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