block/mq-deadline.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  MQ Deadline i/o scheduler - adaptation of the legacy deadline scheduler,
   4  *  for the blk-mq scheduling framework
   5  *
   6  *  Copyright (C) 2016 Jens Axboe <axboe@kernel.dk>
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/fs.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/blk-mq.h>
  12 #include <linux/elevator.h>
  13 #include <linux/bio.h>
  14 #include <linux/module.h>
  15 #include <linux/slab.h>
  16 #include <linux/init.h>
  17 #include <linux/compiler.h>
  18 #include <linux/rbtree.h>
  19 #include <linux/sbitmap.h>
  20
  21 #include <trace/events/block.h>
  22
  23 #include "blk.h"
  24 #include "blk-mq.h"
  25 #include "blk-mq-debugfs.h"
  26 #include "blk-mq-tag.h"
  27 #include "blk-mq-sched.h"
  28
  29 /*
  30  * See Documentation/block/deadline-iosched.rst
  31  */
  32 static const int read_expire = HZ / 2;  /* max time before a read is submitted. */
  33 static const int write_expire = 5 * HZ; /* ditto for writes, these limits are SOFT! */
  34 /*
  35  * Time after which to dispatch lower priority requests even if higher
  36  * priority requests are pending.
  37  */
  38 static const int 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 /* I/O statistics per I/O priority. */
  60 struct io_stats_per_prio {
  61         local_t inserted;
  62         local_t merged;
  63         local_t dispatched;
  64         local_t completed;
  65 };
  66
  67 /* I/O statistics for all I/O priorities (enum dd_prio). */
  68 struct io_stats {
  69         struct io_stats_per_prio stats[DD_PRIO_COUNT];
  70 };
  71
  72 /*
  73  * Deadline scheduler data per I/O priority (enum dd_prio). Requests are
  74  * present on both sort_list[] and fifo_list[].
  75  */
  76 struct dd_per_prio {
  77         struct list_head dispatch;
  78         struct rb_root sort_list[DD_DIR_COUNT];
  79         struct list_head fifo_list[DD_DIR_COUNT];
  80         /* Next request in FIFO order. Read, write or both are NULL. */
  81         struct request *next_rq[DD_DIR_COUNT];
  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         struct io_stats __percpu *stats;
  97
  98         /*
  99          * settings that change how the i/o scheduler behaves
 100          */
 101         int fifo_expire[DD_DIR_COUNT];
 102         int fifo_batch;
 103         int writes_starved;
 104         int front_merges;
 105         u32 async_depth;
 106         int aging_expire;
 107
 108         spinlock_t lock;
 109         spinlock_t zone_lock;
 110 };
 111
 112 /* Count one event of type 'event_type' and with I/O priority 'prio' */
 113 #define dd_count(dd, event_type, prio) do {                             \
 114         struct io_stats *io_stats = get_cpu_ptr((dd)->stats);           \
 115                                                                         \
 116         BUILD_BUG_ON(!__same_type((dd), struct deadline_data *));       \
 117         BUILD_BUG_ON(!__same_type((prio), enum dd_prio));               \
 118         local_inc(&io_stats->stats[(prio)].event_type);                 \
 119         put_cpu_ptr(io_stats);                                          \
 120 } while (0)
 121
 122 /*
 123  * Returns the total number of dd_count(dd, event_type, prio) calls across all
 124  * CPUs. No locking or barriers since it is fine if the returned sum is slightly
 125  * outdated.
 126  */
 127 #define dd_sum(dd, event_type, prio) ({                                 \
 128         unsigned int cpu;                                               \
 129         u32 sum = 0;                                                    \
 130                                                                         \
 131         BUILD_BUG_ON(!__same_type((dd), struct deadline_data *));       \
 132         BUILD_BUG_ON(!__same_type((prio), enum dd_prio));               \
 133         for_each_present_cpu(cpu)                                       \
 134                 sum += local_read(&per_cpu_ptr((dd)->stats, cpu)->      \
 135                                   stats[(prio)].event_type);            \
 136         sum;                                                            \
 137 })
 138
 139 /* Maps an I/O priority class to a deadline scheduler priority. */
 140 static const enum dd_prio ioprio_class_to_prio[] = {
 141         [IOPRIO_CLASS_NONE]     = DD_BE_PRIO,
 142         [IOPRIO_CLASS_RT]       = DD_RT_PRIO,
 143         [IOPRIO_CLASS_BE]       = DD_BE_PRIO,
 144         [IOPRIO_CLASS_IDLE]     = DD_IDLE_PRIO,
 145 };
 146
 147 static inline struct rb_root *
 148 deadline_rb_root(struct dd_per_prio *per_prio, struct request *rq)
 149 {
 150         return &per_prio->sort_list[rq_data_dir(rq)];
 151 }
 152
 153 /*
 154  * Returns the I/O priority class (IOPRIO_CLASS_*) that has been assigned to a
 155  * request.
 156  */
 157 static u8 dd_rq_ioclass(struct request *rq)
 158 {
 159         return IOPRIO_PRIO_CLASS(req_get_ioprio(rq));
 160 }
 161
 162 /*
 163  * get the request after `rq' in sector-sorted order
 164  */
 165 static inline struct request *
 166 deadline_latter_request(struct request *rq)
 167 {
 168         struct rb_node *node = rb_next(&rq->rb_node);
 169
 170         if (node)
 171                 return rb_entry_rq(node);
 172
 173         return NULL;
 174 }
 175
 176 static void
 177 deadline_add_rq_rb(struct dd_per_prio *per_prio, struct request *rq)
 178 {
 179         struct rb_root *root = deadline_rb_root(per_prio, rq);
 180
 181         elv_rb_add(root, rq);
 182 }
 183
 184 static inline void
 185 deadline_del_rq_rb(struct dd_per_prio *per_prio, struct request *rq)
 186 {
 187         const enum dd_data_dir data_dir = rq_data_dir(rq);
 188
 189         if (per_prio->next_rq[data_dir] == rq)
 190                 per_prio->next_rq[data_dir] = deadline_latter_request(rq);
 191
 192         elv_rb_del(deadline_rb_root(per_prio, rq), rq);
 193 }
 194
 195 /*
 196  * remove rq from rbtree and fifo.
 197  */
 198 static void deadline_remove_request(struct request_queue *q,
 199                                     struct dd_per_prio *per_prio,
 200                                     struct request *rq)
 201 {
 202         list_del_init(&rq->queuelist);
 203
 204         /*
 205          * We might not be on the rbtree, if we are doing an insert merge
 206          */
 207         if (!RB_EMPTY_NODE(&rq->rb_node))
 208                 deadline_del_rq_rb(per_prio, rq);
 209
 210         elv_rqhash_del(q, rq);
 211         if (q->last_merge == rq)
 212                 q->last_merge = NULL;
 213 }
 214
 215 static void dd_request_merged(struct request_queue *q, struct request *req,
 216                               enum elv_merge type)
 217 {
 218         struct deadline_data *dd = q->elevator->elevator_data;
 219         const u8 ioprio_class = dd_rq_ioclass(req);
 220         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 221         struct dd_per_prio *per_prio = &dd->per_prio[prio];
 222
 223         /*
 224          * if the merge was a front merge, we need to reposition request
 225          */
 226         if (type == ELEVATOR_FRONT_MERGE) {
 227                 elv_rb_del(deadline_rb_root(per_prio, req), req);
 228                 deadline_add_rq_rb(per_prio, req);
 229         }
 230 }
 231
 232 /*
 233  * Callback function that is invoked after @next has been merged into @req.
 234  */
 235 static void dd_merged_requests(struct request_queue *q, struct request *req,
 236                                struct request *next)
 237 {
 238         struct deadline_data *dd = q->elevator->elevator_data;
 239         const u8 ioprio_class = dd_rq_ioclass(next);
 240         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 241
 242         dd_count(dd, merged, prio);
 243
 244         /*
 245          * if next expires before rq, assign its expire time to rq
 246          * and move into next position (next will be deleted) in fifo
 247          */
 248         if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
 249                 if (time_before((unsigned long)next->fifo_time,
 250                                 (unsigned long)req->fifo_time)) {
 251                         list_move(&req->queuelist, &next->queuelist);
 252                         req->fifo_time = next->fifo_time;
 253                 }
 254         }
 255
 256         /*
 257          * kill knowledge of next, this one is a goner
 258          */
 259         deadline_remove_request(q, &dd->per_prio[prio], next);
 260 }
 261
 262 /*
 263  * move an entry to dispatch queue
 264  */
 265 static void
 266 deadline_move_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 267                       struct request *rq)
 268 {
 269         const enum dd_data_dir data_dir = rq_data_dir(rq);
 270
 271         per_prio->next_rq[data_dir] = deadline_latter_request(rq);
 272
 273         /*
 274          * take it off the sort and fifo list
 275          */
 276         deadline_remove_request(rq->q, per_prio, rq);
 277 }
 278
 279 /* Number of requests queued for a given priority level. */
 280 static u32 dd_queued(struct deadline_data *dd, enum dd_prio prio)
 281 {
 282         return dd_sum(dd, inserted, prio) - dd_sum(dd, completed, prio);
 283 }
 284
 285 /*
 286  * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
 287  * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
 288  */
 289 static inline int deadline_check_fifo(struct dd_per_prio *per_prio,
 290                                       enum dd_data_dir data_dir)
 291 {
 292         struct request *rq = rq_entry_fifo(per_prio->fifo_list[data_dir].next);
 293
 294         /*
 295          * rq is expired!
 296          */
 297         if (time_after_eq(jiffies, (unsigned long)rq->fifo_time))
 298                 return 1;
 299
 300         return 0;
 301 }
 302
 303 /*
 304  * For the specified data direction, return the next request to
 305  * dispatch using arrival ordered lists.
 306  */
 307 static struct request *
 308 deadline_fifo_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 309                       enum dd_data_dir data_dir)
 310 {
 311         struct request *rq;
 312         unsigned long flags;
 313
 314         if (list_empty(&per_prio->fifo_list[data_dir]))
 315                 return NULL;
 316
 317         rq = rq_entry_fifo(per_prio->fifo_list[data_dir].next);
 318         if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q))
 319                 return rq;
 320
 321         /*
 322          * Look for a write request that can be dispatched, that is one with
 323          * an unlocked target zone.
 324          */
 325         spin_lock_irqsave(&dd->zone_lock, flags);
 326         list_for_each_entry(rq, &per_prio->fifo_list[DD_WRITE], queuelist) {
 327                 if (blk_req_can_dispatch_to_zone(rq))
 328                         goto out;
 329         }
 330         rq = NULL;
 331 out:
 332         spin_unlock_irqrestore(&dd->zone_lock, flags);
 333
 334         return rq;
 335 }
 336
 337 /*
 338  * For the specified data direction, return the next request to
 339  * dispatch using sector position sorted lists.
 340  */
 341 static struct request *
 342 deadline_next_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
 343                       enum dd_data_dir data_dir)
 344 {
 345         struct request *rq;
 346         unsigned long flags;
 347
 348         rq = per_prio->next_rq[data_dir];
 349         if (!rq)
 350                 return NULL;
 351
 352         if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q))
 353                 return rq;
 354
 355         /*
 356          * Look for a write request that can be dispatched, that is one with
 357          * an unlocked target zone.
 358          */
 359         spin_lock_irqsave(&dd->zone_lock, flags);
 360         while (rq) {
 361                 if (blk_req_can_dispatch_to_zone(rq))
 362                         break;
 363                 rq = deadline_latter_request(rq);
 364         }
 365         spin_unlock_irqrestore(&dd->zone_lock, flags);
 366
 367         return rq;
 368 }
 369
 370 /*
 371  * deadline_dispatch_requests selects the best request according to
 372  * read/write expire, fifo_batch, etc and with a start time <= @latest.
 373  */
 374 static struct request *__dd_dispatch_request(struct deadline_data *dd,
 375                                              struct dd_per_prio *per_prio,
 376                                              u64 latest_start_ns)
 377 {
 378         struct request *rq, *next_rq;
 379         enum dd_data_dir data_dir;
 380         enum dd_prio prio;
 381         u8 ioprio_class;
 382
 383         lockdep_assert_held(&dd->lock);
 384
 385         if (!list_empty(&per_prio->dispatch)) {
 386                 rq = list_first_entry(&per_prio->dispatch, struct request,
 387                                       queuelist);
 388                 if (rq->start_time_ns > latest_start_ns)
 389                         return NULL;
 390                 list_del_init(&rq->queuelist);
 391                 goto done;
 392         }
 393
 394         /*
 395          * batches are currently reads XOR writes
 396          */
 397         rq = deadline_next_request(dd, per_prio, dd->last_dir);
 398         if (rq && dd->batching < dd->fifo_batch)
 399                 /* we have a next request are still entitled to batch */
 400                 goto dispatch_request;
 401
 402         /*
 403          * at this point we are not running a batch. select the appropriate
 404          * data direction (read / write)
 405          */
 406
 407         if (!list_empty(&per_prio->fifo_list[DD_READ])) {
 408                 BUG_ON(RB_EMPTY_ROOT(&per_prio->sort_list[DD_READ]));
 409
 410                 if (deadline_fifo_request(dd, per_prio, DD_WRITE) &&
 411                     (dd->starved++ >= dd->writes_starved))
 412                         goto dispatch_writes;
 413
 414                 data_dir = DD_READ;
 415
 416                 goto dispatch_find_request;
 417         }
 418
 419         /*
 420          * there are either no reads or writes have been starved
 421          */
 422
 423         if (!list_empty(&per_prio->fifo_list[DD_WRITE])) {
 424 dispatch_writes:
 425                 BUG_ON(RB_EMPTY_ROOT(&per_prio->sort_list[DD_WRITE]));
 426
 427                 dd->starved = 0;
 428
 429                 data_dir = DD_WRITE;
 430
 431                 goto dispatch_find_request;
 432         }
 433
 434         return NULL;
 435
 436 dispatch_find_request:
 437         /*
 438          * we are not running a batch, find best request for selected data_dir
 439          */
 440         next_rq = deadline_next_request(dd, per_prio, data_dir);
 441         if (deadline_check_fifo(per_prio, data_dir) || !next_rq) {
 442                 /*
 443                  * A deadline has expired, the last request was in the other
 444                  * direction, or we have run out of higher-sectored requests.
 445                  * Start again from the request with the earliest expiry time.
 446                  */
 447                 rq = deadline_fifo_request(dd, per_prio, data_dir);
 448         } else {
 449                 /*
 450                  * The last req was the same dir and we have a next request in
 451                  * sort order. No expired requests so continue on from here.
 452                  */
 453                 rq = next_rq;
 454         }
 455
 456         /*
 457          * For a zoned block device, if we only have writes queued and none of
 458          * them can be dispatched, rq will be NULL.
 459          */
 460         if (!rq)
 461                 return NULL;
 462
 463         dd->last_dir = data_dir;
 464         dd->batching = 0;
 465
 466 dispatch_request:
 467         if (rq->start_time_ns > latest_start_ns)
 468                 return NULL;
 469         /*
 470          * rq is the selected appropriate request.
 471          */
 472         dd->batching++;
 473         deadline_move_request(dd, per_prio, rq);
 474 done:
 475         ioprio_class = dd_rq_ioclass(rq);
 476         prio = ioprio_class_to_prio[ioprio_class];
 477         dd_count(dd, dispatched, prio);
 478         /*
 479          * If the request needs its target zone locked, do it.
 480          */
 481         blk_req_zone_write_lock(rq);
 482         rq->rq_flags |= RQF_STARTED;
 483         return rq;
 484 }
 485
 486 /*
 487  * Called from blk_mq_run_hw_queue() -> __blk_mq_sched_dispatch_requests().
 488  *
 489  * One confusing aspect here is that we get called for a specific
 490  * hardware queue, but we may return a request that is for a
 491  * different hardware queue. This is because mq-deadline has shared
 492  * state for all hardware queues, in terms of sorting, FIFOs, etc.
 493  */
 494 static struct request *dd_dispatch_request(struct blk_mq_hw_ctx *hctx)
 495 {
 496         struct deadline_data *dd = hctx->queue->elevator->elevator_data;
 497         const u64 now_ns = ktime_get_ns();
 498         struct request *rq = NULL;
 499         enum dd_prio prio;
 500
 501         spin_lock(&dd->lock);
 502         /*
 503          * Start with dispatching requests whose deadline expired more than
 504          * aging_expire jiffies ago.
 505          */
 506         for (prio = DD_BE_PRIO; prio <= DD_PRIO_MAX; prio++) {
 507                 rq = __dd_dispatch_request(dd, &dd->per_prio[prio], now_ns -
 508                                            jiffies_to_nsecs(dd->aging_expire));
 509                 if (rq)
 510                         goto unlock;
 511         }
 512         /*
 513          * Next, dispatch requests in priority order. Ignore lower priority
 514          * requests if any higher priority requests are pending.
 515          */
 516         for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 517                 rq = __dd_dispatch_request(dd, &dd->per_prio[prio], now_ns);
 518                 if (rq || dd_queued(dd, prio))
 519                         break;
 520         }
 521
 522 unlock:
 523         spin_unlock(&dd->lock);
 524
 525         return rq;
 526 }
 527
 528 /*
 529  * Called by __blk_mq_alloc_request(). The shallow_depth value set by this
 530  * function is used by __blk_mq_get_tag().
 531  */
 532 static void dd_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
 533 {
 534         struct deadline_data *dd = data->q->elevator->elevator_data;
 535
 536         /* Do not throttle synchronous reads. */
 537         if (op_is_sync(op) && !op_is_write(op))
 538                 return;
 539
 540         /*
 541          * Throttle asynchronous requests and writes such that these requests
 542          * do not block the allocation of synchronous requests.
 543          */
 544         data->shallow_depth = dd->async_depth;
 545 }
 546
 547 /* Called by blk_mq_update_nr_requests(). */
 548 static void dd_depth_updated(struct blk_mq_hw_ctx *hctx)
 549 {
 550         struct request_queue *q = hctx->queue;
 551         struct deadline_data *dd = q->elevator->elevator_data;
 552         struct blk_mq_tags *tags = hctx->sched_tags;
 553
 554         dd->async_depth = max(1UL, 3 * q->nr_requests / 4);
 555
 556         sbitmap_queue_min_shallow_depth(tags->bitmap_tags, dd->async_depth);
 557 }
 558
 559 /* Called by blk_mq_init_hctx() and blk_mq_init_sched(). */
 560 static int dd_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 561 {
 562         dd_depth_updated(hctx);
 563         return 0;
 564 }
 565
 566 static void dd_exit_sched(struct elevator_queue *e)
 567 {
 568         struct deadline_data *dd = e->elevator_data;
 569         enum dd_prio prio;
 570
 571         for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 572                 struct dd_per_prio *per_prio = &dd->per_prio[prio];
 573
 574                 WARN_ON_ONCE(!list_empty(&per_prio->fifo_list[DD_READ]));
 575                 WARN_ON_ONCE(!list_empty(&per_prio->fifo_list[DD_WRITE]));
 576         }
 577
 578         free_percpu(dd->stats);
 579
 580         kfree(dd);
 581 }
 582
 583 /*
 584  * initialize elevator private data (deadline_data).
 585  */
 586 static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
 587 {
 588         struct deadline_data *dd;
 589         struct elevator_queue *eq;
 590         enum dd_prio prio;
 591         int ret = -ENOMEM;
 592
 593         eq = elevator_alloc(q, e);
 594         if (!eq)
 595                 return ret;
 596
 597         dd = kzalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
 598         if (!dd)
 599                 goto put_eq;
 600
 601         eq->elevator_data = dd;
 602
 603         dd->stats = alloc_percpu_gfp(typeof(*dd->stats),
 604                                      GFP_KERNEL | __GFP_ZERO);
 605         if (!dd->stats)
 606                 goto free_dd;
 607
 608         for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 609                 struct dd_per_prio *per_prio = &dd->per_prio[prio];
 610
 611                 INIT_LIST_HEAD(&per_prio->dispatch);
 612                 INIT_LIST_HEAD(&per_prio->fifo_list[DD_READ]);
 613                 INIT_LIST_HEAD(&per_prio->fifo_list[DD_WRITE]);
 614                 per_prio->sort_list[DD_READ] = RB_ROOT;
 615                 per_prio->sort_list[DD_WRITE] = RB_ROOT;
 616         }
 617         dd->fifo_expire[DD_READ] = read_expire;
 618         dd->fifo_expire[DD_WRITE] = write_expire;
 619         dd->writes_starved = writes_starved;
 620         dd->front_merges = 1;
 621         dd->last_dir = DD_WRITE;
 622         dd->fifo_batch = fifo_batch;
 623         dd->aging_expire = aging_expire;
 624         spin_lock_init(&dd->lock);
 625         spin_lock_init(&dd->zone_lock);
 626
 627         q->elevator = eq;
 628         return 0;
 629
 630 free_dd:
 631         kfree(dd);
 632
 633 put_eq:
 634         kobject_put(&eq->kobj);
 635         return ret;
 636 }
 637
 638 /*
 639  * Try to merge @bio into an existing request. If @bio has been merged into
 640  * an existing request, store the pointer to that request into *@rq.
 641  */
 642 static int dd_request_merge(struct request_queue *q, struct request **rq,
 643                             struct bio *bio)
 644 {
 645         struct deadline_data *dd = q->elevator->elevator_data;
 646         const u8 ioprio_class = IOPRIO_PRIO_CLASS(bio->bi_ioprio);
 647         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 648         struct dd_per_prio *per_prio = &dd->per_prio[prio];
 649         sector_t sector = bio_end_sector(bio);
 650         struct request *__rq;
 651
 652         if (!dd->front_merges)
 653                 return ELEVATOR_NO_MERGE;
 654
 655         __rq = elv_rb_find(&per_prio->sort_list[bio_data_dir(bio)], sector);
 656         if (__rq) {
 657                 BUG_ON(sector != blk_rq_pos(__rq));
 658
 659                 if (elv_bio_merge_ok(__rq, bio)) {
 660                         *rq = __rq;
 661                         return ELEVATOR_FRONT_MERGE;
 662                 }
 663         }
 664
 665         return ELEVATOR_NO_MERGE;
 666 }
 667
 668 /*
 669  * Attempt to merge a bio into an existing request. This function is called
 670  * before @bio is associated with a request.
 671  */
 672 static bool dd_bio_merge(struct request_queue *q, struct bio *bio,
 673                 unsigned int nr_segs)
 674 {
 675         struct deadline_data *dd = q->elevator->elevator_data;
 676         struct request *free = NULL;
 677         bool ret;
 678
 679         spin_lock(&dd->lock);
 680         ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
 681         spin_unlock(&dd->lock);
 682
 683         if (free)
 684                 blk_mq_free_request(free);
 685
 686         return ret;
 687 }
 688
 689 /*
 690  * add rq to rbtree and fifo
 691  */
 692 static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 693                               bool at_head)
 694 {
 695         struct request_queue *q = hctx->queue;
 696         struct deadline_data *dd = q->elevator->elevator_data;
 697         const enum dd_data_dir data_dir = rq_data_dir(rq);
 698         u16 ioprio = req_get_ioprio(rq);
 699         u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
 700         struct dd_per_prio *per_prio;
 701         enum dd_prio prio;
 702         LIST_HEAD(free);
 703
 704         lockdep_assert_held(&dd->lock);
 705
 706         /*
 707          * This may be a requeue of a write request that has locked its
 708          * target zone. If it is the case, this releases the zone lock.
 709          */
 710         blk_req_zone_write_unlock(rq);
 711
 712         prio = ioprio_class_to_prio[ioprio_class];
 713         dd_count(dd, inserted, prio);
 714         rq->elv.priv[0] = (void *)(uintptr_t)1;
 715
 716         if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
 717                 blk_mq_free_requests(&free);
 718                 return;
 719         }
 720
 721         trace_block_rq_insert(rq);
 722
 723         per_prio = &dd->per_prio[prio];
 724         if (at_head) {
 725                 list_add(&rq->queuelist, &per_prio->dispatch);
 726         } else {
 727                 deadline_add_rq_rb(per_prio, rq);
 728
 729                 if (rq_mergeable(rq)) {
 730                         elv_rqhash_add(q, rq);
 731                         if (!q->last_merge)
 732                                 q->last_merge = rq;
 733                 }
 734
 735                 /*
 736                  * set expire time and add to fifo list
 737                  */
 738                 rq->fifo_time = jiffies + dd->fifo_expire[data_dir];
 739                 list_add_tail(&rq->queuelist, &per_prio->fifo_list[data_dir]);
 740         }
 741 }
 742
 743 /*
 744  * Called from blk_mq_sched_insert_request() or blk_mq_sched_insert_requests().
 745  */
 746 static void dd_insert_requests(struct blk_mq_hw_ctx *hctx,
 747                                struct list_head *list, bool at_head)
 748 {
 749         struct request_queue *q = hctx->queue;
 750         struct deadline_data *dd = q->elevator->elevator_data;
 751
 752         spin_lock(&dd->lock);
 753         while (!list_empty(list)) {
 754                 struct request *rq;
 755
 756                 rq = list_first_entry(list, struct request, queuelist);
 757                 list_del_init(&rq->queuelist);
 758                 dd_insert_request(hctx, rq, at_head);
 759         }
 760         spin_unlock(&dd->lock);
 761 }
 762
 763 /* Callback from inside blk_mq_rq_ctx_init(). */
 764 static void dd_prepare_request(struct request *rq)
 765 {
 766         rq->elv.priv[0] = NULL;
 767 }
 768
 769 /*
 770  * Callback from inside blk_mq_free_request().
 771  *
 772  * For zoned block devices, write unlock the target zone of
 773  * completed write requests. Do this while holding the zone lock
 774  * spinlock so that the zone is never unlocked while deadline_fifo_request()
 775  * or deadline_next_request() are executing. This function is called for
 776  * all requests, whether or not these requests complete successfully.
 777  *
 778  * For a zoned block device, __dd_dispatch_request() may have stopped
 779  * dispatching requests if all the queued requests are write requests directed
 780  * at zones that are already locked due to on-going write requests. To ensure
 781  * write request dispatch progress in this case, mark the queue as needing a
 782  * restart to ensure that the queue is run again after completion of the
 783  * request and zones being unlocked.
 784  */
 785 static void dd_finish_request(struct request *rq)
 786 {
 787         struct request_queue *q = rq->q;
 788         struct deadline_data *dd = q->elevator->elevator_data;
 789         const u8 ioprio_class = dd_rq_ioclass(rq);
 790         const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 791         struct dd_per_prio *per_prio = &dd->per_prio[prio];
 792
 793         /*
 794          * The block layer core may call dd_finish_request() without having
 795          * called dd_insert_requests(). Hence only update statistics for
 796          * requests for which dd_insert_requests() has been called. See also
 797          * blk_mq_request_bypass_insert().
 798          */
 799         if (rq->elv.priv[0])
 800                 dd_count(dd, completed, prio);
 801
 802         if (blk_queue_is_zoned(q)) {
 803                 unsigned long flags;
 804
 805                 spin_lock_irqsave(&dd->zone_lock, flags);
 806                 blk_req_zone_write_unlock(rq);
 807                 if (!list_empty(&per_prio->fifo_list[DD_WRITE]))
 808                         blk_mq_sched_mark_restart_hctx(rq->mq_hctx);
 809                 spin_unlock_irqrestore(&dd->zone_lock, flags);
 810         }
 811 }
 812
 813 static bool dd_has_work_for_prio(struct dd_per_prio *per_prio)
 814 {
 815         return !list_empty_careful(&per_prio->dispatch) ||
 816                 !list_empty_careful(&per_prio->fifo_list[DD_READ]) ||
 817                 !list_empty_careful(&per_prio->fifo_list[DD_WRITE]);
 818 }
 819
 820 static bool dd_has_work(struct blk_mq_hw_ctx *hctx)
 821 {
 822         struct deadline_data *dd = hctx->queue->elevator->elevator_data;
 823         enum dd_prio prio;
 824
 825         for (prio = 0; prio <= DD_PRIO_MAX; prio++)
 826                 if (dd_has_work_for_prio(&dd->per_prio[prio]))
 827                         return true;
 828
 829         return false;
 830 }
 831
 832 /*
 833  * sysfs parts below
 834  */
 835 #define SHOW_INT(__FUNC, __VAR)                                         \
 836 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
 837 {                                                                       \
 838         struct deadline_data *dd = e->elevator_data;                    \
 839                                                                         \
 840         return sysfs_emit(page, "%d\n", __VAR);                         \
 841 }
 842 #define SHOW_JIFFIES(__FUNC, __VAR) SHOW_INT(__FUNC, jiffies_to_msecs(__VAR))
 843 SHOW_JIFFIES(deadline_read_expire_show, dd->fifo_expire[DD_READ]);
 844 SHOW_JIFFIES(deadline_write_expire_show, dd->fifo_expire[DD_WRITE]);
 845 SHOW_JIFFIES(deadline_aging_expire_show, dd->aging_expire);
 846 SHOW_INT(deadline_writes_starved_show, dd->writes_starved);
 847 SHOW_INT(deadline_front_merges_show, dd->front_merges);
 848 SHOW_INT(deadline_async_depth_show, dd->front_merges);
 849 SHOW_INT(deadline_fifo_batch_show, dd->fifo_batch);
 850 #undef SHOW_INT
 851 #undef SHOW_JIFFIES
 852
 853 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
 854 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
 855 {                                                                       \
 856         struct deadline_data *dd = e->elevator_data;                    \
 857         int __data, __ret;                                              \
 858                                                                         \
 859         __ret = kstrtoint(page, 0, &__data);                            \
 860         if (__ret < 0)                                                  \
 861                 return __ret;                                           \
 862         if (__data < (MIN))                                             \
 863                 __data = (MIN);                                         \
 864         else if (__data > (MAX))                                        \
 865                 __data = (MAX);                                         \
 866         *(__PTR) = __CONV(__data);                                      \
 867         return count;                                                   \
 868 }
 869 #define STORE_INT(__FUNC, __PTR, MIN, MAX)                              \
 870         STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, )
 871 #define STORE_JIFFIES(__FUNC, __PTR, MIN, MAX)                          \
 872         STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, msecs_to_jiffies)
 873 STORE_JIFFIES(deadline_read_expire_store, &dd->fifo_expire[DD_READ], 0, INT_MAX);
 874 STORE_JIFFIES(deadline_write_expire_store, &dd->fifo_expire[DD_WRITE], 0, INT_MAX);
 875 STORE_JIFFIES(deadline_aging_expire_store, &dd->aging_expire, 0, INT_MAX);
 876 STORE_INT(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX);
 877 STORE_INT(deadline_front_merges_store, &dd->front_merges, 0, 1);
 878 STORE_INT(deadline_async_depth_store, &dd->front_merges, 1, INT_MAX);
 879 STORE_INT(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX);
 880 #undef STORE_FUNCTION
 881 #undef STORE_INT
 882 #undef STORE_JIFFIES
 883
 884 #define DD_ATTR(name) \
 885         __ATTR(name, 0644, deadline_##name##_show, deadline_##name##_store)
 886
 887 static struct elv_fs_entry deadline_attrs[] = {
 888         DD_ATTR(read_expire),
 889         DD_ATTR(write_expire),
 890         DD_ATTR(writes_starved),
 891         DD_ATTR(front_merges),
 892         DD_ATTR(async_depth),
 893         DD_ATTR(fifo_batch),
 894         DD_ATTR(aging_expire),
 895         __ATTR_NULL
 896 };
 897
 898 #ifdef CONFIG_BLK_DEBUG_FS
 899 #define DEADLINE_DEBUGFS_DDIR_ATTRS(prio, data_dir, name)               \
 900 static void *deadline_##name##_fifo_start(struct seq_file *m,           \
 901                                           loff_t *pos)                  \
 902         __acquires(&dd->lock)                                           \
 903 {                                                                       \
 904         struct request_queue *q = m->private;                           \
 905         struct deadline_data *dd = q->elevator->elevator_data;          \
 906         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
 907                                                                         \
 908         spin_lock(&dd->lock);                                           \
 909         return seq_list_start(&per_prio->fifo_list[data_dir], *pos);    \
 910 }                                                                       \
 911                                                                         \
 912 static void *deadline_##name##_fifo_next(struct seq_file *m, void *v,   \
 913                                          loff_t *pos)                   \
 914 {                                                                       \
 915         struct request_queue *q = m->private;                           \
 916         struct deadline_data *dd = q->elevator->elevator_data;          \
 917         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
 918                                                                         \
 919         return seq_list_next(v, &per_prio->fifo_list[data_dir], pos);   \
 920 }                                                                       \
 921                                                                         \
 922 static void deadline_##name##_fifo_stop(struct seq_file *m, void *v)    \
 923         __releases(&dd->lock)                                           \
 924 {                                                                       \
 925         struct request_queue *q = m->private;                           \
 926         struct deadline_data *dd = q->elevator->elevator_data;          \
 927                                                                         \
 928         spin_unlock(&dd->lock);                                         \
 929 }                                                                       \
 930                                                                         \
 931 static const struct seq_operations deadline_##name##_fifo_seq_ops = {   \
 932         .start  = deadline_##name##_fifo_start,                         \
 933         .next   = deadline_##name##_fifo_next,                          \
 934         .stop   = deadline_##name##_fifo_stop,                          \
 935         .show   = blk_mq_debugfs_rq_show,                               \
 936 };                                                                      \
 937                                                                         \
 938 static int deadline_##name##_next_rq_show(void *data,                   \
 939                                           struct seq_file *m)           \
 940 {                                                                       \
 941         struct request_queue *q = data;                                 \
 942         struct deadline_data *dd = q->elevator->elevator_data;          \
 943         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
 944         struct request *rq = per_prio->next_rq[data_dir];               \
 945                                                                         \
 946         if (rq)                                                         \
 947                 __blk_mq_debugfs_rq_show(m, rq);                        \
 948         return 0;                                                       \
 949 }
 950
 951 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_RT_PRIO, DD_READ, read0);
 952 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_RT_PRIO, DD_WRITE, write0);
 953 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_BE_PRIO, DD_READ, read1);
 954 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_BE_PRIO, DD_WRITE, write1);
 955 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_IDLE_PRIO, DD_READ, read2);
 956 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_IDLE_PRIO, DD_WRITE, write2);
 957 #undef DEADLINE_DEBUGFS_DDIR_ATTRS
 958
 959 static int deadline_batching_show(void *data, struct seq_file *m)
 960 {
 961         struct request_queue *q = data;
 962         struct deadline_data *dd = q->elevator->elevator_data;
 963
 964         seq_printf(m, "%u\n", dd->batching);
 965         return 0;
 966 }
 967
 968 static int deadline_starved_show(void *data, struct seq_file *m)
 969 {
 970         struct request_queue *q = data;
 971         struct deadline_data *dd = q->elevator->elevator_data;
 972
 973         seq_printf(m, "%u\n", dd->starved);
 974         return 0;
 975 }
 976
 977 static int dd_async_depth_show(void *data, struct seq_file *m)
 978 {
 979         struct request_queue *q = data;
 980         struct deadline_data *dd = q->elevator->elevator_data;
 981
 982         seq_printf(m, "%u\n", dd->async_depth);
 983         return 0;
 984 }
 985
 986 static int dd_queued_show(void *data, struct seq_file *m)
 987 {
 988         struct request_queue *q = data;
 989         struct deadline_data *dd = q->elevator->elevator_data;
 990
 991         seq_printf(m, "%u %u %u\n", dd_queued(dd, DD_RT_PRIO),
 992                    dd_queued(dd, DD_BE_PRIO),
 993                    dd_queued(dd, DD_IDLE_PRIO));
 994         return 0;
 995 }
 996
 997 /* Number of requests owned by the block driver for a given priority. */
 998 static u32 dd_owned_by_driver(struct deadline_data *dd, enum dd_prio prio)
 999 {
1000         return dd_sum(dd, dispatched, prio) + dd_sum(dd, merged, prio)
1001                 - dd_sum(dd, completed, prio);
1002 }
1003
1004 static int dd_owned_by_driver_show(void *data, struct seq_file *m)
1005 {
1006         struct request_queue *q = data;
1007         struct deadline_data *dd = q->elevator->elevator_data;
1008
1009         seq_printf(m, "%u %u %u\n", dd_owned_by_driver(dd, DD_RT_PRIO),
1010                    dd_owned_by_driver(dd, DD_BE_PRIO),
1011                    dd_owned_by_driver(dd, DD_IDLE_PRIO));
1012         return 0;
1013 }
1014
1015 #define DEADLINE_DISPATCH_ATTR(prio)                                    \
1016 static void *deadline_dispatch##prio##_start(struct seq_file *m,        \
1017                                              loff_t *pos)               \
1018         __acquires(&dd->lock)                                           \
1019 {                                                                       \
1020         struct request_queue *q = m->private;                           \
1021         struct deadline_data *dd = q->elevator->elevator_data;          \
1022         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
1023                                                                         \
1024         spin_lock(&dd->lock);                                           \
1025         return seq_list_start(&per_prio->dispatch, *pos);               \
1026 }                                                                       \
1027                                                                         \
1028 static void *deadline_dispatch##prio##_next(struct seq_file *m,         \
1029                                             void *v, loff_t *pos)       \
1030 {                                                                       \
1031         struct request_queue *q = m->private;                           \
1032         struct deadline_data *dd = q->elevator->elevator_data;          \
1033         struct dd_per_prio *per_prio = &dd->per_prio[prio];             \
1034                                                                         \
1035         return seq_list_next(v, &per_prio->dispatch, pos);              \
1036 }                                                                       \
1037                                                                         \
1038 static void deadline_dispatch##prio##_stop(struct seq_file *m, void *v) \
1039         __releases(&dd->lock)                                           \
1040 {                                                                       \
1041         struct request_queue *q = m->private;                           \
1042         struct deadline_data *dd = q->elevator->elevator_data;          \
1043                                                                         \
1044         spin_unlock(&dd->lock);                                         \
1045 }                                                                       \
1046                                                                         \
1047 static const struct seq_operations deadline_dispatch##prio##_seq_ops = { \
1048         .start  = deadline_dispatch##prio##_start,                      \
1049         .next   = deadline_dispatch##prio##_next,                       \
1050         .stop   = deadline_dispatch##prio##_stop,                       \
1051         .show   = blk_mq_debugfs_rq_show,                               \
1052 }
1053
1054 DEADLINE_DISPATCH_ATTR(0);
1055 DEADLINE_DISPATCH_ATTR(1);
1056 DEADLINE_DISPATCH_ATTR(2);
1057 #undef DEADLINE_DISPATCH_ATTR
1058
1059 #define DEADLINE_QUEUE_DDIR_ATTRS(name)                                 \
1060         {#name "_fifo_list", 0400,                                      \
1061                         .seq_ops = &deadline_##name##_fifo_seq_ops}
1062 #define DEADLINE_NEXT_RQ_ATTR(name)                                     \
1063         {#name "_next_rq", 0400, deadline_##name##_next_rq_show}
1064 static const struct blk_mq_debugfs_attr deadline_queue_debugfs_attrs[] = {
1065         DEADLINE_QUEUE_DDIR_ATTRS(read0),
1066         DEADLINE_QUEUE_DDIR_ATTRS(write0),
1067         DEADLINE_QUEUE_DDIR_ATTRS(read1),
1068         DEADLINE_QUEUE_DDIR_ATTRS(write1),
1069         DEADLINE_QUEUE_DDIR_ATTRS(read2),
1070         DEADLINE_QUEUE_DDIR_ATTRS(write2),
1071         DEADLINE_NEXT_RQ_ATTR(read0),
1072         DEADLINE_NEXT_RQ_ATTR(write0),
1073         DEADLINE_NEXT_RQ_ATTR(read1),
1074         DEADLINE_NEXT_RQ_ATTR(write1),
1075         DEADLINE_NEXT_RQ_ATTR(read2),
1076         DEADLINE_NEXT_RQ_ATTR(write2),
1077         {"batching", 0400, deadline_batching_show},
1078         {"starved", 0400, deadline_starved_show},
1079         {"async_depth", 0400, dd_async_depth_show},
1080         {"dispatch0", 0400, .seq_ops = &deadline_dispatch0_seq_ops},
1081         {"dispatch1", 0400, .seq_ops = &deadline_dispatch1_seq_ops},
1082         {"dispatch2", 0400, .seq_ops = &deadline_dispatch2_seq_ops},
1083         {"owned_by_driver", 0400, dd_owned_by_driver_show},
1084         {"queued", 0400, dd_queued_show},
1085         {},
1086 };
1087 #undef DEADLINE_QUEUE_DDIR_ATTRS
1088 #endif
1089
1090 static struct elevator_type mq_deadline = {
1091         .ops = {
1092                 .depth_updated          = dd_depth_updated,
1093                 .limit_depth            = dd_limit_depth,
1094                 .insert_requests        = dd_insert_requests,
1095                 .dispatch_request       = dd_dispatch_request,
1096                 .prepare_request        = dd_prepare_request,
1097                 .finish_request         = dd_finish_request,
1098                 .next_request           = elv_rb_latter_request,
1099                 .former_request         = elv_rb_former_request,
1100                 .bio_merge              = dd_bio_merge,
1101                 .request_merge          = dd_request_merge,
1102                 .requests_merged        = dd_merged_requests,
1103                 .request_merged         = dd_request_merged,
1104                 .has_work               = dd_has_work,
1105                 .init_sched             = dd_init_sched,
1106                 .exit_sched             = dd_exit_sched,
1107                 .init_hctx              = dd_init_hctx,
1108         },
1109
1110 #ifdef CONFIG_BLK_DEBUG_FS
1111         .queue_debugfs_attrs = deadline_queue_debugfs_attrs,
1112 #endif
1113         .elevator_attrs = deadline_attrs,
1114         .elevator_name = "mq-deadline",
1115         .elevator_alias = "deadline",
1116         .elevator_features = ELEVATOR_F_ZBD_SEQ_WRITE,
1117         .elevator_owner = THIS_MODULE,
1118 };
1119 MODULE_ALIAS("mq-deadline-iosched");
1120
1121 static int __init deadline_init(void)
1122 {
1123         return elv_register(&mq_deadline);
1124 }
1125
1126 static void __exit deadline_exit(void)
1127 {
1128         elv_unregister(&mq_deadline);
1129 }
1130
1131 module_init(deadline_init);
1132 module_exit(deadline_exit);
1133
1134 MODULE_AUTHOR("Jens Axboe, Damien Le Moal and Bart Van Assche");
1135 MODULE_LICENSE("GPL");
1136 MODULE_DESCRIPTION("MQ deadline IO scheduler");