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 "blk.h"
  22 #include "blk-mq.h"
  23 #include "blk-mq-debugfs.h"
  24 #include "blk-mq-tag.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 static const int writes_starved = 2;    /* max times reads can starve a write */
  33 static const int fifo_batch = 16;       /* # of sequential requests treated as one
  34                                      by the above parameters. For throughput. */
  35
  36 struct deadline_data {
  37         /*
  38          * run time data
  39          */
  40
  41         /*
  42          * requests (deadline_rq s) are present on both sort_list and fifo_list
  43          */
  44         struct rb_root sort_list[2];
  45         struct list_head fifo_list[2];
  46
  47         /*
  48          * next in sort order. read, write or both are NULL
  49          */
  50         struct request *next_rq[2];
  51         unsigned int batching;          /* number of sequential requests made */
  52         unsigned int starved;           /* times reads have starved writes */
  53
  54         /*
  55          * settings that change how the i/o scheduler behaves
  56          */
  57         int fifo_expire[2];
  58         int fifo_batch;
  59         int writes_starved;
  60         int front_merges;
  61
  62         spinlock_t lock;
  63         spinlock_t zone_lock;
  64         struct list_head dispatch;
  65 };
  66
  67 static inline struct rb_root *
  68 deadline_rb_root(struct deadline_data *dd, struct request *rq)
  69 {
  70         return &dd->sort_list[rq_data_dir(rq)];
  71 }
  72
  73 /*
  74  * get the request after `rq' in sector-sorted order
  75  */
  76 static inline struct request *
  77 deadline_latter_request(struct request *rq)
  78 {
  79         struct rb_node *node = rb_next(&rq->rb_node);
  80
  81         if (node)
  82                 return rb_entry_rq(node);
  83
  84         return NULL;
  85 }
  86
  87 static void
  88 deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
  89 {
  90         struct rb_root *root = deadline_rb_root(dd, rq);
  91
  92         elv_rb_add(root, rq);
  93 }
  94
  95 static inline void
  96 deadline_del_rq_rb(struct deadline_data *dd, struct request *rq)
  97 {
  98         const int data_dir = rq_data_dir(rq);
  99
 100         if (dd->next_rq[data_dir] == rq)
 101                 dd->next_rq[data_dir] = deadline_latter_request(rq);
 102
 103         elv_rb_del(deadline_rb_root(dd, rq), rq);
 104 }
 105
 106 /*
 107  * remove rq from rbtree and fifo.
 108  */
 109 static void deadline_remove_request(struct request_queue *q, struct request *rq)
 110 {
 111         struct deadline_data *dd = q->elevator->elevator_data;
 112
 113         list_del_init(&rq->queuelist);
 114
 115         /*
 116          * We might not be on the rbtree, if we are doing an insert merge
 117          */
 118         if (!RB_EMPTY_NODE(&rq->rb_node))
 119                 deadline_del_rq_rb(dd, rq);
 120
 121         elv_rqhash_del(q, rq);
 122         if (q->last_merge == rq)
 123                 q->last_merge = NULL;
 124 }
 125
 126 static void dd_request_merged(struct request_queue *q, struct request *req,
 127                               enum elv_merge type)
 128 {
 129         struct deadline_data *dd = q->elevator->elevator_data;
 130
 131         /*
 132          * if the merge was a front merge, we need to reposition request
 133          */
 134         if (type == ELEVATOR_FRONT_MERGE) {
 135                 elv_rb_del(deadline_rb_root(dd, req), req);
 136                 deadline_add_rq_rb(dd, req);
 137         }
 138 }
 139
 140 static void dd_merged_requests(struct request_queue *q, struct request *req,
 141                                struct request *next)
 142 {
 143         /*
 144          * if next expires before rq, assign its expire time to rq
 145          * and move into next position (next will be deleted) in fifo
 146          */
 147         if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
 148                 if (time_before((unsigned long)next->fifo_time,
 149                                 (unsigned long)req->fifo_time)) {
 150                         list_move(&req->queuelist, &next->queuelist);
 151                         req->fifo_time = next->fifo_time;
 152                 }
 153         }
 154
 155         /*
 156          * kill knowledge of next, this one is a goner
 157          */
 158         deadline_remove_request(q, next);
 159 }
 160
 161 /*
 162  * move an entry to dispatch queue
 163  */
 164 static void
 165 deadline_move_request(struct deadline_data *dd, struct request *rq)
 166 {
 167         const int data_dir = rq_data_dir(rq);
 168
 169         dd->next_rq[READ] = NULL;
 170         dd->next_rq[WRITE] = NULL;
 171         dd->next_rq[data_dir] = deadline_latter_request(rq);
 172
 173         /*
 174          * take it off the sort and fifo list
 175          */
 176         deadline_remove_request(rq->q, rq);
 177 }
 178
 179 /*
 180  * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
 181  * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
 182  */
 183 static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
 184 {
 185         struct request *rq = rq_entry_fifo(dd->fifo_list[ddir].next);
 186
 187         /*
 188          * rq is expired!
 189          */
 190         if (time_after_eq(jiffies, (unsigned long)rq->fifo_time))
 191                 return 1;
 192
 193         return 0;
 194 }
 195
 196 /*
 197  * For the specified data direction, return the next request to
 198  * dispatch using arrival ordered lists.
 199  */
 200 static struct request *
 201 deadline_fifo_request(struct deadline_data *dd, int data_dir)
 202 {
 203         struct request *rq;
 204         unsigned long flags;
 205
 206         if (WARN_ON_ONCE(data_dir != READ && data_dir != WRITE))
 207                 return NULL;
 208
 209         if (list_empty(&dd->fifo_list[data_dir]))
 210                 return NULL;
 211
 212         rq = rq_entry_fifo(dd->fifo_list[data_dir].next);
 213         if (data_dir == READ || !blk_queue_is_zoned(rq->q))
 214                 return rq;
 215
 216         /*
 217          * Look for a write request that can be dispatched, that is one with
 218          * an unlocked target zone.
 219          */
 220         spin_lock_irqsave(&dd->zone_lock, flags);
 221         list_for_each_entry(rq, &dd->fifo_list[WRITE], queuelist) {
 222                 if (blk_req_can_dispatch_to_zone(rq))
 223                         goto out;
 224         }
 225         rq = NULL;
 226 out:
 227         spin_unlock_irqrestore(&dd->zone_lock, flags);
 228
 229         return rq;
 230 }
 231
 232 /*
 233  * For the specified data direction, return the next request to
 234  * dispatch using sector position sorted lists.
 235  */
 236 static struct request *
 237 deadline_next_request(struct deadline_data *dd, int data_dir)
 238 {
 239         struct request *rq;
 240         unsigned long flags;
 241
 242         if (WARN_ON_ONCE(data_dir != READ && data_dir != WRITE))
 243                 return NULL;
 244
 245         rq = dd->next_rq[data_dir];
 246         if (!rq)
 247                 return NULL;
 248
 249         if (data_dir == READ || !blk_queue_is_zoned(rq->q))
 250                 return rq;
 251
 252         /*
 253          * Look for a write request that can be dispatched, that is one with
 254          * an unlocked target zone.
 255          */
 256         spin_lock_irqsave(&dd->zone_lock, flags);
 257         while (rq) {
 258                 if (blk_req_can_dispatch_to_zone(rq))
 259                         break;
 260                 rq = deadline_latter_request(rq);
 261         }
 262         spin_unlock_irqrestore(&dd->zone_lock, flags);
 263
 264         return rq;
 265 }
 266
 267 /*
 268  * deadline_dispatch_requests selects the best request according to
 269  * read/write expire, fifo_batch, etc
 270  */
 271 static struct request *__dd_dispatch_request(struct deadline_data *dd)
 272 {
 273         struct request *rq, *next_rq;
 274         bool reads, writes;
 275         int data_dir;
 276
 277         if (!list_empty(&dd->dispatch)) {
 278                 rq = list_first_entry(&dd->dispatch, struct request, queuelist);
 279                 list_del_init(&rq->queuelist);
 280                 goto done;
 281         }
 282
 283         reads = !list_empty(&dd->fifo_list[READ]);
 284         writes = !list_empty(&dd->fifo_list[WRITE]);
 285
 286         /*
 287          * batches are currently reads XOR writes
 288          */
 289         rq = deadline_next_request(dd, WRITE);
 290         if (!rq)
 291                 rq = deadline_next_request(dd, READ);
 292
 293         if (rq && dd->batching < dd->fifo_batch)
 294                 /* we have a next request are still entitled to batch */
 295                 goto dispatch_request;
 296
 297         /*
 298          * at this point we are not running a batch. select the appropriate
 299          * data direction (read / write)
 300          */
 301
 302         if (reads) {
 303                 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[READ]));
 304
 305                 if (deadline_fifo_request(dd, WRITE) &&
 306                     (dd->starved++ >= dd->writes_starved))
 307                         goto dispatch_writes;
 308
 309                 data_dir = READ;
 310
 311                 goto dispatch_find_request;
 312         }
 313
 314         /*
 315          * there are either no reads or writes have been starved
 316          */
 317
 318         if (writes) {
 319 dispatch_writes:
 320                 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[WRITE]));
 321
 322                 dd->starved = 0;
 323
 324                 data_dir = WRITE;
 325
 326                 goto dispatch_find_request;
 327         }
 328
 329         return NULL;
 330
 331 dispatch_find_request:
 332         /*
 333          * we are not running a batch, find best request for selected data_dir
 334          */
 335         next_rq = deadline_next_request(dd, data_dir);
 336         if (deadline_check_fifo(dd, data_dir) || !next_rq) {
 337                 /*
 338                  * A deadline has expired, the last request was in the other
 339                  * direction, or we have run out of higher-sectored requests.
 340                  * Start again from the request with the earliest expiry time.
 341                  */
 342                 rq = deadline_fifo_request(dd, data_dir);
 343         } else {
 344                 /*
 345                  * The last req was the same dir and we have a next request in
 346                  * sort order. No expired requests so continue on from here.
 347                  */
 348                 rq = next_rq;
 349         }
 350
 351         /*
 352          * For a zoned block device, if we only have writes queued and none of
 353          * them can be dispatched, rq will be NULL.
 354          */
 355         if (!rq)
 356                 return NULL;
 357
 358         dd->batching = 0;
 359
 360 dispatch_request:
 361         /*
 362          * rq is the selected appropriate request.
 363          */
 364         dd->batching++;
 365         deadline_move_request(dd, rq);
 366 done:
 367         /*
 368          * If the request needs its target zone locked, do it.
 369          */
 370         blk_req_zone_write_lock(rq);
 371         rq->rq_flags |= RQF_STARTED;
 372         return rq;
 373 }
 374
 375 /*
 376  * One confusing aspect here is that we get called for a specific
 377  * hardware queue, but we may return a request that is for a
 378  * different hardware queue. This is because mq-deadline has shared
 379  * state for all hardware queues, in terms of sorting, FIFOs, etc.
 380  */
 381 static struct request *dd_dispatch_request(struct blk_mq_hw_ctx *hctx)
 382 {
 383         struct deadline_data *dd = hctx->queue->elevator->elevator_data;
 384         struct request *rq;
 385
 386         spin_lock(&dd->lock);
 387         rq = __dd_dispatch_request(dd);
 388         spin_unlock(&dd->lock);
 389         if (rq)
 390                 atomic_dec(&rq->mq_hctx->elevator_queued);
 391
 392         return rq;
 393 }
 394
 395 static void dd_exit_queue(struct elevator_queue *e)
 396 {
 397         struct deadline_data *dd = e->elevator_data;
 398
 399         BUG_ON(!list_empty(&dd->fifo_list[READ]));
 400         BUG_ON(!list_empty(&dd->fifo_list[WRITE]));
 401
 402         kfree(dd);
 403 }
 404
 405 /*
 406  * initialize elevator private data (deadline_data).
 407  */
 408 static int dd_init_queue(struct request_queue *q, struct elevator_type *e)
 409 {
 410         struct deadline_data *dd;
 411         struct elevator_queue *eq;
 412
 413         eq = elevator_alloc(q, e);
 414         if (!eq)
 415                 return -ENOMEM;
 416
 417         dd = kzalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
 418         if (!dd) {
 419                 kobject_put(&eq->kobj);
 420                 return -ENOMEM;
 421         }
 422         eq->elevator_data = dd;
 423
 424         INIT_LIST_HEAD(&dd->fifo_list[READ]);
 425         INIT_LIST_HEAD(&dd->fifo_list[WRITE]);
 426         dd->sort_list[READ] = RB_ROOT;
 427         dd->sort_list[WRITE] = RB_ROOT;
 428         dd->fifo_expire[READ] = read_expire;
 429         dd->fifo_expire[WRITE] = write_expire;
 430         dd->writes_starved = writes_starved;
 431         dd->front_merges = 1;
 432         dd->fifo_batch = fifo_batch;
 433         spin_lock_init(&dd->lock);
 434         spin_lock_init(&dd->zone_lock);
 435         INIT_LIST_HEAD(&dd->dispatch);
 436
 437         q->elevator = eq;
 438         return 0;
 439 }
 440
 441 static int dd_request_merge(struct request_queue *q, struct request **rq,
 442                             struct bio *bio)
 443 {
 444         struct deadline_data *dd = q->elevator->elevator_data;
 445         sector_t sector = bio_end_sector(bio);
 446         struct request *__rq;
 447
 448         if (!dd->front_merges)
 449                 return ELEVATOR_NO_MERGE;
 450
 451         __rq = elv_rb_find(&dd->sort_list[bio_data_dir(bio)], sector);
 452         if (__rq) {
 453                 BUG_ON(sector != blk_rq_pos(__rq));
 454
 455                 if (elv_bio_merge_ok(__rq, bio)) {
 456                         *rq = __rq;
 457                         if (blk_discard_mergable(__rq))
 458                                 return ELEVATOR_DISCARD_MERGE;
 459                         return ELEVATOR_FRONT_MERGE;
 460                 }
 461         }
 462
 463         return ELEVATOR_NO_MERGE;
 464 }
 465
 466 static bool dd_bio_merge(struct request_queue *q, struct bio *bio,
 467                 unsigned int nr_segs)
 468 {
 469         struct deadline_data *dd = q->elevator->elevator_data;
 470         struct request *free = NULL;
 471         bool ret;
 472
 473         spin_lock(&dd->lock);
 474         ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
 475         spin_unlock(&dd->lock);
 476
 477         if (free)
 478                 blk_mq_free_request(free);
 479
 480         return ret;
 481 }
 482
 483 /*
 484  * add rq to rbtree and fifo
 485  */
 486 static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 487                               bool at_head)
 488 {
 489         struct request_queue *q = hctx->queue;
 490         struct deadline_data *dd = q->elevator->elevator_data;
 491         const int data_dir = rq_data_dir(rq);
 492
 493         /*
 494          * This may be a requeue of a write request that has locked its
 495          * target zone. If it is the case, this releases the zone lock.
 496          */
 497         blk_req_zone_write_unlock(rq);
 498
 499         if (blk_mq_sched_try_insert_merge(q, rq))
 500                 return;
 501
 502         blk_mq_sched_request_inserted(rq);
 503
 504         if (at_head || blk_rq_is_passthrough(rq)) {
 505                 if (at_head)
 506                         list_add(&rq->queuelist, &dd->dispatch);
 507                 else
 508                         list_add_tail(&rq->queuelist, &dd->dispatch);
 509         } else {
 510                 deadline_add_rq_rb(dd, rq);
 511
 512                 if (rq_mergeable(rq)) {
 513                         elv_rqhash_add(q, rq);
 514                         if (!q->last_merge)
 515                                 q->last_merge = rq;
 516                 }
 517
 518                 /*
 519                  * set expire time and add to fifo list
 520                  */
 521                 rq->fifo_time = jiffies + dd->fifo_expire[data_dir];
 522                 list_add_tail(&rq->queuelist, &dd->fifo_list[data_dir]);
 523         }
 524 }
 525
 526 static void dd_insert_requests(struct blk_mq_hw_ctx *hctx,
 527                                struct list_head *list, bool at_head)
 528 {
 529         struct request_queue *q = hctx->queue;
 530         struct deadline_data *dd = q->elevator->elevator_data;
 531
 532         spin_lock(&dd->lock);
 533         while (!list_empty(list)) {
 534                 struct request *rq;
 535
 536                 rq = list_first_entry(list, struct request, queuelist);
 537                 list_del_init(&rq->queuelist);
 538                 dd_insert_request(hctx, rq, at_head);
 539                 atomic_inc(&hctx->elevator_queued);
 540         }
 541         spin_unlock(&dd->lock);
 542 }
 543
 544 /*
 545  * Nothing to do here. This is defined only to ensure that .finish_request
 546  * method is called upon request completion.
 547  */
 548 static void dd_prepare_request(struct request *rq)
 549 {
 550 }
 551
 552 /*
 553  * For zoned block devices, write unlock the target zone of
 554  * completed write requests. Do this while holding the zone lock
 555  * spinlock so that the zone is never unlocked while deadline_fifo_request()
 556  * or deadline_next_request() are executing. This function is called for
 557  * all requests, whether or not these requests complete successfully.
 558  *
 559  * For a zoned block device, __dd_dispatch_request() may have stopped
 560  * dispatching requests if all the queued requests are write requests directed
 561  * at zones that are already locked due to on-going write requests. To ensure
 562  * write request dispatch progress in this case, mark the queue as needing a
 563  * restart to ensure that the queue is run again after completion of the
 564  * request and zones being unlocked.
 565  */
 566 static void dd_finish_request(struct request *rq)
 567 {
 568         struct request_queue *q = rq->q;
 569
 570         if (blk_queue_is_zoned(q)) {
 571                 struct deadline_data *dd = q->elevator->elevator_data;
 572                 unsigned long flags;
 573
 574                 spin_lock_irqsave(&dd->zone_lock, flags);
 575                 blk_req_zone_write_unlock(rq);
 576                 if (!list_empty(&dd->fifo_list[WRITE]))
 577                         blk_mq_sched_mark_restart_hctx(rq->mq_hctx);
 578                 spin_unlock_irqrestore(&dd->zone_lock, flags);
 579         }
 580 }
 581
 582 static bool dd_has_work(struct blk_mq_hw_ctx *hctx)
 583 {
 584         struct deadline_data *dd = hctx->queue->elevator->elevator_data;
 585
 586         if (!atomic_read(&hctx->elevator_queued))
 587                 return false;
 588
 589         return !list_empty_careful(&dd->dispatch) ||
 590                 !list_empty_careful(&dd->fifo_list[0]) ||
 591                 !list_empty_careful(&dd->fifo_list[1]);
 592 }
 593
 594 /*
 595  * sysfs parts below
 596  */
 597 static ssize_t
 598 deadline_var_show(int var, char *page)
 599 {
 600         return sprintf(page, "%d\n", var);
 601 }
 602
 603 static void
 604 deadline_var_store(int *var, const char *page)
 605 {
 606         char *p = (char *) page;
 607
 608         *var = simple_strtol(p, &p, 10);
 609 }
 610
 611 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
 612 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
 613 {                                                                       \
 614         struct deadline_data *dd = e->elevator_data;                    \
 615         int __data = __VAR;                                             \
 616         if (__CONV)                                                     \
 617                 __data = jiffies_to_msecs(__data);                      \
 618         return deadline_var_show(__data, (page));                       \
 619 }
 620 SHOW_FUNCTION(deadline_read_expire_show, dd->fifo_expire[READ], 1);
 621 SHOW_FUNCTION(deadline_write_expire_show, dd->fifo_expire[WRITE], 1);
 622 SHOW_FUNCTION(deadline_writes_starved_show, dd->writes_starved, 0);
 623 SHOW_FUNCTION(deadline_front_merges_show, dd->front_merges, 0);
 624 SHOW_FUNCTION(deadline_fifo_batch_show, dd->fifo_batch, 0);
 625 #undef SHOW_FUNCTION
 626
 627 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
 628 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
 629 {                                                                       \
 630         struct deadline_data *dd = e->elevator_data;                    \
 631         int __data;                                                     \
 632         deadline_var_store(&__data, (page));                            \
 633         if (__data < (MIN))                                             \
 634                 __data = (MIN);                                         \
 635         else if (__data > (MAX))                                        \
 636                 __data = (MAX);                                         \
 637         if (__CONV)                                                     \
 638                 *(__PTR) = msecs_to_jiffies(__data);                    \
 639         else                                                            \
 640                 *(__PTR) = __data;                                      \
 641         return count;                                                   \
 642 }
 643 STORE_FUNCTION(deadline_read_expire_store, &dd->fifo_expire[READ], 0, INT_MAX, 1);
 644 STORE_FUNCTION(deadline_write_expire_store, &dd->fifo_expire[WRITE], 0, INT_MAX, 1);
 645 STORE_FUNCTION(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX, 0);
 646 STORE_FUNCTION(deadline_front_merges_store, &dd->front_merges, 0, 1, 0);
 647 STORE_FUNCTION(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX, 0);
 648 #undef STORE_FUNCTION
 649
 650 #define DD_ATTR(name) \
 651         __ATTR(name, 0644, deadline_##name##_show, deadline_##name##_store)
 652
 653 static struct elv_fs_entry deadline_attrs[] = {
 654         DD_ATTR(read_expire),
 655         DD_ATTR(write_expire),
 656         DD_ATTR(writes_starved),
 657         DD_ATTR(front_merges),
 658         DD_ATTR(fifo_batch),
 659         __ATTR_NULL
 660 };
 661
 662 #ifdef CONFIG_BLK_DEBUG_FS
 663 #define DEADLINE_DEBUGFS_DDIR_ATTRS(ddir, name)                         \
 664 static void *deadline_##name##_fifo_start(struct seq_file *m,           \
 665                                           loff_t *pos)                  \
 666         __acquires(&dd->lock)                                           \
 667 {                                                                       \
 668         struct request_queue *q = m->private;                           \
 669         struct deadline_data *dd = q->elevator->elevator_data;          \
 670                                                                         \
 671         spin_lock(&dd->lock);                                           \
 672         return seq_list_start(&dd->fifo_list[ddir], *pos);              \
 673 }                                                                       \
 674                                                                         \
 675 static void *deadline_##name##_fifo_next(struct seq_file *m, void *v,   \
 676                                          loff_t *pos)                   \
 677 {                                                                       \
 678         struct request_queue *q = m->private;                           \
 679         struct deadline_data *dd = q->elevator->elevator_data;          \
 680                                                                         \
 681         return seq_list_next(v, &dd->fifo_list[ddir], pos);             \
 682 }                                                                       \
 683                                                                         \
 684 static void deadline_##name##_fifo_stop(struct seq_file *m, void *v)    \
 685         __releases(&dd->lock)                                           \
 686 {                                                                       \
 687         struct request_queue *q = m->private;                           \
 688         struct deadline_data *dd = q->elevator->elevator_data;          \
 689                                                                         \
 690         spin_unlock(&dd->lock);                                         \
 691 }                                                                       \
 692                                                                         \
 693 static const struct seq_operations deadline_##name##_fifo_seq_ops = {   \
 694         .start  = deadline_##name##_fifo_start,                         \
 695         .next   = deadline_##name##_fifo_next,                          \
 696         .stop   = deadline_##name##_fifo_stop,                          \
 697         .show   = blk_mq_debugfs_rq_show,                               \
 698 };                                                                      \
 699                                                                         \
 700 static int deadline_##name##_next_rq_show(void *data,                   \
 701                                           struct seq_file *m)           \
 702 {                                                                       \
 703         struct request_queue *q = data;                                 \
 704         struct deadline_data *dd = q->elevator->elevator_data;          \
 705         struct request *rq = dd->next_rq[ddir];                         \
 706                                                                         \
 707         if (rq)                                                         \
 708                 __blk_mq_debugfs_rq_show(m, rq);                        \
 709         return 0;                                                       \
 710 }
 711 DEADLINE_DEBUGFS_DDIR_ATTRS(READ, read)
 712 DEADLINE_DEBUGFS_DDIR_ATTRS(WRITE, write)
 713 #undef DEADLINE_DEBUGFS_DDIR_ATTRS
 714
 715 static int deadline_batching_show(void *data, struct seq_file *m)
 716 {
 717         struct request_queue *q = data;
 718         struct deadline_data *dd = q->elevator->elevator_data;
 719
 720         seq_printf(m, "%u\n", dd->batching);
 721         return 0;
 722 }
 723
 724 static int deadline_starved_show(void *data, struct seq_file *m)
 725 {
 726         struct request_queue *q = data;
 727         struct deadline_data *dd = q->elevator->elevator_data;
 728
 729         seq_printf(m, "%u\n", dd->starved);
 730         return 0;
 731 }
 732
 733 static void *deadline_dispatch_start(struct seq_file *m, loff_t *pos)
 734         __acquires(&dd->lock)
 735 {
 736         struct request_queue *q = m->private;
 737         struct deadline_data *dd = q->elevator->elevator_data;
 738
 739         spin_lock(&dd->lock);
 740         return seq_list_start(&dd->dispatch, *pos);
 741 }
 742
 743 static void *deadline_dispatch_next(struct seq_file *m, void *v, loff_t *pos)
 744 {
 745         struct request_queue *q = m->private;
 746         struct deadline_data *dd = q->elevator->elevator_data;
 747
 748         return seq_list_next(v, &dd->dispatch, pos);
 749 }
 750
 751 static void deadline_dispatch_stop(struct seq_file *m, void *v)
 752         __releases(&dd->lock)
 753 {
 754         struct request_queue *q = m->private;
 755         struct deadline_data *dd = q->elevator->elevator_data;
 756
 757         spin_unlock(&dd->lock);
 758 }
 759
 760 static const struct seq_operations deadline_dispatch_seq_ops = {
 761         .start  = deadline_dispatch_start,
 762         .next   = deadline_dispatch_next,
 763         .stop   = deadline_dispatch_stop,
 764         .show   = blk_mq_debugfs_rq_show,
 765 };
 766
 767 #define DEADLINE_QUEUE_DDIR_ATTRS(name)                                         \
 768         {#name "_fifo_list", 0400, .seq_ops = &deadline_##name##_fifo_seq_ops}, \
 769         {#name "_next_rq", 0400, deadline_##name##_next_rq_show}
 770 static const struct blk_mq_debugfs_attr deadline_queue_debugfs_attrs[] = {
 771         DEADLINE_QUEUE_DDIR_ATTRS(read),
 772         DEADLINE_QUEUE_DDIR_ATTRS(write),
 773         {"batching", 0400, deadline_batching_show},
 774         {"starved", 0400, deadline_starved_show},
 775         {"dispatch", 0400, .seq_ops = &deadline_dispatch_seq_ops},
 776         {},
 777 };
 778 #undef DEADLINE_QUEUE_DDIR_ATTRS
 779 #endif
 780
 781 static struct elevator_type mq_deadline = {
 782         .ops = {
 783                 .insert_requests        = dd_insert_requests,
 784                 .dispatch_request       = dd_dispatch_request,
 785                 .prepare_request        = dd_prepare_request,
 786                 .finish_request         = dd_finish_request,
 787                 .next_request           = elv_rb_latter_request,
 788                 .former_request         = elv_rb_former_request,
 789                 .bio_merge              = dd_bio_merge,
 790                 .request_merge          = dd_request_merge,
 791                 .requests_merged        = dd_merged_requests,
 792                 .request_merged         = dd_request_merged,
 793                 .has_work               = dd_has_work,
 794                 .init_sched             = dd_init_queue,
 795                 .exit_sched             = dd_exit_queue,
 796         },
 797
 798 #ifdef CONFIG_BLK_DEBUG_FS
 799         .queue_debugfs_attrs = deadline_queue_debugfs_attrs,
 800 #endif
 801         .elevator_attrs = deadline_attrs,
 802         .elevator_name = "mq-deadline",
 803         .elevator_alias = "deadline",
 804         .elevator_features = ELEVATOR_F_ZBD_SEQ_WRITE,
 805         .elevator_owner = THIS_MODULE,
 806 };
 807 MODULE_ALIAS("mq-deadline-iosched");
 808
 809 static int __init deadline_init(void)
 810 {
 811         return elv_register(&mq_deadline);
 812 }
 813
 814 static void __exit deadline_exit(void)
 815 {
 816         elv_unregister(&mq_deadline);
 817 }
 818
 819 module_init(deadline_init);
 820 module_exit(deadline_exit);
 821
 822 MODULE_AUTHOR("Jens Axboe");
 823 MODULE_LICENSE("GPL");
 824 MODULE_DESCRIPTION("MQ deadline IO scheduler");