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_dev          *dev;
  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         unsigned int            nr_ddevs;
  43
  44         unsigned int            flags;
  45
  46         /* Zoned block device information */
  47         struct dmz_dev          *dev;
  48
  49         /* For metadata handling */
  50         struct dmz_metadata     *metadata;
  51
  52         /* For chunk work */
  53         struct radix_tree_root  chunk_rxtree;
  54         struct workqueue_struct *chunk_wq;
  55         struct mutex            chunk_lock;
  56
  57         /* For cloned BIOs to zones */
  58         struct bio_set          bio_set;
  59
  60         /* For flush */
  61         spinlock_t              flush_lock;
  62         struct bio_list         flush_list;
  63         struct delayed_work     flush_work;
  64         struct workqueue_struct *flush_wq;
  65 };
  66
  67 /*
  68  * Flush intervals (seconds).
  69  */
  70 #define DMZ_FLUSH_PERIOD        (10 * HZ)
  71
  72 /*
  73  * Target BIO completion.
  74  */
  75 static inline void dmz_bio_endio(struct bio *bio, blk_status_t status)
  76 {
  77         struct dmz_bioctx *bioctx =
  78                 dm_per_bio_data(bio, sizeof(struct dmz_bioctx));
  79
  80         if (status != BLK_STS_OK && bio->bi_status == BLK_STS_OK)
  81                 bio->bi_status = status;
  82         if (bioctx->dev && bio->bi_status != BLK_STS_OK)
  83                 bioctx->dev->flags |= DMZ_CHECK_BDEV;
  84
  85         if (refcount_dec_and_test(&bioctx->ref)) {
  86                 struct dm_zone *zone = bioctx->zone;
  87
  88                 if (zone) {
  89                         if (bio->bi_status != BLK_STS_OK &&
  90                             bio_op(bio) == REQ_OP_WRITE &&
  91                             dmz_is_seq(zone))
  92                                 set_bit(DMZ_SEQ_WRITE_ERR, &zone->flags);
  93                         dmz_deactivate_zone(zone);
  94                 }
  95                 bio_endio(bio);
  96         }
  97 }
  98
  99 /*
 100  * Completion callback for an internally cloned target BIO. This terminates the
 101  * target BIO when there are no more references to its context.
 102  */
 103 static void dmz_clone_endio(struct bio *clone)
 104 {
 105         struct dmz_bioctx *bioctx = clone->bi_private;
 106         blk_status_t status = clone->bi_status;
 107
 108         bio_put(clone);
 109         dmz_bio_endio(bioctx->bio, status);
 110 }
 111
 112 /*
 113  * Issue a clone of a target BIO. The clone may only partially process the
 114  * original target BIO.
 115  */
 116 static int dmz_submit_bio(struct dmz_target *dmz, struct dm_zone *zone,
 117                           struct bio *bio, sector_t chunk_block,
 118                           unsigned int nr_blocks)
 119 {
 120         struct dmz_bioctx *bioctx =
 121                 dm_per_bio_data(bio, sizeof(struct dmz_bioctx));
 122         struct dmz_dev *dev = zone->dev;
 123         struct bio *clone;
 124
 125         if (dev->flags & DMZ_BDEV_DYING)
 126                 return -EIO;
 127
 128         clone = bio_clone_fast(bio, GFP_NOIO, &dmz->bio_set);
 129         if (!clone)
 130                 return -ENOMEM;
 131
 132         bio_set_dev(clone, dev->bdev);
 133         bioctx->dev = dev;
 134         clone->bi_iter.bi_sector =
 135                 dmz_start_sect(dmz->metadata, zone) + dmz_blk2sect(chunk_block);
 136         clone->bi_iter.bi_size = dmz_blk2sect(nr_blocks) << SECTOR_SHIFT;
 137         clone->bi_end_io = dmz_clone_endio;
 138         clone->bi_private = bioctx;
 139
 140         bio_advance(bio, clone->bi_iter.bi_size);
 141
 142         refcount_inc(&bioctx->ref);
 143         submit_bio_noacct(clone);
 144
 145         if (bio_op(bio) == REQ_OP_WRITE && dmz_is_seq(zone))
 146                 zone->wp_block += nr_blocks;
 147
 148         return 0;
 149 }
 150
 151 /*
 152  * Zero out pages of discarded blocks accessed by a read BIO.
 153  */
 154 static void dmz_handle_read_zero(struct dmz_target *dmz, struct bio *bio,
 155                                  sector_t chunk_block, unsigned int nr_blocks)
 156 {
 157         unsigned int size = nr_blocks << DMZ_BLOCK_SHIFT;
 158
 159         /* Clear nr_blocks */
 160         swap(bio->bi_iter.bi_size, size);
 161         zero_fill_bio(bio);
 162         swap(bio->bi_iter.bi_size, size);
 163
 164         bio_advance(bio, size);
 165 }
 166
 167 /*
 168  * Process a read BIO.
 169  */
 170 static int dmz_handle_read(struct dmz_target *dmz, struct dm_zone *zone,
 171                            struct bio *bio)
 172 {
 173         struct dmz_metadata *zmd = dmz->metadata;
 174         sector_t chunk_block = dmz_chunk_block(zmd, dmz_bio_block(bio));
 175         unsigned int nr_blocks = dmz_bio_blocks(bio);
 176         sector_t end_block = chunk_block + nr_blocks;
 177         struct dm_zone *rzone, *bzone;
 178         int ret;
 179
 180         /* Read into unmapped chunks need only zeroing the BIO buffer */
 181         if (!zone) {
 182                 zero_fill_bio(bio);
 183                 return 0;
 184         }
 185
 186         DMDEBUG("(%s): READ chunk %llu -> %s zone %u, block %llu, %u blocks",
 187                 dmz_metadata_label(zmd),
 188                 (unsigned long long)dmz_bio_chunk(zmd, bio),
 189                 (dmz_is_rnd(zone) ? "RND" :
 190                  (dmz_is_cache(zone) ? "CACHE" : "SEQ")),
 191                 zone->id,
 192                 (unsigned long long)chunk_block, nr_blocks);
 193
 194         /* Check block validity to determine the read location */
 195         bzone = zone->bzone;
 196         while (chunk_block < end_block) {
 197                 nr_blocks = 0;
 198                 if (dmz_is_rnd(zone) || dmz_is_cache(zone) ||
 199                     chunk_block < zone->wp_block) {
 200                         /* Test block validity in the data zone */
 201                         ret = dmz_block_valid(zmd, zone, chunk_block);
 202                         if (ret < 0)
 203                                 return ret;
 204                         if (ret > 0) {
 205                                 /* Read data zone blocks */
 206                                 nr_blocks = ret;
 207                                 rzone = zone;
 208                         }
 209                 }
 210
 211                 /*
 212                  * No valid blocks found in the data zone.
 213                  * Check the buffer zone, if there is one.
 214                  */
 215                 if (!nr_blocks && bzone) {
 216                         ret = dmz_block_valid(zmd, bzone, chunk_block);
 217                         if (ret < 0)
 218                                 return ret;
 219                         if (ret > 0) {
 220                                 /* Read buffer zone blocks */
 221                                 nr_blocks = ret;
 222                                 rzone = bzone;
 223                         }
 224                 }
 225
 226                 if (nr_blocks) {
 227                         /* Valid blocks found: read them */
 228                         nr_blocks = min_t(unsigned int, nr_blocks,
 229                                           end_block - chunk_block);
 230                         ret = dmz_submit_bio(dmz, rzone, bio,
 231                                              chunk_block, nr_blocks);
 232                         if (ret)
 233                                 return ret;
 234                         chunk_block += nr_blocks;
 235                 } else {
 236                         /* No valid block: zeroout the current BIO block */
 237                         dmz_handle_read_zero(dmz, bio, chunk_block, 1);
 238                         chunk_block++;
 239                 }
 240         }
 241
 242         return 0;
 243 }
 244
 245 /*
 246  * Write blocks directly in a data zone, at the write pointer.
 247  * If a buffer zone is assigned, invalidate the blocks written
 248  * in place.
 249  */
 250 static int dmz_handle_direct_write(struct dmz_target *dmz,
 251                                    struct dm_zone *zone, struct bio *bio,
 252                                    sector_t chunk_block,
 253                                    unsigned int nr_blocks)
 254 {
 255         struct dmz_metadata *zmd = dmz->metadata;
 256         struct dm_zone *bzone = zone->bzone;
 257         int ret;
 258
 259         if (dmz_is_readonly(zone))
 260                 return -EROFS;
 261
 262         /* Submit write */
 263         ret = dmz_submit_bio(dmz, zone, bio, chunk_block, nr_blocks);
 264         if (ret)
 265                 return ret;
 266
 267         /*
 268          * Validate the blocks in the data zone and invalidate
 269          * in the buffer zone, if there is one.
 270          */
 271         ret = dmz_validate_blocks(zmd, zone, chunk_block, nr_blocks);
 272         if (ret == 0 && bzone)
 273                 ret = dmz_invalidate_blocks(zmd, bzone, chunk_block, nr_blocks);
 274
 275         return ret;
 276 }
 277
 278 /*
 279  * Write blocks in the buffer zone of @zone.
 280  * If no buffer zone is assigned yet, get one.
 281  * Called with @zone write locked.
 282  */
 283 static int dmz_handle_buffered_write(struct dmz_target *dmz,
 284                                      struct dm_zone *zone, struct bio *bio,
 285                                      sector_t chunk_block,
 286                                      unsigned int nr_blocks)
 287 {
 288         struct dmz_metadata *zmd = dmz->metadata;
 289         struct dm_zone *bzone;
 290         int ret;
 291
 292         /* Get the buffer zone. One will be allocated if needed */
 293         bzone = dmz_get_chunk_buffer(zmd, zone);
 294         if (IS_ERR(bzone))
 295                 return PTR_ERR(bzone);
 296
 297         if (dmz_is_readonly(bzone))
 298                 return -EROFS;
 299
 300         /* Submit write */
 301         ret = dmz_submit_bio(dmz, bzone, bio, chunk_block, nr_blocks);
 302         if (ret)
 303                 return ret;
 304
 305         /*
 306          * Validate the blocks in the buffer zone
 307          * and invalidate in the data zone.
 308          */
 309         ret = dmz_validate_blocks(zmd, bzone, chunk_block, nr_blocks);
 310         if (ret == 0 && chunk_block < zone->wp_block)
 311                 ret = dmz_invalidate_blocks(zmd, zone, chunk_block, nr_blocks);
 312
 313         return ret;
 314 }
 315
 316 /*
 317  * Process a write BIO.
 318  */
 319 static int dmz_handle_write(struct dmz_target *dmz, struct dm_zone *zone,
 320                             struct bio *bio)
 321 {
 322         struct dmz_metadata *zmd = dmz->metadata;
 323         sector_t chunk_block = dmz_chunk_block(zmd, dmz_bio_block(bio));
 324         unsigned int nr_blocks = dmz_bio_blocks(bio);
 325
 326         if (!zone)
 327                 return -ENOSPC;
 328
 329         DMDEBUG("(%s): WRITE chunk %llu -> %s zone %u, block %llu, %u blocks",
 330                 dmz_metadata_label(zmd),
 331                 (unsigned long long)dmz_bio_chunk(zmd, bio),
 332                 (dmz_is_rnd(zone) ? "RND" :
 333                  (dmz_is_cache(zone) ? "CACHE" : "SEQ")),
 334                 zone->id,
 335                 (unsigned long long)chunk_block, nr_blocks);
 336
 337         if (dmz_is_rnd(zone) || dmz_is_cache(zone) ||
 338             chunk_block == zone->wp_block) {
 339                 /*
 340                  * zone is a random zone or it is a sequential zone
 341                  * and the BIO is aligned to the zone write pointer:
 342                  * direct write the zone.
 343                  */
 344                 return dmz_handle_direct_write(dmz, zone, bio,
 345                                                chunk_block, nr_blocks);
 346         }
 347
 348         /*
 349          * This is an unaligned write in a sequential zone:
 350          * use buffered write.
 351          */
 352         return dmz_handle_buffered_write(dmz, zone, bio, chunk_block, nr_blocks);
 353 }
 354
 355 /*
 356  * Process a discard BIO.
 357  */
 358 static int dmz_handle_discard(struct dmz_target *dmz, struct dm_zone *zone,
 359                               struct bio *bio)
 360 {
 361         struct dmz_metadata *zmd = dmz->metadata;
 362         sector_t block = dmz_bio_block(bio);
 363         unsigned int nr_blocks = dmz_bio_blocks(bio);
 364         sector_t chunk_block = dmz_chunk_block(zmd, block);
 365         int ret = 0;
 366
 367         /* For unmapped chunks, there is nothing to do */
 368         if (!zone)
 369                 return 0;
 370
 371         if (dmz_is_readonly(zone))
 372                 return -EROFS;
 373
 374         DMDEBUG("(%s): DISCARD chunk %llu -> zone %u, block %llu, %u blocks",
 375                 dmz_metadata_label(dmz->metadata),
 376                 (unsigned long long)dmz_bio_chunk(zmd, bio),
 377                 zone->id,
 378                 (unsigned long long)chunk_block, nr_blocks);
 379
 380         /*
 381          * Invalidate blocks in the data zone and its
 382          * buffer zone if one is mapped.
 383          */
 384         if (dmz_is_rnd(zone) || dmz_is_cache(zone) ||
 385             chunk_block < zone->wp_block)
 386                 ret = dmz_invalidate_blocks(zmd, zone, chunk_block, nr_blocks);
 387         if (ret == 0 && zone->bzone)
 388                 ret = dmz_invalidate_blocks(zmd, zone->bzone,
 389                                             chunk_block, nr_blocks);
 390         return ret;
 391 }
 392
 393 /*
 394  * Process a BIO.
 395  */
 396 static void dmz_handle_bio(struct dmz_target *dmz, struct dm_chunk_work *cw,
 397                            struct bio *bio)
 398 {
 399         struct dmz_bioctx *bioctx =
 400                 dm_per_bio_data(bio, sizeof(struct dmz_bioctx));
 401         struct dmz_metadata *zmd = dmz->metadata;
 402         struct dm_zone *zone;
 403         int ret;
 404
 405         dmz_lock_metadata(zmd);
 406
 407         /*
 408          * Get the data zone mapping the chunk. There may be no
 409          * mapping for read and discard. If a mapping is obtained,
 410          + the zone returned will be set to active state.
 411          */
 412         zone = dmz_get_chunk_mapping(zmd, dmz_bio_chunk(zmd, bio),
 413                                      bio_op(bio));
 414         if (IS_ERR(zone)) {
 415                 ret = PTR_ERR(zone);
 416                 goto out;
 417         }
 418
 419         /* Process the BIO */
 420         if (zone) {
 421                 dmz_activate_zone(zone);
 422                 bioctx->zone = zone;
 423                 dmz_reclaim_bio_acc(zone->dev->reclaim);
 424         }
 425
 426         switch (bio_op(bio)) {
 427         case REQ_OP_READ:
 428                 ret = dmz_handle_read(dmz, zone, bio);
 429                 break;
 430         case REQ_OP_WRITE:
 431                 ret = dmz_handle_write(dmz, zone, bio);
 432                 break;
 433         case REQ_OP_DISCARD:
 434         case REQ_OP_WRITE_ZEROES:
 435                 ret = dmz_handle_discard(dmz, zone, bio);
 436                 break;
 437         default:
 438                 DMERR("(%s): Unsupported BIO operation 0x%x",
 439                       dmz_metadata_label(dmz->metadata), bio_op(bio));
 440                 ret = -EIO;
 441         }
 442
 443         /*
 444          * Release the chunk mapping. This will check that the mapping
 445          * is still valid, that is, that the zone used still has valid blocks.
 446          */
 447         if (zone)
 448                 dmz_put_chunk_mapping(zmd, zone);
 449 out:
 450         dmz_bio_endio(bio, errno_to_blk_status(ret));
 451
 452         dmz_unlock_metadata(zmd);
 453 }
 454
 455 /*
 456  * Increment a chunk reference counter.
 457  */
 458 static inline void dmz_get_chunk_work(struct dm_chunk_work *cw)
 459 {
 460         refcount_inc(&cw->refcount);
 461 }
 462
 463 /*
 464  * Decrement a chunk work reference count and
 465  * free it if it becomes 0.
 466  */
 467 static void dmz_put_chunk_work(struct dm_chunk_work *cw)
 468 {
 469         if (refcount_dec_and_test(&cw->refcount)) {
 470                 WARN_ON(!bio_list_empty(&cw->bio_list));
 471                 radix_tree_delete(&cw->target->chunk_rxtree, cw->chunk);
 472                 kfree(cw);
 473         }
 474 }
 475
 476 /*
 477  * Chunk BIO work function.
 478  */
 479 static void dmz_chunk_work(struct work_struct *work)
 480 {
 481         struct dm_chunk_work *cw = container_of(work, struct dm_chunk_work, work);
 482         struct dmz_target *dmz = cw->target;
 483         struct bio *bio;
 484
 485         mutex_lock(&dmz->chunk_lock);
 486
 487         /* Process the chunk BIOs */
 488         while ((bio = bio_list_pop(&cw->bio_list))) {
 489                 mutex_unlock(&dmz->chunk_lock);
 490                 dmz_handle_bio(dmz, cw, bio);
 491                 mutex_lock(&dmz->chunk_lock);
 492                 dmz_put_chunk_work(cw);
 493         }
 494
 495         /* Queueing the work incremented the work refcount */
 496         dmz_put_chunk_work(cw);
 497
 498         mutex_unlock(&dmz->chunk_lock);
 499 }
 500
 501 /*
 502  * Flush work.
 503  */
 504 static void dmz_flush_work(struct work_struct *work)
 505 {
 506         struct dmz_target *dmz = container_of(work, struct dmz_target, flush_work.work);
 507         struct bio *bio;
 508         int ret;
 509
 510         /* Flush dirty metadata blocks */
 511         ret = dmz_flush_metadata(dmz->metadata);
 512         if (ret)
 513                 DMDEBUG("(%s): Metadata flush failed, rc=%d",
 514                         dmz_metadata_label(dmz->metadata), ret);
 515
 516         /* Process queued flush requests */
 517         while (1) {
 518                 spin_lock(&dmz->flush_lock);
 519                 bio = bio_list_pop(&dmz->flush_list);
 520                 spin_unlock(&dmz->flush_lock);
 521
 522                 if (!bio)
 523                         break;
 524
 525                 dmz_bio_endio(bio, errno_to_blk_status(ret));
 526         }
 527
 528         queue_delayed_work(dmz->flush_wq, &dmz->flush_work, DMZ_FLUSH_PERIOD);
 529 }
 530
 531 /*
 532  * Get a chunk work and start it to process a new BIO.
 533  * If the BIO chunk has no work yet, create one.
 534  */
 535 static int dmz_queue_chunk_work(struct dmz_target *dmz, struct bio *bio)
 536 {
 537         unsigned int chunk = dmz_bio_chunk(dmz->metadata, bio);
 538         struct dm_chunk_work *cw;
 539         int ret = 0;
 540
 541         mutex_lock(&dmz->chunk_lock);
 542
 543         /* Get the BIO chunk work. If one is not active yet, create one */
 544         cw = radix_tree_lookup(&dmz->chunk_rxtree, chunk);
 545         if (cw) {
 546                 dmz_get_chunk_work(cw);
 547         } else {
 548                 /* Create a new chunk work */
 549                 cw = kmalloc(sizeof(struct dm_chunk_work), GFP_NOIO);
 550                 if (unlikely(!cw)) {
 551                         ret = -ENOMEM;
 552                         goto out;
 553                 }
 554
 555                 INIT_WORK(&cw->work, dmz_chunk_work);
 556                 refcount_set(&cw->refcount, 1);
 557                 cw->target = dmz;
 558                 cw->chunk = chunk;
 559                 bio_list_init(&cw->bio_list);
 560
 561                 ret = radix_tree_insert(&dmz->chunk_rxtree, chunk, cw);
 562                 if (unlikely(ret)) {
 563                         kfree(cw);
 564                         goto out;
 565                 }
 566         }
 567
 568         bio_list_add(&cw->bio_list, bio);
 569
 570         if (queue_work(dmz->chunk_wq, &cw->work))
 571                 dmz_get_chunk_work(cw);
 572 out:
 573         mutex_unlock(&dmz->chunk_lock);
 574         return ret;
 575 }
 576
 577 /*
 578  * Check if the backing device is being removed. If it's on the way out,
 579  * start failing I/O. Reclaim and metadata components also call this
 580  * function to cleanly abort operation in the event of such failure.
 581  */
 582 bool dmz_bdev_is_dying(struct dmz_dev *dmz_dev)
 583 {
 584         if (dmz_dev->flags & DMZ_BDEV_DYING)
 585                 return true;
 586
 587         if (dmz_dev->flags & DMZ_CHECK_BDEV)
 588                 return !dmz_check_bdev(dmz_dev);
 589
 590         if (blk_queue_dying(bdev_get_queue(dmz_dev->bdev))) {
 591                 dmz_dev_warn(dmz_dev, "Backing device queue dying");
 592                 dmz_dev->flags |= DMZ_BDEV_DYING;
 593         }
 594
 595         return dmz_dev->flags & DMZ_BDEV_DYING;
 596 }
 597
 598 /*
 599  * Check the backing device availability. This detects such events as
 600  * backing device going offline due to errors, media removals, etc.
 601  * This check is less efficient than dmz_bdev_is_dying() and should
 602  * only be performed as a part of error handling.
 603  */
 604 bool dmz_check_bdev(struct dmz_dev *dmz_dev)
 605 {
 606         struct gendisk *disk;
 607
 608         dmz_dev->flags &= ~DMZ_CHECK_BDEV;
 609
 610         if (dmz_bdev_is_dying(dmz_dev))
 611                 return false;
 612
 613         disk = dmz_dev->bdev->bd_disk;
 614         if (disk->fops->check_events &&
 615             disk->fops->check_events(disk, 0) & DISK_EVENT_MEDIA_CHANGE) {
 616                 dmz_dev_warn(dmz_dev, "Backing device offline");
 617                 dmz_dev->flags |= DMZ_BDEV_DYING;
 618         }
 619
 620         return !(dmz_dev->flags & DMZ_BDEV_DYING);
 621 }
 622
 623 /*
 624  * Process a new BIO.
 625  */
 626 static int dmz_map(struct dm_target *ti, struct bio *bio)
 627 {
 628         struct dmz_target *dmz = ti->private;
 629         struct dmz_metadata *zmd = dmz->metadata;
 630         struct dmz_bioctx *bioctx = dm_per_bio_data(bio, sizeof(struct dmz_bioctx));
 631         sector_t sector = bio->bi_iter.bi_sector;
 632         unsigned int nr_sectors = bio_sectors(bio);
 633         sector_t chunk_sector;
 634         int ret;
 635
 636         if (dmz_dev_is_dying(zmd))
 637                 return DM_MAPIO_KILL;
 638
 639         DMDEBUG("(%s): BIO op %d sector %llu + %u => chunk %llu, block %llu, %u blocks",
 640                 dmz_metadata_label(zmd),
 641                 bio_op(bio), (unsigned long long)sector, nr_sectors,
 642                 (unsigned long long)dmz_bio_chunk(zmd, bio),
 643                 (unsigned long long)dmz_chunk_block(zmd, dmz_bio_block(bio)),
 644                 (unsigned int)dmz_bio_blocks(bio));
 645
 646         if (!nr_sectors && bio_op(bio) != REQ_OP_WRITE)
 647                 return DM_MAPIO_REMAPPED;
 648
 649         /* The BIO should be block aligned */
 650         if ((nr_sectors & DMZ_BLOCK_SECTORS_MASK) || (sector & DMZ_BLOCK_SECTORS_MASK))
 651                 return DM_MAPIO_KILL;
 652
 653         /* Initialize the BIO context */
 654         bioctx->dev = NULL;
 655         bioctx->zone = NULL;
 656         bioctx->bio = bio;
 657         refcount_set(&bioctx->ref, 1);
 658
 659         /* Set the BIO pending in the flush list */
 660         if (!nr_sectors && bio_op(bio) == REQ_OP_WRITE) {
 661                 spin_lock(&dmz->flush_lock);
 662                 bio_list_add(&dmz->flush_list, bio);
 663                 spin_unlock(&dmz->flush_lock);
 664                 mod_delayed_work(dmz->flush_wq, &dmz->flush_work, 0);
 665                 return DM_MAPIO_SUBMITTED;
 666         }
 667
 668         /* Split zone BIOs to fit entirely into a zone */
 669         chunk_sector = sector & (dmz_zone_nr_sectors(zmd) - 1);
 670         if (chunk_sector + nr_sectors > dmz_zone_nr_sectors(zmd))
 671                 dm_accept_partial_bio(bio, dmz_zone_nr_sectors(zmd) - chunk_sector);
 672
 673         /* Now ready to handle this BIO */
 674         ret = dmz_queue_chunk_work(dmz, bio);
 675         if (ret) {
 676                 DMDEBUG("(%s): BIO op %d, can't process chunk %llu, err %i",
 677                         dmz_metadata_label(zmd),
 678                         bio_op(bio), (u64)dmz_bio_chunk(zmd, bio),
 679                         ret);
 680                 return DM_MAPIO_REQUEUE;
 681         }
 682
 683         return DM_MAPIO_SUBMITTED;
 684 }
 685
 686 /*
 687  * Get zoned device information.
 688  */
 689 static int dmz_get_zoned_device(struct dm_target *ti, char *path,
 690                                 int idx, int nr_devs)
 691 {
 692         struct dmz_target *dmz = ti->private;
 693         struct dm_dev *ddev;
 694         struct dmz_dev *dev;
 695         int ret;
 696         struct block_device *bdev;
 697
 698         /* Get the target device */
 699         ret = dm_get_device(ti, path, dm_table_get_mode(ti->table), &ddev);
 700         if (ret) {
 701                 ti->error = "Get target device failed";
 702                 return ret;
 703         }
 704
 705         bdev = ddev->bdev;
 706         if (bdev_zoned_model(bdev) == BLK_ZONED_NONE) {
 707                 if (nr_devs == 1) {
 708                         ti->error = "Invalid regular device";
 709                         goto err;
 710                 }
 711                 if (idx != 0) {
 712                         ti->error = "First device must be a regular device";
 713                         goto err;
 714                 }
 715                 if (dmz->ddev[0]) {
 716                         ti->error = "Too many regular devices";
 717                         goto err;
 718                 }
 719                 dev = &dmz->dev[idx];
 720                 dev->flags = DMZ_BDEV_REGULAR;
 721         } else {
 722                 if (dmz->ddev[idx]) {
 723                         ti->error = "Too many zoned devices";
 724                         goto err;
 725                 }
 726                 if (nr_devs > 1 && idx == 0) {
 727                         ti->error = "First device must be a regular device";
 728                         goto err;
 729                 }
 730                 dev = &dmz->dev[idx];
 731         }
 732         dev->bdev = bdev;
 733         dev->dev_idx = idx;
 734         (void)bdevname(dev->bdev, dev->name);
 735
 736         dev->capacity = i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
 737         if (ti->begin) {
 738                 ti->error = "Partial mapping is not supported";
 739                 goto err;
 740         }
 741
 742         dmz->ddev[idx] = ddev;
 743
 744         return 0;
 745 err:
 746         dm_put_device(ti, ddev);
 747         return -EINVAL;
 748 }
 749
 750 /*
 751  * Cleanup zoned device information.
 752  */
 753 static void dmz_put_zoned_device(struct dm_target *ti)
 754 {
 755         struct dmz_target *dmz = ti->private;
 756         int i;
 757
 758         for (i = 0; i < dmz->nr_ddevs; i++) {
 759                 if (dmz->ddev[i]) {
 760                         dm_put_device(ti, dmz->ddev[i]);
 761                         dmz->ddev[i] = NULL;
 762                 }
 763         }
 764 }
 765
 766 static int dmz_fixup_devices(struct dm_target *ti)
 767 {
 768         struct dmz_target *dmz = ti->private;
 769         struct dmz_dev *reg_dev, *zoned_dev;
 770         struct request_queue *q;
 771         sector_t zone_nr_sectors = 0;
 772         int i;
 773
 774         /*
 775          * When we have more than on devices, the first one must be a
 776          * regular block device and the others zoned block devices.
 777          */
 778         if (dmz->nr_ddevs > 1) {
 779                 reg_dev = &dmz->dev[0];
 780                 if (!(reg_dev->flags & DMZ_BDEV_REGULAR)) {
 781                         ti->error = "Primary disk is not a regular device";
 782                         return -EINVAL;
 783                 }
 784                 for (i = 1; i < dmz->nr_ddevs; i++) {
 785                         zoned_dev = &dmz->dev[i];
 786                         if (zoned_dev->flags & DMZ_BDEV_REGULAR) {
 787                                 ti->error = "Secondary disk is not a zoned device";
 788                                 return -EINVAL;
 789                         }
 790                         q = bdev_get_queue(zoned_dev->bdev);
 791                         if (zone_nr_sectors &&
 792                             zone_nr_sectors != blk_queue_zone_sectors(q)) {
 793                                 ti->error = "Zone nr sectors mismatch";
 794                                 return -EINVAL;
 795                         }
 796                         zone_nr_sectors = blk_queue_zone_sectors(q);
 797                         zoned_dev->zone_nr_sectors = zone_nr_sectors;
 798                         zoned_dev->nr_zones =
 799                                 blkdev_nr_zones(zoned_dev->bdev->bd_disk);
 800                 }
 801         } else {
 802                 reg_dev = NULL;
 803                 zoned_dev = &dmz->dev[0];
 804                 if (zoned_dev->flags & DMZ_BDEV_REGULAR) {
 805                         ti->error = "Disk is not a zoned device";
 806                         return -EINVAL;
 807                 }
 808                 q = bdev_get_queue(zoned_dev->bdev);
 809                 zoned_dev->zone_nr_sectors = blk_queue_zone_sectors(q);
 810                 zoned_dev->nr_zones = blkdev_nr_zones(zoned_dev->bdev->bd_disk);
 811         }
 812
 813         if (reg_dev) {
 814                 sector_t zone_offset;
 815
 816                 reg_dev->zone_nr_sectors = zone_nr_sectors;
 817                 reg_dev->nr_zones =
 818                         DIV_ROUND_UP_SECTOR_T(reg_dev->capacity,
 819                                               reg_dev->zone_nr_sectors);
 820                 reg_dev->zone_offset = 0;
 821                 zone_offset = reg_dev->nr_zones;
 822                 for (i = 1; i < dmz->nr_ddevs; i++) {
 823                         dmz->dev[i].zone_offset = zone_offset;
 824                         zone_offset += dmz->dev[i].nr_zones;
 825                 }
 826         }
 827         return 0;
 828 }
 829
 830 /*
 831  * Setup target.
 832  */
 833 static int dmz_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 834 {
 835         struct dmz_target *dmz;
 836         int ret, i;
 837
 838         /* Check arguments */
 839         if (argc < 1) {
 840                 ti->error = "Invalid argument count";
 841                 return -EINVAL;
 842         }
 843
 844         /* Allocate and initialize the target descriptor */
 845         dmz = kzalloc(sizeof(struct dmz_target), GFP_KERNEL);
 846         if (!dmz) {
 847                 ti->error = "Unable to allocate the zoned target descriptor";
 848                 return -ENOMEM;
 849         }
 850         dmz->dev = kcalloc(argc, sizeof(struct dmz_dev), GFP_KERNEL);
 851         if (!dmz->dev) {
 852                 ti->error = "Unable to allocate the zoned device descriptors";
 853                 kfree(dmz);
 854                 return -ENOMEM;
 855         }
 856         dmz->ddev = kcalloc(argc, sizeof(struct dm_dev *), GFP_KERNEL);
 857         if (!dmz->ddev) {
 858                 ti->error = "Unable to allocate the dm device descriptors";
 859                 ret = -ENOMEM;
 860                 goto err;
 861         }
 862         dmz->nr_ddevs = argc;
 863
 864         ti->private = dmz;
 865
 866         /* Get the target zoned block device */
 867         for (i = 0; i < argc; i++) {
 868                 ret = dmz_get_zoned_device(ti, argv[i], i, argc);
 869                 if (ret)
 870                         goto err_dev;
 871         }
 872         ret = dmz_fixup_devices(ti);
 873         if (ret)
 874                 goto err_dev;
 875
 876         /* Initialize metadata */
 877         ret = dmz_ctr_metadata(dmz->dev, argc, &dmz->metadata,
 878                                dm_table_device_name(ti->table));
 879         if (ret) {
 880                 ti->error = "Metadata initialization failed";
 881                 goto err_dev;
 882         }
 883
 884         /* Set target (no write same support) */
 885         ti->max_io_len = dmz_zone_nr_sectors(dmz->metadata);
 886         ti->num_flush_bios = 1;
 887         ti->num_discard_bios = 1;
 888         ti->num_write_zeroes_bios = 1;
 889         ti->per_io_data_size = sizeof(struct dmz_bioctx);
 890         ti->flush_supported = true;
 891         ti->discards_supported = true;
 892
 893         /* The exposed capacity is the number of chunks that can be mapped */
 894         ti->len = (sector_t)dmz_nr_chunks(dmz->metadata) <<
 895                 dmz_zone_nr_sectors_shift(dmz->metadata);
 896
 897         /* Zone BIO */
 898         ret = bioset_init(&dmz->bio_set, DMZ_MIN_BIOS, 0, 0);
 899         if (ret) {
 900                 ti->error = "Create BIO set failed";
 901                 goto err_meta;
 902         }
 903
 904         /* Chunk BIO work */
 905         mutex_init(&dmz->chunk_lock);
 906         INIT_RADIX_TREE(&dmz->chunk_rxtree, GFP_NOIO);
 907         dmz->chunk_wq = alloc_workqueue("dmz_cwq_%s",
 908                                         WQ_MEM_RECLAIM | WQ_UNBOUND, 0,
 909                                         dmz_metadata_label(dmz->metadata));
 910         if (!dmz->chunk_wq) {
 911                 ti->error = "Create chunk workqueue failed";
 912                 ret = -ENOMEM;
 913                 goto err_bio;
 914         }
 915
 916         /* Flush work */
 917         spin_lock_init(&dmz->flush_lock);
 918         bio_list_init(&dmz->flush_list);
 919         INIT_DELAYED_WORK(&dmz->flush_work, dmz_flush_work);
 920         dmz->flush_wq = alloc_ordered_workqueue("dmz_fwq_%s", WQ_MEM_RECLAIM,
 921                                                 dmz_metadata_label(dmz->metadata));
 922         if (!dmz->flush_wq) {
 923                 ti->error = "Create flush workqueue failed";
 924                 ret = -ENOMEM;
 925                 goto err_cwq;
 926         }
 927         mod_delayed_work(dmz->flush_wq, &dmz->flush_work, DMZ_FLUSH_PERIOD);
 928
 929         /* Initialize reclaim */
 930         for (i = 0; i < dmz->nr_ddevs; i++) {
 931                 ret = dmz_ctr_reclaim(dmz->metadata, &dmz->dev[i].reclaim, i);
 932                 if (ret) {
 933                         ti->error = "Zone reclaim initialization failed";
 934                         goto err_fwq;
 935                 }
 936         }
 937
 938         DMINFO("(%s): Target device: %llu 512-byte logical sectors (%llu blocks)",
 939                dmz_metadata_label(dmz->metadata),
 940                (unsigned long long)ti->len,
 941                (unsigned long long)dmz_sect2blk(ti->len));
 942
 943         return 0;
 944 err_fwq:
 945         destroy_workqueue(dmz->flush_wq);
 946 err_cwq:
 947         destroy_workqueue(dmz->chunk_wq);
 948 err_bio:
 949         mutex_destroy(&dmz->chunk_lock);
 950         bioset_exit(&dmz->bio_set);
 951 err_meta:
 952         dmz_dtr_metadata(dmz->metadata);
 953 err_dev:
 954         dmz_put_zoned_device(ti);
 955 err:
 956         kfree(dmz->dev);
 957         kfree(dmz);
 958
 959         return ret;
 960 }
 961
 962 /*
 963  * Cleanup target.
 964  */
 965 static void dmz_dtr(struct dm_target *ti)
 966 {
 967         struct dmz_target *dmz = ti->private;
 968         int i;
 969
 970         flush_workqueue(dmz->chunk_wq);
 971         destroy_workqueue(dmz->chunk_wq);
 972
 973         for (i = 0; i < dmz->nr_ddevs; i++)
 974                 dmz_dtr_reclaim(dmz->dev[i].reclaim);
 975
 976         cancel_delayed_work_sync(&dmz->flush_work);
 977         destroy_workqueue(dmz->flush_wq);
 978
 979         (void) dmz_flush_metadata(dmz->metadata);
 980
 981         dmz_dtr_metadata(dmz->metadata);
 982
 983         bioset_exit(&dmz->bio_set);
 984
 985         dmz_put_zoned_device(ti);
 986
 987         mutex_destroy(&dmz->chunk_lock);
 988
 989         kfree(dmz->dev);
 990         kfree(dmz);
 991 }
 992
 993 /*
 994  * Setup target request queue limits.
 995  */
 996 static void dmz_io_hints(struct dm_target *ti, struct queue_limits *limits)
 997 {
 998         struct dmz_target *dmz = ti->private;
 999         unsigned int chunk_sectors = dmz_zone_nr_sectors(dmz->metadata);
1000
1001         limits->logical_block_size = DMZ_BLOCK_SIZE;
1002         limits->physical_block_size = DMZ_BLOCK_SIZE;
1003
1004         blk_limits_io_min(limits, DMZ_BLOCK_SIZE);
1005         blk_limits_io_opt(limits, DMZ_BLOCK_SIZE);
1006
1007         limits->discard_alignment = DMZ_BLOCK_SIZE;
1008         limits->discard_granularity = DMZ_BLOCK_SIZE;
1009         limits->max_discard_sectors = chunk_sectors;
1010         limits->max_hw_discard_sectors = chunk_sectors;
1011         limits->max_write_zeroes_sectors = chunk_sectors;
1012
1013         /* FS hint to try to align to the device zone size */
1014         limits->chunk_sectors = chunk_sectors;
1015         limits->max_sectors = chunk_sectors;
1016
1017         /* We are exposing a drive-managed zoned block device */
1018         limits->zoned = BLK_ZONED_NONE;
1019 }
1020
1021 /*
1022  * Pass on ioctl to the backend device.
1023  */
1024 static int dmz_prepare_ioctl(struct dm_target *ti, struct block_device **bdev)
1025 {
1026         struct dmz_target *dmz = ti->private;
1027         struct dmz_dev *dev = &dmz->dev[0];
1028
1029         if (!dmz_check_bdev(dev))
1030                 return -EIO;
1031
1032         *bdev = dev->bdev;
1033
1034         return 0;
1035 }
1036
1037 /*
1038  * Stop works on suspend.
1039  */
1040 static void dmz_suspend(struct dm_target *ti)
1041 {
1042         struct dmz_target *dmz = ti->private;
1043         int i;
1044
1045         flush_workqueue(dmz->chunk_wq);
1046         for (i = 0; i < dmz->nr_ddevs; i++)
1047                 dmz_suspend_reclaim(dmz->dev[i].reclaim);
1048         cancel_delayed_work_sync(&dmz->flush_work);
1049 }
1050
1051 /*
1052  * Restart works on resume or if suspend failed.
1053  */
1054 static void dmz_resume(struct dm_target *ti)
1055 {
1056         struct dmz_target *dmz = ti->private;
1057         int i;
1058
1059         queue_delayed_work(dmz->flush_wq, &dmz->flush_work, DMZ_FLUSH_PERIOD);
1060         for (i = 0; i < dmz->nr_ddevs; i++)
1061                 dmz_resume_reclaim(dmz->dev[i].reclaim);
1062 }
1063
1064 static int dmz_iterate_devices(struct dm_target *ti,
1065                                iterate_devices_callout_fn fn, void *data)
1066 {
1067         struct dmz_target *dmz = ti->private;
1068         unsigned int zone_nr_sectors = dmz_zone_nr_sectors(dmz->metadata);
1069         sector_t capacity;
1070         int i, r;
1071
1072         for (i = 0; i < dmz->nr_ddevs; i++) {
1073                 capacity = dmz->dev[i].capacity & ~(zone_nr_sectors - 1);
1074                 r = fn(ti, dmz->ddev[i], 0, capacity, data);
1075                 if (r)
1076                         break;
1077         }
1078         return r;
1079 }
1080
1081 static void dmz_status(struct dm_target *ti, status_type_t type,
1082                        unsigned int status_flags, char *result,
1083                        unsigned int maxlen)
1084 {
1085         struct dmz_target *dmz = ti->private;
1086         ssize_t sz = 0;
1087         char buf[BDEVNAME_SIZE];
1088         struct dmz_dev *dev;
1089         int i;
1090
1091         switch (type) {
1092         case STATUSTYPE_INFO:
1093                 DMEMIT("%u zones %u/%u cache",
1094                        dmz_nr_zones(dmz->metadata),
1095                        dmz_nr_unmap_cache_zones(dmz->metadata),
1096                        dmz_nr_cache_zones(dmz->metadata));
1097                 for (i = 0; i < dmz->nr_ddevs; i++) {
1098                         /*
1099                          * For a multi-device setup the first device
1100                          * contains only cache zones.
1101                          */
1102                         if ((i == 0) &&
1103                             (dmz_nr_cache_zones(dmz->metadata) > 0))
1104                                 continue;
1105                         DMEMIT(" %u/%u random %u/%u sequential",
1106                                dmz_nr_unmap_rnd_zones(dmz->metadata, i),
1107                                dmz_nr_rnd_zones(dmz->metadata, i),
1108                                dmz_nr_unmap_seq_zones(dmz->metadata, i),
1109                                dmz_nr_seq_zones(dmz->metadata, i));
1110                 }
1111                 break;
1112         case STATUSTYPE_TABLE:
1113                 dev = &dmz->dev[0];
1114                 format_dev_t(buf, dev->bdev->bd_dev);
1115                 DMEMIT("%s", buf);
1116                 for (i = 1; i < dmz->nr_ddevs; i++) {
1117                         dev = &dmz->dev[i];
1118                         format_dev_t(buf, dev->bdev->bd_dev);
1119                         DMEMIT(" %s", buf);
1120                 }
1121                 break;
1122         }
1123         return;
1124 }
1125
1126 static int dmz_message(struct dm_target *ti, unsigned int argc, char **argv,
1127                        char *result, unsigned int maxlen)
1128 {
1129         struct dmz_target *dmz = ti->private;
1130         int r = -EINVAL;
1131
1132         if (!strcasecmp(argv[0], "reclaim")) {
1133                 int i;
1134
1135                 for (i = 0; i < dmz->nr_ddevs; i++)
1136                         dmz_schedule_reclaim(dmz->dev[i].reclaim);
1137                 r = 0;
1138         } else
1139                 DMERR("unrecognized message %s", argv[0]);
1140         return r;
1141 }
1142
1143 static struct target_type dmz_type = {
1144         .name            = "zoned",
1145         .version         = {2, 0, 0},
1146         .features        = DM_TARGET_SINGLETON | DM_TARGET_MIXED_ZONED_MODEL,
1147         .module          = THIS_MODULE,
1148         .ctr             = dmz_ctr,
1149         .dtr             = dmz_dtr,
1150         .map             = dmz_map,
1151         .io_hints        = dmz_io_hints,
1152         .prepare_ioctl   = dmz_prepare_ioctl,
1153         .postsuspend     = dmz_suspend,
1154         .resume          = dmz_resume,
1155         .iterate_devices = dmz_iterate_devices,
1156         .status          = dmz_status,
1157         .message         = dmz_message,
1158 };
1159
1160 static int __init dmz_init(void)
1161 {
1162         return dm_register_target(&dmz_type);
1163 }
1164
1165 static void __exit dmz_exit(void)
1166 {
1167         dm_unregister_target(&dmz_type);
1168 }
1169
1170 module_init(dmz_init);
1171 module_exit(dmz_exit);
1172
1173 MODULE_DESCRIPTION(DM_NAME " target for zoned block devices");
1174 MODULE_AUTHOR("Damien Le Moal <damien.lemoal@wdc.com>");
1175 MODULE_LICENSE("GPL");