone up if we are above the normal limit. 152 */ 153 if (inflight && inflight >= limit) 154 return; 155 156 if (waitqueue_active(&rqw->wait)) { 157 int diff = limit - inflight; 158 159 if (!inflight || diff >= rwb->wb_background / 2) 160 wake_up_all(&rqw->wait); 161 } 162 } 163 164 /* 165 * Called on completion of a request. Note that it's also called when 166 * a request is merged, when the request gets freed. 167 */ 168 void wbt_done(struct rq_wb *rwb, struct blk_issue_stat *stat) 169 { 170 if (!rwb) 171 return; 172 173 if (!wbt_is_tracked(stat)) { 174 if (rwb->sync_cookie == stat) { 175 rwb->sync_issue = 0; 176 rwb->sync_cookie = NULL; 177 } 178 179 if (wbt_is_read(stat)) 180 wb_timestamp(rwb, &rwb->last_comp); 181 wbt_clear_state(stat); 182 } else { 183 WARN_ON_ONCE(stat == rwb->sync_cookie); 184 __wbt_done(rwb, wbt_stat_to_mask(stat)); 185 wbt_clear_state(stat); 186 } 187 } 188 189 /* 190 * Return true, if we can't increase the depth further by scaling 191 */ 192 static bool calc_wb_limits(struct rq_wb *rwb) 193 { 194 unsigned int depth; 195 bool ret = false; 196 197 if (!rwb->min_lat_nsec) { 198 rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0; 199 return false; 200 } 201 202 /* 203 * For QD=1 devices, this is a special case. It's important for those 204 * to have one request ready when one completes, so force a depth of 205 * 2 for those devices. On the backend, it'll be a depth of 1 anyway, 206 * since the device can't have more than that in flight. If we're 207 * scaling down, then keep a setting of 1/1/1. 208 */ 209 if (rwb->queue_depth == 1) { 210 if (rwb->scale_step > 0) 211 rwb->wb_max = rwb->wb_normal = 1; 212 else { 213 rwb->wb_max = rwb->wb_normal = 2; 214 ret = true; 215 } 216 rwb->wb_background = 1; 217 } else { 218 /* 219 * scale_step == 0 is our default state. If we have suffered 220 * latency spikes, step will be > 0, and we shrink the 221 * allowed write depths. If step is < 0, we're only doing 222 * writes, and we allow a temporarily higher depth to 223 * increase performance. 224 */ 225 depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth); 226 if (rwb->scale_step > 0) 227 depth = 1 + ((depth - 1) >> min(31, rwb->scale_step)); 228 else if (rwb->scale_step < 0) { 229 unsigned int maxd = 3 * rwb->queue_depth / 4; 230 231 depth = 1 + ((depth - 1) << -rwb->scale_step); 232 if (depth > maxd) { 233 depth = maxd; 234 ret = true; 235 } 236 } 237 238 /* 239 * Set our max/normal/bg queue depths based on how far 240 * we have scaled down (->scale_step). 241 */ 242 rwb->wb_max = depth; 243 rwb->wb_normal = (rwb->wb_max + 1) / 2; 244 rwb->wb_background = (rwb->wb_max + 3) / 4; 245 } 246 247 return ret; 248 } 249 250 static inline bool stat_sample_valid(struct blk_rq_stat *stat) 251 { 252 /* 253 * We need at least one read sample, and a minimum of 254 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know 255 * that it's writes impacting us, and not just some sole read on 256 * a device that is in a lower power state. 257 */ 258 return stat[BLK_STAT_READ].nr_samples >= 1 && 259 stat[BLK_STAT_WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES; 260 } 261 262 static u64 rwb_sync_issue_lat(struct rq_wb *rwb) 263 { 264 u64 now, issue = ACCESS_ONCE(rwb->sync_issue); 265 266 if (!issue || !rwb->sync_cookie) 267 return 0; 268 269 now = ktime_to_ns(ktime_get()); 270 return now - issue; 271 } 272 273 enum { 274 LAT_OK = 1, 275 LAT_UNKNOWN, 276 LAT_UNKNOWN_WRITES, 277 LAT_EXCEEDED, 278 }; 279 280 static int __latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) 281 { 282 struct backing_dev_info *bdi = &rwb->queue->backing_dev_info; 283 u64 thislat; 284 285 /* 286 * If our stored sync issue exceeds the window size, or it 287 * exceeds our min target AND we haven't logged any entries, 288 * flag the latency as exceeded. wbt works off completion latencies, 289 * but for a flooded device, a single sync IO can take a long time 290 * to complete after being issued. If this time exceeds our 291 * monitoring window AND we didn't see any other completions in that 292 * window, then count that sync IO as a violation of the latency. 293 */ 294 thislat = rwb_sync_issue_lat(rwb); 295 if (thislat > rwb->cur_win_nsec || 296 (thislat > rwb->min_lat_nsec && !stat[BLK_STAT_READ].nr_samples)) { 297 trace_wbt_lat(bdi, thislat); 298 return LAT_EXCEEDED; 299 } 300 301 /* 302 * No read/write mix, if stat isn't valid 303 */ 304 if (!stat_sample_valid(stat)) { 305 /* 306 * If we had writes in this stat window and the window is 307 * current, we're only doing writes. If a task recently 308 * waited or still has writes in flights, consider us doing 309 * just writes as well. 310 */ 311 if ((stat[BLK_STAT_WRITE].nr_samples && blk_stat_is_current(stat)) || 312 wb_recent_wait(rwb) || wbt_inflight(rwb)) 313 return LAT_UNKNOWN_WRITES; 314 return LAT_UNKNOWN; 315 } 316 317 /* 318 * If the 'min' latency exceeds our target, step down. 319 */ 320 if (stat[BLK_STAT_READ].min > rwb->min_lat_nsec) { 321 trace_wbt_lat(bdi, stat[BLK_STAT_READ].min); 322 trace_wbt_stat(bdi, stat); 323 return LAT_EXCEEDED; 324 } 325 326 if (rwb->scale_step) 327 trace_wbt_stat(bdi, stat); 328 329 return LAT_OK; 330 } 331 332 static int latency_exceeded(struct rq_wb *rwb) 333 { 334 struct blk_rq_stat stat[2]; 335 336 blk_queue_stat_get(rwb->queue, stat); 337 return __latency_exceeded(rwb, stat); 338 } 339 340 static void rwb_trace_step(struct rq_wb *rwb, const char *msg) 341 { 342 struct backing_dev_info *bdi = &rwb->queue->backing_dev_info; 343 344 trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec, 345 rwb->wb_background, rwb->wb_normal, rwb->wb_max); 346 } 347 348 static void scale_up(struct rq_wb *rwb) 349 { 350 /* 351 * Hit max in previous round, stop here 352 */ 353 if (rwb->scaled_max) 354 return; 355 356 rwb->scale_step--; 357 rwb->unknown_cnt = 0; 358 blk_stat_clear(rwb->queue); 359 360 rwb->scaled_max = calc_wb_limits(rwb); 361 362 rwb_wake_all(rwb); 363 364 rwb_trace_step(rwb, "step up"); 365 } 366 367 /* 368 * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we 369 * had a latency violation. 370 */ 371 static void scale_down(struct rq_wb *rwb, bool hard_throttle) 372 { 373 /* 374 * Stop scaling down when we've hit the limit. This also prevents 375 * ->scale_step from going to crazy values, if the device can't 376 * keep up. 377 */ 378 if (rwb->wb_max == 1) 379 return; 380 381 if (rwb->scale_step < 0 && hard_throttle) 382 rwb->scale_step = 0; 383 else 384 rwb->scale_step++; 385 386 rwb->scaled_max = false; 387 rwb->unknown_cnt = 0; 388 blk_stat_clear(rwb->queue); 389 calc_wb_limits(rwb); 390 rwb_trace_step(rwb, "step down"); 391 } 392 393 static void rwb_arm_timer(struct rq_wb *rwb) 394 { 395 unsigned long expires; 396 397 if (rwb->scale_step > 0) { 398 /* 399 * We should speed this up, using some variant of a fast 400 * integer inverse square root calculation. Since we only do 401 * this for every window expiration, it's not a huge deal, 402 * though. 403 */ 404 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, 405 int_sqrt((rwb->scale_step + 1) << 8)); 406 } else { 407 /* 408 * For step < 0, we don't want to increase/decrease the 409 * window size. 410 */ 411 rwb->cur_win_nsec = rwb->win_nsec; 412 } 413 414 expires = jiffies + nsecs_to_jiffies(rwb->cur_win_nsec); 415 mod_timer(&rwb->window_timer, expires); 416 } 417 418 static void wb_timer_fn(unsigned long data) 419 { 420 struct rq_wb *rwb = (struct rq_wb *) data; 421 unsigned int inflight = wbt_inflight(rwb); 422 int status; 423 424 status = latency_exceeded(rwb); 425 426 trace_wbt_timer(&rwb->queue->backing_dev_info, status, rwb->scale_step, 427 inflight); 428 429 /* 430 * If we exceeded the latency target, step down. If we did not, 431 * step one level up. If we don't know enough to say either exceeded 432 * or ok, then don't do anything. 433 */ 434 switch (status) { 435 case LAT_EXCEEDED: 436 scale_down(rwb, true); 437 break; 438 case LAT_OK: 439 scale_up(rwb); 440 break; 441 case LAT_UNKNOWN_WRITES: 442 /* 443 * We started a the center step, but don't have a valid 444 * read/write sample, but we do have writes going on. 445 * Allow step to go negative, to increase write perf. 446 */ 447 scale_up(rwb); 448 break; 449 case LAT_UNKNOWN: 450 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) 451 break; 452 /* 453 * We get here when previously scaled reduced depth, and we 454 * currently don't have a valid read/write sample. For that 455 * case, slowly return to center state (step == 0). 456 */ 457 if (rwb->scale_step > 0) 458 scale_up(rwb); 459 else if (rwb->scale_step < 0) 460 scale_down(rwb, false); 461 break; 462 default: 463 break; 464 } 465 466 /* 467 * Re-arm timer, if we have IO in flight 468 */ 469 if (rwb->scale_step || inflight) 470 rwb_arm_timer(rwb); 471 } 472 473 void wbt_update_limits(struct rq_wb *rwb) 474 { 475 rwb->scale_step = 0; 476 rwb->scaled_max = false; 477 calc_wb_limits(rwb); 478 479 rwb_wake_all(rwb); 480 } 481 482 static bool close_io(struct rq_wb *rwb) 483 { 484 const unsigned long now = jiffies; 485 486 return time_before(now, rwb->last_issue + HZ / 10) || 487 time_before(now, rwb->last_comp + HZ / 10); 488 } 489 490 #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) 491 492 static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) 493 { 494 unsigned int limit; 495 496 /* 497 * At this point we know it's a buffered write. If this is 498 * kswapd trying to free memory, or REQ_SYNC is set, set, then 499 * it's WB_SYNC_ALL writeback, and we'll use the max limit for 500 * that. If the write is marked as a background write, then use 501 * the idle limit, or go to normal if we haven't had competing 502 * IO for a bit. 503 */ 504 if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) 505 limit = rwb->wb_max; 506 else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { 507 /* 508 * If less than 100ms since we completed unrelated IO, 509 * limit us to half the depth for background writeback. 510 */ 511 limit = rwb->wb_background; 512 } else 513 limit = rwb->wb_normal; 514 515 return limit; 516 } 517 518 static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw, 519 wait_queue_t *wait, unsigned long rw) 520 { 521 /* 522 * inc it here even if disabled, since we'll dec it at completion. 523 * this only happens if the task was sleeping in __wbt_wait(), 524 * and someone turned it off at the same time. 525 */ 526 if (!rwb_enabled(rwb)) { 527 atomic_inc(&rqw->inflight); 528 return true; 529 } 530 531 /* 532 * If the waitqueue is already active and we are not the next 533 * in line to be woken up, wait for our turn. 534 */ 535 if (waitqueue_active(&rqw->wait) && 536 rqw->wait.task_list.next != &wait->task_list) 537 return false; 538 539 return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw)); 540 } 541 542 /* 543 * Block if we will exceed our limit, or if we are currently waiting for 544 * the timer to kick off queuing again. 545 */ 546 static void __wbt_wait(struct rq_wb *rwb, unsigned long rw, spinlock_t *lock) 547 { 548 struct rq_wait *rqw = get_rq_wait(rwb, current_is_kswapd()); 549 DEFINE_WAIT(wait); 550 551 if (may_queue(rwb, rqw, &wait, rw)) 552 return; 553 554 do { 555 prepare_to_wait_exclusive(&rqw->wait, &wait, 556 TASK_UNINTERRUPTIBLE); 557 558 if (may_queue(rwb, rqw, &wait, rw)) 559 break; 560 561 if (lock) 562 spin_unlock_irq(lock); 563 564 io_schedule(); 565 566 if (lock) 567 spin_lock_irq(lock); 568 } while (1); 569 570 finish_wait(&rqw->wait, &wait); 571 } 572 573 static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio) 574 { 575 const int op = bio_op(bio); 576 577 /* 578 * If not a WRITE, do nothing 579 */ 580 if (op != REQ_OP_WRITE) 581 return false; 582 583 /* 584 * Don't throttle WRITE_ODIRECT 585 */ 586 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == (REQ_SYNC | REQ_IDLE)) 587 return false; 588 589 return true; 590 } 591 592 /* 593 * Returns true if the IO request should be accounted, false if not. 594 * May sleep, if we have exceeded the writeback limits. Caller can pass 595 * in an irq held spinlock, if it holds one when calling this function. 596 * If we do sleep, we'll release and re-grab it. 597 */ 598 unsigned int wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock) 599 { 600 unsigned int ret = 0; 601 602 if (!rwb_enabled(rwb)) 603 return 0; 604 605 if (bio_op(bio) == REQ_OP_READ) 606 ret = WBT_READ; 607 608 if (!wbt_should_throttle(rwb, bio)) { 609 if (ret & WBT_READ) 610 wb_timestamp(rwb, &rwb->last_issue); 611 return ret; 612 } 613 614 __wbt_wait(rwb, bio->bi_opf, lock); 615 616 if (!timer_pending(&rwb->window_timer)) 617 rwb_arm_timer(rwb); 618 619 if (current_is_kswapd()) 620 ret |= WBT_KSWAPD; 621 622 return ret | WBT_TRACKED; 623 } 624 625 void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat) 626 { 627 if (!rwb_enabled(rwb)) 628 return; 629 630 /* 631 * Track sync issue, in case it takes a long time to complete. Allows 632 * us to react quicker, if a sync IO takes a long time to complete. 633 * Note that this is just a hint. 'stat' can go away when the 634 * request completes, so it's important we never dereference it. We 635 * only use the address to compare with, which is why we store the 636 * sync_issue time locally. 637 */ 638 if (wbt_is_read(stat) && !rwb->sync_issue) { 639 rwb->sync_cookie = stat; 640 rwb->sync_issue = blk_stat_time(stat); 641 } 642 } 643 644 void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat) 645 { 646 if (!rwb_enabled(rwb)) 647 return; 648 if (stat == rwb->sync_cookie) { 649 rwb->sync_issue = 0; 650 rwb->sync_cookie = NULL; 651 } 652 } 653 654 void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth) 655 { 656 if (rwb) { 657 rwb->queue_depth = depth; 658 wbt_update_limits(rwb); 659 } 660 } 661 662 void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on) 663 { 664 if (rwb) 665 rwb->wc = write_cache_on; 666 } 667 668 /* 669 * Disable wbt, if enabled by default. Only called from CFQ, if we have 670 * cgroups enabled 671 */ 672 void wbt_disable_default(struct request_queue *q) 673 { 674 struct rq_wb *rwb = q->rq_wb; 675 676 if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT) { 677 del_timer_sync(&rwb->window_timer); 678 rwb->win_nsec = rwb->min_lat_nsec = 0; 679 wbt_update_limits(rwb); 680 } 681 } 682 EXPORT_SYMBOL_GPL(wbt_disable_default); 683 684 u64 wbt_default_latency_nsec(struct request_queue *q) 685 { 686 /* 687 * We default to 2msec for non-rotational storage, and 75msec 688 * for rotational storage. 689 */ 690 if (blk_queue_nonrot(q)) 691 return 2000000ULL; 692 else 693 return 75000000ULL; 694 } 695 696 int wbt_init(struct request_queue *q) 697 { 698 struct rq_wb *rwb; 699 int i; 700 701 /* 702 * For now, we depend on the stats window being larger than 703 * our monitoring window. Ensure that this isn't inadvertently 704 * violated. 705 */ 706 BUILD_BUG_ON(RWB_WINDOW_NSEC > BLK_STAT_NSEC); 707 BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS); 708 709 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); 710 if (!rwb) 711 return -ENOMEM; 712 713 for (i = 0; i < WBT_NUM_RWQ; i++) { 714 atomic_set(&rwb->rq_wait[i].inflight, 0); 715 init_waitqueue_head(&rwb->rq_wait[i].wait); 716 } 717 718 setup_timer(&rwb->window_timer, wb_timer_fn, (unsigned long) rwb); 719 rwb->wc = 1; 720 rwb->queue_depth = RWB_DEF_DEPTH; 721 rwb->last_comp = rwb->last_issue = jiffies; 722 rwb->queue = q; 723 rwb->win_nsec = RWB_WINDOW_NSEC; 724 rwb->enable_state = WBT_STATE_ON_DEFAULT; 725 wbt_update_limits(rwb); 726 727 /* 728 * Assign rwb, and turn on stats tracking for this queue 729 */ 730 q->rq_wb = rwb; 731 blk_stat_enable(q); 732 733 rwb->min_lat_nsec = wbt_default_latency_nsec(q); 734 735 wbt_set_queue_depth(rwb, blk_queue_depth(q)); 736 wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); 737 738 return 0; 739 } 740 741 void wbt_exit(struct request_queue *q) 742 { 743 struct rq_wb *rwb = q->rq_wb; 744 745 if (rwb) { 746 del_timer_sync(&rwb->window_timer); 747 q->rq_wb = NULL; 748 kfree(rwb); 749 } 750 } 751