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