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