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