drivers/md/dm-zoned-target.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2017 Western Digital Corporation or its affiliates.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include "dm-zoned.h"
   9
  10 #include <linux/module.h>
  11
  12 #define DM_MSG_PREFIX           "zoned"
  13
  14 #define DMZ_MIN_BIOS            8192
  15
  16 /*
  17  * Zone BIO context.
  18  */
  19 struct dmz_bioctx {
  20         struct dmz_target       *target;
  21         struct dm_zone          *zone;
  22         struct bio              *bio;
  23         refcount_t              ref;
  24 };
  25
  26 /*
  27  * Chunk work descriptor.
  28  */
  29 struct dm_chunk_work {
  30         struct work_struct      work;
  31         refcount_t              refcount;
  32         struct dmz_target       *target;
  33         unsigned int            chunk;
  34         struct bio_list         bio_list;
  35 };
  36
  37 /*
  38  * Target descriptor.
  39  */
  40 struct dmz_target {
  41         struct dm_dev           *ddev;
  42
  43         unsigned long           flags;
  44
  45         /* Zoned block device information */
  46         struct dmz_dev          *dev;
  47
  48         /* For metadata handling */
  49         struct dmz_metadata     *metadata;
  50
  51         /* For reclaim */
  52         struct dmz_reclaim      *reclaim;
  53
  54         /* For chunk work */
  55         struct radix_tree_root  chunk_rxtree;
  56         struct workqueue_struct *chunk_wq;
  57         struct mutex            chunk_lock;
  58
  59         /* For cloned BIOs to zones */
  60         struct bio_set          bio_set;
  61
  62         /* For flush */
  63         spinlock_t              flush_lock;
  64         struct bio_list         flush_list;
  65         struct delayed_work     flush_work;
  66         struct workqueue_struct *flush_wq;
  67 };
  68
  69 /*
  70  * Flush intervals (seconds).
  71  */
  72 #define DMZ_FLUSH_PERIOD        (10 * HZ)
  73
  74 /*
  75  * Target BIO completion.
  76  */
  77 static inline void dmz_bio_endio(struct bio *bio, blk_status_t status)
  78 {
  79         struct dmz_bioctx *bioctx = dm_per_bio_data(bio, sizeof(struct dmz_bioctx));
  80
  81         if (status != BLK_STS_OK && bio->bi_status == BLK_STS_OK)
  82                 bio->bi_status = status;
  83         if (bio->bi_status != BLK_STS_OK)
  84                 bioctx->target->dev->flags |= DMZ_CHECK_BDEV;
  85
  86         if (refcount_dec_and_test(&bioctx->ref)) {
  87                 struct dm_zone *zone = bioctx->zone;
  88
  89                 if (zone) {
  90                         if (bio->bi_status != BLK_STS_OK &&
  91                             bio_op(bio) == REQ_OP_WRITE &&
  92                             dmz_is_seq(zone))
  93                                 set_bit(DMZ_SEQ_WRITE_ERR, &zone->flags);
  94                         dmz_deactivate_zone(zone);
  95                 }
  96                 bio_endio(bio);
  97         }
  98 }
  99
 100 /*
 101  * Completion callback for an internally cloned target BIO. This terminates the
 102  * target BIO when there are no more references to its context.
 103  */
 104 static void dmz_clone_endio(struct bio *clone)
 105 {
 106         struct dmz_bioctx *bioctx = clone->bi_private;
 107         blk_status_t status = clone->bi_status;
 108
 109         bio_put(clone);
 110         dmz_bio_endio(bioctx->bio, status);
 111 }
 112
 113 /*
 114  * Issue a clone of a target BIO. The clone may only partially process the
 115  * original target BIO.
 116  */
 117 static int dmz_submit_bio(struct dmz_target *dmz, struct dm_zone *zone,
 118                           struct bio *bio, sector_t chunk_block,
 119                           unsigned int nr_blocks)
 120 {
 121         struct dmz_bioctx *bioctx = dm_per_bio_data(bio, sizeof(struct dmz_bioctx));
 122         struct bio *clone;
 123
 124         clone = bio_clone_fast(bio, GFP_NOIO, &dmz->bio_set);
 125         if (!clone)
 126                 return -ENOMEM;
 127
 128         bio_set_dev(clone, dmz->dev->bdev);
 129         clone->bi_iter.bi_sector =
 130                 dmz_start_sect(dmz->metadata, zone) + dmz_blk2sect(chunk_block);
 131         clone->bi_iter.bi_size = dmz_blk2sect(nr_blocks) << SECTOR_SHIFT;
 132         clone->bi_end_io = dmz_clone_endio;
 133         clone->bi_private = bioctx;
 134
 135         bio_advance(bio, clone->bi_iter.bi_size);
 136
 137         refcount_inc(&bioctx->ref);
 138         generic_make_request(clone);
 139
 140         if (bio_op(bio) == REQ_OP_WRITE && dmz_is_seq(zone))
 141                 zone->wp_block += nr_blocks;
 142
 143         return 0;
 144 }
 145
 146 /*
 147  * Zero out pages of discarded blocks accessed by a read BIO.
 148  */
 149 static void dmz_handle_read_zero(struct dmz_target *dmz, struct bio *bio,
 150                                  sector_t chunk_block, unsigned int nr_blocks)
 151 {
 152         unsigned int size = nr_blocks << DMZ_BLOCK_SHIFT;
 153
 154         /* Clear nr_blocks */
 155         swap(bio->bi_iter.bi_size, size);
 156         zero_fill_bio(bio);
 157         swap(bio->bi_iter.bi_size, size);
 158
 159         bio_advance(bio, size);
 160 }
 161
 162 /*
 163  * Process a read BIO.
 164  */
 165 static int dmz_handle_read(struct dmz_target *dmz, struct dm_zone *zone,
 166                            struct bio *bio)
 167 {
 168         sector_t chunk_block = dmz_chunk_block(dmz->dev, dmz_bio_block(bio));
 169         unsigned int nr_blocks = dmz_bio_blocks(bio);
 170         sector_t end_block = chunk_block + nr_blocks;
 171         struct dm_zone *rzone, *bzone;
 172         int ret;
 173
 174         /* Read into unmapped chunks need only zeroing the BIO buffer */
 175         if (!zone) {
 176                 zero_fill_bio(bio);
 177                 return 0;
 178         }
 179
 180         dmz_dev_debug(dmz->dev, "READ chunk %llu -> %s zone %u, block %llu, %u blocks",
 181                       (unsigned long long)dmz_bio_chunk(dmz->dev, bio),
 182                       (dmz_is_rnd(zone) ? "RND" : "SEQ"),
 183                       dmz_id(dmz->metadata, zone),
 184                       (unsigned long long)chunk_block, nr_blocks);
 185
 186         /* Check block validity to determine the read location */
 187         bzone = zone->bzone;
 188         while (chunk_block < end_block) {
 189                 nr_blocks = 0;
 190                 if (dmz_is_rnd(zone) || chunk_block < zone->wp_block) {
 191                         /* Test block validity in the data zone */
 192                         ret = dmz_block_valid(dmz->metadata, zone, chunk_block);
 193                         if (ret < 0)
 194                                 return ret;
 195                         if (ret > 0) {
 196                                 /* Read data zone blocks */
 197                                 nr_blocks = ret;
 198                                 rzone = zone;
 199                         }
 200                 }
 201
 202                 /*
 203                  * No valid blocks found in the data zone.
 204                  * Check the buffer zone, if there is one.
 205                  */
 206                 if (!nr_blocks && bzone) {
 207                         ret = dmz_block_valid(dmz->metadata, bzone, chunk_block);
 208                         if (ret < 0)
 209                                 return ret;
 210                         if (ret > 0) {
 211                                 /* Read buffer zone blocks */
 212                                 nr_blocks = ret;
 213                                 rzone = bzone;
 214                         }
 215                 }
 216
 217                 if (nr_blocks) {
 218                         /* Valid blocks found: read them */
 219                         nr_blocks = min_t(unsigned int, nr_blocks, end_block - chunk_block);
 220                         ret = dmz_submit_bio(dmz, rzone, bio, chunk_block, nr_blocks);
 221                         if (ret)
 222                                 return ret;
 223                         chunk_block += nr_blocks;
 224                 } else {
 225                         /* No valid block: zeroout the current BIO block */
 226                         dmz_handle_read_zero(dmz, bio, chunk_block, 1);
 227                         chunk_block++;
 228                 }
 229         }
 230
 231         return 0;
 232 }
 233
 234 /*
 235  * Write blocks directly in a data zone, at the write pointer.
 236  * If a buffer zone is assigned, invalidate the blocks written
 237  * in place.
 238  */
 239 static int dmz_handle_direct_write(struct dmz_target *dmz,
 240                                    struct dm_zone *zone, struct bio *bio,
 241                                    sector_t chunk_block,
 242                                    unsigned int nr_blocks)
 243 {
 244         struct dmz_metadata *zmd = dmz->metadata;
 245         struct dm_zone *bzone = zone->bzone;
 246         int ret;
 247
 248         if (dmz_is_readonly(zone))
 249                 return -EROFS;
 250
 251         /* Submit write */
 252         ret = dmz_submit_bio(dmz, zone, bio, chunk_block, nr_blocks);
 253         if (ret)
 254                 return ret;
 255
 256         /*
 257          * Validate the blocks in the data zone and invalidate
 258          * in the buffer zone, if there is one.
 259          */
 260         ret = dmz_validate_blocks(zmd, zone, chunk_block, nr_blocks);
 261         if (ret == 0 && bzone)
 262                 ret = dmz_invalidate_blocks(zmd, bzone, chunk_block, nr_blocks);
 263
 264         return ret;
 265 }
 266
 267 /*
 268  * Write blocks in the buffer zone of @zone.
 269  * If no buffer zone is assigned yet, get one.
 270  * Called with @zone write locked.
 271  */
 272 static int dmz_handle_buffered_write(struct dmz_target *dmz,
 273                                      struct dm_zone *zone, struct bio *bio,
 274                                      sector_t chunk_block,
 275                                      unsigned int nr_blocks)
 276 {
 277         struct dmz_metadata *zmd = dmz->metadata;
 278         struct dm_zone *bzone;
 279         int ret;
 280
 281         /* Get the buffer zone. One will be allocated if needed */
 282         bzone = dmz_get_chunk_buffer(zmd, zone);
 283         if (IS_ERR(bzone))
 284                 return PTR_ERR(bzone);
 285
 286         if (dmz_is_readonly(bzone))
 287                 return -EROFS;
 288
 289         /* Submit write */
 290         ret = dmz_submit_bio(dmz, bzone, bio, chunk_block, nr_blocks);
 291         if (ret)
 292                 return ret;
 293
 294         /*
 295          * Validate the blocks in the buffer zone
 296          * and invalidate in the data zone.
 297          */
 298         ret = dmz_validate_blocks(zmd, bzone, chunk_block, nr_blocks);
 299         if (ret == 0 && chunk_block < zone->wp_block)
 300                 ret = dmz_invalidate_blocks(zmd, zone, chunk_block, nr_blocks);
 301
 302         return ret;
 303 }
 304
 305 /*
 306  * Process a write BIO.
 307  */
 308 static int dmz_handle_write(struct dmz_target *dmz, struct dm_zone *zone,
 309                             struct bio *bio)
 310 {
 311         sector_t chunk_block = dmz_chunk_block(dmz->dev, dmz_bio_block(bio));
 312         unsigned int nr_blocks = dmz_bio_blocks(bio);
 313
 314         if (!zone)
 315                 return -ENOSPC;
 316
 317         dmz_dev_debug(dmz->dev, "WRITE chunk %llu -> %s zone %u, block %llu, %u blocks",
 318                       (unsigned long long)dmz_bio_chunk(dmz->dev, bio),
 319                       (dmz_is_rnd(zone) ? "RND" : "SEQ"),
 320                       dmz_id(dmz->metadata, zone),
 321                       (unsigned long long)chunk_block, nr_blocks);
 322
 323         if (dmz_is_rnd(zone) || chunk_block == zone->wp_block) {
 324                 /*
 325                  * zone is a random zone or it is a sequential zone
 326                  * and the BIO is aligned to the zone write pointer:
 327                  * direct write the zone.
 328                  */
 329                 return dmz_handle_direct_write(dmz, zone, bio, chunk_block, nr_blocks);
 330         }
 331
 332         /*
 333          * This is an unaligned write in a sequential zone:
 334          * use buffered write.
 335          */
 336         return dmz_handle_buffered_write(dmz, zone, bio, chunk_block, nr_blocks);
 337 }
 338
 339 /*
 340  * Process a discard BIO.
 341  */
 342 static int dmz_handle_discard(struct dmz_target *dmz, struct dm_zone *zone,
 343                               struct bio *bio)
 344 {
 345         struct dmz_metadata *zmd = dmz->metadata;
 346         sector_t block = dmz_bio_block(bio);
 347         unsigned int nr_blocks = dmz_bio_blocks(bio);
 348         sector_t chunk_block = dmz_chunk_block(dmz->dev, block);
 349         int ret = 0;
 350
 351         /* For unmapped chunks, there is nothing to do */
 352         if (!zone)
 353                 return 0;
 354
 355         if (dmz_is_readonly(zone))
 356                 return -EROFS;
 357
 358         dmz_dev_debug(dmz->dev, "DISCARD chunk %llu -> zone %u, block %llu, %u blocks",
 359                       (unsigned long long)dmz_bio_chunk(dmz->dev, bio),
 360                       dmz_id(zmd, zone),
 361                       (unsigned long long)chunk_block, nr_blocks);
 362
 363         /*
 364          * Invalidate blocks in the data zone and its
 365          * buffer zone if one is mapped.
 366          */
 367         if (dmz_is_rnd(zone) || chunk_block < zone->wp_block)
 368                 ret = dmz_invalidate_blocks(zmd, zone, chunk_block, nr_blocks);
 369         if (ret == 0 && zone->bzone)
 370                 ret = dmz_invalidate_blocks(zmd, zone->bzone,
 371                                             chunk_block, nr_blocks);
 372         return ret;
 373 }
 374
 375 /*
 376  * Process a BIO.
 377  */
 378 static void dmz_handle_bio(struct dmz_target *dmz, struct dm_chunk_work *cw,
 379                            struct bio *bio)
 380 {
 381         struct dmz_bioctx *bioctx = dm_per_bio_data(bio, sizeof(struct dmz_bioctx));
 382         struct dmz_metadata *zmd = dmz->metadata;
 383         struct dm_zone *zone;
 384         int ret;
 385
 386         /*
 387          * Write may trigger a zone allocation. So make sure the
 388          * allocation can succeed.
 389          */
 390         if (bio_op(bio) == REQ_OP_WRITE)
 391                 dmz_schedule_reclaim(dmz->reclaim);
 392
 393         dmz_lock_metadata(zmd);
 394
 395         if (dmz->dev->flags & DMZ_BDEV_DYING) {
 396                 ret = -EIO;
 397                 goto out;
 398         }
 399
 400         /*
 401          * Get the data zone mapping the chunk. There may be no
 402          * mapping for read and discard. If a mapping is obtained,
 403          + the zone returned will be set to active state.
 404          */
 405         zone = dmz_get_chunk_mapping(zmd, dmz_bio_chunk(dmz->dev, bio),
 406                                      bio_op(bio));
 407         if (IS_ERR(zone)) {
 408                 ret = PTR_ERR(zone);
 409                 goto out;
 410         }
 411
 412         /* Process the BIO */
 413         if (zone) {
 414                 dmz_activate_zone(zone);
 415                 bioctx->zone = zone;
 416         }
 417
 418         switch (bio_op(bio)) {
 419         case REQ_OP_READ:
 420                 ret = dmz_handle_read(dmz, zone, bio);
 421                 break;
 422         case REQ_OP_WRITE:
 423                 ret = dmz_handle_write(dmz, zone, bio);
 424                 break;
 425         case REQ_OP_DISCARD:
 426         case REQ_OP_WRITE_ZEROES:
 427                 ret = dmz_handle_discard(dmz, zone, bio);
 428                 break;
 429         default:
 430                 dmz_dev_err(dmz->dev, "Unsupported BIO operation 0x%x",
 431                             bio_op(bio));
 432                 ret = -EIO;
 433         }
 434
 435         /*
 436          * Release the chunk mapping. This will check that the mapping
 437          * is still valid, that is, that the zone used still has valid blocks.
 438          */
 439         if (zone)
 440                 dmz_put_chunk_mapping(zmd, zone);
 441 out:
 442         dmz_bio_endio(bio, errno_to_blk_status(ret));
 443
 444         dmz_unlock_metadata(zmd);
 445 }
 446
 447 /*
 448  * Increment a chunk reference counter.
 449  */
 450 static inline void dmz_get_chunk_work(struct dm_chunk_work *cw)
 451 {
 452         refcount_inc(&cw->refcount);
 453 }
 454
 455 /*
 456  * Decrement a chunk work reference count and
 457  * free it if it becomes 0.
 458  */
 459 static void dmz_put_chunk_work(struct dm_chunk_work *cw)
 460 {
 461         if (refcount_dec_and_test(&cw->refcount)) {
 462                 WARN_ON(!bio_list_empty(&cw->bio_list));
 463                 radix_tree_delete(&cw->target->chunk_rxtree, cw->chunk);
 464                 kfree(cw);
 465         }
 466 }
 467
 468 /*
 469  * Chunk BIO work function.
 470  */
 471 static void dmz_chunk_work(struct work_struct *work)
 472 {
 473         struct dm_chunk_work *cw = container_of(work, struct dm_chunk_work, work);
 474         struct dmz_target *dmz = cw->target;
 475         struct bio *bio;
 476
 477         mutex_lock(&dmz->chunk_lock);
 478
 479         /* Process the chunk BIOs */
 480         while ((bio = bio_list_pop(&cw->bio_list))) {
 481                 mutex_unlock(&dmz->chunk_lock);
 482                 dmz_handle_bio(dmz, cw, bio);
 483                 mutex_lock(&dmz->chunk_lock);
 484                 dmz_put_chunk_work(cw);
 485         }
 486
 487         /* Queueing the work incremented the work refcount */
 488         dmz_put_chunk_work(cw);
 489
 490         mutex_unlock(&dmz->chunk_lock);
 491 }
 492
 493 /*
 494  * Flush work.
 495  */
 496 static void dmz_flush_work(struct work_struct *work)
 497 {
 498         struct dmz_target *dmz = container_of(work, struct dmz_target, flush_work.work);
 499         struct bio *bio;
 500         int ret;
 501
 502         /* Flush dirty metadata blocks */
 503         ret = dmz_flush_metadata(dmz->metadata);
 504         if (ret)
 505                 dmz_dev_debug(dmz->dev, "Metadata flush failed, rc=%d\n", ret);
 506
 507         /* Process queued flush requests */
 508         while (1) {
 509                 spin_lock(&dmz->flush_lock);
 510                 bio = bio_list_pop(&dmz->flush_list);
 511                 spin_unlock(&dmz->flush_lock);
 512
 513                 if (!bio)
 514                         break;
 515
 516                 dmz_bio_endio(bio, errno_to_blk_status(ret));
 517         }
 518
 519         queue_delayed_work(dmz->flush_wq, &dmz->flush_work, DMZ_FLUSH_PERIOD);
 520 }
 521
 522 /*
 523  * Get a chunk work and start it to process a new BIO.
 524  * If the BIO chunk has no work yet, create one.
 525  */
 526 static int dmz_queue_chunk_work(struct dmz_target *dmz, struct bio *bio)
 527 {
 528         unsigned int chunk = dmz_bio_chunk(dmz->dev, bio);
 529         struct dm_chunk_work *cw;
 530         int ret = 0;
 531
 532         mutex_lock(&dmz->chunk_lock);
 533
 534         /* Get the BIO chunk work. If one is not active yet, create one */
 535         cw = radix_tree_lookup(&dmz->chunk_rxtree, chunk);
 536         if (!cw) {
 537
 538                 /* Create a new chunk work */
 539                 cw = kmalloc(sizeof(struct dm_chunk_work), GFP_NOIO);
 540                 if (unlikely(!cw)) {
 541                         ret = -ENOMEM;
 542                         goto out;
 543                 }
 544
 545                 INIT_WORK(&cw->work, dmz_chunk_work);
 546                 refcount_set(&cw->refcount, 0);
 547                 cw->target = dmz;
 548                 cw->chunk = chunk;
 549                 bio_list_init(&cw->bio_list);
 550
 551                 ret = radix_tree_insert(&dmz->chunk_rxtree, chunk, cw);
 552                 if (unlikely(ret)) {
 553                         kfree(cw);
 554                         goto out;
 555                 }
 556         }
 557
 558         bio_list_add(&cw->bio_list, bio);
 559         dmz_get_chunk_work(cw);
 560
 561         dmz_reclaim_bio_acc(dmz->reclaim);
 562         if (queue_work(dmz->chunk_wq, &cw->work))
 563                 dmz_get_chunk_work(cw);
 564 out:
 565         mutex_unlock(&dmz->chunk_lock);
 566         return ret;
 567 }
 568
 569 /*
 570  * Check if the backing device is being removed. If it's on the way out,
 571  * start failing I/O. Reclaim and metadata components also call this
 572  * function to cleanly abort operation in the event of such failure.
 573  */
 574 bool dmz_bdev_is_dying(struct dmz_dev *dmz_dev)
 575 {
 576         if (dmz_dev->flags & DMZ_BDEV_DYING)
 577                 return true;
 578
 579         if (dmz_dev->flags & DMZ_CHECK_BDEV)
 580                 return !dmz_check_bdev(dmz_dev);
 581
 582         if (blk_queue_dying(bdev_get_queue(dmz_dev->bdev))) {
 583                 dmz_dev_warn(dmz_dev, "Backing device queue dying");
 584                 dmz_dev->flags |= DMZ_BDEV_DYING;
 585         }
 586
 587         return dmz_dev->flags & DMZ_BDEV_DYING;
 588 }
 589
 590 /*
 591  * Check the backing device availability. This detects such events as
 592  * backing device going offline due to errors, media removals, etc.
 593  * This check is less efficient than dmz_bdev_is_dying() and should
 594  * only be performed as a part of error handling.
 595  */
 596 bool dmz_check_bdev(struct dmz_dev *dmz_dev)
 597 {
 598         struct gendisk *disk;
 599
 600         dmz_dev->flags &= ~DMZ_CHECK_BDEV;
 601
 602         if (dmz_bdev_is_dying(dmz_dev))
 603                 return false;
 604
 605         disk = dmz_dev->bdev->bd_disk;
 606         if (disk->fops->check_events &&
 607             disk->fops->check_events(disk, 0) & DISK_EVENT_MEDIA_CHANGE) {
 608                 dmz_dev_warn(dmz_dev, "Backing device offline");
 609                 dmz_dev->flags |= DMZ_BDEV_DYING;
 610         }
 611
 612         return !(dmz_dev->flags & DMZ_BDEV_DYING);
 613 }
 614
 615 /*
 616  * Process a new BIO.
 617  */
 618 static int dmz_map(struct dm_target *ti, struct bio *bio)
 619 {
 620         struct dmz_target *dmz = ti->private;
 621         struct dmz_dev *dev = dmz->dev;
 622         struct dmz_bioctx *bioctx = dm_per_bio_data(bio, sizeof(struct dmz_bioctx));
 623         sector_t sector = bio->bi_iter.bi_sector;
 624         unsigned int nr_sectors = bio_sectors(bio);
 625         sector_t chunk_sector;
 626         int ret;
 627
 628         if (dmz_bdev_is_dying(dmz->dev))
 629                 return DM_MAPIO_KILL;
 630
 631         dmz_dev_debug(dev, "BIO op %d sector %llu + %u => chunk %llu, block %llu, %u blocks",
 632                       bio_op(bio), (unsigned long long)sector, nr_sectors,
 633                       (unsigned long long)dmz_bio_chunk(dmz->dev, bio),
 634                       (unsigned long long)dmz_chunk_block(dmz->dev, dmz_bio_block(bio)),
 635                       (unsigned int)dmz_bio_blocks(bio));
 636
 637         bio_set_dev(bio, dev->bdev);
 638
 639         if (!nr_sectors && bio_op(bio) != REQ_OP_WRITE)
 640                 return DM_MAPIO_REMAPPED;
 641
 642         /* The BIO should be block aligned */
 643         if ((nr_sectors & DMZ_BLOCK_SECTORS_MASK) || (sector & DMZ_BLOCK_SECTORS_MASK))
 644                 return DM_MAPIO_KILL;
 645
 646         /* Initialize the BIO context */
 647         bioctx->target = dmz;
 648         bioctx->zone = NULL;
 649         bioctx->bio = bio;
 650         refcount_set(&bioctx->ref, 1);
 651
 652         /* Set the BIO pending in the flush list */
 653         if (!nr_sectors && bio_op(bio) == REQ_OP_WRITE) {
 654                 spin_lock(&dmz->flush_lock);
 655                 bio_list_add(&dmz->flush_list, bio);
 656                 spin_unlock(&dmz->flush_lock);
 657                 mod_delayed_work(dmz->flush_wq, &dmz->flush_work, 0);
 658                 return DM_MAPIO_SUBMITTED;
 659         }
 660
 661         /* Split zone BIOs to fit entirely into a zone */
 662         chunk_sector = sector & (dev->zone_nr_sectors - 1);
 663         if (chunk_sector + nr_sectors > dev->zone_nr_sectors)
 664                 dm_accept_partial_bio(bio, dev->zone_nr_sectors - chunk_sector);
 665
 666         /* Now ready to handle this BIO */
 667         ret = dmz_queue_chunk_work(dmz, bio);
 668         if (ret) {
 669                 dmz_dev_debug(dmz->dev,
 670                               "BIO op %d, can't process chunk %llu, err %i\n",
 671                               bio_op(bio), (u64)dmz_bio_chunk(dmz->dev, bio),
 672                               ret);
 673                 return DM_MAPIO_REQUEUE;
 674         }
 675
 676         return DM_MAPIO_SUBMITTED;
 677 }
 678
 679 /*
 680  * Get zoned device information.
 681  */
 682 static int dmz_get_zoned_device(struct dm_target *ti, char *path)
 683 {
 684         struct dmz_target *dmz = ti->private;
 685         struct request_queue *q;
 686         struct dmz_dev *dev;
 687         sector_t aligned_capacity;
 688         int ret;
 689
 690         /* Get the target device */
 691         ret = dm_get_device(ti, path, dm_table_get_mode(ti->table), &dmz->ddev);
 692         if (ret) {
 693                 ti->error = "Get target device failed";
 694                 dmz->ddev = NULL;
 695                 return ret;
 696         }
 697
 698         dev = kzalloc(sizeof(struct dmz_dev), GFP_KERNEL);
 699         if (!dev) {
 700                 ret = -ENOMEM;
 701                 goto err;
 702         }
 703
 704         dev->bdev = dmz->ddev->bdev;
 705         (void)bdevname(dev->bdev, dev->name);
 706
 707         if (bdev_zoned_model(dev->bdev) == BLK_ZONED_NONE) {
 708                 ti->error = "Not a zoned block device";
 709                 ret = -EINVAL;
 710                 goto err;
 711         }
 712
 713         q = bdev_get_queue(dev->bdev);
 714         dev->capacity = i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
 715         aligned_capacity = dev->capacity &
 716                                 ~((sector_t)blk_queue_zone_sectors(q) - 1);
 717         if (ti->begin ||
 718             ((ti->len != dev->capacity) && (ti->len != aligned_capacity))) {
 719                 ti->error = "Partial mapping not supported";
 720                 ret = -EINVAL;
 721                 goto err;
 722         }
 723
 724         dev->zone_nr_sectors = blk_queue_zone_sectors(q);
 725         dev->zone_nr_sectors_shift = ilog2(dev->zone_nr_sectors);
 726
 727         dev->zone_nr_blocks = dmz_sect2blk(dev->zone_nr_sectors);
 728         dev->zone_nr_blocks_shift = ilog2(dev->zone_nr_blocks);
 729
 730         dev->nr_zones = blkdev_nr_zones(dev->bdev);
 731
 732         dmz->dev = dev;
 733
 734         return 0;
 735 err:
 736         dm_put_device(ti, dmz->ddev);
 737         kfree(dev);
 738
 739         return ret;
 740 }
 741
 742 /*
 743  * Cleanup zoned device information.
 744  */
 745 static void dmz_put_zoned_device(struct dm_target *ti)
 746 {
 747         struct dmz_target *dmz = ti->private;
 748
 749         dm_put_device(ti, dmz->ddev);
 750         kfree(dmz->dev);
 751         dmz->dev = NULL;
 752 }
 753
 754 /*
 755  * Setup target.
 756  */
 757 static int dmz_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 758 {
 759         struct dmz_target *dmz;
 760         struct dmz_dev *dev;
 761         int ret;
 762
 763         /* Check arguments */
 764         if (argc != 1) {
 765                 ti->error = "Invalid argument count";
 766                 return -EINVAL;
 767         }
 768
 769         /* Allocate and initialize the target descriptor */
 770         dmz = kzalloc(sizeof(struct dmz_target), GFP_KERNEL);
 771         if (!dmz) {
 772                 ti->error = "Unable to allocate the zoned target descriptor";
 773                 return -ENOMEM;
 774         }
 775         ti->private = dmz;
 776
 777         /* Get the target zoned block device */
 778         ret = dmz_get_zoned_device(ti, argv[0]);
 779         if (ret) {
 780                 dmz->ddev = NULL;
 781                 goto err;
 782         }
 783
 784         /* Initialize metadata */
 785         dev = dmz->dev;
 786         ret = dmz_ctr_metadata(dev, &dmz->metadata);
 787         if (ret) {
 788                 ti->error = "Metadata initialization failed";
 789                 goto err_dev;
 790         }
 791
 792         /* Set target (no write same support) */
 793         ti->max_io_len = dev->zone_nr_sectors << 9;
 794         ti->num_flush_bios = 1;
 795         ti->num_discard_bios = 1;
 796         ti->num_write_zeroes_bios = 1;
 797         ti->per_io_data_size = sizeof(struct dmz_bioctx);
 798         ti->flush_supported = true;
 799         ti->discards_supported = true;
 800
 801         /* The exposed capacity is the number of chunks that can be mapped */
 802         ti->len = (sector_t)dmz_nr_chunks(dmz->metadata) << dev->zone_nr_sectors_shift;
 803
 804         /* Zone BIO */
 805         ret = bioset_init(&dmz->bio_set, DMZ_MIN_BIOS, 0, 0);
 806         if (ret) {
 807                 ti->error = "Create BIO set failed";
 808                 goto err_meta;
 809         }
 810
 811         /* Chunk BIO work */
 812         mutex_init(&dmz->chunk_lock);
 813         INIT_RADIX_TREE(&dmz->chunk_rxtree, GFP_NOIO);
 814         dmz->chunk_wq = alloc_workqueue("dmz_cwq_%s", WQ_MEM_RECLAIM | WQ_UNBOUND,
 815                                         0, dev->name);
 816         if (!dmz->chunk_wq) {
 817                 ti->error = "Create chunk workqueue failed";
 818                 ret = -ENOMEM;
 819                 goto err_bio;
 820         }
 821
 822         /* Flush work */
 823         spin_lock_init(&dmz->flush_lock);
 824         bio_list_init(&dmz->flush_list);
 825         INIT_DELAYED_WORK(&dmz->flush_work, dmz_flush_work);
 826         dmz->flush_wq = alloc_ordered_workqueue("dmz_fwq_%s", WQ_MEM_RECLAIM,
 827                                                 dev->name);
 828         if (!dmz->flush_wq) {
 829                 ti->error = "Create flush workqueue failed";
 830                 ret = -ENOMEM;
 831                 goto err_cwq;
 832         }
 833         mod_delayed_work(dmz->flush_wq, &dmz->flush_work, DMZ_FLUSH_PERIOD);
 834
 835         /* Initialize reclaim */
 836         ret = dmz_ctr_reclaim(dev, dmz->metadata, &dmz->reclaim);
 837         if (ret) {
 838                 ti->error = "Zone reclaim initialization failed";
 839                 goto err_fwq;
 840         }
 841
 842         dmz_dev_info(dev, "Target device: %llu 512-byte logical sectors (%llu blocks)",
 843                      (unsigned long long)ti->len,
 844                      (unsigned long long)dmz_sect2blk(ti->len));
 845
 846         return 0;
 847 err_fwq:
 848         destroy_workqueue(dmz->flush_wq);
 849 err_cwq:
 850         destroy_workqueue(dmz->chunk_wq);
 851 err_bio:
 852         mutex_destroy(&dmz->chunk_lock);
 853         bioset_exit(&dmz->bio_set);
 854 err_meta:
 855         dmz_dtr_metadata(dmz->metadata);
 856 err_dev:
 857         dmz_put_zoned_device(ti);
 858 err:
 859         kfree(dmz);
 860
 861         return ret;
 862 }
 863
 864 /*
 865  * Cleanup target.
 866  */
 867 static void dmz_dtr(struct dm_target *ti)
 868 {
 869         struct dmz_target *dmz = ti->private;
 870
 871         flush_workqueue(dmz->chunk_wq);
 872         destroy_workqueue(dmz->chunk_wq);
 873
 874         dmz_dtr_reclaim(dmz->reclaim);
 875
 876         cancel_delayed_work_sync(&dmz->flush_work);
 877         destroy_workqueue(dmz->flush_wq);
 878
 879         (void) dmz_flush_metadata(dmz->metadata);
 880
 881         dmz_dtr_metadata(dmz->metadata);
 882
 883         bioset_exit(&dmz->bio_set);
 884
 885         dmz_put_zoned_device(ti);
 886
 887         mutex_destroy(&dmz->chunk_lock);
 888
 889         kfree(dmz);
 890 }
 891
 892 /*
 893  * Setup target request queue limits.
 894  */
 895 static void dmz_io_hints(struct dm_target *ti, struct queue_limits *limits)
 896 {
 897         struct dmz_target *dmz = ti->private;
 898         unsigned int chunk_sectors = dmz->dev->zone_nr_sectors;
 899
 900         limits->logical_block_size = DMZ_BLOCK_SIZE;
 901         limits->physical_block_size = DMZ_BLOCK_SIZE;
 902
 903         blk_limits_io_min(limits, DMZ_BLOCK_SIZE);
 904         blk_limits_io_opt(limits, DMZ_BLOCK_SIZE);
 905
 906         limits->discard_alignment = DMZ_BLOCK_SIZE;
 907         limits->discard_granularity = DMZ_BLOCK_SIZE;
 908         limits->max_discard_sectors = chunk_sectors;
 909         limits->max_hw_discard_sectors = chunk_sectors;
 910         limits->max_write_zeroes_sectors = chunk_sectors;
 911
 912         /* FS hint to try to align to the device zone size */
 913         limits->chunk_sectors = chunk_sectors;
 914         limits->max_sectors = chunk_sectors;
 915
 916         /* We are exposing a drive-managed zoned block device */
 917         limits->zoned = BLK_ZONED_NONE;
 918 }
 919
 920 /*
 921  * Pass on ioctl to the backend device.
 922  */
 923 static int dmz_prepare_ioctl(struct dm_target *ti, struct block_device **bdev)
 924 {
 925         struct dmz_target *dmz = ti->private;
 926
 927         if (!dmz_check_bdev(dmz->dev))
 928                 return -EIO;
 929
 930         *bdev = dmz->dev->bdev;
 931
 932         return 0;
 933 }
 934
 935 /*
 936  * Stop works on suspend.
 937  */
 938 static void dmz_suspend(struct dm_target *ti)
 939 {
 940         struct dmz_target *dmz = ti->private;
 941
 942         flush_workqueue(dmz->chunk_wq);
 943         dmz_suspend_reclaim(dmz->reclaim);
 944         cancel_delayed_work_sync(&dmz->flush_work);
 945 }
 946
 947 /*
 948  * Restart works on resume or if suspend failed.
 949  */
 950 static void dmz_resume(struct dm_target *ti)
 951 {
 952         struct dmz_target *dmz = ti->private;
 953
 954         queue_delayed_work(dmz->flush_wq, &dmz->flush_work, DMZ_FLUSH_PERIOD);
 955         dmz_resume_reclaim(dmz->reclaim);
 956 }
 957
 958 static int dmz_iterate_devices(struct dm_target *ti,
 959                                iterate_devices_callout_fn fn, void *data)
 960 {
 961         struct dmz_target *dmz = ti->private;
 962         struct dmz_dev *dev = dmz->dev;
 963         sector_t capacity = dev->capacity & ~(dev->zone_nr_sectors - 1);
 964
 965         return fn(ti, dmz->ddev, 0, capacity, data);
 966 }
 967
 968 static struct target_type dmz_type = {
 969         .name            = "zoned",
 970         .version         = {1, 0, 0},
 971         .features        = DM_TARGET_SINGLETON | DM_TARGET_ZONED_HM,
 972         .module          = THIS_MODULE,
 973         .ctr             = dmz_ctr,
 974         .dtr             = dmz_dtr,
 975         .map             = dmz_map,
 976         .io_hints        = dmz_io_hints,
 977         .prepare_ioctl   = dmz_prepare_ioctl,
 978         .postsuspend     = dmz_suspend,
 979         .resume          = dmz_resume,
 980         .iterate_devices = dmz_iterate_devices,
 981 };
 982
 983 static int __init dmz_init(void)
 984 {
 985         return dm_register_target(&dmz_type);
 986 }
 987
 988 static void __exit dmz_exit(void)
 989 {
 990         dm_unregister_target(&dmz_type);
 991 }
 992
 993 module_init(dmz_init);
 994 module_exit(dmz_exit);
 995
 996 MODULE_DESCRIPTION(DM_NAME " target for zoned block devices");
 997 MODULE_AUTHOR("Damien Le Moal <damien.lemoal@wdc.com>");
 998 MODULE_LICENSE("GPL");