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