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