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