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