fs/btrfs/dev-replace.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) STRATO AG 2012.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/bio.h>
   8 #include <linux/slab.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/kthread.h>
  11 #include <linux/math64.h>
  12 #include "misc.h"
  13 #include "ctree.h"
  14 #include "disk-io.h"
  15 #include "transaction.h"
  16 #include "volumes.h"
  17 #include "async-thread.h"
  18 #include "dev-replace.h"
  19 #include "sysfs.h"
  20 #include "zoned.h"
  21 #include "block-group.h"
  22 #include "fs.h"
  23 #include "accessors.h"
  24 #include "scrub.h"
  25
  26 /*
  27  * Device replace overview
  28  *
  29  * [Objective]
  30  * To copy all extents (both new and on-disk) from source device to target
  31  * device, while still keeping the filesystem read-write.
  32  *
  33  * [Method]
  34  * There are two main methods involved:
  35  *
  36  * - Write duplication
  37  *
  38  *   All new writes will be written to both target and source devices, so even
  39  *   if replace gets canceled, sources device still contains up-to-date data.
  40  *
  41  *   Location:          handle_ops_on_dev_replace() from btrfs_map_block()
  42  *   Start:             btrfs_dev_replace_start()
  43  *   End:               btrfs_dev_replace_finishing()
  44  *   Content:           Latest data/metadata
  45  *
  46  * - Copy existing extents
  47  *
  48  *   This happens by reusing scrub facility, as scrub also iterates through
  49  *   existing extents from commit root.
  50  *
  51  *   Location:          scrub_write_block_to_dev_replace() from
  52  *                      scrub_block_complete()
  53  *   Content:           Data/meta from commit root.
  54  *
  55  * Due to the content difference, we need to avoid nocow write when dev-replace
  56  * is happening.  This is done by marking the block group read-only and waiting
  57  * for NOCOW writes.
  58  *
  59  * After replace is done, the finishing part is done by swapping the target and
  60  * source devices.
  61  *
  62  *   Location:          btrfs_dev_replace_update_device_in_mapping_tree() from
  63  *                      btrfs_dev_replace_finishing()
  64  */
  65
  66 static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
  67                                        int scrub_ret);
  68 static int btrfs_dev_replace_kthread(void *data);
  69
  70 int btrfs_init_dev_replace(struct btrfs_fs_info *fs_info)
  71 {
  72         struct btrfs_dev_lookup_args args = { .devid = BTRFS_DEV_REPLACE_DEVID };
  73         struct btrfs_key key;
  74         struct btrfs_root *dev_root = fs_info->dev_root;
  75         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
  76         struct extent_buffer *eb;
  77         int slot;
  78         int ret = 0;
  79         BTRFS_PATH_AUTO_FREE(path);
  80         int item_size;
  81         struct btrfs_dev_replace_item *ptr;
  82         u64 src_devid;
  83
  84         if (!dev_root)
  85                 return 0;
  86
  87         path = btrfs_alloc_path();
  88         if (!path)
  89                 return -ENOMEM;
  90
  91         key.objectid = 0;
  92         key.type = BTRFS_DEV_REPLACE_KEY;
  93         key.offset = 0;
  94         ret = btrfs_search_slot(NULL, dev_root, &key, path, 0, 0);
  95         if (ret) {
  96 no_valid_dev_replace_entry_found:
  97                 /*
  98                  * We don't have a replace item or it's corrupted.  If there is
  99                  * a replace target, fail the mount.
 100                  */
 101                 if (btrfs_find_device(fs_info->fs_devices, &args)) {
 102                         btrfs_err(fs_info,
 103                         "found replace target device without a valid replace item");
 104                         return -EUCLEAN;
 105                 }
 106                 dev_replace->replace_state =
 107                         BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED;
 108                 dev_replace->cont_reading_from_srcdev_mode =
 109                     BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_ALWAYS;
 110                 dev_replace->time_started = 0;
 111                 dev_replace->time_stopped = 0;
 112                 atomic64_set(&dev_replace->num_write_errors, 0);
 113                 atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
 114                 dev_replace->cursor_left = 0;
 115                 dev_replace->committed_cursor_left = 0;
 116                 dev_replace->cursor_left_last_write_of_item = 0;
 117                 dev_replace->cursor_right = 0;
 118                 dev_replace->srcdev = NULL;
 119                 dev_replace->tgtdev = NULL;
 120                 dev_replace->is_valid = 0;
 121                 dev_replace->item_needs_writeback = 0;
 122                 return 0;
 123         }
 124         slot = path->slots[0];
 125         eb = path->nodes[0];
 126         item_size = btrfs_item_size(eb, slot);
 127         ptr = btrfs_item_ptr(eb, slot, struct btrfs_dev_replace_item);
 128
 129         if (item_size != sizeof(struct btrfs_dev_replace_item)) {
 130                 btrfs_warn(fs_info,
 131                         "dev_replace entry found has unexpected size, ignore entry");
 132                 goto no_valid_dev_replace_entry_found;
 133         }
 134
 135         src_devid = btrfs_dev_replace_src_devid(eb, ptr);
 136         dev_replace->cont_reading_from_srcdev_mode =
 137                 btrfs_dev_replace_cont_reading_from_srcdev_mode(eb, ptr);
 138         dev_replace->replace_state = btrfs_dev_replace_replace_state(eb, ptr);
 139         dev_replace->time_started = btrfs_dev_replace_time_started(eb, ptr);
 140         dev_replace->time_stopped =
 141                 btrfs_dev_replace_time_stopped(eb, ptr);
 142         atomic64_set(&dev_replace->num_write_errors,
 143                      btrfs_dev_replace_num_write_errors(eb, ptr));
 144         atomic64_set(&dev_replace->num_uncorrectable_read_errors,
 145                      btrfs_dev_replace_num_uncorrectable_read_errors(eb, ptr));
 146         dev_replace->cursor_left = btrfs_dev_replace_cursor_left(eb, ptr);
 147         dev_replace->committed_cursor_left = dev_replace->cursor_left;
 148         dev_replace->cursor_left_last_write_of_item = dev_replace->cursor_left;
 149         dev_replace->cursor_right = btrfs_dev_replace_cursor_right(eb, ptr);
 150         dev_replace->is_valid = 1;
 151
 152         dev_replace->item_needs_writeback = 0;
 153         switch (dev_replace->replace_state) {
 154         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
 155         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
 156         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
 157                 /*
 158                  * We don't have an active replace item but if there is a
 159                  * replace target, fail the mount.
 160                  */
 161                 if (btrfs_find_device(fs_info->fs_devices, &args)) {
 162                         btrfs_err(fs_info,
 163 "replace without active item, run 'device scan --forget' on the target device");
 164                         ret = -EUCLEAN;
 165                 } else {
 166                         dev_replace->srcdev = NULL;
 167                         dev_replace->tgtdev = NULL;
 168                 }
 169                 break;
 170         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
 171         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
 172                 dev_replace->tgtdev = btrfs_find_device(fs_info->fs_devices, &args);
 173                 args.devid = src_devid;
 174                 dev_replace->srcdev = btrfs_find_device(fs_info->fs_devices, &args);
 175
 176                 /*
 177                  * allow 'btrfs dev replace_cancel' if src/tgt device is
 178                  * missing
 179                  */
 180                 if (!dev_replace->srcdev &&
 181                     !btrfs_test_opt(fs_info, DEGRADED)) {
 182                         ret = -EIO;
 183                         btrfs_warn(fs_info,
 184                            "cannot mount because device replace operation is ongoing and");
 185                         btrfs_warn(fs_info,
 186                            "srcdev (devid %llu) is missing, need to run 'btrfs dev scan'?",
 187                            src_devid);
 188                 }
 189                 if (!dev_replace->tgtdev &&
 190                     !btrfs_test_opt(fs_info, DEGRADED)) {
 191                         ret = -EIO;
 192                         btrfs_warn(fs_info,
 193                            "cannot mount because device replace operation is ongoing and");
 194                         btrfs_warn(fs_info,
 195                            "tgtdev (devid %llu) is missing, need to run 'btrfs dev scan'?",
 196                                 BTRFS_DEV_REPLACE_DEVID);
 197                 }
 198                 if (dev_replace->tgtdev) {
 199                         if (dev_replace->srcdev) {
 200                                 dev_replace->tgtdev->total_bytes =
 201                                         dev_replace->srcdev->total_bytes;
 202                                 dev_replace->tgtdev->disk_total_bytes =
 203                                         dev_replace->srcdev->disk_total_bytes;
 204                                 dev_replace->tgtdev->commit_total_bytes =
 205                                         dev_replace->srcdev->commit_total_bytes;
 206                                 dev_replace->tgtdev->bytes_used =
 207                                         dev_replace->srcdev->bytes_used;
 208                                 dev_replace->tgtdev->commit_bytes_used =
 209                                         dev_replace->srcdev->commit_bytes_used;
 210                         }
 211                         set_bit(BTRFS_DEV_STATE_REPLACE_TGT,
 212                                 &dev_replace->tgtdev->dev_state);
 213
 214                         WARN_ON(fs_info->fs_devices->rw_devices == 0);
 215                         dev_replace->tgtdev->io_width = fs_info->sectorsize;
 216                         dev_replace->tgtdev->io_align = fs_info->sectorsize;
 217                         dev_replace->tgtdev->sector_size = fs_info->sectorsize;
 218                         dev_replace->tgtdev->fs_info = fs_info;
 219                         set_bit(BTRFS_DEV_STATE_IN_FS_METADATA,
 220                                 &dev_replace->tgtdev->dev_state);
 221                 }
 222                 break;
 223         }
 224
 225         return ret;
 226 }
 227
 228 /*
 229  * Initialize a new device for device replace target from a given source dev
 230  * and path.
 231  *
 232  * Return 0 and new device in @device_out, otherwise return < 0
 233  */
 234 static int btrfs_init_dev_replace_tgtdev(struct btrfs_fs_info *fs_info,
 235                                   const char *device_path,
 236                                   struct btrfs_device *srcdev,
 237                                   struct btrfs_device **device_out)
 238 {
 239         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 240         struct btrfs_device *device;
 241         struct file *bdev_file;
 242         struct block_device *bdev;
 243         u64 devid = BTRFS_DEV_REPLACE_DEVID;
 244         int ret = 0;
 245
 246         *device_out = NULL;
 247         if (srcdev->fs_devices->seeding) {
 248                 btrfs_err(fs_info, "the filesystem is a seed filesystem!");
 249                 return -EINVAL;
 250         }
 251
 252         bdev_file = bdev_file_open_by_path(device_path, BLK_OPEN_WRITE,
 253                                         fs_info->bdev_holder, NULL);
 254         if (IS_ERR(bdev_file)) {
 255                 btrfs_err(fs_info, "target device %s is invalid!", device_path);
 256                 return PTR_ERR(bdev_file);
 257         }
 258         bdev = file_bdev(bdev_file);
 259
 260         if (!btrfs_check_device_zone_type(fs_info, bdev)) {
 261                 btrfs_err(fs_info,
 262                 "dev-replace: zoned type of target device mismatch with filesystem");
 263                 ret = -EINVAL;
 264                 goto error;
 265         }
 266
 267         sync_blockdev(bdev);
 268
 269         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 270                 if (device->bdev == bdev) {
 271                         btrfs_err(fs_info,
 272                                   "target device is in the filesystem!");
 273                         ret = -EEXIST;
 274                         goto error;
 275                 }
 276         }
 277
 278
 279         if (bdev_nr_bytes(bdev) < btrfs_device_get_total_bytes(srcdev)) {
 280                 btrfs_err(fs_info,
 281                           "target device is smaller than source device!");
 282                 ret = -EINVAL;
 283                 goto error;
 284         }
 285
 286
 287         device = btrfs_alloc_device(NULL, &devid, NULL, device_path);
 288         if (IS_ERR(device)) {
 289                 ret = PTR_ERR(device);
 290                 goto error;
 291         }
 292
 293         ret = lookup_bdev(device_path, &device->devt);
 294         if (ret)
 295                 goto error;
 296
 297         set_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
 298         device->generation = 0;
 299         device->io_width = fs_info->sectorsize;
 300         device->io_align = fs_info->sectorsize;
 301         device->sector_size = fs_info->sectorsize;
 302         device->total_bytes = btrfs_device_get_total_bytes(srcdev);
 303         device->disk_total_bytes = btrfs_device_get_disk_total_bytes(srcdev);
 304         device->bytes_used = btrfs_device_get_bytes_used(srcdev);
 305         device->commit_total_bytes = srcdev->commit_total_bytes;
 306         device->commit_bytes_used = device->bytes_used;
 307         device->fs_info = fs_info;
 308         device->bdev = bdev;
 309         device->bdev_file = bdev_file;
 310         set_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
 311         set_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state);
 312         device->dev_stats_valid = 1;
 313         set_blocksize(bdev_file, BTRFS_BDEV_BLOCKSIZE);
 314         device->fs_devices = fs_devices;
 315
 316         ret = btrfs_get_dev_zone_info(device, false);
 317         if (ret)
 318                 goto error;
 319
 320         mutex_lock(&fs_devices->device_list_mutex);
 321         list_add(&device->dev_list, &fs_devices->devices);
 322         fs_devices->num_devices++;
 323         fs_devices->open_devices++;
 324         mutex_unlock(&fs_devices->device_list_mutex);
 325
 326         *device_out = device;
 327         return 0;
 328
 329 error:
 330         fput(bdev_file);
 331         return ret;
 332 }
 333
 334 /*
 335  * called from commit_transaction. Writes changed device replace state to
 336  * disk.
 337  */
 338 int btrfs_run_dev_replace(struct btrfs_trans_handle *trans)
 339 {
 340         struct btrfs_fs_info *fs_info = trans->fs_info;
 341         int ret;
 342         struct btrfs_root *dev_root = fs_info->dev_root;
 343         BTRFS_PATH_AUTO_FREE(path);
 344         struct btrfs_key key;
 345         struct extent_buffer *eb;
 346         struct btrfs_dev_replace_item *ptr;
 347         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 348
 349         down_read(&dev_replace->rwsem);
 350         if (!dev_replace->is_valid ||
 351             !dev_replace->item_needs_writeback) {
 352                 up_read(&dev_replace->rwsem);
 353                 return 0;
 354         }
 355         up_read(&dev_replace->rwsem);
 356
 357         key.objectid = 0;
 358         key.type = BTRFS_DEV_REPLACE_KEY;
 359         key.offset = 0;
 360
 361         path = btrfs_alloc_path();
 362         if (!path)
 363                 return -ENOMEM;
 364
 365         ret = btrfs_search_slot(trans, dev_root, &key, path, -1, 1);
 366         if (ret < 0) {
 367                 btrfs_warn(fs_info,
 368                            "error %d while searching for dev_replace item!",
 369                            ret);
 370                 return ret;
 371         }
 372
 373         if (ret == 0 &&
 374             btrfs_item_size(path->nodes[0], path->slots[0]) < sizeof(*ptr)) {
 375                 /*
 376                  * need to delete old one and insert a new one.
 377                  * Since no attempt is made to recover any old state, if the
 378                  * dev_replace state is 'running', the data on the target
 379                  * drive is lost.
 380                  * It would be possible to recover the state: just make sure
 381                  * that the beginning of the item is never changed and always
 382                  * contains all the essential information. Then read this
 383                  * minimal set of information and use it as a base for the
 384                  * new state.
 385                  */
 386                 ret = btrfs_del_item(trans, dev_root, path);
 387                 if (ret != 0) {
 388                         btrfs_warn(fs_info,
 389                                    "delete too small dev_replace item failed %d!",
 390                                    ret);
 391                         return ret;
 392                 }
 393                 ret = 1;
 394         }
 395
 396         if (ret == 1) {
 397                 /* need to insert a new item */
 398                 btrfs_release_path(path);
 399                 ret = btrfs_insert_empty_item(trans, dev_root, path,
 400                                               &key, sizeof(*ptr));
 401                 if (ret < 0) {
 402                         btrfs_warn(fs_info,
 403                                    "insert dev_replace item failed %d!", ret);
 404                         return ret;
 405                 }
 406         }
 407
 408         eb = path->nodes[0];
 409         ptr = btrfs_item_ptr(eb, path->slots[0],
 410                              struct btrfs_dev_replace_item);
 411
 412         down_write(&dev_replace->rwsem);
 413         if (dev_replace->srcdev)
 414                 btrfs_set_dev_replace_src_devid(eb, ptr,
 415                         dev_replace->srcdev->devid);
 416         else
 417                 btrfs_set_dev_replace_src_devid(eb, ptr, (u64)-1);
 418         btrfs_set_dev_replace_cont_reading_from_srcdev_mode(eb, ptr,
 419                 dev_replace->cont_reading_from_srcdev_mode);
 420         btrfs_set_dev_replace_replace_state(eb, ptr,
 421                 dev_replace->replace_state);
 422         btrfs_set_dev_replace_time_started(eb, ptr, dev_replace->time_started);
 423         btrfs_set_dev_replace_time_stopped(eb, ptr, dev_replace->time_stopped);
 424         btrfs_set_dev_replace_num_write_errors(eb, ptr,
 425                 atomic64_read(&dev_replace->num_write_errors));
 426         btrfs_set_dev_replace_num_uncorrectable_read_errors(eb, ptr,
 427                 atomic64_read(&dev_replace->num_uncorrectable_read_errors));
 428         dev_replace->cursor_left_last_write_of_item =
 429                 dev_replace->cursor_left;
 430         btrfs_set_dev_replace_cursor_left(eb, ptr,
 431                 dev_replace->cursor_left_last_write_of_item);
 432         btrfs_set_dev_replace_cursor_right(eb, ptr,
 433                 dev_replace->cursor_right);
 434         dev_replace->item_needs_writeback = 0;
 435         up_write(&dev_replace->rwsem);
 436
 437         return ret;
 438 }
 439
 440 static int mark_block_group_to_copy(struct btrfs_fs_info *fs_info,
 441                                     struct btrfs_device *src_dev)
 442 {
 443         struct btrfs_path *path;
 444         struct btrfs_key key;
 445         struct btrfs_key found_key;
 446         struct btrfs_root *root = fs_info->dev_root;
 447         struct btrfs_dev_extent *dev_extent = NULL;
 448         struct btrfs_block_group *cache;
 449         struct btrfs_trans_handle *trans;
 450         int iter_ret = 0;
 451         int ret = 0;
 452         u64 chunk_offset;
 453
 454         /* Do not use "to_copy" on non zoned filesystem for now */
 455         if (!btrfs_is_zoned(fs_info))
 456                 return 0;
 457
 458         mutex_lock(&fs_info->chunk_mutex);
 459
 460         /* Ensure we don't have pending new block group */
 461         spin_lock(&fs_info->trans_lock);
 462         while (fs_info->running_transaction &&
 463                !list_empty(&fs_info->running_transaction->dev_update_list)) {
 464                 spin_unlock(&fs_info->trans_lock);
 465                 mutex_unlock(&fs_info->chunk_mutex);
 466                 trans = btrfs_attach_transaction(root);
 467                 if (IS_ERR(trans)) {
 468                         ret = PTR_ERR(trans);
 469                         mutex_lock(&fs_info->chunk_mutex);
 470                         if (ret == -ENOENT) {
 471                                 spin_lock(&fs_info->trans_lock);
 472                                 continue;
 473                         } else {
 474                                 goto unlock;
 475                         }
 476                 }
 477
 478                 ret = btrfs_commit_transaction(trans);
 479                 mutex_lock(&fs_info->chunk_mutex);
 480                 if (ret)
 481                         goto unlock;
 482
 483                 spin_lock(&fs_info->trans_lock);
 484         }
 485         spin_unlock(&fs_info->trans_lock);
 486
 487         path = btrfs_alloc_path();
 488         if (!path) {
 489                 ret = -ENOMEM;
 490                 goto unlock;
 491         }
 492
 493         path->reada = READA_FORWARD;
 494         path->search_commit_root = 1;
 495         path->skip_locking = 1;
 496
 497         key.objectid = src_dev->devid;
 498         key.type = BTRFS_DEV_EXTENT_KEY;
 499         key.offset = 0;
 500
 501         btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
 502                 struct extent_buffer *leaf = path->nodes[0];
 503
 504                 if (found_key.objectid != src_dev->devid)
 505                         break;
 506
 507                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
 508                         break;
 509
 510                 if (found_key.offset < key.offset)
 511                         break;
 512
 513                 dev_extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
 514
 515                 chunk_offset = btrfs_dev_extent_chunk_offset(leaf, dev_extent);
 516
 517                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
 518                 if (!cache)
 519                         continue;
 520
 521                 set_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags);
 522                 btrfs_put_block_group(cache);
 523         }
 524         if (iter_ret < 0)
 525                 ret = iter_ret;
 526
 527         btrfs_free_path(path);
 528 unlock:
 529         mutex_unlock(&fs_info->chunk_mutex);
 530
 531         return ret;
 532 }
 533
 534 bool btrfs_finish_block_group_to_copy(struct btrfs_device *srcdev,
 535                                       struct btrfs_block_group *cache,
 536                                       u64 physical)
 537 {
 538         struct btrfs_fs_info *fs_info = cache->fs_info;
 539         struct btrfs_chunk_map *map;
 540         u64 chunk_offset = cache->start;
 541         int num_extents, cur_extent;
 542         int i;
 543
 544         /* Do not use "to_copy" on non zoned filesystem for now */
 545         if (!btrfs_is_zoned(fs_info))
 546                 return true;
 547
 548         spin_lock(&cache->lock);
 549         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
 550                 spin_unlock(&cache->lock);
 551                 return true;
 552         }
 553         spin_unlock(&cache->lock);
 554
 555         map = btrfs_get_chunk_map(fs_info, chunk_offset, 1);
 556         ASSERT(!IS_ERR(map));
 557
 558         num_extents = 0;
 559         cur_extent = 0;
 560         for (i = 0; i < map->num_stripes; i++) {
 561                 /* We have more device extent to copy */
 562                 if (srcdev != map->stripes[i].dev)
 563                         continue;
 564
 565                 num_extents++;
 566                 if (physical == map->stripes[i].physical)
 567                         cur_extent = i;
 568         }
 569
 570         btrfs_free_chunk_map(map);
 571
 572         if (num_extents > 1 && cur_extent < num_extents - 1) {
 573                 /*
 574                  * Has more stripes on this device. Keep this block group
 575                  * readonly until we finish all the stripes.
 576                  */
 577                 return false;
 578         }
 579
 580         /* Last stripe on this device */
 581         clear_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags);
 582
 583         return true;
 584 }
 585
 586 static int btrfs_dev_replace_start(struct btrfs_fs_info *fs_info,
 587                 const char *tgtdev_name, u64 srcdevid, const char *srcdev_name,
 588                 int read_src)
 589 {
 590         struct btrfs_root *root = fs_info->dev_root;
 591         struct btrfs_trans_handle *trans;
 592         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 593         int ret;
 594         struct btrfs_device *tgt_device = NULL;
 595         struct btrfs_device *src_device = NULL;
 596
 597         src_device = btrfs_find_device_by_devspec(fs_info, srcdevid,
 598                                                   srcdev_name);
 599         if (IS_ERR(src_device))
 600                 return PTR_ERR(src_device);
 601
 602         if (btrfs_pinned_by_swapfile(fs_info, src_device)) {
 603                 btrfs_warn_in_rcu(fs_info,
 604           "cannot replace device %s (devid %llu) due to active swapfile",
 605                         btrfs_dev_name(src_device), src_device->devid);
 606                 return -ETXTBSY;
 607         }
 608
 609         /*
 610          * Here we commit the transaction to make sure commit_total_bytes
 611          * of all the devices are updated.
 612          */
 613         trans = btrfs_attach_transaction(root);
 614         if (!IS_ERR(trans)) {
 615                 ret = btrfs_commit_transaction(trans);
 616                 if (ret)
 617                         return ret;
 618         } else if (PTR_ERR(trans) != -ENOENT) {
 619                 return PTR_ERR(trans);
 620         }
 621
 622         ret = btrfs_init_dev_replace_tgtdev(fs_info, tgtdev_name,
 623                                             src_device, &tgt_device);
 624         if (ret)
 625                 return ret;
 626
 627         ret = mark_block_group_to_copy(fs_info, src_device);
 628         if (ret)
 629                 return ret;
 630
 631         down_write(&dev_replace->rwsem);
 632         dev_replace->replace_task = current;
 633         switch (dev_replace->replace_state) {
 634         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
 635         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
 636         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
 637                 break;
 638         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
 639         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
 640                 DEBUG_WARN("unexpected STARTED ot SUSPENDED dev-replace state");
 641                 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_ALREADY_STARTED;
 642                 up_write(&dev_replace->rwsem);
 643                 goto leave;
 644         }
 645
 646         dev_replace->cont_reading_from_srcdev_mode = read_src;
 647         dev_replace->srcdev = src_device;
 648         dev_replace->tgtdev = tgt_device;
 649
 650         btrfs_info_in_rcu(fs_info,
 651                       "dev_replace from %s (devid %llu) to %s started",
 652                       btrfs_dev_name(src_device),
 653                       src_device->devid,
 654                       btrfs_dev_name(tgt_device));
 655
 656         /*
 657          * from now on, the writes to the srcdev are all duplicated to
 658          * go to the tgtdev as well (refer to btrfs_map_block()).
 659          */
 660         dev_replace->replace_state = BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
 661         dev_replace->time_started = ktime_get_real_seconds();
 662         dev_replace->cursor_left = 0;
 663         dev_replace->committed_cursor_left = 0;
 664         dev_replace->cursor_left_last_write_of_item = 0;
 665         dev_replace->cursor_right = 0;
 666         dev_replace->is_valid = 1;
 667         dev_replace->item_needs_writeback = 1;
 668         atomic64_set(&dev_replace->num_write_errors, 0);
 669         atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
 670         up_write(&dev_replace->rwsem);
 671
 672         ret = btrfs_sysfs_add_device(tgt_device);
 673         if (ret)
 674                 btrfs_err(fs_info, "kobj add dev failed %d", ret);
 675
 676         btrfs_wait_ordered_roots(fs_info, U64_MAX, NULL);
 677
 678         /*
 679          * Commit dev_replace state and reserve 1 item for it.
 680          * This is crucial to ensure we won't miss copying extents for new block
 681          * groups that are allocated after we started the device replace, and
 682          * must be done after setting up the device replace state.
 683          */
 684         trans = btrfs_start_transaction(root, 1);
 685         if (IS_ERR(trans)) {
 686                 ret = PTR_ERR(trans);
 687                 down_write(&dev_replace->rwsem);
 688                 dev_replace->replace_state =
 689                         BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED;
 690                 dev_replace->srcdev = NULL;
 691                 dev_replace->tgtdev = NULL;
 692                 up_write(&dev_replace->rwsem);
 693                 goto leave;
 694         }
 695
 696         ret = btrfs_commit_transaction(trans);
 697         WARN_ON(ret);
 698
 699         /* the disk copy procedure reuses the scrub code */
 700         ret = btrfs_scrub_dev(fs_info, src_device->devid, 0,
 701                               btrfs_device_get_total_bytes(src_device),
 702                               &dev_replace->scrub_progress, 0, 1);
 703
 704         ret = btrfs_dev_replace_finishing(fs_info, ret);
 705         if (ret == -EINPROGRESS)
 706                 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS;
 707
 708         return ret;
 709
 710 leave:
 711         btrfs_destroy_dev_replace_tgtdev(tgt_device);
 712         return ret;
 713 }
 714
 715 static int btrfs_check_replace_dev_names(struct btrfs_ioctl_dev_replace_args *args)
 716 {
 717         if (args->start.srcdevid == 0) {
 718                 if (memchr(args->start.srcdev_name, 0,
 719                            sizeof(args->start.srcdev_name)) == NULL)
 720                         return -ENAMETOOLONG;
 721         } else {
 722                 args->start.srcdev_name[0] = 0;
 723         }
 724
 725         if (memchr(args->start.tgtdev_name, 0,
 726                    sizeof(args->start.tgtdev_name)) == NULL)
 727             return -ENAMETOOLONG;
 728
 729         return 0;
 730 }
 731
 732 int btrfs_dev_replace_by_ioctl(struct btrfs_fs_info *fs_info,
 733                             struct btrfs_ioctl_dev_replace_args *args)
 734 {
 735         int ret;
 736
 737         switch (args->start.cont_reading_from_srcdev_mode) {
 738         case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_ALWAYS:
 739         case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_AVOID:
 740                 break;
 741         default:
 742                 return -EINVAL;
 743         }
 744         ret = btrfs_check_replace_dev_names(args);
 745         if (ret < 0)
 746                 return ret;
 747
 748         ret = btrfs_dev_replace_start(fs_info, args->start.tgtdev_name,
 749                                         args->start.srcdevid,
 750                                         args->start.srcdev_name,
 751                                         args->start.cont_reading_from_srcdev_mode);
 752         args->result = ret;
 753         /* don't warn if EINPROGRESS, someone else might be running scrub */
 754         if (ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS ||
 755             ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR)
 756                 return 0;
 757
 758         return ret;
 759 }
 760
 761 /*
 762  * blocked until all in-flight bios operations are finished.
 763  */
 764 static void btrfs_rm_dev_replace_blocked(struct btrfs_fs_info *fs_info)
 765 {
 766         set_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
 767         wait_event(fs_info->dev_replace.replace_wait, !percpu_counter_sum(
 768                    &fs_info->dev_replace.bio_counter));
 769 }
 770
 771 /*
 772  * we have removed target device, it is safe to allow new bios request.
 773  */
 774 static void btrfs_rm_dev_replace_unblocked(struct btrfs_fs_info *fs_info)
 775 {
 776         clear_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
 777         wake_up(&fs_info->dev_replace.replace_wait);
 778 }
 779
 780 /*
 781  * When finishing the device replace, before swapping the source device with the
 782  * target device we must update the chunk allocation state in the target device,
 783  * as it is empty because replace works by directly copying the chunks and not
 784  * through the normal chunk allocation path.
 785  */
 786 static int btrfs_set_target_alloc_state(struct btrfs_device *srcdev,
 787                                         struct btrfs_device *tgtdev)
 788 {
 789         struct extent_state *cached_state = NULL;
 790         u64 start = 0;
 791         u64 found_start;
 792         u64 found_end;
 793         int ret = 0;
 794
 795         lockdep_assert_held(&srcdev->fs_info->chunk_mutex);
 796
 797         while (btrfs_find_first_extent_bit(&srcdev->alloc_state, start,
 798                                            &found_start, &found_end,
 799                                            CHUNK_ALLOCATED, &cached_state)) {
 800                 ret = btrfs_set_extent_bit(&tgtdev->alloc_state, found_start,
 801                                            found_end, CHUNK_ALLOCATED, NULL);
 802                 if (ret)
 803                         break;
 804                 start = found_end + 1;
 805         }
 806
 807         btrfs_free_extent_state(cached_state);
 808         return ret;
 809 }
 810
 811 static void btrfs_dev_replace_update_device_in_mapping_tree(
 812                                                 struct btrfs_fs_info *fs_info,
 813                                                 struct btrfs_device *srcdev,
 814                                                 struct btrfs_device *tgtdev)
 815 {
 816         struct rb_node *node;
 817
 818         /*
 819          * The chunk mutex must be held so that no new chunks can be created
 820          * while we are updating existing chunks. This guarantees we don't miss
 821          * any new chunk that gets created for a range that falls before the
 822          * range of the last chunk we processed.
 823          */
 824         lockdep_assert_held(&fs_info->chunk_mutex);
 825
 826         write_lock(&fs_info->mapping_tree_lock);
 827         node = rb_first_cached(&fs_info->mapping_tree);
 828         while (node) {
 829                 struct rb_node *next = rb_next(node);
 830                 struct btrfs_chunk_map *map;
 831                 u64 next_start;
 832
 833                 map = rb_entry(node, struct btrfs_chunk_map, rb_node);
 834                 next_start = map->start + map->chunk_len;
 835
 836                 for (int i = 0; i < map->num_stripes; i++)
 837                         if (srcdev == map->stripes[i].dev)
 838                                 map->stripes[i].dev = tgtdev;
 839
 840                 if (cond_resched_rwlock_write(&fs_info->mapping_tree_lock)) {
 841                         map = btrfs_find_chunk_map_nolock(fs_info, next_start, U64_MAX);
 842                         if (!map)
 843                                 break;
 844                         node = &map->rb_node;
 845                         /*
 846                          * Drop the lookup reference since we are holding the
 847                          * lock in write mode and no one can remove the chunk
 848                          * map from the tree and drop its tree reference.
 849                          */
 850                         btrfs_free_chunk_map(map);
 851                 } else {
 852                         node = next;
 853                 }
 854         }
 855         write_unlock(&fs_info->mapping_tree_lock);
 856 }
 857
 858 static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
 859                                        int scrub_ret)
 860 {
 861         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 862         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 863         struct btrfs_device *tgt_device;
 864         struct btrfs_device *src_device;
 865         struct btrfs_root *root = fs_info->tree_root;
 866         u8 uuid_tmp[BTRFS_UUID_SIZE];
 867         struct btrfs_trans_handle *trans;
 868         int ret = 0;
 869
 870         /* don't allow cancel or unmount to disturb the finishing procedure */
 871         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
 872
 873         down_read(&dev_replace->rwsem);
 874         /* was the operation canceled, or is it finished? */
 875         if (dev_replace->replace_state !=
 876             BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED) {
 877                 up_read(&dev_replace->rwsem);
 878                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 879                 return 0;
 880         }
 881
 882         tgt_device = dev_replace->tgtdev;
 883         src_device = dev_replace->srcdev;
 884         up_read(&dev_replace->rwsem);
 885
 886         /*
 887          * flush all outstanding I/O and inode extent mappings before the
 888          * copy operation is declared as being finished
 889          */
 890         ret = btrfs_start_delalloc_roots(fs_info, LONG_MAX, false);
 891         if (ret) {
 892                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 893                 return ret;
 894         }
 895         btrfs_wait_ordered_roots(fs_info, U64_MAX, NULL);
 896
 897         /*
 898          * We have to use this loop approach because at this point src_device
 899          * has to be available for transaction commit to complete, yet new
 900          * chunks shouldn't be allocated on the device.
 901          */
 902         while (1) {
 903                 trans = btrfs_start_transaction(root, 0);
 904                 if (IS_ERR(trans)) {
 905                         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 906                         return PTR_ERR(trans);
 907                 }
 908                 ret = btrfs_commit_transaction(trans);
 909                 WARN_ON(ret);
 910
 911                 /* Prevent write_all_supers() during the finishing procedure */
 912                 mutex_lock(&fs_devices->device_list_mutex);
 913                 /* Prevent new chunks being allocated on the source device */
 914                 mutex_lock(&fs_info->chunk_mutex);
 915
 916                 if (!list_empty(&src_device->post_commit_list)) {
 917                         mutex_unlock(&fs_devices->device_list_mutex);
 918                         mutex_unlock(&fs_info->chunk_mutex);
 919                 } else {
 920                         break;
 921                 }
 922         }
 923
 924         down_write(&dev_replace->rwsem);
 925         dev_replace->replace_state =
 926                 scrub_ret ? BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED
 927                           : BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED;
 928         dev_replace->tgtdev = NULL;
 929         dev_replace->srcdev = NULL;
 930         dev_replace->time_stopped = ktime_get_real_seconds();
 931         dev_replace->item_needs_writeback = 1;
 932
 933         /*
 934          * Update allocation state in the new device and replace the old device
 935          * with the new one in the mapping tree.
 936          */
 937         if (!scrub_ret) {
 938                 scrub_ret = btrfs_set_target_alloc_state(src_device, tgt_device);
 939                 if (scrub_ret)
 940                         goto error;
 941                 btrfs_dev_replace_update_device_in_mapping_tree(fs_info,
 942                                                                 src_device,
 943                                                                 tgt_device);
 944         } else {
 945                 if (scrub_ret != -ECANCELED)
 946                         btrfs_err_in_rcu(fs_info,
 947                                  "btrfs_scrub_dev(%s, %llu, %s) failed %d",
 948                                  btrfs_dev_name(src_device),
 949                                  src_device->devid,
 950                                  btrfs_dev_name(tgt_device), scrub_ret);
 951 error:
 952                 up_write(&dev_replace->rwsem);
 953                 mutex_unlock(&fs_info->chunk_mutex);
 954                 mutex_unlock(&fs_devices->device_list_mutex);
 955                 btrfs_rm_dev_replace_blocked(fs_info);
 956                 if (tgt_device)
 957                         btrfs_destroy_dev_replace_tgtdev(tgt_device);
 958                 btrfs_rm_dev_replace_unblocked(fs_info);
 959                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 960
 961                 return scrub_ret;
 962         }
 963
 964         btrfs_info_in_rcu(fs_info,
 965                           "dev_replace from %s (devid %llu) to %s finished",
 966                           btrfs_dev_name(src_device),
 967                           src_device->devid,
 968                           btrfs_dev_name(tgt_device));
 969         clear_bit(BTRFS_DEV_STATE_REPLACE_TGT, &tgt_device->dev_state);
 970         tgt_device->devid = src_device->devid;
 971         src_device->devid = BTRFS_DEV_REPLACE_DEVID;
 972         memcpy(uuid_tmp, tgt_device->uuid, sizeof(uuid_tmp));
 973         memcpy(tgt_device->uuid, src_device->uuid, sizeof(tgt_device->uuid));
 974         memcpy(src_device->uuid, uuid_tmp, sizeof(src_device->uuid));
 975         btrfs_device_set_total_bytes(tgt_device, src_device->total_bytes);
 976         btrfs_device_set_disk_total_bytes(tgt_device,
 977                                           src_device->disk_total_bytes);
 978         btrfs_device_set_bytes_used(tgt_device, src_device->bytes_used);
 979         tgt_device->commit_bytes_used = src_device->bytes_used;
 980
 981         btrfs_assign_next_active_device(src_device, tgt_device);
 982
 983         list_add(&tgt_device->dev_alloc_list, &fs_devices->alloc_list);
 984         fs_devices->rw_devices++;
 985
 986         dev_replace->replace_task = NULL;
 987         up_write(&dev_replace->rwsem);
 988         btrfs_rm_dev_replace_blocked(fs_info);
 989
 990         btrfs_rm_dev_replace_remove_srcdev(src_device);
 991
 992         btrfs_rm_dev_replace_unblocked(fs_info);
 993
 994         /*
 995          * Increment dev_stats_ccnt so that btrfs_run_dev_stats() will
 996          * update on-disk dev stats value during commit transaction
 997          */
 998         atomic_inc(&tgt_device->dev_stats_ccnt);
 999
1000         /*
1001          * this is again a consistent state where no dev_replace procedure
1002          * is running, the target device is part of the filesystem, the
1003          * source device is not part of the filesystem anymore and its 1st
1004          * superblock is scratched out so that it is no longer marked to
1005          * belong to this filesystem.
1006          */
1007         mutex_unlock(&fs_info->chunk_mutex);
1008         mutex_unlock(&fs_devices->device_list_mutex);
1009
1010         /* replace the sysfs entry */
1011         btrfs_sysfs_remove_device(src_device);
1012         btrfs_sysfs_update_devid(tgt_device);
1013         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &src_device->dev_state))
1014                 btrfs_scratch_superblocks(fs_info, src_device);
1015
1016         /* write back the superblocks */
1017         trans = btrfs_start_transaction(root, 0);
1018         if (!IS_ERR(trans))
1019                 btrfs_commit_transaction(trans);
1020
1021         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1022
1023         btrfs_rm_dev_replace_free_srcdev(src_device);
1024
1025         return 0;
1026 }
1027
1028 /*
1029  * Read progress of device replace status according to the state and last
1030  * stored position. The value format is the same as for
1031  * btrfs_dev_replace::progress_1000
1032  */
1033 static u64 btrfs_dev_replace_progress(struct btrfs_fs_info *fs_info)
1034 {
1035         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1036         u64 ret = 0;
1037
1038         switch (dev_replace->replace_state) {
1039         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1040         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1041                 ret = 0;
1042                 break;
1043         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1044                 ret = 1000;
1045                 break;
1046         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1047         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1048                 ret = div64_u64(dev_replace->cursor_left,
1049                                 div_u64(btrfs_device_get_total_bytes(
1050                                                 dev_replace->srcdev), 1000));
1051                 break;
1052         }
1053
1054         return ret;
1055 }
1056
1057 void btrfs_dev_replace_status(struct btrfs_fs_info *fs_info,
1058                               struct btrfs_ioctl_dev_replace_args *args)
1059 {
1060         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1061
1062         down_read(&dev_replace->rwsem);
1063         /* even if !dev_replace_is_valid, the values are good enough for
1064          * the replace_status ioctl */
1065         args->result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1066         args->status.replace_state = dev_replace->replace_state;
1067         args->status.time_started = dev_replace->time_started;
1068         args->status.time_stopped = dev_replace->time_stopped;
1069         args->status.num_write_errors =
1070                 atomic64_read(&dev_replace->num_write_errors);
1071         args->status.num_uncorrectable_read_errors =
1072                 atomic64_read(&dev_replace->num_uncorrectable_read_errors);
1073         args->status.progress_1000 = btrfs_dev_replace_progress(fs_info);
1074         up_read(&dev_replace->rwsem);
1075 }
1076
1077 int btrfs_dev_replace_cancel(struct btrfs_fs_info *fs_info)
1078 {
1079         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1080         struct btrfs_device *tgt_device = NULL;
1081         struct btrfs_device *src_device = NULL;
1082         struct btrfs_trans_handle *trans;
1083         struct btrfs_root *root = fs_info->tree_root;
1084         int result;
1085         int ret;
1086
1087         if (sb_rdonly(fs_info->sb))
1088                 return -EROFS;
1089
1090         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1091         down_write(&dev_replace->rwsem);
1092         switch (dev_replace->replace_state) {
1093         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1094         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1095         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1096                 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1097                 up_write(&dev_replace->rwsem);
1098                 break;
1099         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1100                 tgt_device = dev_replace->tgtdev;
1101                 src_device = dev_replace->srcdev;
1102                 up_write(&dev_replace->rwsem);
1103                 ret = btrfs_scrub_cancel(fs_info);
1104                 if (ret < 0) {
1105                         result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1106                 } else {
1107                         result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1108                         /*
1109                          * btrfs_dev_replace_finishing() will handle the
1110                          * cleanup part
1111                          */
1112                         btrfs_info_in_rcu(fs_info,
1113                                 "dev_replace from %s (devid %llu) to %s canceled",
1114                                 btrfs_dev_name(src_device), src_device->devid,
1115                                 btrfs_dev_name(tgt_device));
1116                 }
1117                 break;
1118         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1119                 /*
1120                  * Scrub doing the replace isn't running so we need to do the
1121                  * cleanup step of btrfs_dev_replace_finishing() here
1122                  */
1123                 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1124                 tgt_device = dev_replace->tgtdev;
1125                 src_device = dev_replace->srcdev;
1126                 dev_replace->tgtdev = NULL;
1127                 dev_replace->srcdev = NULL;
1128                 dev_replace->replace_state =
1129                                 BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED;
1130                 dev_replace->time_stopped = ktime_get_real_seconds();
1131                 dev_replace->item_needs_writeback = 1;
1132
1133                 up_write(&dev_replace->rwsem);
1134
1135                 /* Scrub for replace must not be running in suspended state */
1136                 btrfs_scrub_cancel(fs_info);
1137
1138                 trans = btrfs_start_transaction(root, 0);
1139                 if (IS_ERR(trans)) {
1140                         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1141                         return PTR_ERR(trans);
1142                 }
1143                 ret = btrfs_commit_transaction(trans);
1144                 WARN_ON(ret);
1145
1146                 btrfs_info_in_rcu(fs_info,
1147                 "suspended dev_replace from %s (devid %llu) to %s canceled",
1148                         btrfs_dev_name(src_device), src_device->devid,
1149                         btrfs_dev_name(tgt_device));
1150
1151                 if (tgt_device)
1152                         btrfs_destroy_dev_replace_tgtdev(tgt_device);
1153                 break;
1154         default:
1155                 up_write(&dev_replace->rwsem);
1156                 result = -EINVAL;
1157         }
1158
1159         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1160         return result;
1161 }
1162
1163 void btrfs_dev_replace_suspend_for_unmount(struct btrfs_fs_info *fs_info)
1164 {
1165         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1166
1167         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1168         down_write(&dev_replace->rwsem);
1169
1170         switch (dev_replace->replace_state) {
1171         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1172         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1173         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1174         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1175                 break;
1176         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1177                 dev_replace->replace_state =
1178                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1179                 dev_replace->time_stopped = ktime_get_real_seconds();
1180                 dev_replace->item_needs_writeback = 1;
1181                 btrfs_info(fs_info, "suspending dev_replace for unmount");
1182                 break;
1183         }
1184
1185         up_write(&dev_replace->rwsem);
1186         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1187 }
1188
1189 /* resume dev_replace procedure that was interrupted by unmount */
1190 int btrfs_resume_dev_replace_async(struct btrfs_fs_info *fs_info)
1191 {
1192         struct task_struct *task;
1193         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1194
1195         down_write(&dev_replace->rwsem);
1196
1197         switch (dev_replace->replace_state) {
1198         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1199         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1200         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1201                 up_write(&dev_replace->rwsem);
1202                 return 0;
1203         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1204                 break;
1205         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1206                 dev_replace->replace_state =
1207                         BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
1208                 break;
1209         }
1210         if (!dev_replace->tgtdev || !dev_replace->tgtdev->bdev) {
1211                 btrfs_info(fs_info,
1212                            "cannot continue dev_replace, tgtdev is missing");
1213                 btrfs_info(fs_info,
1214                            "you may cancel the operation after 'mount -o degraded'");
1215                 dev_replace->replace_state =
1216                                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1217                 up_write(&dev_replace->rwsem);
1218                 return 0;
1219         }
1220         up_write(&dev_replace->rwsem);
1221
1222         /*
1223          * This could collide with a paused balance, but the exclusive op logic
1224          * should never allow both to start and pause. We don't want to allow
1225          * dev-replace to start anyway.
1226          */
1227         if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_DEV_REPLACE)) {
1228                 down_write(&dev_replace->rwsem);
1229                 dev_replace->replace_state =
1230                                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1231                 up_write(&dev_replace->rwsem);
1232                 btrfs_info(fs_info,
1233                 "cannot resume dev-replace, other exclusive operation running");
1234                 return 0;
1235         }
1236
1237         task = kthread_run(btrfs_dev_replace_kthread, fs_info, "btrfs-devrepl");
1238         return PTR_ERR_OR_ZERO(task);
1239 }
1240
1241 static int btrfs_dev_replace_kthread(void *data)
1242 {
1243         struct btrfs_fs_info *fs_info = data;
1244         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1245         u64 progress;
1246         int ret;
1247
1248         progress = btrfs_dev_replace_progress(fs_info);
1249         progress = div_u64(progress, 10);
1250         btrfs_info_in_rcu(fs_info,
1251                 "continuing dev_replace from %s (devid %llu) to target %s @%u%%",
1252                 btrfs_dev_name(dev_replace->srcdev),
1253                 dev_replace->srcdev->devid,
1254                 btrfs_dev_name(dev_replace->tgtdev),
1255                 (unsigned int)progress);
1256
1257         ret = btrfs_scrub_dev(fs_info, dev_replace->srcdev->devid,
1258                               dev_replace->committed_cursor_left,
1259                               btrfs_device_get_total_bytes(dev_replace->srcdev),
1260                               &dev_replace->scrub_progress, 0, 1);
1261         ret = btrfs_dev_replace_finishing(fs_info, ret);
1262         WARN_ON(ret && ret != -ECANCELED);
1263
1264         btrfs_exclop_finish(fs_info);
1265         return 0;
1266 }
1267
1268 bool __pure btrfs_dev_replace_is_ongoing(struct btrfs_dev_replace *dev_replace)
1269 {
1270         if (!dev_replace->is_valid)
1271                 return false;
1272
1273         switch (dev_replace->replace_state) {
1274         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1275         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1276         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1277                 return false;
1278         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1279         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1280                 /*
1281                  * return true even if tgtdev is missing (this is
1282                  * something that can happen if the dev_replace
1283                  * procedure is suspended by an umount and then
1284                  * the tgtdev is missing (or "btrfs dev scan") was
1285                  * not called and the filesystem is remounted
1286                  * in degraded state. This does not stop the
1287                  * dev_replace procedure. It needs to be canceled
1288                  * manually if the cancellation is wanted.
1289                  */
1290                 break;
1291         }
1292         return true;
1293 }
1294
1295 void btrfs_bio_counter_sub(struct btrfs_fs_info *fs_info, s64 amount)
1296 {
1297         percpu_counter_sub(&fs_info->dev_replace.bio_counter, amount);
1298         cond_wake_up_nomb(&fs_info->dev_replace.replace_wait);
1299 }
1300
1301 void btrfs_bio_counter_inc_blocked(struct btrfs_fs_info *fs_info)
1302 {
1303         while (1) {
1304                 percpu_counter_inc(&fs_info->dev_replace.bio_counter);
1305                 if (likely(!test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1306                                      &fs_info->fs_state)))
1307                         break;
1308
1309                 btrfs_bio_counter_dec(fs_info);
1310                 wait_event(fs_info->dev_replace.replace_wait,
1311                            !test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1312                                      &fs_info->fs_state));
1313         }
1314 }