block/blk-wbt.c

/* [<][>][^][v][top][bottom][index][help] */
This source file includes following definitions.
wbt_clear_state
wbt_is_tracked
wbt_is_read
rwb_enabled
wb_timestamp
wb_recent_wait
get_rq_wait
rwb_wake_all
wbt_rqw_done
__wbt_done
wbt_done
stat_sample_valid
rwb_sync_issue_lat
latency_exceeded
rwb_trace_step
calc_wb_limits
scale_up
scale_down
rwb_arm_timer
wb_timer_fn
__wbt_update_limits
wbt_update_limits
wbt_get_min_lat
wbt_set_min_lat
close_io
get_limit
wbt_inflight_cb
wbt_cleanup_cb
__wbt_wait
wbt_should_throttle
bio_to_wbt_flags
wbt_cleanup
wbt_wait
wbt_track
wbt_issue
wbt_requeue
wbt_set_write_cache
wbt_enable_default
wbt_default_latency_nsec
wbt_data_dir
wbt_queue_depth_changed
wbt_exit
wbt_disable_default
wbt_curr_win_nsec_show
wbt_enabled_show
wbt_id_show
wbt_inflight_show
wbt_min_lat_nsec_show
wbt_unknown_cnt_show
wbt_normal_show
wbt_background_show
wbt_init
   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * buffered writeback throttling. loosely based on CoDel. We can't drop
   4  * packets for IO scheduling, so the logic is something like this:
   5  *
   6  * - Monitor latencies in a defined window of time.
   7  * - If the minimum latency in the above window exceeds some target, increment
   8  *   scaling step and scale down queue depth by a factor of 2x. The monitoring
   9  *   window is then shrunk to 100 / sqrt(scaling step + 1).
  10  * - For any window where we don't have solid data on what the latencies
  11  *   look like, retain status quo.
  12  * - If latencies look good, decrement scaling step.
  13  * - If we're only doing writes, allow the scaling step to go negative. This
  14  *   will temporarily boost write performance, snapping back to a stable
  15  *   scaling step of 0 if reads show up or the heavy writers finish. Unlike
  16  *   positive scaling steps where we shrink the monitoring window, a negative
  17  *   scaling step retains the default step==0 window size.
  18  *
  19  * Copyright (C) 2016 Jens Axboe
  20  *
  21  */
  22 #include <linux/kernel.h>
  23 #include <linux/blk_types.h>
  24 #include <linux/slab.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/swap.h>
  27 
  28 #include "blk-wbt.h"
  29 #include "blk-rq-qos.h"
  30 
  31 #define CREATE_TRACE_POINTS
  32 #include <trace/events/wbt.h>
  33 
  34 static inline void wbt_clear_state(struct request *rq)
  35 {
  36         rq->wbt_flags = 0;
  37 }
  38 
  39 static inline enum wbt_flags wbt_flags(struct request *rq)
  40 {
  41         return rq->wbt_flags;
  42 }
  43 
  44 static inline bool wbt_is_tracked(struct request *rq)
  45 {
  46         return rq->wbt_flags & WBT_TRACKED;
  47 }
  48 
  49 static inline bool wbt_is_read(struct request *rq)
  50 {
  51         return rq->wbt_flags & WBT_READ;
  52 }
  53 
  54 enum {
  55         /*
  56          * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
  57          * from here depending on device stats
  58          */
  59         RWB_DEF_DEPTH   = 16,
  60 
  61         /*
  62          * 100msec window
  63          */
  64         RWB_WINDOW_NSEC         = 100 * 1000 * 1000ULL,
  65 
  66         /*
  67          * Disregard stats, if we don't meet this minimum
  68          */
  69         RWB_MIN_WRITE_SAMPLES   = 3,
  70 
  71         /*
  72          * If we have this number of consecutive windows with not enough
  73          * information to scale up or down, scale up.
  74          */
  75         RWB_UNKNOWN_BUMP        = 5,
  76 };
  77 
  78 static inline bool rwb_enabled(struct rq_wb *rwb)
  79 {
  80         return rwb && rwb->wb_normal != 0;
  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->rqos.q->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,
 105                                           enum wbt_flags wb_acct)
 106 {
 107         if (wb_acct & WBT_KSWAPD)
 108                 return &rwb->rq_wait[WBT_RWQ_KSWAPD];
 109         else if (wb_acct & WBT_DISCARD)
 110                 return &rwb->rq_wait[WBT_RWQ_DISCARD];
 111 
 112         return &rwb->rq_wait[WBT_RWQ_BG];
 113 }
 114 
 115 static void rwb_wake_all(struct rq_wb *rwb)
 116 {
 117         int i;
 118 
 119         for (i = 0; i < WBT_NUM_RWQ; i++) {
 120                 struct rq_wait *rqw = &rwb->rq_wait[i];
 121 
 122                 if (wq_has_sleeper(&rqw->wait))
 123                         wake_up_all(&rqw->wait);
 124         }
 125 }
 126 
 127 static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw,
 128                          enum wbt_flags wb_acct)
 129 {
 130         int inflight, limit;
 131 
 132         inflight = atomic_dec_return(&rqw->inflight);
 133 
 134         /*
 135          * wbt got disabled with IO in flight. Wake up any potential
 136          * waiters, we don't have to do more than that.
 137          */
 138         if (unlikely(!rwb_enabled(rwb))) {
 139                 rwb_wake_all(rwb);
 140                 return;
 141         }
 142 
 143         /*
 144          * For discards, our limit is always the background. For writes, if
 145          * the device does write back caching, drop further down before we
 146          * wake people up.
 147          */
 148         if (wb_acct & WBT_DISCARD)
 149                 limit = rwb->wb_background;
 150         else if (rwb->wc && !wb_recent_wait(rwb))
 151                 limit = 0;
 152         else
 153                 limit = rwb->wb_normal;
 154 
 155         /*
 156          * Don't wake anyone up if we are above the normal limit.
 157          */
 158         if (inflight && inflight >= limit)
 159                 return;
 160 
 161         if (wq_has_sleeper(&rqw->wait)) {
 162                 int diff = limit - inflight;
 163 
 164                 if (!inflight || diff >= rwb->wb_background / 2)
 165                         wake_up_all(&rqw->wait);
 166         }
 167 }
 168 
 169 static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct)
 170 {
 171         struct rq_wb *rwb = RQWB(rqos);
 172         struct rq_wait *rqw;
 173 
 174         if (!(wb_acct & WBT_TRACKED))
 175                 return;
 176 
 177         rqw = get_rq_wait(rwb, wb_acct);
 178         wbt_rqw_done(rwb, rqw, wb_acct);
 179 }
 180 
 181 /*
 182  * Called on completion of a request. Note that it's also called when
 183  * a request is merged, when the request gets freed.
 184  */
 185 static void wbt_done(struct rq_qos *rqos, struct request *rq)
 186 {
 187         struct rq_wb *rwb = RQWB(rqos);
 188 
 189         if (!wbt_is_tracked(rq)) {
 190                 if (rwb->sync_cookie == rq) {
 191                         rwb->sync_issue = 0;
 192                         rwb->sync_cookie = NULL;
 193                 }
 194 
 195                 if (wbt_is_read(rq))
 196                         wb_timestamp(rwb, &rwb->last_comp);
 197         } else {
 198                 WARN_ON_ONCE(rq == rwb->sync_cookie);
 199                 __wbt_done(rqos, wbt_flags(rq));
 200         }
 201         wbt_clear_state(rq);
 202 }
 203 
 204 static inline bool stat_sample_valid(struct blk_rq_stat *stat)
 205 {
 206         /*
 207          * We need at least one read sample, and a minimum of
 208          * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
 209          * that it's writes impacting us, and not just some sole read on
 210          * a device that is in a lower power state.
 211          */
 212         return (stat[READ].nr_samples >= 1 &&
 213                 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
 214 }
 215 
 216 static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
 217 {
 218         u64 now, issue = READ_ONCE(rwb->sync_issue);
 219 
 220         if (!issue || !rwb->sync_cookie)
 221                 return 0;
 222 
 223         now = ktime_to_ns(ktime_get());
 224         return now - issue;
 225 }
 226 
 227 enum {
 228         LAT_OK = 1,
 229         LAT_UNKNOWN,
 230         LAT_UNKNOWN_WRITES,
 231         LAT_EXCEEDED,
 232 };
 233 
 234 static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
 235 {
 236         struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info;
 237         struct rq_depth *rqd = &rwb->rq_depth;
 238         u64 thislat;
 239 
 240         /*
 241          * If our stored sync issue exceeds the window size, or it
 242          * exceeds our min target AND we haven't logged any entries,
 243          * flag the latency as exceeded. wbt works off completion latencies,
 244          * but for a flooded device, a single sync IO can take a long time
 245          * to complete after being issued. If this time exceeds our
 246          * monitoring window AND we didn't see any other completions in that
 247          * window, then count that sync IO as a violation of the latency.
 248          */
 249         thislat = rwb_sync_issue_lat(rwb);
 250         if (thislat > rwb->cur_win_nsec ||
 251             (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
 252                 trace_wbt_lat(bdi, thislat);
 253                 return LAT_EXCEEDED;
 254         }
 255 
 256         /*
 257          * No read/write mix, if stat isn't valid
 258          */
 259         if (!stat_sample_valid(stat)) {
 260                 /*
 261                  * If we had writes in this stat window and the window is
 262                  * current, we're only doing writes. If a task recently
 263                  * waited or still has writes in flights, consider us doing
 264                  * just writes as well.
 265                  */
 266                 if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
 267                     wbt_inflight(rwb))
 268                         return LAT_UNKNOWN_WRITES;
 269                 return LAT_UNKNOWN;
 270         }
 271 
 272         /*
 273          * If the 'min' latency exceeds our target, step down.
 274          */
 275         if (stat[READ].min > rwb->min_lat_nsec) {
 276                 trace_wbt_lat(bdi, stat[READ].min);
 277                 trace_wbt_stat(bdi, stat);
 278                 return LAT_EXCEEDED;
 279         }
 280 
 281         if (rqd->scale_step)
 282                 trace_wbt_stat(bdi, stat);
 283 
 284         return LAT_OK;
 285 }
 286 
 287 static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
 288 {
 289         struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info;
 290         struct rq_depth *rqd = &rwb->rq_depth;
 291 
 292         trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,
 293                         rwb->wb_background, rwb->wb_normal, rqd->max_depth);
 294 }
 295 
 296 static void calc_wb_limits(struct rq_wb *rwb)
 297 {
 298         if (rwb->min_lat_nsec == 0) {
 299                 rwb->wb_normal = rwb->wb_background = 0;
 300         } else if (rwb->rq_depth.max_depth <= 2) {
 301                 rwb->wb_normal = rwb->rq_depth.max_depth;
 302                 rwb->wb_background = 1;
 303         } else {
 304                 rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;
 305                 rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;
 306         }
 307 }
 308 
 309 static void scale_up(struct rq_wb *rwb)
 310 {
 311         if (!rq_depth_scale_up(&rwb->rq_depth))
 312                 return;
 313         calc_wb_limits(rwb);
 314         rwb->unknown_cnt = 0;
 315         rwb_wake_all(rwb);
 316         rwb_trace_step(rwb, "scale up");
 317 }
 318 
 319 static void scale_down(struct rq_wb *rwb, bool hard_throttle)
 320 {
 321         if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle))
 322                 return;
 323         calc_wb_limits(rwb);
 324         rwb->unknown_cnt = 0;
 325         rwb_trace_step(rwb, "scale down");
 326 }
 327 
 328 static void rwb_arm_timer(struct rq_wb *rwb)
 329 {
 330         struct rq_depth *rqd = &rwb->rq_depth;
 331 
 332         if (rqd->scale_step > 0) {
 333                 /*
 334                  * We should speed this up, using some variant of a fast
 335                  * integer inverse square root calculation. Since we only do
 336                  * this for every window expiration, it's not a huge deal,
 337                  * though.
 338                  */
 339                 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
 340                                         int_sqrt((rqd->scale_step + 1) << 8));
 341         } else {
 342                 /*
 343                  * For step < 0, we don't want to increase/decrease the
 344                  * window size.
 345                  */
 346                 rwb->cur_win_nsec = rwb->win_nsec;
 347         }
 348 
 349         blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
 350 }
 351 
 352 static void wb_timer_fn(struct blk_stat_callback *cb)
 353 {
 354         struct rq_wb *rwb = cb->data;
 355         struct rq_depth *rqd = &rwb->rq_depth;
 356         unsigned int inflight = wbt_inflight(rwb);
 357         int status;
 358 
 359         status = latency_exceeded(rwb, cb->stat);
 360 
 361         trace_wbt_timer(rwb->rqos.q->backing_dev_info, status, rqd->scale_step,
 362                         inflight);
 363 
 364         /*
 365          * If we exceeded the latency target, step down. If we did not,
 366          * step one level up. If we don't know enough to say either exceeded
 367          * or ok, then don't do anything.
 368          */
 369         switch (status) {
 370         case LAT_EXCEEDED:
 371                 scale_down(rwb, true);
 372                 break;
 373         case LAT_OK:
 374                 scale_up(rwb);
 375                 break;
 376         case LAT_UNKNOWN_WRITES:
 377                 /*
 378                  * We started a the center step, but don't have a valid
 379                  * read/write sample, but we do have writes going on.
 380                  * Allow step to go negative, to increase write perf.
 381                  */
 382                 scale_up(rwb);
 383                 break;
 384         case LAT_UNKNOWN:
 385                 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
 386                         break;
 387                 /*
 388                  * We get here when previously scaled reduced depth, and we
 389                  * currently don't have a valid read/write sample. For that
 390                  * case, slowly return to center state (step == 0).
 391                  */
 392                 if (rqd->scale_step > 0)
 393                         scale_up(rwb);
 394                 else if (rqd->scale_step < 0)
 395                         scale_down(rwb, false);
 396                 break;
 397         default:
 398                 break;
 399         }
 400 
 401         /*
 402          * Re-arm timer, if we have IO in flight
 403          */
 404         if (rqd->scale_step || inflight)
 405                 rwb_arm_timer(rwb);
 406 }
 407 
 408 static void __wbt_update_limits(struct rq_wb *rwb)
 409 {
 410         struct rq_depth *rqd = &rwb->rq_depth;
 411 
 412         rqd->scale_step = 0;
 413         rqd->scaled_max = false;
 414 
 415         rq_depth_calc_max_depth(rqd);
 416         calc_wb_limits(rwb);
 417 
 418         rwb_wake_all(rwb);
 419 }
 420 
 421 void wbt_update_limits(struct request_queue *q)
 422 {
 423         struct rq_qos *rqos = wbt_rq_qos(q);
 424         if (!rqos)
 425                 return;
 426         __wbt_update_limits(RQWB(rqos));
 427 }
 428 
 429 u64 wbt_get_min_lat(struct request_queue *q)
 430 {
 431         struct rq_qos *rqos = wbt_rq_qos(q);
 432         if (!rqos)
 433                 return 0;
 434         return RQWB(rqos)->min_lat_nsec;
 435 }
 436 
 437 void wbt_set_min_lat(struct request_queue *q, u64 val)
 438 {
 439         struct rq_qos *rqos = wbt_rq_qos(q);
 440         if (!rqos)
 441                 return;
 442         RQWB(rqos)->min_lat_nsec = val;
 443         RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;
 444         __wbt_update_limits(RQWB(rqos));
 445 }
 446 
 447 
 448 static bool close_io(struct rq_wb *rwb)
 449 {
 450         const unsigned long now = jiffies;
 451 
 452         return time_before(now, rwb->last_issue + HZ / 10) ||
 453                 time_before(now, rwb->last_comp + HZ / 10);
 454 }
 455 
 456 #define REQ_HIPRIO      (REQ_SYNC | REQ_META | REQ_PRIO)
 457 
 458 static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
 459 {
 460         unsigned int limit;
 461 
 462         /*
 463          * If we got disabled, just return UINT_MAX. This ensures that
 464          * we'll properly inc a new IO, and dec+wakeup at the end.
 465          */
 466         if (!rwb_enabled(rwb))
 467                 return UINT_MAX;
 468 
 469         if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD)
 470                 return rwb->wb_background;
 471 
 472         /*
 473          * At this point we know it's a buffered write. If this is
 474          * kswapd trying to free memory, or REQ_SYNC is set, then
 475          * it's WB_SYNC_ALL writeback, and we'll use the max limit for
 476          * that. If the write is marked as a background write, then use
 477          * the idle limit, or go to normal if we haven't had competing
 478          * IO for a bit.
 479          */
 480         if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
 481                 limit = rwb->rq_depth.max_depth;
 482         else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
 483                 /*
 484                  * If less than 100ms since we completed unrelated IO,
 485                  * limit us to half the depth for background writeback.
 486                  */
 487                 limit = rwb->wb_background;
 488         } else
 489                 limit = rwb->wb_normal;
 490 
 491         return limit;
 492 }
 493 
 494 struct wbt_wait_data {
 495         struct rq_wb *rwb;
 496         enum wbt_flags wb_acct;
 497         unsigned long rw;
 498 };
 499 
 500 static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data)
 501 {
 502         struct wbt_wait_data *data = private_data;
 503         return rq_wait_inc_below(rqw, get_limit(data->rwb, data->rw));
 504 }
 505 
 506 static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data)
 507 {
 508         struct wbt_wait_data *data = private_data;
 509         wbt_rqw_done(data->rwb, rqw, data->wb_acct);
 510 }
 511 
 512 /*
 513  * Block if we will exceed our limit, or if we are currently waiting for
 514  * the timer to kick off queuing again.
 515  */
 516 static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct,
 517                        unsigned long rw)
 518 {
 519         struct rq_wait *rqw = get_rq_wait(rwb, wb_acct);
 520         struct wbt_wait_data data = {
 521                 .rwb = rwb,
 522                 .wb_acct = wb_acct,
 523                 .rw = rw,
 524         };
 525 
 526         rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb);
 527 }
 528 
 529 static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
 530 {
 531         switch (bio_op(bio)) {
 532         case REQ_OP_WRITE:
 533                 /*
 534                  * Don't throttle WRITE_ODIRECT
 535                  */
 536                 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
 537                     (REQ_SYNC | REQ_IDLE))
 538                         return false;
 539                 /* fallthrough */
 540         case REQ_OP_DISCARD:
 541                 return true;
 542         default:
 543                 return false;
 544         }
 545 }
 546 
 547 static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio)
 548 {
 549         enum wbt_flags flags = 0;
 550 
 551         if (!rwb_enabled(rwb))
 552                 return 0;
 553 
 554         if (bio_op(bio) == REQ_OP_READ) {
 555                 flags = WBT_READ;
 556         } else if (wbt_should_throttle(rwb, bio)) {
 557                 if (current_is_kswapd())
 558                         flags |= WBT_KSWAPD;
 559                 if (bio_op(bio) == REQ_OP_DISCARD)
 560                         flags |= WBT_DISCARD;
 561                 flags |= WBT_TRACKED;
 562         }
 563         return flags;
 564 }
 565 
 566 static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio)
 567 {
 568         struct rq_wb *rwb = RQWB(rqos);
 569         enum wbt_flags flags = bio_to_wbt_flags(rwb, bio);
 570         __wbt_done(rqos, flags);
 571 }
 572 
 573 /*
 574  * Returns true if the IO request should be accounted, false if not.
 575  * May sleep, if we have exceeded the writeback limits. Caller can pass
 576  * in an irq held spinlock, if it holds one when calling this function.
 577  * If we do sleep, we'll release and re-grab it.
 578  */
 579 static void wbt_wait(struct rq_qos *rqos, struct bio *bio)
 580 {
 581         struct rq_wb *rwb = RQWB(rqos);
 582         enum wbt_flags flags;
 583 
 584         flags = bio_to_wbt_flags(rwb, bio);
 585         if (!(flags & WBT_TRACKED)) {
 586                 if (flags & WBT_READ)
 587                         wb_timestamp(rwb, &rwb->last_issue);
 588                 return;
 589         }
 590 
 591         __wbt_wait(rwb, flags, bio->bi_opf);
 592 
 593         if (!blk_stat_is_active(rwb->cb))
 594                 rwb_arm_timer(rwb);
 595 }
 596 
 597 static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
 598 {
 599         struct rq_wb *rwb = RQWB(rqos);
 600         rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);
 601 }
 602 
 603 static void wbt_issue(struct rq_qos *rqos, struct request *rq)
 604 {
 605         struct rq_wb *rwb = RQWB(rqos);
 606 
 607         if (!rwb_enabled(rwb))
 608                 return;
 609 
 610         /*
 611          * Track sync issue, in case it takes a long time to complete. Allows us
 612          * to react quicker, if a sync IO takes a long time to complete. Note
 613          * that this is just a hint. The request can go away when it completes,
 614          * so it's important we never dereference it. We only use the address to
 615          * compare with, which is why we store the sync_issue time locally.
 616          */
 617         if (wbt_is_read(rq) && !rwb->sync_issue) {
 618                 rwb->sync_cookie = rq;
 619                 rwb->sync_issue = rq->io_start_time_ns;
 620         }
 621 }
 622 
 623 static void wbt_requeue(struct rq_qos *rqos, struct request *rq)
 624 {
 625         struct rq_wb *rwb = RQWB(rqos);
 626         if (!rwb_enabled(rwb))
 627                 return;
 628         if (rq == rwb->sync_cookie) {
 629                 rwb->sync_issue = 0;
 630                 rwb->sync_cookie = NULL;
 631         }
 632 }
 633 
 634 void wbt_set_write_cache(struct request_queue *q, bool write_cache_on)
 635 {
 636         struct rq_qos *rqos = wbt_rq_qos(q);
 637         if (rqos)
 638                 RQWB(rqos)->wc = write_cache_on;
 639 }
 640 
 641 /*
 642  * Enable wbt if defaults are configured that way
 643  */
 644 void wbt_enable_default(struct request_queue *q)
 645 {
 646         struct rq_qos *rqos = wbt_rq_qos(q);
 647         /* Throttling already enabled? */
 648         if (rqos)
 649                 return;
 650 
 651         /* Queue not registered? Maybe shutting down... */
 652         if (!blk_queue_registered(q))
 653                 return;
 654 
 655         if (queue_is_mq(q) && IS_ENABLED(CONFIG_BLK_WBT_MQ))
 656                 wbt_init(q);
 657 }
 658 EXPORT_SYMBOL_GPL(wbt_enable_default);
 659 
 660 u64 wbt_default_latency_nsec(struct request_queue *q)
 661 {
 662         /*
 663          * We default to 2msec for non-rotational storage, and 75msec
 664          * for rotational storage.
 665          */
 666         if (blk_queue_nonrot(q))
 667                 return 2000000ULL;
 668         else
 669                 return 75000000ULL;
 670 }
 671 
 672 static int wbt_data_dir(const struct request *rq)
 673 {
 674         const int op = req_op(rq);
 675 
 676         if (op == REQ_OP_READ)
 677                 return READ;
 678         else if (op_is_write(op))
 679                 return WRITE;
 680 
 681         /* don't account */
 682         return -1;
 683 }
 684 
 685 static void wbt_queue_depth_changed(struct rq_qos *rqos)
 686 {
 687         RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->q);
 688         __wbt_update_limits(RQWB(rqos));
 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         .queue_depth_changed = wbt_queue_depth_changed,
 814         .exit = wbt_exit,
 815 #ifdef CONFIG_BLK_DEBUG_FS
 816         .debugfs_attrs = wbt_debugfs_attrs,
 817 #endif
 818 };
 819 
 820 int wbt_init(struct request_queue *q)
 821 {
 822         struct rq_wb *rwb;
 823         int i;
 824 
 825         rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
 826         if (!rwb)
 827                 return -ENOMEM;
 828 
 829         rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
 830         if (!rwb->cb) {
 831                 kfree(rwb);
 832                 return -ENOMEM;
 833         }
 834 
 835         for (i = 0; i < WBT_NUM_RWQ; i++)
 836                 rq_wait_init(&rwb->rq_wait[i]);
 837 
 838         rwb->rqos.id = RQ_QOS_WBT;
 839         rwb->rqos.ops = &wbt_rqos_ops;
 840         rwb->rqos.q = q;
 841         rwb->last_comp = rwb->last_issue = jiffies;
 842         rwb->win_nsec = RWB_WINDOW_NSEC;
 843         rwb->enable_state = WBT_STATE_ON_DEFAULT;
 844         rwb->wc = 1;
 845         rwb->rq_depth.default_depth = RWB_DEF_DEPTH;
 846         __wbt_update_limits(rwb);
 847 
 848         /*
 849          * Assign rwb and add the stats callback.
 850          */
 851         rq_qos_add(q, &rwb->rqos);
 852         blk_stat_add_callback(q, rwb->cb);
 853 
 854         rwb->min_lat_nsec = wbt_default_latency_nsec(q);
 855 
 856         wbt_queue_depth_changed(&rwb->rqos);
 857         wbt_set_write_cache(q, test_bit(QUEUE_FLAG_WC, &q->queue_flags));
 858 
 859         return 0;
 860 }
/* [<][>][^][v][top][bottom][index][help] */
root/block/blk-wbt.c

DEFINITIONS
