block/mq-deadline-main.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 #include "mq-deadline-cgroup.h"
  29
  30 /*
  31  * See Documentation/block/deadline-iosched.rst
  32  */
  33 static const int read_expire = HZ / 2;  /* max time before a read is submitted. */
  34 static const int write_expire = 5 * HZ; /* ditto for writes, these limits are SOFT! */
  35 /*
  36  * Time after which to dispatch lower priority requests even if higher
  37  * priority requests are pending.
  38  */
  39 static const int aging_expire = 10 * HZ;
  40 static const int writes_starved = 2;    /* max times reads can starve a write */
  41 static const int fifo_batch = 16;       /* # of sequential requests treated as one
  42                                      by the above parameters. For throughput. */
  43
  44 enum dd_data_dir {
  45         DD_READ         = READ,
  46         DD_WRITE        = WRITE,
  47 };
  48
  49 enum { DD_DIR_COUNT = 2 };
  50
  51 enum dd_prio {
  52         DD_RT_PRIO      = 0,
  53         DD_BE_PRIO      = 1,
  54         DD_IDLE_PRIO    = 2,
  55         DD_PRIO_MAX     = 2,
  56 };
  57
  58 enum { DD_PRIO_COUNT = 3 };
  59
  60 /* I/O statistics for all I/O priorities (enum dd_prio). */
  61 struct io_stats {
  62         struct io_stats_per_prio stats[DD_PRIO_COUNT];
  63 };
  64
  65 /*
  66  * Deadline scheduler data per I/O priority (enum dd_prio). Requests are
  67  * present on both sort_list[] and fifo_list[].
  68  */
  69 struct dd_per_prio {
  70         struct list_head dispatch;
  71         struct rb_root sort_list[DD_DIR_COUNT];
  72         struct list_head fifo_list[DD_DIR_COUNT];
  73         /* Next request in FIFO order. Read, write or both are NULL. */
  74         struct request *next_rq[DD_DIR_COUNT];
  75 };
  76
  77 struct deadline_data {
  78         /*
  79          * run time data
  80          */
  81
  82         /* Request queue that owns this data structure. */
  83         struct request_queue *queue;
  84
  85         struct dd_per_prio per_prio[DD_PRIO_COUNT];
  86
  87         /* Data direction of latest dispatched request. */
  88         enum dd_data_dir last_dir;
  89         unsigned int batching;          /* number of sequential requests made */
  90         unsigned int starved;           /* times reads have starved writes */
  91
  92         struct io_stats __percpu *stats;
  93
  94         /*
  95          * settings that change how the i/o scheduler behaves
  96          */
  97         int fifo_expire[DD_DIR_COUNT];
  98         int fifo_batch;
  99         int writes_starved;
 100         int front_merges;
 101         u32 async_depth;
 102         int aging_expire;
 103
 104         spinlock_t lock;
 105         spinlock_t zone_lock;
 106 };
 107
 108 /* Count one event of type 'event_type' and with I/O priority 'prio' */
 109 #define dd_count(dd, event_type, prio) do {                             \
 110         struct io_stats *io_stats = get_cpu_ptr((dd)->stats);           \
 111                                                                         \
 112         BUILD_BUG_ON(!__same_type((dd), struct deadline_data *));       \
 113         BUILD_BUG_ON(!__same_type((prio), enum dd_prio));               \
 114         local_inc(&io_stats->stats[(prio)].event_type);                 \
 115         put_cpu_ptr(io_stats);                                          \
 116 } while (0)
 117
 118 /*
 119  * Returns the total number of dd_count(dd, event_type, prio) calls across all
 120  * CPUs. No locking or barriers since it is fine if the returned sum is slightly
 121  * outdated.
 122  */
 123 #define dd_sum(dd, event_type, prio) ({                                 \
 124         unsigned int cpu;                                               \
 125         u32 sum = 0;                                                    \
 126                                                                         \
 127         BUILD_BUG_ON(!__same_type((dd), struct deadline_data *));       \
 128         BUILD_BUG_ON(!__same_type((prio), enum dd_prio));               \
 129         for_each_present_cpu(cpu)                                       \
 130                 sum += local_read(&per_cpu_ptr((dd)->stats, cpu)->      \
 131                                   stats[(prio)].event_type);            \
 132         sum;                                                            \
 133 })
 134
 135 /* Maps an I/O priority class to a deadline scheduler priority. */
 136 static const enum dd_prio ioprio_class_to_prio[] = {
 137         [IOPRIO_CLASS_NONE]     = DD_BE_PRIO,
 138         [IOPRIO_CLASS_RT]       = DD_RT_PRIO,
 139         [IOPRIO_CLASS_BE]       = DD_BE_PRIO,
 140         [IOPRIO_CLASS_IDLE]     = DD_IDLE_PRIO,
 141 };
 142
 143 static inline struct rb_root *
 144 deadline_rb_root(struct dd_per_prio *per_prio, struct request *rq)
 145 {
 146         return &per_prio->sort_list[rq_data_dir(rq)];
 147 }
 148
 149 /*
 150  * Returns the I/O priority class (IOPRIO_CLASS_*) that has been assigned to a
 151  * request.
 152  */
 153 static u8 dd_rq_ioclass(struct request *rq)
 154 {
 155         return IOPRIO_PRIO_CLASS(req_get_ioprio(rq));
 156 }
 157
 158 /*
 159  * get the request after `rq' in sector-sorted order
 160  */
 161 static inline struct request *
 162 deadline_latter_request(struct request *rq)
 163 {
 164         struct rb_node *node = rb_next(&rq->rb_node);
 165
 166         if (node)
 167                 return rb_entry_rq(node);
 168
 169         return NULL;
 170 }
 171
 172 static void
 173 deadline_add_rq_rb(struct dd_per_prio *per_prio, struct request *rq)
 174 {
 175         struct rb_root *root = deadline_rb_root(per_prio, rq);
 176
 177         elv_rb_add(root, rq);
 178 }
 179
 180 static inline void
 181 deadline_del_rq_rb(struct dd_per_prio *per_prio, struct request *rq)
 182 {
 183         const enum dd_data_dir data_dir = rq_data_dir(rq);
 184
 185         if (per_prio->next_rq[data_dir] == rq)
 186                 per_prio->next_rq[data_dir] = deadline_latter_request(rq);
 187
 188         elv_rb_del(deadline_rb_root(per_prio, rq), rq);
 189 }
 190
 191 /*
 192  * remove rq from rbtree and fifo.
 193  */
 194 static void deadline_remove_request(struct request_queue *q,
 195                                     struct dd_per_prio *per_prio,
 196                                     struct request *rq)
 197 {
 198         list_del_init(&rq->queuelist);
 199
 200         /*
 201          * We might not be on the rbtree, if we are doing an insert merge
 202          */
 203         if (!RB_EMPTY_NODE(&rq->rb_node))
 204                 deadline_del_rq_rb(per_prio, rq);
 205
 206         elv_rqhash_del(q, rq);
 207         if (q->last_merge == rq)
 208                 q->last_merge = NULL;
 209 }
 210
 211 static void dd_request_merged(struct request_queue *q, struct request *req,
 212                               enum elv_merge type)
 213 {
 214         struct deadline_data *dd = q->elevator->elevator_data;
 215         const u8 ioprio_class = dd_rq_ioclass(req);
 216         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 217         struct dd_per_prio *per_prio = &dd->per_prio[prio];
 218
 219         /*
 220          * if the merge was a front merge, we need to reposition request
 221          */
 222         if (type == ELEVATOR_FRONT_MERGE) {
 223                 elv_rb_del(deadline_rb_root(per_prio, req), req);
 224                 deadline_add_rq_rb(per_prio, req);
 225         }
 226 }
 227
 228 /*
 229  * Callback function that is invoked after @next has been merged into @req.
 230  */
 231 static void dd_merged_requests(struct request_queue *q, struct request *req,
 232                                struct request *next)
 233 {
 234         struct deadline_data *dd = q->elevator->elevator_data;
 235         const u8 ioprio_class = dd_rq_ioclass(next);
 236         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 237         struct dd_blkcg *blkcg = next->elv.priv[0];
 238
 239         dd_count(dd, merged, prio);
 240         ddcg_count(blkcg, merged, ioprio_class);
 241
 242         /*
 243          * if next expires before rq, assign its expire time to rq
 244          * and move into next position (next will be deleted) in fifo
 245          */
 246         if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
 247                 if (time_before((unsigned long)next->fifo_time,
 248                                 (unsigned long)req->fifo_time)) {
 249                         list_move(&req->queuelist, &next->queuelist);
 250                         req->fifo_time = next->fifo_time;
 251                 }
 252         }
 253
 254         /*
 255          * kill knowledge of next, this one is a goner
 256          */
 257         deadline_remove_request(q, &dd->per_prio[prio], next);
 258 }
 259
 260 /*
 261  * move an entry to dispatch queue
 262  */
 263 static void
 264 deadline_move_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 265                       struct request *rq)
 266 {
 267         const enum dd_data_dir data_dir = rq_data_dir(rq);
 268
 269         per_prio->next_rq[data_dir] = deadline_latter_request(rq);
 270
 271         /*
 272          * take it off the sort and fifo list
 273          */
 274         deadline_remove_request(rq->q, per_prio, rq);
 275 }
 276
 277 /* Number of requests queued for a given priority level. */
 278 static u32 dd_queued(struct deadline_data *dd, enum dd_prio prio)
 279 {
 280         return dd_sum(dd, inserted, prio) - dd_sum(dd, completed, prio);
 281 }
 282
 283 /*
 284  * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
 285  * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
 286  */
 287 static inline int deadline_check_fifo(struct dd_per_prio *per_prio,
 288                                       enum dd_data_dir data_dir)
 289 {
 290         struct request *rq = rq_entry_fifo(per_prio->fifo_list[data_dir].next);
 291
 292         /*
 293          * rq is expired!
 294          */
 295         if (time_after_eq(jiffies, (unsigned long)rq->fifo_time))
 296                 return 1;
 297
 298         return 0;
 299 }
 300
 301 /*
 302  * For the specified data direction, return the next request to
 303  * dispatch using arrival ordered lists.
 304  */
 305 static struct request *
 306 deadline_fifo_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 307                       enum dd_data_dir data_dir)
 308 {
 309         struct request *rq;
 310         unsigned long flags;
 311
 312         if (list_empty(&per_prio->fifo_list[data_dir]))
 313                 return NULL;
 314
 315         rq = rq_entry_fifo(per_prio->fifo_list[data_dir].next);
 316         if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q))
 317                 return rq;
 318
 319         /*
 320          * Look for a write request that can be dispatched, that is one with
 321          * an unlocked target zone.
 322          */
 323         spin_lock_irqsave(&dd->zone_lock, flags);
 324         list_for_each_entry(rq, &per_prio->fifo_list[DD_WRITE], queuelist) {
 325                 if (blk_req_can_dispatch_to_zone(rq))
 326                         goto out;
 327         }
 328         rq = NULL;
 329 out:
 330         spin_unlock_irqrestore(&dd->zone_lock, flags);
 331
 332         return rq;
 333 }
 334
 335 /*
 336  * For the specified data direction, return the next request to
 337  * dispatch using sector position sorted lists.
 338  */
 339 static struct request *
 340 deadline_next_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 341                       enum dd_data_dir data_dir)
 342 {
 343         struct request *rq;
 344         unsigned long flags;
 345
 346         rq = per_prio->next_rq[data_dir];
 347         if (!rq)
 348                 return NULL;
 349
 350         if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q))
 351                 return rq;
 352
 353         /*
 354          * Look for a write request that can be dispatched, that is one with
 355          * an unlocked target zone.
 356          */
 357         spin_lock_irqsave(&dd->zone_lock, flags);
 358         while (rq) {
 359                 if (blk_req_can_dispatch_to_zone(rq))
 360                         break;
 361                 rq = deadline_latter_request(rq);
 362         }
 363         spin_unlock_irqrestore(&dd->zone_lock, flags);
 364
 365         return rq;
 366 }
 367
 368 /*
 369  * deadline_dispatch_requests selects the best request according to
 370  * read/write expire, fifo_batch, etc and with a start time <= @latest.
 371  */
 372 static struct request *__dd_dispatch_request(struct deadline_data *dd,
 373                                              struct dd_per_prio *per_prio,
 374                                              u64 latest_start_ns)
 375 {
 376         struct request *rq, *next_rq;
 377         enum dd_data_dir data_dir;
 378         struct dd_blkcg *blkcg;
 379         enum dd_prio prio;
 380         u8 ioprio_class;
 381
 382         lockdep_assert_held(&dd->lock);
 383
 384         if (!list_empty(&per_prio->dispatch)) {
 385                 rq = list_first_entry(&per_prio->dispatch, struct request,
 386                                       queuelist);
 387                 if (rq->start_time_ns > latest_start_ns)
 388                         return NULL;
 389                 list_del_init(&rq->queuelist);
 390                 goto done;
 391         }
 392
 393         /*
 394          * batches are currently reads XOR writes
 395          */
 396         rq = deadline_next_request(dd, per_prio, dd->last_dir);
 397         if (rq && dd->batching < dd->fifo_batch)
 398                 /* we have a next request are still entitled to batch */
 399                 goto dispatch_request;
 400
 401         /*
 402          * at this point we are not running a batch. select the appropriate
 403          * data direction (read / write)
 404          */
 405
 406         if (!list_empty(&per_prio->fifo_list[DD_READ])) {
 407                 BUG_ON(RB_EMPTY_ROOT(&per_prio->sort_list[DD_READ]));
 408
 409                 if (deadline_fifo_request(dd, per_prio, DD_WRITE) &&
 410                     (dd->starved++ >= dd->writes_starved))
 411                         goto dispatch_writes;
 412
 413                 data_dir = DD_READ;
 414
 415                 goto dispatch_find_request;
 416         }
 417
 418         /*
 419          * there are either no reads or writes have been starved
 420          */
 421
 422         if (!list_empty(&per_prio->fifo_list[DD_WRITE])) {
 423 dispatch_writes:
 424                 BUG_ON(RB_EMPTY_ROOT(&per_prio->sort_list[DD_WRITE]));
 425
 426                 dd->starved = 0;
 427
 428                 data_dir = DD_WRITE;
 429
 430                 goto dispatch_find_request;
 431         }
 432
 433         return NULL;
 434
 435 dispatch_find_request:
 436         /*
 437          * we are not running a batch, find best request for selected data_dir
 438          */
 439         next_rq = deadline_next_request(dd, per_prio, data_dir);
 440         if (deadline_check_fifo(per_prio, data_dir) || !next_rq) {
 441                 /*
 442                  * A deadline has expired, the last request was in the other
 443                  * direction, or we have run out of higher-sectored requests.
 444                  * Start again from the request with the earliest expiry time.
 445                  */
 446                 rq = deadline_fifo_request(dd, per_prio, data_dir);
 447         } else {
 448                 /*
 449                  * The last req was the same dir and we have a next request in
 450                  * sort order. No expired requests so continue on from here.
 451                  */
 452                 rq = next_rq;
 453         }
 454
 455         /*
 456          * For a zoned block device, if we only have writes queued and none of
 457          * them can be dispatched, rq will be NULL.
 458          */
 459         if (!rq)
 460                 return NULL;
 461
 462         dd->last_dir = data_dir;
 463         dd->batching = 0;
 464
 465 dispatch_request:
 466         if (rq->start_time_ns > latest_start_ns)
 467                 return NULL;
 468         /*
 469          * rq is the selected appropriate request.
 470          */
 471         dd->batching++;
 472         deadline_move_request(dd, per_prio, rq);
 473 done:
 474         ioprio_class = dd_rq_ioclass(rq);
 475         prio = ioprio_class_to_prio[ioprio_class];
 476         dd_count(dd, dispatched, prio);
 477         blkcg = rq->elv.priv[0];
 478         ddcg_count(blkcg, dispatched, ioprio_class);
 479         /*
 480          * If the request needs its target zone locked, do it.
 481          */
 482         blk_req_zone_write_lock(rq);
 483         rq->rq_flags |= RQF_STARTED;
 484         return rq;
 485 }
 486
 487 /*
 488  * Called from blk_mq_run_hw_queue() -> __blk_mq_sched_dispatch_requests().
 489  *
 490  * One confusing aspect here is that we get called for a specific
 491  * hardware queue, but we may return a request that is for a
 492  * different hardware queue. This is because mq-deadline has shared
 493  * state for all hardware queues, in terms of sorting, FIFOs, etc.
 494  */
 495 static struct request *dd_dispatch_request(struct blk_mq_hw_ctx *hctx)
 496 {
 497         struct deadline_data *dd = hctx->queue->elevator->elevator_data;
 498         const u64 now_ns = ktime_get_ns();
 499         struct request *rq = NULL;
 500         enum dd_prio prio;
 501
 502         spin_lock(&dd->lock);
 503         /*
 504          * Start with dispatching requests whose deadline expired more than
 505          * aging_expire jiffies ago.
 506          */
 507         for (prio = DD_BE_PRIO; prio <= DD_PRIO_MAX; prio++) {
 508                 rq = __dd_dispatch_request(dd, &dd->per_prio[prio], now_ns -
 509                                            jiffies_to_nsecs(dd->aging_expire));
 510                 if (rq)
 511                         goto unlock;
 512         }
 513         /*
 514          * Next, dispatch requests in priority order. Ignore lower priority
 515          * requests if any higher priority requests are pending.
 516          */
 517         for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 518                 rq = __dd_dispatch_request(dd, &dd->per_prio[prio], now_ns);
 519                 if (rq || dd_queued(dd, prio))
 520                         break;
 521         }
 522
 523 unlock:
 524         spin_unlock(&dd->lock);
 525
 526         return rq;
 527 }
 528
 529 /*
 530  * Called by __blk_mq_alloc_request(). The shallow_depth value set by this
 531  * function is used by __blk_mq_get_tag().
 532  */
 533 static void dd_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
 534 {
 535         struct deadline_data *dd = data->q->elevator->elevator_data;
 536
 537         /* Do not throttle synchronous reads. */
 538         if (op_is_sync(op) && !op_is_write(op))
 539                 return;
 540
 541         /*
 542          * Throttle asynchronous requests and writes such that these requests
 543          * do not block the allocation of synchronous requests.
 544          */
 545         data->shallow_depth = dd->async_depth;
 546 }
 547
 548 /* Called by blk_mq_update_nr_requests(). */
 549 static void dd_depth_updated(struct blk_mq_hw_ctx *hctx)
 550 {
 551         struct request_queue *q = hctx->queue;
 552         struct deadline_data *dd = q->elevator->elevator_data;
 553         struct blk_mq_tags *tags = hctx->sched_tags;
 554
 555         dd->async_depth = max(1UL, 3 * q->nr_requests / 4);
 556
 557         sbitmap_queue_min_shallow_depth(tags->bitmap_tags, dd->async_depth);
 558 }
 559
 560 /* Called by blk_mq_init_hctx() and blk_mq_init_sched(). */
 561 static int dd_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 562 {
 563         dd_depth_updated(hctx);
 564         return 0;
 565 }
 566
 567 static void dd_exit_sched(struct elevator_queue *e)
 568 {
 569         struct deadline_data *dd = e->elevator_data;
 570         enum dd_prio prio;
 571
 572         dd_deactivate_policy(dd->queue);
 573
 574         for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 575                 struct dd_per_prio *per_prio = &dd->per_prio[prio];
 576
 577                 WARN_ON_ONCE(!list_empty(&per_prio->fifo_list[DD_READ]));
 578                 WARN_ON_ONCE(!list_empty(&per_prio->fifo_list[DD_WRITE]));
 579         }
 580
 581         free_percpu(dd->stats);
 582
 583         kfree(dd);
 584 }
 585
 586 /*
 587  * Initialize elevator private data (deadline_data) and associate with blkcg.
 588  */
 589 static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
 590 {
 591         struct deadline_data *dd;
 592         struct elevator_queue *eq;
 593         enum dd_prio prio;
 594         int ret = -ENOMEM;
 595
 596         /*
 597          * Initialization would be very tricky if the queue is not frozen,
 598          * hence the warning statement below.
 599          */
 600         WARN_ON_ONCE(!percpu_ref_is_zero(&q->q_usage_counter));
 601
 602         eq = elevator_alloc(q, e);
 603         if (!eq)
 604                 return ret;
 605
 606         dd = kzalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
 607         if (!dd)
 608                 goto put_eq;
 609
 610         eq->elevator_data = dd;
 611
 612         dd->stats = alloc_percpu_gfp(typeof(*dd->stats),
 613                                      GFP_KERNEL | __GFP_ZERO);
 614         if (!dd->stats)
 615                 goto free_dd;
 616
 617         dd->queue = q;
 618
 619         for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 620                 struct dd_per_prio *per_prio = &dd->per_prio[prio];
 621
 622                 INIT_LIST_HEAD(&per_prio->dispatch);
 623                 INIT_LIST_HEAD(&per_prio->fifo_list[DD_READ]);
 624                 INIT_LIST_HEAD(&per_prio->fifo_list[DD_WRITE]);
 625                 per_prio->sort_list[DD_READ] = RB_ROOT;
 626                 per_prio->sort_list[DD_WRITE] = RB_ROOT;
 627         }
 628         dd->fifo_expire[DD_READ] = read_expire;
 629         dd->fifo_expire[DD_WRITE] = write_expire;
 630         dd->writes_starved = writes_starved;
 631         dd->front_merges = 1;
 632         dd->last_dir = DD_WRITE;
 633         dd->fifo_batch = fifo_batch;
 634         dd->aging_expire = aging_expire;
 635         spin_lock_init(&dd->lock);
 636         spin_lock_init(&dd->zone_lock);
 637
 638         ret = dd_activate_policy(q);
 639         if (ret)
 640                 goto free_stats;
 641
 642         ret = 0;
 643         q->elevator = eq;
 644         return 0;
 645
 646 free_stats:
 647         free_percpu(dd->stats);
 648
 649 free_dd:
 650         kfree(dd);
 651
 652 put_eq:
 653         kobject_put(&eq->kobj);
 654         return ret;
 655 }
 656
 657 /*
 658  * Try to merge @bio into an existing request. If @bio has been merged into
 659  * an existing request, store the pointer to that request into *@rq.
 660  */
 661 static int dd_request_merge(struct request_queue *q, struct request **rq,
 662                             struct bio *bio)
 663 {
 664         struct deadline_data *dd = q->elevator->elevator_data;
 665         const u8 ioprio_class = IOPRIO_PRIO_CLASS(bio->bi_ioprio);
 666         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 667         struct dd_per_prio *per_prio = &dd->per_prio[prio];
 668         sector_t sector = bio_end_sector(bio);
 669         struct request *__rq;
 670
 671         if (!dd->front_merges)
 672                 return ELEVATOR_NO_MERGE;
 673
 674         __rq = elv_rb_find(&per_prio->sort_list[bio_data_dir(bio)], sector);
 675         if (__rq) {
 676                 BUG_ON(sector != blk_rq_pos(__rq));
 677
 678                 if (elv_bio_merge_ok(__rq, bio)) {
 679                         *rq = __rq;
 680                         return ELEVATOR_FRONT_MERGE;
 681                 }
 682         }
 683
 684         return ELEVATOR_NO_MERGE;
 685 }
 686
 687 /*
 688  * Attempt to merge a bio into an existing request. This function is called
 689  * before @bio is associated with a request.
 690  */
 691 static bool dd_bio_merge(struct request_queue *q, struct bio *bio,
 692                 unsigned int nr_segs)
 693 {
 694         struct deadline_data *dd = q->elevator->elevator_data;
 695         struct request *free = NULL;
 696         bool ret;
 697
 698         spin_lock(&dd->lock);
 699         ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
 700         spin_unlock(&dd->lock);
 701
 702         if (free)
 703                 blk_mq_free_request(free);
 704
 705         return ret;
 706 }
 707
 708 /*
 709  * add rq to rbtree and fifo
 710  */
 711 static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 712                               bool at_head)
 713 {
 714         struct request_queue *q = hctx->queue;
 715         struct deadline_data *dd = q->elevator->elevator_data;
 716         const enum dd_data_dir data_dir = rq_data_dir(rq);
 717         u16 ioprio = req_get_ioprio(rq);
 718         u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
 719         struct dd_per_prio *per_prio;
 720         enum dd_prio prio;
 721         struct dd_blkcg *blkcg;
 722         LIST_HEAD(free);
 723
 724         lockdep_assert_held(&dd->lock);
 725
 726         /*
 727          * This may be a requeue of a write request that has locked its
 728          * target zone. If it is the case, this releases the zone lock.
 729          */
 730         blk_req_zone_write_unlock(rq);
 731
 732         /*
 733          * If a block cgroup has been associated with the submitter and if an
 734          * I/O priority has been set in the associated block cgroup, use the
 735          * lowest of the cgroup priority and the request priority for the
 736          * request. If no priority has been set in the request, use the cgroup
 737          * priority.
 738          */
 739         prio = ioprio_class_to_prio[ioprio_class];
 740         dd_count(dd, inserted, prio);
 741         blkcg = dd_blkcg_from_bio(rq->bio);
 742         ddcg_count(blkcg, inserted, ioprio_class);
 743         rq->elv.priv[0] = blkcg;
 744
 745         if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
 746                 blk_mq_free_requests(&free);
 747                 return;
 748         }
 749
 750         trace_block_rq_insert(rq);
 751
 752         per_prio = &dd->per_prio[prio];
 753         if (at_head) {
 754                 list_add(&rq->queuelist, &per_prio->dispatch);
 755         } else {
 756                 deadline_add_rq_rb(per_prio, rq);
 757
 758                 if (rq_mergeable(rq)) {
 759                         elv_rqhash_add(q, rq);
 760                         if (!q->last_merge)
 761                                 q->last_merge = rq;
 762                 }
 763
 764                 /*
 765                  * set expire time and add to fifo list
 766                  */
 767                 rq->fifo_time = jiffies + dd->fifo_expire[data_dir];
 768                 list_add_tail(&rq->queuelist, &per_prio->fifo_list[data_dir]);
 769         }
 770 }
 771
 772 /*
 773  * Called from blk_mq_sched_insert_request() or blk_mq_sched_insert_requests().
 774  */
 775 static void dd_insert_requests(struct blk_mq_hw_ctx *hctx,
 776                                struct list_head *list, bool at_head)
 777 {
 778         struct request_queue *q = hctx->queue;
 779         struct deadline_data *dd = q->elevator->elevator_data;
 780
 781         spin_lock(&dd->lock);
 782         while (!list_empty(list)) {
 783                 struct request *rq;
 784
 785                 rq = list_first_entry(list, struct request, queuelist);
 786                 list_del_init(&rq->queuelist);
 787                 dd_insert_request(hctx, rq, at_head);
 788         }
 789         spin_unlock(&dd->lock);
 790 }
 791
 792 /* Callback from inside blk_mq_rq_ctx_init(). */
 793 static void dd_prepare_request(struct request *rq)
 794 {
 795         rq->elv.priv[0] = NULL;
 796 }
 797
 798 /*
 799  * Callback from inside blk_mq_free_request().
 800  *
 801  * For zoned block devices, write unlock the target zone of
 802  * completed write requests. Do this while holding the zone lock
 803  * spinlock so that the zone is never unlocked while deadline_fifo_request()
 804  * or deadline_next_request() are executing. This function is called for
 805  * all requests, whether or not these requests complete successfully.
 806  *
 807  * For a zoned block device, __dd_dispatch_request() may have stopped
 808  * dispatching requests if all the queued requests are write requests directed
 809  * at zones that are already locked due to on-going write requests. To ensure
 810  * write request dispatch progress in this case, mark the queue as needing a
 811  * restart to ensure that the queue is run again after completion of the
 812  * request and zones being unlocked.
 813  */
 814 static void dd_finish_request(struct request *rq)
 815 {
 816         struct request_queue *q = rq->q;
 817         struct deadline_data *dd = q->elevator->elevator_data;
 818         struct dd_blkcg *blkcg = rq->elv.priv[0];
 819         const u8 ioprio_class = dd_rq_ioclass(rq);
 820         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 821         struct dd_per_prio *per_prio = &dd->per_prio[prio];
 822
 823         dd_count(dd, completed, prio);
 824         ddcg_count(blkcg, completed, ioprio_class);
 825
 826         if (blk_queue_is_zoned(q)) {
 827                 unsigned long flags;
 828
 829                 spin_lock_irqsave(&dd->zone_lock, flags);
 830                 blk_req_zone_write_unlock(rq);
 831                 if (!list_empty(&per_prio->fifo_list[DD_WRITE]))
 832                         blk_mq_sched_mark_restart_hctx(rq->mq_hctx);
 833                 spin_unlock_irqrestore(&dd->zone_lock, flags);
 834         }
 835 }
 836
 837 static bool dd_has_work_for_prio(struct dd_per_prio *per_prio)
 838 {
 839         return !list_empty_careful(&per_prio->dispatch) ||
 840                 !list_empty_careful(&per_prio->fifo_list[DD_READ]) ||
 841                 !list_empty_careful(&per_prio->fifo_list[DD_WRITE]);
 842 }
 843
 844 static bool dd_has_work(struct blk_mq_hw_ctx *hctx)
 845 {
 846         struct deadline_data *dd = hctx->queue->elevator->elevator_data;
 847         enum dd_prio prio;
 848
 849         for (prio = 0; prio <= DD_PRIO_MAX; prio++)
 850                 if (dd_has_work_for_prio(&dd->per_prio[prio]))
 851                         return true;
 852
 853         return false;
 854 }
 855
 856 /*
 857  * sysfs parts below
 858  */
 859 #define SHOW_INT(__FUNC, __VAR)                                         \
 860 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
 861 {                                                                       \
 862         struct deadline_data *dd = e->elevator_data;                    \
 863                                                                         \
 864         return sysfs_emit(page, "%d\n", __VAR);                         \
 865 }
 866 #define SHOW_JIFFIES(__FUNC, __VAR) SHOW_INT(__FUNC, jiffies_to_msecs(__VAR))
 867 SHOW_JIFFIES(deadline_read_expire_show, dd->fifo_expire[DD_READ]);
 868 SHOW_JIFFIES(deadline_write_expire_show, dd->fifo_expire[DD_WRITE]);
 869 SHOW_JIFFIES(deadline_aging_expire_show, dd->aging_expire);
 870 SHOW_INT(deadline_writes_starved_show, dd->writes_starved);
 871 SHOW_INT(deadline_front_merges_show, dd->front_merges);
 872 SHOW_INT(deadline_async_depth_show, dd->front_merges);
 873 SHOW_INT(deadline_fifo_batch_show, dd->fifo_batch);
 874 #undef SHOW_INT
 875 #undef SHOW_JIFFIES
 876
 877 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
 878 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
 879 {                                                                       \
 880         struct deadline_data *dd = e->elevator_data;                    \
 881         int __data, __ret;                                              \
 882                                                                         \
 883         __ret = kstrtoint(page, 0, &__data);                            \
 884         if (__ret < 0)                                                  \
 885                 return __ret;                                           \
 886         if (__data < (MIN))                                             \
 887                 __data = (MIN);                                         \
 888         else if (__data > (MAX))                                        \
 889                 __data = (MAX);                                         \
 890         *(__PTR) = __CONV(__data);                                      \
 891         return count;                                                   \
 892 }
 893 #define STORE_INT(__FUNC, __PTR, MIN, MAX)                              \
 894         STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, )
 895 #define STORE_JIFFIES(__FUNC, __PTR, MIN, MAX)                          \
 896         STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, msecs_to_jiffies)
 897 STORE_JIFFIES(deadline_read_expire_store, &dd->fifo_expire[DD_READ], 0, INT_MAX);
 898 STORE_JIFFIES(deadline_write_expire_store, &dd->fifo_expire[DD_WRITE], 0, INT_MAX);
 899 STORE_JIFFIES(deadline_aging_expire_store, &dd->aging_expire, 0, INT_MAX);
 900 STORE_INT(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX);
 901 STORE_INT(deadline_front_merges_store, &dd->front_merges, 0, 1);
 902 STORE_INT(deadline_async_depth_store, &dd->front_merges, 1, INT_MAX);
 903 STORE_INT(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX);
 904 #undef STORE_FUNCTION
 905 #undef STORE_INT
 906 #undef STORE_JIFFIES
 907
 908 #define DD_ATTR(name) \
 909         __ATTR(name, 0644, deadline_##name##_show, deadline_##name##_store)
 910
 911 static struct elv_fs_entry deadline_attrs[] = {
 912         DD_ATTR(read_expire),
 913         DD_ATTR(write_expire),
 914         DD_ATTR(writes_starved),
 915         DD_ATTR(front_merges),
 916         DD_ATTR(async_depth),
 917         DD_ATTR(fifo_batch),
 918         DD_ATTR(aging_expire),
 919         __ATTR_NULL
 920 };
 921
 922 #ifdef CONFIG_BLK_DEBUG_FS
 923 #define DEADLINE_DEBUGFS_DDIR_ATTRS(prio, data_dir, name)               \
 924 static void *deadline_##name##_fifo_start(struct seq_file *m,           \
 925                                           loff_t *pos)                  \
 926         __acquires(&dd->lock)                                           \
 927 {                                                                       \
 928         struct request_queue *q = m->private;                           \
 929         struct deadline_data *dd = q->elevator->elevator_data;          \
 930         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
 931                                                                         \
 932         spin_lock(&dd->lock);                                           \
 933         return seq_list_start(&per_prio->fifo_list[data_dir], *pos);    \
 934 }                                                                       \
 935                                                                         \
 936 static void *deadline_##name##_fifo_next(struct seq_file *m, void *v,   \
 937                                          loff_t *pos)                   \
 938 {                                                                       \
 939         struct request_queue *q = m->private;                           \
 940         struct deadline_data *dd = q->elevator->elevator_data;          \
 941         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
 942                                                                         \
 943         return seq_list_next(v, &per_prio->fifo_list[data_dir], pos);   \
 944 }                                                                       \
 945                                                                         \
 946 static void deadline_##name##_fifo_stop(struct seq_file *m, void *v)    \
 947         __releases(&dd->lock)                                           \
 948 {                                                                       \
 949         struct request_queue *q = m->private;                           \
 950         struct deadline_data *dd = q->elevator->elevator_data;          \
 951                                                                         \
 952         spin_unlock(&dd->lock);                                         \
 953 }                                                                       \
 954                                                                         \
 955 static const struct seq_operations deadline_##name##_fifo_seq_ops = {   \
 956         .start  = deadline_##name##_fifo_start,                         \
 957         .next   = deadline_##name##_fifo_next,                          \
 958         .stop   = deadline_##name##_fifo_stop,                          \
 959         .show   = blk_mq_debugfs_rq_show,                               \
 960 };                                                                      \
 961                                                                         \
 962 static int deadline_##name##_next_rq_show(void *data,                   \
 963                                           struct seq_file *m)           \
 964 {                                                                       \
 965         struct request_queue *q = data;                                 \
 966         struct deadline_data *dd = q->elevator->elevator_data;          \
 967         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
 968         struct request *rq = per_prio->next_rq[data_dir];               \
 969                                                                         \
 970         if (rq)                                                         \
 971                 __blk_mq_debugfs_rq_show(m, rq);                        \
 972         return 0;                                                       \
 973 }
 974
 975 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_RT_PRIO, DD_READ, read0);
 976 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_RT_PRIO, DD_WRITE, write0);
 977 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_BE_PRIO, DD_READ, read1);
 978 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_BE_PRIO, DD_WRITE, write1);
 979 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_IDLE_PRIO, DD_READ, read2);
 980 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_IDLE_PRIO, DD_WRITE, write2);
 981 #undef DEADLINE_DEBUGFS_DDIR_ATTRS
 982
 983 static int deadline_batching_show(void *data, struct seq_file *m)
 984 {
 985         struct request_queue *q = data;
 986         struct deadline_data *dd = q->elevator->elevator_data;
 987
 988         seq_printf(m, "%u\n", dd->batching);
 989         return 0;
 990 }
 991
 992 static int deadline_starved_show(void *data, struct seq_file *m)
 993 {
 994         struct request_queue *q = data;
 995         struct deadline_data *dd = q->elevator->elevator_data;
 996
 997         seq_printf(m, "%u\n", dd->starved);
 998         return 0;
 999 }
1000
1001 static int dd_async_depth_show(void *data, struct seq_file *m)
1002 {
1003         struct request_queue *q = data;
1004         struct deadline_data *dd = q->elevator->elevator_data;
1005
1006         seq_printf(m, "%u\n", dd->async_depth);
1007         return 0;
1008 }
1009
1010 static int dd_queued_show(void *data, struct seq_file *m)
1011 {
1012         struct request_queue *q = data;
1013         struct deadline_data *dd = q->elevator->elevator_data;
1014
1015         seq_printf(m, "%u %u %u\n", dd_queued(dd, DD_RT_PRIO),
1016                    dd_queued(dd, DD_BE_PRIO),
1017                    dd_queued(dd, DD_IDLE_PRIO));
1018         return 0;
1019 }
1020
1021 /* Number of requests owned by the block driver for a given priority. */
1022 static u32 dd_owned_by_driver(struct deadline_data *dd, enum dd_prio prio)
1023 {
1024         return dd_sum(dd, dispatched, prio) + dd_sum(dd, merged, prio)
1025                 - dd_sum(dd, completed, prio);
1026 }
1027
1028 static int dd_owned_by_driver_show(void *data, struct seq_file *m)
1029 {
1030         struct request_queue *q = data;
1031         struct deadline_data *dd = q->elevator->elevator_data;
1032
1033         seq_printf(m, "%u %u %u\n", dd_owned_by_driver(dd, DD_RT_PRIO),
1034                    dd_owned_by_driver(dd, DD_BE_PRIO),
1035                    dd_owned_by_driver(dd, DD_IDLE_PRIO));
1036         return 0;
1037 }
1038
1039 #define DEADLINE_DISPATCH_ATTR(prio)                                    \
1040 static void *deadline_dispatch##prio##_start(struct seq_file *m,        \
1041                                              loff_t *pos)               \
1042         __acquires(&dd->lock)                                           \
1043 {                                                                       \
1044         struct request_queue *q = m->private;                           \
1045         struct deadline_data *dd = q->elevator->elevator_data;          \
1046         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
1047                                                                         \
1048         spin_lock(&dd->lock);                                           \
1049         return seq_list_start(&per_prio->dispatch, *pos);               \
1050 }                                                                       \
1051                                                                         \
1052 static void *deadline_dispatch##prio##_next(struct seq_file *m,         \
1053                                             void *v, loff_t *pos)       \
1054 {                                                                       \
1055         struct request_queue *q = m->private;                           \
1056         struct deadline_data *dd = q->elevator->elevator_data;          \
1057         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
1058                                                                         \
1059         return seq_list_next(v, &per_prio->dispatch, pos);              \
1060 }                                                                       \
1061                                                                         \
1062 static void deadline_dispatch##prio##_stop(struct seq_file *m, void *v) \
1063         __releases(&dd->lock)                                           \
1064 {                                                                       \
1065         struct request_queue *q = m->private;                           \
1066         struct deadline_data *dd = q->elevator->elevator_data;          \
1067                                                                         \
1068         spin_unlock(&dd->lock);                                         \
1069 }                                                                       \
1070                                                                         \
1071 static const struct seq_operations deadline_dispatch##prio##_seq_ops = { \
1072         .start  = deadline_dispatch##prio##_start,                      \
1073         .next   = deadline_dispatch##prio##_next,                       \
1074         .stop   = deadline_dispatch##prio##_stop,                       \
1075         .show   = blk_mq_debugfs_rq_show,                               \
1076 }
1077
1078 DEADLINE_DISPATCH_ATTR(0);
1079 DEADLINE_DISPATCH_ATTR(1);
1080 DEADLINE_DISPATCH_ATTR(2);
1081 #undef DEADLINE_DISPATCH_ATTR
1082
1083 #define DEADLINE_QUEUE_DDIR_ATTRS(name)                                 \
1084         {#name "_fifo_list", 0400,                                      \
1085                         .seq_ops = &deadline_##name##_fifo_seq_ops}
1086 #define DEADLINE_NEXT_RQ_ATTR(name)                                     \
1087         {#name "_next_rq", 0400, deadline_##name##_next_rq_show}
1088 static const struct blk_mq_debugfs_attr deadline_queue_debugfs_attrs[] = {
1089         DEADLINE_QUEUE_DDIR_ATTRS(read0),
1090         DEADLINE_QUEUE_DDIR_ATTRS(write0),
1091         DEADLINE_QUEUE_DDIR_ATTRS(read1),
1092         DEADLINE_QUEUE_DDIR_ATTRS(write1),
1093         DEADLINE_QUEUE_DDIR_ATTRS(read2),
1094         DEADLINE_QUEUE_DDIR_ATTRS(write2),
1095         DEADLINE_NEXT_RQ_ATTR(read0),
1096         DEADLINE_NEXT_RQ_ATTR(write0),
1097         DEADLINE_NEXT_RQ_ATTR(read1),
1098         DEADLINE_NEXT_RQ_ATTR(write1),
1099         DEADLINE_NEXT_RQ_ATTR(read2),
1100         DEADLINE_NEXT_RQ_ATTR(write2),
1101         {"batching", 0400, deadline_batching_show},
1102         {"starved", 0400, deadline_starved_show},
1103         {"async_depth", 0400, dd_async_depth_show},
1104         {"dispatch0", 0400, .seq_ops = &deadline_dispatch0_seq_ops},
1105         {"dispatch1", 0400, .seq_ops = &deadline_dispatch1_seq_ops},
1106         {"dispatch2", 0400, .seq_ops = &deadline_dispatch2_seq_ops},
1107         {"owned_by_driver", 0400, dd_owned_by_driver_show},
1108         {"queued", 0400, dd_queued_show},
1109         {},
1110 };
1111 #undef DEADLINE_QUEUE_DDIR_ATTRS
1112 #endif
1113
1114 static struct elevator_type mq_deadline = {
1115         .ops = {
1116                 .depth_updated          = dd_depth_updated,
1117                 .limit_depth            = dd_limit_depth,
1118                 .insert_requests        = dd_insert_requests,
1119                 .dispatch_request       = dd_dispatch_request,
1120                 .prepare_request        = dd_prepare_request,
1121                 .finish_request         = dd_finish_request,
1122                 .next_request           = elv_rb_latter_request,
1123                 .former_request         = elv_rb_former_request,
1124                 .bio_merge              = dd_bio_merge,
1125                 .request_merge          = dd_request_merge,
1126                 .requests_merged        = dd_merged_requests,
1127                 .request_merged         = dd_request_merged,
1128                 .has_work               = dd_has_work,
1129                 .init_sched             = dd_init_sched,
1130                 .exit_sched             = dd_exit_sched,
1131                 .init_hctx              = dd_init_hctx,
1132         },
1133
1134 #ifdef CONFIG_BLK_DEBUG_FS
1135         .queue_debugfs_attrs = deadline_queue_debugfs_attrs,
1136 #endif
1137         .elevator_attrs = deadline_attrs,
1138         .elevator_name = "mq-deadline",
1139         .elevator_alias = "deadline",
1140         .elevator_features = ELEVATOR_F_ZBD_SEQ_WRITE,
1141         .elevator_owner = THIS_MODULE,
1142 };
1143 MODULE_ALIAS("mq-deadline-iosched");
1144
1145 static int __init deadline_init(void)
1146 {
1147         int ret;
1148
1149         ret = elv_register(&mq_deadline);
1150         if (ret)
1151                 goto out;
1152         ret = dd_blkcg_init();
1153         if (ret)
1154                 goto unreg;
1155
1156 out:
1157         return ret;
1158
1159 unreg:
1160         elv_unregister(&mq_deadline);
1161         goto out;
1162 }
1163
1164 static void __exit deadline_exit(void)
1165 {
1166         dd_blkcg_exit();
1167         elv_unregister(&mq_deadline);
1168 }
1169
1170 module_init(deadline_init);
1171 module_exit(deadline_exit);
1172
1173 MODULE_AUTHOR("Jens Axboe, Damien Le Moal and Bart Van Assche");
1174 MODULE_LICENSE("GPL");
1175 MODULE_DESCRIPTION("MQ deadline IO scheduler");