drivers/md/dm-rq.c

   1 /*
   2  * Copyright (C) 2016 Red Hat, Inc. All rights reserved.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-core.h"
   8 #include "dm-rq.h"
   9
  10 #include <linux/elevator.h> /* for rq_end_sector() */
  11 #include <linux/blk-mq.h>
  12
  13 #define DM_MSG_PREFIX "core-rq"
  14
  15 #define DM_MQ_NR_HW_QUEUES 1
  16 #define DM_MQ_QUEUE_DEPTH 2048
  17 static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
  18 static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
  19
  20 /*
  21  * Request-based DM's mempools' reserved IOs set by the user.
  22  */
  23 #define RESERVED_REQUEST_BASED_IOS      256
  24 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
  25
  26 static bool use_blk_mq = IS_ENABLED(CONFIG_DM_MQ_DEFAULT);
  27
  28 bool dm_use_blk_mq_default(void)
  29 {
  30         return use_blk_mq;
  31 }
  32
  33 bool dm_use_blk_mq(struct mapped_device *md)
  34 {
  35         return md->use_blk_mq;
  36 }
  37 EXPORT_SYMBOL_GPL(dm_use_blk_mq);
  38
  39 unsigned dm_get_reserved_rq_based_ios(void)
  40 {
  41         return __dm_get_module_param(&reserved_rq_based_ios,
  42                                      RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
  43 }
  44 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
  45
  46 static unsigned dm_get_blk_mq_nr_hw_queues(void)
  47 {
  48         return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
  49 }
  50
  51 static unsigned dm_get_blk_mq_queue_depth(void)
  52 {
  53         return __dm_get_module_param(&dm_mq_queue_depth,
  54                                      DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
  55 }
  56
  57 int dm_request_based(struct mapped_device *md)
  58 {
  59         return blk_queue_stackable(md->queue);
  60 }
  61
  62 static void dm_old_start_queue(struct request_queue *q)
  63 {
  64         unsigned long flags;
  65
  66         spin_lock_irqsave(q->queue_lock, flags);
  67         if (blk_queue_stopped(q))
  68                 blk_start_queue(q);
  69         spin_unlock_irqrestore(q->queue_lock, flags);
  70 }
  71
  72 static void dm_mq_start_queue(struct request_queue *q)
  73 {
  74         blk_mq_unquiesce_queue(q);
  75         blk_mq_kick_requeue_list(q);
  76 }
  77
  78 void dm_start_queue(struct request_queue *q)
  79 {
  80         if (!q->mq_ops)
  81                 dm_old_start_queue(q);
  82         else
  83                 dm_mq_start_queue(q);
  84 }
  85
  86 static void dm_old_stop_queue(struct request_queue *q)
  87 {
  88         unsigned long flags;
  89
  90         spin_lock_irqsave(q->queue_lock, flags);
  91         if (!blk_queue_stopped(q))
  92                 blk_stop_queue(q);
  93         spin_unlock_irqrestore(q->queue_lock, flags);
  94 }
  95
  96 static void dm_mq_stop_queue(struct request_queue *q)
  97 {
  98         blk_mq_quiesce_queue(q);
  99 }
 100
 101 void dm_stop_queue(struct request_queue *q)
 102 {
 103         if (!q->mq_ops)
 104                 dm_old_stop_queue(q);
 105         else
 106                 dm_mq_stop_queue(q);
 107 }
 108
 109 /*
 110  * Partial completion handling for request-based dm
 111  */
 112 static void end_clone_bio(struct bio *clone)
 113 {
 114         struct dm_rq_clone_bio_info *info =
 115                 container_of(clone, struct dm_rq_clone_bio_info, clone);
 116         struct dm_rq_target_io *tio = info->tio;
 117         unsigned int nr_bytes = info->orig->bi_iter.bi_size;
 118         blk_status_t error = clone->bi_status;
 119         bool is_last = !clone->bi_next;
 120
 121         bio_put(clone);
 122
 123         if (tio->error)
 124                 /*
 125                  * An error has already been detected on the request.
 126                  * Once error occurred, just let clone->end_io() handle
 127                  * the remainder.
 128                  */
 129                 return;
 130         else if (error) {
 131                 /*
 132                  * Don't notice the error to the upper layer yet.
 133                  * The error handling decision is made by the target driver,
 134                  * when the request is completed.
 135                  */
 136                 tio->error = error;
 137                 goto exit;
 138         }
 139
 140         /*
 141          * I/O for the bio successfully completed.
 142          * Notice the data completion to the upper layer.
 143          */
 144         tio->completed += nr_bytes;
 145
 146         /*
 147          * Update the original request.
 148          * Do not use blk_end_request() here, because it may complete
 149          * the original request before the clone, and break the ordering.
 150          */
 151         if (is_last)
 152  exit:
 153                 blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
 154 }
 155
 156 static struct dm_rq_target_io *tio_from_request(struct request *rq)
 157 {
 158         return blk_mq_rq_to_pdu(rq);
 159 }
 160
 161 static void rq_end_stats(struct mapped_device *md, struct request *orig)
 162 {
 163         if (unlikely(dm_stats_used(&md->stats))) {
 164                 struct dm_rq_target_io *tio = tio_from_request(orig);
 165                 tio->duration_jiffies = jiffies - tio->duration_jiffies;
 166                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 167                                     blk_rq_pos(orig), tio->n_sectors, true,
 168                                     tio->duration_jiffies, &tio->stats_aux);
 169         }
 170 }
 171
 172 /*
 173  * Don't touch any member of the md after calling this function because
 174  * the md may be freed in dm_put() at the end of this function.
 175  * Or do dm_get() before calling this function and dm_put() later.
 176  */
 177 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
 178 {
 179         struct request_queue *q = md->queue;
 180         unsigned long flags;
 181
 182         atomic_dec(&md->pending[rw]);
 183
 184         /* nudge anyone waiting on suspend queue */
 185         if (!md_in_flight(md))
 186                 wake_up(&md->wait);
 187
 188         /*
 189          * Run this off this callpath, as drivers could invoke end_io while
 190          * inside their request_fn (and holding the queue lock). Calling
 191          * back into ->request_fn() could deadlock attempting to grab the
 192          * queue lock again.
 193          */
 194         if (!q->mq_ops && run_queue) {
 195                 spin_lock_irqsave(q->queue_lock, flags);
 196                 blk_run_queue_async(q);
 197                 spin_unlock_irqrestore(q->queue_lock, flags);
 198         }
 199
 200         /*
 201          * dm_put() must be at the end of this function. See the comment above
 202          */
 203         dm_put(md);
 204 }
 205
 206 /*
 207  * Complete the clone and the original request.
 208  * Must be called without clone's queue lock held,
 209  * see end_clone_request() for more details.
 210  */
 211 static void dm_end_request(struct request *clone, blk_status_t error)
 212 {
 213         int rw = rq_data_dir(clone);
 214         struct dm_rq_target_io *tio = clone->end_io_data;
 215         struct mapped_device *md = tio->md;
 216         struct request *rq = tio->orig;
 217
 218         blk_rq_unprep_clone(clone);
 219         tio->ti->type->release_clone_rq(clone);
 220
 221         rq_end_stats(md, rq);
 222         if (!rq->q->mq_ops)
 223                 blk_end_request_all(rq, error);
 224         else
 225                 blk_mq_end_request(rq, error);
 226         rq_completed(md, rw, true);
 227 }
 228
 229 /*
 230  * Requeue the original request of a clone.
 231  */
 232 static void dm_old_requeue_request(struct request *rq, unsigned long delay_ms)
 233 {
 234         struct request_queue *q = rq->q;
 235         unsigned long flags;
 236
 237         spin_lock_irqsave(q->queue_lock, flags);
 238         blk_requeue_request(q, rq);
 239         blk_delay_queue(q, delay_ms);
 240         spin_unlock_irqrestore(q->queue_lock, flags);
 241 }
 242
 243 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
 244 {
 245         blk_mq_delay_kick_requeue_list(q, msecs);
 246 }
 247
 248 void dm_mq_kick_requeue_list(struct mapped_device *md)
 249 {
 250         __dm_mq_kick_requeue_list(dm_get_md_queue(md), 0);
 251 }
 252 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
 253
 254 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
 255 {
 256         blk_mq_requeue_request(rq, false);
 257         __dm_mq_kick_requeue_list(rq->q, msecs);
 258 }
 259
 260 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
 261 {
 262         struct mapped_device *md = tio->md;
 263         struct request *rq = tio->orig;
 264         int rw = rq_data_dir(rq);
 265         unsigned long delay_ms = delay_requeue ? 100 : 0;
 266
 267         rq_end_stats(md, rq);
 268         if (tio->clone) {
 269                 blk_rq_unprep_clone(tio->clone);
 270                 tio->ti->type->release_clone_rq(tio->clone);
 271         }
 272
 273         if (!rq->q->mq_ops)
 274                 dm_old_requeue_request(rq, delay_ms);
 275         else
 276                 dm_mq_delay_requeue_request(rq, delay_ms);
 277
 278         rq_completed(md, rw, false);
 279 }
 280
 281 static void dm_done(struct request *clone, blk_status_t error, bool mapped)
 282 {
 283         int r = DM_ENDIO_DONE;
 284         struct dm_rq_target_io *tio = clone->end_io_data;
 285         dm_request_endio_fn rq_end_io = NULL;
 286
 287         if (tio->ti) {
 288                 rq_end_io = tio->ti->type->rq_end_io;
 289
 290                 if (mapped && rq_end_io)
 291                         r = rq_end_io(tio->ti, clone, error, &tio->info);
 292         }
 293
 294         if (unlikely(error == BLK_STS_TARGET)) {
 295                 if (req_op(clone) == REQ_OP_WRITE_SAME &&
 296                     !clone->q->limits.max_write_same_sectors)
 297                         disable_write_same(tio->md);
 298                 if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
 299                     !clone->q->limits.max_write_zeroes_sectors)
 300                         disable_write_zeroes(tio->md);
 301         }
 302
 303         switch (r) {
 304         case DM_ENDIO_DONE:
 305                 /* The target wants to complete the I/O */
 306                 dm_end_request(clone, error);
 307                 break;
 308         case DM_ENDIO_INCOMPLETE:
 309                 /* The target will handle the I/O */
 310                 return;
 311         case DM_ENDIO_REQUEUE:
 312                 /* The target wants to requeue the I/O */
 313                 dm_requeue_original_request(tio, false);
 314                 break;
 315         default:
 316                 DMWARN("unimplemented target endio return value: %d", r);
 317                 BUG();
 318         }
 319 }
 320
 321 /*
 322  * Request completion handler for request-based dm
 323  */
 324 static void dm_softirq_done(struct request *rq)
 325 {
 326         bool mapped = true;
 327         struct dm_rq_target_io *tio = tio_from_request(rq);
 328         struct request *clone = tio->clone;
 329         int rw;
 330
 331         if (!clone) {
 332                 struct mapped_device *md = tio->md;
 333
 334                 rq_end_stats(md, rq);
 335                 rw = rq_data_dir(rq);
 336                 if (!rq->q->mq_ops)
 337                         blk_end_request_all(rq, tio->error);
 338                 else
 339                         blk_mq_end_request(rq, tio->error);
 340                 rq_completed(md, rw, false);
 341                 return;
 342         }
 343
 344         if (rq->rq_flags & RQF_FAILED)
 345                 mapped = false;
 346
 347         dm_done(clone, tio->error, mapped);
 348 }
 349
 350 /*
 351  * Complete the clone and the original request with the error status
 352  * through softirq context.
 353  */
 354 static void dm_complete_request(struct request *rq, blk_status_t error)
 355 {
 356         struct dm_rq_target_io *tio = tio_from_request(rq);
 357
 358         tio->error = error;
 359         if (!rq->q->mq_ops)
 360                 blk_complete_request(rq);
 361         else
 362                 blk_mq_complete_request(rq);
 363 }
 364
 365 /*
 366  * Complete the not-mapped clone and the original request with the error status
 367  * through softirq context.
 368  * Target's rq_end_io() function isn't called.
 369  * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
 370  */
 371 static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
 372 {
 373         rq->rq_flags |= RQF_FAILED;
 374         dm_complete_request(rq, error);
 375 }
 376
 377 /*
 378  * Called with the clone's queue lock held (in the case of .request_fn)
 379  */
 380 static void end_clone_request(struct request *clone, blk_status_t error)
 381 {
 382         struct dm_rq_target_io *tio = clone->end_io_data;
 383
 384         /*
 385          * Actual request completion is done in a softirq context which doesn't
 386          * hold the clone's queue lock.  Otherwise, deadlock could occur because:
 387          *     - another request may be submitted by the upper level driver
 388          *       of the stacking during the completion
 389          *     - the submission which requires queue lock may be done
 390          *       against this clone's queue
 391          */
 392         dm_complete_request(tio->orig, error);
 393 }
 394
 395 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
 396 {
 397         blk_status_t r;
 398
 399         if (blk_queue_io_stat(clone->q))
 400                 clone->rq_flags |= RQF_IO_STAT;
 401
 402         clone->start_time = jiffies;
 403         r = blk_insert_cloned_request(clone->q, clone);
 404         if (r)
 405                 /* must complete clone in terms of original request */
 406                 dm_complete_request(rq, r);
 407 }
 408
 409 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
 410                                  void *data)
 411 {
 412         struct dm_rq_target_io *tio = data;
 413         struct dm_rq_clone_bio_info *info =
 414                 container_of(bio, struct dm_rq_clone_bio_info, clone);
 415
 416         info->orig = bio_orig;
 417         info->tio = tio;
 418         bio->bi_end_io = end_clone_bio;
 419
 420         return 0;
 421 }
 422
 423 static int setup_clone(struct request *clone, struct request *rq,
 424                        struct dm_rq_target_io *tio, gfp_t gfp_mask)
 425 {
 426         int r;
 427
 428         r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
 429                               dm_rq_bio_constructor, tio);
 430         if (r)
 431                 return r;
 432
 433         clone->end_io = end_clone_request;
 434         clone->end_io_data = tio;
 435
 436         tio->clone = clone;
 437
 438         return 0;
 439 }
 440
 441 static void map_tio_request(struct kthread_work *work);
 442
 443 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
 444                      struct mapped_device *md)
 445 {
 446         tio->md = md;
 447         tio->ti = NULL;
 448         tio->clone = NULL;
 449         tio->orig = rq;
 450         tio->error = 0;
 451         tio->completed = 0;
 452         /*
 453          * Avoid initializing info for blk-mq; it passes
 454          * target-specific data through info.ptr
 455          * (see: dm_mq_init_request)
 456          */
 457         if (!md->init_tio_pdu)
 458                 memset(&tio->info, 0, sizeof(tio->info));
 459         if (md->kworker_task)
 460                 kthread_init_work(&tio->work, map_tio_request);
 461 }
 462
 463 /*
 464  * Returns:
 465  * DM_MAPIO_*       : the request has been processed as indicated
 466  * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
 467  * < 0              : the request was completed due to failure
 468  */
 469 static int map_request(struct dm_rq_target_io *tio)
 470 {
 471         int r;
 472         struct dm_target *ti = tio->ti;
 473         struct mapped_device *md = tio->md;
 474         struct request *rq = tio->orig;
 475         struct request *clone = NULL;
 476
 477         r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
 478         switch (r) {
 479         case DM_MAPIO_SUBMITTED:
 480                 /* The target has taken the I/O to submit by itself later */
 481                 break;
 482         case DM_MAPIO_REMAPPED:
 483                 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
 484                         /* -ENOMEM */
 485                         ti->type->release_clone_rq(clone);
 486                         return DM_MAPIO_REQUEUE;
 487                 }
 488
 489                 /* The target has remapped the I/O so dispatch it */
 490                 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
 491                                      blk_rq_pos(rq));
 492                 dm_dispatch_clone_request(clone, rq);
 493                 break;
 494         case DM_MAPIO_REQUEUE:
 495                 /* The target wants to requeue the I/O */
 496                 break;
 497         case DM_MAPIO_DELAY_REQUEUE:
 498                 /* The target wants to requeue the I/O after a delay */
 499                 dm_requeue_original_request(tio, true);
 500                 break;
 501         case DM_MAPIO_KILL:
 502                 /* The target wants to complete the I/O */
 503                 dm_kill_unmapped_request(rq, BLK_STS_IOERR);
 504                 break;
 505         default:
 506                 DMWARN("unimplemented target map return value: %d", r);
 507                 BUG();
 508         }
 509
 510         return r;
 511 }
 512
 513 static void dm_start_request(struct mapped_device *md, struct request *orig)
 514 {
 515         if (!orig->q->mq_ops)
 516                 blk_start_request(orig);
 517         else
 518                 blk_mq_start_request(orig);
 519         atomic_inc(&md->pending[rq_data_dir(orig)]);
 520
 521         if (md->seq_rq_merge_deadline_usecs) {
 522                 md->last_rq_pos = rq_end_sector(orig);
 523                 md->last_rq_rw = rq_data_dir(orig);
 524                 md->last_rq_start_time = ktime_get();
 525         }
 526
 527         if (unlikely(dm_stats_used(&md->stats))) {
 528                 struct dm_rq_target_io *tio = tio_from_request(orig);
 529                 tio->duration_jiffies = jiffies;
 530                 tio->n_sectors = blk_rq_sectors(orig);
 531                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 532                                     blk_rq_pos(orig), tio->n_sectors, false, 0,
 533                                     &tio->stats_aux);
 534         }
 535
 536         /*
 537          * Hold the md reference here for the in-flight I/O.
 538          * We can't rely on the reference count by device opener,
 539          * because the device may be closed during the request completion
 540          * when all bios are completed.
 541          * See the comment in rq_completed() too.
 542          */
 543         dm_get(md);
 544 }
 545
 546 static int __dm_rq_init_rq(struct mapped_device *md, struct request *rq)
 547 {
 548         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 549
 550         /*
 551          * Must initialize md member of tio, otherwise it won't
 552          * be available in dm_mq_queue_rq.
 553          */
 554         tio->md = md;
 555
 556         if (md->init_tio_pdu) {
 557                 /* target-specific per-io data is immediately after the tio */
 558                 tio->info.ptr = tio + 1;
 559         }
 560
 561         return 0;
 562 }
 563
 564 static int dm_rq_init_rq(struct request_queue *q, struct request *rq, gfp_t gfp)
 565 {
 566         return __dm_rq_init_rq(q->rq_alloc_data, rq);
 567 }
 568
 569 static void map_tio_request(struct kthread_work *work)
 570 {
 571         struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
 572
 573         if (map_request(tio) == DM_MAPIO_REQUEUE)
 574                 dm_requeue_original_request(tio, false);
 575 }
 576
 577 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
 578 {
 579         return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
 580 }
 581
 582 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
 583
 584 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
 585                                                      const char *buf, size_t count)
 586 {
 587         unsigned deadline;
 588
 589         if (dm_get_md_type(md) != DM_TYPE_REQUEST_BASED)
 590                 return count;
 591
 592         if (kstrtouint(buf, 10, &deadline))
 593                 return -EINVAL;
 594
 595         if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
 596                 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
 597
 598         md->seq_rq_merge_deadline_usecs = deadline;
 599
 600         return count;
 601 }
 602
 603 static bool dm_old_request_peeked_before_merge_deadline(struct mapped_device *md)
 604 {
 605         ktime_t kt_deadline;
 606
 607         if (!md->seq_rq_merge_deadline_usecs)
 608                 return false;
 609
 610         kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
 611         kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
 612
 613         return !ktime_after(ktime_get(), kt_deadline);
 614 }
 615
 616 /*
 617  * q->request_fn for old request-based dm.
 618  * Called with the queue lock held.
 619  */
 620 static void dm_old_request_fn(struct request_queue *q)
 621 {
 622         struct mapped_device *md = q->queuedata;
 623         struct dm_target *ti = md->immutable_target;
 624         struct request *rq;
 625         struct dm_rq_target_io *tio;
 626         sector_t pos = 0;
 627
 628         if (unlikely(!ti)) {
 629                 int srcu_idx;
 630                 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
 631
 632                 if (unlikely(!map)) {
 633                         dm_put_live_table(md, srcu_idx);
 634                         return;
 635                 }
 636                 ti = dm_table_find_target(map, pos);
 637                 dm_put_live_table(md, srcu_idx);
 638         }
 639
 640         /*
 641          * For suspend, check blk_queue_stopped() and increment
 642          * ->pending within a single queue_lock not to increment the
 643          * number of in-flight I/Os after the queue is stopped in
 644          * dm_suspend().
 645          */
 646         while (!blk_queue_stopped(q)) {
 647                 rq = blk_peek_request(q);
 648                 if (!rq)
 649                         return;
 650
 651                 /* always use block 0 to find the target for flushes for now */
 652                 pos = 0;
 653                 if (req_op(rq) != REQ_OP_FLUSH)
 654                         pos = blk_rq_pos(rq);
 655
 656                 if ((dm_old_request_peeked_before_merge_deadline(md) &&
 657                      md_in_flight(md) && rq->bio && !bio_multiple_segments(rq->bio) &&
 658                      md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
 659                     (ti->type->busy && ti->type->busy(ti))) {
 660                         blk_delay_queue(q, 10);
 661                         return;
 662                 }
 663
 664                 dm_start_request(md, rq);
 665
 666                 tio = tio_from_request(rq);
 667                 init_tio(tio, rq, md);
 668                 /* Establish tio->ti before queuing work (map_tio_request) */
 669                 tio->ti = ti;
 670                 kthread_queue_work(&md->kworker, &tio->work);
 671                 BUG_ON(!irqs_disabled());
 672         }
 673 }
 674
 675 /*
 676  * Fully initialize a .request_fn request-based queue.
 677  */
 678 int dm_old_init_request_queue(struct mapped_device *md, struct dm_table *t)
 679 {
 680         struct dm_target *immutable_tgt;
 681
 682         /* Fully initialize the queue */
 683         md->queue->cmd_size = sizeof(struct dm_rq_target_io);
 684         md->queue->rq_alloc_data = md;
 685         md->queue->request_fn = dm_old_request_fn;
 686         md->queue->init_rq_fn = dm_rq_init_rq;
 687
 688         immutable_tgt = dm_table_get_immutable_target(t);
 689         if (immutable_tgt && immutable_tgt->per_io_data_size) {
 690                 /* any target-specific per-io data is immediately after the tio */
 691                 md->queue->cmd_size += immutable_tgt->per_io_data_size;
 692                 md->init_tio_pdu = true;
 693         }
 694         if (blk_init_allocated_queue(md->queue) < 0)
 695                 return -EINVAL;
 696
 697         /* disable dm_old_request_fn's merge heuristic by default */
 698         md->seq_rq_merge_deadline_usecs = 0;
 699
 700         dm_init_normal_md_queue(md);
 701         blk_queue_softirq_done(md->queue, dm_softirq_done);
 702
 703         /* Initialize the request-based DM worker thread */
 704         kthread_init_worker(&md->kworker);
 705         md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
 706                                        "kdmwork-%s", dm_device_name(md));
 707         if (IS_ERR(md->kworker_task)) {
 708                 int error = PTR_ERR(md->kworker_task);
 709                 md->kworker_task = NULL;
 710                 return error;
 711         }
 712
 713         elv_register_queue(md->queue);
 714
 715         return 0;
 716 }
 717
 718 static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
 719                 unsigned int hctx_idx, unsigned int numa_node)
 720 {
 721         return __dm_rq_init_rq(set->driver_data, rq);
 722 }
 723
 724 static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
 725                           const struct blk_mq_queue_data *bd)
 726 {
 727         struct request *rq = bd->rq;
 728         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 729         struct mapped_device *md = tio->md;
 730         struct dm_target *ti = md->immutable_target;
 731
 732         if (unlikely(!ti)) {
 733                 int srcu_idx;
 734                 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
 735
 736                 ti = dm_table_find_target(map, 0);
 737                 dm_put_live_table(md, srcu_idx);
 738         }
 739
 740         if (ti->type->busy && ti->type->busy(ti))
 741                 return BLK_STS_RESOURCE;
 742
 743         dm_start_request(md, rq);
 744
 745         /* Init tio using md established in .init_request */
 746         init_tio(tio, rq, md);
 747
 748         /*
 749          * Establish tio->ti before calling map_request().
 750          */
 751         tio->ti = ti;
 752
 753         /* Direct call is fine since .queue_rq allows allocations */
 754         if (map_request(tio) == DM_MAPIO_REQUEUE) {
 755                 /* Undo dm_start_request() before requeuing */
 756                 rq_end_stats(md, rq);
 757                 rq_completed(md, rq_data_dir(rq), false);
 758                 blk_mq_delay_run_hw_queue(hctx, 100/*ms*/);
 759                 return BLK_STS_RESOURCE;
 760         }
 761
 762         return BLK_STS_OK;
 763 }
 764
 765 static const struct blk_mq_ops dm_mq_ops = {
 766         .queue_rq = dm_mq_queue_rq,
 767         .complete = dm_softirq_done,
 768         .init_request = dm_mq_init_request,
 769 };
 770
 771 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
 772 {
 773         struct request_queue *q;
 774         struct dm_target *immutable_tgt;
 775         int err;
 776
 777         if (!dm_table_all_blk_mq_devices(t)) {
 778                 DMERR("request-based dm-mq may only be stacked on blk-mq device(s)");
 779                 return -EINVAL;
 780         }
 781
 782         md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
 783         if (!md->tag_set)
 784                 return -ENOMEM;
 785
 786         md->tag_set->ops = &dm_mq_ops;
 787         md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
 788         md->tag_set->numa_node = md->numa_node_id;
 789         md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
 790         md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
 791         md->tag_set->driver_data = md;
 792
 793         md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
 794         immutable_tgt = dm_table_get_immutable_target(t);
 795         if (immutable_tgt && immutable_tgt->per_io_data_size) {
 796                 /* any target-specific per-io data is immediately after the tio */
 797                 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
 798                 md->init_tio_pdu = true;
 799         }
 800
 801         err = blk_mq_alloc_tag_set(md->tag_set);
 802         if (err)
 803                 goto out_kfree_tag_set;
 804
 805         q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
 806         if (IS_ERR(q)) {
 807                 err = PTR_ERR(q);
 808                 goto out_tag_set;
 809         }
 810         dm_init_md_queue(md);
 811
 812         /* backfill 'mq' sysfs registration normally done in blk_register_queue */
 813         err = blk_mq_register_dev(disk_to_dev(md->disk), q);
 814         if (err)
 815                 goto out_cleanup_queue;
 816
 817         return 0;
 818
 819 out_cleanup_queue:
 820         blk_cleanup_queue(q);
 821 out_tag_set:
 822         blk_mq_free_tag_set(md->tag_set);
 823 out_kfree_tag_set:
 824         kfree(md->tag_set);
 825         md->tag_set = NULL;
 826
 827         return err;
 828 }
 829
 830 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
 831 {
 832         if (md->tag_set) {
 833                 blk_mq_free_tag_set(md->tag_set);
 834                 kfree(md->tag_set);
 835                 md->tag_set = NULL;
 836         }
 837 }
 838
 839 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
 840 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
 841
 842 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
 843 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
 844
 845 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
 846 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
 847
 848 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
 849 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");