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_key key;
  74         struct btrfs_root *dev_root = fs_info->dev_root;
  75         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
  76         struct extent_buffer *eb;
  77         int slot;
  78         int ret = 0;
  79         struct btrfs_path *path = NULL;
  80         int item_size;
  81         struct btrfs_dev_replace_item *ptr;
  82         u64 src_devid;
  83
  84         if (!dev_root)
  85                 return 0;
  86
  87         path = btrfs_alloc_path();
  88         if (!path) {
  89                 ret = -ENOMEM;
  90                 goto out;
  91         }
  92
  93         key.objectid = 0;
  94         key.type = BTRFS_DEV_REPLACE_KEY;
  95         key.offset = 0;
  96         ret = btrfs_search_slot(NULL, dev_root, &key, path, 0, 0);
  97         if (ret) {
  98 no_valid_dev_replace_entry_found:
  99                 /*
 100                  * We don't have a replace item or it's corrupted.  If there is
 101                  * a replace target, fail the mount.
 102                  */
 103                 if (btrfs_find_device(fs_info->fs_devices,
 104                                       BTRFS_DEV_REPLACE_DEVID, NULL, NULL)) {
 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,
 167                                       BTRFS_DEV_REPLACE_DEVID, NULL, NULL)) {
 168                         btrfs_err(fs_info,
 169                         "replace devid present without an active replace item");
 170                         ret = -EUCLEAN;
 171                 } else {
 172                         dev_replace->srcdev = NULL;
 173                         dev_replace->tgtdev = NULL;
 174                 }
 175                 break;
 176         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
 177         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
 178                 dev_replace->srcdev = btrfs_find_device(fs_info->fs_devices,
 179                                                 src_devid, NULL, NULL);
 180                 dev_replace->tgtdev = btrfs_find_device(fs_info->fs_devices,
 181                                                         BTRFS_DEV_REPLACE_DEVID,
 182                                                         NULL, NULL);
 183                 /*
 184                  * allow 'btrfs dev replace_cancel' if src/tgt device is
 185                  * missing
 186                  */
 187                 if (!dev_replace->srcdev &&
 188                     !btrfs_test_opt(fs_info, DEGRADED)) {
 189                         ret = -EIO;
 190                         btrfs_warn(fs_info,
 191                            "cannot mount because device replace operation is ongoing and");
 192                         btrfs_warn(fs_info,
 193                            "srcdev (devid %llu) is missing, need to run 'btrfs dev scan'?",
 194                            src_devid);
 195                 }
 196                 if (!dev_replace->tgtdev &&
 197                     !btrfs_test_opt(fs_info, DEGRADED)) {
 198                         ret = -EIO;
 199                         btrfs_warn(fs_info,
 200                            "cannot mount because device replace operation is ongoing and");
 201                         btrfs_warn(fs_info,
 202                            "tgtdev (devid %llu) is missing, need to run 'btrfs dev scan'?",
 203                                 BTRFS_DEV_REPLACE_DEVID);
 204                 }
 205                 if (dev_replace->tgtdev) {
 206                         if (dev_replace->srcdev) {
 207                                 dev_replace->tgtdev->total_bytes =
 208                                         dev_replace->srcdev->total_bytes;
 209                                 dev_replace->tgtdev->disk_total_bytes =
 210                                         dev_replace->srcdev->disk_total_bytes;
 211                                 dev_replace->tgtdev->commit_total_bytes =
 212                                         dev_replace->srcdev->commit_total_bytes;
 213                                 dev_replace->tgtdev->bytes_used =
 214                                         dev_replace->srcdev->bytes_used;
 215                                 dev_replace->tgtdev->commit_bytes_used =
 216                                         dev_replace->srcdev->commit_bytes_used;
 217                         }
 218                         set_bit(BTRFS_DEV_STATE_REPLACE_TGT,
 219                                 &dev_replace->tgtdev->dev_state);
 220
 221                         WARN_ON(fs_info->fs_devices->rw_devices == 0);
 222                         dev_replace->tgtdev->io_width = fs_info->sectorsize;
 223                         dev_replace->tgtdev->io_align = fs_info->sectorsize;
 224                         dev_replace->tgtdev->sector_size = fs_info->sectorsize;
 225                         dev_replace->tgtdev->fs_info = fs_info;
 226                         set_bit(BTRFS_DEV_STATE_IN_FS_METADATA,
 227                                 &dev_replace->tgtdev->dev_state);
 228                 }
 229                 break;
 230         }
 231
 232 out:
 233         btrfs_free_path(path);
 234         return ret;
 235 }
 236
 237 /*
 238  * Initialize a new device for device replace target from a given source dev
 239  * and path.
 240  *
 241  * Return 0 and new device in @device_out, otherwise return < 0
 242  */
 243 static int btrfs_init_dev_replace_tgtdev(struct btrfs_fs_info *fs_info,
 244                                   const char *device_path,
 245                                   struct btrfs_device *srcdev,
 246                                   struct btrfs_device **device_out)
 247 {
 248         struct btrfs_device *device;
 249         struct block_device *bdev;
 250         struct rcu_string *name;
 251         u64 devid = BTRFS_DEV_REPLACE_DEVID;
 252         int ret = 0;
 253
 254         *device_out = NULL;
 255         if (srcdev->fs_devices->seeding) {
 256                 btrfs_err(fs_info, "the filesystem is a seed filesystem!");
 257                 return -EINVAL;
 258         }
 259
 260         bdev = blkdev_get_by_path(device_path, FMODE_WRITE | FMODE_EXCL,
 261                                   fs_info->bdev_holder);
 262         if (IS_ERR(bdev)) {
 263                 btrfs_err(fs_info, "target device %s is invalid!", device_path);
 264                 return PTR_ERR(bdev);
 265         }
 266
 267         if (!btrfs_check_device_zone_type(fs_info, bdev)) {
 268                 btrfs_err(fs_info,
 269                 "dev-replace: zoned type of target device mismatch with filesystem");
 270                 ret = -EINVAL;
 271                 goto error;
 272         }
 273
 274         sync_blockdev(bdev);
 275
 276         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 277                 if (device->bdev == bdev) {
 278                         btrfs_err(fs_info,
 279                                   "target device is in the filesystem!");
 280                         ret = -EEXIST;
 281                         goto error;
 282                 }
 283         }
 284
 285
 286         if (i_size_read(bdev->bd_inode) <
 287             btrfs_device_get_total_bytes(srcdev)) {
 288                 btrfs_err(fs_info,
 289                           "target device is smaller than source device!");
 290                 ret = -EINVAL;
 291                 goto error;
 292         }
 293
 294
 295         device = btrfs_alloc_device(NULL, &devid, NULL);
 296         if (IS_ERR(device)) {
 297                 ret = PTR_ERR(device);
 298                 goto error;
 299         }
 300
 301         name = rcu_string_strdup(device_path, GFP_KERNEL);
 302         if (!name) {
 303                 btrfs_free_device(device);
 304                 ret = -ENOMEM;
 305                 goto error;
 306         }
 307         rcu_assign_pointer(device->name, name);
 308
 309         set_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
 310         device->generation = 0;
 311         device->io_width = fs_info->sectorsize;
 312         device->io_align = fs_info->sectorsize;
 313         device->sector_size = fs_info->sectorsize;
 314         device->total_bytes = btrfs_device_get_total_bytes(srcdev);
 315         device->disk_total_bytes = btrfs_device_get_disk_total_bytes(srcdev);
 316         device->bytes_used = btrfs_device_get_bytes_used(srcdev);
 317         device->commit_total_bytes = srcdev->commit_total_bytes;
 318         device->commit_bytes_used = device->bytes_used;
 319         device->fs_info = fs_info;
 320         device->bdev = bdev;
 321         set_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
 322         set_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state);
 323         device->mode = FMODE_EXCL;
 324         device->dev_stats_valid = 1;
 325         set_blocksize(device->bdev, BTRFS_BDEV_BLOCKSIZE);
 326         device->fs_devices = fs_info->fs_devices;
 327
 328         ret = btrfs_get_dev_zone_info(device);
 329         if (ret)
 330                 goto error;
 331
 332         mutex_lock(&fs_info->fs_devices->device_list_mutex);
 333         list_add(&device->dev_list, &fs_info->fs_devices->devices);
 334         fs_info->fs_devices->num_devices++;
 335         fs_info->fs_devices->open_devices++;
 336         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
 337
 338         *device_out = device;
 339         return 0;
 340
 341 error:
 342         blkdev_put(bdev, FMODE_EXCL);
 343         return ret;
 344 }
 345
 346 /*
 347  * called from commit_transaction. Writes changed device replace state to
 348  * disk.
 349  */
 350 int btrfs_run_dev_replace(struct btrfs_trans_handle *trans)
 351 {
 352         struct btrfs_fs_info *fs_info = trans->fs_info;
 353         int ret;
 354         struct btrfs_root *dev_root = fs_info->dev_root;
 355         struct btrfs_path *path;
 356         struct btrfs_key key;
 357         struct extent_buffer *eb;
 358         struct btrfs_dev_replace_item *ptr;
 359         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 360
 361         down_read(&dev_replace->rwsem);
 362         if (!dev_replace->is_valid ||
 363             !dev_replace->item_needs_writeback) {
 364                 up_read(&dev_replace->rwsem);
 365                 return 0;
 366         }
 367         up_read(&dev_replace->rwsem);
 368
 369         key.objectid = 0;
 370         key.type = BTRFS_DEV_REPLACE_KEY;
 371         key.offset = 0;
 372
 373         path = btrfs_alloc_path();
 374         if (!path) {
 375                 ret = -ENOMEM;
 376                 goto out;
 377         }
 378         ret = btrfs_search_slot(trans, dev_root, &key, path, -1, 1);
 379         if (ret < 0) {
 380                 btrfs_warn(fs_info,
 381                            "error %d while searching for dev_replace item!",
 382                            ret);
 383                 goto out;
 384         }
 385
 386         if (ret == 0 &&
 387             btrfs_item_size_nr(path->nodes[0], path->slots[0]) < sizeof(*ptr)) {
 388                 /*
 389                  * need to delete old one and insert a new one.
 390                  * Since no attempt is made to recover any old state, if the
 391                  * dev_replace state is 'running', the data on the target
 392                  * drive is lost.
 393                  * It would be possible to recover the state: just make sure
 394                  * that the beginning of the item is never changed and always
 395                  * contains all the essential information. Then read this
 396                  * minimal set of information and use it as a base for the
 397                  * new state.
 398                  */
 399                 ret = btrfs_del_item(trans, dev_root, path);
 400                 if (ret != 0) {
 401                         btrfs_warn(fs_info,
 402                                    "delete too small dev_replace item failed %d!",
 403                                    ret);
 404                         goto out;
 405                 }
 406                 ret = 1;
 407         }
 408
 409         if (ret == 1) {
 410                 /* need to insert a new item */
 411                 btrfs_release_path(path);
 412                 ret = btrfs_insert_empty_item(trans, dev_root, path,
 413                                               &key, sizeof(*ptr));
 414                 if (ret < 0) {
 415                         btrfs_warn(fs_info,
 416                                    "insert dev_replace item failed %d!", ret);
 417                         goto out;
 418                 }
 419         }
 420
 421         eb = path->nodes[0];
 422         ptr = btrfs_item_ptr(eb, path->slots[0],
 423                              struct btrfs_dev_replace_item);
 424
 425         down_write(&dev_replace->rwsem);
 426         if (dev_replace->srcdev)
 427                 btrfs_set_dev_replace_src_devid(eb, ptr,
 428                         dev_replace->srcdev->devid);
 429         else
 430                 btrfs_set_dev_replace_src_devid(eb, ptr, (u64)-1);
 431         btrfs_set_dev_replace_cont_reading_from_srcdev_mode(eb, ptr,
 432                 dev_replace->cont_reading_from_srcdev_mode);
 433         btrfs_set_dev_replace_replace_state(eb, ptr,
 434                 dev_replace->replace_state);
 435         btrfs_set_dev_replace_time_started(eb, ptr, dev_replace->time_started);
 436         btrfs_set_dev_replace_time_stopped(eb, ptr, dev_replace->time_stopped);
 437         btrfs_set_dev_replace_num_write_errors(eb, ptr,
 438                 atomic64_read(&dev_replace->num_write_errors));
 439         btrfs_set_dev_replace_num_uncorrectable_read_errors(eb, ptr,
 440                 atomic64_read(&dev_replace->num_uncorrectable_read_errors));
 441         dev_replace->cursor_left_last_write_of_item =
 442                 dev_replace->cursor_left;
 443         btrfs_set_dev_replace_cursor_left(eb, ptr,
 444                 dev_replace->cursor_left_last_write_of_item);
 445         btrfs_set_dev_replace_cursor_right(eb, ptr,
 446                 dev_replace->cursor_right);
 447         dev_replace->item_needs_writeback = 0;
 448         up_write(&dev_replace->rwsem);
 449
 450         btrfs_mark_buffer_dirty(eb);
 451
 452 out:
 453         btrfs_free_path(path);
 454
 455         return ret;
 456 }
 457
 458 static char* btrfs_dev_name(struct btrfs_device *device)
 459 {
 460         if (!device || test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
 461                 return "<missing disk>";
 462         else
 463                 return rcu_str_deref(device->name);
 464 }
 465
 466 static int mark_block_group_to_copy(struct btrfs_fs_info *fs_info,
 467                                     struct btrfs_device *src_dev)
 468 {
 469         struct btrfs_path *path;
 470         struct btrfs_key key;
 471         struct btrfs_key found_key;
 472         struct btrfs_root *root = fs_info->dev_root;
 473         struct btrfs_dev_extent *dev_extent = NULL;
 474         struct btrfs_block_group *cache;
 475         struct btrfs_trans_handle *trans;
 476         int ret = 0;
 477         u64 chunk_offset;
 478
 479         /* Do not use "to_copy" on non zoned filesystem for now */
 480         if (!btrfs_is_zoned(fs_info))
 481                 return 0;
 482
 483         mutex_lock(&fs_info->chunk_mutex);
 484
 485         /* Ensure we don't have pending new block group */
 486         spin_lock(&fs_info->trans_lock);
 487         while (fs_info->running_transaction &&
 488                !list_empty(&fs_info->running_transaction->dev_update_list)) {
 489                 spin_unlock(&fs_info->trans_lock);
 490                 mutex_unlock(&fs_info->chunk_mutex);
 491                 trans = btrfs_attach_transaction(root);
 492                 if (IS_ERR(trans)) {
 493                         ret = PTR_ERR(trans);
 494                         mutex_lock(&fs_info->chunk_mutex);
 495                         if (ret == -ENOENT) {
 496                                 spin_lock(&fs_info->trans_lock);
 497                                 continue;
 498                         } else {
 499                                 goto unlock;
 500                         }
 501                 }
 502
 503                 ret = btrfs_commit_transaction(trans);
 504                 mutex_lock(&fs_info->chunk_mutex);
 505                 if (ret)
 506                         goto unlock;
 507
 508                 spin_lock(&fs_info->trans_lock);
 509         }
 510         spin_unlock(&fs_info->trans_lock);
 511
 512         path = btrfs_alloc_path();
 513         if (!path) {
 514                 ret = -ENOMEM;
 515                 goto unlock;
 516         }
 517
 518         path->reada = READA_FORWARD;
 519         path->search_commit_root = 1;
 520         path->skip_locking = 1;
 521
 522         key.objectid = src_dev->devid;
 523         key.type = BTRFS_DEV_EXTENT_KEY;
 524         key.offset = 0;
 525
 526         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 527         if (ret < 0)
 528                 goto free_path;
 529         if (ret > 0) {
 530                 if (path->slots[0] >=
 531                     btrfs_header_nritems(path->nodes[0])) {
 532                         ret = btrfs_next_leaf(root, path);
 533                         if (ret < 0)
 534                                 goto free_path;
 535                         if (ret > 0) {
 536                                 ret = 0;
 537                                 goto free_path;
 538                         }
 539                 } else {
 540                         ret = 0;
 541                 }
 542         }
 543
 544         while (1) {
 545                 struct extent_buffer *leaf = path->nodes[0];
 546                 int slot = path->slots[0];
 547
 548                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
 549
 550                 if (found_key.objectid != src_dev->devid)
 551                         break;
 552
 553                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
 554                         break;
 555
 556                 if (found_key.offset < key.offset)
 557                         break;
 558
 559                 dev_extent = btrfs_item_ptr(leaf, slot, struct btrfs_dev_extent);
 560
 561                 chunk_offset = btrfs_dev_extent_chunk_offset(leaf, dev_extent);
 562
 563                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
 564                 if (!cache)
 565                         goto skip;
 566
 567                 spin_lock(&cache->lock);
 568                 cache->to_copy = 1;
 569                 spin_unlock(&cache->lock);
 570
 571                 btrfs_put_block_group(cache);
 572
 573 skip:
 574                 ret = btrfs_next_item(root, path);
 575                 if (ret != 0) {
 576                         if (ret > 0)
 577                                 ret = 0;
 578                         break;
 579                 }
 580         }
 581
 582 free_path:
 583         btrfs_free_path(path);
 584 unlock:
 585         mutex_unlock(&fs_info->chunk_mutex);
 586
 587         return ret;
 588 }
 589
 590 bool btrfs_finish_block_group_to_copy(struct btrfs_device *srcdev,
 591                                       struct btrfs_block_group *cache,
 592                                       u64 physical)
 593 {
 594         struct btrfs_fs_info *fs_info = cache->fs_info;
 595         struct extent_map *em;
 596         struct map_lookup *map;
 597         u64 chunk_offset = cache->start;
 598         int num_extents, cur_extent;
 599         int i;
 600
 601         /* Do not use "to_copy" on non zoned filesystem for now */
 602         if (!btrfs_is_zoned(fs_info))
 603                 return true;
 604
 605         spin_lock(&cache->lock);
 606         if (cache->removed) {
 607                 spin_unlock(&cache->lock);
 608                 return true;
 609         }
 610         spin_unlock(&cache->lock);
 611
 612         em = btrfs_get_chunk_map(fs_info, chunk_offset, 1);
 613         ASSERT(!IS_ERR(em));
 614         map = em->map_lookup;
 615
 616         num_extents = cur_extent = 0;
 617         for (i = 0; i < map->num_stripes; i++) {
 618                 /* We have more device extent to copy */
 619                 if (srcdev != map->stripes[i].dev)
 620                         continue;
 621
 622                 num_extents++;
 623                 if (physical == map->stripes[i].physical)
 624                         cur_extent = i;
 625         }
 626
 627         free_extent_map(em);
 628
 629         if (num_extents > 1 && cur_extent < num_extents - 1) {
 630                 /*
 631                  * Has more stripes on this device. Keep this block group
 632                  * readonly until we finish all the stripes.
 633                  */
 634                 return false;
 635         }
 636
 637         /* Last stripe on this device */
 638         spin_lock(&cache->lock);
 639         cache->to_copy = 0;
 640         spin_unlock(&cache->lock);
 641
 642         return true;
 643 }
 644
 645 static int btrfs_dev_replace_start(struct btrfs_fs_info *fs_info,
 646                 const char *tgtdev_name, u64 srcdevid, const char *srcdev_name,
 647                 int read_src)
 648 {
 649         struct btrfs_root *root = fs_info->dev_root;
 650         struct btrfs_trans_handle *trans;
 651         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 652         int ret;
 653         struct btrfs_device *tgt_device = NULL;
 654         struct btrfs_device *src_device = NULL;
 655
 656         src_device = btrfs_find_device_by_devspec(fs_info, srcdevid,
 657                                                   srcdev_name);
 658         if (IS_ERR(src_device))
 659                 return PTR_ERR(src_device);
 660
 661         if (btrfs_pinned_by_swapfile(fs_info, src_device)) {
 662                 btrfs_warn_in_rcu(fs_info,
 663           "cannot replace device %s (devid %llu) due to active swapfile",
 664                         btrfs_dev_name(src_device), src_device->devid);
 665                 return -ETXTBSY;
 666         }
 667
 668         /*
 669          * Here we commit the transaction to make sure commit_total_bytes
 670          * of all the devices are updated.
 671          */
 672         trans = btrfs_attach_transaction(root);
 673         if (!IS_ERR(trans)) {
 674                 ret = btrfs_commit_transaction(trans);
 675                 if (ret)
 676                         return ret;
 677         } else if (PTR_ERR(trans) != -ENOENT) {
 678                 return PTR_ERR(trans);
 679         }
 680
 681         ret = btrfs_init_dev_replace_tgtdev(fs_info, tgtdev_name,
 682                                             src_device, &tgt_device);
 683         if (ret)
 684                 return ret;
 685
 686         ret = mark_block_group_to_copy(fs_info, src_device);
 687         if (ret)
 688                 return ret;
 689
 690         down_write(&dev_replace->rwsem);
 691         switch (dev_replace->replace_state) {
 692         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
 693         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
 694         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
 695                 break;
 696         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
 697         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
 698                 ASSERT(0);
 699                 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_ALREADY_STARTED;
 700                 up_write(&dev_replace->rwsem);
 701                 goto leave;
 702         }
 703
 704         dev_replace->cont_reading_from_srcdev_mode = read_src;
 705         dev_replace->srcdev = src_device;
 706         dev_replace->tgtdev = tgt_device;
 707
 708         btrfs_info_in_rcu(fs_info,
 709                       "dev_replace from %s (devid %llu) to %s started",
 710                       btrfs_dev_name(src_device),
 711                       src_device->devid,
 712                       rcu_str_deref(tgt_device->name));
 713
 714         /*
 715          * from now on, the writes to the srcdev are all duplicated to
 716          * go to the tgtdev as well (refer to btrfs_map_block()).
 717          */
 718         dev_replace->replace_state = BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
 719         dev_replace->time_started = ktime_get_real_seconds();
 720         dev_replace->cursor_left = 0;
 721         dev_replace->committed_cursor_left = 0;
 722         dev_replace->cursor_left_last_write_of_item = 0;
 723         dev_replace->cursor_right = 0;
 724         dev_replace->is_valid = 1;
 725         dev_replace->item_needs_writeback = 1;
 726         atomic64_set(&dev_replace->num_write_errors, 0);
 727         atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
 728         up_write(&dev_replace->rwsem);
 729
 730         ret = btrfs_sysfs_add_device(tgt_device);
 731         if (ret)
 732                 btrfs_err(fs_info, "kobj add dev failed %d", ret);
 733
 734         btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
 735
 736         /* Commit dev_replace state and reserve 1 item for it. */
 737         trans = btrfs_start_transaction(root, 1);
 738         if (IS_ERR(trans)) {
 739                 ret = PTR_ERR(trans);
 740                 down_write(&dev_replace->rwsem);
 741                 dev_replace->replace_state =
 742                         BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED;
 743                 dev_replace->srcdev = NULL;
 744                 dev_replace->tgtdev = NULL;
 745                 up_write(&dev_replace->rwsem);
 746                 goto leave;
 747         }
 748
 749         ret = btrfs_commit_transaction(trans);
 750         WARN_ON(ret);
 751
 752         /* the disk copy procedure reuses the scrub code */
 753         ret = btrfs_scrub_dev(fs_info, src_device->devid, 0,
 754                               btrfs_device_get_total_bytes(src_device),
 755                               &dev_replace->scrub_progress, 0, 1);
 756
 757         ret = btrfs_dev_replace_finishing(fs_info, ret);
 758         if (ret == -EINPROGRESS)
 759                 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS;
 760
 761         return ret;
 762
 763 leave:
 764         btrfs_destroy_dev_replace_tgtdev(tgt_device);
 765         return ret;
 766 }
 767
 768 int btrfs_dev_replace_by_ioctl(struct btrfs_fs_info *fs_info,
 769                             struct btrfs_ioctl_dev_replace_args *args)
 770 {
 771         int ret;
 772
 773         switch (args->start.cont_reading_from_srcdev_mode) {
 774         case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_ALWAYS:
 775         case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_AVOID:
 776                 break;
 777         default:
 778                 return -EINVAL;
 779         }
 780
 781         if ((args->start.srcdevid == 0 && args->start.srcdev_name[0] == '\0') ||
 782             args->start.tgtdev_name[0] == '\0')
 783                 return -EINVAL;
 784
 785         ret = btrfs_dev_replace_start(fs_info, args->start.tgtdev_name,
 786                                         args->start.srcdevid,
 787                                         args->start.srcdev_name,
 788                                         args->start.cont_reading_from_srcdev_mode);
 789         args->result = ret;
 790         /* don't warn if EINPROGRESS, someone else might be running scrub */
 791         if (ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS ||
 792             ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR)
 793                 return 0;
 794
 795         return ret;
 796 }
 797
 798 /*
 799  * blocked until all in-flight bios operations are finished.
 800  */
 801 static void btrfs_rm_dev_replace_blocked(struct btrfs_fs_info *fs_info)
 802 {
 803         set_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
 804         wait_event(fs_info->dev_replace.replace_wait, !percpu_counter_sum(
 805                    &fs_info->dev_replace.bio_counter));
 806 }
 807
 808 /*
 809  * we have removed target device, it is safe to allow new bios request.
 810  */
 811 static void btrfs_rm_dev_replace_unblocked(struct btrfs_fs_info *fs_info)
 812 {
 813         clear_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
 814         wake_up(&fs_info->dev_replace.replace_wait);
 815 }
 816
 817 /*
 818  * When finishing the device replace, before swapping the source device with the
 819  * target device we must update the chunk allocation state in the target device,
 820  * as it is empty because replace works by directly copying the chunks and not
 821  * through the normal chunk allocation path.
 822  */
 823 static int btrfs_set_target_alloc_state(struct btrfs_device *srcdev,
 824                                         struct btrfs_device *tgtdev)
 825 {
 826         struct extent_state *cached_state = NULL;
 827         u64 start = 0;
 828         u64 found_start;
 829         u64 found_end;
 830         int ret = 0;
 831
 832         lockdep_assert_held(&srcdev->fs_info->chunk_mutex);
 833
 834         while (!find_first_extent_bit(&srcdev->alloc_state, start,
 835                                       &found_start, &found_end,
 836                                       CHUNK_ALLOCATED, &cached_state)) {
 837                 ret = set_extent_bits(&tgtdev->alloc_state, found_start,
 838                                       found_end, CHUNK_ALLOCATED);
 839                 if (ret)
 840                         break;
 841                 start = found_end + 1;
 842         }
 843
 844         free_extent_state(cached_state);
 845         return ret;
 846 }
 847
 848 static void btrfs_dev_replace_update_device_in_mapping_tree(
 849                                                 struct btrfs_fs_info *fs_info,
 850                                                 struct btrfs_device *srcdev,
 851                                                 struct btrfs_device *tgtdev)
 852 {
 853         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
 854         struct extent_map *em;
 855         struct map_lookup *map;
 856         u64 start = 0;
 857         int i;
 858
 859         write_lock(&em_tree->lock);
 860         do {
 861                 em = lookup_extent_mapping(em_tree, start, (u64)-1);
 862                 if (!em)
 863                         break;
 864                 map = em->map_lookup;
 865                 for (i = 0; i < map->num_stripes; i++)
 866                         if (srcdev == map->stripes[i].dev)
 867                                 map->stripes[i].dev = tgtdev;
 868                 start = em->start + em->len;
 869                 free_extent_map(em);
 870         } while (start);
 871         write_unlock(&em_tree->lock);
 872 }
 873
 874 static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
 875                                        int scrub_ret)
 876 {
 877         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
 878         struct btrfs_device *tgt_device;
 879         struct btrfs_device *src_device;
 880         struct btrfs_root *root = fs_info->tree_root;
 881         u8 uuid_tmp[BTRFS_UUID_SIZE];
 882         struct btrfs_trans_handle *trans;
 883         int ret = 0;
 884
 885         /* don't allow cancel or unmount to disturb the finishing procedure */
 886         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
 887
 888         down_read(&dev_replace->rwsem);
 889         /* was the operation canceled, or is it finished? */
 890         if (dev_replace->replace_state !=
 891             BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED) {
 892                 up_read(&dev_replace->rwsem);
 893                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 894                 return 0;
 895         }
 896
 897         tgt_device = dev_replace->tgtdev;
 898         src_device = dev_replace->srcdev;
 899         up_read(&dev_replace->rwsem);
 900
 901         /*
 902          * flush all outstanding I/O and inode extent mappings before the
 903          * copy operation is declared as being finished
 904          */
 905         ret = btrfs_start_delalloc_roots(fs_info, LONG_MAX, false);
 906         if (ret) {
 907                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 908                 return ret;
 909         }
 910         btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
 911
 912         if (!scrub_ret)
 913                 btrfs_reada_remove_dev(src_device);
 914
 915         /*
 916          * We have to use this loop approach because at this point src_device
 917          * has to be available for transaction commit to complete, yet new
 918          * chunks shouldn't be allocated on the device.
 919          */
 920         while (1) {
 921                 trans = btrfs_start_transaction(root, 0);
 922                 if (IS_ERR(trans)) {
 923                         btrfs_reada_undo_remove_dev(src_device);
 924                         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 925                         return PTR_ERR(trans);
 926                 }
 927                 ret = btrfs_commit_transaction(trans);
 928                 WARN_ON(ret);
 929
 930                 /* Prevent write_all_supers() during the finishing procedure */
 931                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
 932                 /* Prevent new chunks being allocated on the source device */
 933                 mutex_lock(&fs_info->chunk_mutex);
 934
 935                 if (!list_empty(&src_device->post_commit_list)) {
 936                         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
 937                         mutex_unlock(&fs_info->chunk_mutex);
 938                 } else {
 939                         break;
 940                 }
 941         }
 942
 943         down_write(&dev_replace->rwsem);
 944         dev_replace->replace_state =
 945                 scrub_ret ? BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED
 946                           : BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED;
 947         dev_replace->tgtdev = NULL;
 948         dev_replace->srcdev = NULL;
 949         dev_replace->time_stopped = ktime_get_real_seconds();
 950         dev_replace->item_needs_writeback = 1;
 951
 952         /*
 953          * Update allocation state in the new device and replace the old device
 954          * with the new one in the mapping tree.
 955          */
 956         if (!scrub_ret) {
 957                 scrub_ret = btrfs_set_target_alloc_state(src_device, tgt_device);
 958                 if (scrub_ret)
 959                         goto error;
 960                 btrfs_dev_replace_update_device_in_mapping_tree(fs_info,
 961                                                                 src_device,
 962                                                                 tgt_device);
 963         } else {
 964                 if (scrub_ret != -ECANCELED)
 965                         btrfs_err_in_rcu(fs_info,
 966                                  "btrfs_scrub_dev(%s, %llu, %s) failed %d",
 967                                  btrfs_dev_name(src_device),
 968                                  src_device->devid,
 969                                  rcu_str_deref(tgt_device->name), scrub_ret);
 970 error:
 971                 up_write(&dev_replace->rwsem);
 972                 mutex_unlock(&fs_info->chunk_mutex);
 973                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
 974                 btrfs_reada_undo_remove_dev(src_device);
 975                 btrfs_rm_dev_replace_blocked(fs_info);
 976                 if (tgt_device)
 977                         btrfs_destroy_dev_replace_tgtdev(tgt_device);
 978                 btrfs_rm_dev_replace_unblocked(fs_info);
 979                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
 980
 981                 return scrub_ret;
 982         }
 983
 984         btrfs_info_in_rcu(fs_info,
 985                           "dev_replace from %s (devid %llu) to %s finished",
 986                           btrfs_dev_name(src_device),
 987                           src_device->devid,
 988                           rcu_str_deref(tgt_device->name));
 989         clear_bit(BTRFS_DEV_STATE_REPLACE_TGT, &tgt_device->dev_state);
 990         tgt_device->devid = src_device->devid;
 991         src_device->devid = BTRFS_DEV_REPLACE_DEVID;
 992         memcpy(uuid_tmp, tgt_device->uuid, sizeof(uuid_tmp));
 993         memcpy(tgt_device->uuid, src_device->uuid, sizeof(tgt_device->uuid));
 994         memcpy(src_device->uuid, uuid_tmp, sizeof(src_device->uuid));
 995         btrfs_device_set_total_bytes(tgt_device, src_device->total_bytes);
 996         btrfs_device_set_disk_total_bytes(tgt_device,
 997                                           src_device->disk_total_bytes);
 998         btrfs_device_set_bytes_used(tgt_device, src_device->bytes_used);
 999         tgt_device->commit_bytes_used = src_device->bytes_used;
1000
1001         btrfs_assign_next_active_device(src_device, tgt_device);
1002
1003         list_add(&tgt_device->dev_alloc_list, &fs_info->fs_devices->alloc_list);
1004         fs_info->fs_devices->rw_devices++;
1005
1006         up_write(&dev_replace->rwsem);
1007         btrfs_rm_dev_replace_blocked(fs_info);
1008
1009         btrfs_rm_dev_replace_remove_srcdev(src_device);
1010
1011         btrfs_rm_dev_replace_unblocked(fs_info);
1012
1013         /*
1014          * Increment dev_stats_ccnt so that btrfs_run_dev_stats() will
1015          * update on-disk dev stats value during commit transaction
1016          */
1017         atomic_inc(&tgt_device->dev_stats_ccnt);
1018
1019         /*
1020          * this is again a consistent state where no dev_replace procedure
1021          * is running, the target device is part of the filesystem, the
1022          * source device is not part of the filesystem anymore and its 1st
1023          * superblock is scratched out so that it is no longer marked to
1024          * belong to this filesystem.
1025          */
1026         mutex_unlock(&fs_info->chunk_mutex);
1027         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1028
1029         /* replace the sysfs entry */
1030         btrfs_sysfs_remove_device(src_device);
1031         btrfs_sysfs_update_devid(tgt_device);
1032         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &src_device->dev_state))
1033                 btrfs_scratch_superblocks(fs_info, src_device->bdev,
1034                                           src_device->name->str);
1035
1036         /* write back the superblocks */
1037         trans = btrfs_start_transaction(root, 0);
1038         if (!IS_ERR(trans))
1039                 btrfs_commit_transaction(trans);
1040
1041         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1042
1043         btrfs_rm_dev_replace_free_srcdev(src_device);
1044
1045         return 0;
1046 }
1047
1048 /*
1049  * Read progress of device replace status according to the state and last
1050  * stored position. The value format is the same as for
1051  * btrfs_dev_replace::progress_1000
1052  */
1053 static u64 btrfs_dev_replace_progress(struct btrfs_fs_info *fs_info)
1054 {
1055         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1056         u64 ret = 0;
1057
1058         switch (dev_replace->replace_state) {
1059         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1060         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1061                 ret = 0;
1062                 break;
1063         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1064                 ret = 1000;
1065                 break;
1066         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1067         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1068                 ret = div64_u64(dev_replace->cursor_left,
1069                                 div_u64(btrfs_device_get_total_bytes(
1070                                                 dev_replace->srcdev), 1000));
1071                 break;
1072         }
1073
1074         return ret;
1075 }
1076
1077 void btrfs_dev_replace_status(struct btrfs_fs_info *fs_info,
1078                               struct btrfs_ioctl_dev_replace_args *args)
1079 {
1080         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1081
1082         down_read(&dev_replace->rwsem);
1083         /* even if !dev_replace_is_valid, the values are good enough for
1084          * the replace_status ioctl */
1085         args->result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1086         args->status.replace_state = dev_replace->replace_state;
1087         args->status.time_started = dev_replace->time_started;
1088         args->status.time_stopped = dev_replace->time_stopped;
1089         args->status.num_write_errors =
1090                 atomic64_read(&dev_replace->num_write_errors);
1091         args->status.num_uncorrectable_read_errors =
1092                 atomic64_read(&dev_replace->num_uncorrectable_read_errors);
1093         args->status.progress_1000 = btrfs_dev_replace_progress(fs_info);
1094         up_read(&dev_replace->rwsem);
1095 }
1096
1097 int btrfs_dev_replace_cancel(struct btrfs_fs_info *fs_info)
1098 {
1099         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1100         struct btrfs_device *tgt_device = NULL;
1101         struct btrfs_device *src_device = NULL;
1102         struct btrfs_trans_handle *trans;
1103         struct btrfs_root *root = fs_info->tree_root;
1104         int result;
1105         int ret;
1106
1107         if (sb_rdonly(fs_info->sb))
1108                 return -EROFS;
1109
1110         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1111         down_write(&dev_replace->rwsem);
1112         switch (dev_replace->replace_state) {
1113         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1114         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1115         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1116                 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1117                 up_write(&dev_replace->rwsem);
1118                 break;
1119         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1120                 tgt_device = dev_replace->tgtdev;
1121                 src_device = dev_replace->srcdev;
1122                 up_write(&dev_replace->rwsem);
1123                 ret = btrfs_scrub_cancel(fs_info);
1124                 if (ret < 0) {
1125                         result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1126                 } else {
1127                         result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1128                         /*
1129                          * btrfs_dev_replace_finishing() will handle the
1130                          * cleanup part
1131                          */
1132                         btrfs_info_in_rcu(fs_info,
1133                                 "dev_replace from %s (devid %llu) to %s canceled",
1134                                 btrfs_dev_name(src_device), src_device->devid,
1135                                 btrfs_dev_name(tgt_device));
1136                 }
1137                 break;
1138         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1139                 /*
1140                  * Scrub doing the replace isn't running so we need to do the
1141                  * cleanup step of btrfs_dev_replace_finishing() here
1142                  */
1143                 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1144                 tgt_device = dev_replace->tgtdev;
1145                 src_device = dev_replace->srcdev;
1146                 dev_replace->tgtdev = NULL;
1147                 dev_replace->srcdev = NULL;
1148                 dev_replace->replace_state =
1149                                 BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED;
1150                 dev_replace->time_stopped = ktime_get_real_seconds();
1151                 dev_replace->item_needs_writeback = 1;
1152
1153                 up_write(&dev_replace->rwsem);
1154
1155                 /* Scrub for replace must not be running in suspended state */
1156                 ret = btrfs_scrub_cancel(fs_info);
1157                 ASSERT(ret != -ENOTCONN);
1158
1159                 trans = btrfs_start_transaction(root, 0);
1160                 if (IS_ERR(trans)) {
1161                         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1162                         return PTR_ERR(trans);
1163                 }
1164                 ret = btrfs_commit_transaction(trans);
1165                 WARN_ON(ret);
1166
1167                 btrfs_info_in_rcu(fs_info,
1168                 "suspended dev_replace from %s (devid %llu) to %s canceled",
1169                         btrfs_dev_name(src_device), src_device->devid,
1170                         btrfs_dev_name(tgt_device));
1171
1172                 if (tgt_device)
1173                         btrfs_destroy_dev_replace_tgtdev(tgt_device);
1174                 break;
1175         default:
1176                 up_write(&dev_replace->rwsem);
1177                 result = -EINVAL;
1178         }
1179
1180         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1181         return result;
1182 }
1183
1184 void btrfs_dev_replace_suspend_for_unmount(struct btrfs_fs_info *fs_info)
1185 {
1186         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1187
1188         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1189         down_write(&dev_replace->rwsem);
1190
1191         switch (dev_replace->replace_state) {
1192         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1193         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1194         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1195         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1196                 break;
1197         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1198                 dev_replace->replace_state =
1199                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1200                 dev_replace->time_stopped = ktime_get_real_seconds();
1201                 dev_replace->item_needs_writeback = 1;
1202                 btrfs_info(fs_info, "suspending dev_replace for unmount");
1203                 break;
1204         }
1205
1206         up_write(&dev_replace->rwsem);
1207         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1208 }
1209
1210 /* resume dev_replace procedure that was interrupted by unmount */
1211 int btrfs_resume_dev_replace_async(struct btrfs_fs_info *fs_info)
1212 {
1213         struct task_struct *task;
1214         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1215
1216         down_write(&dev_replace->rwsem);
1217
1218         switch (dev_replace->replace_state) {
1219         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1220         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1221         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1222                 up_write(&dev_replace->rwsem);
1223                 return 0;
1224         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1225                 break;
1226         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1227                 dev_replace->replace_state =
1228                         BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
1229                 break;
1230         }
1231         if (!dev_replace->tgtdev || !dev_replace->tgtdev->bdev) {
1232                 btrfs_info(fs_info,
1233                            "cannot continue dev_replace, tgtdev is missing");
1234                 btrfs_info(fs_info,
1235                            "you may cancel the operation after 'mount -o degraded'");
1236                 dev_replace->replace_state =
1237                                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1238                 up_write(&dev_replace->rwsem);
1239                 return 0;
1240         }
1241         up_write(&dev_replace->rwsem);
1242
1243         /*
1244          * This could collide with a paused balance, but the exclusive op logic
1245          * should never allow both to start and pause. We don't want to allow
1246          * dev-replace to start anyway.
1247          */
1248         if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_DEV_REPLACE)) {
1249                 down_write(&dev_replace->rwsem);
1250                 dev_replace->replace_state =
1251                                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1252                 up_write(&dev_replace->rwsem);
1253                 btrfs_info(fs_info,
1254                 "cannot resume dev-replace, other exclusive operation running");
1255                 return 0;
1256         }
1257
1258         task = kthread_run(btrfs_dev_replace_kthread, fs_info, "btrfs-devrepl");
1259         return PTR_ERR_OR_ZERO(task);
1260 }
1261
1262 static int btrfs_dev_replace_kthread(void *data)
1263 {
1264         struct btrfs_fs_info *fs_info = data;
1265         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1266         u64 progress;
1267         int ret;
1268
1269         progress = btrfs_dev_replace_progress(fs_info);
1270         progress = div_u64(progress, 10);
1271         btrfs_info_in_rcu(fs_info,
1272                 "continuing dev_replace from %s (devid %llu) to target %s @%u%%",
1273                 btrfs_dev_name(dev_replace->srcdev),
1274                 dev_replace->srcdev->devid,
1275                 btrfs_dev_name(dev_replace->tgtdev),
1276                 (unsigned int)progress);
1277
1278         ret = btrfs_scrub_dev(fs_info, dev_replace->srcdev->devid,
1279                               dev_replace->committed_cursor_left,
1280                               btrfs_device_get_total_bytes(dev_replace->srcdev),
1281                               &dev_replace->scrub_progress, 0, 1);
1282         ret = btrfs_dev_replace_finishing(fs_info, ret);
1283         WARN_ON(ret && ret != -ECANCELED);
1284
1285         btrfs_exclop_finish(fs_info);
1286         return 0;
1287 }
1288
1289 int __pure btrfs_dev_replace_is_ongoing(struct btrfs_dev_replace *dev_replace)
1290 {
1291         if (!dev_replace->is_valid)
1292                 return 0;
1293
1294         switch (dev_replace->replace_state) {
1295         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1296         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1297         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1298                 return 0;
1299         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1300         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1301                 /*
1302                  * return true even if tgtdev is missing (this is
1303                  * something that can happen if the dev_replace
1304                  * procedure is suspended by an umount and then
1305                  * the tgtdev is missing (or "btrfs dev scan") was
1306                  * not called and the filesystem is remounted
1307                  * in degraded state. This does not stop the
1308                  * dev_replace procedure. It needs to be canceled
1309                  * manually if the cancellation is wanted.
1310                  */
1311                 break;
1312         }
1313         return 1;
1314 }
1315
1316 void btrfs_bio_counter_inc_noblocked(struct btrfs_fs_info *fs_info)
1317 {
1318         percpu_counter_inc(&fs_info->dev_replace.bio_counter);
1319 }
1320
1321 void btrfs_bio_counter_sub(struct btrfs_fs_info *fs_info, s64 amount)
1322 {
1323         percpu_counter_sub(&fs_info->dev_replace.bio_counter, amount);
1324         cond_wake_up_nomb(&fs_info->dev_replace.replace_wait);
1325 }
1326
1327 void btrfs_bio_counter_inc_blocked(struct btrfs_fs_info *fs_info)
1328 {
1329         while (1) {
1330                 percpu_counter_inc(&fs_info->dev_replace.bio_counter);
1331                 if (likely(!test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1332                                      &fs_info->fs_state)))
1333                         break;
1334
1335                 btrfs_bio_counter_dec(fs_info);
1336                 wait_event(fs_info->dev_replace.replace_wait,
1337                            !test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1338                                      &fs_info->fs_state));
1339         }
1340 }