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_nr(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_device *device;
 247         struct block_device *bdev;
 248         struct rcu_string *name;
 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 = blkdev_get_by_path(device_path, FMODE_WRITE | FMODE_EXCL,
 259                                   fs_info->bdev_holder);
 260         if (IS_ERR(bdev)) {
 261                 btrfs_err(fs_info, "target device %s is invalid!", device_path);
 262                 return PTR_ERR(bdev);
 263         }
 264
 265         if (!btrfs_check_device_zone_type(fs_info, bdev)) {
 266                 btrfs_err(fs_info,
 267                 "dev-replace: zoned type of target device mismatch with filesystem");
 268                 ret = -EINVAL;
 269                 goto error;
 270         }
 271
 272         sync_blockdev(bdev);
 273
 274         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 275                 if (device->bdev == bdev) {
 276                         btrfs_err(fs_info,
 277                                   "target device is in the filesystem!");
 278                         ret = -EEXIST;
 279                         goto error;
 280                 }
 281         }
 282
 283
 284         if (i_size_read(bdev->bd_inode) <
 285             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
 307         set_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
 308         device->generation = 0;
 309         device->io_width = fs_info->sectorsize;
 310         device->io_align = fs_info->sectorsize;
 311         device->sector_size = fs_info->sectorsize;
 312         device->total_bytes = btrfs_device_get_total_bytes(srcdev);
 313         device->disk_total_bytes = btrfs_device_get_disk_total_bytes(srcdev);
 314         device->bytes_used = btrfs_device_get_bytes_used(srcdev);
 315         device->commit_total_bytes = srcdev->commit_total_bytes;
 316         device->commit_bytes_used = device->bytes_used;
 317         device->fs_info = fs_info;
 318         device->bdev = bdev;
 319         set_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
 320         set_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state);
 321         device->mode = FMODE_EXCL;
 322         device->dev_stats_valid = 1;
 323         set_blocksize(device->bdev, BTRFS_BDEV_BLOCKSIZE);
 324         device->fs_devices = fs_info->fs_devices;
 325
 326         ret = btrfs_get_dev_zone_info(device, false);
 327         if (ret)
 328                 goto error;
 329
 330         mutex_lock(&fs_info->fs_devices->device_list_mutex);
 331         list_add(&device->dev_list, &fs_info->fs_devices->devices);
 332         fs_info->fs_devices->num_devices++;
 333         fs_info->fs_devices->open_devices++;
 334         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
 335
 336         *device_out = device;
 337         return 0;
 338
 339 error:
 340         blkdev_put(bdev, FMODE_EXCL);
 341         return ret;
 342 }
 343
 344 /*
 345  * called from commit_transaction. Writes changed device replace state to
 346  * disk.
 347  */
 348 int btrfs_run_dev_replace(struct btrfs_trans_handle *trans)
 349 {
 350         struct btrfs_fs_info *fs_info = trans->fs_info;
 351         int ret;
 352         struct btrfs_root *dev_root = fs_info->dev_root;
 353         struct btrfs_path *path;
 354         struct btrfs_key key;
 355         struct extent_buffer *eb;
 356         struct btrfs_dev_replace_item *ptr;
 357         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 358
 359         down_read(&dev_replace->rwsem);
 360         if (!dev_replace->is_valid ||
 361             !dev_replace->item_needs_writeback) {
 362                 up_read(&dev_replace->rwsem);
 363                 return 0;
 364         }
 365         up_read(&dev_replace->rwsem);
 366
 367         key.objectid = 0;
 368         key.type = BTRFS_DEV_REPLACE_KEY;
 369         key.offset = 0;
 370
 371         path = btrfs_alloc_path();
 372         if (!path) {
 373                 ret = -ENOMEM;
 374                 goto out;
 375         }
 376         ret = btrfs_search_slot(trans, dev_root, &key, path, -1, 1);
 377         if (ret < 0) {
 378                 btrfs_warn(fs_info,
 379                            "error %d while searching for dev_replace item!",
 380                            ret);
 381                 goto out;
 382         }
 383
 384         if (ret == 0 &&
 385             btrfs_item_size_nr(path->nodes[0], path->slots[0]) < sizeof(*ptr)) {
 386                 /*
 387                  * need to delete old one and insert a new one.
 388                  * Since no attempt is made to recover any old state, if the
 389                  * dev_replace state is 'running', the data on the target
 390                  * drive is lost.
 391                  * It would be possible to recover the state: just make sure
 392                  * that the beginning of the item is never changed and always
 393                  * contains all the essential information. Then read this
 394                  * minimal set of information and use it as a base for the
 395                  * new state.
 396                  */
 397                 ret = btrfs_del_item(trans, dev_root, path);
 398                 if (ret != 0) {
 399                         btrfs_warn(fs_info,
 400                                    "delete too small dev_replace item failed %d!",
 401                                    ret);
 402                         goto out;
 403                 }
 404                 ret = 1;
 405         }
 406
 407         if (ret == 1) {
 408                 /* need to insert a new item */
 409                 btrfs_release_path(path);
 410                 ret = btrfs_insert_empty_item(trans, dev_root, path,
 411                                               &key, sizeof(*ptr));
 412                 if (ret < 0) {
 413                         btrfs_warn(fs_info,
 414                                    "insert dev_replace item failed %d!", ret);
 415                         goto out;
 416                 }
 417         }
 418
 419         eb = path->nodes[0];
 420         ptr = btrfs_item_ptr(eb, path->slots[0],
 421                              struct btrfs_dev_replace_item);
 422
 423         down_write(&dev_replace->rwsem);
 424         if (dev_replace->srcdev)
 425                 btrfs_set_dev_replace_src_devid(eb, ptr,
 426                         dev_replace->srcdev->devid);
 427         else
 428                 btrfs_set_dev_replace_src_devid(eb, ptr, (u64)-1);
 429         btrfs_set_dev_replace_cont_reading_from_srcdev_mode(eb, ptr,
 430                 dev_replace->cont_reading_from_srcdev_mode);
 431         btrfs_set_dev_replace_replace_state(eb, ptr,
 432                 dev_replace->replace_state);
 433         btrfs_set_dev_replace_time_started(eb, ptr, dev_replace->time_started);
 434         btrfs_set_dev_replace_time_stopped(eb, ptr, dev_replace->time_stopped);
 435         btrfs_set_dev_replace_num_write_errors(eb, ptr,
 436                 atomic64_read(&dev_replace->num_write_errors));
 437         btrfs_set_dev_replace_num_uncorrectable_read_errors(eb, ptr,
 438                 atomic64_read(&dev_replace->num_uncorrectable_read_errors));
 439         dev_replace->cursor_left_last_write_of_item =
 440                 dev_replace->cursor_left;
 441         btrfs_set_dev_replace_cursor_left(eb, ptr,
 442                 dev_replace->cursor_left_last_write_of_item);
 443         btrfs_set_dev_replace_cursor_right(eb, ptr,
 444                 dev_replace->cursor_right);
 445         dev_replace->item_needs_writeback = 0;
 446         up_write(&dev_replace->rwsem);
 447
 448         btrfs_mark_buffer_dirty(eb);
 449
 450 out:
 451         btrfs_free_path(path);
 452
 453         return ret;
 454 }
 455
 456 static char* btrfs_dev_name(struct btrfs_device *device)
 457 {
 458         if (!device || test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
 459                 return "<missing disk>";
 460         else
 461                 return rcu_str_deref(device->name);
 462 }
 463
 464 static int mark_block_group_to_copy(struct btrfs_fs_info *fs_info,
 465                                     struct btrfs_device *src_dev)
 466 {
 467         struct btrfs_path *path;
 468         struct btrfs_key key;
 469         struct btrfs_key found_key;
 470         struct btrfs_root *root = fs_info->dev_root;
 471         struct btrfs_dev_extent *dev_extent = NULL;
 472         struct btrfs_block_group *cache;
 473         struct btrfs_trans_handle *trans;
 474         int ret = 0;
 475         u64 chunk_offset;
 476
 477         /* Do not use "to_copy" on non zoned filesystem for now */
 478         if (!btrfs_is_zoned(fs_info))
 479                 return 0;
 480
 481         mutex_lock(&fs_info->chunk_mutex);
 482
 483         /* Ensure we don't have pending new block group */
 484         spin_lock(&fs_info->trans_lock);
 485         while (fs_info->running_transaction &&
 486                !list_empty(&fs_info->running_transaction->dev_update_list)) {
 487                 spin_unlock(&fs_info->trans_lock);
 488                 mutex_unlock(&fs_info->chunk_mutex);
 489                 trans = btrfs_attach_transaction(root);
 490                 if (IS_ERR(trans)) {
 491                         ret = PTR_ERR(trans);
 492                         mutex_lock(&fs_info->chunk_mutex);
 493                         if (ret == -ENOENT) {
 494                                 spin_lock(&fs_info->trans_lock);
 495                                 continue;
 496                         } else {
 497                                 goto unlock;
 498                         }
 499                 }
 500
 501                 ret = btrfs_commit_transaction(trans);
 502                 mutex_lock(&fs_info->chunk_mutex);
 503                 if (ret)
 504                         goto unlock;
 505
 506                 spin_lock(&fs_info->trans_lock);
 507         }
 508         spin_unlock(&fs_info->trans_lock);
 509
 510         path = btrfs_alloc_path();
 511         if (!path) {
 512                 ret = -ENOMEM;
 513                 goto unlock;
 514         }
 515
 516         path->reada = READA_FORWARD;
 517         path->search_commit_root = 1;
 518         path->skip_locking = 1;
 519
 520         key.objectid = src_dev->devid;
 521         key.type = BTRFS_DEV_EXTENT_KEY;
 522         key.offset = 0;
 523
 524         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 525         if (ret < 0)
 526                 goto free_path;
 527         if (ret > 0) {
 528                 if (path->slots[0] >=
 529                     btrfs_header_nritems(path->nodes[0])) {
 530                         ret = btrfs_next_leaf(root, path);
 531                         if (ret < 0)
 532                                 goto free_path;
 533                         if (ret > 0) {
 534                                 ret = 0;
 535                                 goto free_path;
 536                         }
 537                 } else {
 538                         ret = 0;
 539                 }
 540         }
 541
 542         while (1) {
 543                 struct extent_buffer *leaf = path->nodes[0];
 544                 int slot = path->slots[0];
 545
 546                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
 547
 548                 if (found_key.objectid != src_dev->devid)
 549                         break;
 550
 551                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
 552                         break;
 553
 554                 if (found_key.offset < key.offset)
 555                         break;
 556
 557                 dev_extent = btrfs_item_ptr(leaf, slot, struct btrfs_dev_extent);
 558
 559                 chunk_offset = btrfs_dev_extent_chunk_offset(leaf, dev_extent);
 560
 561                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
 562                 if (!cache)
 563                         goto skip;
 564
 565                 spin_lock(&cache->lock);
 566                 cache->to_copy = 1;
 567                 spin_unlock(&cache->lock);
 568
 569                 btrfs_put_block_group(cache);
 570
 571 skip:
 572                 ret = btrfs_next_item(root, path);
 573                 if (ret != 0) {
 574                         if (ret > 0)
 575                                 ret = 0;
 576                         break;
 577                 }
 578         }
 579
 580 free_path:
 581         btrfs_free_path(path);
 582 unlock:
 583         mutex_unlock(&fs_info->chunk_mutex);
 584
 585         return ret;
 586 }
 587
 588 bool btrfs_finish_block_group_to_copy(struct btrfs_device *srcdev,
 589                                       struct btrfs_block_group *cache,
 590                                       u64 physical)
 591 {
 592         struct btrfs_fs_info *fs_info = cache->fs_info;
 593         struct extent_map *em;
 594         struct map_lookup *map;
 595         u64 chunk_offset = cache->start;
 596         int num_extents, cur_extent;
 597         int i;
 598
 599         /* Do not use "to_copy" on non zoned filesystem for now */
 600         if (!btrfs_is_zoned(fs_info))
 601                 return true;
 602
 603         spin_lock(&cache->lock);
 604         if (cache->removed) {
 605                 spin_unlock(&cache->lock);
 606                 return true;
 607         }
 608         spin_unlock(&cache->lock);
 609
 610         em = btrfs_get_chunk_map(fs_info, chunk_offset, 1);
 611         ASSERT(!IS_ERR(em));
 612         map = em->map_lookup;
 613
 614         num_extents = cur_extent = 0;
 615         for (i = 0; i < map->num_stripes; i++) {
 616                 /* We have more device extent to copy */
 617                 if (srcdev != map->stripes[i].dev)
 618                         continue;
 619
 620                 num_extents++;
 621                 if (physical == map->stripes[i].physical)
 622                         cur_extent = i;
 623         }
 624
 625         free_extent_map(em);
 626
 627         if (num_extents > 1 && cur_extent < num_extents - 1) {
 628                 /*
 629                  * Has more stripes on this device. Keep this block group
 630                  * readonly until we finish all the stripes.
 631                  */
 632                 return false;
 633         }
 634
 635         /* Last stripe on this device */
 636         spin_lock(&cache->lock);
 637         cache->to_copy = 0;
 638         spin_unlock(&cache->lock);
 639
 640         return true;
 641 }
 642
 643 static int btrfs_dev_replace_start(struct btrfs_fs_info *fs_info,
 644                 const char *tgtdev_name, u64 srcdevid, const char *srcdev_name,
 645                 int read_src)
 646 {
 647         struct btrfs_root *root = fs_info->dev_root;
 648         struct btrfs_trans_handle *trans;
 649         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 650         int ret;
 651         struct btrfs_device *tgt_device = NULL;
 652         struct btrfs_device *src_device = NULL;
 653
 654         src_device = btrfs_find_device_by_devspec(fs_info, srcdevid,
 655                                                   srcdev_name);
 656         if (IS_ERR(src_device))
 657                 return PTR_ERR(src_device);
 658
 659         if (btrfs_pinned_by_swapfile(fs_info, src_device)) {
 660                 btrfs_warn_in_rcu(fs_info,
 661           "cannot replace device %s (devid %llu) due to active swapfile",
 662                         btrfs_dev_name(src_device), src_device->devid);
 663                 return -ETXTBSY;
 664         }
 665
 666         /*
 667          * Here we commit the transaction to make sure commit_total_bytes
 668          * of all the devices are updated.
 669          */
 670         trans = btrfs_attach_transaction(root);
 671         if (!IS_ERR(trans)) {
 672                 ret = btrfs_commit_transaction(trans);
 673                 if (ret)
 674                         return ret;
 675         } else if (PTR_ERR(trans) != -ENOENT) {
 676                 return PTR_ERR(trans);
 677         }
 678
 679         ret = btrfs_init_dev_replace_tgtdev(fs_info, tgtdev_name,
 680                                             src_device, &tgt_device);
 681         if (ret)
 682                 return ret;
 683
 684         ret = mark_block_group_to_copy(fs_info, src_device);
 685         if (ret)
 686                 return ret;
 687
 688         down_write(&dev_replace->rwsem);
 689         switch (dev_replace->replace_state) {
 690         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
 691         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
 692         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
 693                 break;
 694         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
 695         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
 696                 ASSERT(0);
 697                 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_ALREADY_STARTED;
 698                 up_write(&dev_replace->rwsem);
 699                 goto leave;
 700         }
 701
 702         dev_replace->cont_reading_from_srcdev_mode = read_src;
 703         dev_replace->srcdev = src_device;
 704         dev_replace->tgtdev = tgt_device;
 705
 706         btrfs_info_in_rcu(fs_info,
 707                       "dev_replace from %s (devid %llu) to %s started",
 708                       btrfs_dev_name(src_device),
 709                       src_device->devid,
 710                       rcu_str_deref(tgt_device->name));
 711
 712         /*
 713          * from now on, the writes to the srcdev are all duplicated to
 714          * go to the tgtdev as well (refer to btrfs_map_block()).
 715          */
 716         dev_replace->replace_state = BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
 717         dev_replace->time_started = ktime_get_real_seconds();
 718         dev_replace->cursor_left = 0;
 719         dev_replace->committed_cursor_left = 0;
 720         dev_replace->cursor_left_last_write_of_item = 0;
 721         dev_replace->cursor_right = 0;
 722         dev_replace->is_valid = 1;
 723         dev_replace->item_needs_writeback = 1;
 724         atomic64_set(&dev_replace->num_write_errors, 0);
 725         atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
 726         up_write(&dev_replace->rwsem);
 727
 728         ret = btrfs_sysfs_add_device(tgt_device);
 729         if (ret)
 730                 btrfs_err(fs_info, "kobj add dev failed %d", ret);
 731
 732         btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
 733
 734         /* Commit dev_replace state and reserve 1 item for it. */
 735         trans = btrfs_start_transaction(root, 1);
 736         if (IS_ERR(trans)) {
 737                 ret = PTR_ERR(trans);
 738                 down_write(&dev_replace->rwsem);
 739                 dev_replace->replace_state =
 740                         BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED;
 741                 dev_replace->srcdev = NULL;
 742                 dev_replace->tgtdev = NULL;
 743                 up_write(&dev_replace->rwsem);
 744                 goto leave;
 745         }
 746
 747         ret = btrfs_commit_transaction(trans);
 748         WARN_ON(ret);
 749
 750         /* the disk copy procedure reuses the scrub code */
 751         ret = btrfs_scrub_dev(fs_info, src_device->devid, 0,
 752                               btrfs_device_get_total_bytes(src_device),
 753                               &dev_replace->scrub_progress, 0, 1);
 754
 755         ret = btrfs_dev_replace_finishing(fs_info, ret);
 756         if (ret == -EINPROGRESS)
 757                 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS;
 758
 759         return ret;
 760
 761 leave:
 762         btrfs_destroy_dev_replace_tgtdev(tgt_device);
 763         return ret;
 764 }
 765
 766 int btrfs_dev_replace_by_ioctl(struct btrfs_fs_info *fs_info,
 767                             struct btrfs_ioctl_dev_replace_args *args)
 768 {
 769         int ret;
 770
 771         switch (args->start.cont_reading_from_srcdev_mode) {
 772         case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_ALWAYS:
 773         case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_AVOID:
 774                 break;
 775         default:
 776                 return -EINVAL;
 777         }
 778
 779         if ((args->start.srcdevid == 0 && args->start.srcdev_name[0] == '\0') ||
 780             args->start.tgtdev_name[0] == '\0')
 781                 return -EINVAL;
 782
 783         ret = btrfs_dev_replace_start(fs_info, args->start.tgtdev_name,
 784                                         args->start.srcdevid,
 785                                         args->start.srcdev_name,
 786                                         args->start.cont_reading_from_srcdev_mode);
 787         args->result = ret;
 788         /* don't warn if EINPROGRESS, someone else might be running scrub */
 789         if (ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS ||
 790             ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR)
 791                 return 0;
 792
 793         return ret;
 794 }
 795
 796 /*
 797  * blocked until all in-flight bios operations are finished.
 798  */
 799 static void btrfs_rm_dev_replace_blocked(struct btrfs_fs_info *fs_info)
 800 {
 801         set_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
 802         wait_event(fs_info->dev_replace.replace_wait, !percpu_counter_sum(
 803                    &fs_info->dev_replace.bio_counter));
 804 }
 805
 806 /*
 807  * we have removed target device, it is safe to allow new bios request.
 808  */
 809 static void btrfs_rm_dev_replace_unblocked(struct btrfs_fs_info *fs_info)
 810 {
 811         clear_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
 812         wake_up(&fs_info->dev_replace.replace_wait);
 813 }
 814
 815 /*
 816  * When finishing the device replace, before swapping the source device with the
 817  * target device we must update the chunk allocation state in the target device,
 818  * as it is empty because replace works by directly copying the chunks and not
 819  * through the normal chunk allocation path.
 820  */
 821 static int btrfs_set_target_alloc_state(struct btrfs_device *srcdev,
 822                                         struct btrfs_device *tgtdev)
 823 {
 824         struct extent_state *cached_state = NULL;
 825         u64 start = 0;
 826         u64 found_start;
 827         u64 found_end;
 828         int ret = 0;
 829
 830         lockdep_assert_held(&srcdev->fs_info->chunk_mutex);
 831
 832         while (!find_first_extent_bit(&srcdev->alloc_state, start,
 833                                       &found_start, &found_end,
 834                                       CHUNK_ALLOCATED, &cached_state)) {
 835                 ret = set_extent_bits(&tgtdev->alloc_state, found_start,
 836                                       found_end, CHUNK_ALLOCATED);
 837                 if (ret)
 838                         break;
 839                 start = found_end + 1;
 840         }
 841
 842         free_extent_state(cached_state);
 843         return ret;
 844 }
 845
 846 static void btrfs_dev_replace_update_device_in_mapping_tree(
 847                                                 struct btrfs_fs_info *fs_info,
 848                                                 struct btrfs_device *srcdev,
 849                                                 struct btrfs_device *tgtdev)
 850 {
 851         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
 852         struct extent_map *em;
 853         struct map_lookup *map;
 854         u64 start = 0;
 855         int i;
 856
 857         write_lock(&em_tree->lock);
 858         do {
 859                 em = lookup_extent_mapping(em_tree, start, (u64)-1);
 860                 if (!em)
 861                         break;
 862                 map = em->map_lookup;
 863                 for (i = 0; i < map->num_stripes; i++)
 864                         if (srcdev == map->stripes[i].dev)
 865                                 map->stripes[i].dev = tgtdev;
 866                 start = em->start + em->len;
 867                 free_extent_map(em);
 868         } while (start);
 869         write_unlock(&em_tree->lock);
 870 }
 871
 872 static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
 873                                        int scrub_ret)
 874 {
 875         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 876         struct btrfs_device *tgt_device;
 877         struct btrfs_device *src_device;
 878         struct btrfs_root *root = fs_info->tree_root;
 879         u8 uuid_tmp[BTRFS_UUID_SIZE];
 880         struct btrfs_trans_handle *trans;
 881         int ret = 0;
 882
 883         /* don't allow cancel or unmount to disturb the finishing procedure */
 884         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
 885
 886         down_read(&dev_replace->rwsem);
 887         /* was the operation canceled, or is it finished? */
 888         if (dev_replace->replace_state !=
 889             BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED) {
 890                 up_read(&dev_replace->rwsem);
 891                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 892                 return 0;
 893         }
 894
 895         tgt_device = dev_replace->tgtdev;
 896         src_device = dev_replace->srcdev;
 897         up_read(&dev_replace->rwsem);
 898
 899         /*
 900          * flush all outstanding I/O and inode extent mappings before the
 901          * copy operation is declared as being finished
 902          */
 903         ret = btrfs_start_delalloc_roots(fs_info, LONG_MAX, false);
 904         if (ret) {
 905                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 906                 return ret;
 907         }
 908         btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
 909
 910         if (!scrub_ret)
 911                 btrfs_reada_remove_dev(src_device);
 912
 913         /*
 914          * We have to use this loop approach because at this point src_device
 915          * has to be available for transaction commit to complete, yet new
 916          * chunks shouldn't be allocated on the device.
 917          */
 918         while (1) {
 919                 trans = btrfs_start_transaction(root, 0);
 920                 if (IS_ERR(trans)) {
 921                         btrfs_reada_undo_remove_dev(src_device);
 922                         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 923                         return PTR_ERR(trans);
 924                 }
 925                 ret = btrfs_commit_transaction(trans);
 926                 WARN_ON(ret);
 927
 928                 /* Prevent write_all_supers() during the finishing procedure */
 929                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
 930                 /* Prevent new chunks being allocated on the source device */
 931                 mutex_lock(&fs_info->chunk_mutex);
 932
 933                 if (!list_empty(&src_device->post_commit_list)) {
 934                         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
 935                         mutex_unlock(&fs_info->chunk_mutex);
 936                 } else {
 937                         break;
 938                 }
 939         }
 940
 941         down_write(&dev_replace->rwsem);
 942         dev_replace->replace_state =
 943                 scrub_ret ? BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED
 944                           : BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED;
 945         dev_replace->tgtdev = NULL;
 946         dev_replace->srcdev = NULL;
 947         dev_replace->time_stopped = ktime_get_real_seconds();
 948         dev_replace->item_needs_writeback = 1;
 949
 950         /*
 951          * Update allocation state in the new device and replace the old device
 952          * with the new one in the mapping tree.
 953          */
 954         if (!scrub_ret) {
 955                 scrub_ret = btrfs_set_target_alloc_state(src_device, tgt_device);
 956                 if (scrub_ret)
 957                         goto error;
 958                 btrfs_dev_replace_update_device_in_mapping_tree(fs_info,
 959                                                                 src_device,
 960                                                                 tgt_device);
 961         } else {
 962                 if (scrub_ret != -ECANCELED)
 963                         btrfs_err_in_rcu(fs_info,
 964                                  "btrfs_scrub_dev(%s, %llu, %s) failed %d",
 965                                  btrfs_dev_name(src_device),
 966                                  src_device->devid,
 967                                  rcu_str_deref(tgt_device->name), scrub_ret);
 968 error:
 969                 up_write(&dev_replace->rwsem);
 970                 mutex_unlock(&fs_info->chunk_mutex);
 971                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
 972                 btrfs_reada_undo_remove_dev(src_device);
 973                 btrfs_rm_dev_replace_blocked(fs_info);
 974                 if (tgt_device)
 975                         btrfs_destroy_dev_replace_tgtdev(tgt_device);
 976                 btrfs_rm_dev_replace_unblocked(fs_info);
 977                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 978
 979                 return scrub_ret;
 980         }
 981
 982         btrfs_info_in_rcu(fs_info,
 983                           "dev_replace from %s (devid %llu) to %s finished",
 984                           btrfs_dev_name(src_device),
 985                           src_device->devid,
 986                           rcu_str_deref(tgt_device->name));
 987         clear_bit(BTRFS_DEV_STATE_REPLACE_TGT, &tgt_device->dev_state);
 988         tgt_device->devid = src_device->devid;
 989         src_device->devid = BTRFS_DEV_REPLACE_DEVID;
 990         memcpy(uuid_tmp, tgt_device->uuid, sizeof(uuid_tmp));
 991         memcpy(tgt_device->uuid, src_device->uuid, sizeof(tgt_device->uuid));
 992         memcpy(src_device->uuid, uuid_tmp, sizeof(src_device->uuid));
 993         btrfs_device_set_total_bytes(tgt_device, src_device->total_bytes);
 994         btrfs_device_set_disk_total_bytes(tgt_device,
 995                                           src_device->disk_total_bytes);
 996         btrfs_device_set_bytes_used(tgt_device, src_device->bytes_used);
 997         tgt_device->commit_bytes_used = src_device->bytes_used;
 998
 999         btrfs_assign_next_active_device(src_device, tgt_device);
1000
1001         list_add(&tgt_device->dev_alloc_list, &fs_info->fs_devices->alloc_list);
1002         fs_info->fs_devices->rw_devices++;
1003
1004         up_write(&dev_replace->rwsem);
1005         btrfs_rm_dev_replace_blocked(fs_info);
1006
1007         btrfs_rm_dev_replace_remove_srcdev(src_device);
1008
1009         btrfs_rm_dev_replace_unblocked(fs_info);
1010
1011         /*
1012          * Increment dev_stats_ccnt so that btrfs_run_dev_stats() will
1013          * update on-disk dev stats value during commit transaction
1014          */
1015         atomic_inc(&tgt_device->dev_stats_ccnt);
1016
1017         /*
1018          * this is again a consistent state where no dev_replace procedure
1019          * is running, the target device is part of the filesystem, the
1020          * source device is not part of the filesystem anymore and its 1st
1021          * superblock is scratched out so that it is no longer marked to
1022          * belong to this filesystem.
1023          */
1024         mutex_unlock(&fs_info->chunk_mutex);
1025         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1026
1027         /* replace the sysfs entry */
1028         btrfs_sysfs_remove_device(src_device);
1029         btrfs_sysfs_update_devid(tgt_device);
1030         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &src_device->dev_state))
1031                 btrfs_scratch_superblocks(fs_info, src_device->bdev,
1032                                           src_device->name->str);
1033
1034         /* write back the superblocks */
1035         trans = btrfs_start_transaction(root, 0);
1036         if (!IS_ERR(trans))
1037                 btrfs_commit_transaction(trans);
1038
1039         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1040
1041         btrfs_rm_dev_replace_free_srcdev(src_device);
1042
1043         return 0;
1044 }
1045
1046 /*
1047  * Read progress of device replace status according to the state and last
1048  * stored position. The value format is the same as for
1049  * btrfs_dev_replace::progress_1000
1050  */
1051 static u64 btrfs_dev_replace_progress(struct btrfs_fs_info *fs_info)
1052 {
1053         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1054         u64 ret = 0;
1055
1056         switch (dev_replace->replace_state) {
1057         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1058         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1059                 ret = 0;
1060                 break;
1061         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1062                 ret = 1000;
1063                 break;
1064         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1065         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1066                 ret = div64_u64(dev_replace->cursor_left,
1067                                 div_u64(btrfs_device_get_total_bytes(
1068                                                 dev_replace->srcdev), 1000));
1069                 break;
1070         }
1071
1072         return ret;
1073 }
1074
1075 void btrfs_dev_replace_status(struct btrfs_fs_info *fs_info,
1076                               struct btrfs_ioctl_dev_replace_args *args)
1077 {
1078         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1079
1080         down_read(&dev_replace->rwsem);
1081         /* even if !dev_replace_is_valid, the values are good enough for
1082          * the replace_status ioctl */
1083         args->result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1084         args->status.replace_state = dev_replace->replace_state;
1085         args->status.time_started = dev_replace->time_started;
1086         args->status.time_stopped = dev_replace->time_stopped;
1087         args->status.num_write_errors =
1088                 atomic64_read(&dev_replace->num_write_errors);
1089         args->status.num_uncorrectable_read_errors =
1090                 atomic64_read(&dev_replace->num_uncorrectable_read_errors);
1091         args->status.progress_1000 = btrfs_dev_replace_progress(fs_info);
1092         up_read(&dev_replace->rwsem);
1093 }
1094
1095 int btrfs_dev_replace_cancel(struct btrfs_fs_info *fs_info)
1096 {
1097         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1098         struct btrfs_device *tgt_device = NULL;
1099         struct btrfs_device *src_device = NULL;
1100         struct btrfs_trans_handle *trans;
1101         struct btrfs_root *root = fs_info->tree_root;
1102         int result;
1103         int ret;
1104
1105         if (sb_rdonly(fs_info->sb))
1106                 return -EROFS;
1107
1108         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1109         down_write(&dev_replace->rwsem);
1110         switch (dev_replace->replace_state) {
1111         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1112         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1113         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1114                 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1115                 up_write(&dev_replace->rwsem);
1116                 break;
1117         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1118                 tgt_device = dev_replace->tgtdev;
1119                 src_device = dev_replace->srcdev;
1120                 up_write(&dev_replace->rwsem);
1121                 ret = btrfs_scrub_cancel(fs_info);
1122                 if (ret < 0) {
1123                         result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1124                 } else {
1125                         result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1126                         /*
1127                          * btrfs_dev_replace_finishing() will handle the
1128                          * cleanup part
1129                          */
1130                         btrfs_info_in_rcu(fs_info,
1131                                 "dev_replace from %s (devid %llu) to %s canceled",
1132                                 btrfs_dev_name(src_device), src_device->devid,
1133                                 btrfs_dev_name(tgt_device));
1134                 }
1135                 break;
1136         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1137                 /*
1138                  * Scrub doing the replace isn't running so we need to do the
1139                  * cleanup step of btrfs_dev_replace_finishing() here
1140                  */
1141                 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1142                 tgt_device = dev_replace->tgtdev;
1143                 src_device = dev_replace->srcdev;
1144                 dev_replace->tgtdev = NULL;
1145                 dev_replace->srcdev = NULL;
1146                 dev_replace->replace_state =
1147                                 BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED;
1148                 dev_replace->time_stopped = ktime_get_real_seconds();
1149                 dev_replace->item_needs_writeback = 1;
1150
1151                 up_write(&dev_replace->rwsem);
1152
1153                 /* Scrub for replace must not be running in suspended state */
1154                 btrfs_scrub_cancel(fs_info);
1155
1156                 trans = btrfs_start_transaction(root, 0);
1157                 if (IS_ERR(trans)) {
1158                         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1159                         return PTR_ERR(trans);
1160                 }
1161                 ret = btrfs_commit_transaction(trans);
1162                 WARN_ON(ret);
1163
1164                 btrfs_info_in_rcu(fs_info,
1165                 "suspended dev_replace from %s (devid %llu) to %s canceled",
1166                         btrfs_dev_name(src_device), src_device->devid,
1167                         btrfs_dev_name(tgt_device));
1168
1169                 if (tgt_device)
1170                         btrfs_destroy_dev_replace_tgtdev(tgt_device);
1171                 break;
1172         default:
1173                 up_write(&dev_replace->rwsem);
1174                 result = -EINVAL;
1175         }
1176
1177         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1178         return result;
1179 }
1180
1181 void btrfs_dev_replace_suspend_for_unmount(struct btrfs_fs_info *fs_info)
1182 {
1183         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1184
1185         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
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         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1193                 break;
1194         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1195                 dev_replace->replace_state =
1196                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1197                 dev_replace->time_stopped = ktime_get_real_seconds();
1198                 dev_replace->item_needs_writeback = 1;
1199                 btrfs_info(fs_info, "suspending dev_replace for unmount");
1200                 break;
1201         }
1202
1203         up_write(&dev_replace->rwsem);
1204         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1205 }
1206
1207 /* resume dev_replace procedure that was interrupted by unmount */
1208 int btrfs_resume_dev_replace_async(struct btrfs_fs_info *fs_info)
1209 {
1210         struct task_struct *task;
1211         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1212
1213         down_write(&dev_replace->rwsem);
1214
1215         switch (dev_replace->replace_state) {
1216         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1217         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1218         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1219                 up_write(&dev_replace->rwsem);
1220                 return 0;
1221         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1222                 break;
1223         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1224                 dev_replace->replace_state =
1225                         BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
1226                 break;
1227         }
1228         if (!dev_replace->tgtdev || !dev_replace->tgtdev->bdev) {
1229                 btrfs_info(fs_info,
1230                            "cannot continue dev_replace, tgtdev is missing");
1231                 btrfs_info(fs_info,
1232                            "you may cancel the operation after 'mount -o degraded'");
1233                 dev_replace->replace_state =
1234                                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1235                 up_write(&dev_replace->rwsem);
1236                 return 0;
1237         }
1238         up_write(&dev_replace->rwsem);
1239
1240         /*
1241          * This could collide with a paused balance, but the exclusive op logic
1242          * should never allow both to start and pause. We don't want to allow
1243          * dev-replace to start anyway.
1244          */
1245         if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_DEV_REPLACE)) {
1246                 down_write(&dev_replace->rwsem);
1247                 dev_replace->replace_state =
1248                                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1249                 up_write(&dev_replace->rwsem);
1250                 btrfs_info(fs_info,
1251                 "cannot resume dev-replace, other exclusive operation running");
1252                 return 0;
1253         }
1254
1255         task = kthread_run(btrfs_dev_replace_kthread, fs_info, "btrfs-devrepl");
1256         return PTR_ERR_OR_ZERO(task);
1257 }
1258
1259 static int btrfs_dev_replace_kthread(void *data)
1260 {
1261         struct btrfs_fs_info *fs_info = data;
1262         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1263         u64 progress;
1264         int ret;
1265
1266         progress = btrfs_dev_replace_progress(fs_info);
1267         progress = div_u64(progress, 10);
1268         btrfs_info_in_rcu(fs_info,
1269                 "continuing dev_replace from %s (devid %llu) to target %s @%u%%",
1270                 btrfs_dev_name(dev_replace->srcdev),
1271                 dev_replace->srcdev->devid,
1272                 btrfs_dev_name(dev_replace->tgtdev),
1273                 (unsigned int)progress);
1274
1275         ret = btrfs_scrub_dev(fs_info, dev_replace->srcdev->devid,
1276                               dev_replace->committed_cursor_left,
1277                               btrfs_device_get_total_bytes(dev_replace->srcdev),
1278                               &dev_replace->scrub_progress, 0, 1);
1279         ret = btrfs_dev_replace_finishing(fs_info, ret);
1280         WARN_ON(ret && ret != -ECANCELED);
1281
1282         btrfs_exclop_finish(fs_info);
1283         return 0;
1284 }
1285
1286 int __pure btrfs_dev_replace_is_ongoing(struct btrfs_dev_replace *dev_replace)
1287 {
1288         if (!dev_replace->is_valid)
1289                 return 0;
1290
1291         switch (dev_replace->replace_state) {
1292         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1293         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1294         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1295                 return 0;
1296         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1297         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1298                 /*
1299                  * return true even if tgtdev is missing (this is
1300                  * something that can happen if the dev_replace
1301                  * procedure is suspended by an umount and then
1302                  * the tgtdev is missing (or "btrfs dev scan") was
1303                  * not called and the filesystem is remounted
1304                  * in degraded state. This does not stop the
1305                  * dev_replace procedure. It needs to be canceled
1306                  * manually if the cancellation is wanted.
1307                  */
1308                 break;
1309         }
1310         return 1;
1311 }
1312
1313 void btrfs_bio_counter_inc_noblocked(struct btrfs_fs_info *fs_info)
1314 {
1315         percpu_counter_inc(&fs_info->dev_replace.bio_counter);
1316 }
1317
1318 void btrfs_bio_counter_sub(struct btrfs_fs_info *fs_info, s64 amount)
1319 {
1320         percpu_counter_sub(&fs_info->dev_replace.bio_counter, amount);
1321         cond_wake_up_nomb(&fs_info->dev_replace.replace_wait);
1322 }
1323
1324 void btrfs_bio_counter_inc_blocked(struct btrfs_fs_info *fs_info)
1325 {
1326         while (1) {
1327                 percpu_counter_inc(&fs_info->dev_replace.bio_counter);
1328                 if (likely(!test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1329                                      &fs_info->fs_state)))
1330                         break;
1331
1332                 btrfs_bio_counter_dec(fs_info);
1333                 wait_event(fs_info->dev_replace.replace_wait,
1334                            !test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1335                                      &fs_info->fs_state));
1336         }
1337 }