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