block/mq-deadline.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  MQ Deadline i/o scheduler - adaptation of the legacy deadline scheduler,
   4  *  for the blk-mq scheduling framework
   5  *
   6  *  Copyright (C) 2016 Jens Axboe <axboe@kernel.dk>
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/fs.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/blk-mq.h>
  12 #include <linux/bio.h>
  13 #include <linux/module.h>
  14 #include <linux/slab.h>
  15 #include <linux/init.h>
  16 #include <linux/compiler.h>
  17 #include <linux/rbtree.h>
  18 #include <linux/sbitmap.h>
  19
  20 #include <trace/events/block.h>
  21
  22 #include "elevator.h"
  23 #include "blk.h"
  24 #include "blk-mq.h"
  25 #include "blk-mq-debugfs.h"
  26 #include "blk-mq-tag.h"
  27 #include "blk-mq-sched.h"
  28
  29 /*
  30  * See Documentation/block/deadline-iosched.rst
  31  */
  32 static const int read_expire = HZ / 2;  /* max time before a read is submitted. */
  33 static const int write_expire = 5 * HZ; /* ditto for writes, these limits are SOFT! */
  34 /*
  35  * Time after which to dispatch lower priority requests even if higher
  36  * priority requests are pending.
  37  */
  38 static const int prio_aging_expire = 10 * HZ;
  39 static const int writes_starved = 2;    /* max times reads can starve a write */
  40 static const int fifo_batch = 16;       /* # of sequential requests treated as one
  41                                      by the above parameters. For throughput. */
  42
  43 enum dd_data_dir {
  44         DD_READ         = READ,
  45         DD_WRITE        = WRITE,
  46 };
  47
  48 enum { DD_DIR_COUNT = 2 };
  49
  50 enum dd_prio {
  51         DD_RT_PRIO      = 0,
  52         DD_BE_PRIO      = 1,
  53         DD_IDLE_PRIO    = 2,
  54         DD_PRIO_MAX     = 2,
  55 };
  56
  57 enum { DD_PRIO_COUNT = 3 };
  58
  59 /*
  60  * I/O statistics per I/O priority. It is fine if these counters overflow.
  61  * What matters is that these counters are at least as wide as
  62  * log2(max_outstanding_requests).
  63  */
  64 struct io_stats_per_prio {
  65         uint32_t inserted;
  66         uint32_t merged;
  67         uint32_t dispatched;
  68         atomic_t completed;
  69 };
  70
  71 /*
  72  * Deadline scheduler data per I/O priority (enum dd_prio). Requests are
  73  * present on both sort_list[] and fifo_list[].
  74  */
  75 struct dd_per_prio {
  76         struct list_head dispatch;
  77         struct rb_root sort_list[DD_DIR_COUNT];
  78         struct list_head fifo_list[DD_DIR_COUNT];
  79         /* Next request in FIFO order. Read, write or both are NULL. */
  80         struct request *next_rq[DD_DIR_COUNT];
  81         struct io_stats_per_prio stats;
  82 };
  83
  84 struct deadline_data {
  85         /*
  86          * run time data
  87          */
  88
  89         struct dd_per_prio per_prio[DD_PRIO_COUNT];
  90
  91         /* Data direction of latest dispatched request. */
  92         enum dd_data_dir last_dir;
  93         unsigned int batching;          /* number of sequential requests made */
  94         unsigned int starved;           /* times reads have starved writes */
  95
  96         /*
  97          * settings that change how the i/o scheduler behaves
  98          */
  99         int fifo_expire[DD_DIR_COUNT];
 100         int fifo_batch;
 101         int writes_starved;
 102         int front_merges;
 103         u32 async_depth;
 104         int prio_aging_expire;
 105
 106         spinlock_t lock;
 107         spinlock_t zone_lock;
 108 };
 109
 110 /* Maps an I/O priority class to a deadline scheduler priority. */
 111 static const enum dd_prio ioprio_class_to_prio[] = {
 112         [IOPRIO_CLASS_NONE]     = DD_BE_PRIO,
 113         [IOPRIO_CLASS_RT]       = DD_RT_PRIO,
 114         [IOPRIO_CLASS_BE]       = DD_BE_PRIO,
 115         [IOPRIO_CLASS_IDLE]     = DD_IDLE_PRIO,
 116 };
 117
 118 static inline struct rb_root *
 119 deadline_rb_root(struct dd_per_prio *per_prio, struct request *rq)
 120 {
 121         return &per_prio->sort_list[rq_data_dir(rq)];
 122 }
 123
 124 /*
 125  * Returns the I/O priority class (IOPRIO_CLASS_*) that has been assigned to a
 126  * request.
 127  */
 128 static u8 dd_rq_ioclass(struct request *rq)
 129 {
 130         return IOPRIO_PRIO_CLASS(req_get_ioprio(rq));
 131 }
 132
 133 /*
 134  * get the request before `rq' in sector-sorted order
 135  */
 136 static inline struct request *
 137 deadline_earlier_request(struct request *rq)
 138 {
 139         struct rb_node *node = rb_prev(&rq->rb_node);
 140
 141         if (node)
 142                 return rb_entry_rq(node);
 143
 144         return NULL;
 145 }
 146
 147 /*
 148  * get the request after `rq' in sector-sorted order
 149  */
 150 static inline struct request *
 151 deadline_latter_request(struct request *rq)
 152 {
 153         struct rb_node *node = rb_next(&rq->rb_node);
 154
 155         if (node)
 156                 return rb_entry_rq(node);
 157
 158         return NULL;
 159 }
 160
 161 static void
 162 deadline_add_rq_rb(struct dd_per_prio *per_prio, struct request *rq)
 163 {
 164         struct rb_root *root = deadline_rb_root(per_prio, rq);
 165
 166         elv_rb_add(root, rq);
 167 }
 168
 169 static inline void
 170 deadline_del_rq_rb(struct dd_per_prio *per_prio, struct request *rq)
 171 {
 172         const enum dd_data_dir data_dir = rq_data_dir(rq);
 173
 174         if (per_prio->next_rq[data_dir] == rq)
 175                 per_prio->next_rq[data_dir] = deadline_latter_request(rq);
 176
 177         elv_rb_del(deadline_rb_root(per_prio, rq), rq);
 178 }
 179
 180 /*
 181  * remove rq from rbtree and fifo.
 182  */
 183 static void deadline_remove_request(struct request_queue *q,
 184                                     struct dd_per_prio *per_prio,
 185                                     struct request *rq)
 186 {
 187         list_del_init(&rq->queuelist);
 188
 189         /*
 190          * We might not be on the rbtree, if we are doing an insert merge
 191          */
 192         if (!RB_EMPTY_NODE(&rq->rb_node))
 193                 deadline_del_rq_rb(per_prio, rq);
 194
 195         elv_rqhash_del(q, rq);
 196         if (q->last_merge == rq)
 197                 q->last_merge = NULL;
 198 }
 199
 200 static void dd_request_merged(struct request_queue *q, struct request *req,
 201                               enum elv_merge type)
 202 {
 203         struct deadline_data *dd = q->elevator->elevator_data;
 204         const u8 ioprio_class = dd_rq_ioclass(req);
 205         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 206         struct dd_per_prio *per_prio = &dd->per_prio[prio];
 207
 208         /*
 209          * if the merge was a front merge, we need to reposition request
 210          */
 211         if (type == ELEVATOR_FRONT_MERGE) {
 212                 elv_rb_del(deadline_rb_root(per_prio, req), req);
 213                 deadline_add_rq_rb(per_prio, req);
 214         }
 215 }
 216
 217 /*
 218  * Callback function that is invoked after @next has been merged into @req.
 219  */
 220 static void dd_merged_requests(struct request_queue *q, struct request *req,
 221                                struct request *next)
 222 {
 223         struct deadline_data *dd = q->elevator->elevator_data;
 224         const u8 ioprio_class = dd_rq_ioclass(next);
 225         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 226
 227         lockdep_assert_held(&dd->lock);
 228
 229         dd->per_prio[prio].stats.merged++;
 230
 231         /*
 232          * if next expires before rq, assign its expire time to rq
 233          * and move into next position (next will be deleted) in fifo
 234          */
 235         if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
 236                 if (time_before((unsigned long)next->fifo_time,
 237                                 (unsigned long)req->fifo_time)) {
 238                         list_move(&req->queuelist, &next->queuelist);
 239                         req->fifo_time = next->fifo_time;
 240                 }
 241         }
 242
 243         /*
 244          * kill knowledge of next, this one is a goner
 245          */
 246         deadline_remove_request(q, &dd->per_prio[prio], next);
 247 }
 248
 249 /*
 250  * move an entry to dispatch queue
 251  */
 252 static void
 253 deadline_move_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 254                       struct request *rq)
 255 {
 256         const enum dd_data_dir data_dir = rq_data_dir(rq);
 257
 258         per_prio->next_rq[data_dir] = deadline_latter_request(rq);
 259
 260         /*
 261          * take it off the sort and fifo list
 262          */
 263         deadline_remove_request(rq->q, per_prio, rq);
 264 }
 265
 266 /* Number of requests queued for a given priority level. */
 267 static u32 dd_queued(struct deadline_data *dd, enum dd_prio prio)
 268 {
 269         const struct io_stats_per_prio *stats = &dd->per_prio[prio].stats;
 270
 271         lockdep_assert_held(&dd->lock);
 272
 273         return stats->inserted - atomic_read(&stats->completed);
 274 }
 275
 276 /*
 277  * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
 278  * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
 279  */
 280 static inline int deadline_check_fifo(struct dd_per_prio *per_prio,
 281                                       enum dd_data_dir data_dir)
 282 {
 283         struct request *rq = rq_entry_fifo(per_prio->fifo_list[data_dir].next);
 284
 285         /*
 286          * rq is expired!
 287          */
 288         if (time_after_eq(jiffies, (unsigned long)rq->fifo_time))
 289                 return 1;
 290
 291         return 0;
 292 }
 293
 294 /*
 295  * Check if rq has a sequential request preceding it.
 296  */
 297 static bool deadline_is_seq_write(struct deadline_data *dd, struct request *rq)
 298 {
 299         struct request *prev = deadline_earlier_request(rq);
 300
 301         if (!prev)
 302                 return false;
 303
 304         return blk_rq_pos(prev) + blk_rq_sectors(prev) == blk_rq_pos(rq);
 305 }
 306
 307 /*
 308  * Skip all write requests that are sequential from @rq, even if we cross
 309  * a zone boundary.
 310  */
 311 static struct request *deadline_skip_seq_writes(struct deadline_data *dd,
 312                                                 struct request *rq)
 313 {
 314         sector_t pos = blk_rq_pos(rq);
 315         sector_t skipped_sectors = 0;
 316
 317         while (rq) {
 318                 if (blk_rq_pos(rq) != pos + skipped_sectors)
 319                         break;
 320                 skipped_sectors += blk_rq_sectors(rq);
 321                 rq = deadline_latter_request(rq);
 322         }
 323
 324         return rq;
 325 }
 326
 327 /*
 328  * For the specified data direction, return the next request to
 329  * dispatch using arrival ordered lists.
 330  */
 331 static struct request *
 332 deadline_fifo_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 333                       enum dd_data_dir data_dir)
 334 {
 335         struct request *rq;
 336         unsigned long flags;
 337
 338         if (list_empty(&per_prio->fifo_list[data_dir]))
 339                 return NULL;
 340
 341         rq = rq_entry_fifo(per_prio->fifo_list[data_dir].next);
 342         if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q))
 343                 return rq;
 344
 345         /*
 346          * Look for a write request that can be dispatched, that is one with
 347          * an unlocked target zone. For some HDDs, breaking a sequential
 348          * write stream can lead to lower throughput, so make sure to preserve
 349          * sequential write streams, even if that stream crosses into the next
 350          * zones and these zones are unlocked.
 351          */
 352         spin_lock_irqsave(&dd->zone_lock, flags);
 353         list_for_each_entry(rq, &per_prio->fifo_list[DD_WRITE], queuelist) {
 354                 if (blk_req_can_dispatch_to_zone(rq) &&
 355                     (blk_queue_nonrot(rq->q) ||
 356                      !deadline_is_seq_write(dd, rq)))
 357                         goto out;
 358         }
 359         rq = NULL;
 360 out:
 361         spin_unlock_irqrestore(&dd->zone_lock, flags);
 362
 363         return rq;
 364 }
 365
 366 /*
 367  * For the specified data direction, return the next request to
 368  * dispatch using sector position sorted lists.
 369  */
 370 static struct request *
 371 deadline_next_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 372                       enum dd_data_dir data_dir)
 373 {
 374         struct request *rq;
 375         unsigned long flags;
 376
 377         rq = per_prio->next_rq[data_dir];
 378         if (!rq)
 379                 return NULL;
 380
 381         if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q))
 382                 return rq;
 383
 384         /*
 385          * Look for a write request that can be dispatched, that is one with
 386          * an unlocked target zone. For some HDDs, breaking a sequential
 387          * write stream can lead to lower throughput, so make sure to preserve
 388          * sequential write streams, even if that stream crosses into the next
 389          * zones and these zones are unlocked.
 390          */
 391         spin_lock_irqsave(&dd->zone_lock, flags);
 392         while (rq) {
 393                 if (blk_req_can_dispatch_to_zone(rq))
 394                         break;
 395                 if (blk_queue_nonrot(rq->q))
 396                         rq = deadline_latter_request(rq);
 397                 else
 398                         rq = deadline_skip_seq_writes(dd, rq);
 399         }
 400         spin_unlock_irqrestore(&dd->zone_lock, flags);
 401
 402         return rq;
 403 }
 404
 405 /*
 406  * Returns true if and only if @rq started after @latest_start where
 407  * @latest_start is in jiffies.
 408  */
 409 static bool started_after(struct deadline_data *dd, struct request *rq,
 410                           unsigned long latest_start)
 411 {
 412         unsigned long start_time = (unsigned long)rq->fifo_time;
 413
 414         start_time -= dd->fifo_expire[rq_data_dir(rq)];
 415
 416         return time_after(start_time, latest_start);
 417 }
 418
 419 /*
 420  * deadline_dispatch_requests selects the best request according to
 421  * read/write expire, fifo_batch, etc and with a start time <= @latest_start.
 422  */
 423 static struct request *__dd_dispatch_request(struct deadline_data *dd,
 424                                              struct dd_per_prio *per_prio,
 425                                              unsigned long latest_start)
 426 {
 427         struct request *rq, *next_rq;
 428         enum dd_data_dir data_dir;
 429         enum dd_prio prio;
 430         u8 ioprio_class;
 431
 432         lockdep_assert_held(&dd->lock);
 433
 434         if (!list_empty(&per_prio->dispatch)) {
 435                 rq = list_first_entry(&per_prio->dispatch, struct request,
 436                                       queuelist);
 437                 if (started_after(dd, rq, latest_start))
 438                         return NULL;
 439                 list_del_init(&rq->queuelist);
 440                 goto done;
 441         }
 442
 443         /*
 444          * batches are currently reads XOR writes
 445          */
 446         rq = deadline_next_request(dd, per_prio, dd->last_dir);
 447         if (rq && dd->batching < dd->fifo_batch)
 448                 /* we have a next request are still entitled to batch */
 449                 goto dispatch_request;
 450
 451         /*
 452          * at this point we are not running a batch. select the appropriate
 453          * data direction (read / write)
 454          */
 455
 456         if (!list_empty(&per_prio->fifo_list[DD_READ])) {
 457                 BUG_ON(RB_EMPTY_ROOT(&per_prio->sort_list[DD_READ]));
 458
 459                 if (deadline_fifo_request(dd, per_prio, DD_WRITE) &&
 460                     (dd->starved++ >= dd->writes_starved))
 461                         goto dispatch_writes;
 462
 463                 data_dir = DD_READ;
 464
 465                 goto dispatch_find_request;
 466         }
 467
 468         /*
 469          * there are either no reads or writes have been starved
 470          */
 471
 472         if (!list_empty(&per_prio->fifo_list[DD_WRITE])) {
 473 dispatch_writes:
 474                 BUG_ON(RB_EMPTY_ROOT(&per_prio->sort_list[DD_WRITE]));
 475
 476                 dd->starved = 0;
 477
 478                 data_dir = DD_WRITE;
 479
 480                 goto dispatch_find_request;
 481         }
 482
 483         return NULL;
 484
 485 dispatch_find_request:
 486         /*
 487          * we are not running a batch, find best request for selected data_dir
 488          */
 489         next_rq = deadline_next_request(dd, per_prio, data_dir);
 490         if (deadline_check_fifo(per_prio, data_dir) || !next_rq) {
 491                 /*
 492                  * A deadline has expired, the last request was in the other
 493                  * direction, or we have run out of higher-sectored requests.
 494                  * Start again from the request with the earliest expiry time.
 495                  */
 496                 rq = deadline_fifo_request(dd, per_prio, data_dir);
 497         } else {
 498                 /*
 499                  * The last req was the same dir and we have a next request in
 500                  * sort order. No expired requests so continue on from here.
 501                  */
 502                 rq = next_rq;
 503         }
 504
 505         /*
 506          * For a zoned block device, if we only have writes queued and none of
 507          * them can be dispatched, rq will be NULL.
 508          */
 509         if (!rq)
 510                 return NULL;
 511
 512         dd->last_dir = data_dir;
 513         dd->batching = 0;
 514
 515 dispatch_request:
 516         if (started_after(dd, rq, latest_start))
 517                 return NULL;
 518
 519         /*
 520          * rq is the selected appropriate request.
 521          */
 522         dd->batching++;
 523         deadline_move_request(dd, per_prio, rq);
 524 done:
 525         ioprio_class = dd_rq_ioclass(rq);
 526         prio = ioprio_class_to_prio[ioprio_class];
 527         dd->per_prio[prio].stats.dispatched++;
 528         /*
 529          * If the request needs its target zone locked, do it.
 530          */
 531         blk_req_zone_write_lock(rq);
 532         rq->rq_flags |= RQF_STARTED;
 533         return rq;
 534 }
 535
 536 /*
 537  * Check whether there are any requests with priority other than DD_RT_PRIO
 538  * that were inserted more than prio_aging_expire jiffies ago.
 539  */
 540 static struct request *dd_dispatch_prio_aged_requests(struct deadline_data *dd,
 541                                                       unsigned long now)
 542 {
 543         struct request *rq;
 544         enum dd_prio prio;
 545         int prio_cnt;
 546
 547         lockdep_assert_held(&dd->lock);
 548
 549         prio_cnt = !!dd_queued(dd, DD_RT_PRIO) + !!dd_queued(dd, DD_BE_PRIO) +
 550                    !!dd_queued(dd, DD_IDLE_PRIO);
 551         if (prio_cnt < 2)
 552                 return NULL;
 553
 554         for (prio = DD_BE_PRIO; prio <= DD_PRIO_MAX; prio++) {
 555                 rq = __dd_dispatch_request(dd, &dd->per_prio[prio],
 556                                            now - dd->prio_aging_expire);
 557                 if (rq)
 558                         return rq;
 559         }
 560
 561         return NULL;
 562 }
 563
 564 /*
 565  * Called from blk_mq_run_hw_queue() -> __blk_mq_sched_dispatch_requests().
 566  *
 567  * One confusing aspect here is that we get called for a specific
 568  * hardware queue, but we may return a request that is for a
 569  * different hardware queue. This is because mq-deadline has shared
 570  * state for all hardware queues, in terms of sorting, FIFOs, etc.
 571  */
 572 static struct request *dd_dispatch_request(struct blk_mq_hw_ctx *hctx)
 573 {
 574         struct deadline_data *dd = hctx->queue->elevator->elevator_data;
 575         const unsigned long now = jiffies;
 576         struct request *rq;
 577         enum dd_prio prio;
 578
 579         spin_lock(&dd->lock);
 580         rq = dd_dispatch_prio_aged_requests(dd, now);
 581         if (rq)
 582                 goto unlock;
 583
 584         /*
 585          * Next, dispatch requests in priority order. Ignore lower priority
 586          * requests if any higher priority requests are pending.
 587          */
 588         for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 589                 rq = __dd_dispatch_request(dd, &dd->per_prio[prio], now);
 590                 if (rq || dd_queued(dd, prio))
 591                         break;
 592         }
 593
 594 unlock:
 595         spin_unlock(&dd->lock);
 596
 597         return rq;
 598 }
 599
 600 /*
 601  * Called by __blk_mq_alloc_request(). The shallow_depth value set by this
 602  * function is used by __blk_mq_get_tag().
 603  */
 604 static void dd_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
 605 {
 606         struct deadline_data *dd = data->q->elevator->elevator_data;
 607
 608         /* Do not throttle synchronous reads. */
 609         if (op_is_sync(opf) && !op_is_write(opf))
 610                 return;
 611
 612         /*
 613          * Throttle asynchronous requests and writes such that these requests
 614          * do not block the allocation of synchronous requests.
 615          */
 616         data->shallow_depth = dd->async_depth;
 617 }
 618
 619 /* Called by blk_mq_update_nr_requests(). */
 620 static void dd_depth_updated(struct blk_mq_hw_ctx *hctx)
 621 {
 622         struct request_queue *q = hctx->queue;
 623         struct deadline_data *dd = q->elevator->elevator_data;
 624         struct blk_mq_tags *tags = hctx->sched_tags;
 625         unsigned int shift = tags->bitmap_tags.sb.shift;
 626
 627         dd->async_depth = max(1U, 3 * (1U << shift)  / 4);
 628
 629         sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, dd->async_depth);
 630 }
 631
 632 /* Called by blk_mq_init_hctx() and blk_mq_init_sched(). */
 633 static int dd_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 634 {
 635         dd_depth_updated(hctx);
 636         return 0;
 637 }
 638
 639 static void dd_exit_sched(struct elevator_queue *e)
 640 {
 641         struct deadline_data *dd = e->elevator_data;
 642         enum dd_prio prio;
 643
 644         for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 645                 struct dd_per_prio *per_prio = &dd->per_prio[prio];
 646                 const struct io_stats_per_prio *stats = &per_prio->stats;
 647                 uint32_t queued;
 648
 649                 WARN_ON_ONCE(!list_empty(&per_prio->fifo_list[DD_READ]));
 650                 WARN_ON_ONCE(!list_empty(&per_prio->fifo_list[DD_WRITE]));
 651
 652                 spin_lock(&dd->lock);
 653                 queued = dd_queued(dd, prio);
 654                 spin_unlock(&dd->lock);
 655
 656                 WARN_ONCE(queued != 0,
 657                           "statistics for priority %d: i %u m %u d %u c %u\n",
 658                           prio, stats->inserted, stats->merged,
 659                           stats->dispatched, atomic_read(&stats->completed));
 660         }
 661
 662         kfree(dd);
 663 }
 664
 665 /*
 666  * initialize elevator private data (deadline_data).
 667  */
 668 static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
 669 {
 670         struct deadline_data *dd;
 671         struct elevator_queue *eq;
 672         enum dd_prio prio;
 673         int ret = -ENOMEM;
 674
 675         eq = elevator_alloc(q, e);
 676         if (!eq)
 677                 return ret;
 678
 679         dd = kzalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
 680         if (!dd)
 681                 goto put_eq;
 682
 683         eq->elevator_data = dd;
 684
 685         for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 686                 struct dd_per_prio *per_prio = &dd->per_prio[prio];
 687
 688                 INIT_LIST_HEAD(&per_prio->dispatch);
 689                 INIT_LIST_HEAD(&per_prio->fifo_list[DD_READ]);
 690                 INIT_LIST_HEAD(&per_prio->fifo_list[DD_WRITE]);
 691                 per_prio->sort_list[DD_READ] = RB_ROOT;
 692                 per_prio->sort_list[DD_WRITE] = RB_ROOT;
 693         }
 694         dd->fifo_expire[DD_READ] = read_expire;
 695         dd->fifo_expire[DD_WRITE] = write_expire;
 696         dd->writes_starved = writes_starved;
 697         dd->front_merges = 1;
 698         dd->last_dir = DD_WRITE;
 699         dd->fifo_batch = fifo_batch;
 700         dd->prio_aging_expire = prio_aging_expire;
 701         spin_lock_init(&dd->lock);
 702         spin_lock_init(&dd->zone_lock);
 703
 704         /* We dispatch from request queue wide instead of hw queue */
 705         blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
 706
 707         q->elevator = eq;
 708         return 0;
 709
 710 put_eq:
 711         kobject_put(&eq->kobj);
 712         return ret;
 713 }
 714
 715 /*
 716  * Try to merge @bio into an existing request. If @bio has been merged into
 717  * an existing request, store the pointer to that request into *@rq.
 718  */
 719 static int dd_request_merge(struct request_queue *q, struct request **rq,
 720                             struct bio *bio)
 721 {
 722         struct deadline_data *dd = q->elevator->elevator_data;
 723         const u8 ioprio_class = IOPRIO_PRIO_CLASS(bio->bi_ioprio);
 724         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 725         struct dd_per_prio *per_prio = &dd->per_prio[prio];
 726         sector_t sector = bio_end_sector(bio);
 727         struct request *__rq;
 728
 729         if (!dd->front_merges)
 730                 return ELEVATOR_NO_MERGE;
 731
 732         __rq = elv_rb_find(&per_prio->sort_list[bio_data_dir(bio)], sector);
 733         if (__rq) {
 734                 BUG_ON(sector != blk_rq_pos(__rq));
 735
 736                 if (elv_bio_merge_ok(__rq, bio)) {
 737                         *rq = __rq;
 738                         if (blk_discard_mergable(__rq))
 739                                 return ELEVATOR_DISCARD_MERGE;
 740                         return ELEVATOR_FRONT_MERGE;
 741                 }
 742         }
 743
 744         return ELEVATOR_NO_MERGE;
 745 }
 746
 747 /*
 748  * Attempt to merge a bio into an existing request. This function is called
 749  * before @bio is associated with a request.
 750  */
 751 static bool dd_bio_merge(struct request_queue *q, struct bio *bio,
 752                 unsigned int nr_segs)
 753 {
 754         struct deadline_data *dd = q->elevator->elevator_data;
 755         struct request *free = NULL;
 756         bool ret;
 757
 758         spin_lock(&dd->lock);
 759         ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
 760         spin_unlock(&dd->lock);
 761
 762         if (free)
 763                 blk_mq_free_request(free);
 764
 765         return ret;
 766 }
 767
 768 /*
 769  * add rq to rbtree and fifo
 770  */
 771 static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 772                               bool at_head)
 773 {
 774         struct request_queue *q = hctx->queue;
 775         struct deadline_data *dd = q->elevator->elevator_data;
 776         const enum dd_data_dir data_dir = rq_data_dir(rq);
 777         u16 ioprio = req_get_ioprio(rq);
 778         u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
 779         struct dd_per_prio *per_prio;
 780         enum dd_prio prio;
 781         LIST_HEAD(free);
 782
 783         lockdep_assert_held(&dd->lock);
 784
 785         /*
 786          * This may be a requeue of a write request that has locked its
 787          * target zone. If it is the case, this releases the zone lock.
 788          */
 789         blk_req_zone_write_unlock(rq);
 790
 791         prio = ioprio_class_to_prio[ioprio_class];
 792         per_prio = &dd->per_prio[prio];
 793         if (!rq->elv.priv[0]) {
 794                 per_prio->stats.inserted++;
 795                 rq->elv.priv[0] = (void *)(uintptr_t)1;
 796         }
 797
 798         if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
 799                 blk_mq_free_requests(&free);
 800                 return;
 801         }
 802
 803         trace_block_rq_insert(rq);
 804
 805         if (at_head) {
 806                 list_add(&rq->queuelist, &per_prio->dispatch);
 807                 rq->fifo_time = jiffies;
 808         } else {
 809                 deadline_add_rq_rb(per_prio, rq);
 810
 811                 if (rq_mergeable(rq)) {
 812                         elv_rqhash_add(q, rq);
 813                         if (!q->last_merge)
 814                                 q->last_merge = rq;
 815                 }
 816
 817                 /*
 818                  * set expire time and add to fifo list
 819                  */
 820                 rq->fifo_time = jiffies + dd->fifo_expire[data_dir];
 821                 list_add_tail(&rq->queuelist, &per_prio->fifo_list[data_dir]);
 822         }
 823 }
 824
 825 /*
 826  * Called from blk_mq_sched_insert_request() or blk_mq_sched_insert_requests().
 827  */
 828 static void dd_insert_requests(struct blk_mq_hw_ctx *hctx,
 829                                struct list_head *list, bool at_head)
 830 {
 831         struct request_queue *q = hctx->queue;
 832         struct deadline_data *dd = q->elevator->elevator_data;
 833
 834         spin_lock(&dd->lock);
 835         while (!list_empty(list)) {
 836                 struct request *rq;
 837
 838                 rq = list_first_entry(list, struct request, queuelist);
 839                 list_del_init(&rq->queuelist);
 840                 dd_insert_request(hctx, rq, at_head);
 841         }
 842         spin_unlock(&dd->lock);
 843 }
 844
 845 /* Callback from inside blk_mq_rq_ctx_init(). */
 846 static void dd_prepare_request(struct request *rq)
 847 {
 848         rq->elv.priv[0] = NULL;
 849 }
 850
 851 static bool dd_has_write_work(struct blk_mq_hw_ctx *hctx)
 852 {
 853         struct deadline_data *dd = hctx->queue->elevator->elevator_data;
 854         enum dd_prio p;
 855
 856         for (p = 0; p <= DD_PRIO_MAX; p++)
 857                 if (!list_empty_careful(&dd->per_prio[p].fifo_list[DD_WRITE]))
 858                         return true;
 859
 860         return false;
 861 }
 862
 863 /*
 864  * Callback from inside blk_mq_free_request().
 865  *
 866  * For zoned block devices, write unlock the target zone of
 867  * completed write requests. Do this while holding the zone lock
 868  * spinlock so that the zone is never unlocked while deadline_fifo_request()
 869  * or deadline_next_request() are executing. This function is called for
 870  * all requests, whether or not these requests complete successfully.
 871  *
 872  * For a zoned block device, __dd_dispatch_request() may have stopped
 873  * dispatching requests if all the queued requests are write requests directed
 874  * at zones that are already locked due to on-going write requests. To ensure
 875  * write request dispatch progress in this case, mark the queue as needing a
 876  * restart to ensure that the queue is run again after completion of the
 877  * request and zones being unlocked.
 878  */
 879 static void dd_finish_request(struct request *rq)
 880 {
 881         struct request_queue *q = rq->q;
 882         struct deadline_data *dd = q->elevator->elevator_data;
 883         const u8 ioprio_class = dd_rq_ioclass(rq);
 884         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 885         struct dd_per_prio *per_prio = &dd->per_prio[prio];
 886
 887         /*
 888          * The block layer core may call dd_finish_request() without having
 889          * called dd_insert_requests(). Skip requests that bypassed I/O
 890          * scheduling. See also blk_mq_request_bypass_insert().
 891          */
 892         if (!rq->elv.priv[0])
 893                 return;
 894
 895         atomic_inc(&per_prio->stats.completed);
 896
 897         if (blk_queue_is_zoned(q)) {
 898                 unsigned long flags;
 899
 900                 spin_lock_irqsave(&dd->zone_lock, flags);
 901                 blk_req_zone_write_unlock(rq);
 902                 spin_unlock_irqrestore(&dd->zone_lock, flags);
 903
 904                 if (dd_has_write_work(rq->mq_hctx))
 905                         blk_mq_sched_mark_restart_hctx(rq->mq_hctx);
 906         }
 907 }
 908
 909 static bool dd_has_work_for_prio(struct dd_per_prio *per_prio)
 910 {
 911         return !list_empty_careful(&per_prio->dispatch) ||
 912                 !list_empty_careful(&per_prio->fifo_list[DD_READ]) ||
 913                 !list_empty_careful(&per_prio->fifo_list[DD_WRITE]);
 914 }
 915
 916 static bool dd_has_work(struct blk_mq_hw_ctx *hctx)
 917 {
 918         struct deadline_data *dd = hctx->queue->elevator->elevator_data;
 919         enum dd_prio prio;
 920
 921         for (prio = 0; prio <= DD_PRIO_MAX; prio++)
 922                 if (dd_has_work_for_prio(&dd->per_prio[prio]))
 923                         return true;
 924
 925         return false;
 926 }
 927
 928 /*
 929  * sysfs parts below
 930  */
 931 #define SHOW_INT(__FUNC, __VAR)                                         \
 932 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
 933 {                                                                       \
 934         struct deadline_data *dd = e->elevator_data;                    \
 935                                                                         \
 936         return sysfs_emit(page, "%d\n", __VAR);                         \
 937 }
 938 #define SHOW_JIFFIES(__FUNC, __VAR) SHOW_INT(__FUNC, jiffies_to_msecs(__VAR))
 939 SHOW_JIFFIES(deadline_read_expire_show, dd->fifo_expire[DD_READ]);
 940 SHOW_JIFFIES(deadline_write_expire_show, dd->fifo_expire[DD_WRITE]);
 941 SHOW_JIFFIES(deadline_prio_aging_expire_show, dd->prio_aging_expire);
 942 SHOW_INT(deadline_writes_starved_show, dd->writes_starved);
 943 SHOW_INT(deadline_front_merges_show, dd->front_merges);
 944 SHOW_INT(deadline_async_depth_show, dd->async_depth);
 945 SHOW_INT(deadline_fifo_batch_show, dd->fifo_batch);
 946 #undef SHOW_INT
 947 #undef SHOW_JIFFIES
 948
 949 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
 950 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
 951 {                                                                       \
 952         struct deadline_data *dd = e->elevator_data;                    \
 953         int __data, __ret;                                              \
 954                                                                         \
 955         __ret = kstrtoint(page, 0, &__data);                            \
 956         if (__ret < 0)                                                  \
 957                 return __ret;                                           \
 958         if (__data < (MIN))                                             \
 959                 __data = (MIN);                                         \
 960         else if (__data > (MAX))                                        \
 961                 __data = (MAX);                                         \
 962         *(__PTR) = __CONV(__data);                                      \
 963         return count;                                                   \
 964 }
 965 #define STORE_INT(__FUNC, __PTR, MIN, MAX)                              \
 966         STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, )
 967 #define STORE_JIFFIES(__FUNC, __PTR, MIN, MAX)                          \
 968         STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, msecs_to_jiffies)
 969 STORE_JIFFIES(deadline_read_expire_store, &dd->fifo_expire[DD_READ], 0, INT_MAX);
 970 STORE_JIFFIES(deadline_write_expire_store, &dd->fifo_expire[DD_WRITE], 0, INT_MAX);
 971 STORE_JIFFIES(deadline_prio_aging_expire_store, &dd->prio_aging_expire, 0, INT_MAX);
 972 STORE_INT(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX);
 973 STORE_INT(deadline_front_merges_store, &dd->front_merges, 0, 1);
 974 STORE_INT(deadline_async_depth_store, &dd->async_depth, 1, INT_MAX);
 975 STORE_INT(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX);
 976 #undef STORE_FUNCTION
 977 #undef STORE_INT
 978 #undef STORE_JIFFIES
 979
 980 #define DD_ATTR(name) \
 981         __ATTR(name, 0644, deadline_##name##_show, deadline_##name##_store)
 982
 983 static struct elv_fs_entry deadline_attrs[] = {
 984         DD_ATTR(read_expire),
 985         DD_ATTR(write_expire),
 986         DD_ATTR(writes_starved),
 987         DD_ATTR(front_merges),
 988         DD_ATTR(async_depth),
 989         DD_ATTR(fifo_batch),
 990         DD_ATTR(prio_aging_expire),
 991         __ATTR_NULL
 992 };
 993
 994 #ifdef CONFIG_BLK_DEBUG_FS
 995 #define DEADLINE_DEBUGFS_DDIR_ATTRS(prio, data_dir, name)               \
 996 static void *deadline_##name##_fifo_start(struct seq_file *m,           \
 997                                           loff_t *pos)                  \
 998         __acquires(&dd->lock)                                           \
 999 {                                                                       \
1000         struct request_queue *q = m->private;                           \
1001         struct deadline_data *dd = q->elevator->elevator_data;          \
1002         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
1003                                                                         \
1004         spin_lock(&dd->lock);                                           \
1005         return seq_list_start(&per_prio->fifo_list[data_dir], *pos);    \
1006 }                                                                       \
1007                                                                         \
1008 static void *deadline_##name##_fifo_next(struct seq_file *m, void *v,   \
1009                                          loff_t *pos)                   \
1010 {                                                                       \
1011         struct request_queue *q = m->private;                           \
1012         struct deadline_data *dd = q->elevator->elevator_data;          \
1013         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
1014                                                                         \
1015         return seq_list_next(v, &per_prio->fifo_list[data_dir], pos);   \
1016 }                                                                       \
1017                                                                         \
1018 static void deadline_##name##_fifo_stop(struct seq_file *m, void *v)    \
1019         __releases(&dd->lock)                                           \
1020 {                                                                       \
1021         struct request_queue *q = m->private;                           \
1022         struct deadline_data *dd = q->elevator->elevator_data;          \
1023                                                                         \
1024         spin_unlock(&dd->lock);                                         \
1025 }                                                                       \
1026                                                                         \
1027 static const struct seq_operations deadline_##name##_fifo_seq_ops = {   \
1028         .start  = deadline_##name##_fifo_start,                         \
1029         .next   = deadline_##name##_fifo_next,                          \
1030         .stop   = deadline_##name##_fifo_stop,                          \
1031         .show   = blk_mq_debugfs_rq_show,                               \
1032 };                                                                      \
1033                                                                         \
1034 static int deadline_##name##_next_rq_show(void *data,                   \
1035                                           struct seq_file *m)           \
1036 {                                                                       \
1037         struct request_queue *q = data;                                 \
1038         struct deadline_data *dd = q->elevator->elevator_data;          \
1039         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
1040         struct request *rq = per_prio->next_rq[data_dir];               \
1041                                                                         \
1042         if (rq)                                                         \
1043                 __blk_mq_debugfs_rq_show(m, rq);                        \
1044         return 0;                                                       \
1045 }
1046
1047 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_RT_PRIO, DD_READ, read0);
1048 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_RT_PRIO, DD_WRITE, write0);
1049 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_BE_PRIO, DD_READ, read1);
1050 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_BE_PRIO, DD_WRITE, write1);
1051 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_IDLE_PRIO, DD_READ, read2);
1052 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_IDLE_PRIO, DD_WRITE, write2);
1053 #undef DEADLINE_DEBUGFS_DDIR_ATTRS
1054
1055 static int deadline_batching_show(void *data, struct seq_file *m)
1056 {
1057         struct request_queue *q = data;
1058         struct deadline_data *dd = q->elevator->elevator_data;
1059
1060         seq_printf(m, "%u\n", dd->batching);
1061         return 0;
1062 }
1063
1064 static int deadline_starved_show(void *data, struct seq_file *m)
1065 {
1066         struct request_queue *q = data;
1067         struct deadline_data *dd = q->elevator->elevator_data;
1068
1069         seq_printf(m, "%u\n", dd->starved);
1070         return 0;
1071 }
1072
1073 static int dd_async_depth_show(void *data, struct seq_file *m)
1074 {
1075         struct request_queue *q = data;
1076         struct deadline_data *dd = q->elevator->elevator_data;
1077
1078         seq_printf(m, "%u\n", dd->async_depth);
1079         return 0;
1080 }
1081
1082 static int dd_queued_show(void *data, struct seq_file *m)
1083 {
1084         struct request_queue *q = data;
1085         struct deadline_data *dd = q->elevator->elevator_data;
1086         u32 rt, be, idle;
1087
1088         spin_lock(&dd->lock);
1089         rt = dd_queued(dd, DD_RT_PRIO);
1090         be = dd_queued(dd, DD_BE_PRIO);
1091         idle = dd_queued(dd, DD_IDLE_PRIO);
1092         spin_unlock(&dd->lock);
1093
1094         seq_printf(m, "%u %u %u\n", rt, be, idle);
1095
1096         return 0;
1097 }
1098
1099 /* Number of requests owned by the block driver for a given priority. */
1100 static u32 dd_owned_by_driver(struct deadline_data *dd, enum dd_prio prio)
1101 {
1102         const struct io_stats_per_prio *stats = &dd->per_prio[prio].stats;
1103
1104         lockdep_assert_held(&dd->lock);
1105
1106         return stats->dispatched + stats->merged -
1107                 atomic_read(&stats->completed);
1108 }
1109
1110 static int dd_owned_by_driver_show(void *data, struct seq_file *m)
1111 {
1112         struct request_queue *q = data;
1113         struct deadline_data *dd = q->elevator->elevator_data;
1114         u32 rt, be, idle;
1115
1116         spin_lock(&dd->lock);
1117         rt = dd_owned_by_driver(dd, DD_RT_PRIO);
1118         be = dd_owned_by_driver(dd, DD_BE_PRIO);
1119         idle = dd_owned_by_driver(dd, DD_IDLE_PRIO);
1120         spin_unlock(&dd->lock);
1121
1122         seq_printf(m, "%u %u %u\n", rt, be, idle);
1123
1124         return 0;
1125 }
1126
1127 #define DEADLINE_DISPATCH_ATTR(prio)                                    \
1128 static void *deadline_dispatch##prio##_start(struct seq_file *m,        \
1129                                              loff_t *pos)               \
1130         __acquires(&dd->lock)                                           \
1131 {                                                                       \
1132         struct request_queue *q = m->private;                           \
1133         struct deadline_data *dd = q->elevator->elevator_data;          \
1134         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
1135                                                                         \
1136         spin_lock(&dd->lock);                                           \
1137         return seq_list_start(&per_prio->dispatch, *pos);               \
1138 }                                                                       \
1139                                                                         \
1140 static void *deadline_dispatch##prio##_next(struct seq_file *m,         \
1141                                             void *v, loff_t *pos)       \
1142 {                                                                       \
1143         struct request_queue *q = m->private;                           \
1144         struct deadline_data *dd = q->elevator->elevator_data;          \
1145         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
1146                                                                         \
1147         return seq_list_next(v, &per_prio->dispatch, pos);              \
1148 }                                                                       \
1149                                                                         \
1150 static void deadline_dispatch##prio##_stop(struct seq_file *m, void *v) \
1151         __releases(&dd->lock)                                           \
1152 {                                                                       \
1153         struct request_queue *q = m->private;                           \
1154         struct deadline_data *dd = q->elevator->elevator_data;          \
1155                                                                         \
1156         spin_unlock(&dd->lock);                                         \
1157 }                                                                       \
1158                                                                         \
1159 static const struct seq_operations deadline_dispatch##prio##_seq_ops = { \
1160         .start  = deadline_dispatch##prio##_start,                      \
1161         .next   = deadline_dispatch##prio##_next,                       \
1162         .stop   = deadline_dispatch##prio##_stop,                       \
1163         .show   = blk_mq_debugfs_rq_show,                               \
1164 }
1165
1166 DEADLINE_DISPATCH_ATTR(0);
1167 DEADLINE_DISPATCH_ATTR(1);
1168 DEADLINE_DISPATCH_ATTR(2);
1169 #undef DEADLINE_DISPATCH_ATTR
1170
1171 #define DEADLINE_QUEUE_DDIR_ATTRS(name)                                 \
1172         {#name "_fifo_list", 0400,                                      \
1173                         .seq_ops = &deadline_##name##_fifo_seq_ops}
1174 #define DEADLINE_NEXT_RQ_ATTR(name)                                     \
1175         {#name "_next_rq", 0400, deadline_##name##_next_rq_show}
1176 static const struct blk_mq_debugfs_attr deadline_queue_debugfs_attrs[] = {
1177         DEADLINE_QUEUE_DDIR_ATTRS(read0),
1178         DEADLINE_QUEUE_DDIR_ATTRS(write0),
1179         DEADLINE_QUEUE_DDIR_ATTRS(read1),
1180         DEADLINE_QUEUE_DDIR_ATTRS(write1),
1181         DEADLINE_QUEUE_DDIR_ATTRS(read2),
1182         DEADLINE_QUEUE_DDIR_ATTRS(write2),
1183         DEADLINE_NEXT_RQ_ATTR(read0),
1184         DEADLINE_NEXT_RQ_ATTR(write0),
1185         DEADLINE_NEXT_RQ_ATTR(read1),
1186         DEADLINE_NEXT_RQ_ATTR(write1),
1187         DEADLINE_NEXT_RQ_ATTR(read2),
1188         DEADLINE_NEXT_RQ_ATTR(write2),
1189         {"batching", 0400, deadline_batching_show},
1190         {"starved", 0400, deadline_starved_show},
1191         {"async_depth", 0400, dd_async_depth_show},
1192         {"dispatch0", 0400, .seq_ops = &deadline_dispatch0_seq_ops},
1193         {"dispatch1", 0400, .seq_ops = &deadline_dispatch1_seq_ops},
1194         {"dispatch2", 0400, .seq_ops = &deadline_dispatch2_seq_ops},
1195         {"owned_by_driver", 0400, dd_owned_by_driver_show},
1196         {"queued", 0400, dd_queued_show},
1197         {},
1198 };
1199 #undef DEADLINE_QUEUE_DDIR_ATTRS
1200 #endif
1201
1202 static struct elevator_type mq_deadline = {
1203         .ops = {
1204                 .depth_updated          = dd_depth_updated,
1205                 .limit_depth            = dd_limit_depth,
1206                 .insert_requests        = dd_insert_requests,
1207                 .dispatch_request       = dd_dispatch_request,
1208                 .prepare_request        = dd_prepare_request,
1209                 .finish_request         = dd_finish_request,
1210                 .next_request           = elv_rb_latter_request,
1211                 .former_request         = elv_rb_former_request,
1212                 .bio_merge              = dd_bio_merge,
1213                 .request_merge          = dd_request_merge,
1214                 .requests_merged        = dd_merged_requests,
1215                 .request_merged         = dd_request_merged,
1216                 .has_work               = dd_has_work,
1217                 .init_sched             = dd_init_sched,
1218                 .exit_sched             = dd_exit_sched,
1219                 .init_hctx              = dd_init_hctx,
1220         },
1221
1222 #ifdef CONFIG_BLK_DEBUG_FS
1223         .queue_debugfs_attrs = deadline_queue_debugfs_attrs,
1224 #endif
1225         .elevator_attrs = deadline_attrs,
1226         .elevator_name = "mq-deadline",
1227         .elevator_alias = "deadline",
1228         .elevator_features = ELEVATOR_F_ZBD_SEQ_WRITE,
1229         .elevator_owner = THIS_MODULE,
1230 };
1231 MODULE_ALIAS("mq-deadline-iosched");
1232
1233 static int __init deadline_init(void)
1234 {
1235         return elv_register(&mq_deadline);
1236 }
1237
1238 static void __exit deadline_exit(void)
1239 {
1240         elv_unregister(&mq_deadline);
1241 }
1242
1243 module_init(deadline_init);
1244 module_exit(deadline_exit);
1245
1246 MODULE_AUTHOR("Jens Axboe, Damien Le Moal and Bart Van Assche");
1247 MODULE_LICENSE("GPL");
1248 MODULE_DESCRIPTION("MQ deadline IO scheduler");