xref: /linux/fs/btrfs/dev-replace.c (revision ea8a163e02d6925773129e2dd86e419e491b791d)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) STRATO AG 2012.  All rights reserved.
4  */
5 
6 #include <linux/sched.h>
7 #include <linux/bio.h>
8 #include <linux/slab.h>
9 #include <linux/blkdev.h>
10 #include <linux/kthread.h>
11 #include <linux/math64.h>
12 #include "misc.h"
13 #include "ctree.h"
14 #include "extent_map.h"
15 #include "disk-io.h"
16 #include "transaction.h"
17 #include "print-tree.h"
18 #include "volumes.h"
19 #include "async-thread.h"
20 #include "check-integrity.h"
21 #include "rcu-string.h"
22 #include "dev-replace.h"
23 #include "sysfs.h"
24 #include "zoned.h"
25 #include "block-group.h"
26 
27 /*
28  * Device replace overview
29  *
30  * [Objective]
31  * To copy all extents (both new and on-disk) from source device to target
32  * device, while still keeping the filesystem read-write.
33  *
34  * [Method]
35  * There are two main methods involved:
36  *
37  * - Write duplication
38  *
39  *   All new writes will be written to both target and source devices, so even
40  *   if replace gets canceled, sources device still contains up-to-date data.
41  *
42  *   Location:		handle_ops_on_dev_replace() from __btrfs_map_block()
43  *   Start:		btrfs_dev_replace_start()
44  *   End:		btrfs_dev_replace_finishing()
45  *   Content:		Latest data/metadata
46  *
47  * - Copy existing extents
48  *
49  *   This happens by re-using scrub facility, as scrub also iterates through
50  *   existing extents from commit root.
51  *
52  *   Location:		scrub_write_block_to_dev_replace() from
53  *   			scrub_block_complete()
54  *   Content:		Data/meta from commit root.
55  *
56  * Due to the content difference, we need to avoid nocow write when dev-replace
57  * is happening.  This is done by marking the block group read-only and waiting
58  * for NOCOW writes.
59  *
60  * After replace is done, the finishing part is done by swapping the target and
61  * source devices.
62  *
63  *   Location:		btrfs_dev_replace_update_device_in_mapping_tree() from
64  *   			btrfs_dev_replace_finishing()
65  */
66 
67 static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
68 				       int scrub_ret);
69 static int btrfs_dev_replace_kthread(void *data);
70 
71 int btrfs_init_dev_replace(struct btrfs_fs_info *fs_info)
72 {
73 	struct btrfs_dev_lookup_args args = { .devid = BTRFS_DEV_REPLACE_DEVID };
74 	struct btrfs_key key;
75 	struct btrfs_root *dev_root = fs_info->dev_root;
76 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
77 	struct extent_buffer *eb;
78 	int slot;
79 	int ret = 0;
80 	struct btrfs_path *path = NULL;
81 	int item_size;
82 	struct btrfs_dev_replace_item *ptr;
83 	u64 src_devid;
84 
85 	if (!dev_root)
86 		return 0;
87 
88 	path = btrfs_alloc_path();
89 	if (!path) {
90 		ret = -ENOMEM;
91 		goto out;
92 	}
93 
94 	key.objectid = 0;
95 	key.type = BTRFS_DEV_REPLACE_KEY;
96 	key.offset = 0;
97 	ret = btrfs_search_slot(NULL, dev_root, &key, path, 0, 0);
98 	if (ret) {
99 no_valid_dev_replace_entry_found:
100 		/*
101 		 * We don't have a replace item or it's corrupted.  If there is
102 		 * a replace target, fail the mount.
103 		 */
104 		if (btrfs_find_device(fs_info->fs_devices, &args)) {
105 			btrfs_err(fs_info,
106 			"found replace target device without a valid replace item");
107 			ret = -EUCLEAN;
108 			goto out;
109 		}
110 		ret = 0;
111 		dev_replace->replace_state =
112 			BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED;
113 		dev_replace->cont_reading_from_srcdev_mode =
114 		    BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_ALWAYS;
115 		dev_replace->time_started = 0;
116 		dev_replace->time_stopped = 0;
117 		atomic64_set(&dev_replace->num_write_errors, 0);
118 		atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
119 		dev_replace->cursor_left = 0;
120 		dev_replace->committed_cursor_left = 0;
121 		dev_replace->cursor_left_last_write_of_item = 0;
122 		dev_replace->cursor_right = 0;
123 		dev_replace->srcdev = NULL;
124 		dev_replace->tgtdev = NULL;
125 		dev_replace->is_valid = 0;
126 		dev_replace->item_needs_writeback = 0;
127 		goto out;
128 	}
129 	slot = path->slots[0];
130 	eb = path->nodes[0];
131 	item_size = btrfs_item_size_nr(eb, slot);
132 	ptr = btrfs_item_ptr(eb, slot, struct btrfs_dev_replace_item);
133 
134 	if (item_size != sizeof(struct btrfs_dev_replace_item)) {
135 		btrfs_warn(fs_info,
136 			"dev_replace entry found has unexpected size, ignore entry");
137 		goto no_valid_dev_replace_entry_found;
138 	}
139 
140 	src_devid = btrfs_dev_replace_src_devid(eb, ptr);
141 	dev_replace->cont_reading_from_srcdev_mode =
142 		btrfs_dev_replace_cont_reading_from_srcdev_mode(eb, ptr);
143 	dev_replace->replace_state = btrfs_dev_replace_replace_state(eb, ptr);
144 	dev_replace->time_started = btrfs_dev_replace_time_started(eb, ptr);
145 	dev_replace->time_stopped =
146 		btrfs_dev_replace_time_stopped(eb, ptr);
147 	atomic64_set(&dev_replace->num_write_errors,
148 		     btrfs_dev_replace_num_write_errors(eb, ptr));
149 	atomic64_set(&dev_replace->num_uncorrectable_read_errors,
150 		     btrfs_dev_replace_num_uncorrectable_read_errors(eb, ptr));
151 	dev_replace->cursor_left = btrfs_dev_replace_cursor_left(eb, ptr);
152 	dev_replace->committed_cursor_left = dev_replace->cursor_left;
153 	dev_replace->cursor_left_last_write_of_item = dev_replace->cursor_left;
154 	dev_replace->cursor_right = btrfs_dev_replace_cursor_right(eb, ptr);
155 	dev_replace->is_valid = 1;
156 
157 	dev_replace->item_needs_writeback = 0;
158 	switch (dev_replace->replace_state) {
159 	case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
160 	case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
161 	case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
162 		/*
163 		 * We don't have an active replace item but if there is a
164 		 * replace target, fail the mount.
165 		 */
166 		if (btrfs_find_device(fs_info->fs_devices, &args)) {
167 			btrfs_err(fs_info,
168 			"replace devid present without an active replace item");
169 			ret = -EUCLEAN;
170 		} else {
171 			dev_replace->srcdev = NULL;
172 			dev_replace->tgtdev = NULL;
173 		}
174 		break;
175 	case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
176 	case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
177 		dev_replace->tgtdev = btrfs_find_device(fs_info->fs_devices, &args);
178 		args.devid = src_devid;
179 		dev_replace->srcdev = btrfs_find_device(fs_info->fs_devices, &args);
180 
181 		/*
182 		 * allow 'btrfs dev replace_cancel' if src/tgt device is
183 		 * missing
184 		 */
185 		if (!dev_replace->srcdev &&
186 		    !btrfs_test_opt(fs_info, DEGRADED)) {
187 			ret = -EIO;
188 			btrfs_warn(fs_info,
189 			   "cannot mount because device replace operation is ongoing and");
190 			btrfs_warn(fs_info,
191 			   "srcdev (devid %llu) is missing, need to run 'btrfs dev scan'?",
192 			   src_devid);
193 		}
194 		if (!dev_replace->tgtdev &&
195 		    !btrfs_test_opt(fs_info, DEGRADED)) {
196 			ret = -EIO;
197 			btrfs_warn(fs_info,
198 			   "cannot mount because device replace operation is ongoing and");
199 			btrfs_warn(fs_info,
200 			   "tgtdev (devid %llu) is missing, need to run 'btrfs dev scan'?",
201 				BTRFS_DEV_REPLACE_DEVID);
202 		}
203 		if (dev_replace->tgtdev) {
204 			if (dev_replace->srcdev) {
205 				dev_replace->tgtdev->total_bytes =
206 					dev_replace->srcdev->total_bytes;
207 				dev_replace->tgtdev->disk_total_bytes =
208 					dev_replace->srcdev->disk_total_bytes;
209 				dev_replace->tgtdev->commit_total_bytes =
210 					dev_replace->srcdev->commit_total_bytes;
211 				dev_replace->tgtdev->bytes_used =
212 					dev_replace->srcdev->bytes_used;
213 				dev_replace->tgtdev->commit_bytes_used =
214 					dev_replace->srcdev->commit_bytes_used;
215 			}
216 			set_bit(BTRFS_DEV_STATE_REPLACE_TGT,
217 				&dev_replace->tgtdev->dev_state);
218 
219 			WARN_ON(fs_info->fs_devices->rw_devices == 0);
220 			dev_replace->tgtdev->io_width = fs_info->sectorsize;
221 			dev_replace->tgtdev->io_align = fs_info->sectorsize;
222 			dev_replace->tgtdev->sector_size = fs_info->sectorsize;
223 			dev_replace->tgtdev->fs_info = fs_info;
224 			set_bit(BTRFS_DEV_STATE_IN_FS_METADATA,
225 				&dev_replace->tgtdev->dev_state);
226 		}
227 		break;
228 	}
229 
230 out:
231 	btrfs_free_path(path);
232 	return ret;
233 }
234 
235 /*
236  * Initialize a new device for device replace target from a given source dev
237  * and path.
238  *
239  * Return 0 and new device in @device_out, otherwise return < 0
240  */
241 static int btrfs_init_dev_replace_tgtdev(struct btrfs_fs_info *fs_info,
242 				  const char *device_path,
243 				  struct btrfs_device *srcdev,
244 				  struct btrfs_device **device_out)
245 {
246 	struct btrfs_device *device;
247 	struct block_device *bdev;
248 	struct rcu_string *name;
249 	u64 devid = BTRFS_DEV_REPLACE_DEVID;
250 	int ret = 0;
251 
252 	*device_out = NULL;
253 	if (srcdev->fs_devices->seeding) {
254 		btrfs_err(fs_info, "the filesystem is a seed filesystem!");
255 		return -EINVAL;
256 	}
257 
258 	bdev = blkdev_get_by_path(device_path, FMODE_WRITE | FMODE_EXCL,
259 				  fs_info->bdev_holder);
260 	if (IS_ERR(bdev)) {
261 		btrfs_err(fs_info, "target device %s is invalid!", device_path);
262 		return PTR_ERR(bdev);
263 	}
264 
265 	if (!btrfs_check_device_zone_type(fs_info, bdev)) {
266 		btrfs_err(fs_info,
267 		"dev-replace: zoned type of target device mismatch with filesystem");
268 		ret = -EINVAL;
269 		goto error;
270 	}
271 
272 	sync_blockdev(bdev);
273 
274 	list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
275 		if (device->bdev == bdev) {
276 			btrfs_err(fs_info,
277 				  "target device is in the filesystem!");
278 			ret = -EEXIST;
279 			goto error;
280 		}
281 	}
282 
283 
284 	if (bdev_nr_bytes(bdev) < btrfs_device_get_total_bytes(srcdev)) {
285 		btrfs_err(fs_info,
286 			  "target device is smaller than source device!");
287 		ret = -EINVAL;
288 		goto error;
289 	}
290 
291 
292 	device = btrfs_alloc_device(NULL, &devid, NULL);
293 	if (IS_ERR(device)) {
294 		ret = PTR_ERR(device);
295 		goto error;
296 	}
297 
298 	name = rcu_string_strdup(device_path, GFP_KERNEL);
299 	if (!name) {
300 		btrfs_free_device(device);
301 		ret = -ENOMEM;
302 		goto error;
303 	}
304 	rcu_assign_pointer(device->name, name);
305 
306 	set_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state);
307 	device->generation = 0;
308 	device->io_width = fs_info->sectorsize;
309 	device->io_align = fs_info->sectorsize;
310 	device->sector_size = fs_info->sectorsize;
311 	device->total_bytes = btrfs_device_get_total_bytes(srcdev);
312 	device->disk_total_bytes = btrfs_device_get_disk_total_bytes(srcdev);
313 	device->bytes_used = btrfs_device_get_bytes_used(srcdev);
314 	device->commit_total_bytes = srcdev->commit_total_bytes;
315 	device->commit_bytes_used = device->bytes_used;
316 	device->fs_info = fs_info;
317 	device->bdev = bdev;
318 	set_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state);
319 	set_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state);
320 	device->mode = FMODE_EXCL;
321 	device->dev_stats_valid = 1;
322 	set_blocksize(device->bdev, BTRFS_BDEV_BLOCKSIZE);
323 	device->fs_devices = fs_info->fs_devices;
324 
325 	ret = btrfs_get_dev_zone_info(device);
326 	if (ret)
327 		goto error;
328 
329 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
330 	list_add(&device->dev_list, &fs_info->fs_devices->devices);
331 	fs_info->fs_devices->num_devices++;
332 	fs_info->fs_devices->open_devices++;
333 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
334 
335 	*device_out = device;
336 	return 0;
337 
338 error:
339 	blkdev_put(bdev, FMODE_EXCL);
340 	return ret;
341 }
342 
343 /*
344  * called from commit_transaction. Writes changed device replace state to
345  * disk.
346  */
347 int btrfs_run_dev_replace(struct btrfs_trans_handle *trans)
348 {
349 	struct btrfs_fs_info *fs_info = trans->fs_info;
350 	int ret;
351 	struct btrfs_root *dev_root = fs_info->dev_root;
352 	struct btrfs_path *path;
353 	struct btrfs_key key;
354 	struct extent_buffer *eb;
355 	struct btrfs_dev_replace_item *ptr;
356 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
357 
358 	down_read(&dev_replace->rwsem);
359 	if (!dev_replace->is_valid ||
360 	    !dev_replace->item_needs_writeback) {
361 		up_read(&dev_replace->rwsem);
362 		return 0;
363 	}
364 	up_read(&dev_replace->rwsem);
365 
366 	key.objectid = 0;
367 	key.type = BTRFS_DEV_REPLACE_KEY;
368 	key.offset = 0;
369 
370 	path = btrfs_alloc_path();
371 	if (!path) {
372 		ret = -ENOMEM;
373 		goto out;
374 	}
375 	ret = btrfs_search_slot(trans, dev_root, &key, path, -1, 1);
376 	if (ret < 0) {
377 		btrfs_warn(fs_info,
378 			   "error %d while searching for dev_replace item!",
379 			   ret);
380 		goto out;
381 	}
382 
383 	if (ret == 0 &&
384 	    btrfs_item_size_nr(path->nodes[0], path->slots[0]) < sizeof(*ptr)) {
385 		/*
386 		 * need to delete old one and insert a new one.
387 		 * Since no attempt is made to recover any old state, if the
388 		 * dev_replace state is 'running', the data on the target
389 		 * drive is lost.
390 		 * It would be possible to recover the state: just make sure
391 		 * that the beginning of the item is never changed and always
392 		 * contains all the essential information. Then read this
393 		 * minimal set of information and use it as a base for the
394 		 * new state.
395 		 */
396 		ret = btrfs_del_item(trans, dev_root, path);
397 		if (ret != 0) {
398 			btrfs_warn(fs_info,
399 				   "delete too small dev_replace item failed %d!",
400 				   ret);
401 			goto out;
402 		}
403 		ret = 1;
404 	}
405 
406 	if (ret == 1) {
407 		/* need to insert a new item */
408 		btrfs_release_path(path);
409 		ret = btrfs_insert_empty_item(trans, dev_root, path,
410 					      &key, sizeof(*ptr));
411 		if (ret < 0) {
412 			btrfs_warn(fs_info,
413 				   "insert dev_replace item failed %d!", ret);
414 			goto out;
415 		}
416 	}
417 
418 	eb = path->nodes[0];
419 	ptr = btrfs_item_ptr(eb, path->slots[0],
420 			     struct btrfs_dev_replace_item);
421 
422 	down_write(&dev_replace->rwsem);
423 	if (dev_replace->srcdev)
424 		btrfs_set_dev_replace_src_devid(eb, ptr,
425 			dev_replace->srcdev->devid);
426 	else
427 		btrfs_set_dev_replace_src_devid(eb, ptr, (u64)-1);
428 	btrfs_set_dev_replace_cont_reading_from_srcdev_mode(eb, ptr,
429 		dev_replace->cont_reading_from_srcdev_mode);
430 	btrfs_set_dev_replace_replace_state(eb, ptr,
431 		dev_replace->replace_state);
432 	btrfs_set_dev_replace_time_started(eb, ptr, dev_replace->time_started);
433 	btrfs_set_dev_replace_time_stopped(eb, ptr, dev_replace->time_stopped);
434 	btrfs_set_dev_replace_num_write_errors(eb, ptr,
435 		atomic64_read(&dev_replace->num_write_errors));
436 	btrfs_set_dev_replace_num_uncorrectable_read_errors(eb, ptr,
437 		atomic64_read(&dev_replace->num_uncorrectable_read_errors));
438 	dev_replace->cursor_left_last_write_of_item =
439 		dev_replace->cursor_left;
440 	btrfs_set_dev_replace_cursor_left(eb, ptr,
441 		dev_replace->cursor_left_last_write_of_item);
442 	btrfs_set_dev_replace_cursor_right(eb, ptr,
443 		dev_replace->cursor_right);
444 	dev_replace->item_needs_writeback = 0;
445 	up_write(&dev_replace->rwsem);
446 
447 	btrfs_mark_buffer_dirty(eb);
448 
449 out:
450 	btrfs_free_path(path);
451 
452 	return ret;
453 }
454 
455 static char* btrfs_dev_name(struct btrfs_device *device)
456 {
457 	if (!device || test_bit(BTRFS_DEV_STATE_MISSING, &device->dev_state))
458 		return "<missing disk>";
459 	else
460 		return rcu_str_deref(device->name);
461 }
462 
463 static int mark_block_group_to_copy(struct btrfs_fs_info *fs_info,
464 				    struct btrfs_device *src_dev)
465 {
466 	struct btrfs_path *path;
467 	struct btrfs_key key;
468 	struct btrfs_key found_key;
469 	struct btrfs_root *root = fs_info->dev_root;
470 	struct btrfs_dev_extent *dev_extent = NULL;
471 	struct btrfs_block_group *cache;
472 	struct btrfs_trans_handle *trans;
473 	int ret = 0;
474 	u64 chunk_offset;
475 
476 	/* Do not use "to_copy" on non zoned filesystem for now */
477 	if (!btrfs_is_zoned(fs_info))
478 		return 0;
479 
480 	mutex_lock(&fs_info->chunk_mutex);
481 
482 	/* Ensure we don't have pending new block group */
483 	spin_lock(&fs_info->trans_lock);
484 	while (fs_info->running_transaction &&
485 	       !list_empty(&fs_info->running_transaction->dev_update_list)) {
486 		spin_unlock(&fs_info->trans_lock);
487 		mutex_unlock(&fs_info->chunk_mutex);
488 		trans = btrfs_attach_transaction(root);
489 		if (IS_ERR(trans)) {
490 			ret = PTR_ERR(trans);
491 			mutex_lock(&fs_info->chunk_mutex);
492 			if (ret == -ENOENT) {
493 				spin_lock(&fs_info->trans_lock);
494 				continue;
495 			} else {
496 				goto unlock;
497 			}
498 		}
499 
500 		ret = btrfs_commit_transaction(trans);
501 		mutex_lock(&fs_info->chunk_mutex);
502 		if (ret)
503 			goto unlock;
504 
505 		spin_lock(&fs_info->trans_lock);
506 	}
507 	spin_unlock(&fs_info->trans_lock);
508 
509 	path = btrfs_alloc_path();
510 	if (!path) {
511 		ret = -ENOMEM;
512 		goto unlock;
513 	}
514 
515 	path->reada = READA_FORWARD;
516 	path->search_commit_root = 1;
517 	path->skip_locking = 1;
518 
519 	key.objectid = src_dev->devid;
520 	key.type = BTRFS_DEV_EXTENT_KEY;
521 	key.offset = 0;
522 
523 	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
524 	if (ret < 0)
525 		goto free_path;
526 	if (ret > 0) {
527 		if (path->slots[0] >=
528 		    btrfs_header_nritems(path->nodes[0])) {
529 			ret = btrfs_next_leaf(root, path);
530 			if (ret < 0)
531 				goto free_path;
532 			if (ret > 0) {
533 				ret = 0;
534 				goto free_path;
535 			}
536 		} else {
537 			ret = 0;
538 		}
539 	}
540 
541 	while (1) {
542 		struct extent_buffer *leaf = path->nodes[0];
543 		int slot = path->slots[0];
544 
545 		btrfs_item_key_to_cpu(leaf, &found_key, slot);
546 
547 		if (found_key.objectid != src_dev->devid)
548 			break;
549 
550 		if (found_key.type != BTRFS_DEV_EXTENT_KEY)
551 			break;
552 
553 		if (found_key.offset < key.offset)
554 			break;
555 
556 		dev_extent = btrfs_item_ptr(leaf, slot, struct btrfs_dev_extent);
557 
558 		chunk_offset = btrfs_dev_extent_chunk_offset(leaf, dev_extent);
559 
560 		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
561 		if (!cache)
562 			goto skip;
563 
564 		spin_lock(&cache->lock);
565 		cache->to_copy = 1;
566 		spin_unlock(&cache->lock);
567 
568 		btrfs_put_block_group(cache);
569 
570 skip:
571 		ret = btrfs_next_item(root, path);
572 		if (ret != 0) {
573 			if (ret > 0)
574 				ret = 0;
575 			break;
576 		}
577 	}
578 
579 free_path:
580 	btrfs_free_path(path);
581 unlock:
582 	mutex_unlock(&fs_info->chunk_mutex);
583 
584 	return ret;
585 }
586 
587 bool btrfs_finish_block_group_to_copy(struct btrfs_device *srcdev,
588 				      struct btrfs_block_group *cache,
589 				      u64 physical)
590 {
591 	struct btrfs_fs_info *fs_info = cache->fs_info;
592 	struct extent_map *em;
593 	struct map_lookup *map;
594 	u64 chunk_offset = cache->start;
595 	int num_extents, cur_extent;
596 	int i;
597 
598 	/* Do not use "to_copy" on non zoned filesystem for now */
599 	if (!btrfs_is_zoned(fs_info))
600 		return true;
601 
602 	spin_lock(&cache->lock);
603 	if (cache->removed) {
604 		spin_unlock(&cache->lock);
605 		return true;
606 	}
607 	spin_unlock(&cache->lock);
608 
609 	em = btrfs_get_chunk_map(fs_info, chunk_offset, 1);
610 	ASSERT(!IS_ERR(em));
611 	map = em->map_lookup;
612 
613 	num_extents = cur_extent = 0;
614 	for (i = 0; i < map->num_stripes; i++) {
615 		/* We have more device extent to copy */
616 		if (srcdev != map->stripes[i].dev)
617 			continue;
618 
619 		num_extents++;
620 		if (physical == map->stripes[i].physical)
621 			cur_extent = i;
622 	}
623 
624 	free_extent_map(em);
625 
626 	if (num_extents > 1 && cur_extent < num_extents - 1) {
627 		/*
628 		 * Has more stripes on this device. Keep this block group
629 		 * readonly until we finish all the stripes.
630 		 */
631 		return false;
632 	}
633 
634 	/* Last stripe on this device */
635 	spin_lock(&cache->lock);
636 	cache->to_copy = 0;
637 	spin_unlock(&cache->lock);
638 
639 	return true;
640 }
641 
642 static int btrfs_dev_replace_start(struct btrfs_fs_info *fs_info,
643 		const char *tgtdev_name, u64 srcdevid, const char *srcdev_name,
644 		int read_src)
645 {
646 	struct btrfs_root *root = fs_info->dev_root;
647 	struct btrfs_trans_handle *trans;
648 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
649 	int ret;
650 	struct btrfs_device *tgt_device = NULL;
651 	struct btrfs_device *src_device = NULL;
652 
653 	src_device = btrfs_find_device_by_devspec(fs_info, srcdevid,
654 						  srcdev_name);
655 	if (IS_ERR(src_device))
656 		return PTR_ERR(src_device);
657 
658 	if (btrfs_pinned_by_swapfile(fs_info, src_device)) {
659 		btrfs_warn_in_rcu(fs_info,
660 	  "cannot replace device %s (devid %llu) due to active swapfile",
661 			btrfs_dev_name(src_device), src_device->devid);
662 		return -ETXTBSY;
663 	}
664 
665 	/*
666 	 * Here we commit the transaction to make sure commit_total_bytes
667 	 * of all the devices are updated.
668 	 */
669 	trans = btrfs_attach_transaction(root);
670 	if (!IS_ERR(trans)) {
671 		ret = btrfs_commit_transaction(trans);
672 		if (ret)
673 			return ret;
674 	} else if (PTR_ERR(trans) != -ENOENT) {
675 		return PTR_ERR(trans);
676 	}
677 
678 	ret = btrfs_init_dev_replace_tgtdev(fs_info, tgtdev_name,
679 					    src_device, &tgt_device);
680 	if (ret)
681 		return ret;
682 
683 	ret = mark_block_group_to_copy(fs_info, src_device);
684 	if (ret)
685 		return ret;
686 
687 	down_write(&dev_replace->rwsem);
688 	switch (dev_replace->replace_state) {
689 	case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
690 	case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
691 	case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
692 		break;
693 	case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
694 	case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
695 		ASSERT(0);
696 		ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_ALREADY_STARTED;
697 		up_write(&dev_replace->rwsem);
698 		goto leave;
699 	}
700 
701 	dev_replace->cont_reading_from_srcdev_mode = read_src;
702 	dev_replace->srcdev = src_device;
703 	dev_replace->tgtdev = tgt_device;
704 
705 	btrfs_info_in_rcu(fs_info,
706 		      "dev_replace from %s (devid %llu) to %s started",
707 		      btrfs_dev_name(src_device),
708 		      src_device->devid,
709 		      rcu_str_deref(tgt_device->name));
710 
711 	/*
712 	 * from now on, the writes to the srcdev are all duplicated to
713 	 * go to the tgtdev as well (refer to btrfs_map_block()).
714 	 */
715 	dev_replace->replace_state = BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
716 	dev_replace->time_started = ktime_get_real_seconds();
717 	dev_replace->cursor_left = 0;
718 	dev_replace->committed_cursor_left = 0;
719 	dev_replace->cursor_left_last_write_of_item = 0;
720 	dev_replace->cursor_right = 0;
721 	dev_replace->is_valid = 1;
722 	dev_replace->item_needs_writeback = 1;
723 	atomic64_set(&dev_replace->num_write_errors, 0);
724 	atomic64_set(&dev_replace->num_uncorrectable_read_errors, 0);
725 	up_write(&dev_replace->rwsem);
726 
727 	ret = btrfs_sysfs_add_device(tgt_device);
728 	if (ret)
729 		btrfs_err(fs_info, "kobj add dev failed %d", ret);
730 
731 	btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
732 
733 	/* Commit dev_replace state and reserve 1 item for it. */
734 	trans = btrfs_start_transaction(root, 1);
735 	if (IS_ERR(trans)) {
736 		ret = PTR_ERR(trans);
737 		down_write(&dev_replace->rwsem);
738 		dev_replace->replace_state =
739 			BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED;
740 		dev_replace->srcdev = NULL;
741 		dev_replace->tgtdev = NULL;
742 		up_write(&dev_replace->rwsem);
743 		goto leave;
744 	}
745 
746 	ret = btrfs_commit_transaction(trans);
747 	WARN_ON(ret);
748 
749 	/* the disk copy procedure reuses the scrub code */
750 	ret = btrfs_scrub_dev(fs_info, src_device->devid, 0,
751 			      btrfs_device_get_total_bytes(src_device),
752 			      &dev_replace->scrub_progress, 0, 1);
753 
754 	ret = btrfs_dev_replace_finishing(fs_info, ret);
755 	if (ret == -EINPROGRESS)
756 		ret = BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS;
757 
758 	return ret;
759 
760 leave:
761 	btrfs_destroy_dev_replace_tgtdev(tgt_device);
762 	return ret;
763 }
764 
765 int btrfs_dev_replace_by_ioctl(struct btrfs_fs_info *fs_info,
766 			    struct btrfs_ioctl_dev_replace_args *args)
767 {
768 	int ret;
769 
770 	switch (args->start.cont_reading_from_srcdev_mode) {
771 	case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_ALWAYS:
772 	case BTRFS_IOCTL_DEV_REPLACE_CONT_READING_FROM_SRCDEV_MODE_AVOID:
773 		break;
774 	default:
775 		return -EINVAL;
776 	}
777 
778 	if ((args->start.srcdevid == 0 && args->start.srcdev_name[0] == '\0') ||
779 	    args->start.tgtdev_name[0] == '\0')
780 		return -EINVAL;
781 
782 	ret = btrfs_dev_replace_start(fs_info, args->start.tgtdev_name,
783 					args->start.srcdevid,
784 					args->start.srcdev_name,
785 					args->start.cont_reading_from_srcdev_mode);
786 	args->result = ret;
787 	/* don't warn if EINPROGRESS, someone else might be running scrub */
788 	if (ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_SCRUB_INPROGRESS ||
789 	    ret == BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR)
790 		return 0;
791 
792 	return ret;
793 }
794 
795 /*
796  * blocked until all in-flight bios operations are finished.
797  */
798 static void btrfs_rm_dev_replace_blocked(struct btrfs_fs_info *fs_info)
799 {
800 	set_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
801 	wait_event(fs_info->dev_replace.replace_wait, !percpu_counter_sum(
802 		   &fs_info->dev_replace.bio_counter));
803 }
804 
805 /*
806  * we have removed target device, it is safe to allow new bios request.
807  */
808 static void btrfs_rm_dev_replace_unblocked(struct btrfs_fs_info *fs_info)
809 {
810 	clear_bit(BTRFS_FS_STATE_DEV_REPLACING, &fs_info->fs_state);
811 	wake_up(&fs_info->dev_replace.replace_wait);
812 }
813 
814 /*
815  * When finishing the device replace, before swapping the source device with the
816  * target device we must update the chunk allocation state in the target device,
817  * as it is empty because replace works by directly copying the chunks and not
818  * through the normal chunk allocation path.
819  */
820 static int btrfs_set_target_alloc_state(struct btrfs_device *srcdev,
821 					struct btrfs_device *tgtdev)
822 {
823 	struct extent_state *cached_state = NULL;
824 	u64 start = 0;
825 	u64 found_start;
826 	u64 found_end;
827 	int ret = 0;
828 
829 	lockdep_assert_held(&srcdev->fs_info->chunk_mutex);
830 
831 	while (!find_first_extent_bit(&srcdev->alloc_state, start,
832 				      &found_start, &found_end,
833 				      CHUNK_ALLOCATED, &cached_state)) {
834 		ret = set_extent_bits(&tgtdev->alloc_state, found_start,
835 				      found_end, CHUNK_ALLOCATED);
836 		if (ret)
837 			break;
838 		start = found_end + 1;
839 	}
840 
841 	free_extent_state(cached_state);
842 	return ret;
843 }
844 
845 static void btrfs_dev_replace_update_device_in_mapping_tree(
846 						struct btrfs_fs_info *fs_info,
847 						struct btrfs_device *srcdev,
848 						struct btrfs_device *tgtdev)
849 {
850 	struct extent_map_tree *em_tree = &fs_info->mapping_tree;
851 	struct extent_map *em;
852 	struct map_lookup *map;
853 	u64 start = 0;
854 	int i;
855 
856 	write_lock(&em_tree->lock);
857 	do {
858 		em = lookup_extent_mapping(em_tree, start, (u64)-1);
859 		if (!em)
860 			break;
861 		map = em->map_lookup;
862 		for (i = 0; i < map->num_stripes; i++)
863 			if (srcdev == map->stripes[i].dev)
864 				map->stripes[i].dev = tgtdev;
865 		start = em->start + em->len;
866 		free_extent_map(em);
867 	} while (start);
868 	write_unlock(&em_tree->lock);
869 }
870 
871 static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
872 				       int scrub_ret)
873 {
874 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
875 	struct btrfs_device *tgt_device;
876 	struct btrfs_device *src_device;
877 	struct btrfs_root *root = fs_info->tree_root;
878 	u8 uuid_tmp[BTRFS_UUID_SIZE];
879 	struct btrfs_trans_handle *trans;
880 	int ret = 0;
881 
882 	/* don't allow cancel or unmount to disturb the finishing procedure */
883 	mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
884 
885 	down_read(&dev_replace->rwsem);
886 	/* was the operation canceled, or is it finished? */
887 	if (dev_replace->replace_state !=
888 	    BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED) {
889 		up_read(&dev_replace->rwsem);
890 		mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
891 		return 0;
892 	}
893 
894 	tgt_device = dev_replace->tgtdev;
895 	src_device = dev_replace->srcdev;
896 	up_read(&dev_replace->rwsem);
897 
898 	/*
899 	 * flush all outstanding I/O and inode extent mappings before the
900 	 * copy operation is declared as being finished
901 	 */
902 	ret = btrfs_start_delalloc_roots(fs_info, LONG_MAX, false);
903 	if (ret) {
904 		mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
905 		return ret;
906 	}
907 	btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
908 
909 	if (!scrub_ret)
910 		btrfs_reada_remove_dev(src_device);
911 
912 	/*
913 	 * We have to use this loop approach because at this point src_device
914 	 * has to be available for transaction commit to complete, yet new
915 	 * chunks shouldn't be allocated on the device.
916 	 */
917 	while (1) {
918 		trans = btrfs_start_transaction(root, 0);
919 		if (IS_ERR(trans)) {
920 			btrfs_reada_undo_remove_dev(src_device);
921 			mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
922 			return PTR_ERR(trans);
923 		}
924 		ret = btrfs_commit_transaction(trans);
925 		WARN_ON(ret);
926 
927 		/* Prevent write_all_supers() during the finishing procedure */
928 		mutex_lock(&fs_info->fs_devices->device_list_mutex);
929 		/* Prevent new chunks being allocated on the source device */
930 		mutex_lock(&fs_info->chunk_mutex);
931 
932 		if (!list_empty(&src_device->post_commit_list)) {
933 			mutex_unlock(&fs_info->fs_devices->device_list_mutex);
934 			mutex_unlock(&fs_info->chunk_mutex);
935 		} else {
936 			break;
937 		}
938 	}
939 
940 	down_write(&dev_replace->rwsem);
941 	dev_replace->replace_state =
942 		scrub_ret ? BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED
943 			  : BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED;
944 	dev_replace->tgtdev = NULL;
945 	dev_replace->srcdev = NULL;
946 	dev_replace->time_stopped = ktime_get_real_seconds();
947 	dev_replace->item_needs_writeback = 1;
948 
949 	/*
950 	 * Update allocation state in the new device and replace the old device
951 	 * with the new one in the mapping tree.
952 	 */
953 	if (!scrub_ret) {
954 		scrub_ret = btrfs_set_target_alloc_state(src_device, tgt_device);
955 		if (scrub_ret)
956 			goto error;
957 		btrfs_dev_replace_update_device_in_mapping_tree(fs_info,
958 								src_device,
959 								tgt_device);
960 	} else {
961 		if (scrub_ret != -ECANCELED)
962 			btrfs_err_in_rcu(fs_info,
963 				 "btrfs_scrub_dev(%s, %llu, %s) failed %d",
964 				 btrfs_dev_name(src_device),
965 				 src_device->devid,
966 				 rcu_str_deref(tgt_device->name), scrub_ret);
967 error:
968 		up_write(&dev_replace->rwsem);
969 		mutex_unlock(&fs_info->chunk_mutex);
970 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
971 		btrfs_reada_undo_remove_dev(src_device);
972 		btrfs_rm_dev_replace_blocked(fs_info);
973 		if (tgt_device)
974 			btrfs_destroy_dev_replace_tgtdev(tgt_device);
975 		btrfs_rm_dev_replace_unblocked(fs_info);
976 		mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
977 
978 		return scrub_ret;
979 	}
980 
981 	btrfs_info_in_rcu(fs_info,
982 			  "dev_replace from %s (devid %llu) to %s finished",
983 			  btrfs_dev_name(src_device),
984 			  src_device->devid,
985 			  rcu_str_deref(tgt_device->name));
986 	clear_bit(BTRFS_DEV_STATE_REPLACE_TGT, &tgt_device->dev_state);
987 	tgt_device->devid = src_device->devid;
988 	src_device->devid = BTRFS_DEV_REPLACE_DEVID;
989 	memcpy(uuid_tmp, tgt_device->uuid, sizeof(uuid_tmp));
990 	memcpy(tgt_device->uuid, src_device->uuid, sizeof(tgt_device->uuid));
991 	memcpy(src_device->uuid, uuid_tmp, sizeof(src_device->uuid));
992 	btrfs_device_set_total_bytes(tgt_device, src_device->total_bytes);
993 	btrfs_device_set_disk_total_bytes(tgt_device,
994 					  src_device->disk_total_bytes);
995 	btrfs_device_set_bytes_used(tgt_device, src_device->bytes_used);
996 	tgt_device->commit_bytes_used = src_device->bytes_used;
997 
998 	btrfs_assign_next_active_device(src_device, tgt_device);
999 
1000 	list_add(&tgt_device->dev_alloc_list, &fs_info->fs_devices->alloc_list);
1001 	fs_info->fs_devices->rw_devices++;
1002 
1003 	up_write(&dev_replace->rwsem);
1004 	btrfs_rm_dev_replace_blocked(fs_info);
1005 
1006 	btrfs_rm_dev_replace_remove_srcdev(src_device);
1007 
1008 	btrfs_rm_dev_replace_unblocked(fs_info);
1009 
1010 	/*
1011 	 * Increment dev_stats_ccnt so that btrfs_run_dev_stats() will
1012 	 * update on-disk dev stats value during commit transaction
1013 	 */
1014 	atomic_inc(&tgt_device->dev_stats_ccnt);
1015 
1016 	/*
1017 	 * this is again a consistent state where no dev_replace procedure
1018 	 * is running, the target device is part of the filesystem, the
1019 	 * source device is not part of the filesystem anymore and its 1st
1020 	 * superblock is scratched out so that it is no longer marked to
1021 	 * belong to this filesystem.
1022 	 */
1023 	mutex_unlock(&fs_info->chunk_mutex);
1024 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1025 
1026 	/* replace the sysfs entry */
1027 	btrfs_sysfs_remove_device(src_device);
1028 	btrfs_sysfs_update_devid(tgt_device);
1029 	if (test_bit(BTRFS_DEV_STATE_WRITEABLE, &src_device->dev_state))
1030 		btrfs_scratch_superblocks(fs_info, src_device->bdev,
1031 					  src_device->name->str);
1032 
1033 	/* write back the superblocks */
1034 	trans = btrfs_start_transaction(root, 0);
1035 	if (!IS_ERR(trans))
1036 		btrfs_commit_transaction(trans);
1037 
1038 	mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1039 
1040 	btrfs_rm_dev_replace_free_srcdev(src_device);
1041 
1042 	return 0;
1043 }
1044 
1045 /*
1046  * Read progress of device replace status according to the state and last
1047  * stored position. The value format is the same as for
1048  * btrfs_dev_replace::progress_1000
1049  */
1050 static u64 btrfs_dev_replace_progress(struct btrfs_fs_info *fs_info)
1051 {
1052 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1053 	u64 ret = 0;
1054 
1055 	switch (dev_replace->replace_state) {
1056 	case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1057 	case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1058 		ret = 0;
1059 		break;
1060 	case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1061 		ret = 1000;
1062 		break;
1063 	case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1064 	case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1065 		ret = div64_u64(dev_replace->cursor_left,
1066 				div_u64(btrfs_device_get_total_bytes(
1067 						dev_replace->srcdev), 1000));
1068 		break;
1069 	}
1070 
1071 	return ret;
1072 }
1073 
1074 void btrfs_dev_replace_status(struct btrfs_fs_info *fs_info,
1075 			      struct btrfs_ioctl_dev_replace_args *args)
1076 {
1077 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1078 
1079 	down_read(&dev_replace->rwsem);
1080 	/* even if !dev_replace_is_valid, the values are good enough for
1081 	 * the replace_status ioctl */
1082 	args->result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1083 	args->status.replace_state = dev_replace->replace_state;
1084 	args->status.time_started = dev_replace->time_started;
1085 	args->status.time_stopped = dev_replace->time_stopped;
1086 	args->status.num_write_errors =
1087 		atomic64_read(&dev_replace->num_write_errors);
1088 	args->status.num_uncorrectable_read_errors =
1089 		atomic64_read(&dev_replace->num_uncorrectable_read_errors);
1090 	args->status.progress_1000 = btrfs_dev_replace_progress(fs_info);
1091 	up_read(&dev_replace->rwsem);
1092 }
1093 
1094 int btrfs_dev_replace_cancel(struct btrfs_fs_info *fs_info)
1095 {
1096 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1097 	struct btrfs_device *tgt_device = NULL;
1098 	struct btrfs_device *src_device = NULL;
1099 	struct btrfs_trans_handle *trans;
1100 	struct btrfs_root *root = fs_info->tree_root;
1101 	int result;
1102 	int ret;
1103 
1104 	if (sb_rdonly(fs_info->sb))
1105 		return -EROFS;
1106 
1107 	mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1108 	down_write(&dev_replace->rwsem);
1109 	switch (dev_replace->replace_state) {
1110 	case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1111 	case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1112 	case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1113 		result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1114 		up_write(&dev_replace->rwsem);
1115 		break;
1116 	case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1117 		tgt_device = dev_replace->tgtdev;
1118 		src_device = dev_replace->srcdev;
1119 		up_write(&dev_replace->rwsem);
1120 		ret = btrfs_scrub_cancel(fs_info);
1121 		if (ret < 0) {
1122 			result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NOT_STARTED;
1123 		} else {
1124 			result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1125 			/*
1126 			 * btrfs_dev_replace_finishing() will handle the
1127 			 * cleanup part
1128 			 */
1129 			btrfs_info_in_rcu(fs_info,
1130 				"dev_replace from %s (devid %llu) to %s canceled",
1131 				btrfs_dev_name(src_device), src_device->devid,
1132 				btrfs_dev_name(tgt_device));
1133 		}
1134 		break;
1135 	case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1136 		/*
1137 		 * Scrub doing the replace isn't running so we need to do the
1138 		 * cleanup step of btrfs_dev_replace_finishing() here
1139 		 */
1140 		result = BTRFS_IOCTL_DEV_REPLACE_RESULT_NO_ERROR;
1141 		tgt_device = dev_replace->tgtdev;
1142 		src_device = dev_replace->srcdev;
1143 		dev_replace->tgtdev = NULL;
1144 		dev_replace->srcdev = NULL;
1145 		dev_replace->replace_state =
1146 				BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED;
1147 		dev_replace->time_stopped = ktime_get_real_seconds();
1148 		dev_replace->item_needs_writeback = 1;
1149 
1150 		up_write(&dev_replace->rwsem);
1151 
1152 		/* Scrub for replace must not be running in suspended state */
1153 		ret = btrfs_scrub_cancel(fs_info);
1154 		ASSERT(ret != -ENOTCONN);
1155 
1156 		trans = btrfs_start_transaction(root, 0);
1157 		if (IS_ERR(trans)) {
1158 			mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1159 			return PTR_ERR(trans);
1160 		}
1161 		ret = btrfs_commit_transaction(trans);
1162 		WARN_ON(ret);
1163 
1164 		btrfs_info_in_rcu(fs_info,
1165 		"suspended dev_replace from %s (devid %llu) to %s canceled",
1166 			btrfs_dev_name(src_device), src_device->devid,
1167 			btrfs_dev_name(tgt_device));
1168 
1169 		if (tgt_device)
1170 			btrfs_destroy_dev_replace_tgtdev(tgt_device);
1171 		break;
1172 	default:
1173 		up_write(&dev_replace->rwsem);
1174 		result = -EINVAL;
1175 	}
1176 
1177 	mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1178 	return result;
1179 }
1180 
1181 void btrfs_dev_replace_suspend_for_unmount(struct btrfs_fs_info *fs_info)
1182 {
1183 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1184 
1185 	mutex_lock(&dev_replace->lock_finishing_cancel_unmount);
1186 	down_write(&dev_replace->rwsem);
1187 
1188 	switch (dev_replace->replace_state) {
1189 	case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1190 	case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1191 	case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1192 	case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1193 		break;
1194 	case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1195 		dev_replace->replace_state =
1196 			BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1197 		dev_replace->time_stopped = ktime_get_real_seconds();
1198 		dev_replace->item_needs_writeback = 1;
1199 		btrfs_info(fs_info, "suspending dev_replace for unmount");
1200 		break;
1201 	}
1202 
1203 	up_write(&dev_replace->rwsem);
1204 	mutex_unlock(&dev_replace->lock_finishing_cancel_unmount);
1205 }
1206 
1207 /* resume dev_replace procedure that was interrupted by unmount */
1208 int btrfs_resume_dev_replace_async(struct btrfs_fs_info *fs_info)
1209 {
1210 	struct task_struct *task;
1211 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1212 
1213 	down_write(&dev_replace->rwsem);
1214 
1215 	switch (dev_replace->replace_state) {
1216 	case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1217 	case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1218 	case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1219 		up_write(&dev_replace->rwsem);
1220 		return 0;
1221 	case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1222 		break;
1223 	case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1224 		dev_replace->replace_state =
1225 			BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED;
1226 		break;
1227 	}
1228 	if (!dev_replace->tgtdev || !dev_replace->tgtdev->bdev) {
1229 		btrfs_info(fs_info,
1230 			   "cannot continue dev_replace, tgtdev is missing");
1231 		btrfs_info(fs_info,
1232 			   "you may cancel the operation after 'mount -o degraded'");
1233 		dev_replace->replace_state =
1234 					BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1235 		up_write(&dev_replace->rwsem);
1236 		return 0;
1237 	}
1238 	up_write(&dev_replace->rwsem);
1239 
1240 	/*
1241 	 * This could collide with a paused balance, but the exclusive op logic
1242 	 * should never allow both to start and pause. We don't want to allow
1243 	 * dev-replace to start anyway.
1244 	 */
1245 	if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_DEV_REPLACE)) {
1246 		down_write(&dev_replace->rwsem);
1247 		dev_replace->replace_state =
1248 					BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED;
1249 		up_write(&dev_replace->rwsem);
1250 		btrfs_info(fs_info,
1251 		"cannot resume dev-replace, other exclusive operation running");
1252 		return 0;
1253 	}
1254 
1255 	task = kthread_run(btrfs_dev_replace_kthread, fs_info, "btrfs-devrepl");
1256 	return PTR_ERR_OR_ZERO(task);
1257 }
1258 
1259 static int btrfs_dev_replace_kthread(void *data)
1260 {
1261 	struct btrfs_fs_info *fs_info = data;
1262 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1263 	u64 progress;
1264 	int ret;
1265 
1266 	progress = btrfs_dev_replace_progress(fs_info);
1267 	progress = div_u64(progress, 10);
1268 	btrfs_info_in_rcu(fs_info,
1269 		"continuing dev_replace from %s (devid %llu) to target %s @%u%%",
1270 		btrfs_dev_name(dev_replace->srcdev),
1271 		dev_replace->srcdev->devid,
1272 		btrfs_dev_name(dev_replace->tgtdev),
1273 		(unsigned int)progress);
1274 
1275 	ret = btrfs_scrub_dev(fs_info, dev_replace->srcdev->devid,
1276 			      dev_replace->committed_cursor_left,
1277 			      btrfs_device_get_total_bytes(dev_replace->srcdev),
1278 			      &dev_replace->scrub_progress, 0, 1);
1279 	ret = btrfs_dev_replace_finishing(fs_info, ret);
1280 	WARN_ON(ret && ret != -ECANCELED);
1281 
1282 	btrfs_exclop_finish(fs_info);
1283 	return 0;
1284 }
1285 
1286 int __pure btrfs_dev_replace_is_ongoing(struct btrfs_dev_replace *dev_replace)
1287 {
1288 	if (!dev_replace->is_valid)
1289 		return 0;
1290 
1291 	switch (dev_replace->replace_state) {
1292 	case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
1293 	case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
1294 	case BTRFS_IOCTL_DEV_REPLACE_STATE_CANCELED:
1295 		return 0;
1296 	case BTRFS_IOCTL_DEV_REPLACE_STATE_STARTED:
1297 	case BTRFS_IOCTL_DEV_REPLACE_STATE_SUSPENDED:
1298 		/*
1299 		 * return true even if tgtdev is missing (this is
1300 		 * something that can happen if the dev_replace
1301 		 * procedure is suspended by an umount and then
1302 		 * the tgtdev is missing (or "btrfs dev scan") was
1303 		 * not called and the filesystem is remounted
1304 		 * in degraded state. This does not stop the
1305 		 * dev_replace procedure. It needs to be canceled
1306 		 * manually if the cancellation is wanted.
1307 		 */
1308 		break;
1309 	}
1310 	return 1;
1311 }
1312 
1313 void btrfs_bio_counter_inc_noblocked(struct btrfs_fs_info *fs_info)
1314 {
1315 	percpu_counter_inc(&fs_info->dev_replace.bio_counter);
1316 }
1317 
1318 void btrfs_bio_counter_sub(struct btrfs_fs_info *fs_info, s64 amount)
1319 {
1320 	percpu_counter_sub(&fs_info->dev_replace.bio_counter, amount);
1321 	cond_wake_up_nomb(&fs_info->dev_replace.replace_wait);
1322 }
1323 
1324 void btrfs_bio_counter_inc_blocked(struct btrfs_fs_info *fs_info)
1325 {
1326 	while (1) {
1327 		percpu_counter_inc(&fs_info->dev_replace.bio_counter);
1328 		if (likely(!test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1329 				     &fs_info->fs_state)))
1330 			break;
1331 
1332 		btrfs_bio_counter_dec(fs_info);
1333 		wait_event(fs_info->dev_replace.replace_wait,
1334 			   !test_bit(BTRFS_FS_STATE_DEV_REPLACING,
1335 				     &fs_info->fs_state));
1336 	}
1337 }
1338