Merge tag 'i3c/for-6.6' of git://git.kernel.org/pub/scm/linux/kernel/git/i3c/linux
[platform/kernel/linux-rpi.git] / 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 "dev-replace.h"
22 #include "sysfs.h"
23 #include "zoned.h"
24 #include "block-group.h"
25 #include "fs.h"
26 #include "accessors.h"
27 #include "scrub.h"
28
29 /*
30  * Device replace overview
31  *
32  * [Objective]
33  * To copy all extents (both new and on-disk) from source device to target
34  * device, while still keeping the filesystem read-write.
35  *
36  * [Method]
37  * There are two main methods involved:
38  *
39  * - Write duplication
40  *
41  *   All new writes will be written to both target and source devices, so even
42  *   if replace gets canceled, sources device still contains up-to-date data.
43  *
44  *   Location:          handle_ops_on_dev_replace() from btrfs_map_block()
45  *   Start:             btrfs_dev_replace_start()
46  *   End:               btrfs_dev_replace_finishing()
47  *   Content:           Latest data/metadata
48  *
49  * - Copy existing extents
50  *
51  *   This happens by re-using scrub facility, as scrub also iterates through
52  *   existing extents from commit root.
53  *
54  *   Location:          scrub_write_block_to_dev_replace() from
55  *                      scrub_block_complete()
56  *   Content:           Data/meta from commit root.
57  *
58  * Due to the content difference, we need to avoid nocow write when dev-replace
59  * is happening.  This is done by marking the block group read-only and waiting
60  * for NOCOW writes.
61  *
62  * After replace is done, the finishing part is done by swapping the target and
63  * source devices.
64  *
65  *   Location:          btrfs_dev_replace_update_device_in_mapping_tree() from
66  *                      btrfs_dev_replace_finishing()
67  */
68
69 static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
70                                        int scrub_ret);
71 static int btrfs_dev_replace_kthread(void *data);
72
73 int btrfs_init_dev_replace(struct btrfs_fs_info *fs_info)
74 {
75         struct btrfs_dev_lookup_args args = { .devid = BTRFS_DEV_REPLACE_DEVID };
76         struct btrfs_key key;
77         struct btrfs_root *dev_root = fs_info->dev_root;
78         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
79         struct extent_buffer *eb;
80         int slot;
81         int ret = 0;
82         struct btrfs_path *path = NULL;
83         int item_size;
84         struct btrfs_dev_replace_item *ptr;
85         u64 src_devid;
86
87         if (!dev_root)
88                 return 0;
89
90         path = btrfs_alloc_path();
91         if (!path) {
92                 ret = -ENOMEM;
93                 goto out;
94         }
95
96         key.objectid = 0;
97         key.type = BTRFS_DEV_REPLACE_KEY;
98         key.offset = 0;
99         ret = btrfs_search_slot(NULL, dev_root, &key, path, 0, 0);
100         if (ret) {
101 no_valid_dev_replace_entry_found:
102                 /*
103                  * We don't have a replace item or it's corrupted.  If there is
104                  * a replace target, fail the mount.
105                  */
106                 if (btrfs_find_device(fs_info->fs_devices, &args)) {
107                         btrfs_err(fs_info,
108                         "found replace target device without a valid replace item");
109                         ret = -EUCLEAN;
110                         goto out;
111                 }
112                 ret = 0;
113                 dev_replace->replace_state =
114                         BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED;
115                 dev_replace->cont_reading_from_srcdev_mode =
116                     BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_ALWAYS;
117                 dev_replace->time_started = 0;
118                 dev_replace->time_stopped = 0;
119                 atomic64_set(&dev_replace->num_write_errors, 0);
120                 atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
121                 dev_replace->cursor_left = 0;
122                 dev_replace->committed_cursor_left = 0;
123                 dev_replace->cursor_left_last_write_of_item = 0;
124                 dev_replace->cursor_right = 0;
125                 dev_replace->srcdev = NULL;
126                 dev_replace->tgtdev = NULL;
127                 dev_replace->is_valid = 0;
128                 dev_replace->item_needs_writeback = 0;
129                 goto out;
130         }
131         slot = path->slots[0];
132         eb = path->nodes[0];
133         item_size = btrfs_item_size(eb, slot);
134         ptr = btrfs_item_ptr(eb, slot, struct btrfs_dev_replace_item);
135
136         if (item_size != sizeof(struct btrfs_dev_replace_item)) {
137                 btrfs_warn(fs_info,
138                         "dev_replace entry found has unexpected size, ignore entry");
139                 goto no_valid_dev_replace_entry_found;
140         }
141
142         src_devid = btrfs_dev_replace_src_devid(eb, ptr);
143         dev_replace->cont_reading_from_srcdev_mode =
144                 btrfs_dev_replace_cont_reading_from_srcdev_mode(eb, ptr);
145         dev_replace->replace_state = btrfs_dev_replace_replace_state(eb, ptr);
146         dev_replace->time_started = btrfs_dev_replace_time_started(eb, ptr);
147         dev_replace->time_stopped =
148                 btrfs_dev_replace_time_stopped(eb, ptr);
149         atomic64_set(&dev_replace->num_write_errors,
150                      btrfs_dev_replace_num_write_errors(eb, ptr));
151         atomic64_set(&dev_replace->num_uncorrectable_read_errors,
152                      btrfs_dev_replace_num_uncorrectable_read_errors(eb, ptr));
153         dev_replace->cursor_left = btrfs_dev_replace_cursor_left(eb, ptr);
154         dev_replace->committed_cursor_left = dev_replace->cursor_left;
155         dev_replace->cursor_left_last_write_of_item = dev_replace->cursor_left;
156         dev_replace->cursor_right = btrfs_dev_replace_cursor_right(eb, ptr);
157         dev_replace->is_valid = 1;
158
159         dev_replace->item_needs_writeback = 0;
160         switch (dev_replace->replace_state) {
161         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
162         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
163         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
164                 /*
165                  * We don't have an active replace item but if there is a
166                  * replace target, fail the mount.
167                  */
168                 if (btrfs_find_device(fs_info->fs_devices, &args)) {
169                         btrfs_err(fs_info,
170 "replace without active item, run 'device scan --forget' on the target device");
171                         ret = -EUCLEAN;
172                 } else {
173                         dev_replace->srcdev = NULL;
174                         dev_replace->tgtdev = NULL;
175                 }
176                 break;
177         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
178         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
179                 dev_replace->tgtdev = btrfs_find_device(fs_info->fs_devices, &args);
180                 args.devid = src_devid;
181                 dev_replace->srcdev = btrfs_find_device(fs_info->fs_devices, &args);
182
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_fs_devices *fs_devices = fs_info->fs_devices;
249         struct btrfs_device *device;
250         struct block_device *bdev;
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, BLK_OPEN_WRITE,
261                                   fs_info->bdev_holder, NULL);
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_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 (bdev_nr_bytes(bdev) < btrfs_device_get_total_bytes(srcdev)) {
287                 btrfs_err(fs_info,
288                           "target device is smaller than source device!");
289                 ret = -EINVAL;
290                 goto error;
291         }
292
293
294         device = btrfs_alloc_device(NULL, &devid, NULL, device_path);
295         if (IS_ERR(device)) {
296                 ret = PTR_ERR(device);
297                 goto error;
298         }
299
300         ret = lookup_bdev(device_path, &device->devt);
301         if (ret)
302                 goto error;
303
304         set_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
305         device->generation = 0;
306         device->io_width = fs_info->sectorsize;
307         device->io_align = fs_info->sectorsize;
308         device->sector_size = fs_info->sectorsize;
309         device->total_bytes = btrfs_device_get_total_bytes(srcdev);
310         device->disk_total_bytes = btrfs_device_get_disk_total_bytes(srcdev);
311         device->bytes_used = btrfs_device_get_bytes_used(srcdev);
312         device->commit_total_bytes = srcdev->commit_total_bytes;
313         device->commit_bytes_used = device->bytes_used;
314         device->fs_info = fs_info;
315         device->bdev = bdev;
316         set_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
317         set_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state);
318         device->holder = fs_info->bdev_holder;
319         device->dev_stats_valid = 1;
320         set_blocksize(device->bdev, BTRFS_BDEV_BLOCKSIZE);
321         device->fs_devices = fs_devices;
322
323         ret = btrfs_get_dev_zone_info(device, false);
324         if (ret)
325                 goto error;
326
327         mutex_lock(&fs_devices->device_list_mutex);
328         list_add(&device->dev_list, &fs_devices->devices);
329         fs_devices->num_devices++;
330         fs_devices->open_devices++;
331         mutex_unlock(&fs_devices->device_list_mutex);
332
333         *device_out = device;
334         return 0;
335
336 error:
337         blkdev_put(bdev, fs_info->bdev_holder);
338         return ret;
339 }
340
341 /*
342  * called from commit_transaction. Writes changed device replace state to
343  * disk.
344  */
345 int btrfs_run_dev_replace(struct btrfs_trans_handle *trans)
346 {
347         struct btrfs_fs_info *fs_info = trans->fs_info;
348         int ret;
349         struct btrfs_root *dev_root = fs_info->dev_root;
350         struct btrfs_path *path;
351         struct btrfs_key key;
352         struct extent_buffer *eb;
353         struct btrfs_dev_replace_item *ptr;
354         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
355
356         down_read(&dev_replace->rwsem);
357         if (!dev_replace->is_valid ||
358             !dev_replace->item_needs_writeback) {
359                 up_read(&dev_replace->rwsem);
360                 return 0;
361         }
362         up_read(&dev_replace->rwsem);
363
364         key.objectid = 0;
365         key.type = BTRFS_DEV_REPLACE_KEY;
366         key.offset = 0;
367
368         path = btrfs_alloc_path();
369         if (!path) {
370                 ret = -ENOMEM;
371                 goto out;
372         }
373         ret = btrfs_search_slot(trans, dev_root, &key, path, -1, 1);
374         if (ret < 0) {
375                 btrfs_warn(fs_info,
376                            "error %d while searching for dev_replace item!",
377                            ret);
378                 goto out;
379         }
380
381         if (ret == 0 &&
382             btrfs_item_size(path->nodes[0], path->slots[0]) < sizeof(*ptr)) {
383                 /*
384                  * need to delete old one and insert a new one.
385                  * Since no attempt is made to recover any old state, if the
386                  * dev_replace state is 'running', the data on the target
387                  * drive is lost.
388                  * It would be possible to recover the state: just make sure
389                  * that the beginning of the item is never changed and always
390                  * contains all the essential information. Then read this
391                  * minimal set of information and use it as a base for the
392                  * new state.
393                  */
394                 ret = btrfs_del_item(trans, dev_root, path);
395                 if (ret != 0) {
396                         btrfs_warn(fs_info,
397                                    "delete too small dev_replace item failed %d!",
398                                    ret);
399                         goto out;
400                 }
401                 ret = 1;
402         }
403
404         if (ret == 1) {
405                 /* need to insert a new item */
406                 btrfs_release_path(path);
407                 ret = btrfs_insert_empty_item(trans, dev_root, path,
408                                               &key, sizeof(*ptr));
409                 if (ret < 0) {
410                         btrfs_warn(fs_info,
411                                    "insert dev_replace item failed %d!", ret);
412                         goto out;
413                 }
414         }
415
416         eb = path->nodes[0];
417         ptr = btrfs_item_ptr(eb, path->slots[0],
418                              struct btrfs_dev_replace_item);
419
420         down_write(&dev_replace->rwsem);
421         if (dev_replace->srcdev)
422                 btrfs_set_dev_replace_src_devid(eb, ptr,
423                         dev_replace->srcdev->devid);
424         else
425                 btrfs_set_dev_replace_src_devid(eb, ptr, (u64)-1);
426         btrfs_set_dev_replace_cont_reading_from_srcdev_mode(eb, ptr,
427                 dev_replace->cont_reading_from_srcdev_mode);
428         btrfs_set_dev_replace_replace_state(eb, ptr,
429                 dev_replace->replace_state);
430         btrfs_set_dev_replace_time_started(eb, ptr, dev_replace->time_started);
431         btrfs_set_dev_replace_time_stopped(eb, ptr, dev_replace->time_stopped);
432         btrfs_set_dev_replace_num_write_errors(eb, ptr,
433                 atomic64_read(&dev_replace->num_write_errors));
434         btrfs_set_dev_replace_num_uncorrectable_read_errors(eb, ptr,
435                 atomic64_read(&dev_replace->num_uncorrectable_read_errors));
436         dev_replace->cursor_left_last_write_of_item =
437                 dev_replace->cursor_left;
438         btrfs_set_dev_replace_cursor_left(eb, ptr,
439                 dev_replace->cursor_left_last_write_of_item);
440         btrfs_set_dev_replace_cursor_right(eb, ptr,
441                 dev_replace->cursor_right);
442         dev_replace->item_needs_writeback = 0;
443         up_write(&dev_replace->rwsem);
444
445         btrfs_mark_buffer_dirty(eb);
446
447 out:
448         btrfs_free_path(path);
449
450         return ret;
451 }
452
453 static int mark_block_group_to_copy(struct btrfs_fs_info *fs_info,
454                                     struct btrfs_device *src_dev)
455 {
456         struct btrfs_path *path;
457         struct btrfs_key key;
458         struct btrfs_key found_key;
459         struct btrfs_root *root = fs_info->dev_root;
460         struct btrfs_dev_extent *dev_extent = NULL;
461         struct btrfs_block_group *cache;
462         struct btrfs_trans_handle *trans;
463         int iter_ret = 0;
464         int ret = 0;
465         u64 chunk_offset;
466
467         /* Do not use "to_copy" on non zoned filesystem for now */
468         if (!btrfs_is_zoned(fs_info))
469                 return 0;
470
471         mutex_lock(&fs_info->chunk_mutex);
472
473         /* Ensure we don't have pending new block group */
474         spin_lock(&fs_info->trans_lock);
475         while (fs_info->running_transaction &&
476                !list_empty(&fs_info->running_transaction->dev_update_list)) {
477                 spin_unlock(&fs_info->trans_lock);
478                 mutex_unlock(&fs_info->chunk_mutex);
479                 trans = btrfs_attach_transaction(root);
480                 if (IS_ERR(trans)) {
481                         ret = PTR_ERR(trans);
482                         mutex_lock(&fs_info->chunk_mutex);
483                         if (ret == -ENOENT) {
484                                 spin_lock(&fs_info->trans_lock);
485                                 continue;
486                         } else {
487                                 goto unlock;
488                         }
489                 }
490
491                 ret = btrfs_commit_transaction(trans);
492                 mutex_lock(&fs_info->chunk_mutex);
493                 if (ret)
494                         goto unlock;
495
496                 spin_lock(&fs_info->trans_lock);
497         }
498         spin_unlock(&fs_info->trans_lock);
499
500         path = btrfs_alloc_path();
501         if (!path) {
502                 ret = -ENOMEM;
503                 goto unlock;
504         }
505
506         path->reada = READA_FORWARD;
507         path->search_commit_root = 1;
508         path->skip_locking = 1;
509
510         key.objectid = src_dev->devid;
511         key.type = BTRFS_DEV_EXTENT_KEY;
512         key.offset = 0;
513
514         btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
515                 struct extent_buffer *leaf = path->nodes[0];
516
517                 if (found_key.objectid != src_dev->devid)
518                         break;
519
520                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
521                         break;
522
523                 if (found_key.offset < key.offset)
524                         break;
525
526                 dev_extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
527
528                 chunk_offset = btrfs_dev_extent_chunk_offset(leaf, dev_extent);
529
530                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
531                 if (!cache)
532                         continue;
533
534                 set_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags);
535                 btrfs_put_block_group(cache);
536         }
537         if (iter_ret < 0)
538                 ret = iter_ret;
539
540         btrfs_free_path(path);
541 unlock:
542         mutex_unlock(&fs_info->chunk_mutex);
543
544         return ret;
545 }
546
547 bool btrfs_finish_block_group_to_copy(struct btrfs_device *srcdev,
548                                       struct btrfs_block_group *cache,
549                                       u64 physical)
550 {
551         struct btrfs_fs_info *fs_info = cache->fs_info;
552         struct extent_map *em;
553         struct map_lookup *map;
554         u64 chunk_offset = cache->start;
555         int num_extents, cur_extent;
556         int i;
557
558         /* Do not use "to_copy" on non zoned filesystem for now */
559         if (!btrfs_is_zoned(fs_info))
560                 return true;
561
562         spin_lock(&cache->lock);
563         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
564                 spin_unlock(&cache->lock);
565                 return true;
566         }
567         spin_unlock(&cache->lock);
568
569         em = btrfs_get_chunk_map(fs_info, chunk_offset, 1);
570         ASSERT(!IS_ERR(em));
571         map = em->map_lookup;
572
573         num_extents = 0;
574         cur_extent = 0;
575         for (i = 0; i < map->num_stripes; i++) {
576                 /* We have more device extent to copy */
577                 if (srcdev != map->stripes[i].dev)
578                         continue;
579
580                 num_extents++;
581                 if (physical == map->stripes[i].physical)
582                         cur_extent = i;
583         }
584
585         free_extent_map(em);
586
587         if (num_extents > 1 && cur_extent < num_extents - 1) {
588                 /*
589                  * Has more stripes on this device. Keep this block group
590                  * readonly until we finish all the stripes.
591                  */
592                 return false;
593         }
594
595         /* Last stripe on this device */
596         clear_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags);
597
598         return true;
599 }
600
601 static int btrfs_dev_replace_start(struct btrfs_fs_info *fs_info,
602                 const char *tgtdev_name, u64 srcdevid, const char *srcdev_name,
603                 int read_src)
604 {
605         struct btrfs_root *root = fs_info->dev_root;
606         struct btrfs_trans_handle *trans;
607         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
608         int ret;
609         struct btrfs_device *tgt_device = NULL;
610         struct btrfs_device *src_device = NULL;
611
612         src_device = btrfs_find_device_by_devspec(fs_info, srcdevid,
613                                                   srcdev_name);
614         if (IS_ERR(src_device))
615                 return PTR_ERR(src_device);
616
617         if (btrfs_pinned_by_swapfile(fs_info, src_device)) {
618                 btrfs_warn_in_rcu(fs_info,
619           "cannot replace device %s (devid %llu) due to active swapfile",
620                         btrfs_dev_name(src_device), src_device->devid);
621                 return -ETXTBSY;
622         }
623
624         /*
625          * Here we commit the transaction to make sure commit_total_bytes
626          * of all the devices are updated.
627          */
628         trans = btrfs_attach_transaction(root);
629         if (!IS_ERR(trans)) {
630                 ret = btrfs_commit_transaction(trans);
631                 if (ret)
632                         return ret;
633         } else if (PTR_ERR(trans) != -ENOENT) {
634                 return PTR_ERR(trans);
635         }
636
637         ret = btrfs_init_dev_replace_tgtdev(fs_info, tgtdev_name,
638                                             src_device, &tgt_device);
639         if (ret)
640                 return ret;
641
642         ret = mark_block_group_to_copy(fs_info, src_device);
643         if (ret)
644                 return ret;
645
646         down_write(&dev_replace->rwsem);
647         switch (dev_replace->replace_state) {
648         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
649         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
650         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
651                 break;
652         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
653         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
654                 ASSERT(0);
655                 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_ALREADY_STARTED;
656                 up_write(&dev_replace->rwsem);
657                 goto leave;
658         }
659
660         dev_replace->cont_reading_from_srcdev_mode = read_src;
661         dev_replace->srcdev = src_device;
662         dev_replace->tgtdev = tgt_device;
663
664         btrfs_info_in_rcu(fs_info,
665                       "dev_replace from %s (devid %llu) to %s started",
666                       btrfs_dev_name(src_device),
667                       src_device->devid,
668                       btrfs_dev_name(tgt_device));
669
670         /*
671          * from now on, the writes to the srcdev are all duplicated to
672          * go to the tgtdev as well (refer to btrfs_map_block()).
673          */
674         dev_replace->replace_state = BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
675         dev_replace->time_started = ktime_get_real_seconds();
676         dev_replace->cursor_left = 0;
677         dev_replace->committed_cursor_left = 0;
678         dev_replace->cursor_left_last_write_of_item = 0;
679         dev_replace->cursor_right = 0;
680         dev_replace->is_valid = 1;
681         dev_replace->item_needs_writeback = 1;
682         atomic64_set(&dev_replace->num_write_errors, 0);
683         atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
684         up_write(&dev_replace->rwsem);
685
686         ret = btrfs_sysfs_add_device(tgt_device);
687         if (ret)
688                 btrfs_err(fs_info, "kobj add dev failed %d", ret);
689
690         btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
691
692         /*
693          * Commit dev_replace state and reserve 1 item for it.
694          * This is crucial to ensure we won't miss copying extents for new block
695          * groups that are allocated after we started the device replace, and
696          * must be done after setting up the device replace state.
697          */
698         trans = btrfs_start_transaction(root, 1);
699         if (IS_ERR(trans)) {
700                 ret = PTR_ERR(trans);
701                 down_write(&dev_replace->rwsem);
702                 dev_replace->replace_state =
703                         BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED;
704                 dev_replace->srcdev = NULL;
705                 dev_replace->tgtdev = NULL;
706                 up_write(&dev_replace->rwsem);
707                 goto leave;
708         }
709
710         ret = btrfs_commit_transaction(trans);
711         WARN_ON(ret);
712
713         /* the disk copy procedure reuses the scrub code */
714         ret = btrfs_scrub_dev(fs_info, src_device->devid, 0,
715                               btrfs_device_get_total_bytes(src_device),
716                               &dev_replace->scrub_progress, 0, 1);
717
718         ret = btrfs_dev_replace_finishing(fs_info, ret);
719         if (ret == -EINPROGRESS)
720                 ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS;
721
722         return ret;
723
724 leave:
725         btrfs_destroy_dev_replace_tgtdev(tgt_device);
726         return ret;
727 }
728
729 int btrfs_dev_replace_by_ioctl(struct btrfs_fs_info *fs_info,
730                             struct btrfs_ioctl_dev_replace_args *args)
731 {
732         int ret;
733
734         switch (args->start.cont_reading_from_srcdev_mode) {
735         case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_ALWAYS:
736         case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_AVOID:
737                 break;
738         default:
739                 return -EINVAL;
740         }
741
742         if ((args->start.srcdevid == 0 && args->start.srcdev_name[0] == '\0') ||
743             args->start.tgtdev_name[0] == '\0')
744                 return -EINVAL;
745
746         ret = btrfs_dev_replace_start(fs_info, args->start.tgtdev_name,
747                                         args->start.srcdevid,
748                                         args->start.srcdev_name,
749                                         args->start.cont_reading_from_srcdev_mode);
750         args->result = ret;
751         /* don't warn if EINPROGRESS, someone else might be running scrub */
752         if (ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS ||
753             ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR)
754                 return 0;
755
756         return ret;
757 }
758
759 /*
760  * blocked until all in-flight bios operations are finished.
761  */
762 static void btrfs_rm_dev_replace_blocked(struct btrfs_fs_info *fs_info)
763 {
764         set_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
765         wait_event(fs_info->dev_replace.replace_wait, !percpu_counter_sum(
766                    &fs_info->dev_replace.bio_counter));
767 }
768
769 /*
770  * we have removed target device, it is safe to allow new bios request.
771  */
772 static void btrfs_rm_dev_replace_unblocked(struct btrfs_fs_info *fs_info)
773 {
774         clear_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
775         wake_up(&fs_info->dev_replace.replace_wait);
776 }
777
778 /*
779  * When finishing the device replace, before swapping the source device with the
780  * target device we must update the chunk allocation state in the target device,
781  * as it is empty because replace works by directly copying the chunks and not
782  * through the normal chunk allocation path.
783  */
784 static int btrfs_set_target_alloc_state(struct btrfs_device *srcdev,
785                                         struct btrfs_device *tgtdev)
786 {
787         struct extent_state *cached_state = NULL;
788         u64 start = 0;
789         u64 found_start;
790         u64 found_end;
791         int ret = 0;
792
793         lockdep_assert_held(&srcdev->fs_info->chunk_mutex);
794
795         while (find_first_extent_bit(&srcdev->alloc_state, start,
796                                      &found_start, &found_end,
797                                      CHUNK_ALLOCATED, &cached_state)) {
798                 ret = set_extent_bit(&tgtdev->alloc_state, found_start,
799                                      found_end, CHUNK_ALLOCATED, NULL);
800                 if (ret)
801                         break;
802                 start = found_end + 1;
803         }
804
805         free_extent_state(cached_state);
806         return ret;
807 }
808
809 static void btrfs_dev_replace_update_device_in_mapping_tree(
810                                                 struct btrfs_fs_info *fs_info,
811                                                 struct btrfs_device *srcdev,
812                                                 struct btrfs_device *tgtdev)
813 {
814         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
815         struct extent_map *em;
816         struct map_lookup *map;
817         u64 start = 0;
818         int i;
819
820         write_lock(&em_tree->lock);
821         do {
822                 em = lookup_extent_mapping(em_tree, start, (u64)-1);
823                 if (!em)
824                         break;
825                 map = em->map_lookup;
826                 for (i = 0; i < map->num_stripes; i++)
827                         if (srcdev == map->stripes[i].dev)
828                                 map->stripes[i].dev = tgtdev;
829                 start = em->start + em->len;
830                 free_extent_map(em);
831         } while (start);
832         write_unlock(&em_tree->lock);
833 }
834
835 static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
836                                        int scrub_ret)
837 {
838         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
839         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
840         struct btrfs_device *tgt_device;
841         struct btrfs_device *src_device;
842         struct btrfs_root *root = fs_info->tree_root;
843         u8 uuid_tmp[BTRFS_UUID_SIZE];
844         struct btrfs_trans_handle *trans;
845         int ret = 0;
846
847         /* don't allow cancel or unmount to disturb the finishing procedure */
848         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
849
850         down_read(&dev_replace->rwsem);
851         /* was the operation canceled, or is it finished? */
852         if (dev_replace->replace_state !=
853             BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED) {
854                 up_read(&dev_replace->rwsem);
855                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
856                 return 0;
857         }
858
859         tgt_device = dev_replace->tgtdev;
860         src_device = dev_replace->srcdev;
861         up_read(&dev_replace->rwsem);
862
863         /*
864          * flush all outstanding I/O and inode extent mappings before the
865          * copy operation is declared as being finished
866          */
867         ret = btrfs_start_delalloc_roots(fs_info, LONG_MAX, false);
868         if (ret) {
869                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
870                 return ret;
871         }
872         btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
873
874         /*
875          * We have to use this loop approach because at this point src_device
876          * has to be available for transaction commit to complete, yet new
877          * chunks shouldn't be allocated on the device.
878          */
879         while (1) {
880                 trans = btrfs_start_transaction(root, 0);
881                 if (IS_ERR(trans)) {
882                         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
883                         return PTR_ERR(trans);
884                 }
885                 ret = btrfs_commit_transaction(trans);
886                 WARN_ON(ret);
887
888                 /* Prevent write_all_supers() during the finishing procedure */
889                 mutex_lock(&fs_devices->device_list_mutex);
890                 /* Prevent new chunks being allocated on the source device */
891                 mutex_lock(&fs_info->chunk_mutex);
892
893                 if (!list_empty(&src_device->post_commit_list)) {
894                         mutex_unlock(&fs_devices->device_list_mutex);
895                         mutex_unlock(&fs_info->chunk_mutex);
896                 } else {
897                         break;
898                 }
899         }
900
901         down_write(&dev_replace->rwsem);
902         dev_replace->replace_state =
903                 scrub_ret ? BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED
904                           : BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED;
905         dev_replace->tgtdev = NULL;
906         dev_replace->srcdev = NULL;
907         dev_replace->time_stopped = ktime_get_real_seconds();
908         dev_replace->item_needs_writeback = 1;
909
910         /*
911          * Update allocation state in the new device and replace the old device
912          * with the new one in the mapping tree.
913          */
914         if (!scrub_ret) {
915                 scrub_ret = btrfs_set_target_alloc_state(src_device, tgt_device);
916                 if (scrub_ret)
917                         goto error;
918                 btrfs_dev_replace_update_device_in_mapping_tree(fs_info,
919                                                                 src_device,
920                                                                 tgt_device);
921         } else {
922                 if (scrub_ret != -ECANCELED)
923                         btrfs_err_in_rcu(fs_info,
924                                  "btrfs_scrub_dev(%s, %llu, %s) failed %d",
925                                  btrfs_dev_name(src_device),
926                                  src_device->devid,
927                                  btrfs_dev_name(tgt_device), scrub_ret);
928 error:
929                 up_write(&dev_replace->rwsem);
930                 mutex_unlock(&fs_info->chunk_mutex);
931                 mutex_unlock(&fs_devices->device_list_mutex);
932                 btrfs_rm_dev_replace_blocked(fs_info);
933                 if (tgt_device)
934                         btrfs_destroy_dev_replace_tgtdev(tgt_device);
935                 btrfs_rm_dev_replace_unblocked(fs_info);
936                 mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
937
938                 return scrub_ret;
939         }
940
941         btrfs_info_in_rcu(fs_info,
942                           "dev_replace from %s (devid %llu) to %s finished",
943                           btrfs_dev_name(src_device),
944                           src_device->devid,
945                           btrfs_dev_name(tgt_device));
946         clear_bit(BTRFS_DEV_STATE_REPLACE_TGT, &tgt_device->dev_state);
947         tgt_device->devid = src_device->devid;
948         src_device->devid = BTRFS_DEV_REPLACE_DEVID;
949         memcpy(uuid_tmp, tgt_device->uuid, sizeof(uuid_tmp));
950         memcpy(tgt_device->uuid, src_device->uuid, sizeof(tgt_device->uuid));
951         memcpy(src_device->uuid, uuid_tmp, sizeof(src_device->uuid));
952         btrfs_device_set_total_bytes(tgt_device, src_device->total_bytes);
953         btrfs_device_set_disk_total_bytes(tgt_device,
954                                           src_device->disk_total_bytes);
955         btrfs_device_set_bytes_used(tgt_device, src_device->bytes_used);
956         tgt_device->commit_bytes_used = src_device->bytes_used;
957
958         btrfs_assign_next_active_device(src_device, tgt_device);
959
960         list_add(&tgt_device->dev_alloc_list, &fs_devices->alloc_list);
961         fs_devices->rw_devices++;
962
963         up_write(&dev_replace->rwsem);
964         btrfs_rm_dev_replace_blocked(fs_info);
965
966         btrfs_rm_dev_replace_remove_srcdev(src_device);
967
968         btrfs_rm_dev_replace_unblocked(fs_info);
969
970         /*
971          * Increment dev_stats_ccnt so that btrfs_run_dev_stats() will
972          * update on-disk dev stats value during commit transaction
973          */
974         atomic_inc(&tgt_device->dev_stats_ccnt);
975
976         /*
977          * this is again a consistent state where no dev_replace procedure
978          * is running, the target device is part of the filesystem, the
979          * source device is not part of the filesystem anymore and its 1st
980          * superblock is scratched out so that it is no longer marked to
981          * belong to this filesystem.
982          */
983         mutex_unlock(&fs_info->chunk_mutex);
984         mutex_unlock(&fs_devices->device_list_mutex);
985
986         /* replace the sysfs entry */
987         btrfs_sysfs_remove_device(src_device);
988         btrfs_sysfs_update_devid(tgt_device);
989         if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &src_device->dev_state))
990                 btrfs_scratch_superblocks(fs_info, src_device->bdev,
991                                           src_device->name->str);
992
993         /* write back the superblocks */
994         trans = btrfs_start_transaction(root, 0);
995         if (!IS_ERR(trans))
996                 btrfs_commit_transaction(trans);
997
998         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
999
1000         btrfs_rm_dev_replace_free_srcdev(src_device);
1001
1002         return 0;
1003 }
1004
1005 /*
1006  * Read progress of device replace status according to the state and last
1007  * stored position. The value format is the same as for
1008  * btrfs_dev_replace::progress_1000
1009  */
1010 static u64 btrfs_dev_replace_progress(struct btrfs_fs_info *fs_info)
1011 {
1012         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1013         u64 ret = 0;
1014
1015         switch (dev_replace->replace_state) {
1016         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1017         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1018                 ret = 0;
1019                 break;
1020         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1021                 ret = 1000;
1022                 break;
1023         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1024         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1025                 ret = div64_u64(dev_replace->cursor_left,
1026                                 div_u64(btrfs_device_get_total_bytes(
1027                                                 dev_replace->srcdev), 1000));
1028                 break;
1029         }
1030
1031         return ret;
1032 }
1033
1034 void btrfs_dev_replace_status(struct btrfs_fs_info *fs_info,
1035                               struct btrfs_ioctl_dev_replace_args *args)
1036 {
1037         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1038
1039         down_read(&dev_replace->rwsem);
1040         /* even if !dev_replace_is_valid, the values are good enough for
1041          * the replace_status ioctl */
1042         args->result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1043         args->status.replace_state = dev_replace->replace_state;
1044         args->status.time_started = dev_replace->time_started;
1045         args->status.time_stopped = dev_replace->time_stopped;
1046         args->status.num_write_errors =
1047                 atomic64_read(&dev_replace->num_write_errors);
1048         args->status.num_uncorrectable_read_errors =
1049                 atomic64_read(&dev_replace->num_uncorrectable_read_errors);
1050         args->status.progress_1000 = btrfs_dev_replace_progress(fs_info);
1051         up_read(&dev_replace->rwsem);
1052 }
1053
1054 int btrfs_dev_replace_cancel(struct btrfs_fs_info *fs_info)
1055 {
1056         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1057         struct btrfs_device *tgt_device = NULL;
1058         struct btrfs_device *src_device = NULL;
1059         struct btrfs_trans_handle *trans;
1060         struct btrfs_root *root = fs_info->tree_root;
1061         int result;
1062         int ret;
1063
1064         if (sb_rdonly(fs_info->sb))
1065                 return -EROFS;
1066
1067         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1068         down_write(&dev_replace->rwsem);
1069         switch (dev_replace->replace_state) {
1070         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1071         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1072         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1073                 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1074                 up_write(&dev_replace->rwsem);
1075                 break;
1076         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1077                 tgt_device = dev_replace->tgtdev;
1078                 src_device = dev_replace->srcdev;
1079                 up_write(&dev_replace->rwsem);
1080                 ret = btrfs_scrub_cancel(fs_info);
1081                 if (ret < 0) {
1082                         result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1083                 } else {
1084                         result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1085                         /*
1086                          * btrfs_dev_replace_finishing() will handle the
1087                          * cleanup part
1088                          */
1089                         btrfs_info_in_rcu(fs_info,
1090                                 "dev_replace from %s (devid %llu) to %s canceled",
1091                                 btrfs_dev_name(src_device), src_device->devid,
1092                                 btrfs_dev_name(tgt_device));
1093                 }
1094                 break;
1095         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1096                 /*
1097                  * Scrub doing the replace isn't running so we need to do the
1098                  * cleanup step of btrfs_dev_replace_finishing() here
1099                  */
1100                 result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1101                 tgt_device = dev_replace->tgtdev;
1102                 src_device = dev_replace->srcdev;
1103                 dev_replace->tgtdev = NULL;
1104                 dev_replace->srcdev = NULL;
1105                 dev_replace->replace_state =
1106                                 BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED;
1107                 dev_replace->time_stopped = ktime_get_real_seconds();
1108                 dev_replace->item_needs_writeback = 1;
1109
1110                 up_write(&dev_replace->rwsem);
1111
1112                 /* Scrub for replace must not be running in suspended state */
1113                 btrfs_scrub_cancel(fs_info);
1114
1115                 trans = btrfs_start_transaction(root, 0);
1116                 if (IS_ERR(trans)) {
1117                         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1118                         return PTR_ERR(trans);
1119                 }
1120                 ret = btrfs_commit_transaction(trans);
1121                 WARN_ON(ret);
1122
1123                 btrfs_info_in_rcu(fs_info,
1124                 "suspended dev_replace from %s (devid %llu) to %s canceled",
1125                         btrfs_dev_name(src_device), src_device->devid,
1126                         btrfs_dev_name(tgt_device));
1127
1128                 if (tgt_device)
1129                         btrfs_destroy_dev_replace_tgtdev(tgt_device);
1130                 break;
1131         default:
1132                 up_write(&dev_replace->rwsem);
1133                 result = -EINVAL;
1134         }
1135
1136         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1137         return result;
1138 }
1139
1140 void btrfs_dev_replace_suspend_for_unmount(struct btrfs_fs_info *fs_info)
1141 {
1142         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1143
1144         mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1145         down_write(&dev_replace->rwsem);
1146
1147         switch (dev_replace->replace_state) {
1148         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1149         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1150         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1151         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1152                 break;
1153         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1154                 dev_replace->replace_state =
1155                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1156                 dev_replace->time_stopped = ktime_get_real_seconds();
1157                 dev_replace->item_needs_writeback = 1;
1158                 btrfs_info(fs_info, "suspending dev_replace for unmount");
1159                 break;
1160         }
1161
1162         up_write(&dev_replace->rwsem);
1163         mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1164 }
1165
1166 /* resume dev_replace procedure that was interrupted by unmount */
1167 int btrfs_resume_dev_replace_async(struct btrfs_fs_info *fs_info)
1168 {
1169         struct task_struct *task;
1170         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1171
1172         down_write(&dev_replace->rwsem);
1173
1174         switch (dev_replace->replace_state) {
1175         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1176         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1177         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1178                 up_write(&dev_replace->rwsem);
1179                 return 0;
1180         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1181                 break;
1182         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1183                 dev_replace->replace_state =
1184                         BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
1185                 break;
1186         }
1187         if (!dev_replace->tgtdev || !dev_replace->tgtdev->bdev) {
1188                 btrfs_info(fs_info,
1189                            "cannot continue dev_replace, tgtdev is missing");
1190                 btrfs_info(fs_info,
1191                            "you may cancel the operation after 'mount -o degraded'");
1192                 dev_replace->replace_state =
1193                                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1194                 up_write(&dev_replace->rwsem);
1195                 return 0;
1196         }
1197         up_write(&dev_replace->rwsem);
1198
1199         /*
1200          * This could collide with a paused balance, but the exclusive op logic
1201          * should never allow both to start and pause. We don't want to allow
1202          * dev-replace to start anyway.
1203          */
1204         if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_DEV_REPLACE)) {
1205                 down_write(&dev_replace->rwsem);
1206                 dev_replace->replace_state =
1207                                         BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1208                 up_write(&dev_replace->rwsem);
1209                 btrfs_info(fs_info,
1210                 "cannot resume dev-replace, other exclusive operation running");
1211                 return 0;
1212         }
1213
1214         task = kthread_run(btrfs_dev_replace_kthread, fs_info, "btrfs-devrepl");
1215         return PTR_ERR_OR_ZERO(task);
1216 }
1217
1218 static int btrfs_dev_replace_kthread(void *data)
1219 {
1220         struct btrfs_fs_info *fs_info = data;
1221         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1222         u64 progress;
1223         int ret;
1224
1225         progress = btrfs_dev_replace_progress(fs_info);
1226         progress = div_u64(progress, 10);
1227         btrfs_info_in_rcu(fs_info,
1228                 "continuing dev_replace from %s (devid %llu) to target %s @%u%%",
1229                 btrfs_dev_name(dev_replace->srcdev),
1230                 dev_replace->srcdev->devid,
1231                 btrfs_dev_name(dev_replace->tgtdev),
1232                 (unsigned int)progress);
1233
1234         ret = btrfs_scrub_dev(fs_info, dev_replace->srcdev->devid,
1235                               dev_replace->committed_cursor_left,
1236                               btrfs_device_get_total_bytes(dev_replace->srcdev),
1237                               &dev_replace->scrub_progress, 0, 1);
1238         ret = btrfs_dev_replace_finishing(fs_info, ret);
1239         WARN_ON(ret && ret != -ECANCELED);
1240
1241         btrfs_exclop_finish(fs_info);
1242         return 0;
1243 }
1244
1245 int __pure btrfs_dev_replace_is_ongoing(struct btrfs_dev_replace *dev_replace)
1246 {
1247         if (!dev_replace->is_valid)
1248                 return 0;
1249
1250         switch (dev_replace->replace_state) {
1251         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1252         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1253         case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1254                 return 0;
1255         case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1256         case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1257                 /*
1258                  * return true even if tgtdev is missing (this is
1259                  * something that can happen if the dev_replace
1260                  * procedure is suspended by an umount and then
1261                  * the tgtdev is missing (or "btrfs dev scan") was
1262                  * not called and the filesystem is remounted
1263                  * in degraded state. This does not stop the
1264                  * dev_replace procedure. It needs to be canceled
1265                  * manually if the cancellation is wanted.
1266                  */
1267                 break;
1268         }
1269         return 1;
1270 }
1271
1272 void btrfs_bio_counter_sub(struct btrfs_fs_info *fs_info, s64 amount)
1273 {
1274         percpu_counter_sub(&fs_info->dev_replace.bio_counter, amount);
1275         cond_wake_up_nomb(&fs_info->dev_replace.replace_wait);
1276 }
1277
1278 void btrfs_bio_counter_inc_blocked(struct btrfs_fs_info *fs_info)
1279 {
1280         while (1) {
1281                 percpu_counter_inc(&fs_info->dev_replace.bio_counter);
1282                 if (likely(!test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1283                                      &fs_info->fs_state)))
1284                         break;
1285
1286                 btrfs_bio_counter_dec(fs_info);
1287                 wait_event(fs_info->dev_replace.replace_wait,
1288                            !test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1289                                      &fs_info->fs_state));
1290         }
1291 }