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         struct btrfs_path *path = NULL;
  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                 ret = -ENOMEM;
  90                 goto out;
  91         }
  92
  93         key.objectid = 0;
  94         key.type = BTRFS_DEV_REPLACE_KEY;
  95         key.offset = 0;
  96         ret = btrfs_search_slot(NULL, dev_root, &key, path, 0, 0);
  97         if (ret) {
  98 no_valid_dev_replace_entry_found:
  99                 /*
 100                  * We don't have a replace item or it's corrupted.  If there is
 101                  * a replace target, fail the mount.
 102                  */
 103                 if (btrfs_find_device(fs_info->fs_devices, &args)) {
 104                         btrfs_err(fs_info,
 105                         "found replace target device without a valid replace item");
 106                         ret = -EUCLEAN;
 107                         goto out;
 108                 }
 109                 ret = 0;
 110                 dev_replace->replace_state =
 111                         BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED;
 112                 dev_replace->cont_reading_from_srcdev_mode =
 113                     BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_ALWAYS;
 114                 dev_replace->time_started = 0;
 115                 dev_replace->time_stopped = 0;
 116                 atomic64_set(&dev_replace->num_write_errors, 0);
 117                 atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
 118                 dev_replace->cursor_left = 0;
 119                 dev_replace->committed_cursor_left = 0;
 120                 dev_replace->cursor_left_last_write_of_item = 0;
 121                 dev_replace->cursor_right = 0;
 122                 dev_replace->srcdev = NULL;
 123                 dev_replace->tgtdev = NULL;
 124                 dev_replace->is_valid = 0;
 125                 dev_replace->item_needs_writeback = 0;
 126                 goto out;
 127         }
 128         slot = path->slots[0];
 129         eb = path->nodes[0];
 130         item_size = btrfs_item_size(eb, slot);
 131         ptr = btrfs_item_ptr(eb, slot, struct btrfs_dev_replace_item);
 132
 133         if (item_size != sizeof(struct btrfs_dev_replace_item)) {
 134                 btrfs_warn(fs_info,
 135                         "dev_replace entry found has unexpected size, ignore entry");
 136                 goto no_valid_dev_replace_entry_found;
 137         }
 138
 139         src_devid = btrfs_dev_replace_src_devid(eb, ptr);
 140         dev_replace->cont_reading_from_srcdev_mode =
 141                 btrfs_dev_replace_cont_reading_from_srcdev_mode(eb, ptr);
 142         dev_replace->replace_state = btrfs_dev_replace_replace_state(eb, ptr);
 143         dev_replace->time_started = btrfs_dev_replace_time_started(eb, ptr);
 144         dev_replace->time_stopped =
 145                 btrfs_dev_replace_time_stopped(eb, ptr);
 146         atomic64_set(&dev_replace->num_write_errors,
 147                      btrfs_dev_replace_num_write_errors(eb, ptr));
 148         atomic64_set(&dev_replace->num_uncorrectable_read_errors,
 149                      btrfs_dev_replace_num_uncorrectable_read_errors(eb, ptr));
 150         dev_replace->cursor_left = btrfs_dev_replace_cursor_left(eb, ptr);
 151         dev_replace->committed_cursor_left = dev_replace->cursor_left;
 152         dev_replace->cursor_left_last_write_of_item = dev_replace->cursor_left;
 153         dev_replace->cursor_right = btrfs_dev_replace_cursor_right(eb, ptr);
 154         dev_replace->is_valid = 1;
 155
 156         dev_replace->item_needs_writeback = 0;
 157         switch (dev_replace->replace_state) {
 158         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
 159         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
 160         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
 161                 /*
 162                  * We don't have an active replace item but if there is a
 163                  * replace target, fail the mount.
 164                  */
 165                 if (btrfs_find_device(fs_info->fs_devices, &args)) {
 166                         btrfs_err(fs_info,
 167 "replace without active item, run 'device scan --forget' on the target device");
 168                         ret = -EUCLEAN;
 169                 } else {
 170                         dev_replace->srcdev = NULL;
 171                         dev_replace->tgtdev = NULL;
 172                 }
 173                 break;
 174         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
 175         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
 176                 dev_replace->tgtdev = btrfs_find_device(fs_info->fs_devices, &args);
 177                 args.devid = src_devid;
 178                 dev_replace->srcdev = btrfs_find_device(fs_info->fs_devices, &args);
 179
 180                 /*
 181                  * allow 'btrfs dev replace_cancel' if src/tgt device is
 182                  * missing
 183                  */
 184                 if (!dev_replace->srcdev &&
 185                     !btrfs_test_opt(fs_info, DEGRADED)) {
 186                         ret = -EIO;
 187                         btrfs_warn(fs_info,
 188                            "cannot mount because device replace operation is ongoing and");
 189                         btrfs_warn(fs_info,
 190                            "srcdev (devid %llu) is missing, need to run 'btrfs dev scan'?",
 191                            src_devid);
 192                 }
 193                 if (!dev_replace->tgtdev &&
 194                     !btrfs_test_opt(fs_info, DEGRADED)) {
 195                         ret = -EIO;
 196                         btrfs_warn(fs_info,
 197                            "cannot mount because device replace operation is ongoing and");
 198                         btrfs_warn(fs_info,
 199                            "tgtdev (devid %llu) is missing, need to run 'btrfs dev scan'?",
 200                                 BTRFS_DEV_REPLACE_DEVID);
 201                 }
 202                 if (dev_replace->tgtdev) {
 203                         if (dev_replace->srcdev) {
 204                                 dev_replace->tgtdev->total_bytes =
 205                                         dev_replace->srcdev->total_bytes;
 206                                 dev_replace->tgtdev->disk_total_bytes =
 207                                         dev_replace->srcdev->disk_total_bytes;
 208                                 dev_replace->tgtdev->commit_total_bytes =
 209                                         dev_replace->srcdev->commit_total_bytes;
 210                                 dev_replace->tgtdev->bytes_used =
 211                                         dev_replace->srcdev->bytes_used;
 212                                 dev_replace->tgtdev->commit_bytes_used =
 213                                         dev_replace->srcdev->commit_bytes_used;
 214                         }
 215                         set_bit(BTRFS_DEV_STATE_REPLACE_TGT,
 216                                 &dev_replace->tgtdev->dev_state);
 217
 218                         WARN_ON(fs_info->fs_devices->rw_devices == 0);
 219                         dev_replace->tgtdev->io_width = fs_info->sectorsize;
 220                         dev_replace->tgtdev->io_align = fs_info->sectorsize;
 221                         dev_replace->tgtdev->sector_size = fs_info->sectorsize;
 222                         dev_replace->tgtdev->fs_info = fs_info;
 223                         set_bit(BTRFS_DEV_STATE_IN_FS_METADATA,
 224                                 &dev_replace->tgtdev->dev_state);
 225                 }
 226                 break;
 227         }
 228
 229 out:
 230         btrfs_free_path(path);
 231         return ret;
 232 }
 233
 234 /*
 235  * Initialize a new device for device replace target from a given source dev
 236  * and path.
 237  *
 238  * Return 0 and new device in @device_out, otherwise return < 0
 239  */
 240 static int btrfs_init_dev_replace_tgtdev(struct btrfs_fs_info *fs_info,
 241                                   const char *device_path,
 242                                   struct btrfs_device *srcdev,
 243                                   struct btrfs_device **device_out)
 244 {
 245         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 246         struct btrfs_device *device;
 247         struct file *bdev_file;
 248         struct block_device *bdev;
 249         u64 devid = BTRFS_DEV_REPLACE_DEVID;
 250         int ret = 0;
 251
 252         *device_out = NULL;
 253         if (srcdev->fs_devices->seeding) {
 254                 btrfs_err(fs_info, "the filesystem is a seed filesystem!");
 255                 return -EINVAL;
 256         }
 257
 258         bdev_file = bdev_file_open_by_path(device_path, BLK_OPEN_WRITE,
 259                                         fs_info->bdev_holder, NULL);
 260         if (IS_ERR(bdev_file)) {
 261                 btrfs_err(fs_info, "target device %s is invalid!", device_path);
 262                 return PTR_ERR(bdev_file);
 263         }
 264         bdev = file_bdev(bdev_file);
 265
 266         if (!btrfs_check_device_zone_type(fs_info, bdev)) {
 267                 btrfs_err(fs_info,
 268                 "dev-replace: zoned type of target device mismatch with filesystem");
 269                 ret = -EINVAL;
 270                 goto error;
 271         }
 272
 273         sync_blockdev(bdev);
 274
 275         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 276                 if (device->bdev == bdev) {
 277                         btrfs_err(fs_info,
 278                                   "target device is in the filesystem!");
 279                         ret = -EEXIST;
 280                         goto error;
 281                 }
 282         }
 283
 284
 285         if (bdev_nr_bytes(bdev) < btrfs_device_get_total_bytes(srcdev)) {
 286                 btrfs_err(fs_info,
 287                           "target device is smaller than source device!");
 288                 ret = -EINVAL;
 289                 goto error;
 290         }
 291
 292
 293         device = btrfs_alloc_device(NULL, &devid, NULL, device_path);
 294         if (IS_ERR(device)) {
 295                 ret = PTR_ERR(device);
 296                 goto error;
 297         }
 298
 299         ret = lookup_bdev(device_path, &device->devt);
 300         if (ret)
 301                 goto error;
 302
 303         set_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
 304         device->generation = 0;
 305         device->io_width = fs_info->sectorsize;
 306         device->io_align = fs_info->sectorsize;
 307         device->sector_size = fs_info->sectorsize;
 308         device->total_bytes = btrfs_device_get_total_bytes(srcdev);
 309         device->disk_total_bytes = btrfs_device_get_disk_total_bytes(srcdev);
 310         device->bytes_used = btrfs_device_get_bytes_used(srcdev);
 311         device->commit_total_bytes = srcdev->commit_total_bytes;
 312         device->commit_bytes_used = device->bytes_used;
 313         device->fs_info = fs_info;
 314         device->bdev = bdev;
 315         device->bdev_file = bdev_file;
 316         set_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
 317         set_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state);
 318         device->dev_stats_valid = 1;
 319         set_blocksize(bdev_file, BTRFS_BDEV_BLOCKSIZE);
 320         device->fs_devices = fs_devices;
 321
 322         ret = btrfs_get_dev_zone_info(device, false);
 323         if (ret)
 324                 goto error;
 325
 326         mutex_lock(&fs_devices->device_list_mutex);
 327         list_add(&device->dev_list, &fs_devices->devices);
 328         fs_devices->num_devices++;
 329         fs_devices->open_devices++;
 330         mutex_unlock(&fs_devices->device_list_mutex);
 331
 332         *device_out = device;
 333         return 0;
 334
 335 error:
 336         fput(bdev_file);
 337         return ret;
 338 }
 339
 340 /*
 341  * called from commit_transaction. Writes changed device replace state to
 342  * disk.
 343  */
 344 int btrfs_run_dev_replace(struct btrfs_trans_handle *trans)
 345 {
 346         struct btrfs_fs_info *fs_info = trans->fs_info;
 347         int ret;
 348         struct btrfs_root *dev_root = fs_info->dev_root;
 349         struct btrfs_path *path;
 350         struct btrfs_key key;
 351         struct extent_buffer *eb;
 352         struct btrfs_dev_replace_item *ptr;
 353         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 354
 355         down_read(&dev_replace->rwsem);
 356         if (!dev_replace->is_valid ||
 357             !dev_replace->item_needs_writeback) {
 358                 up_read(&dev_replace->rwsem);
 359                 return 0;
 360         }
 361         up_read(&dev_replace->rwsem);
 362
 363         key.objectid = 0;
 364         key.type = BTRFS_DEV_REPLACE_KEY;
 365         key.offset = 0;
 366
 367         path = btrfs_alloc_path();
 368         if (!path) {
 369                 ret = -ENOMEM;
 370                 goto out;
 371         }
 372         ret = btrfs_search_slot(trans, dev_root, &key, path, -1, 1);
 373         if (ret < 0) {
 374                 btrfs_warn(fs_info,
 375                            "error %d while searching for dev_replace item!",
 376                            ret);
 377                 goto out;
 378         }
 379
 380         if (ret == 0 &&
 381             btrfs_item_size(path->nodes[0], path->slots[0]) < sizeof(*ptr)) {
 382                 /*
 383                  * need to delete old one and insert a new one.
 384                  * Since no attempt is made to recover any old state, if the
 385                  * dev_replace state is 'running', the data on the target
 386                  * drive is lost.
 387                  * It would be possible to recover the state: just make sure
 388                  * that the beginning of the item is never changed and always
 389                  * contains all the essential information. Then read this
 390                  * minimal set of information and use it as a base for the
 391                  * new state.
 392                  */
 393                 ret = btrfs_del_item(trans, dev_root, path);
 394                 if (ret != 0) {
 395                         btrfs_warn(fs_info,
 396                                    "delete too small dev_replace item failed %d!",
 397                                    ret);
 398                         goto out;
 399                 }
 400                 ret = 1;
 401         }
 402
 403         if (ret == 1) {
 404                 /* need to insert a new item */
 405                 btrfs_release_path(path);
 406                 ret = btrfs_insert_empty_item(trans, dev_root, path,
 407                                               &key, sizeof(*ptr));
 408                 if (ret < 0) {
 409                         btrfs_warn(fs_info,
 410                                    "insert dev_replace item failed %d!", ret);
 411                         goto out;
 412                 }
 413         }
 414
 415         eb = path->nodes[0];
 416         ptr = btrfs_item_ptr(eb, path->slots[0],
 417                              struct btrfs_dev_replace_item);
 418
 419         down_write(&dev_replace->rwsem);
 420         if (dev_replace->srcdev)
 421                 btrfs_set_dev_replace_src_devid(eb, ptr,
 422                         dev_replace->srcdev->devid);
 423         else
 424                 btrfs_set_dev_replace_src_devid(eb, ptr, (u64)-1);
 425         btrfs_set_dev_replace_cont_reading_from_srcdev_mode(eb, ptr,
 426                 dev_replace->cont_reading_from_srcdev_mode);
 427         btrfs_set_dev_replace_replace_state(eb, ptr,
 428                 dev_replace->replace_state);
 429         btrfs_set_dev_replace_time_started(eb, ptr, dev_replace->time_started);
 430         btrfs_set_dev_replace_time_stopped(eb, ptr, dev_replace->time_stopped);
 431         btrfs_set_dev_replace_num_write_errors(eb, ptr,
 432                 atomic64_read(&dev_replace->num_write_errors));
 433         btrfs_set_dev_replace_num_uncorrectable_read_errors(eb, ptr,
 434                 atomic64_read(&dev_replace->num_uncorrectable_read_errors));
 435         dev_replace->cursor_left_last_write_of_item =
 436                 dev_replace->cursor_left;
 437         btrfs_set_dev_replace_cursor_left(eb, ptr,
 438                 dev_replace->cursor_left_last_write_of_item);
 439         btrfs_set_dev_replace_cursor_right(eb, ptr,
 440                 dev_replace->cursor_right);
 441         dev_replace->item_needs_writeback = 0;
 442         up_write(&dev_replace->rwsem);
 443
 444         btrfs_mark_buffer_dirty(trans, eb);
 445
 446 out:
 447         btrfs_free_path(path);
 448
 449         return ret;
 450 }
 451
 452 static int mark_block_group_to_copy(struct btrfs_fs_info *fs_info,
 453                                     struct btrfs_device *src_dev)
 454 {
 455         struct btrfs_path *path;
 456         struct btrfs_key key;
 457         struct btrfs_key found_key;
 458         struct btrfs_root *root = fs_info->dev_root;
 459         struct btrfs_dev_extent *dev_extent = NULL;
 460         struct btrfs_block_group *cache;
 461         struct btrfs_trans_handle *trans;
 462         int iter_ret = 0;
 463         int ret = 0;
 464         u64 chunk_offset;
 465
 466         /* Do not use "to_copy" on non zoned filesystem for now */
 467         if (!btrfs_is_zoned(fs_info))
 468                 return 0;
 469
 470         mutex_lock(&fs_info->chunk_mutex);
 471
 472         /* Ensure we don't have pending new block group */
 473         spin_lock(&fs_info->trans_lock);
 474         while (fs_info->running_transaction &&
 475                !list_empty(&fs_info->running_transaction->dev_update_list)) {
 476                 spin_unlock(&fs_info->trans_lock);
 477                 mutex_unlock(&fs_info->chunk_mutex);
 478                 trans = btrfs_attach_transaction(root);
 479                 if (IS_ERR(trans)) {
 480                         ret = PTR_ERR(trans);
 481                         mutex_lock(&fs_info->chunk_mutex);
 482                         if (ret == -ENOENT) {
 483                                 spin_lock(&fs_info->trans_lock);
 484                                 continue;
 485                         } else {
 486                                 goto unlock;
 487                         }
 488                 }
 489
 490                 ret = btrfs_commit_transaction(trans);
 491                 mutex_lock(&fs_info->chunk_mutex);
 492                 if (ret)
 493                         goto unlock;
 494
 495                 spin_lock(&fs_info->trans_lock);
 496         }
 497         spin_unlock(&fs_info->trans_lock);
 498
 499         path = btrfs_alloc_path();
 500         if (!path) {
 501                 ret = -ENOMEM;
 502                 goto unlock;
 503         }
 504
 505         path->reada = READA_FORWARD;
 506         path->search_commit_root = 1;
 507         path->skip_locking = 1;
 508
 509         key.objectid = src_dev->devid;
 510         key.type = BTRFS_DEV_EXTENT_KEY;
 511         key.offset = 0;
 512
 513         btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
 514                 struct extent_buffer *leaf = path->nodes[0];
 515
 516                 if (found_key.objectid != src_dev->devid)
 517                         break;
 518
 519                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
 520                         break;
 521
 522                 if (found_key.offset < key.offset)
 523                         break;
 524
 525                 dev_extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
 526
 527                 chunk_offset = btrfs_dev_extent_chunk_offset(leaf, dev_extent);
 528
 529                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
 530                 if (!cache)
 531                         continue;
 532
 533                 set_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags);
 534                 btrfs_put_block_group(cache);
 535         }
 536         if (iter_ret < 0)
 537                 ret = iter_ret;
 538
 539         btrfs_free_path(path);
 540 unlock:
 541         mutex_unlock(&fs_info->chunk_mutex);
 542
 543         return ret;
 544 }
 545
 546 bool btrfs_finish_block_group_to_copy(struct btrfs_device *srcdev,
 547                                       struct btrfs_block_group *cache,
 548                                       u64 physical)
 549 {
 550         struct btrfs_fs_info *fs_info = cache->fs_info;
 551         struct btrfs_chunk_map *map;
 552         u64 chunk_offset = cache->start;
 553         int num_extents, cur_extent;
 554         int i;
 555
 556         /* Do not use "to_copy" on non zoned filesystem for now */
 557         if (!btrfs_is_zoned(fs_info))
 558                 return true;
 559
 560         spin_lock(&cache->lock);
 561         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
 562                 spin_unlock(&cache->lock);
 563                 return true;
 564         }
 565         spin_unlock(&cache->lock);
 566
 567         map = btrfs_get_chunk_map(fs_info, chunk_offset, 1);
 568         ASSERT(!IS_ERR(map));
 569
 570         num_extents = 0;
 571         cur_extent = 0;
 572         for (i = 0; i < map->num_stripes; i++) {
 573                 /* We have more device extent to copy */
 574                 if (srcdev != map->stripes[i].dev)
 575                         continue;
 576
 577                 num_extents++;
 578                 if (physical == map->stripes[i].physical)
 579                         cur_extent = i;
 580         }
 581
 582         btrfs_free_chunk_map(map);
 583
 584         if (num_extents > 1 && cur_extent < num_extents - 1) {
 585                 /*
 586                  * Has more stripes on this device. Keep this block group
 587                  * readonly until we finish all the stripes.
 588                  */
 589                 return false;
 590         }
 591
 592         /* Last stripe on this device */
 593         clear_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags);
 594
 595         return true;
 596 }
 597
 598 static int btrfs_dev_replace_start(struct btrfs_fs_info *fs_info,
 599                 const char *tgtdev_name, u64 srcdevid, const char *srcdev_name,
 600                 int read_src)
 601 {
 602         struct btrfs_root *root = fs_info->dev_root;
 603         struct btrfs_trans_handle *trans;
 604         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 605         int ret;
 606         struct btrfs_device *tgt_device = NULL;
 607         struct btrfs_device *src_device = NULL;
 608
 609         src_device = btrfs_find_device_by_devspec(fs_info, srcdevid,
 610                                                   srcdev_name);
 611         if (IS_ERR(src_device))
 612                 return PTR_ERR(src_device);
 613
 614         if (btrfs_pinned_by_swapfile(fs_info, src_device)) {
 615                 btrfs_warn_in_rcu(fs_info,
 616           "cannot replace device %s (devid %llu) due to active swapfile",
 617                         btrfs_dev_name(src_device), src_device->devid);
 618                 return -ETXTBSY;
 619         }
 620
 621         /*
 622          * Here we commit the transaction to make sure commit_total_bytes
 623          * of all the devices are updated.
 624          */
 625         trans = btrfs_attach_transaction(root);
 626         if (!IS_ERR(trans)) {
 627                 ret = btrfs_commit_transaction(trans);
 628                 if (ret)
 629                         return ret;
 630         } else if (PTR_ERR(trans) != -ENOENT) {
 631                 return PTR_ERR(trans);
 632         }
 633
 634         ret = btrfs_init_dev_replace_tgtdev(fs_info, tgtdev_name,
 635                                             src_device, &tgt_device);
 636         if (ret)
 637                 return ret;
 638
 639         ret = mark_block_group_to_copy(fs_info, src_device);
 640         if (ret)
 641                 return ret;
 642
 643         down_write(&dev_replace->rwsem);
 644         dev_replace->replace_task = current;
 645         switch (dev_replace->replace_state) {
 646         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
 647         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
 648         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
 649                 break;
 650         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
 651         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
 652                 ASSERT(0);
 653                 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_ALREADY_STARTED;
 654                 up_write(&dev_replace->rwsem);
 655                 goto leave;
 656         }
 657
 658         dev_replace->cont_reading_from_srcdev_mode = read_src;
 659         dev_replace->srcdev = src_device;
 660         dev_replace->tgtdev = tgt_device;
 661
 662         btrfs_info_in_rcu(fs_info,
 663                       "dev_replace from %s (devid %llu) to %s started",
 664                       btrfs_dev_name(src_device),
 665                       src_device->devid,
 666                       btrfs_dev_name(tgt_device));
 667
 668         /*
 669          * from now on, the writes to the srcdev are all duplicated to
 670          * go to the tgtdev as well (refer to btrfs_map_block()).
 671          */
 672         dev_replace->replace_state = BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
 673         dev_replace->time_started = ktime_get_real_seconds();
 674         dev_replace->cursor_left = 0;
 675         dev_replace->committed_cursor_left = 0;
 676         dev_replace->cursor_left_last_write_of_item = 0;
 677         dev_replace->cursor_right = 0;
 678         dev_replace->is_valid = 1;
 679         dev_replace->item_needs_writeback = 1;
 680         atomic64_set(&dev_replace->num_write_errors, 0);
 681         atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
 682         up_write(&dev_replace->rwsem);
 683
 684         ret = btrfs_sysfs_add_device(tgt_device);
 685         if (ret)
 686                 btrfs_err(fs_info, "kobj add dev failed %d", ret);
 687
 688         btrfs_wait_ordered_roots(fs_info, U64_MAX, NULL);
 689
 690         /*
 691          * Commit dev_replace state and reserve 1 item for it.
 692          * This is crucial to ensure we won't miss copying extents for new block
 693          * groups that are allocated after we started the device replace, and
 694          * must be done after setting up the device replace state.
 695          */
 696         trans = btrfs_start_transaction(root, 1);
 697         if (IS_ERR(trans)) {
 698                 ret = PTR_ERR(trans);
 699                 down_write(&dev_replace->rwsem);
 700                 dev_replace->replace_state =
 701                         BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED;
 702                 dev_replace->srcdev = NULL;
 703                 dev_replace->tgtdev = NULL;
 704                 up_write(&dev_replace->rwsem);
 705                 goto leave;
 706         }
 707
 708         ret = btrfs_commit_transaction(trans);
 709         WARN_ON(ret);
 710
 711         /* the disk copy procedure reuses the scrub code */
 712         ret = btrfs_scrub_dev(fs_info, src_device->devid, 0,
 713                               btrfs_device_get_total_bytes(src_device),
 714                               &dev_replace->scrub_progress, 0, 1);
 715
 716         ret = btrfs_dev_replace_finishing(fs_info, ret);
 717         if (ret == -EINPROGRESS)
 718                 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS;
 719
 720         return ret;
 721
 722 leave:
 723         btrfs_destroy_dev_replace_tgtdev(tgt_device);
 724         return ret;
 725 }
 726
 727 static int btrfs_check_replace_dev_names(struct btrfs_ioctl_dev_replace_args *args)
 728 {
 729         if (args->start.srcdevid == 0) {
 730                 if (memchr(args->start.srcdev_name, 0,
 731                            sizeof(args->start.srcdev_name)) == NULL)
 732                         return -ENAMETOOLONG;
 733         } else {
 734                 args->start.srcdev_name[0] = 0;
 735         }
 736
 737         if (memchr(args->start.tgtdev_name, 0,
 738                    sizeof(args->start.tgtdev_name)) == NULL)
 739             return -ENAMETOOLONG;
 740
 741         return 0;
 742 }
 743
 744 int btrfs_dev_replace_by_ioctl(struct btrfs_fs_info *fs_info,
 745                             struct btrfs_ioctl_dev_replace_args *args)
 746 {
 747         int ret;
 748
 749         switch (args->start.cont_reading_from_srcdev_mode) {
 750         case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_ALWAYS:
 751         case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_AVOID:
 752                 break;
 753         default:
 754                 return -EINVAL;
 755         }
 756         ret = btrfs_check_replace_dev_names(args);
 757         if (ret < 0)
 758                 return ret;
 759
 760         ret = btrfs_dev_replace_start(fs_info, args->start.tgtdev_name,
 761                                         args->start.srcdevid,
 762                                         args->start.srcdev_name,
 763                                         args->start.cont_reading_from_srcdev_mode);
 764         args->result = ret;
 765         /* don't warn if EINPROGRESS, someone else might be running scrub */
 766         if (ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS ||
 767             ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR)
 768                 return 0;
 769
 770         return ret;
 771 }
 772
 773 /*
 774  * blocked until all in-flight bios operations are finished.
 775  */
 776 static void btrfs_rm_dev_replace_blocked(struct btrfs_fs_info *fs_info)
 777 {
 778         set_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
 779         wait_event(fs_info->dev_replace.replace_wait, !percpu_counter_sum(
 780                    &fs_info->dev_replace.bio_counter));
 781 }
 782
 783 /*
 784  * we have removed target device, it is safe to allow new bios request.
 785  */
 786 static void btrfs_rm_dev_replace_unblocked(struct btrfs_fs_info *fs_info)
 787 {
 788         clear_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
 789         wake_up(&fs_info->dev_replace.replace_wait);
 790 }
 791
 792 /*
 793  * When finishing the device replace, before swapping the source device with the
 794  * target device we must update the chunk allocation state in the target device,
 795  * as it is empty because replace works by directly copying the chunks and not
 796  * through the normal chunk allocation path.
 797  */
 798 static int btrfs_set_target_alloc_state(struct btrfs_device *srcdev,
 799                                         struct btrfs_device *tgtdev)
 800 {
 801         struct extent_state *cached_state = NULL;
 802         u64 start = 0;
 803         u64 found_start;
 804         u64 found_end;
 805         int ret = 0;
 806
 807         lockdep_assert_held(&srcdev->fs_info->chunk_mutex);
 808
 809         while (find_first_extent_bit(&srcdev->alloc_state, start,
 810                                      &found_start, &found_end,
 811                                      CHUNK_ALLOCATED, &cached_state)) {
 812                 ret = set_extent_bit(&tgtdev->alloc_state, found_start,
 813                                      found_end, CHUNK_ALLOCATED, NULL);
 814                 if (ret)
 815                         break;
 816                 start = found_end + 1;
 817         }
 818
 819         free_extent_state(cached_state);
 820         return ret;
 821 }
 822
 823 static void btrfs_dev_replace_update_device_in_mapping_tree(
 824                                                 struct btrfs_fs_info *fs_info,
 825                                                 struct btrfs_device *srcdev,
 826                                                 struct btrfs_device *tgtdev)
 827 {
 828         struct rb_node *node;
 829
 830         /*
 831          * The chunk mutex must be held so that no new chunks can be created
 832          * while we are updating existing chunks. This guarantees we don't miss
 833          * any new chunk that gets created for a range that falls before the
 834          * range of the last chunk we processed.
 835          */
 836         lockdep_assert_held(&fs_info->chunk_mutex);
 837
 838         write_lock(&fs_info->mapping_tree_lock);
 839         node = rb_first_cached(&fs_info->mapping_tree);
 840         while (node) {
 841                 struct rb_node *next = rb_next(node);
 842                 struct btrfs_chunk_map *map;
 843                 u64 next_start;
 844
 845                 map = rb_entry(node, struct btrfs_chunk_map, rb_node);
 846                 next_start = map->start + map->chunk_len;
 847
 848                 for (int i = 0; i < map->num_stripes; i++)
 849                         if (srcdev == map->stripes[i].dev)
 850                                 map->stripes[i].dev = tgtdev;
 851
 852                 if (cond_resched_rwlock_write(&fs_info->mapping_tree_lock)) {
 853                         map = btrfs_find_chunk_map_nolock(fs_info, next_start, U64_MAX);
 854                         if (!map)
 855                                 break;
 856                         node = &map->rb_node;
 857                         /*
 858                          * Drop the lookup reference since we are holding the
 859                          * lock in write mode and no one can remove the chunk
 860                          * map from the tree and drop its tree reference.
 861                          */
 862                         btrfs_free_chunk_map(map);
 863                 } else {
 864                         node = next;
 865                 }
 866         }
 867         write_unlock(&fs_info->mapping_tree_lock);
 868 }
 869
 870 static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
 871                                        int scrub_ret)
 872 {
 873         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 874         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 875         struct btrfs_device *tgt_device;
 876         struct btrfs_device *src_device;
 877         struct btrfs_root *root = fs_info->tree_root;
 878         u8 uuid_tmp[BTRFS_UUID_SIZE];
 879         struct btrfs_trans_handle *trans;
 880         int ret = 0;
 881
 882         /* don't allow cancel or unmount to disturb the finishing procedure */
 883         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
 884
 885         down_read(&dev_replace->rwsem);
 886         /* was the operation canceled, or is it finished? */
 887         if (dev_replace->replace_state !=
 888             BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED) {
 889                 up_read(&dev_replace->rwsem);
 890                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 891                 return 0;
 892         }
 893
 894         tgt_device = dev_replace->tgtdev;
 895         src_device = dev_replace->srcdev;
 896         up_read(&dev_replace->rwsem);
 897
 898         /*
 899          * flush all outstanding I/O and inode extent mappings before the
 900          * copy operation is declared as being finished
 901          */
 902         ret = btrfs_start_delalloc_roots(fs_info, LONG_MAX, false);
 903         if (ret) {
 904                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 905                 return ret;
 906         }
 907         btrfs_wait_ordered_roots(fs_info, U64_MAX, NULL);
 908
 909         /*
 910          * We have to use this loop approach because at this point src_device
 911          * has to be available for transaction commit to complete, yet new
 912          * chunks shouldn't be allocated on the device.
 913          */
 914         while (1) {
 915                 trans = btrfs_start_transaction(root, 0);
 916                 if (IS_ERR(trans)) {
 917                         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 918                         return PTR_ERR(trans);
 919                 }
 920                 ret = btrfs_commit_transaction(trans);
 921                 WARN_ON(ret);
 922
 923                 /* Prevent write_all_supers() during the finishing procedure */
 924                 mutex_lock(&fs_devices->device_list_mutex);
 925                 /* Prevent new chunks being allocated on the source device */
 926                 mutex_lock(&fs_info->chunk_mutex);
 927
 928                 if (!list_empty(&src_device->post_commit_list)) {
 929                         mutex_unlock(&fs_devices->device_list_mutex);
 930                         mutex_unlock(&fs_info->chunk_mutex);
 931                 } else {
 932                         break;
 933                 }
 934         }
 935
 936         down_write(&dev_replace->rwsem);
 937         dev_replace->replace_state =
 938                 scrub_ret ? BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED
 939                           : BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED;
 940         dev_replace->tgtdev = NULL;
 941         dev_replace->srcdev = NULL;
 942         dev_replace->time_stopped = ktime_get_real_seconds();
 943         dev_replace->item_needs_writeback = 1;
 944
 945         /*
 946          * Update allocation state in the new device and replace the old device
 947          * with the new one in the mapping tree.
 948          */
 949         if (!scrub_ret) {
 950                 scrub_ret = btrfs_set_target_alloc_state(src_device, tgt_device);
 951                 if (scrub_ret)
 952                         goto error;
 953                 btrfs_dev_replace_update_device_in_mapping_tree(fs_info,
 954                                                                 src_device,
 955                                                                 tgt_device);
 956         } else {
 957                 if (scrub_ret != -ECANCELED)
 958                         btrfs_err_in_rcu(fs_info,
 959                                  "btrfs_scrub_dev(%s, %llu, %s) failed %d",
 960                                  btrfs_dev_name(src_device),
 961                                  src_device->devid,
 962                                  btrfs_dev_name(tgt_device), scrub_ret);
 963 error:
 964                 up_write(&dev_replace->rwsem);
 965                 mutex_unlock(&fs_info->chunk_mutex);
 966                 mutex_unlock(&fs_devices->device_list_mutex);
 967                 btrfs_rm_dev_replace_blocked(fs_info);
 968                 if (tgt_device)
 969                         btrfs_destroy_dev_replace_tgtdev(tgt_device);
 970                 btrfs_rm_dev_replace_unblocked(fs_info);
 971                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 972
 973                 return scrub_ret;
 974         }
 975
 976         btrfs_info_in_rcu(fs_info,
 977                           "dev_replace from %s (devid %llu) to %s finished",
 978                           btrfs_dev_name(src_device),
 979                           src_device->devid,
 980                           btrfs_dev_name(tgt_device));
 981         clear_bit(BTRFS_DEV_STATE_REPLACE_TGT, &tgt_device->dev_state);
 982         tgt_device->devid = src_device->devid;
 983         src_device->devid = BTRFS_DEV_REPLACE_DEVID;
 984         memcpy(uuid_tmp, tgt_device->uuid, sizeof(uuid_tmp));
 985         memcpy(tgt_device->uuid, src_device->uuid, sizeof(tgt_device->uuid));
 986         memcpy(src_device->uuid, uuid_tmp, sizeof(src_device->uuid));
 987         btrfs_device_set_total_bytes(tgt_device, src_device->total_bytes);
 988         btrfs_device_set_disk_total_bytes(tgt_device,
 989                                           src_device->disk_total_bytes);
 990         btrfs_device_set_bytes_used(tgt_device, src_device->bytes_used);
 991         tgt_device->commit_bytes_used = src_device->bytes_used;
 992
 993         btrfs_assign_next_active_device(src_device, tgt_device);
 994
 995         list_add(&tgt_device->dev_alloc_list, &fs_devices->alloc_list);
 996         fs_devices->rw_devices++;
 997
 998         dev_replace->replace_task = NULL;
 999         up_write(&dev_replace->rwsem);
1000         btrfs_rm_dev_replace_blocked(fs_info);
1001
1002         btrfs_rm_dev_replace_remove_srcdev(src_device);
1003
1004         btrfs_rm_dev_replace_unblocked(fs_info);
1005
1006         /*
1007          * Increment dev_stats_ccnt so that btrfs_run_dev_stats() will
1008          * update on-disk dev stats value during commit transaction
1009          */
1010         atomic_inc(&tgt_device->dev_stats_ccnt);
1011
1012         /*
1013          * this is again a consistent state where no dev_replace procedure
1014          * is running, the target device is part of the filesystem, the
1015          * source device is not part of the filesystem anymore and its 1st
1016          * superblock is scratched out so that it is no longer marked to
1017          * belong to this filesystem.
1018          */
1019         mutex_unlock(&fs_info->chunk_mutex);
1020         mutex_unlock(&fs_devices->device_list_mutex);
1021
1022         /* replace the sysfs entry */
1023         btrfs_sysfs_remove_device(src_device);
1024         btrfs_sysfs_update_devid(tgt_device);
1025         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &src_device->dev_state))
1026                 btrfs_scratch_superblocks(fs_info, src_device);
1027
1028         /* write back the superblocks */
1029         trans = btrfs_start_transaction(root, 0);
1030         if (!IS_ERR(trans))
1031                 btrfs_commit_transaction(trans);
1032
1033         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1034
1035         btrfs_rm_dev_replace_free_srcdev(src_device);
1036
1037         return 0;
1038 }
1039
1040 /*
1041  * Read progress of device replace status according to the state and last
1042  * stored position. The value format is the same as for
1043  * btrfs_dev_replace::progress_1000
1044  */
1045 static u64 btrfs_dev_replace_progress(struct btrfs_fs_info *fs_info)
1046 {
1047         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1048         u64 ret = 0;
1049
1050         switch (dev_replace->replace_state) {
1051         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1052         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1053                 ret = 0;
1054                 break;
1055         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1056                 ret = 1000;
1057                 break;
1058         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1059         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1060                 ret = div64_u64(dev_replace->cursor_left,
1061                                 div_u64(btrfs_device_get_total_bytes(
1062                                                 dev_replace->srcdev), 1000));
1063                 break;
1064         }
1065
1066         return ret;
1067 }
1068
1069 void btrfs_dev_replace_status(struct btrfs_fs_info *fs_info,
1070                               struct btrfs_ioctl_dev_replace_args *args)
1071 {
1072         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1073
1074         down_read(&dev_replace->rwsem);
1075         /* even if !dev_replace_is_valid, the values are good enough for
1076          * the replace_status ioctl */
1077         args->result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1078         args->status.replace_state = dev_replace->replace_state;
1079         args->status.time_started = dev_replace->time_started;
1080         args->status.time_stopped = dev_replace->time_stopped;
1081         args->status.num_write_errors =
1082                 atomic64_read(&dev_replace->num_write_errors);
1083         args->status.num_uncorrectable_read_errors =
1084                 atomic64_read(&dev_replace->num_uncorrectable_read_errors);
1085         args->status.progress_1000 = btrfs_dev_replace_progress(fs_info);
1086         up_read(&dev_replace->rwsem);
1087 }
1088
1089 int btrfs_dev_replace_cancel(struct btrfs_fs_info *fs_info)
1090 {
1091         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1092         struct btrfs_device *tgt_device = NULL;
1093         struct btrfs_device *src_device = NULL;
1094         struct btrfs_trans_handle *trans;
1095         struct btrfs_root *root = fs_info->tree_root;
1096         int result;
1097         int ret;
1098
1099         if (sb_rdonly(fs_info->sb))
1100                 return -EROFS;
1101
1102         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1103         down_write(&dev_replace->rwsem);
1104         switch (dev_replace->replace_state) {
1105         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1106         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1107         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1108                 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1109                 up_write(&dev_replace->rwsem);
1110                 break;
1111         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1112                 tgt_device = dev_replace->tgtdev;
1113                 src_device = dev_replace->srcdev;
1114                 up_write(&dev_replace->rwsem);
1115                 ret = btrfs_scrub_cancel(fs_info);
1116                 if (ret < 0) {
1117                         result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1118                 } else {
1119                         result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1120                         /*
1121                          * btrfs_dev_replace_finishing() will handle the
1122                          * cleanup part
1123                          */
1124                         btrfs_info_in_rcu(fs_info,
1125                                 "dev_replace from %s (devid %llu) to %s canceled",
1126                                 btrfs_dev_name(src_device), src_device->devid,
1127                                 btrfs_dev_name(tgt_device));
1128                 }
1129                 break;
1130         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1131                 /*
1132                  * Scrub doing the replace isn't running so we need to do the
1133                  * cleanup step of btrfs_dev_replace_finishing() here
1134                  */
1135                 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1136                 tgt_device = dev_replace->tgtdev;
1137                 src_device = dev_replace->srcdev;
1138                 dev_replace->tgtdev = NULL;
1139                 dev_replace->srcdev = NULL;
1140                 dev_replace->replace_state =
1141                                 BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED;
1142                 dev_replace->time_stopped = ktime_get_real_seconds();
1143                 dev_replace->item_needs_writeback = 1;
1144
1145                 up_write(&dev_replace->rwsem);
1146
1147                 /* Scrub for replace must not be running in suspended state */
1148                 btrfs_scrub_cancel(fs_info);
1149
1150                 trans = btrfs_start_transaction(root, 0);
1151                 if (IS_ERR(trans)) {
1152                         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1153                         return PTR_ERR(trans);
1154                 }
1155                 ret = btrfs_commit_transaction(trans);
1156                 WARN_ON(ret);
1157
1158                 btrfs_info_in_rcu(fs_info,
1159                 "suspended dev_replace from %s (devid %llu) to %s canceled",
1160                         btrfs_dev_name(src_device), src_device->devid,
1161                         btrfs_dev_name(tgt_device));
1162
1163                 if (tgt_device)
1164                         btrfs_destroy_dev_replace_tgtdev(tgt_device);
1165                 break;
1166         default:
1167                 up_write(&dev_replace->rwsem);
1168                 result = -EINVAL;
1169         }
1170
1171         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1172         return result;
1173 }
1174
1175 void btrfs_dev_replace_suspend_for_unmount(struct btrfs_fs_info *fs_info)
1176 {
1177         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1178
1179         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1180         down_write(&dev_replace->rwsem);
1181
1182         switch (dev_replace->replace_state) {
1183         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1184         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1185         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1186         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1187                 break;
1188         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1189                 dev_replace->replace_state =
1190                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1191                 dev_replace->time_stopped = ktime_get_real_seconds();
1192                 dev_replace->item_needs_writeback = 1;
1193                 btrfs_info(fs_info, "suspending dev_replace for unmount");
1194                 break;
1195         }
1196
1197         up_write(&dev_replace->rwsem);
1198         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1199 }
1200
1201 /* resume dev_replace procedure that was interrupted by unmount */
1202 int btrfs_resume_dev_replace_async(struct btrfs_fs_info *fs_info)
1203 {
1204         struct task_struct *task;
1205         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1206
1207         down_write(&dev_replace->rwsem);
1208
1209         switch (dev_replace->replace_state) {
1210         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1211         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1212         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1213                 up_write(&dev_replace->rwsem);
1214                 return 0;
1215         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1216                 break;
1217         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1218                 dev_replace->replace_state =
1219                         BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
1220                 break;
1221         }
1222         if (!dev_replace->tgtdev || !dev_replace->tgtdev->bdev) {
1223                 btrfs_info(fs_info,
1224                            "cannot continue dev_replace, tgtdev is missing");
1225                 btrfs_info(fs_info,
1226                            "you may cancel the operation after 'mount -o degraded'");
1227                 dev_replace->replace_state =
1228                                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1229                 up_write(&dev_replace->rwsem);
1230                 return 0;
1231         }
1232         up_write(&dev_replace->rwsem);
1233
1234         /*
1235          * This could collide with a paused balance, but the exclusive op logic
1236          * should never allow both to start and pause. We don't want to allow
1237          * dev-replace to start anyway.
1238          */
1239         if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_DEV_REPLACE)) {
1240                 down_write(&dev_replace->rwsem);
1241                 dev_replace->replace_state =
1242                                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1243                 up_write(&dev_replace->rwsem);
1244                 btrfs_info(fs_info,
1245                 "cannot resume dev-replace, other exclusive operation running");
1246                 return 0;
1247         }
1248
1249         task = kthread_run(btrfs_dev_replace_kthread, fs_info, "btrfs-devrepl");
1250         return PTR_ERR_OR_ZERO(task);
1251 }
1252
1253 static int btrfs_dev_replace_kthread(void *data)
1254 {
1255         struct btrfs_fs_info *fs_info = data;
1256         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1257         u64 progress;
1258         int ret;
1259
1260         progress = btrfs_dev_replace_progress(fs_info);
1261         progress = div_u64(progress, 10);
1262         btrfs_info_in_rcu(fs_info,
1263                 "continuing dev_replace from %s (devid %llu) to target %s @%u%%",
1264                 btrfs_dev_name(dev_replace->srcdev),
1265                 dev_replace->srcdev->devid,
1266                 btrfs_dev_name(dev_replace->tgtdev),
1267                 (unsigned int)progress);
1268
1269         ret = btrfs_scrub_dev(fs_info, dev_replace->srcdev->devid,
1270                               dev_replace->committed_cursor_left,
1271                               btrfs_device_get_total_bytes(dev_replace->srcdev),
1272                               &dev_replace->scrub_progress, 0, 1);
1273         ret = btrfs_dev_replace_finishing(fs_info, ret);
1274         WARN_ON(ret && ret != -ECANCELED);
1275
1276         btrfs_exclop_finish(fs_info);
1277         return 0;
1278 }
1279
1280 int __pure btrfs_dev_replace_is_ongoing(struct btrfs_dev_replace *dev_replace)
1281 {
1282         if (!dev_replace->is_valid)
1283                 return 0;
1284
1285         switch (dev_replace->replace_state) {
1286         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1287         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1288         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1289                 return 0;
1290         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1291         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1292                 /*
1293                  * return true even if tgtdev is missing (this is
1294                  * something that can happen if the dev_replace
1295                  * procedure is suspended by an umount and then
1296                  * the tgtdev is missing (or "btrfs dev scan") was
1297                  * not called and the filesystem is remounted
1298                  * in degraded state. This does not stop the
1299                  * dev_replace procedure. It needs to be canceled
1300                  * manually if the cancellation is wanted.
1301                  */
1302                 break;
1303         }
1304         return 1;
1305 }
1306
1307 void btrfs_bio_counter_sub(struct btrfs_fs_info *fs_info, s64 amount)
1308 {
1309         percpu_counter_sub(&fs_info->dev_replace.bio_counter, amount);
1310         cond_wake_up_nomb(&fs_info->dev_replace.replace_wait);
1311 }
1312
1313 void btrfs_bio_counter_inc_blocked(struct btrfs_fs_info *fs_info)
1314 {
1315         while (1) {
1316                 percpu_counter_inc(&fs_info->dev_replace.bio_counter);
1317                 if (likely(!test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1318                                      &fs_info->fs_state)))
1319                         break;
1320
1321                 btrfs_bio_counter_dec(fs_info);
1322                 wait_event(fs_info->dev_replace.replace_wait,
1323                            !test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1324                                      &fs_info->fs_state));
1325         }
1326 }