block/mq-deadline.c

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