drivers/md/dm-rq.c

   1 /*
   2  * Copyright (C) 2016 Red Hat, Inc. All rights reserved.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-core.h"
   8 #include "dm-rq.h"
   9
  10 #include <linux/elevator.h> /* for rq_end_sector() */
  11 #include <linux/blk-mq.h>
  12
  13 #define DM_MSG_PREFIX "core-rq"
  14
  15 #define DM_MQ_NR_HW_QUEUES 1
  16 #define DM_MQ_QUEUE_DEPTH 2048
  17 static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
  18 static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
  19
  20 /*
  21  * Request-based DM's mempools' reserved IOs set by the user.
  22  */
  23 #define RESERVED_REQUEST_BASED_IOS      256
  24 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
  25
  26 static bool use_blk_mq = IS_ENABLED(CONFIG_DM_MQ_DEFAULT);
  27
  28 bool dm_use_blk_mq_default(void)
  29 {
  30         return use_blk_mq;
  31 }
  32
  33 bool dm_use_blk_mq(struct mapped_device *md)
  34 {
  35         return md->use_blk_mq;
  36 }
  37 EXPORT_SYMBOL_GPL(dm_use_blk_mq);
  38
  39 unsigned dm_get_reserved_rq_based_ios(void)
  40 {
  41         return __dm_get_module_param(&reserved_rq_based_ios,
  42                                      RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
  43 }
  44 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
  45
  46 static unsigned dm_get_blk_mq_nr_hw_queues(void)
  47 {
  48         return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
  49 }
  50
  51 static unsigned dm_get_blk_mq_queue_depth(void)
  52 {
  53         return __dm_get_module_param(&dm_mq_queue_depth,
  54                                      DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
  55 }
  56
  57 int dm_request_based(struct mapped_device *md)
  58 {
  59         return blk_queue_stackable(md->queue);
  60 }
  61
  62 static void dm_old_start_queue(struct request_queue *q)
  63 {
  64         unsigned long flags;
  65
  66         spin_lock_irqsave(q->queue_lock, flags);
  67         if (blk_queue_stopped(q))
  68                 blk_start_queue(q);
  69         spin_unlock_irqrestore(q->queue_lock, flags);
  70 }
  71
  72 static void dm_mq_start_queue(struct request_queue *q)
  73 {
  74         blk_mq_unquiesce_queue(q);
  75         blk_mq_kick_requeue_list(q);
  76 }
  77
  78 void dm_start_queue(struct request_queue *q)
  79 {
  80         if (!q->mq_ops)
  81                 dm_old_start_queue(q);
  82         else
  83                 dm_mq_start_queue(q);
  84 }
  85
  86 static void dm_old_stop_queue(struct request_queue *q)
  87 {
  88         unsigned long flags;
  89
  90         spin_lock_irqsave(q->queue_lock, flags);
  91         if (!blk_queue_stopped(q))
  92                 blk_stop_queue(q);
  93         spin_unlock_irqrestore(q->queue_lock, flags);
  94 }
  95
  96 static void dm_mq_stop_queue(struct request_queue *q)
  97 {
  98         if (blk_mq_queue_stopped(q))
  99                 return;
 100
 101         blk_mq_quiesce_queue(q);
 102 }
 103
 104 void dm_stop_queue(struct request_queue *q)
 105 {
 106         if (!q->mq_ops)
 107                 dm_old_stop_queue(q);
 108         else
 109                 dm_mq_stop_queue(q);
 110 }
 111
 112 /*
 113  * Partial completion handling for request-based dm
 114  */
 115 static void end_clone_bio(struct bio *clone)
 116 {
 117         struct dm_rq_clone_bio_info *info =
 118                 container_of(clone, struct dm_rq_clone_bio_info, clone);
 119         struct dm_rq_target_io *tio = info->tio;
 120         unsigned int nr_bytes = info->orig->bi_iter.bi_size;
 121         blk_status_t error = clone->bi_status;
 122         bool is_last = !clone->bi_next;
 123
 124         bio_put(clone);
 125
 126         if (tio->error)
 127                 /*
 128                  * An error has already been detected on the request.
 129                  * Once error occurred, just let clone->end_io() handle
 130                  * the remainder.
 131                  */
 132                 return;
 133         else if (error) {
 134                 /*
 135                  * Don't notice the error to the upper layer yet.
 136                  * The error handling decision is made by the target driver,
 137                  * when the request is completed.
 138                  */
 139                 tio->error = error;
 140                 goto exit;
 141         }
 142
 143         /*
 144          * I/O for the bio successfully completed.
 145          * Notice the data completion to the upper layer.
 146          */
 147         tio->completed += nr_bytes;
 148
 149         /*
 150          * Update the original request.
 151          * Do not use blk_end_request() here, because it may complete
 152          * the original request before the clone, and break the ordering.
 153          */
 154         if (is_last)
 155  exit:
 156                 blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
 157 }
 158
 159 static struct dm_rq_target_io *tio_from_request(struct request *rq)
 160 {
 161         return blk_mq_rq_to_pdu(rq);
 162 }
 163
 164 static void rq_end_stats(struct mapped_device *md, struct request *orig)
 165 {
 166         if (unlikely(dm_stats_used(&md->stats))) {
 167                 struct dm_rq_target_io *tio = tio_from_request(orig);
 168                 tio->duration_jiffies = jiffies - tio->duration_jiffies;
 169                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 170                                     blk_rq_pos(orig), tio->n_sectors, true,
 171                                     tio->duration_jiffies, &tio->stats_aux);
 172         }
 173 }
 174
 175 /*
 176  * Don't touch any member of the md after calling this function because
 177  * the md may be freed in dm_put() at the end of this function.
 178  * Or do dm_get() before calling this function and dm_put() later.
 179  */
 180 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
 181 {
 182         struct request_queue *q = md->queue;
 183         unsigned long flags;
 184
 185         atomic_dec(&md->pending[rw]);
 186
 187         /* nudge anyone waiting on suspend queue */
 188         if (!md_in_flight(md))
 189                 wake_up(&md->wait);
 190
 191         /*
 192          * Run this off this callpath, as drivers could invoke end_io while
 193          * inside their request_fn (and holding the queue lock). Calling
 194          * back into ->request_fn() could deadlock attempting to grab the
 195          * queue lock again.
 196          */
 197         if (!q->mq_ops && run_queue) {
 198                 spin_lock_irqsave(q->queue_lock, flags);
 199                 blk_run_queue_async(q);
 200                 spin_unlock_irqrestore(q->queue_lock, flags);
 201         }
 202
 203         /*
 204          * dm_put() must be at the end of this function. See the comment above
 205          */
 206         dm_put(md);
 207 }
 208
 209 /*
 210  * Complete the clone and the original request.
 211  * Must be called without clone's queue lock held,
 212  * see end_clone_request() for more details.
 213  */
 214 static void dm_end_request(struct request *clone, blk_status_t error)
 215 {
 216         int rw = rq_data_dir(clone);
 217         struct dm_rq_target_io *tio = clone->end_io_data;
 218         struct mapped_device *md = tio->md;
 219         struct request *rq = tio->orig;
 220
 221         blk_rq_unprep_clone(clone);
 222         tio->ti->type->release_clone_rq(clone);
 223
 224         rq_end_stats(md, rq);
 225         if (!rq->q->mq_ops)
 226                 blk_end_request_all(rq, error);
 227         else
 228                 blk_mq_end_request(rq, error);
 229         rq_completed(md, rw, true);
 230 }
 231
 232 /*
 233  * Requeue the original request of a clone.
 234  */
 235 static void dm_old_requeue_request(struct request *rq, unsigned long delay_ms)
 236 {
 237         struct request_queue *q = rq->q;
 238         unsigned long flags;
 239
 240         spin_lock_irqsave(q->queue_lock, flags);
 241         blk_requeue_request(q, rq);
 242         blk_delay_queue(q, delay_ms);
 243         spin_unlock_irqrestore(q->queue_lock, flags);
 244 }
 245
 246 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
 247 {
 248         blk_mq_delay_kick_requeue_list(q, msecs);
 249 }
 250
 251 void dm_mq_kick_requeue_list(struct mapped_device *md)
 252 {
 253         __dm_mq_kick_requeue_list(dm_get_md_queue(md), 0);
 254 }
 255 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
 256
 257 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
 258 {
 259         blk_mq_requeue_request(rq, false);
 260         __dm_mq_kick_requeue_list(rq->q, msecs);
 261 }
 262
 263 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
 264 {
 265         struct mapped_device *md = tio->md;
 266         struct request *rq = tio->orig;
 267         int rw = rq_data_dir(rq);
 268         unsigned long delay_ms = delay_requeue ? 100 : 0;
 269
 270         rq_end_stats(md, rq);
 271         if (tio->clone) {
 272                 blk_rq_unprep_clone(tio->clone);
 273                 tio->ti->type->release_clone_rq(tio->clone);
 274         }
 275
 276         if (!rq->q->mq_ops)
 277                 dm_old_requeue_request(rq, delay_ms);
 278         else
 279                 dm_mq_delay_requeue_request(rq, delay_ms);
 280
 281         rq_completed(md, rw, false);
 282 }
 283
 284 static void dm_done(struct request *clone, blk_status_t error, bool mapped)
 285 {
 286         int r = DM_ENDIO_DONE;
 287         struct dm_rq_target_io *tio = clone->end_io_data;
 288         dm_request_endio_fn rq_end_io = NULL;
 289
 290         if (tio->ti) {
 291                 rq_end_io = tio->ti->type->rq_end_io;
 292
 293                 if (mapped && rq_end_io)
 294                         r = rq_end_io(tio->ti, clone, error, &tio->info);
 295         }
 296
 297         if (unlikely(error == BLK_STS_TARGET)) {
 298                 if (req_op(clone) == REQ_OP_WRITE_SAME &&
 299                     !clone->q->limits.max_write_same_sectors)
 300                         disable_write_same(tio->md);
 301                 if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
 302                     !clone->q->limits.max_write_zeroes_sectors)
 303                         disable_write_zeroes(tio->md);
 304         }
 305
 306         switch (r) {
 307         case DM_ENDIO_DONE:
 308                 /* The target wants to complete the I/O */
 309                 dm_end_request(clone, error);
 310                 break;
 311         case DM_ENDIO_INCOMPLETE:
 312                 /* The target will handle the I/O */
 313                 return;
 314         case DM_ENDIO_REQUEUE:
 315                 /* The target wants to requeue the I/O */
 316                 dm_requeue_original_request(tio, false);
 317                 break;
 318         default:
 319                 DMWARN("unimplemented target endio return value: %d", r);
 320                 BUG();
 321         }
 322 }
 323
 324 /*
 325  * Request completion handler for request-based dm
 326  */
 327 static void dm_softirq_done(struct request *rq)
 328 {
 329         bool mapped = true;
 330         struct dm_rq_target_io *tio = tio_from_request(rq);
 331         struct request *clone = tio->clone;
 332         int rw;
 333
 334         if (!clone) {
 335                 struct mapped_device *md = tio->md;
 336
 337                 rq_end_stats(md, rq);
 338                 rw = rq_data_dir(rq);
 339                 if (!rq->q->mq_ops)
 340                         blk_end_request_all(rq, tio->error);
 341                 else
 342                         blk_mq_end_request(rq, tio->error);
 343                 rq_completed(md, rw, false);
 344                 return;
 345         }
 346
 347         if (rq->rq_flags & RQF_FAILED)
 348                 mapped = false;
 349
 350         dm_done(clone, tio->error, mapped);
 351 }
 352
 353 /*
 354  * Complete the clone and the original request with the error status
 355  * through softirq context.
 356  */
 357 static void dm_complete_request(struct request *rq, blk_status_t error)
 358 {
 359         struct dm_rq_target_io *tio = tio_from_request(rq);
 360
 361         tio->error = error;
 362         if (!rq->q->mq_ops)
 363                 blk_complete_request(rq);
 364         else
 365                 blk_mq_complete_request(rq);
 366 }
 367
 368 /*
 369  * Complete the not-mapped clone and the original request with the error status
 370  * through softirq context.
 371  * Target's rq_end_io() function isn't called.
 372  * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
 373  */
 374 static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
 375 {
 376         rq->rq_flags |= RQF_FAILED;
 377         dm_complete_request(rq, error);
 378 }
 379
 380 /*
 381  * Called with the clone's queue lock held (in the case of .request_fn)
 382  */
 383 static void end_clone_request(struct request *clone, blk_status_t error)
 384 {
 385         struct dm_rq_target_io *tio = clone->end_io_data;
 386
 387         /*
 388          * Actual request completion is done in a softirq context which doesn't
 389          * hold the clone's queue lock.  Otherwise, deadlock could occur because:
 390          *     - another request may be submitted by the upper level driver
 391          *       of the stacking during the completion
 392          *     - the submission which requires queue lock may be done
 393          *       against this clone's queue
 394          */
 395         dm_complete_request(tio->orig, error);
 396 }
 397
 398 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
 399 {
 400         blk_status_t r;
 401
 402         if (blk_queue_io_stat(clone->q))
 403                 clone->rq_flags |= RQF_IO_STAT;
 404
 405         clone->start_time = jiffies;
 406         r = blk_insert_cloned_request(clone->q, clone);
 407         if (r)
 408                 /* must complete clone in terms of original request */
 409                 dm_complete_request(rq, r);
 410 }
 411
 412 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
 413                                  void *data)
 414 {
 415         struct dm_rq_target_io *tio = data;
 416         struct dm_rq_clone_bio_info *info =
 417                 container_of(bio, struct dm_rq_clone_bio_info, clone);
 418
 419         info->orig = bio_orig;
 420         info->tio = tio;
 421         bio->bi_end_io = end_clone_bio;
 422
 423         return 0;
 424 }
 425
 426 static int setup_clone(struct request *clone, struct request *rq,
 427                        struct dm_rq_target_io *tio, gfp_t gfp_mask)
 428 {
 429         int r;
 430
 431         r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
 432                               dm_rq_bio_constructor, tio);
 433         if (r)
 434                 return r;
 435
 436         clone->end_io = end_clone_request;
 437         clone->end_io_data = tio;
 438
 439         tio->clone = clone;
 440
 441         return 0;
 442 }
 443
 444 static void map_tio_request(struct kthread_work *work);
 445
 446 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
 447                      struct mapped_device *md)
 448 {
 449         tio->md = md;
 450         tio->ti = NULL;
 451         tio->clone = NULL;
 452         tio->orig = rq;
 453         tio->error = 0;
 454         tio->completed = 0;
 455         /*
 456          * Avoid initializing info for blk-mq; it passes
 457          * target-specific data through info.ptr
 458          * (see: dm_mq_init_request)
 459          */
 460         if (!md->init_tio_pdu)
 461                 memset(&tio->info, 0, sizeof(tio->info));
 462         if (md->kworker_task)
 463                 kthread_init_work(&tio->work, map_tio_request);
 464 }
 465
 466 /*
 467  * Returns:
 468  * DM_MAPIO_*       : the request has been processed as indicated
 469  * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
 470  * < 0              : the request was completed due to failure
 471  */
 472 static int map_request(struct dm_rq_target_io *tio)
 473 {
 474         int r;
 475         struct dm_target *ti = tio->ti;
 476         struct mapped_device *md = tio->md;
 477         struct request *rq = tio->orig;
 478         struct request *clone = NULL;
 479
 480         r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
 481         switch (r) {
 482         case DM_MAPIO_SUBMITTED:
 483                 /* The target has taken the I/O to submit by itself later */
 484                 break;
 485         case DM_MAPIO_REMAPPED:
 486                 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
 487                         /* -ENOMEM */
 488                         ti->type->release_clone_rq(clone);
 489                         return DM_MAPIO_REQUEUE;
 490                 }
 491
 492                 /* The target has remapped the I/O so dispatch it */
 493                 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
 494                                      blk_rq_pos(rq));
 495                 dm_dispatch_clone_request(clone, rq);
 496                 break;
 497         case DM_MAPIO_REQUEUE:
 498                 /* The target wants to requeue the I/O */
 499                 break;
 500         case DM_MAPIO_DELAY_REQUEUE:
 501                 /* The target wants to requeue the I/O after a delay */
 502                 dm_requeue_original_request(tio, true);
 503                 break;
 504         case DM_MAPIO_KILL:
 505                 /* The target wants to complete the I/O */
 506                 dm_kill_unmapped_request(rq, BLK_STS_IOERR);
 507                 break;
 508         default:
 509                 DMWARN("unimplemented target map return value: %d", r);
 510                 BUG();
 511         }
 512
 513         return r;
 514 }
 515
 516 static void dm_start_request(struct mapped_device *md, struct request *orig)
 517 {
 518         if (!orig->q->mq_ops)
 519                 blk_start_request(orig);
 520         else
 521                 blk_mq_start_request(orig);
 522         atomic_inc(&md->pending[rq_data_dir(orig)]);
 523
 524         if (md->seq_rq_merge_deadline_usecs) {
 525                 md->last_rq_pos = rq_end_sector(orig);
 526                 md->last_rq_rw = rq_data_dir(orig);
 527                 md->last_rq_start_time = ktime_get();
 528         }
 529
 530         if (unlikely(dm_stats_used(&md->stats))) {
 531                 struct dm_rq_target_io *tio = tio_from_request(orig);
 532                 tio->duration_jiffies = jiffies;
 533                 tio->n_sectors = blk_rq_sectors(orig);
 534                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 535                                     blk_rq_pos(orig), tio->n_sectors, false, 0,
 536                                     &tio->stats_aux);
 537         }
 538
 539         /*
 540          * Hold the md reference here for the in-flight I/O.
 541          * We can't rely on the reference count by device opener,
 542          * because the device may be closed during the request completion
 543          * when all bios are completed.
 544          * See the comment in rq_completed() too.
 545          */
 546         dm_get(md);
 547 }
 548
 549 static int __dm_rq_init_rq(struct mapped_device *md, struct request *rq)
 550 {
 551         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 552
 553         /*
 554          * Must initialize md member of tio, otherwise it won't
 555          * be available in dm_mq_queue_rq.
 556          */
 557         tio->md = md;
 558
 559         if (md->init_tio_pdu) {
 560                 /* target-specific per-io data is immediately after the tio */
 561                 tio->info.ptr = tio + 1;
 562         }
 563
 564         return 0;
 565 }
 566
 567 static int dm_rq_init_rq(struct request_queue *q, struct request *rq, gfp_t gfp)
 568 {
 569         return __dm_rq_init_rq(q->rq_alloc_data, rq);
 570 }
 571
 572 static void map_tio_request(struct kthread_work *work)
 573 {
 574         struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
 575
 576         if (map_request(tio) == DM_MAPIO_REQUEUE)
 577                 dm_requeue_original_request(tio, false);
 578 }
 579
 580 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
 581 {
 582         return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
 583 }
 584
 585 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
 586
 587 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
 588                                                      const char *buf, size_t count)
 589 {
 590         unsigned deadline;
 591
 592         if (dm_get_md_type(md) != DM_TYPE_REQUEST_BASED)
 593                 return count;
 594
 595         if (kstrtouint(buf, 10, &deadline))
 596                 return -EINVAL;
 597
 598         if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
 599                 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
 600
 601         md->seq_rq_merge_deadline_usecs = deadline;
 602
 603         return count;
 604 }
 605
 606 static bool dm_old_request_peeked_before_merge_deadline(struct mapped_device *md)
 607 {
 608         ktime_t kt_deadline;
 609
 610         if (!md->seq_rq_merge_deadline_usecs)
 611                 return false;
 612
 613         kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
 614         kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
 615
 616         return !ktime_after(ktime_get(), kt_deadline);
 617 }
 618
 619 /*
 620  * q->request_fn for old request-based dm.
 621  * Called with the queue lock held.
 622  */
 623 static void dm_old_request_fn(struct request_queue *q)
 624 {
 625         struct mapped_device *md = q->queuedata;
 626         struct dm_target *ti = md->immutable_target;
 627         struct request *rq;
 628         struct dm_rq_target_io *tio;
 629         sector_t pos = 0;
 630
 631         if (unlikely(!ti)) {
 632                 int srcu_idx;
 633                 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
 634
 635                 if (unlikely(!map)) {
 636                         dm_put_live_table(md, srcu_idx);
 637                         return;
 638                 }
 639                 ti = dm_table_find_target(map, pos);
 640                 dm_put_live_table(md, srcu_idx);
 641         }
 642
 643         /*
 644          * For suspend, check blk_queue_stopped() and increment
 645          * ->pending within a single queue_lock not to increment the
 646          * number of in-flight I/Os after the queue is stopped in
 647          * dm_suspend().
 648          */
 649         while (!blk_queue_stopped(q)) {
 650                 rq = blk_peek_request(q);
 651                 if (!rq)
 652                         return;
 653
 654                 /* always use block 0 to find the target for flushes for now */
 655                 pos = 0;
 656                 if (req_op(rq) != REQ_OP_FLUSH)
 657                         pos = blk_rq_pos(rq);
 658
 659                 if ((dm_old_request_peeked_before_merge_deadline(md) &&
 660                      md_in_flight(md) && rq->bio && !bio_multiple_segments(rq->bio) &&
 661                      md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
 662                     (ti->type->busy && ti->type->busy(ti))) {
 663                         blk_delay_queue(q, 10);
 664                         return;
 665                 }
 666
 667                 dm_start_request(md, rq);
 668
 669                 tio = tio_from_request(rq);
 670                 init_tio(tio, rq, md);
 671                 /* Establish tio->ti before queuing work (map_tio_request) */
 672                 tio->ti = ti;
 673                 kthread_queue_work(&md->kworker, &tio->work);
 674                 BUG_ON(!irqs_disabled());
 675         }
 676 }
 677
 678 /*
 679  * Fully initialize a .request_fn request-based queue.
 680  */
 681 int dm_old_init_request_queue(struct mapped_device *md, struct dm_table *t)
 682 {
 683         struct dm_target *immutable_tgt;
 684
 685         /* Fully initialize the queue */
 686         md->queue->cmd_size = sizeof(struct dm_rq_target_io);
 687         md->queue->rq_alloc_data = md;
 688         md->queue->request_fn = dm_old_request_fn;
 689         md->queue->init_rq_fn = dm_rq_init_rq;
 690
 691         immutable_tgt = dm_table_get_immutable_target(t);
 692         if (immutable_tgt && immutable_tgt->per_io_data_size) {
 693                 /* any target-specific per-io data is immediately after the tio */
 694                 md->queue->cmd_size += immutable_tgt->per_io_data_size;
 695                 md->init_tio_pdu = true;
 696         }
 697         if (blk_init_allocated_queue(md->queue) < 0)
 698                 return -EINVAL;
 699
 700         /* disable dm_old_request_fn's merge heuristic by default */
 701         md->seq_rq_merge_deadline_usecs = 0;
 702
 703         dm_init_normal_md_queue(md);
 704         blk_queue_softirq_done(md->queue, dm_softirq_done);
 705
 706         /* Initialize the request-based DM worker thread */
 707         kthread_init_worker(&md->kworker);
 708         md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
 709                                        "kdmwork-%s", dm_device_name(md));
 710         if (IS_ERR(md->kworker_task)) {
 711                 int error = PTR_ERR(md->kworker_task);
 712                 md->kworker_task = NULL;
 713                 return error;
 714         }
 715
 716         elv_register_queue(md->queue);
 717
 718         return 0;
 719 }
 720
 721 static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
 722                 unsigned int hctx_idx, unsigned int numa_node)
 723 {
 724         return __dm_rq_init_rq(set->driver_data, rq);
 725 }
 726
 727 static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
 728                           const struct blk_mq_queue_data *bd)
 729 {
 730         struct request *rq = bd->rq;
 731         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 732         struct mapped_device *md = tio->md;
 733         struct dm_target *ti = md->immutable_target;
 734
 735         if (unlikely(!ti)) {
 736                 int srcu_idx;
 737                 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
 738
 739                 ti = dm_table_find_target(map, 0);
 740                 dm_put_live_table(md, srcu_idx);
 741         }
 742
 743         if (ti->type->busy && ti->type->busy(ti))
 744                 return BLK_STS_RESOURCE;
 745
 746         dm_start_request(md, rq);
 747
 748         /* Init tio using md established in .init_request */
 749         init_tio(tio, rq, md);
 750
 751         /*
 752          * Establish tio->ti before calling map_request().
 753          */
 754         tio->ti = ti;
 755
 756         /* Direct call is fine since .queue_rq allows allocations */
 757         if (map_request(tio) == DM_MAPIO_REQUEUE) {
 758                 /* Undo dm_start_request() before requeuing */
 759                 rq_end_stats(md, rq);
 760                 rq_completed(md, rq_data_dir(rq), false);
 761                 blk_mq_delay_run_hw_queue(hctx, 100/*ms*/);
 762                 return BLK_STS_RESOURCE;
 763         }
 764
 765         return BLK_STS_OK;
 766 }
 767
 768 static const struct blk_mq_ops dm_mq_ops = {
 769         .queue_rq = dm_mq_queue_rq,
 770         .complete = dm_softirq_done,
 771         .init_request = dm_mq_init_request,
 772 };
 773
 774 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
 775 {
 776         struct request_queue *q;
 777         struct dm_target *immutable_tgt;
 778         int err;
 779
 780         if (!dm_table_all_blk_mq_devices(t)) {
 781                 DMERR("request-based dm-mq may only be stacked on blk-mq device(s)");
 782                 return -EINVAL;
 783         }
 784
 785         md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
 786         if (!md->tag_set)
 787                 return -ENOMEM;
 788
 789         md->tag_set->ops = &dm_mq_ops;
 790         md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
 791         md->tag_set->numa_node = md->numa_node_id;
 792         md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
 793         md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
 794         md->tag_set->driver_data = md;
 795
 796         md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
 797         immutable_tgt = dm_table_get_immutable_target(t);
 798         if (immutable_tgt && immutable_tgt->per_io_data_size) {
 799                 /* any target-specific per-io data is immediately after the tio */
 800                 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
 801                 md->init_tio_pdu = true;
 802         }
 803
 804         err = blk_mq_alloc_tag_set(md->tag_set);
 805         if (err)
 806                 goto out_kfree_tag_set;
 807
 808         q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
 809         if (IS_ERR(q)) {
 810                 err = PTR_ERR(q);
 811                 goto out_tag_set;
 812         }
 813         dm_init_md_queue(md);
 814
 815         /* backfill 'mq' sysfs registration normally done in blk_register_queue */
 816         err = blk_mq_register_dev(disk_to_dev(md->disk), q);
 817         if (err)
 818                 goto out_cleanup_queue;
 819
 820         return 0;
 821
 822 out_cleanup_queue:
 823         blk_cleanup_queue(q);
 824 out_tag_set:
 825         blk_mq_free_tag_set(md->tag_set);
 826 out_kfree_tag_set:
 827         kfree(md->tag_set);
 828
 829         return err;
 830 }
 831
 832 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
 833 {
 834         if (md->tag_set) {
 835                 blk_mq_free_tag_set(md->tag_set);
 836                 kfree(md->tag_set);
 837         }
 838 }
 839
 840 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
 841 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
 842
 843 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
 844 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
 845
 846 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
 847 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
 848
 849 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
 850 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");