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 #ifdef CONFIG_DM_MQ_DEFAULT
  27 static bool use_blk_mq = true;
  28 #else
  29 static bool use_blk_mq = false;
  30 #endif
  31
  32 bool dm_use_blk_mq_default(void)
  33 {
  34         return use_blk_mq;
  35 }
  36
  37 bool dm_use_blk_mq(struct mapped_device *md)
  38 {
  39         return md->use_blk_mq;
  40 }
  41 EXPORT_SYMBOL_GPL(dm_use_blk_mq);
  42
  43 unsigned dm_get_reserved_rq_based_ios(void)
  44 {
  45         return __dm_get_module_param(&reserved_rq_based_ios,
  46                                      RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
  47 }
  48 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
  49
  50 static unsigned dm_get_blk_mq_nr_hw_queues(void)
  51 {
  52         return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
  53 }
  54
  55 static unsigned dm_get_blk_mq_queue_depth(void)
  56 {
  57         return __dm_get_module_param(&dm_mq_queue_depth,
  58                                      DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
  59 }
  60
  61 int dm_request_based(struct mapped_device *md)
  62 {
  63         return blk_queue_stackable(md->queue);
  64 }
  65
  66 static void dm_old_start_queue(struct request_queue *q)
  67 {
  68         unsigned long flags;
  69
  70         spin_lock_irqsave(q->queue_lock, flags);
  71         if (blk_queue_stopped(q))
  72                 blk_start_queue(q);
  73         spin_unlock_irqrestore(q->queue_lock, flags);
  74 }
  75
  76 static void dm_mq_start_queue(struct request_queue *q)
  77 {
  78         unsigned long flags;
  79
  80         spin_lock_irqsave(q->queue_lock, flags);
  81         queue_flag_clear(QUEUE_FLAG_STOPPED, q);
  82         spin_unlock_irqrestore(q->queue_lock, flags);
  83
  84         blk_mq_start_stopped_hw_queues(q, true);
  85         blk_mq_kick_requeue_list(q);
  86 }
  87
  88 void dm_start_queue(struct request_queue *q)
  89 {
  90         if (!q->mq_ops)
  91                 dm_old_start_queue(q);
  92         else
  93                 dm_mq_start_queue(q);
  94 }
  95
  96 static void dm_old_stop_queue(struct request_queue *q)
  97 {
  98         unsigned long flags;
  99
 100         spin_lock_irqsave(q->queue_lock, flags);
 101         if (!blk_queue_stopped(q))
 102                 blk_stop_queue(q);
 103         spin_unlock_irqrestore(q->queue_lock, flags);
 104 }
 105
 106 static void dm_mq_stop_queue(struct request_queue *q)
 107 {
 108         unsigned long flags;
 109
 110         spin_lock_irqsave(q->queue_lock, flags);
 111         if (blk_queue_stopped(q)) {
 112                 spin_unlock_irqrestore(q->queue_lock, flags);
 113                 return;
 114         }
 115
 116         queue_flag_set(QUEUE_FLAG_STOPPED, q);
 117         spin_unlock_irqrestore(q->queue_lock, flags);
 118
 119         /* Avoid that requeuing could restart the queue. */
 120         blk_mq_cancel_requeue_work(q);
 121         blk_mq_stop_hw_queues(q);
 122 }
 123
 124 void dm_stop_queue(struct request_queue *q)
 125 {
 126         if (!q->mq_ops)
 127                 dm_old_stop_queue(q);
 128         else
 129                 dm_mq_stop_queue(q);
 130 }
 131
 132 static struct dm_rq_target_io *alloc_old_rq_tio(struct mapped_device *md,
 133                                                 gfp_t gfp_mask)
 134 {
 135         return mempool_alloc(md->io_pool, gfp_mask);
 136 }
 137
 138 static void free_old_rq_tio(struct dm_rq_target_io *tio)
 139 {
 140         mempool_free(tio, tio->md->io_pool);
 141 }
 142
 143 static struct request *alloc_old_clone_request(struct mapped_device *md,
 144                                                gfp_t gfp_mask)
 145 {
 146         return mempool_alloc(md->rq_pool, gfp_mask);
 147 }
 148
 149 static void free_old_clone_request(struct mapped_device *md, struct request *rq)
 150 {
 151         mempool_free(rq, md->rq_pool);
 152 }
 153
 154 /*
 155  * Partial completion handling for request-based dm
 156  */
 157 static void end_clone_bio(struct bio *clone)
 158 {
 159         struct dm_rq_clone_bio_info *info =
 160                 container_of(clone, struct dm_rq_clone_bio_info, clone);
 161         struct dm_rq_target_io *tio = info->tio;
 162         struct bio *bio = info->orig;
 163         unsigned int nr_bytes = info->orig->bi_iter.bi_size;
 164         int error = clone->bi_error;
 165
 166         bio_put(clone);
 167
 168         if (tio->error)
 169                 /*
 170                  * An error has already been detected on the request.
 171                  * Once error occurred, just let clone->end_io() handle
 172                  * the remainder.
 173                  */
 174                 return;
 175         else if (error) {
 176                 /*
 177                  * Don't notice the error to the upper layer yet.
 178                  * The error handling decision is made by the target driver,
 179                  * when the request is completed.
 180                  */
 181                 tio->error = error;
 182                 return;
 183         }
 184
 185         /*
 186          * I/O for the bio successfully completed.
 187          * Notice the data completion to the upper layer.
 188          */
 189
 190         /*
 191          * bios are processed from the head of the list.
 192          * So the completing bio should always be rq->bio.
 193          * If it's not, something wrong is happening.
 194          */
 195         if (tio->orig->bio != bio)
 196                 DMERR("bio completion is going in the middle of the request");
 197
 198         /*
 199          * Update the original request.
 200          * Do not use blk_end_request() here, because it may complete
 201          * the original request before the clone, and break the ordering.
 202          */
 203         blk_update_request(tio->orig, 0, nr_bytes);
 204 }
 205
 206 static struct dm_rq_target_io *tio_from_request(struct request *rq)
 207 {
 208         return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
 209 }
 210
 211 static void rq_end_stats(struct mapped_device *md, struct request *orig)
 212 {
 213         if (unlikely(dm_stats_used(&md->stats))) {
 214                 struct dm_rq_target_io *tio = tio_from_request(orig);
 215                 tio->duration_jiffies = jiffies - tio->duration_jiffies;
 216                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 217                                     blk_rq_pos(orig), tio->n_sectors, true,
 218                                     tio->duration_jiffies, &tio->stats_aux);
 219         }
 220 }
 221
 222 /*
 223  * Don't touch any member of the md after calling this function because
 224  * the md may be freed in dm_put() at the end of this function.
 225  * Or do dm_get() before calling this function and dm_put() later.
 226  */
 227 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
 228 {
 229         struct request_queue *q = md->queue;
 230         unsigned long flags;
 231
 232         atomic_dec(&md->pending[rw]);
 233
 234         /* nudge anyone waiting on suspend queue */
 235         if (!md_in_flight(md))
 236                 wake_up(&md->wait);
 237
 238         /*
 239          * Run this off this callpath, as drivers could invoke end_io while
 240          * inside their request_fn (and holding the queue lock). Calling
 241          * back into ->request_fn() could deadlock attempting to grab the
 242          * queue lock again.
 243          */
 244         if (!q->mq_ops && run_queue) {
 245                 spin_lock_irqsave(q->queue_lock, flags);
 246                 blk_run_queue_async(q);
 247                 spin_unlock_irqrestore(q->queue_lock, flags);
 248         }
 249
 250         /*
 251          * dm_put() must be at the end of this function. See the comment above
 252          */
 253         dm_put(md);
 254 }
 255
 256 static void free_rq_clone(struct request *clone)
 257 {
 258         struct dm_rq_target_io *tio = clone->end_io_data;
 259         struct mapped_device *md = tio->md;
 260
 261         blk_rq_unprep_clone(clone);
 262
 263         /*
 264          * It is possible for a clone_old_rq() allocated clone to
 265          * get passed in -- it may not yet have a request_queue.
 266          * This is known to occur if the error target replaces
 267          * a multipath target that has a request_fn queue stacked
 268          * on blk-mq queue(s).
 269          */
 270         if (clone->q && clone->q->mq_ops)
 271                 /* stacked on blk-mq queue(s) */
 272                 tio->ti->type->release_clone_rq(clone);
 273         else if (!md->queue->mq_ops)
 274                 /* request_fn queue stacked on request_fn queue(s) */
 275                 free_old_clone_request(md, clone);
 276
 277         if (!md->queue->mq_ops)
 278                 free_old_rq_tio(tio);
 279 }
 280
 281 /*
 282  * Complete the clone and the original request.
 283  * Must be called without clone's queue lock held,
 284  * see end_clone_request() for more details.
 285  */
 286 static void dm_end_request(struct request *clone, int error)
 287 {
 288         int rw = rq_data_dir(clone);
 289         struct dm_rq_target_io *tio = clone->end_io_data;
 290         struct mapped_device *md = tio->md;
 291         struct request *rq = tio->orig;
 292
 293         if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
 294                 rq->errors = clone->errors;
 295                 rq->resid_len = clone->resid_len;
 296
 297                 if (rq->sense)
 298                         /*
 299                          * We are using the sense buffer of the original
 300                          * request.
 301                          * So setting the length of the sense data is enough.
 302                          */
 303                         rq->sense_len = clone->sense_len;
 304         }
 305
 306         free_rq_clone(clone);
 307         rq_end_stats(md, rq);
 308         if (!rq->q->mq_ops)
 309                 blk_end_request_all(rq, error);
 310         else
 311                 blk_mq_end_request(rq, error);
 312         rq_completed(md, rw, true);
 313 }
 314
 315 static void dm_unprep_request(struct request *rq)
 316 {
 317         struct dm_rq_target_io *tio = tio_from_request(rq);
 318         struct request *clone = tio->clone;
 319
 320         if (!rq->q->mq_ops) {
 321                 rq->special = NULL;
 322                 rq->cmd_flags &= ~REQ_DONTPREP;
 323         }
 324
 325         if (clone)
 326                 free_rq_clone(clone);
 327         else if (!tio->md->queue->mq_ops)
 328                 free_old_rq_tio(tio);
 329 }
 330
 331 /*
 332  * Requeue the original request of a clone.
 333  */
 334 static void dm_old_requeue_request(struct request *rq)
 335 {
 336         struct request_queue *q = rq->q;
 337         unsigned long flags;
 338
 339         spin_lock_irqsave(q->queue_lock, flags);
 340         blk_requeue_request(q, rq);
 341         blk_run_queue_async(q);
 342         spin_unlock_irqrestore(q->queue_lock, flags);
 343 }
 344
 345 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
 346 {
 347         unsigned long flags;
 348
 349         spin_lock_irqsave(q->queue_lock, flags);
 350         if (!blk_queue_stopped(q))
 351                 blk_mq_delay_kick_requeue_list(q, msecs);
 352         spin_unlock_irqrestore(q->queue_lock, flags);
 353 }
 354
 355 void dm_mq_kick_requeue_list(struct mapped_device *md)
 356 {
 357         __dm_mq_kick_requeue_list(dm_get_md_queue(md), 0);
 358 }
 359 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
 360
 361 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
 362 {
 363         blk_mq_requeue_request(rq);
 364         __dm_mq_kick_requeue_list(rq->q, msecs);
 365 }
 366
 367 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
 368 {
 369         struct mapped_device *md = tio->md;
 370         struct request *rq = tio->orig;
 371         int rw = rq_data_dir(rq);
 372
 373         rq_end_stats(md, rq);
 374         dm_unprep_request(rq);
 375
 376         if (!rq->q->mq_ops)
 377                 dm_old_requeue_request(rq);
 378         else
 379                 dm_mq_delay_requeue_request(rq, delay_requeue ? 5000 : 0);
 380
 381         rq_completed(md, rw, false);
 382 }
 383
 384 static void dm_done(struct request *clone, int error, bool mapped)
 385 {
 386         int r = error;
 387         struct dm_rq_target_io *tio = clone->end_io_data;
 388         dm_request_endio_fn rq_end_io = NULL;
 389
 390         if (tio->ti) {
 391                 rq_end_io = tio->ti->type->rq_end_io;
 392
 393                 if (mapped && rq_end_io)
 394                         r = rq_end_io(tio->ti, clone, error, &tio->info);
 395         }
 396
 397         if (unlikely(r == -EREMOTEIO && (req_op(clone) == REQ_OP_WRITE_SAME) &&
 398                      !clone->q->limits.max_write_same_sectors))
 399                 disable_write_same(tio->md);
 400
 401         if (r <= 0)
 402                 /* The target wants to complete the I/O */
 403                 dm_end_request(clone, r);
 404         else if (r == DM_ENDIO_INCOMPLETE)
 405                 /* The target will handle the I/O */
 406                 return;
 407         else if (r == DM_ENDIO_REQUEUE)
 408                 /* The target wants to requeue the I/O */
 409                 dm_requeue_original_request(tio, false);
 410         else {
 411                 DMWARN("unimplemented target endio return value: %d", r);
 412                 BUG();
 413         }
 414 }
 415
 416 /*
 417  * Request completion handler for request-based dm
 418  */
 419 static void dm_softirq_done(struct request *rq)
 420 {
 421         bool mapped = true;
 422         struct dm_rq_target_io *tio = tio_from_request(rq);
 423         struct request *clone = tio->clone;
 424         int rw;
 425
 426         if (!clone) {
 427                 rq_end_stats(tio->md, rq);
 428                 rw = rq_data_dir(rq);
 429                 if (!rq->q->mq_ops) {
 430                         blk_end_request_all(rq, tio->error);
 431                         rq_completed(tio->md, rw, false);
 432                         free_old_rq_tio(tio);
 433                 } else {
 434                         blk_mq_end_request(rq, tio->error);
 435                         rq_completed(tio->md, rw, false);
 436                 }
 437                 return;
 438         }
 439
 440         if (rq->cmd_flags & REQ_FAILED)
 441                 mapped = false;
 442
 443         dm_done(clone, tio->error, mapped);
 444 }
 445
 446 /*
 447  * Complete the clone and the original request with the error status
 448  * through softirq context.
 449  */
 450 static void dm_complete_request(struct request *rq, int error)
 451 {
 452         struct dm_rq_target_io *tio = tio_from_request(rq);
 453
 454         tio->error = error;
 455         if (!rq->q->mq_ops)
 456                 blk_complete_request(rq);
 457         else
 458                 blk_mq_complete_request(rq, error);
 459 }
 460
 461 /*
 462  * Complete the not-mapped clone and the original request with the error status
 463  * through softirq context.
 464  * Target's rq_end_io() function isn't called.
 465  * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
 466  */
 467 static void dm_kill_unmapped_request(struct request *rq, int error)
 468 {
 469         rq->cmd_flags |= REQ_FAILED;
 470         dm_complete_request(rq, error);
 471 }
 472
 473 /*
 474  * Called with the clone's queue lock held (in the case of .request_fn)
 475  */
 476 static void end_clone_request(struct request *clone, int error)
 477 {
 478         struct dm_rq_target_io *tio = clone->end_io_data;
 479
 480         if (!clone->q->mq_ops) {
 481                 /*
 482                  * For just cleaning up the information of the queue in which
 483                  * the clone was dispatched.
 484                  * The clone is *NOT* freed actually here because it is alloced
 485                  * from dm own mempool (REQ_ALLOCED isn't set).
 486                  */
 487                 __blk_put_request(clone->q, clone);
 488         }
 489
 490         /*
 491          * Actual request completion is done in a softirq context which doesn't
 492          * hold the clone's queue lock.  Otherwise, deadlock could occur because:
 493          *     - another request may be submitted by the upper level driver
 494          *       of the stacking during the completion
 495          *     - the submission which requires queue lock may be done
 496          *       against this clone's queue
 497          */
 498         dm_complete_request(tio->orig, error);
 499 }
 500
 501 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
 502 {
 503         int r;
 504
 505         if (blk_queue_io_stat(clone->q))
 506                 clone->cmd_flags |= REQ_IO_STAT;
 507
 508         clone->start_time = jiffies;
 509         r = blk_insert_cloned_request(clone->q, clone);
 510         if (r)
 511                 /* must complete clone in terms of original request */
 512                 dm_complete_request(rq, r);
 513 }
 514
 515 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
 516                                  void *data)
 517 {
 518         struct dm_rq_target_io *tio = data;
 519         struct dm_rq_clone_bio_info *info =
 520                 container_of(bio, struct dm_rq_clone_bio_info, clone);
 521
 522         info->orig = bio_orig;
 523         info->tio = tio;
 524         bio->bi_end_io = end_clone_bio;
 525
 526         return 0;
 527 }
 528
 529 static int setup_clone(struct request *clone, struct request *rq,
 530                        struct dm_rq_target_io *tio, gfp_t gfp_mask)
 531 {
 532         int r;
 533
 534         r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
 535                               dm_rq_bio_constructor, tio);
 536         if (r)
 537                 return r;
 538
 539         clone->cmd = rq->cmd;
 540         clone->cmd_len = rq->cmd_len;
 541         clone->sense = rq->sense;
 542         clone->end_io = end_clone_request;
 543         clone->end_io_data = tio;
 544
 545         tio->clone = clone;
 546
 547         return 0;
 548 }
 549
 550 static struct request *clone_old_rq(struct request *rq, struct mapped_device *md,
 551                                     struct dm_rq_target_io *tio, gfp_t gfp_mask)
 552 {
 553         /*
 554          * Create clone for use with .request_fn request_queue
 555          */
 556         struct request *clone;
 557
 558         clone = alloc_old_clone_request(md, gfp_mask);
 559         if (!clone)
 560                 return NULL;
 561
 562         blk_rq_init(NULL, clone);
 563         if (setup_clone(clone, rq, tio, gfp_mask)) {
 564                 /* -ENOMEM */
 565                 free_old_clone_request(md, clone);
 566                 return NULL;
 567         }
 568
 569         return clone;
 570 }
 571
 572 static void map_tio_request(struct kthread_work *work);
 573
 574 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
 575                      struct mapped_device *md)
 576 {
 577         tio->md = md;
 578         tio->ti = NULL;
 579         tio->clone = NULL;
 580         tio->orig = rq;
 581         tio->error = 0;
 582         /*
 583          * Avoid initializing info for blk-mq; it passes
 584          * target-specific data through info.ptr
 585          * (see: dm_mq_init_request)
 586          */
 587         if (!md->init_tio_pdu)
 588                 memset(&tio->info, 0, sizeof(tio->info));
 589         if (md->kworker_task)
 590                 kthread_init_work(&tio->work, map_tio_request);
 591 }
 592
 593 static struct dm_rq_target_io *dm_old_prep_tio(struct request *rq,
 594                                                struct mapped_device *md,
 595                                                gfp_t gfp_mask)
 596 {
 597         struct dm_rq_target_io *tio;
 598         int srcu_idx;
 599         struct dm_table *table;
 600
 601         tio = alloc_old_rq_tio(md, gfp_mask);
 602         if (!tio)
 603                 return NULL;
 604
 605         init_tio(tio, rq, md);
 606
 607         table = dm_get_live_table(md, &srcu_idx);
 608         /*
 609          * Must clone a request if this .request_fn DM device
 610          * is stacked on .request_fn device(s).
 611          */
 612         if (!dm_table_all_blk_mq_devices(table)) {
 613                 if (!clone_old_rq(rq, md, tio, gfp_mask)) {
 614                         dm_put_live_table(md, srcu_idx);
 615                         free_old_rq_tio(tio);
 616                         return NULL;
 617                 }
 618         }
 619         dm_put_live_table(md, srcu_idx);
 620
 621         return tio;
 622 }
 623
 624 /*
 625  * Called with the queue lock held.
 626  */
 627 static int dm_old_prep_fn(struct request_queue *q, struct request *rq)
 628 {
 629         struct mapped_device *md = q->queuedata;
 630         struct dm_rq_target_io *tio;
 631
 632         if (unlikely(rq->special)) {
 633                 DMWARN("Already has something in rq->special.");
 634                 return BLKPREP_KILL;
 635         }
 636
 637         tio = dm_old_prep_tio(rq, md, GFP_ATOMIC);
 638         if (!tio)
 639                 return BLKPREP_DEFER;
 640
 641         rq->special = tio;
 642         rq->cmd_flags |= REQ_DONTPREP;
 643
 644         return BLKPREP_OK;
 645 }
 646
 647 /*
 648  * Returns:
 649  * DM_MAPIO_*       : the request has been processed as indicated
 650  * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
 651  * < 0              : the request was completed due to failure
 652  */
 653 static int map_request(struct dm_rq_target_io *tio)
 654 {
 655         int r;
 656         struct dm_target *ti = tio->ti;
 657         struct mapped_device *md = tio->md;
 658         struct request *rq = tio->orig;
 659         struct request *clone = NULL;
 660
 661         if (tio->clone) {
 662                 clone = tio->clone;
 663                 r = ti->type->map_rq(ti, clone, &tio->info);
 664                 if (r == DM_MAPIO_DELAY_REQUEUE)
 665                         return DM_MAPIO_REQUEUE; /* .request_fn requeue is always immediate */
 666         } else {
 667                 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
 668                 if (r < 0) {
 669                         /* The target wants to complete the I/O */
 670                         dm_kill_unmapped_request(rq, r);
 671                         return r;
 672                 }
 673                 if (r == DM_MAPIO_REMAPPED &&
 674                     setup_clone(clone, rq, tio, GFP_ATOMIC)) {
 675                         /* -ENOMEM */
 676                         ti->type->release_clone_rq(clone);
 677                         return DM_MAPIO_REQUEUE;
 678                 }
 679         }
 680
 681         switch (r) {
 682         case DM_MAPIO_SUBMITTED:
 683                 /* The target has taken the I/O to submit by itself later */
 684                 break;
 685         case DM_MAPIO_REMAPPED:
 686                 /* The target has remapped the I/O so dispatch it */
 687                 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
 688                                      blk_rq_pos(rq));
 689                 dm_dispatch_clone_request(clone, rq);
 690                 break;
 691         case DM_MAPIO_REQUEUE:
 692                 /* The target wants to requeue the I/O */
 693                 break;
 694         case DM_MAPIO_DELAY_REQUEUE:
 695                 /* The target wants to requeue the I/O after a delay */
 696                 dm_requeue_original_request(tio, true);
 697                 break;
 698         default:
 699                 if (r > 0) {
 700                         DMWARN("unimplemented target map return value: %d", r);
 701                         BUG();
 702                 }
 703
 704                 /* The target wants to complete the I/O */
 705                 dm_kill_unmapped_request(rq, r);
 706         }
 707
 708         return r;
 709 }
 710
 711 static void dm_start_request(struct mapped_device *md, struct request *orig)
 712 {
 713         if (!orig->q->mq_ops)
 714                 blk_start_request(orig);
 715         else
 716                 blk_mq_start_request(orig);
 717         atomic_inc(&md->pending[rq_data_dir(orig)]);
 718
 719         if (md->seq_rq_merge_deadline_usecs) {
 720                 md->last_rq_pos = rq_end_sector(orig);
 721                 md->last_rq_rw = rq_data_dir(orig);
 722                 md->last_rq_start_time = ktime_get();
 723         }
 724
 725         if (unlikely(dm_stats_used(&md->stats))) {
 726                 struct dm_rq_target_io *tio = tio_from_request(orig);
 727                 tio->duration_jiffies = jiffies;
 728                 tio->n_sectors = blk_rq_sectors(orig);
 729                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 730                                     blk_rq_pos(orig), tio->n_sectors, false, 0,
 731                                     &tio->stats_aux);
 732         }
 733
 734         /*
 735          * Hold the md reference here for the in-flight I/O.
 736          * We can't rely on the reference count by device opener,
 737          * because the device may be closed during the request completion
 738          * when all bios are completed.
 739          * See the comment in rq_completed() too.
 740          */
 741         dm_get(md);
 742 }
 743
 744 static void map_tio_request(struct kthread_work *work)
 745 {
 746         struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
 747
 748         if (map_request(tio) == DM_MAPIO_REQUEUE)
 749                 dm_requeue_original_request(tio, false);
 750 }
 751
 752 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
 753 {
 754         return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
 755 }
 756
 757 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
 758
 759 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
 760                                                      const char *buf, size_t count)
 761 {
 762         unsigned deadline;
 763
 764         if (dm_get_md_type(md) != DM_TYPE_REQUEST_BASED)
 765                 return count;
 766
 767         if (kstrtouint(buf, 10, &deadline))
 768                 return -EINVAL;
 769
 770         if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
 771                 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
 772
 773         md->seq_rq_merge_deadline_usecs = deadline;
 774
 775         return count;
 776 }
 777
 778 static bool dm_old_request_peeked_before_merge_deadline(struct mapped_device *md)
 779 {
 780         ktime_t kt_deadline;
 781
 782         if (!md->seq_rq_merge_deadline_usecs)
 783                 return false;
 784
 785         kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
 786         kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
 787
 788         return !ktime_after(ktime_get(), kt_deadline);
 789 }
 790
 791 /*
 792  * q->request_fn for old request-based dm.
 793  * Called with the queue lock held.
 794  */
 795 static void dm_old_request_fn(struct request_queue *q)
 796 {
 797         struct mapped_device *md = q->queuedata;
 798         struct dm_target *ti = md->immutable_target;
 799         struct request *rq;
 800         struct dm_rq_target_io *tio;
 801         sector_t pos = 0;
 802
 803         if (unlikely(!ti)) {
 804                 int srcu_idx;
 805                 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
 806
 807                 if (unlikely(!map)) {
 808                         dm_put_live_table(md, srcu_idx);
 809                         return;
 810                 }
 811                 ti = dm_table_find_target(map, pos);
 812                 dm_put_live_table(md, srcu_idx);
 813         }
 814
 815         /*
 816          * For suspend, check blk_queue_stopped() and increment
 817          * ->pending within a single queue_lock not to increment the
 818          * number of in-flight I/Os after the queue is stopped in
 819          * dm_suspend().
 820          */
 821         while (!blk_queue_stopped(q)) {
 822                 rq = blk_peek_request(q);
 823                 if (!rq)
 824                         return;
 825
 826                 /* always use block 0 to find the target for flushes for now */
 827                 pos = 0;
 828                 if (req_op(rq) != REQ_OP_FLUSH)
 829                         pos = blk_rq_pos(rq);
 830
 831                 if ((dm_old_request_peeked_before_merge_deadline(md) &&
 832                      md_in_flight(md) && rq->bio && rq->bio->bi_vcnt == 1 &&
 833                      md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
 834                     (ti->type->busy && ti->type->busy(ti))) {
 835                         blk_delay_queue(q, 10);
 836                         return;
 837                 }
 838
 839                 dm_start_request(md, rq);
 840
 841                 tio = tio_from_request(rq);
 842                 /* Establish tio->ti before queuing work (map_tio_request) */
 843                 tio->ti = ti;
 844                 kthread_queue_work(&md->kworker, &tio->work);
 845                 BUG_ON(!irqs_disabled());
 846         }
 847 }
 848
 849 /*
 850  * Fully initialize a .request_fn request-based queue.
 851  */
 852 int dm_old_init_request_queue(struct mapped_device *md)
 853 {
 854         /* Fully initialize the queue */
 855         if (!blk_init_allocated_queue(md->queue, dm_old_request_fn, NULL))
 856                 return -EINVAL;
 857
 858         /* disable dm_old_request_fn's merge heuristic by default */
 859         md->seq_rq_merge_deadline_usecs = 0;
 860
 861         dm_init_normal_md_queue(md);
 862         blk_queue_softirq_done(md->queue, dm_softirq_done);
 863         blk_queue_prep_rq(md->queue, dm_old_prep_fn);
 864
 865         /* Initialize the request-based DM worker thread */
 866         kthread_init_worker(&md->kworker);
 867         md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
 868                                        "kdmwork-%s", dm_device_name(md));
 869         if (IS_ERR(md->kworker_task)) {
 870                 int error = PTR_ERR(md->kworker_task);
 871                 md->kworker_task = NULL;
 872                 return error;
 873         }
 874
 875         elv_register_queue(md->queue);
 876
 877         return 0;
 878 }
 879
 880 static int dm_mq_init_request(void *data, struct request *rq,
 881                        unsigned int hctx_idx, unsigned int request_idx,
 882                        unsigned int numa_node)
 883 {
 884         struct mapped_device *md = data;
 885         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 886
 887         /*
 888          * Must initialize md member of tio, otherwise it won't
 889          * be available in dm_mq_queue_rq.
 890          */
 891         tio->md = md;
 892
 893         if (md->init_tio_pdu) {
 894                 /* target-specific per-io data is immediately after the tio */
 895                 tio->info.ptr = tio + 1;
 896         }
 897
 898         return 0;
 899 }
 900
 901 static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
 902                           const struct blk_mq_queue_data *bd)
 903 {
 904         struct request *rq = bd->rq;
 905         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 906         struct mapped_device *md = tio->md;
 907         struct dm_target *ti = md->immutable_target;
 908
 909         if (unlikely(!ti)) {
 910                 int srcu_idx;
 911                 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
 912
 913                 ti = dm_table_find_target(map, 0);
 914                 dm_put_live_table(md, srcu_idx);
 915         }
 916
 917         /*
 918          * On suspend dm_stop_queue() handles stopping the blk-mq
 919          * request_queue BUT: even though the hw_queues are marked
 920          * BLK_MQ_S_STOPPED at that point there is still a race that
 921          * is allowing block/blk-mq.c to call ->queue_rq against a
 922          * hctx that it really shouldn't.  The following check guards
 923          * against this rarity (albeit _not_ race-free).
 924          */
 925         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 926                 return BLK_MQ_RQ_QUEUE_BUSY;
 927
 928         if (ti->type->busy && ti->type->busy(ti))
 929                 return BLK_MQ_RQ_QUEUE_BUSY;
 930
 931         dm_start_request(md, rq);
 932
 933         /* Init tio using md established in .init_request */
 934         init_tio(tio, rq, md);
 935
 936         /*
 937          * Establish tio->ti before calling map_request().
 938          */
 939         tio->ti = ti;
 940
 941         /* Direct call is fine since .queue_rq allows allocations */
 942         if (map_request(tio) == DM_MAPIO_REQUEUE) {
 943                 /* Undo dm_start_request() before requeuing */
 944                 rq_end_stats(md, rq);
 945                 rq_completed(md, rq_data_dir(rq), false);
 946                 return BLK_MQ_RQ_QUEUE_BUSY;
 947         }
 948
 949         return BLK_MQ_RQ_QUEUE_OK;
 950 }
 951
 952 static struct blk_mq_ops dm_mq_ops = {
 953         .queue_rq = dm_mq_queue_rq,
 954         .complete = dm_softirq_done,
 955         .init_request = dm_mq_init_request,
 956 };
 957
 958 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
 959 {
 960         struct request_queue *q;
 961         struct dm_target *immutable_tgt;
 962         int err;
 963
 964         if (!dm_table_all_blk_mq_devices(t)) {
 965                 DMERR("request-based dm-mq may only be stacked on blk-mq device(s)");
 966                 return -EINVAL;
 967         }
 968
 969         md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
 970         if (!md->tag_set)
 971                 return -ENOMEM;
 972
 973         md->tag_set->ops = &dm_mq_ops;
 974         md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
 975         md->tag_set->numa_node = md->numa_node_id;
 976         md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
 977         md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
 978         md->tag_set->driver_data = md;
 979
 980         md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
 981         immutable_tgt = dm_table_get_immutable_target(t);
 982         if (immutable_tgt && immutable_tgt->per_io_data_size) {
 983                 /* any target-specific per-io data is immediately after the tio */
 984                 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
 985                 md->init_tio_pdu = true;
 986         }
 987
 988         err = blk_mq_alloc_tag_set(md->tag_set);
 989         if (err)
 990                 goto out_kfree_tag_set;
 991
 992         q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
 993         if (IS_ERR(q)) {
 994                 err = PTR_ERR(q);
 995                 goto out_tag_set;
 996         }
 997         dm_init_md_queue(md);
 998
 999         /* backfill 'mq' sysfs registration normally done in blk_register_queue */
1000         err = blk_mq_register_dev(disk_to_dev(md->disk), q);
1001         if (err)
1002                 goto out_cleanup_queue;
1003
1004         return 0;
1005
1006 out_cleanup_queue:
1007         blk_cleanup_queue(q);
1008 out_tag_set:
1009         blk_mq_free_tag_set(md->tag_set);
1010 out_kfree_tag_set:
1011         kfree(md->tag_set);
1012         md->tag_set = NULL;
1013
1014         return err;
1015 }
1016
1017 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
1018 {
1019         if (md->tag_set) {
1020                 blk_mq_free_tag_set(md->tag_set);
1021                 kfree(md->tag_set);
1022                 md->tag_set = NULL;
1023         }
1024 }
1025
1026 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
1027 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
1028
1029 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
1030 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
1031
1032 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
1033 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
1034
1035 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
1036 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");