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