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 /*
  16  * One of these is allocated per request.
  17  */
  18 struct dm_rq_target_io {
  19         struct mapped_device *md;
  20         struct dm_target *ti;
  21         struct request *orig, *clone;
  22         struct kthread_work work;
  23         blk_status_t error;
  24         union map_info info;
  25         struct dm_stats_aux stats_aux;
  26         unsigned long duration_jiffies;
  27         unsigned n_sectors;
  28         unsigned completed;
  29 };
  30
  31 #define DM_MQ_NR_HW_QUEUES 1
  32 #define DM_MQ_QUEUE_DEPTH 2048
  33 static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
  34 static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
  35
  36 /*
  37  * Request-based DM's mempools' reserved IOs set by the user.
  38  */
  39 #define RESERVED_REQUEST_BASED_IOS      256
  40 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
  41
  42 unsigned dm_get_reserved_rq_based_ios(void)
  43 {
  44         return __dm_get_module_param(&reserved_rq_based_ios,
  45                                      RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
  46 }
  47 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
  48
  49 static unsigned dm_get_blk_mq_nr_hw_queues(void)
  50 {
  51         return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
  52 }
  53
  54 static unsigned dm_get_blk_mq_queue_depth(void)
  55 {
  56         return __dm_get_module_param(&dm_mq_queue_depth,
  57                                      DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
  58 }
  59
  60 int dm_request_based(struct mapped_device *md)
  61 {
  62         return queue_is_mq(md->queue);
  63 }
  64
  65 void dm_start_queue(struct request_queue *q)
  66 {
  67         blk_mq_unquiesce_queue(q);
  68         blk_mq_kick_requeue_list(q);
  69 }
  70
  71 void dm_stop_queue(struct request_queue *q)
  72 {
  73         blk_mq_quiesce_queue(q);
  74 }
  75
  76 /*
  77  * Partial completion handling for request-based dm
  78  */
  79 static void end_clone_bio(struct bio *clone)
  80 {
  81         struct dm_rq_clone_bio_info *info =
  82                 container_of(clone, struct dm_rq_clone_bio_info, clone);
  83         struct dm_rq_target_io *tio = info->tio;
  84         unsigned int nr_bytes = info->orig->bi_iter.bi_size;
  85         blk_status_t error = clone->bi_status;
  86         bool is_last = !clone->bi_next;
  87
  88         bio_put(clone);
  89
  90         if (tio->error)
  91                 /*
  92                  * An error has already been detected on the request.
  93                  * Once error occurred, just let clone->end_io() handle
  94                  * the remainder.
  95                  */
  96                 return;
  97         else if (error) {
  98                 /*
  99                  * Don't notice the error to the upper layer yet.
 100                  * The error handling decision is made by the target driver,
 101                  * when the request is completed.
 102                  */
 103                 tio->error = error;
 104                 goto exit;
 105         }
 106
 107         /*
 108          * I/O for the bio successfully completed.
 109          * Notice the data completion to the upper layer.
 110          */
 111         tio->completed += nr_bytes;
 112
 113         /*
 114          * Update the original request.
 115          * Do not use blk_mq_end_request() here, because it may complete
 116          * the original request before the clone, and break the ordering.
 117          */
 118         if (is_last)
 119  exit:
 120                 blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
 121 }
 122
 123 static struct dm_rq_target_io *tio_from_request(struct request *rq)
 124 {
 125         return blk_mq_rq_to_pdu(rq);
 126 }
 127
 128 static void rq_end_stats(struct mapped_device *md, struct request *orig)
 129 {
 130         if (unlikely(dm_stats_used(&md->stats))) {
 131                 struct dm_rq_target_io *tio = tio_from_request(orig);
 132                 tio->duration_jiffies = jiffies - tio->duration_jiffies;
 133                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 134                                     blk_rq_pos(orig), tio->n_sectors, true,
 135                                     tio->duration_jiffies, &tio->stats_aux);
 136         }
 137 }
 138
 139 /*
 140  * Don't touch any member of the md after calling this function because
 141  * the md may be freed in dm_put() at the end of this function.
 142  * Or do dm_get() before calling this function and dm_put() later.
 143  */
 144 static void rq_completed(struct mapped_device *md)
 145 {
 146         /*
 147          * dm_put() must be at the end of this function. See the comment above
 148          */
 149         dm_put(md);
 150 }
 151
 152 /*
 153  * Complete the clone and the original request.
 154  * Must be called without clone's queue lock held,
 155  * see end_clone_request() for more details.
 156  */
 157 static void dm_end_request(struct request *clone, blk_status_t error)
 158 {
 159         struct dm_rq_target_io *tio = clone->end_io_data;
 160         struct mapped_device *md = tio->md;
 161         struct request *rq = tio->orig;
 162
 163         blk_rq_unprep_clone(clone);
 164         tio->ti->type->release_clone_rq(clone, NULL);
 165
 166         rq_end_stats(md, rq);
 167         blk_mq_end_request(rq, error);
 168         rq_completed(md);
 169 }
 170
 171 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
 172 {
 173         blk_mq_delay_kick_requeue_list(q, msecs);
 174 }
 175
 176 void dm_mq_kick_requeue_list(struct mapped_device *md)
 177 {
 178         __dm_mq_kick_requeue_list(md->queue, 0);
 179 }
 180 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
 181
 182 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
 183 {
 184         blk_mq_requeue_request(rq, false);
 185         __dm_mq_kick_requeue_list(rq->q, msecs);
 186 }
 187
 188 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
 189 {
 190         struct mapped_device *md = tio->md;
 191         struct request *rq = tio->orig;
 192         unsigned long delay_ms = delay_requeue ? 100 : 0;
 193
 194         rq_end_stats(md, rq);
 195         if (tio->clone) {
 196                 blk_rq_unprep_clone(tio->clone);
 197                 tio->ti->type->release_clone_rq(tio->clone, NULL);
 198         }
 199
 200         dm_mq_delay_requeue_request(rq, delay_ms);
 201         rq_completed(md);
 202 }
 203
 204 static void dm_done(struct request *clone, blk_status_t error, bool mapped)
 205 {
 206         int r = DM_ENDIO_DONE;
 207         struct dm_rq_target_io *tio = clone->end_io_data;
 208         dm_request_endio_fn rq_end_io = NULL;
 209
 210         if (tio->ti) {
 211                 rq_end_io = tio->ti->type->rq_end_io;
 212
 213                 if (mapped && rq_end_io)
 214                         r = rq_end_io(tio->ti, clone, error, &tio->info);
 215         }
 216
 217         if (unlikely(error == BLK_STS_TARGET)) {
 218                 if (req_op(clone) == REQ_OP_DISCARD &&
 219                     !clone->q->limits.max_discard_sectors)
 220                         disable_discard(tio->md);
 221                 else if (req_op(clone) == REQ_OP_WRITE_SAME &&
 222                          !clone->q->limits.max_write_same_sectors)
 223                         disable_write_same(tio->md);
 224                 else if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
 225                          !clone->q->limits.max_write_zeroes_sectors)
 226                         disable_write_zeroes(tio->md);
 227         }
 228
 229         switch (r) {
 230         case DM_ENDIO_DONE:
 231                 /* The target wants to complete the I/O */
 232                 dm_end_request(clone, error);
 233                 break;
 234         case DM_ENDIO_INCOMPLETE:
 235                 /* The target will handle the I/O */
 236                 return;
 237         case DM_ENDIO_REQUEUE:
 238                 /* The target wants to requeue the I/O */
 239                 dm_requeue_original_request(tio, false);
 240                 break;
 241         case DM_ENDIO_DELAY_REQUEUE:
 242                 /* The target wants to requeue the I/O after a delay */
 243                 dm_requeue_original_request(tio, true);
 244                 break;
 245         default:
 246                 DMWARN("unimplemented target endio return value: %d", r);
 247                 BUG();
 248         }
 249 }
 250
 251 /*
 252  * Request completion handler for request-based dm
 253  */
 254 static void dm_softirq_done(struct request *rq)
 255 {
 256         bool mapped = true;
 257         struct dm_rq_target_io *tio = tio_from_request(rq);
 258         struct request *clone = tio->clone;
 259
 260         if (!clone) {
 261                 struct mapped_device *md = tio->md;
 262
 263                 rq_end_stats(md, rq);
 264                 blk_mq_end_request(rq, tio->error);
 265                 rq_completed(md);
 266                 return;
 267         }
 268
 269         if (rq->rq_flags & RQF_FAILED)
 270                 mapped = false;
 271
 272         dm_done(clone, tio->error, mapped);
 273 }
 274
 275 /*
 276  * Complete the clone and the original request with the error status
 277  * through softirq context.
 278  */
 279 static void dm_complete_request(struct request *rq, blk_status_t error)
 280 {
 281         struct dm_rq_target_io *tio = tio_from_request(rq);
 282
 283         tio->error = error;
 284         if (likely(!blk_should_fake_timeout(rq->q)))
 285                 blk_mq_complete_request(rq);
 286 }
 287
 288 /*
 289  * Complete the not-mapped clone and the original request with the error status
 290  * through softirq context.
 291  * Target's rq_end_io() function isn't called.
 292  * This may be used when the target's clone_and_map_rq() function fails.
 293  */
 294 static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
 295 {
 296         rq->rq_flags |= RQF_FAILED;
 297         dm_complete_request(rq, error);
 298 }
 299
 300 static void end_clone_request(struct request *clone, blk_status_t error)
 301 {
 302         struct dm_rq_target_io *tio = clone->end_io_data;
 303
 304         dm_complete_request(tio->orig, error);
 305 }
 306
 307 static blk_status_t dm_dispatch_clone_request(struct request *clone, struct request *rq)
 308 {
 309         blk_status_t r;
 310
 311         if (blk_queue_io_stat(clone->q))
 312                 clone->rq_flags |= RQF_IO_STAT;
 313
 314         clone->start_time_ns = ktime_get_ns();
 315         r = blk_insert_cloned_request(clone->q, clone);
 316         if (r != BLK_STS_OK && r != BLK_STS_RESOURCE && r != BLK_STS_DEV_RESOURCE)
 317                 /* must complete clone in terms of original request */
 318                 dm_complete_request(rq, r);
 319         return r;
 320 }
 321
 322 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
 323                                  void *data)
 324 {
 325         struct dm_rq_target_io *tio = data;
 326         struct dm_rq_clone_bio_info *info =
 327                 container_of(bio, struct dm_rq_clone_bio_info, clone);
 328
 329         info->orig = bio_orig;
 330         info->tio = tio;
 331         bio->bi_end_io = end_clone_bio;
 332
 333         return 0;
 334 }
 335
 336 static int setup_clone(struct request *clone, struct request *rq,
 337                        struct dm_rq_target_io *tio, gfp_t gfp_mask)
 338 {
 339         int r;
 340
 341         r = blk_rq_prep_clone(clone, rq, &tio->md->bs, gfp_mask,
 342                               dm_rq_bio_constructor, tio);
 343         if (r)
 344                 return r;
 345
 346         clone->end_io = end_clone_request;
 347         clone->end_io_data = tio;
 348
 349         tio->clone = clone;
 350
 351         return 0;
 352 }
 353
 354 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
 355                      struct mapped_device *md)
 356 {
 357         tio->md = md;
 358         tio->ti = NULL;
 359         tio->clone = NULL;
 360         tio->orig = rq;
 361         tio->error = 0;
 362         tio->completed = 0;
 363         /*
 364          * Avoid initializing info for blk-mq; it passes
 365          * target-specific data through info.ptr
 366          * (see: dm_mq_init_request)
 367          */
 368         if (!md->init_tio_pdu)
 369                 memset(&tio->info, 0, sizeof(tio->info));
 370 }
 371
 372 /*
 373  * Returns:
 374  * DM_MAPIO_*       : the request has been processed as indicated
 375  * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
 376  * < 0              : the request was completed due to failure
 377  */
 378 static int map_request(struct dm_rq_target_io *tio)
 379 {
 380         int r;
 381         struct dm_target *ti = tio->ti;
 382         struct mapped_device *md = tio->md;
 383         struct request *rq = tio->orig;
 384         struct request *clone = NULL;
 385         blk_status_t ret;
 386
 387         r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
 388         switch (r) {
 389         case DM_MAPIO_SUBMITTED:
 390                 /* The target has taken the I/O to submit by itself later */
 391                 break;
 392         case DM_MAPIO_REMAPPED:
 393                 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
 394                         /* -ENOMEM */
 395                         ti->type->release_clone_rq(clone, &tio->info);
 396                         return DM_MAPIO_REQUEUE;
 397                 }
 398
 399                 /* The target has remapped the I/O so dispatch it */
 400                 trace_block_rq_remap(clone, disk_devt(dm_disk(md)),
 401                                      blk_rq_pos(rq));
 402                 ret = dm_dispatch_clone_request(clone, rq);
 403                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
 404                         blk_rq_unprep_clone(clone);
 405                         blk_mq_cleanup_rq(clone);
 406                         tio->ti->type->release_clone_rq(clone, &tio->info);
 407                         tio->clone = NULL;
 408                         return DM_MAPIO_REQUEUE;
 409                 }
 410                 break;
 411         case DM_MAPIO_REQUEUE:
 412                 /* The target wants to requeue the I/O */
 413                 break;
 414         case DM_MAPIO_DELAY_REQUEUE:
 415                 /* The target wants to requeue the I/O after a delay */
 416                 dm_requeue_original_request(tio, true);
 417                 break;
 418         case DM_MAPIO_KILL:
 419                 /* The target wants to complete the I/O */
 420                 dm_kill_unmapped_request(rq, BLK_STS_IOERR);
 421                 break;
 422         default:
 423                 DMWARN("unimplemented target map return value: %d", r);
 424                 BUG();
 425         }
 426
 427         return r;
 428 }
 429
 430 /* DEPRECATED: previously used for request-based merge heuristic in dm_request_fn() */
 431 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
 432 {
 433         return sprintf(buf, "%u\n", 0);
 434 }
 435
 436 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
 437                                                      const char *buf, size_t count)
 438 {
 439         return count;
 440 }
 441
 442 static void dm_start_request(struct mapped_device *md, struct request *orig)
 443 {
 444         blk_mq_start_request(orig);
 445
 446         if (unlikely(dm_stats_used(&md->stats))) {
 447                 struct dm_rq_target_io *tio = tio_from_request(orig);
 448                 tio->duration_jiffies = jiffies;
 449                 tio->n_sectors = blk_rq_sectors(orig);
 450                 dm_stats_account_io(&md->stats, rq_data_dir(orig),
 451                                     blk_rq_pos(orig), tio->n_sectors, false, 0,
 452                                     &tio->stats_aux);
 453         }
 454
 455         /*
 456          * Hold the md reference here for the in-flight I/O.
 457          * We can't rely on the reference count by device opener,
 458          * because the device may be closed during the request completion
 459          * when all bios are completed.
 460          * See the comment in rq_completed() too.
 461          */
 462         dm_get(md);
 463 }
 464
 465 static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
 466                               unsigned int hctx_idx, unsigned int numa_node)
 467 {
 468         struct mapped_device *md = set->driver_data;
 469         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 470
 471         /*
 472          * Must initialize md member of tio, otherwise it won't
 473          * be available in dm_mq_queue_rq.
 474          */
 475         tio->md = md;
 476
 477         if (md->init_tio_pdu) {
 478                 /* target-specific per-io data is immediately after the tio */
 479                 tio->info.ptr = tio + 1;
 480         }
 481
 482         return 0;
 483 }
 484
 485 static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
 486                           const struct blk_mq_queue_data *bd)
 487 {
 488         struct request *rq = bd->rq;
 489         struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
 490         struct mapped_device *md = tio->md;
 491         struct dm_target *ti = md->immutable_target;
 492
 493         /*
 494          * blk-mq's unquiesce may come from outside events, such as
 495          * elevator switch, updating nr_requests or others, and request may
 496          * come during suspend, so simply ask for blk-mq to requeue it.
 497          */
 498         if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)))
 499                 return BLK_STS_RESOURCE;
 500
 501         if (unlikely(!ti)) {
 502                 int srcu_idx;
 503                 struct dm_table *map;
 504
 505                 map = dm_get_live_table(md, &srcu_idx);
 506                 if (unlikely(!map)) {
 507                         dm_put_live_table(md, srcu_idx);
 508                         return BLK_STS_RESOURCE;
 509                 }
 510                 ti = dm_table_find_target(map, 0);
 511                 dm_put_live_table(md, srcu_idx);
 512         }
 513
 514         if (ti->type->busy && ti->type->busy(ti))
 515                 return BLK_STS_RESOURCE;
 516
 517         dm_start_request(md, rq);
 518
 519         /* Init tio using md established in .init_request */
 520         init_tio(tio, rq, md);
 521
 522         /*
 523          * Establish tio->ti before calling map_request().
 524          */
 525         tio->ti = ti;
 526
 527         /* Direct call is fine since .queue_rq allows allocations */
 528         if (map_request(tio) == DM_MAPIO_REQUEUE) {
 529                 /* Undo dm_start_request() before requeuing */
 530                 rq_end_stats(md, rq);
 531                 rq_completed(md);
 532                 return BLK_STS_RESOURCE;
 533         }
 534
 535         return BLK_STS_OK;
 536 }
 537
 538 static const struct blk_mq_ops dm_mq_ops = {
 539         .queue_rq = dm_mq_queue_rq,
 540         .complete = dm_softirq_done,
 541         .init_request = dm_mq_init_request,
 542 };
 543
 544 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
 545 {
 546         struct dm_target *immutable_tgt;
 547         int err;
 548
 549         md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
 550         if (!md->tag_set)
 551                 return -ENOMEM;
 552
 553         md->tag_set->ops = &dm_mq_ops;
 554         md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
 555         md->tag_set->numa_node = md->numa_node_id;
 556         md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_STACKING;
 557         md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
 558         md->tag_set->driver_data = md;
 559
 560         md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
 561         immutable_tgt = dm_table_get_immutable_target(t);
 562         if (immutable_tgt && immutable_tgt->per_io_data_size) {
 563                 /* any target-specific per-io data is immediately after the tio */
 564                 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
 565                 md->init_tio_pdu = true;
 566         }
 567
 568         err = blk_mq_alloc_tag_set(md->tag_set);
 569         if (err)
 570                 goto out_kfree_tag_set;
 571
 572         err = blk_mq_init_allocated_queue(md->tag_set, md->queue);
 573         if (err)
 574                 goto out_tag_set;
 575         return 0;
 576
 577 out_tag_set:
 578         blk_mq_free_tag_set(md->tag_set);
 579 out_kfree_tag_set:
 580         kfree(md->tag_set);
 581         md->tag_set = NULL;
 582
 583         return err;
 584 }
 585
 586 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
 587 {
 588         if (md->tag_set) {
 589                 blk_mq_free_tag_set(md->tag_set);
 590                 kfree(md->tag_set);
 591                 md->tag_set = NULL;
 592         }
 593 }
 594
 595 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
 596 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
 597
 598 /* Unused, but preserved for userspace compatibility */
 599 static bool use_blk_mq = true;
 600 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
 601 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
 602
 603 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
 604 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
 605
 606 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
 607 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");