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_start_stopped_hw_queues(q, true);
  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         if (blk_mq_queue_stopped(q))
  99                 return;
 100
 101         blk_mq_quiesce_queue(q);
 102 }
 103
 104 void dm_stop_queue(struct request_queue *q)
 105 {
 106         if (!q->mq_ops)
 107                 dm_old_stop_queue(q);
 108         else
 109                 dm_mq_stop_queue(q);
 110 }
 111
 112 /*
 113  * Partial completion handling for request-based dm
 114  */
 115 static void end_clone_bio(struct bio *clone)
 116 {
 117         struct dm_rq_clone_bio_info *info =
 118                 container_of(clone, struct dm_rq_clone_bio_info, clone);
 119         struct dm_rq_target_io *tio = info->tio;
 120         struct bio *bio = info->orig;
 121         unsigned int nr_bytes = info->orig->bi_iter.bi_size;
 122         int error = clone->bi_error;
 123
 124         bio_put(clone);
 125
 126         if (tio->error)
 127                 /*
 128                  * An error has already been detected on the request.
 129                  * Once error occurred, just let clone->end_io() handle
 130                  * the remainder.
 131                  */
 132                 return;
 133         else if (error) {
 134                 /*
 135                  * Don't notice the error to the upper layer yet.
 136                  * The error handling decision is made by the target driver,
 137                  * when the request is completed.
 138                  */
 139                 tio->error = error;
 140                 return;
 141         }
 142
 143         /*
 144          * I/O for the bio successfully completed.
 145          * Notice the data completion to the upper layer.
 146          */
 147
 148         /*
 149          * bios are processed from the head of the list.
 150          * So the completing bio should always be rq->bio.
 151          * If it's not, something wrong is happening.
 152          */
 153         if (tio->orig->bio != bio)
 154                 DMERR("bio completion is going in the middle of the request");
 155
 156         /*
 157          * Update the original request.
 158          * Do not use blk_end_request() here, because it may complete
 159          * the original request before the clone, and break the ordering.
 160          */
 161         blk_update_request(tio->orig, 0, nr_bytes);
 162 }
 163
 164 static struct dm_rq_target_io *tio_from_request(struct request *rq)
 165 {
 166         return blk_mq_rq_to_pdu(rq);
 167 }
 168
 169 static void rq_end_stats(struct mapped_device *md, struct request *orig)
 170 {
 171         if (unlikely(dm_stats_used(&md->stats))) {
 172                 struct dm_rq_target_io *tio = tio_from_request(orig);
 173                 tio->duration_jiffies = jiffies - tio->duration_jiffies;
 174                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 175                                     blk_rq_pos(orig), tio->n_sectors, true,
 176                                     tio->duration_jiffies, &tio->stats_aux);
 177         }
 178 }
 179
 180 /*
 181  * Don't touch any member of the md after calling this function because
 182  * the md may be freed in dm_put() at the end of this function.
 183  * Or do dm_get() before calling this function and dm_put() later.
 184  */
 185 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
 186 {
 187         struct request_queue *q = md->queue;
 188         unsigned long flags;
 189
 190         atomic_dec(&md->pending[rw]);
 191
 192         /* nudge anyone waiting on suspend queue */
 193         if (!md_in_flight(md))
 194                 wake_up(&md->wait);
 195
 196         /*
 197          * Run this off this callpath, as drivers could invoke end_io while
 198          * inside their request_fn (and holding the queue lock). Calling
 199          * back into ->request_fn() could deadlock attempting to grab the
 200          * queue lock again.
 201          */
 202         if (!q->mq_ops && run_queue) {
 203                 spin_lock_irqsave(q->queue_lock, flags);
 204                 blk_run_queue_async(q);
 205                 spin_unlock_irqrestore(q->queue_lock, flags);
 206         }
 207
 208         /*
 209          * dm_put() must be at the end of this function. See the comment above
 210          */
 211         dm_put(md);
 212 }
 213
 214 /*
 215  * Complete the clone and the original request.
 216  * Must be called without clone's queue lock held,
 217  * see end_clone_request() for more details.
 218  */
 219 static void dm_end_request(struct request *clone, int error)
 220 {
 221         int rw = rq_data_dir(clone);
 222         struct dm_rq_target_io *tio = clone->end_io_data;
 223         struct mapped_device *md = tio->md;
 224         struct request *rq = tio->orig;
 225
 226         blk_rq_unprep_clone(clone);
 227         tio->ti->type->release_clone_rq(clone);
 228
 229         rq_end_stats(md, rq);
 230         if (!rq->q->mq_ops)
 231                 blk_end_request_all(rq, error);
 232         else
 233                 blk_mq_end_request(rq, error);
 234         rq_completed(md, rw, true);
 235 }
 236
 237 /*
 238  * Requeue the original request of a clone.
 239  */
 240 static void dm_old_requeue_request(struct request *rq)
 241 {
 242         struct request_queue *q = rq->q;
 243         unsigned long flags;
 244
 245         spin_lock_irqsave(q->queue_lock, flags);
 246         blk_requeue_request(q, rq);
 247         blk_run_queue_async(q);
 248         spin_unlock_irqrestore(q->queue_lock, flags);
 249 }
 250
 251 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
 252 {
 253         blk_mq_delay_kick_requeue_list(q, msecs);
 254 }
 255
 256 void dm_mq_kick_requeue_list(struct mapped_device *md)
 257 {
 258         __dm_mq_kick_requeue_list(dm_get_md_queue(md), 0);
 259 }
 260 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
 261
 262 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
 263 {
 264         blk_mq_requeue_request(rq, false);
 265         __dm_mq_kick_requeue_list(rq->q, msecs);
 266 }
 267
 268 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
 269 {
 270         struct mapped_device *md = tio->md;
 271         struct request *rq = tio->orig;
 272         int rw = rq_data_dir(rq);
 273
 274         rq_end_stats(md, rq);
 275         if (tio->clone) {
 276                 blk_rq_unprep_clone(tio->clone);
 277                 tio->ti->type->release_clone_rq(tio->clone);
 278         }
 279
 280         if (!rq->q->mq_ops)
 281                 dm_old_requeue_request(rq);
 282         else
 283                 dm_mq_delay_requeue_request(rq, delay_requeue ? 100/*ms*/ : 0);
 284
 285         rq_completed(md, rw, false);
 286 }
 287
 288 static void dm_done(struct request *clone, int error, bool mapped)
 289 {
 290         int r = DM_ENDIO_DONE;
 291         struct dm_rq_target_io *tio = clone->end_io_data;
 292         dm_request_endio_fn rq_end_io = NULL;
 293
 294         if (tio->ti) {
 295                 rq_end_io = tio->ti->type->rq_end_io;
 296
 297                 if (mapped && rq_end_io)
 298                         r = rq_end_io(tio->ti, clone, error, &tio->info);
 299         }
 300
 301         if (unlikely(error == -EREMOTEIO)) {
 302                 if (req_op(clone) == REQ_OP_WRITE_SAME &&
 303                     !clone->q->limits.max_write_same_sectors)
 304                         disable_write_same(tio->md);
 305                 if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
 306                     !clone->q->limits.max_write_zeroes_sectors)
 307                         disable_write_zeroes(tio->md);
 308         }
 309
 310         switch (r) {
 311         case DM_ENDIO_DONE:
 312                 /* The target wants to complete the I/O */
 313                 dm_end_request(clone, error);
 314                 break;
 315         case DM_ENDIO_INCOMPLETE:
 316                 /* The target will handle the I/O */
 317                 return;
 318         case DM_ENDIO_REQUEUE:
 319                 /* The target wants to requeue the I/O */
 320                 dm_requeue_original_request(tio, false);
 321                 break;
 322         default:
 323                 DMWARN("unimplemented target endio return value: %d", r);
 324                 BUG();
 325         }
 326 }
 327
 328 /*
 329  * Request completion handler for request-based dm
 330  */
 331 static void dm_softirq_done(struct request *rq)
 332 {
 333         bool mapped = true;
 334         struct dm_rq_target_io *tio = tio_from_request(rq);
 335         struct request *clone = tio->clone;
 336         int rw;
 337
 338         if (!clone) {
 339                 struct mapped_device *md = tio->md;
 340
 341                 rq_end_stats(md, rq);
 342                 rw = rq_data_dir(rq);
 343                 if (!rq->q->mq_ops)
 344                         blk_end_request_all(rq, tio->error);
 345                 else
 346                         blk_mq_end_request(rq, tio->error);
 347                 rq_completed(md, rw, false);
 348                 return;
 349         }
 350
 351         if (rq->rq_flags & RQF_FAILED)
 352                 mapped = false;
 353
 354         dm_done(clone, tio->error, mapped);
 355 }
 356
 357 /*
 358  * Complete the clone and the original request with the error status
 359  * through softirq context.
 360  */
 361 static void dm_complete_request(struct request *rq, int error)
 362 {
 363         struct dm_rq_target_io *tio = tio_from_request(rq);
 364
 365         tio->error = error;
 366         if (!rq->q->mq_ops)
 367                 blk_complete_request(rq);
 368         else
 369                 blk_mq_complete_request(rq);
 370 }
 371
 372 /*
 373  * Complete the not-mapped clone and the original request with the error status
 374  * through softirq context.
 375  * Target's rq_end_io() function isn't called.
 376  * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
 377  */
 378 static void dm_kill_unmapped_request(struct request *rq, int error)
 379 {
 380         rq->rq_flags |= RQF_FAILED;
 381         dm_complete_request(rq, error);
 382 }
 383
 384 /*
 385  * Called with the clone's queue lock held (in the case of .request_fn)
 386  */
 387 static void end_clone_request(struct request *clone, int error)
 388 {
 389         struct dm_rq_target_io *tio = clone->end_io_data;
 390
 391         /*
 392          * Actual request completion is done in a softirq context which doesn't
 393          * hold the clone's queue lock.  Otherwise, deadlock could occur because:
 394          *     - another request may be submitted by the upper level driver
 395          *       of the stacking during the completion
 396          *     - the submission which requires queue lock may be done
 397          *       against this clone's queue
 398          */
 399         dm_complete_request(tio->orig, error);
 400 }
 401
 402 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
 403 {
 404         int r;
 405
 406         if (blk_queue_io_stat(clone->q))
 407                 clone->rq_flags |= RQF_IO_STAT;
 408
 409         clone->start_time = jiffies;
 410         r = blk_insert_cloned_request(clone->q, clone);
 411         if (r)
 412                 /* must complete clone in terms of original request */
 413                 dm_complete_request(rq, r);
 414 }
 415
 416 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
 417                                  void *data)
 418 {
 419         struct dm_rq_target_io *tio = data;
 420         struct dm_rq_clone_bio_info *info =
 421                 container_of(bio, struct dm_rq_clone_bio_info, clone);
 422
 423         info->orig = bio_orig;
 424         info->tio = tio;
 425         bio->bi_end_io = end_clone_bio;
 426
 427         return 0;
 428 }
 429
 430 static int setup_clone(struct request *clone, struct request *rq,
 431                        struct dm_rq_target_io *tio, gfp_t gfp_mask)
 432 {
 433         int r;
 434
 435         r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
 436                               dm_rq_bio_constructor, tio);
 437         if (r)
 438                 return r;
 439
 440         clone->end_io = end_clone_request;
 441         clone->end_io_data = tio;
 442
 443         tio->clone = clone;
 444
 445         return 0;
 446 }
 447
 448 static void map_tio_request(struct kthread_work *work);
 449
 450 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
 451                      struct mapped_device *md)
 452 {
 453         tio->md = md;
 454         tio->ti = NULL;
 455         tio->clone = NULL;
 456         tio->orig = rq;
 457         tio->error = 0;
 458         /*
 459          * Avoid initializing info for blk-mq; it passes
 460          * target-specific data through info.ptr
 461          * (see: dm_mq_init_request)
 462          */
 463         if (!md->init_tio_pdu)
 464                 memset(&tio->info, 0, sizeof(tio->info));
 465         if (md->kworker_task)
 466                 kthread_init_work(&tio->work, map_tio_request);
 467 }
 468
 469 /*
 470  * Returns:
 471  * DM_MAPIO_*       : the request has been processed as indicated
 472  * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
 473  * < 0              : the request was completed due to failure
 474  */
 475 static int map_request(struct dm_rq_target_io *tio)
 476 {
 477         int r;
 478         struct dm_target *ti = tio->ti;
 479         struct mapped_device *md = tio->md;
 480         struct request *rq = tio->orig;
 481         struct request *clone = NULL;
 482
 483         r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
 484         switch (r) {
 485         case DM_MAPIO_SUBMITTED:
 486                 /* The target has taken the I/O to submit by itself later */
 487                 break;
 488         case DM_MAPIO_REMAPPED:
 489                 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
 490                         /* -ENOMEM */
 491                         ti->type->release_clone_rq(clone);
 492                         return DM_MAPIO_REQUEUE;
 493                 }
 494
 495                 /* The target has remapped the I/O so dispatch it */
 496                 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
 497                                      blk_rq_pos(rq));
 498                 dm_dispatch_clone_request(clone, rq);
 499                 break;
 500         case DM_MAPIO_REQUEUE:
 501                 /* The target wants to requeue the I/O */
 502                 break;
 503         case DM_MAPIO_DELAY_REQUEUE:
 504                 /* The target wants to requeue the I/O after a delay */
 505                 dm_requeue_original_request(tio, true);
 506                 break;
 507         case DM_MAPIO_KILL:
 508                 /* The target wants to complete the I/O */
 509                 dm_kill_unmapped_request(rq, -EIO);
 510                 break;
 511         default:
 512                 DMWARN("unimplemented target map return value: %d", r);
 513                 BUG();
 514         }
 515
 516         return r;
 517 }
 518
 519 static void dm_start_request(struct mapped_device *md, struct request *orig)
 520 {
 521         if (!orig->q->mq_ops)
 522                 blk_start_request(orig);
 523         else
 524                 blk_mq_start_request(orig);
 525         atomic_inc(&md->pending[rq_data_dir(orig)]);
 526
 527         if (md->seq_rq_merge_deadline_usecs) {
 528                 md->last_rq_pos = rq_end_sector(orig);
 529                 md->last_rq_rw = rq_data_dir(orig);
 530                 md->last_rq_start_time = ktime_get();
 531         }
 532
 533         if (unlikely(dm_stats_used(&md->stats))) {
 534                 struct dm_rq_target_io *tio = tio_from_request(orig);
 535                 tio->duration_jiffies = jiffies;
 536                 tio->n_sectors = blk_rq_sectors(orig);
 537                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 538                                     blk_rq_pos(orig), tio->n_sectors, false, 0,
 539                                     &tio->stats_aux);
 540         }
 541
 542         /*
 543          * Hold the md reference here for the in-flight I/O.
 544          * We can't rely on the reference count by device opener,
 545          * because the device may be closed during the request completion
 546          * when all bios are completed.
 547          * See the comment in rq_completed() too.
 548          */
 549         dm_get(md);
 550 }
 551
 552 static int __dm_rq_init_rq(struct mapped_device *md, struct request *rq)
 553 {
 554         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 555
 556         /*
 557          * Must initialize md member of tio, otherwise it won't
 558          * be available in dm_mq_queue_rq.
 559          */
 560         tio->md = md;
 561
 562         if (md->init_tio_pdu) {
 563                 /* target-specific per-io data is immediately after the tio */
 564                 tio->info.ptr = tio + 1;
 565         }
 566
 567         return 0;
 568 }
 569
 570 static int dm_rq_init_rq(struct request_queue *q, struct request *rq, gfp_t gfp)
 571 {
 572         return __dm_rq_init_rq(q->rq_alloc_data, rq);
 573 }
 574
 575 static void map_tio_request(struct kthread_work *work)
 576 {
 577         struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
 578
 579         if (map_request(tio) == DM_MAPIO_REQUEUE)
 580                 dm_requeue_original_request(tio, false);
 581 }
 582
 583 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
 584 {
 585         return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
 586 }
 587
 588 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
 589
 590 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
 591                                                      const char *buf, size_t count)
 592 {
 593         unsigned deadline;
 594
 595         if (dm_get_md_type(md) != DM_TYPE_REQUEST_BASED)
 596                 return count;
 597
 598         if (kstrtouint(buf, 10, &deadline))
 599                 return -EINVAL;
 600
 601         if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
 602                 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
 603
 604         md->seq_rq_merge_deadline_usecs = deadline;
 605
 606         return count;
 607 }
 608
 609 static bool dm_old_request_peeked_before_merge_deadline(struct mapped_device *md)
 610 {
 611         ktime_t kt_deadline;
 612
 613         if (!md->seq_rq_merge_deadline_usecs)
 614                 return false;
 615
 616         kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
 617         kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
 618
 619         return !ktime_after(ktime_get(), kt_deadline);
 620 }
 621
 622 /*
 623  * q->request_fn for old request-based dm.
 624  * Called with the queue lock held.
 625  */
 626 static void dm_old_request_fn(struct request_queue *q)
 627 {
 628         struct mapped_device *md = q->queuedata;
 629         struct dm_target *ti = md->immutable_target;
 630         struct request *rq;
 631         struct dm_rq_target_io *tio;
 632         sector_t pos = 0;
 633
 634         if (unlikely(!ti)) {
 635                 int srcu_idx;
 636                 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
 637
 638                 if (unlikely(!map)) {
 639                         dm_put_live_table(md, srcu_idx);
 640                         return;
 641                 }
 642                 ti = dm_table_find_target(map, pos);
 643                 dm_put_live_table(md, srcu_idx);
 644         }
 645
 646         /*
 647          * For suspend, check blk_queue_stopped() and increment
 648          * ->pending within a single queue_lock not to increment the
 649          * number of in-flight I/Os after the queue is stopped in
 650          * dm_suspend().
 651          */
 652         while (!blk_queue_stopped(q)) {
 653                 rq = blk_peek_request(q);
 654                 if (!rq)
 655                         return;
 656
 657                 /* always use block 0 to find the target for flushes for now */
 658                 pos = 0;
 659                 if (req_op(rq) != REQ_OP_FLUSH)
 660                         pos = blk_rq_pos(rq);
 661
 662                 if ((dm_old_request_peeked_before_merge_deadline(md) &&
 663                      md_in_flight(md) && rq->bio && !bio_multiple_segments(rq->bio) &&
 664                      md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
 665                     (ti->type->busy && ti->type->busy(ti))) {
 666                         blk_delay_queue(q, 10);
 667                         return;
 668                 }
 669
 670                 dm_start_request(md, rq);
 671
 672                 tio = tio_from_request(rq);
 673                 init_tio(tio, rq, md);
 674                 /* Establish tio->ti before queuing work (map_tio_request) */
 675                 tio->ti = ti;
 676                 kthread_queue_work(&md->kworker, &tio->work);
 677                 BUG_ON(!irqs_disabled());
 678         }
 679 }
 680
 681 /*
 682  * Fully initialize a .request_fn request-based queue.
 683  */
 684 int dm_old_init_request_queue(struct mapped_device *md, struct dm_table *t)
 685 {
 686         struct dm_target *immutable_tgt;
 687
 688         /* Fully initialize the queue */
 689         md->queue->cmd_size = sizeof(struct dm_rq_target_io);
 690         md->queue->rq_alloc_data = md;
 691         md->queue->request_fn = dm_old_request_fn;
 692         md->queue->init_rq_fn = dm_rq_init_rq;
 693
 694         immutable_tgt = dm_table_get_immutable_target(t);
 695         if (immutable_tgt && immutable_tgt->per_io_data_size) {
 696                 /* any target-specific per-io data is immediately after the tio */
 697                 md->queue->cmd_size += immutable_tgt->per_io_data_size;
 698                 md->init_tio_pdu = true;
 699         }
 700         if (blk_init_allocated_queue(md->queue) < 0)
 701                 return -EINVAL;
 702
 703         /* disable dm_old_request_fn's merge heuristic by default */
 704         md->seq_rq_merge_deadline_usecs = 0;
 705
 706         dm_init_normal_md_queue(md);
 707         blk_queue_softirq_done(md->queue, dm_softirq_done);
 708
 709         /* Initialize the request-based DM worker thread */
 710         kthread_init_worker(&md->kworker);
 711         md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
 712                                        "kdmwork-%s", dm_device_name(md));
 713         if (IS_ERR(md->kworker_task)) {
 714                 int error = PTR_ERR(md->kworker_task);
 715                 md->kworker_task = NULL;
 716                 return error;
 717         }
 718
 719         elv_register_queue(md->queue);
 720
 721         return 0;
 722 }
 723
 724 static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
 725                 unsigned int hctx_idx, unsigned int numa_node)
 726 {
 727         return __dm_rq_init_rq(set->driver_data, rq);
 728 }
 729
 730 static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
 731                           const struct blk_mq_queue_data *bd)
 732 {
 733         struct request *rq = bd->rq;
 734         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 735         struct mapped_device *md = tio->md;
 736         struct dm_target *ti = md->immutable_target;
 737
 738         if (unlikely(!ti)) {
 739                 int srcu_idx;
 740                 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
 741
 742                 ti = dm_table_find_target(map, 0);
 743                 dm_put_live_table(md, srcu_idx);
 744         }
 745
 746         if (ti->type->busy && ti->type->busy(ti))
 747                 return BLK_MQ_RQ_QUEUE_BUSY;
 748
 749         dm_start_request(md, rq);
 750
 751         /* Init tio using md established in .init_request */
 752         init_tio(tio, rq, md);
 753
 754         /*
 755          * Establish tio->ti before calling map_request().
 756          */
 757         tio->ti = ti;
 758
 759         /* Direct call is fine since .queue_rq allows allocations */
 760         if (map_request(tio) == DM_MAPIO_REQUEUE) {
 761                 /* Undo dm_start_request() before requeuing */
 762                 rq_end_stats(md, rq);
 763                 rq_completed(md, rq_data_dir(rq), false);
 764                 blk_mq_delay_run_hw_queue(hctx, 100/*ms*/);
 765                 return BLK_MQ_RQ_QUEUE_BUSY;
 766         }
 767
 768         return BLK_MQ_RQ_QUEUE_OK;
 769 }
 770
 771 static const struct blk_mq_ops dm_mq_ops = {
 772         .queue_rq = dm_mq_queue_rq,
 773         .complete = dm_softirq_done,
 774         .init_request = dm_mq_init_request,
 775 };
 776
 777 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
 778 {
 779         struct request_queue *q;
 780         struct dm_target *immutable_tgt;
 781         int err;
 782
 783         if (!dm_table_all_blk_mq_devices(t)) {
 784                 DMERR("request-based dm-mq may only be stacked on blk-mq device(s)");
 785                 return -EINVAL;
 786         }
 787
 788         md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
 789         if (!md->tag_set)
 790                 return -ENOMEM;
 791
 792         md->tag_set->ops = &dm_mq_ops;
 793         md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
 794         md->tag_set->numa_node = md->numa_node_id;
 795         md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
 796         md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
 797         md->tag_set->driver_data = md;
 798
 799         md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
 800         immutable_tgt = dm_table_get_immutable_target(t);
 801         if (immutable_tgt && immutable_tgt->per_io_data_size) {
 802                 /* any target-specific per-io data is immediately after the tio */
 803                 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
 804                 md->init_tio_pdu = true;
 805         }
 806
 807         err = blk_mq_alloc_tag_set(md->tag_set);
 808         if (err)
 809                 goto out_kfree_tag_set;
 810
 811         q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
 812         if (IS_ERR(q)) {
 813                 err = PTR_ERR(q);
 814                 goto out_tag_set;
 815         }
 816         dm_init_md_queue(md);
 817
 818         /* backfill 'mq' sysfs registration normally done in blk_register_queue */
 819         err = blk_mq_register_dev(disk_to_dev(md->disk), q);
 820         if (err)
 821                 goto out_cleanup_queue;
 822
 823         return 0;
 824
 825 out_cleanup_queue:
 826         blk_cleanup_queue(q);
 827 out_tag_set:
 828         blk_mq_free_tag_set(md->tag_set);
 829 out_kfree_tag_set:
 830         kfree(md->tag_set);
 831
 832         return err;
 833 }
 834
 835 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
 836 {
 837         if (md->tag_set) {
 838                 blk_mq_free_tag_set(md->tag_set);
 839                 kfree(md->tag_set);
 840         }
 841 }
 842
 843 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
 844 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
 845
 846 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
 847 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
 848
 849 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
 850 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
 851
 852 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
 853 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");