block/blk-wbt.c

   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 #include "blk-rq-qos.h"
  29
  30 #define CREATE_TRACE_POINTS
  31 #include <trace/events/wbt.h>
  32
  33 static inline void wbt_clear_state(struct request *rq)
  34 {
  35         rq->wbt_flags = 0;
  36 }
  37
  38 static inline enum wbt_flags wbt_flags(struct request *rq)
  39 {
  40         return rq->wbt_flags;
  41 }
  42
  43 static inline bool wbt_is_tracked(struct request *rq)
  44 {
  45         return rq->wbt_flags & WBT_TRACKED;
  46 }
  47
  48 static inline bool wbt_is_read(struct request *rq)
  49 {
  50         return rq->wbt_flags & WBT_READ;
  51 }
  52
  53 enum {
  54         /*
  55          * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
  56          * from here depending on device stats
  57          */
  58         RWB_DEF_DEPTH   = 16,
  59
  60         /*
  61          * 100msec window
  62          */
  63         RWB_WINDOW_NSEC         = 100 * 1000 * 1000ULL,
  64
  65         /*
  66          * Disregard stats, if we don't meet this minimum
  67          */
  68         RWB_MIN_WRITE_SAMPLES   = 3,
  69
  70         /*
  71          * If we have this number of consecutive windows with not enough
  72          * information to scale up or down, scale up.
  73          */
  74         RWB_UNKNOWN_BUMP        = 5,
  75 };
  76
  77 static inline bool rwb_enabled(struct rq_wb *rwb)
  78 {
  79         return rwb && rwb->wb_normal != 0;
  80 }
  81
  82 static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
  83 {
  84         if (rwb_enabled(rwb)) {
  85                 const unsigned long cur = jiffies;
  86
  87                 if (cur != *var)
  88                         *var = cur;
  89         }
  90 }
  91
  92 /*
  93  * If a task was rate throttled in balance_dirty_pages() within the last
  94  * second or so, use that to indicate a higher cleaning rate.
  95  */
  96 static bool wb_recent_wait(struct rq_wb *rwb)
  97 {
  98         struct bdi_writeback *wb = &rwb->rqos.q->backing_dev_info->wb;
  99
 100         return time_before(jiffies, wb->dirty_sleep + HZ);
 101 }
 102
 103 static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb,
 104                                           enum wbt_flags wb_acct)
 105 {
 106         if (wb_acct & WBT_KSWAPD)
 107                 return &rwb->rq_wait[WBT_RWQ_KSWAPD];
 108         else if (wb_acct & WBT_DISCARD)
 109                 return &rwb->rq_wait[WBT_RWQ_DISCARD];
 110
 111         return &rwb->rq_wait[WBT_RWQ_BG];
 112 }
 113
 114 static void rwb_wake_all(struct rq_wb *rwb)
 115 {
 116         int i;
 117
 118         for (i = 0; i < WBT_NUM_RWQ; i++) {
 119                 struct rq_wait *rqw = &rwb->rq_wait[i];
 120
 121                 if (wq_has_sleeper(&rqw->wait))
 122                         wake_up_all(&rqw->wait);
 123         }
 124 }
 125
 126 static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw,
 127                          enum wbt_flags wb_acct)
 128 {
 129         int inflight, limit;
 130
 131         inflight = atomic_dec_return(&rqw->inflight);
 132
 133         /*
 134          * wbt got disabled with IO in flight. Wake up any potential
 135          * waiters, we don't have to do more than that.
 136          */
 137         if (unlikely(!rwb_enabled(rwb))) {
 138                 rwb_wake_all(rwb);
 139                 return;
 140         }
 141
 142         /*
 143          * For discards, our limit is always the background. For writes, if
 144          * the device does write back caching, drop further down before we
 145          * wake people up.
 146          */
 147         if (wb_acct & WBT_DISCARD)
 148                 limit = rwb->wb_background;
 149         else if (rwb->wc && !wb_recent_wait(rwb))
 150                 limit = 0;
 151         else
 152                 limit = rwb->wb_normal;
 153
 154         /*
 155          * Don't wake anyone up if we are above the normal limit.
 156          */
 157         if (inflight && inflight >= limit)
 158                 return;
 159
 160         if (wq_has_sleeper(&rqw->wait)) {
 161                 int diff = limit - inflight;
 162
 163                 if (!inflight || diff >= rwb->wb_background / 2)
 164                         wake_up_all(&rqw->wait);
 165         }
 166 }
 167
 168 static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct)
 169 {
 170         struct rq_wb *rwb = RQWB(rqos);
 171         struct rq_wait *rqw;
 172
 173         if (!(wb_acct & WBT_TRACKED))
 174                 return;
 175
 176         rqw = get_rq_wait(rwb, wb_acct);
 177         wbt_rqw_done(rwb, rqw, wb_acct);
 178 }
 179
 180 /*
 181  * Called on completion of a request. Note that it's also called when
 182  * a request is merged, when the request gets freed.
 183  */
 184 static void wbt_done(struct rq_qos *rqos, struct request *rq)
 185 {
 186         struct rq_wb *rwb = RQWB(rqos);
 187
 188         if (!wbt_is_tracked(rq)) {
 189                 if (rwb->sync_cookie == rq) {
 190                         rwb->sync_issue = 0;
 191                         rwb->sync_cookie = NULL;
 192                 }
 193
 194                 if (wbt_is_read(rq))
 195                         wb_timestamp(rwb, &rwb->last_comp);
 196         } else {
 197                 WARN_ON_ONCE(rq == rwb->sync_cookie);
 198                 __wbt_done(rqos, wbt_flags(rq));
 199         }
 200         wbt_clear_state(rq);
 201 }
 202
 203 static inline bool stat_sample_valid(struct blk_rq_stat *stat)
 204 {
 205         /*
 206          * We need at least one read sample, and a minimum of
 207          * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
 208          * that it's writes impacting us, and not just some sole read on
 209          * a device that is in a lower power state.
 210          */
 211         return (stat[READ].nr_samples >= 1 &&
 212                 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
 213 }
 214
 215 static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
 216 {
 217         u64 now, issue = READ_ONCE(rwb->sync_issue);
 218
 219         if (!issue || !rwb->sync_cookie)
 220                 return 0;
 221
 222         now = ktime_to_ns(ktime_get());
 223         return now - issue;
 224 }
 225
 226 enum {
 227         LAT_OK = 1,
 228         LAT_UNKNOWN,
 229         LAT_UNKNOWN_WRITES,
 230         LAT_EXCEEDED,
 231 };
 232
 233 static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
 234 {
 235         struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info;
 236         struct rq_depth *rqd = &rwb->rq_depth;
 237         u64 thislat;
 238
 239         /*
 240          * If our stored sync issue exceeds the window size, or it
 241          * exceeds our min target AND we haven't logged any entries,
 242          * flag the latency as exceeded. wbt works off completion latencies,
 243          * but for a flooded device, a single sync IO can take a long time
 244          * to complete after being issued. If this time exceeds our
 245          * monitoring window AND we didn't see any other completions in that
 246          * window, then count that sync IO as a violation of the latency.
 247          */
 248         thislat = rwb_sync_issue_lat(rwb);
 249         if (thislat > rwb->cur_win_nsec ||
 250             (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
 251                 trace_wbt_lat(bdi, thislat);
 252                 return LAT_EXCEEDED;
 253         }
 254
 255         /*
 256          * No read/write mix, if stat isn't valid
 257          */
 258         if (!stat_sample_valid(stat)) {
 259                 /*
 260                  * If we had writes in this stat window and the window is
 261                  * current, we're only doing writes. If a task recently
 262                  * waited or still has writes in flights, consider us doing
 263                  * just writes as well.
 264                  */
 265                 if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
 266                     wbt_inflight(rwb))
 267                         return LAT_UNKNOWN_WRITES;
 268                 return LAT_UNKNOWN;
 269         }
 270
 271         /*
 272          * If the 'min' latency exceeds our target, step down.
 273          */
 274         if (stat[READ].min > rwb->min_lat_nsec) {
 275                 trace_wbt_lat(bdi, stat[READ].min);
 276                 trace_wbt_stat(bdi, stat);
 277                 return LAT_EXCEEDED;
 278         }
 279
 280         if (rqd->scale_step)
 281                 trace_wbt_stat(bdi, stat);
 282
 283         return LAT_OK;
 284 }
 285
 286 static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
 287 {
 288         struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info;
 289         struct rq_depth *rqd = &rwb->rq_depth;
 290
 291         trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,
 292                         rwb->wb_background, rwb->wb_normal, rqd->max_depth);
 293 }
 294
 295 static void calc_wb_limits(struct rq_wb *rwb)
 296 {
 297         if (rwb->min_lat_nsec == 0) {
 298                 rwb->wb_normal = rwb->wb_background = 0;
 299         } else if (rwb->rq_depth.max_depth <= 2) {
 300                 rwb->wb_normal = rwb->rq_depth.max_depth;
 301                 rwb->wb_background = 1;
 302         } else {
 303                 rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;
 304                 rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;
 305         }
 306 }
 307
 308 static void scale_up(struct rq_wb *rwb)
 309 {
 310         rq_depth_scale_up(&rwb->rq_depth);
 311         calc_wb_limits(rwb);
 312         rwb->unknown_cnt = 0;
 313         rwb_wake_all(rwb);
 314         rwb_trace_step(rwb, "scale up");
 315 }
 316
 317 static void scale_down(struct rq_wb *rwb, bool hard_throttle)
 318 {
 319         rq_depth_scale_down(&rwb->rq_depth, hard_throttle);
 320         calc_wb_limits(rwb);
 321         rwb->unknown_cnt = 0;
 322         rwb_trace_step(rwb, "scale down");
 323 }
 324
 325 static void rwb_arm_timer(struct rq_wb *rwb)
 326 {
 327         struct rq_depth *rqd = &rwb->rq_depth;
 328
 329         if (rqd->scale_step > 0) {
 330                 /*
 331                  * We should speed this up, using some variant of a fast
 332                  * integer inverse square root calculation. Since we only do
 333                  * this for every window expiration, it's not a huge deal,
 334                  * though.
 335                  */
 336                 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
 337                                         int_sqrt((rqd->scale_step + 1) << 8));
 338         } else {
 339                 /*
 340                  * For step < 0, we don't want to increase/decrease the
 341                  * window size.
 342                  */
 343                 rwb->cur_win_nsec = rwb->win_nsec;
 344         }
 345
 346         blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
 347 }
 348
 349 static void wb_timer_fn(struct blk_stat_callback *cb)
 350 {
 351         struct rq_wb *rwb = cb->data;
 352         struct rq_depth *rqd = &rwb->rq_depth;
 353         unsigned int inflight = wbt_inflight(rwb);
 354         int status;
 355
 356         status = latency_exceeded(rwb, cb->stat);
 357
 358         trace_wbt_timer(rwb->rqos.q->backing_dev_info, status, rqd->scale_step,
 359                         inflight);
 360
 361         /*
 362          * If we exceeded the latency target, step down. If we did not,
 363          * step one level up. If we don't know enough to say either exceeded
 364          * or ok, then don't do anything.
 365          */
 366         switch (status) {
 367         case LAT_EXCEEDED:
 368                 scale_down(rwb, true);
 369                 break;
 370         case LAT_OK:
 371                 scale_up(rwb);
 372                 break;
 373         case LAT_UNKNOWN_WRITES:
 374                 /*
 375                  * We started a the center step, but don't have a valid
 376                  * read/write sample, but we do have writes going on.
 377                  * Allow step to go negative, to increase write perf.
 378                  */
 379                 scale_up(rwb);
 380                 break;
 381         case LAT_UNKNOWN:
 382                 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
 383                         break;
 384                 /*
 385                  * We get here when previously scaled reduced depth, and we
 386                  * currently don't have a valid read/write sample. For that
 387                  * case, slowly return to center state (step == 0).
 388                  */
 389                 if (rqd->scale_step > 0)
 390                         scale_up(rwb);
 391                 else if (rqd->scale_step < 0)
 392                         scale_down(rwb, false);
 393                 break;
 394         default:
 395                 break;
 396         }
 397
 398         /*
 399          * Re-arm timer, if we have IO in flight
 400          */
 401         if (rqd->scale_step || inflight)
 402                 rwb_arm_timer(rwb);
 403 }
 404
 405 static void __wbt_update_limits(struct rq_wb *rwb)
 406 {
 407         struct rq_depth *rqd = &rwb->rq_depth;
 408
 409         rqd->scale_step = 0;
 410         rqd->scaled_max = false;
 411
 412         rq_depth_calc_max_depth(rqd);
 413         calc_wb_limits(rwb);
 414
 415         rwb_wake_all(rwb);
 416 }
 417
 418 void wbt_update_limits(struct request_queue *q)
 419 {
 420         struct rq_qos *rqos = wbt_rq_qos(q);
 421         if (!rqos)
 422                 return;
 423         __wbt_update_limits(RQWB(rqos));
 424 }
 425
 426 u64 wbt_get_min_lat(struct request_queue *q)
 427 {
 428         struct rq_qos *rqos = wbt_rq_qos(q);
 429         if (!rqos)
 430                 return 0;
 431         return RQWB(rqos)->min_lat_nsec;
 432 }
 433
 434 void wbt_set_min_lat(struct request_queue *q, u64 val)
 435 {
 436         struct rq_qos *rqos = wbt_rq_qos(q);
 437         if (!rqos)
 438                 return;
 439         RQWB(rqos)->min_lat_nsec = val;
 440         RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;
 441         __wbt_update_limits(RQWB(rqos));
 442 }
 443
 444
 445 static bool close_io(struct rq_wb *rwb)
 446 {
 447         const unsigned long now = jiffies;
 448
 449         return time_before(now, rwb->last_issue + HZ / 10) ||
 450                 time_before(now, rwb->last_comp + HZ / 10);
 451 }
 452
 453 #define REQ_HIPRIO      (REQ_SYNC | REQ_META | REQ_PRIO)
 454
 455 static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
 456 {
 457         unsigned int limit;
 458
 459         /*
 460          * If we got disabled, just return UINT_MAX. This ensures that
 461          * we'll properly inc a new IO, and dec+wakeup at the end.
 462          */
 463         if (!rwb_enabled(rwb))
 464                 return UINT_MAX;
 465
 466         if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD)
 467                 return rwb->wb_background;
 468
 469         /*
 470          * At this point we know it's a buffered write. If this is
 471          * kswapd trying to free memory, or REQ_SYNC is set, then
 472          * it's WB_SYNC_ALL writeback, and we'll use the max limit for
 473          * that. If the write is marked as a background write, then use
 474          * the idle limit, or go to normal if we haven't had competing
 475          * IO for a bit.
 476          */
 477         if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
 478                 limit = rwb->rq_depth.max_depth;
 479         else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
 480                 /*
 481                  * If less than 100ms since we completed unrelated IO,
 482                  * limit us to half the depth for background writeback.
 483                  */
 484                 limit = rwb->wb_background;
 485         } else
 486                 limit = rwb->wb_normal;
 487
 488         return limit;
 489 }
 490
 491 struct wbt_wait_data {
 492         struct rq_wb *rwb;
 493         enum wbt_flags wb_acct;
 494         unsigned long rw;
 495 };
 496
 497 static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data)
 498 {
 499         struct wbt_wait_data *data = private_data;
 500         return rq_wait_inc_below(rqw, get_limit(data->rwb, data->rw));
 501 }
 502
 503 static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data)
 504 {
 505         struct wbt_wait_data *data = private_data;
 506         wbt_rqw_done(data->rwb, rqw, data->wb_acct);
 507 }
 508
 509 /*
 510  * Block if we will exceed our limit, or if we are currently waiting for
 511  * the timer to kick off queuing again.
 512  */
 513 static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct,
 514                        unsigned long rw)
 515 {
 516         struct rq_wait *rqw = get_rq_wait(rwb, wb_acct);
 517         struct wbt_wait_data data = {
 518                 .rwb = rwb,
 519                 .wb_acct = wb_acct,
 520                 .rw = rw,
 521         };
 522
 523         rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb);
 524 }
 525
 526 static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
 527 {
 528         switch (bio_op(bio)) {
 529         case REQ_OP_WRITE:
 530                 /*
 531                  * Don't throttle WRITE_ODIRECT
 532                  */
 533                 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
 534                     (REQ_SYNC | REQ_IDLE))
 535                         return false;
 536                 /* fallthrough */
 537         case REQ_OP_DISCARD:
 538                 return true;
 539         default:
 540                 return false;
 541         }
 542 }
 543
 544 static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio)
 545 {
 546         enum wbt_flags flags = 0;
 547
 548         if (!rwb_enabled(rwb))
 549                 return 0;
 550
 551         if (bio_op(bio) == REQ_OP_READ) {
 552                 flags = WBT_READ;
 553         } else if (wbt_should_throttle(rwb, bio)) {
 554                 if (current_is_kswapd())
 555                         flags |= WBT_KSWAPD;
 556                 if (bio_op(bio) == REQ_OP_DISCARD)
 557                         flags |= WBT_DISCARD;
 558                 flags |= WBT_TRACKED;
 559         }
 560         return flags;
 561 }
 562
 563 static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio)
 564 {
 565         struct rq_wb *rwb = RQWB(rqos);
 566         enum wbt_flags flags = bio_to_wbt_flags(rwb, bio);
 567         __wbt_done(rqos, flags);
 568 }
 569
 570 /*
 571  * Returns true if the IO request should be accounted, false if not.
 572  * May sleep, if we have exceeded the writeback limits. Caller can pass
 573  * in an irq held spinlock, if it holds one when calling this function.
 574  * If we do sleep, we'll release and re-grab it.
 575  */
 576 static void wbt_wait(struct rq_qos *rqos, struct bio *bio)
 577 {
 578         struct rq_wb *rwb = RQWB(rqos);
 579         enum wbt_flags flags;
 580
 581         flags = bio_to_wbt_flags(rwb, bio);
 582         if (!(flags & WBT_TRACKED)) {
 583                 if (flags & WBT_READ)
 584                         wb_timestamp(rwb, &rwb->last_issue);
 585                 return;
 586         }
 587
 588         __wbt_wait(rwb, flags, bio->bi_opf);
 589
 590         if (!blk_stat_is_active(rwb->cb))
 591                 rwb_arm_timer(rwb);
 592 }
 593
 594 static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
 595 {
 596         struct rq_wb *rwb = RQWB(rqos);
 597         rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);
 598 }
 599
 600 void wbt_issue(struct rq_qos *rqos, struct request *rq)
 601 {
 602         struct rq_wb *rwb = RQWB(rqos);
 603
 604         if (!rwb_enabled(rwb))
 605                 return;
 606
 607         /*
 608          * Track sync issue, in case it takes a long time to complete. Allows us
 609          * to react quicker, if a sync IO takes a long time to complete. Note
 610          * that this is just a hint. The request can go away when it completes,
 611          * so it's important we never dereference it. We only use the address to
 612          * compare with, which is why we store the sync_issue time locally.
 613          */
 614         if (wbt_is_read(rq) && !rwb->sync_issue) {
 615                 rwb->sync_cookie = rq;
 616                 rwb->sync_issue = rq->io_start_time_ns;
 617         }
 618 }
 619
 620 void wbt_requeue(struct rq_qos *rqos, struct request *rq)
 621 {
 622         struct rq_wb *rwb = RQWB(rqos);
 623         if (!rwb_enabled(rwb))
 624                 return;
 625         if (rq == rwb->sync_cookie) {
 626                 rwb->sync_issue = 0;
 627                 rwb->sync_cookie = NULL;
 628         }
 629 }
 630
 631 void wbt_set_queue_depth(struct request_queue *q, unsigned int depth)
 632 {
 633         struct rq_qos *rqos = wbt_rq_qos(q);
 634         if (rqos) {
 635                 RQWB(rqos)->rq_depth.queue_depth = depth;
 636                 __wbt_update_limits(RQWB(rqos));
 637         }
 638 }
 639
 640 void wbt_set_write_cache(struct request_queue *q, bool write_cache_on)
 641 {
 642         struct rq_qos *rqos = wbt_rq_qos(q);
 643         if (rqos)
 644                 RQWB(rqos)->wc = write_cache_on;
 645 }
 646
 647 /*
 648  * Enable wbt if defaults are configured that way
 649  */
 650 void wbt_enable_default(struct request_queue *q)
 651 {
 652         struct rq_qos *rqos = wbt_rq_qos(q);
 653         /* Throttling already enabled? */
 654         if (rqos)
 655                 return;
 656
 657         /* Queue not registered? Maybe shutting down... */
 658         if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
 659                 return;
 660
 661         if (queue_is_mq(q) && IS_ENABLED(CONFIG_BLK_WBT_MQ))
 662                 wbt_init(q);
 663 }
 664 EXPORT_SYMBOL_GPL(wbt_enable_default);
 665
 666 u64 wbt_default_latency_nsec(struct request_queue *q)
 667 {
 668         /*
 669          * We default to 2msec for non-rotational storage, and 75msec
 670          * for rotational storage.
 671          */
 672         if (blk_queue_nonrot(q))
 673                 return 2000000ULL;
 674         else
 675                 return 75000000ULL;
 676 }
 677
 678 static int wbt_data_dir(const struct request *rq)
 679 {
 680         const int op = req_op(rq);
 681
 682         if (op == REQ_OP_READ)
 683                 return READ;
 684         else if (op_is_write(op))
 685                 return WRITE;
 686
 687         /* don't account */
 688         return -1;
 689 }
 690
 691 static void wbt_exit(struct rq_qos *rqos)
 692 {
 693         struct rq_wb *rwb = RQWB(rqos);
 694         struct request_queue *q = rqos->q;
 695
 696         blk_stat_remove_callback(q, rwb->cb);
 697         blk_stat_free_callback(rwb->cb);
 698         kfree(rwb);
 699 }
 700
 701 /*
 702  * Disable wbt, if enabled by default.
 703  */
 704 void wbt_disable_default(struct request_queue *q)
 705 {
 706         struct rq_qos *rqos = wbt_rq_qos(q);
 707         struct rq_wb *rwb;
 708         if (!rqos)
 709                 return;
 710         rwb = RQWB(rqos);
 711         if (rwb->enable_state == WBT_STATE_ON_DEFAULT) {
 712                 blk_stat_deactivate(rwb->cb);
 713                 rwb->wb_normal = 0;
 714         }
 715 }
 716 EXPORT_SYMBOL_GPL(wbt_disable_default);
 717
 718 #ifdef CONFIG_BLK_DEBUG_FS
 719 static int wbt_curr_win_nsec_show(void *data, struct seq_file *m)
 720 {
 721         struct rq_qos *rqos = data;
 722         struct rq_wb *rwb = RQWB(rqos);
 723
 724         seq_printf(m, "%llu\n", rwb->cur_win_nsec);
 725         return 0;
 726 }
 727
 728 static int wbt_enabled_show(void *data, struct seq_file *m)
 729 {
 730         struct rq_qos *rqos = data;
 731         struct rq_wb *rwb = RQWB(rqos);
 732
 733         seq_printf(m, "%d\n", rwb->enable_state);
 734         return 0;
 735 }
 736
 737 static int wbt_id_show(void *data, struct seq_file *m)
 738 {
 739         struct rq_qos *rqos = data;
 740
 741         seq_printf(m, "%u\n", rqos->id);
 742         return 0;
 743 }
 744
 745 static int wbt_inflight_show(void *data, struct seq_file *m)
 746 {
 747         struct rq_qos *rqos = data;
 748         struct rq_wb *rwb = RQWB(rqos);
 749         int i;
 750
 751         for (i = 0; i < WBT_NUM_RWQ; i++)
 752                 seq_printf(m, "%d: inflight %d\n", i,
 753                            atomic_read(&rwb->rq_wait[i].inflight));
 754         return 0;
 755 }
 756
 757 static int wbt_min_lat_nsec_show(void *data, struct seq_file *m)
 758 {
 759         struct rq_qos *rqos = data;
 760         struct rq_wb *rwb = RQWB(rqos);
 761
 762         seq_printf(m, "%lu\n", rwb->min_lat_nsec);
 763         return 0;
 764 }
 765
 766 static int wbt_unknown_cnt_show(void *data, struct seq_file *m)
 767 {
 768         struct rq_qos *rqos = data;
 769         struct rq_wb *rwb = RQWB(rqos);
 770
 771         seq_printf(m, "%u\n", rwb->unknown_cnt);
 772         return 0;
 773 }
 774
 775 static int wbt_normal_show(void *data, struct seq_file *m)
 776 {
 777         struct rq_qos *rqos = data;
 778         struct rq_wb *rwb = RQWB(rqos);
 779
 780         seq_printf(m, "%u\n", rwb->wb_normal);
 781         return 0;
 782 }
 783
 784 static int wbt_background_show(void *data, struct seq_file *m)
 785 {
 786         struct rq_qos *rqos = data;
 787         struct rq_wb *rwb = RQWB(rqos);
 788
 789         seq_printf(m, "%u\n", rwb->wb_background);
 790         return 0;
 791 }
 792
 793 static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = {
 794         {"curr_win_nsec", 0400, wbt_curr_win_nsec_show},
 795         {"enabled", 0400, wbt_enabled_show},
 796         {"id", 0400, wbt_id_show},
 797         {"inflight", 0400, wbt_inflight_show},
 798         {"min_lat_nsec", 0400, wbt_min_lat_nsec_show},
 799         {"unknown_cnt", 0400, wbt_unknown_cnt_show},
 800         {"wb_normal", 0400, wbt_normal_show},
 801         {"wb_background", 0400, wbt_background_show},
 802         {},
 803 };
 804 #endif
 805
 806 static struct rq_qos_ops wbt_rqos_ops = {
 807         .throttle = wbt_wait,
 808         .issue = wbt_issue,
 809         .track = wbt_track,
 810         .requeue = wbt_requeue,
 811         .done = wbt_done,
 812         .cleanup = wbt_cleanup,
 813         .exit = wbt_exit,
 814 #ifdef CONFIG_BLK_DEBUG_FS
 815         .debugfs_attrs = wbt_debugfs_attrs,
 816 #endif
 817 };
 818
 819 int wbt_init(struct request_queue *q)
 820 {
 821         struct rq_wb *rwb;
 822         int i;
 823
 824         rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
 825         if (!rwb)
 826                 return -ENOMEM;
 827
 828         rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
 829         if (!rwb->cb) {
 830                 kfree(rwb);
 831                 return -ENOMEM;
 832         }
 833
 834         for (i = 0; i < WBT_NUM_RWQ; i++)
 835                 rq_wait_init(&rwb->rq_wait[i]);
 836
 837         rwb->rqos.id = RQ_QOS_WBT;
 838         rwb->rqos.ops = &wbt_rqos_ops;
 839         rwb->rqos.q = q;
 840         rwb->last_comp = rwb->last_issue = jiffies;
 841         rwb->win_nsec = RWB_WINDOW_NSEC;
 842         rwb->enable_state = WBT_STATE_ON_DEFAULT;
 843         rwb->wc = 1;
 844         rwb->rq_depth.default_depth = RWB_DEF_DEPTH;
 845         __wbt_update_limits(rwb);
 846
 847         /*
 848          * Assign rwb and add the stats callback.
 849          */
 850         rq_qos_add(q, &rwb->rqos);
 851         blk_stat_add_callback(q, rwb->cb);
 852
 853         rwb->min_lat_nsec = wbt_default_latency_nsec(q);
 854
 855         wbt_set_queue_depth(q, blk_queue_depth(q));
 856         wbt_set_write_cache(q, test_bit(QUEUE_FLAG_WC, &q->queue_flags));
 857
 858         return 0;
 859 }