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