xref: /linux/fs/btrfs/tree-mod-log.c (revision 87c9c16317882dd6dbbc07e349bc3223e14f3244)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "tree-mod-log.h"
4 #include "disk-io.h"
5 
6 struct tree_mod_root {
7 	u64 logical;
8 	u8 level;
9 };
10 
11 struct tree_mod_elem {
12 	struct rb_node node;
13 	u64 logical;
14 	u64 seq;
15 	enum btrfs_mod_log_op op;
16 
17 	/*
18 	 * This is used for BTRFS_MOD_LOG_KEY_* and BTRFS_MOD_LOG_MOVE_KEYS
19 	 * operations.
20 	 */
21 	int slot;
22 
23 	/* This is used for BTRFS_MOD_LOG_KEY* and BTRFS_MOD_LOG_ROOT_REPLACE. */
24 	u64 generation;
25 
26 	/* Those are used for op == BTRFS_MOD_LOG_KEY_{REPLACE,REMOVE}. */
27 	struct btrfs_disk_key key;
28 	u64 blockptr;
29 
30 	/* This is used for op == BTRFS_MOD_LOG_MOVE_KEYS. */
31 	struct {
32 		int dst_slot;
33 		int nr_items;
34 	} move;
35 
36 	/* This is used for op == BTRFS_MOD_LOG_ROOT_REPLACE. */
37 	struct tree_mod_root old_root;
38 };
39 
40 /*
41  * Pull a new tree mod seq number for our operation.
42  */
43 static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
44 {
45 	return atomic64_inc_return(&fs_info->tree_mod_seq);
46 }
47 
48 /*
49  * This adds a new blocker to the tree mod log's blocker list if the @elem
50  * passed does not already have a sequence number set. So when a caller expects
51  * to record tree modifications, it should ensure to set elem->seq to zero
52  * before calling btrfs_get_tree_mod_seq.
53  * Returns a fresh, unused tree log modification sequence number, even if no new
54  * blocker was added.
55  */
56 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
57 			   struct btrfs_seq_list *elem)
58 {
59 	write_lock(&fs_info->tree_mod_log_lock);
60 	if (!elem->seq) {
61 		elem->seq = btrfs_inc_tree_mod_seq(fs_info);
62 		list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
63 		set_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags);
64 	}
65 	write_unlock(&fs_info->tree_mod_log_lock);
66 
67 	return elem->seq;
68 }
69 
70 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
71 			    struct btrfs_seq_list *elem)
72 {
73 	struct rb_root *tm_root;
74 	struct rb_node *node;
75 	struct rb_node *next;
76 	struct tree_mod_elem *tm;
77 	u64 min_seq = BTRFS_SEQ_LAST;
78 	u64 seq_putting = elem->seq;
79 
80 	if (!seq_putting)
81 		return;
82 
83 	write_lock(&fs_info->tree_mod_log_lock);
84 	list_del(&elem->list);
85 	elem->seq = 0;
86 
87 	if (list_empty(&fs_info->tree_mod_seq_list)) {
88 		clear_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags);
89 	} else {
90 		struct btrfs_seq_list *first;
91 
92 		first = list_first_entry(&fs_info->tree_mod_seq_list,
93 					 struct btrfs_seq_list, list);
94 		if (seq_putting > first->seq) {
95 			/*
96 			 * Blocker with lower sequence number exists, we cannot
97 			 * remove anything from the log.
98 			 */
99 			write_unlock(&fs_info->tree_mod_log_lock);
100 			return;
101 		}
102 		min_seq = first->seq;
103 	}
104 
105 	/*
106 	 * Anything that's lower than the lowest existing (read: blocked)
107 	 * sequence number can be removed from the tree.
108 	 */
109 	tm_root = &fs_info->tree_mod_log;
110 	for (node = rb_first(tm_root); node; node = next) {
111 		next = rb_next(node);
112 		tm = rb_entry(node, struct tree_mod_elem, node);
113 		if (tm->seq >= min_seq)
114 			continue;
115 		rb_erase(node, tm_root);
116 		kfree(tm);
117 	}
118 	write_unlock(&fs_info->tree_mod_log_lock);
119 }
120 
121 /*
122  * Key order of the log:
123  *       node/leaf start address -> sequence
124  *
125  * The 'start address' is the logical address of the *new* root node for root
126  * replace operations, or the logical address of the affected block for all
127  * other operations.
128  */
129 static noinline int tree_mod_log_insert(struct btrfs_fs_info *fs_info,
130 					struct tree_mod_elem *tm)
131 {
132 	struct rb_root *tm_root;
133 	struct rb_node **new;
134 	struct rb_node *parent = NULL;
135 	struct tree_mod_elem *cur;
136 
137 	lockdep_assert_held_write(&fs_info->tree_mod_log_lock);
138 
139 	tm->seq = btrfs_inc_tree_mod_seq(fs_info);
140 
141 	tm_root = &fs_info->tree_mod_log;
142 	new = &tm_root->rb_node;
143 	while (*new) {
144 		cur = rb_entry(*new, struct tree_mod_elem, node);
145 		parent = *new;
146 		if (cur->logical < tm->logical)
147 			new = &((*new)->rb_left);
148 		else if (cur->logical > tm->logical)
149 			new = &((*new)->rb_right);
150 		else if (cur->seq < tm->seq)
151 			new = &((*new)->rb_left);
152 		else if (cur->seq > tm->seq)
153 			new = &((*new)->rb_right);
154 		else
155 			return -EEXIST;
156 	}
157 
158 	rb_link_node(&tm->node, parent, new);
159 	rb_insert_color(&tm->node, tm_root);
160 	return 0;
161 }
162 
163 /*
164  * Determines if logging can be omitted. Returns true if it can. Otherwise, it
165  * returns false with the tree_mod_log_lock acquired. The caller must hold
166  * this until all tree mod log insertions are recorded in the rb tree and then
167  * write unlock fs_info::tree_mod_log_lock.
168  */
169 static inline bool tree_mod_dont_log(struct btrfs_fs_info *fs_info,
170 				    struct extent_buffer *eb)
171 {
172 	if (!test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
173 		return true;
174 	if (eb && btrfs_header_level(eb) == 0)
175 		return true;
176 
177 	write_lock(&fs_info->tree_mod_log_lock);
178 	if (list_empty(&(fs_info)->tree_mod_seq_list)) {
179 		write_unlock(&fs_info->tree_mod_log_lock);
180 		return true;
181 	}
182 
183 	return false;
184 }
185 
186 /* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
187 static inline bool tree_mod_need_log(const struct btrfs_fs_info *fs_info,
188 				    struct extent_buffer *eb)
189 {
190 	if (!test_bit(BTRFS_FS_TREE_MOD_LOG_USERS, &fs_info->flags))
191 		return false;
192 	if (eb && btrfs_header_level(eb) == 0)
193 		return false;
194 
195 	return true;
196 }
197 
198 static struct tree_mod_elem *alloc_tree_mod_elem(struct extent_buffer *eb,
199 						 int slot,
200 						 enum btrfs_mod_log_op op,
201 						 gfp_t flags)
202 {
203 	struct tree_mod_elem *tm;
204 
205 	tm = kzalloc(sizeof(*tm), flags);
206 	if (!tm)
207 		return NULL;
208 
209 	tm->logical = eb->start;
210 	if (op != BTRFS_MOD_LOG_KEY_ADD) {
211 		btrfs_node_key(eb, &tm->key, slot);
212 		tm->blockptr = btrfs_node_blockptr(eb, slot);
213 	}
214 	tm->op = op;
215 	tm->slot = slot;
216 	tm->generation = btrfs_node_ptr_generation(eb, slot);
217 	RB_CLEAR_NODE(&tm->node);
218 
219 	return tm;
220 }
221 
222 int btrfs_tree_mod_log_insert_key(struct extent_buffer *eb, int slot,
223 				  enum btrfs_mod_log_op op, gfp_t flags)
224 {
225 	struct tree_mod_elem *tm;
226 	int ret;
227 
228 	if (!tree_mod_need_log(eb->fs_info, eb))
229 		return 0;
230 
231 	tm = alloc_tree_mod_elem(eb, slot, op, flags);
232 	if (!tm)
233 		return -ENOMEM;
234 
235 	if (tree_mod_dont_log(eb->fs_info, eb)) {
236 		kfree(tm);
237 		return 0;
238 	}
239 
240 	ret = tree_mod_log_insert(eb->fs_info, tm);
241 	write_unlock(&eb->fs_info->tree_mod_log_lock);
242 	if (ret)
243 		kfree(tm);
244 
245 	return ret;
246 }
247 
248 int btrfs_tree_mod_log_insert_move(struct extent_buffer *eb,
249 				   int dst_slot, int src_slot,
250 				   int nr_items)
251 {
252 	struct tree_mod_elem *tm = NULL;
253 	struct tree_mod_elem **tm_list = NULL;
254 	int ret = 0;
255 	int i;
256 	bool locked = false;
257 
258 	if (!tree_mod_need_log(eb->fs_info, eb))
259 		return 0;
260 
261 	tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
262 	if (!tm_list)
263 		return -ENOMEM;
264 
265 	tm = kzalloc(sizeof(*tm), GFP_NOFS);
266 	if (!tm) {
267 		ret = -ENOMEM;
268 		goto free_tms;
269 	}
270 
271 	tm->logical = eb->start;
272 	tm->slot = src_slot;
273 	tm->move.dst_slot = dst_slot;
274 	tm->move.nr_items = nr_items;
275 	tm->op = BTRFS_MOD_LOG_MOVE_KEYS;
276 
277 	for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
278 		tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
279 				BTRFS_MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
280 		if (!tm_list[i]) {
281 			ret = -ENOMEM;
282 			goto free_tms;
283 		}
284 	}
285 
286 	if (tree_mod_dont_log(eb->fs_info, eb))
287 		goto free_tms;
288 	locked = true;
289 
290 	/*
291 	 * When we override something during the move, we log these removals.
292 	 * This can only happen when we move towards the beginning of the
293 	 * buffer, i.e. dst_slot < src_slot.
294 	 */
295 	for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
296 		ret = tree_mod_log_insert(eb->fs_info, tm_list[i]);
297 		if (ret)
298 			goto free_tms;
299 	}
300 
301 	ret = tree_mod_log_insert(eb->fs_info, tm);
302 	if (ret)
303 		goto free_tms;
304 	write_unlock(&eb->fs_info->tree_mod_log_lock);
305 	kfree(tm_list);
306 
307 	return 0;
308 
309 free_tms:
310 	for (i = 0; i < nr_items; i++) {
311 		if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
312 			rb_erase(&tm_list[i]->node, &eb->fs_info->tree_mod_log);
313 		kfree(tm_list[i]);
314 	}
315 	if (locked)
316 		write_unlock(&eb->fs_info->tree_mod_log_lock);
317 	kfree(tm_list);
318 	kfree(tm);
319 
320 	return ret;
321 }
322 
323 static inline int tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
324 				       struct tree_mod_elem **tm_list,
325 				       int nritems)
326 {
327 	int i, j;
328 	int ret;
329 
330 	for (i = nritems - 1; i >= 0; i--) {
331 		ret = tree_mod_log_insert(fs_info, tm_list[i]);
332 		if (ret) {
333 			for (j = nritems - 1; j > i; j--)
334 				rb_erase(&tm_list[j]->node,
335 					 &fs_info->tree_mod_log);
336 			return ret;
337 		}
338 	}
339 
340 	return 0;
341 }
342 
343 int btrfs_tree_mod_log_insert_root(struct extent_buffer *old_root,
344 				   struct extent_buffer *new_root,
345 				   bool log_removal)
346 {
347 	struct btrfs_fs_info *fs_info = old_root->fs_info;
348 	struct tree_mod_elem *tm = NULL;
349 	struct tree_mod_elem **tm_list = NULL;
350 	int nritems = 0;
351 	int ret = 0;
352 	int i;
353 
354 	if (!tree_mod_need_log(fs_info, NULL))
355 		return 0;
356 
357 	if (log_removal && btrfs_header_level(old_root) > 0) {
358 		nritems = btrfs_header_nritems(old_root);
359 		tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
360 				  GFP_NOFS);
361 		if (!tm_list) {
362 			ret = -ENOMEM;
363 			goto free_tms;
364 		}
365 		for (i = 0; i < nritems; i++) {
366 			tm_list[i] = alloc_tree_mod_elem(old_root, i,
367 			    BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
368 			if (!tm_list[i]) {
369 				ret = -ENOMEM;
370 				goto free_tms;
371 			}
372 		}
373 	}
374 
375 	tm = kzalloc(sizeof(*tm), GFP_NOFS);
376 	if (!tm) {
377 		ret = -ENOMEM;
378 		goto free_tms;
379 	}
380 
381 	tm->logical = new_root->start;
382 	tm->old_root.logical = old_root->start;
383 	tm->old_root.level = btrfs_header_level(old_root);
384 	tm->generation = btrfs_header_generation(old_root);
385 	tm->op = BTRFS_MOD_LOG_ROOT_REPLACE;
386 
387 	if (tree_mod_dont_log(fs_info, NULL))
388 		goto free_tms;
389 
390 	if (tm_list)
391 		ret = tree_mod_log_free_eb(fs_info, tm_list, nritems);
392 	if (!ret)
393 		ret = tree_mod_log_insert(fs_info, tm);
394 
395 	write_unlock(&fs_info->tree_mod_log_lock);
396 	if (ret)
397 		goto free_tms;
398 	kfree(tm_list);
399 
400 	return ret;
401 
402 free_tms:
403 	if (tm_list) {
404 		for (i = 0; i < nritems; i++)
405 			kfree(tm_list[i]);
406 		kfree(tm_list);
407 	}
408 	kfree(tm);
409 
410 	return ret;
411 }
412 
413 static struct tree_mod_elem *__tree_mod_log_search(struct btrfs_fs_info *fs_info,
414 						   u64 start, u64 min_seq,
415 						   bool smallest)
416 {
417 	struct rb_root *tm_root;
418 	struct rb_node *node;
419 	struct tree_mod_elem *cur = NULL;
420 	struct tree_mod_elem *found = NULL;
421 
422 	read_lock(&fs_info->tree_mod_log_lock);
423 	tm_root = &fs_info->tree_mod_log;
424 	node = tm_root->rb_node;
425 	while (node) {
426 		cur = rb_entry(node, struct tree_mod_elem, node);
427 		if (cur->logical < start) {
428 			node = node->rb_left;
429 		} else if (cur->logical > start) {
430 			node = node->rb_right;
431 		} else if (cur->seq < min_seq) {
432 			node = node->rb_left;
433 		} else if (!smallest) {
434 			/* We want the node with the highest seq */
435 			if (found)
436 				BUG_ON(found->seq > cur->seq);
437 			found = cur;
438 			node = node->rb_left;
439 		} else if (cur->seq > min_seq) {
440 			/* We want the node with the smallest seq */
441 			if (found)
442 				BUG_ON(found->seq < cur->seq);
443 			found = cur;
444 			node = node->rb_right;
445 		} else {
446 			found = cur;
447 			break;
448 		}
449 	}
450 	read_unlock(&fs_info->tree_mod_log_lock);
451 
452 	return found;
453 }
454 
455 /*
456  * This returns the element from the log with the smallest time sequence
457  * value that's in the log (the oldest log item). Any element with a time
458  * sequence lower than min_seq will be ignored.
459  */
460 static struct tree_mod_elem *tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info,
461 							u64 start, u64 min_seq)
462 {
463 	return __tree_mod_log_search(fs_info, start, min_seq, true);
464 }
465 
466 /*
467  * This returns the element from the log with the largest time sequence
468  * value that's in the log (the most recent log item). Any element with
469  * a time sequence lower than min_seq will be ignored.
470  */
471 static struct tree_mod_elem *tree_mod_log_search(struct btrfs_fs_info *fs_info,
472 						 u64 start, u64 min_seq)
473 {
474 	return __tree_mod_log_search(fs_info, start, min_seq, false);
475 }
476 
477 int btrfs_tree_mod_log_eb_copy(struct extent_buffer *dst,
478 			       struct extent_buffer *src,
479 			       unsigned long dst_offset,
480 			       unsigned long src_offset,
481 			       int nr_items)
482 {
483 	struct btrfs_fs_info *fs_info = dst->fs_info;
484 	int ret = 0;
485 	struct tree_mod_elem **tm_list = NULL;
486 	struct tree_mod_elem **tm_list_add, **tm_list_rem;
487 	int i;
488 	bool locked = false;
489 
490 	if (!tree_mod_need_log(fs_info, NULL))
491 		return 0;
492 
493 	if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
494 		return 0;
495 
496 	tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
497 			  GFP_NOFS);
498 	if (!tm_list)
499 		return -ENOMEM;
500 
501 	tm_list_add = tm_list;
502 	tm_list_rem = tm_list + nr_items;
503 	for (i = 0; i < nr_items; i++) {
504 		tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
505 		    BTRFS_MOD_LOG_KEY_REMOVE, GFP_NOFS);
506 		if (!tm_list_rem[i]) {
507 			ret = -ENOMEM;
508 			goto free_tms;
509 		}
510 
511 		tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
512 						BTRFS_MOD_LOG_KEY_ADD, GFP_NOFS);
513 		if (!tm_list_add[i]) {
514 			ret = -ENOMEM;
515 			goto free_tms;
516 		}
517 	}
518 
519 	if (tree_mod_dont_log(fs_info, NULL))
520 		goto free_tms;
521 	locked = true;
522 
523 	for (i = 0; i < nr_items; i++) {
524 		ret = tree_mod_log_insert(fs_info, tm_list_rem[i]);
525 		if (ret)
526 			goto free_tms;
527 		ret = tree_mod_log_insert(fs_info, tm_list_add[i]);
528 		if (ret)
529 			goto free_tms;
530 	}
531 
532 	write_unlock(&fs_info->tree_mod_log_lock);
533 	kfree(tm_list);
534 
535 	return 0;
536 
537 free_tms:
538 	for (i = 0; i < nr_items * 2; i++) {
539 		if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
540 			rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
541 		kfree(tm_list[i]);
542 	}
543 	if (locked)
544 		write_unlock(&fs_info->tree_mod_log_lock);
545 	kfree(tm_list);
546 
547 	return ret;
548 }
549 
550 int btrfs_tree_mod_log_free_eb(struct extent_buffer *eb)
551 {
552 	struct tree_mod_elem **tm_list = NULL;
553 	int nritems = 0;
554 	int i;
555 	int ret = 0;
556 
557 	if (!tree_mod_need_log(eb->fs_info, eb))
558 		return 0;
559 
560 	nritems = btrfs_header_nritems(eb);
561 	tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
562 	if (!tm_list)
563 		return -ENOMEM;
564 
565 	for (i = 0; i < nritems; i++) {
566 		tm_list[i] = alloc_tree_mod_elem(eb, i,
567 		    BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
568 		if (!tm_list[i]) {
569 			ret = -ENOMEM;
570 			goto free_tms;
571 		}
572 	}
573 
574 	if (tree_mod_dont_log(eb->fs_info, eb))
575 		goto free_tms;
576 
577 	ret = tree_mod_log_free_eb(eb->fs_info, tm_list, nritems);
578 	write_unlock(&eb->fs_info->tree_mod_log_lock);
579 	if (ret)
580 		goto free_tms;
581 	kfree(tm_list);
582 
583 	return 0;
584 
585 free_tms:
586 	for (i = 0; i < nritems; i++)
587 		kfree(tm_list[i]);
588 	kfree(tm_list);
589 
590 	return ret;
591 }
592 
593 /*
594  * Returns the logical address of the oldest predecessor of the given root.
595  * Entries older than time_seq are ignored.
596  */
597 static struct tree_mod_elem *tree_mod_log_oldest_root(struct extent_buffer *eb_root,
598 						      u64 time_seq)
599 {
600 	struct tree_mod_elem *tm;
601 	struct tree_mod_elem *found = NULL;
602 	u64 root_logical = eb_root->start;
603 	bool looped = false;
604 
605 	if (!time_seq)
606 		return NULL;
607 
608 	/*
609 	 * The very last operation that's logged for a root is the replacement
610 	 * operation (if it is replaced at all). This has the logical address
611 	 * of the *new* root, making it the very first operation that's logged
612 	 * for this root.
613 	 */
614 	while (1) {
615 		tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
616 						time_seq);
617 		if (!looped && !tm)
618 			return NULL;
619 		/*
620 		 * If there are no tree operation for the oldest root, we simply
621 		 * return it. This should only happen if that (old) root is at
622 		 * level 0.
623 		 */
624 		if (!tm)
625 			break;
626 
627 		/*
628 		 * If there's an operation that's not a root replacement, we
629 		 * found the oldest version of our root. Normally, we'll find a
630 		 * BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
631 		 */
632 		if (tm->op != BTRFS_MOD_LOG_ROOT_REPLACE)
633 			break;
634 
635 		found = tm;
636 		root_logical = tm->old_root.logical;
637 		looped = true;
638 	}
639 
640 	/* If there's no old root to return, return what we found instead */
641 	if (!found)
642 		found = tm;
643 
644 	return found;
645 }
646 
647 
648 /*
649  * tm is a pointer to the first operation to rewind within eb. Then, all
650  * previous operations will be rewound (until we reach something older than
651  * time_seq).
652  */
653 static void tree_mod_log_rewind(struct btrfs_fs_info *fs_info,
654 				struct extent_buffer *eb,
655 				u64 time_seq,
656 				struct tree_mod_elem *first_tm)
657 {
658 	u32 n;
659 	struct rb_node *next;
660 	struct tree_mod_elem *tm = first_tm;
661 	unsigned long o_dst;
662 	unsigned long o_src;
663 	unsigned long p_size = sizeof(struct btrfs_key_ptr);
664 
665 	n = btrfs_header_nritems(eb);
666 	read_lock(&fs_info->tree_mod_log_lock);
667 	while (tm && tm->seq >= time_seq) {
668 		/*
669 		 * All the operations are recorded with the operator used for
670 		 * the modification. As we're going backwards, we do the
671 		 * opposite of each operation here.
672 		 */
673 		switch (tm->op) {
674 		case BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING:
675 			BUG_ON(tm->slot < n);
676 			fallthrough;
677 		case BTRFS_MOD_LOG_KEY_REMOVE_WHILE_MOVING:
678 		case BTRFS_MOD_LOG_KEY_REMOVE:
679 			btrfs_set_node_key(eb, &tm->key, tm->slot);
680 			btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
681 			btrfs_set_node_ptr_generation(eb, tm->slot,
682 						      tm->generation);
683 			n++;
684 			break;
685 		case BTRFS_MOD_LOG_KEY_REPLACE:
686 			BUG_ON(tm->slot >= n);
687 			btrfs_set_node_key(eb, &tm->key, tm->slot);
688 			btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
689 			btrfs_set_node_ptr_generation(eb, tm->slot,
690 						      tm->generation);
691 			break;
692 		case BTRFS_MOD_LOG_KEY_ADD:
693 			/* if a move operation is needed it's in the log */
694 			n--;
695 			break;
696 		case BTRFS_MOD_LOG_MOVE_KEYS:
697 			o_dst = btrfs_node_key_ptr_offset(tm->slot);
698 			o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
699 			memmove_extent_buffer(eb, o_dst, o_src,
700 					      tm->move.nr_items * p_size);
701 			break;
702 		case BTRFS_MOD_LOG_ROOT_REPLACE:
703 			/*
704 			 * This operation is special. For roots, this must be
705 			 * handled explicitly before rewinding.
706 			 * For non-roots, this operation may exist if the node
707 			 * was a root: root A -> child B; then A gets empty and
708 			 * B is promoted to the new root. In the mod log, we'll
709 			 * have a root-replace operation for B, a tree block
710 			 * that is no root. We simply ignore that operation.
711 			 */
712 			break;
713 		}
714 		next = rb_next(&tm->node);
715 		if (!next)
716 			break;
717 		tm = rb_entry(next, struct tree_mod_elem, node);
718 		if (tm->logical != first_tm->logical)
719 			break;
720 	}
721 	read_unlock(&fs_info->tree_mod_log_lock);
722 	btrfs_set_header_nritems(eb, n);
723 }
724 
725 /*
726  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
727  * is returned. If rewind operations happen, a fresh buffer is returned. The
728  * returned buffer is always read-locked. If the returned buffer is not the
729  * input buffer, the lock on the input buffer is released and the input buffer
730  * is freed (its refcount is decremented).
731  */
732 struct extent_buffer *btrfs_tree_mod_log_rewind(struct btrfs_fs_info *fs_info,
733 						struct btrfs_path *path,
734 						struct extent_buffer *eb,
735 						u64 time_seq)
736 {
737 	struct extent_buffer *eb_rewin;
738 	struct tree_mod_elem *tm;
739 
740 	if (!time_seq)
741 		return eb;
742 
743 	if (btrfs_header_level(eb) == 0)
744 		return eb;
745 
746 	tm = tree_mod_log_search(fs_info, eb->start, time_seq);
747 	if (!tm)
748 		return eb;
749 
750 	if (tm->op == BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
751 		BUG_ON(tm->slot != 0);
752 		eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
753 		if (!eb_rewin) {
754 			btrfs_tree_read_unlock(eb);
755 			free_extent_buffer(eb);
756 			return NULL;
757 		}
758 		btrfs_set_header_bytenr(eb_rewin, eb->start);
759 		btrfs_set_header_backref_rev(eb_rewin,
760 					     btrfs_header_backref_rev(eb));
761 		btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
762 		btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
763 	} else {
764 		eb_rewin = btrfs_clone_extent_buffer(eb);
765 		if (!eb_rewin) {
766 			btrfs_tree_read_unlock(eb);
767 			free_extent_buffer(eb);
768 			return NULL;
769 		}
770 	}
771 
772 	btrfs_tree_read_unlock(eb);
773 	free_extent_buffer(eb);
774 
775 	btrfs_set_buffer_lockdep_class(btrfs_header_owner(eb_rewin),
776 				       eb_rewin, btrfs_header_level(eb_rewin));
777 	btrfs_tree_read_lock(eb_rewin);
778 	tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
779 	WARN_ON(btrfs_header_nritems(eb_rewin) >
780 		BTRFS_NODEPTRS_PER_BLOCK(fs_info));
781 
782 	return eb_rewin;
783 }
784 
785 /*
786  * Rewind the state of @root's root node to the given @time_seq value.
787  * If there are no changes, the current root->root_node is returned. If anything
788  * changed in between, there's a fresh buffer allocated on which the rewind
789  * operations are done. In any case, the returned buffer is read locked.
790  * Returns NULL on error (with no locks held).
791  */
792 struct extent_buffer *btrfs_get_old_root(struct btrfs_root *root, u64 time_seq)
793 {
794 	struct btrfs_fs_info *fs_info = root->fs_info;
795 	struct tree_mod_elem *tm;
796 	struct extent_buffer *eb = NULL;
797 	struct extent_buffer *eb_root;
798 	u64 eb_root_owner = 0;
799 	struct extent_buffer *old;
800 	struct tree_mod_root *old_root = NULL;
801 	u64 old_generation = 0;
802 	u64 logical;
803 	int level;
804 
805 	eb_root = btrfs_read_lock_root_node(root);
806 	tm = tree_mod_log_oldest_root(eb_root, time_seq);
807 	if (!tm)
808 		return eb_root;
809 
810 	if (tm->op == BTRFS_MOD_LOG_ROOT_REPLACE) {
811 		old_root = &tm->old_root;
812 		old_generation = tm->generation;
813 		logical = old_root->logical;
814 		level = old_root->level;
815 	} else {
816 		logical = eb_root->start;
817 		level = btrfs_header_level(eb_root);
818 	}
819 
820 	tm = tree_mod_log_search(fs_info, logical, time_seq);
821 	if (old_root && tm && tm->op != BTRFS_MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
822 		btrfs_tree_read_unlock(eb_root);
823 		free_extent_buffer(eb_root);
824 		old = read_tree_block(fs_info, logical, root->root_key.objectid,
825 				      0, level, NULL);
826 		if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
827 			if (!IS_ERR(old))
828 				free_extent_buffer(old);
829 			btrfs_warn(fs_info,
830 				   "failed to read tree block %llu from get_old_root",
831 				   logical);
832 		} else {
833 			struct tree_mod_elem *tm2;
834 
835 			btrfs_tree_read_lock(old);
836 			eb = btrfs_clone_extent_buffer(old);
837 			/*
838 			 * After the lookup for the most recent tree mod operation
839 			 * above and before we locked and cloned the extent buffer
840 			 * 'old', a new tree mod log operation may have been added.
841 			 * So lookup for a more recent one to make sure the number
842 			 * of mod log operations we replay is consistent with the
843 			 * number of items we have in the cloned extent buffer,
844 			 * otherwise we can hit a BUG_ON when rewinding the extent
845 			 * buffer.
846 			 */
847 			tm2 = tree_mod_log_search(fs_info, logical, time_seq);
848 			btrfs_tree_read_unlock(old);
849 			free_extent_buffer(old);
850 			ASSERT(tm2);
851 			ASSERT(tm2 == tm || tm2->seq > tm->seq);
852 			if (!tm2 || tm2->seq < tm->seq) {
853 				free_extent_buffer(eb);
854 				return NULL;
855 			}
856 			tm = tm2;
857 		}
858 	} else if (old_root) {
859 		eb_root_owner = btrfs_header_owner(eb_root);
860 		btrfs_tree_read_unlock(eb_root);
861 		free_extent_buffer(eb_root);
862 		eb = alloc_dummy_extent_buffer(fs_info, logical);
863 	} else {
864 		eb = btrfs_clone_extent_buffer(eb_root);
865 		btrfs_tree_read_unlock(eb_root);
866 		free_extent_buffer(eb_root);
867 	}
868 
869 	if (!eb)
870 		return NULL;
871 	if (old_root) {
872 		btrfs_set_header_bytenr(eb, eb->start);
873 		btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
874 		btrfs_set_header_owner(eb, eb_root_owner);
875 		btrfs_set_header_level(eb, old_root->level);
876 		btrfs_set_header_generation(eb, old_generation);
877 	}
878 	btrfs_set_buffer_lockdep_class(btrfs_header_owner(eb), eb,
879 				       btrfs_header_level(eb));
880 	btrfs_tree_read_lock(eb);
881 	if (tm)
882 		tree_mod_log_rewind(fs_info, eb, time_seq, tm);
883 	else
884 		WARN_ON(btrfs_header_level(eb) != 0);
885 	WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
886 
887 	return eb;
888 }
889 
890 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
891 {
892 	struct tree_mod_elem *tm;
893 	int level;
894 	struct extent_buffer *eb_root = btrfs_root_node(root);
895 
896 	tm = tree_mod_log_oldest_root(eb_root, time_seq);
897 	if (tm && tm->op == BTRFS_MOD_LOG_ROOT_REPLACE)
898 		level = tm->old_root.level;
899 	else
900 		level = btrfs_header_level(eb_root);
901 
902 	free_extent_buffer(eb_root);
903 
904 	return level;
905 }
906 
907 /*
908  * Return the lowest sequence number in the tree modification log.
909  *
910  * Return the sequence number of the oldest tree modification log user, which
911  * corresponds to the lowest sequence number of all existing users. If there are
912  * no users it returns 0.
913  */
914 u64 btrfs_tree_mod_log_lowest_seq(struct btrfs_fs_info *fs_info)
915 {
916 	u64 ret = 0;
917 
918 	read_lock(&fs_info->tree_mod_log_lock);
919 	if (!list_empty(&fs_info->tree_mod_seq_list)) {
920 		struct btrfs_seq_list *elem;
921 
922 		elem = list_first_entry(&fs_info->tree_mod_seq_list,
923 					struct btrfs_seq_list, list);
924 		ret = elem->seq;
925 	}
926 	read_unlock(&fs_info->tree_mod_log_lock);
927 
928 	return ret;
929 }
930