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