aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorYan, Zheng <zheng.yan@oracle.com>2010-05-16 10:49:59 -0400
committerChris Mason <chris.mason@oracle.com>2010-05-25 10:34:54 -0400
commit3fd0a5585eb98e074fb9934549c8d85c49756c0d (patch)
tree3e7ff9bd9678a5eea62818a2f4a50e19dda91a81
parentefa56464562991b8c24f965199888806bd8c4b38 (diff)
Btrfs: Metadata ENOSPC handling for balance
This patch adds metadata ENOSPC handling for the balance code. It is consisted by following major changes: 1. Avoid COW tree leave in the phrase of merging tree. 2. Handle interaction with snapshot creation. 3. make the backref cache can live across transactions. Signed-off-by: Yan Zheng <zheng.yan@oracle.com> Signed-off-by: Chris Mason <chris.mason@oracle.com>
Diffstat
-rw-r--r--fs/btrfs/ctree.c3
-rw-r--r--fs/btrfs/ctree.h11
-rw-r--r--fs/btrfs/extent-tree.c23
-rw-r--r--fs/btrfs/relocation.c1867
-rw-r--r--fs/btrfs/transaction.c6
5 files changed, 1163 insertions, 747 deletions
diff --git a/fs/btrfs/ctree.c b/fs/btrfs/ctree.c
index 6bee8e5204fb..acd532af8e55 100644
--- a/fs/btrfs/ctree.c
+++ b/fs/btrfs/ctree.c
@@ -447,6 +447,9 @@ static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
447 447
448 update_ref_for_cow(trans, root, buf, cow, &last_ref); 448 update_ref_for_cow(trans, root, buf, cow, &last_ref);
449 449
450 if (root->ref_cows)
451 btrfs_reloc_cow_block(trans, root, buf, cow);
452
450 if (buf == root->node) { 453 if (buf == root->node) {
451 WARN_ON(parent && parent != buf); 454 WARN_ON(parent && parent != buf);
452 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID || 455 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
index 65530837d04b..5ed0223d1cbe 100644
--- a/fs/btrfs/ctree.h
+++ b/fs/btrfs/ctree.h
@@ -2205,7 +2205,8 @@ static inline int btrfs_insert_empty_item(struct btrfs_trans_handle *trans,
2205int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path); 2205int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path);
2206int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path); 2206int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path);
2207int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf); 2207int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf);
2208int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref); 2208int btrfs_drop_snapshot(struct btrfs_root *root,
2209 struct btrfs_block_rsv *block_rsv, int update_ref);
2209int btrfs_drop_subtree(struct btrfs_trans_handle *trans, 2210int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
2210 struct btrfs_root *root, 2211 struct btrfs_root *root,
2211 struct extent_buffer *node, 2212 struct extent_buffer *node,
@@ -2479,4 +2480,12 @@ int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
2479 struct btrfs_root *root); 2480 struct btrfs_root *root);
2480int btrfs_recover_relocation(struct btrfs_root *root); 2481int btrfs_recover_relocation(struct btrfs_root *root);
2481int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len); 2482int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len);
2483void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
2484 struct btrfs_root *root, struct extent_buffer *buf,
2485 struct extent_buffer *cow);
2486void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
2487 struct btrfs_pending_snapshot *pending,
2488 u64 *bytes_to_reserve);
2489void btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
2490 struct btrfs_pending_snapshot *pending);
2482#endif 2491#endif
diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
index a713f69f0c7a..d61a799fe323 100644
--- a/fs/btrfs/extent-tree.c
+++ b/fs/btrfs/extent-tree.c
@@ -5875,7 +5875,8 @@ static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
5875 * also make sure backrefs for the shared block and all lower level 5875 * also make sure backrefs for the shared block and all lower level
5876 * blocks are properly updated. 5876 * blocks are properly updated.
5877 */ 5877 */
5878int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref) 5878int btrfs_drop_snapshot(struct btrfs_root *root,
5879 struct btrfs_block_rsv *block_rsv, int update_ref)
5879{ 5880{
5880 struct btrfs_path *path; 5881 struct btrfs_path *path;
5881 struct btrfs_trans_handle *trans; 5882 struct btrfs_trans_handle *trans;
@@ -5894,6 +5895,8 @@ int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref)
5894 BUG_ON(!wc); 5895 BUG_ON(!wc);
5895 5896
5896 trans = btrfs_start_transaction(tree_root, 0); 5897 trans = btrfs_start_transaction(tree_root, 0);
5898 if (block_rsv)
5899 trans->block_rsv = block_rsv;
5897 5900
5898 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 5901 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
5899 level = btrfs_header_level(root->node); 5902 level = btrfs_header_level(root->node);
@@ -5981,24 +5984,16 @@ int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref)
5981 } 5984 }
5982 5985
5983 BUG_ON(wc->level == 0); 5986 BUG_ON(wc->level == 0);
5984 if (trans->transaction->in_commit || 5987 if (btrfs_should_end_transaction(trans, tree_root)) {
5985 trans->transaction->delayed_refs.flushing) {
5986 ret = btrfs_update_root(trans, tree_root, 5988 ret = btrfs_update_root(trans, tree_root,
5987 &root->root_key, 5989 &root->root_key,
5988 root_item); 5990 root_item);
5989 BUG_ON(ret); 5991 BUG_ON(ret);
5990 5992
5991 btrfs_end_transaction(trans, tree_root); 5993 btrfs_end_transaction_throttle(trans, tree_root);
5992 trans = btrfs_start_transaction(tree_root, 0); 5994 trans = btrfs_start_transaction(tree_root, 0);
5993 if (IS_ERR(trans)) 5995 if (block_rsv)
5994 return PTR_ERR(trans); 5996 trans->block_rsv = block_rsv;
5995 } else {
5996 unsigned long update;
5997 update = trans->delayed_ref_updates;
5998 trans->delayed_ref_updates = 0;
5999 if (update)
6000 btrfs_run_delayed_refs(trans, tree_root,
6001 update);
6002 } 5997 }
6003 } 5998 }
6004 btrfs_release_path(root, path); 5999 btrfs_release_path(root, path);
@@ -6026,7 +6021,7 @@ int btrfs_drop_snapshot(struct btrfs_root *root, int update_ref)
6026 kfree(root); 6021 kfree(root);
6027 } 6022 }
6028out: 6023out:
6029 btrfs_end_transaction(trans, tree_root); 6024 btrfs_end_transaction_throttle(trans, tree_root);
6030 kfree(wc); 6025 kfree(wc);
6031 btrfs_free_path(path); 6026 btrfs_free_path(path);
6032 return err; 6027 return err;
diff --git a/fs/btrfs/relocation.c b/fs/btrfs/relocation.c
index 3943526b7348..05d41e569236 100644
--- a/fs/btrfs/relocation.c
+++ b/fs/btrfs/relocation.c
@@ -44,8 +44,12 @@ struct tree_entry {
44struct backref_node { 44struct backref_node {
45 struct rb_node rb_node; 45 struct rb_node rb_node;
46 u64 bytenr; 46 u64 bytenr;
47 /* objectid tree block owner */ 47
48 u64 new_bytenr;
49 /* objectid of tree block owner, can be not uptodate */
48 u64 owner; 50 u64 owner;
51 /* link to pending, changed or detached list */
52 struct list_head list;
49 /* list of upper level blocks reference this block */ 53 /* list of upper level blocks reference this block */
50 struct list_head upper; 54 struct list_head upper;
51 /* list of child blocks in the cache */ 55 /* list of child blocks in the cache */
@@ -56,9 +60,9 @@ struct backref_node {
56 struct extent_buffer *eb; 60 struct extent_buffer *eb;
57 /* level of tree block */ 61 /* level of tree block */
58 unsigned int level:8; 62 unsigned int level:8;
59 /* 1 if the block is root of old snapshot */ 63 /* is the block in non-reference counted tree */
60 unsigned int old_root:1; 64 unsigned int cowonly:1;
61 /* 1 if no child blocks in the cache */ 65 /* 1 if no child node in the cache */
62 unsigned int lowest:1; 66 unsigned int lowest:1;
63 /* is the extent buffer locked */ 67 /* is the extent buffer locked */
64 unsigned int locked:1; 68 unsigned int locked:1;
@@ -66,6 +70,16 @@ struct backref_node {
66 unsigned int processed:1; 70 unsigned int processed:1;
67 /* have backrefs of this block been checked */ 71 /* have backrefs of this block been checked */
68 unsigned int checked:1; 72 unsigned int checked:1;
73 /*
74 * 1 if corresponding block has been cowed but some upper
75 * level block pointers may not point to the new location
76 */
77 unsigned int pending:1;
78 /*
79 * 1 if the backref node isn't connected to any other
80 * backref node.
81 */
82 unsigned int detached:1;
69}; 83};
70 84
71/* 85/*
@@ -74,7 +88,6 @@ struct backref_node {
74struct backref_edge { 88struct backref_edge {
75 struct list_head list[2]; 89 struct list_head list[2];
76 struct backref_node *node[2]; 90 struct backref_node *node[2];
77 u64 blockptr;
78}; 91};
79 92
80#define LOWER 0 93#define LOWER 0
@@ -83,9 +96,25 @@ struct backref_edge {
83struct backref_cache { 96struct backref_cache {
84 /* red black tree of all backref nodes in the cache */ 97 /* red black tree of all backref nodes in the cache */
85 struct rb_root rb_root; 98 struct rb_root rb_root;
86 /* list of backref nodes with no child block in the cache */ 99 /* for passing backref nodes to btrfs_reloc_cow_block */
100 struct backref_node *path[BTRFS_MAX_LEVEL];
101 /*
102 * list of blocks that have been cowed but some block
103 * pointers in upper level blocks may not reflect the
104 * new location
105 */
87 struct list_head pending[BTRFS_MAX_LEVEL]; 106 struct list_head pending[BTRFS_MAX_LEVEL];
88 spinlock_t lock; 107 /* list of backref nodes with no child node */
108 struct list_head leaves;
109 /* list of blocks that have been cowed in current transaction */
110 struct list_head changed;
111 /* list of detached backref node. */
112 struct list_head detached;
113
114 u64 last_trans;
115
116 int nr_nodes;
117 int nr_edges;
89}; 118};
90 119
91/* 120/*
@@ -113,15 +142,6 @@ struct tree_block {
113 unsigned int key_ready:1; 142 unsigned int key_ready:1;
114}; 143};
115 144
116/* inode vector */
117#define INODEVEC_SIZE 16
118
119struct inodevec {
120 struct list_head list;
121 struct inode *inode[INODEVEC_SIZE];
122 int nr;
123};
124
125#define MAX_EXTENTS 128 145#define MAX_EXTENTS 128
126 146
127struct file_extent_cluster { 147struct file_extent_cluster {
@@ -138,36 +158,43 @@ struct reloc_control {
138 struct btrfs_root *extent_root; 158 struct btrfs_root *extent_root;
139 /* inode for moving data */ 159 /* inode for moving data */
140 struct inode *data_inode; 160 struct inode *data_inode;
141 struct btrfs_workers workers; 161
162 struct btrfs_block_rsv *block_rsv;
163
164 struct backref_cache backref_cache;
165
166 struct file_extent_cluster cluster;
142 /* tree blocks have been processed */ 167 /* tree blocks have been processed */
143 struct extent_io_tree processed_blocks; 168 struct extent_io_tree processed_blocks;
144 /* map start of tree root to corresponding reloc tree */ 169 /* map start of tree root to corresponding reloc tree */
145 struct mapping_tree reloc_root_tree; 170 struct mapping_tree reloc_root_tree;
146 /* list of reloc trees */ 171 /* list of reloc trees */
147 struct list_head reloc_roots; 172 struct list_head reloc_roots;
173 /* size of metadata reservation for merging reloc trees */
174 u64 merging_rsv_size;
175 /* size of relocated tree nodes */
176 u64 nodes_relocated;
177
148 u64 search_start; 178 u64 search_start;
149 u64 extents_found; 179 u64 extents_found;
150 u64 extents_skipped; 180
151 int stage; 181 int block_rsv_retries;
152 int create_reloc_root; 182
183 unsigned int stage:8;
184 unsigned int create_reloc_tree:1;
185 unsigned int merge_reloc_tree:1;
153 unsigned int found_file_extent:1; 186 unsigned int found_file_extent:1;
154 unsigned int found_old_snapshot:1; 187 unsigned int commit_transaction:1;
155}; 188};
156 189
157/* stages of data relocation */ 190/* stages of data relocation */
158#define MOVE_DATA_EXTENTS 0 191#define MOVE_DATA_EXTENTS 0
159#define UPDATE_DATA_PTRS 1 192#define UPDATE_DATA_PTRS 1
160 193
161/* 194static void remove_backref_node(struct backref_cache *cache,
162 * merge reloc tree to corresponding fs tree in worker threads 195 struct backref_node *node);
163 */ 196static void __mark_block_processed(struct reloc_control *rc,
164struct async_merge { 197 struct backref_node *node);
165 struct btrfs_work work;
166 struct reloc_control *rc;
167 struct btrfs_root *root;
168 struct completion *done;
169 atomic_t *num_pending;
170};
171 198
172static void mapping_tree_init(struct mapping_tree *tree) 199static void mapping_tree_init(struct mapping_tree *tree)
173{ 200{
@@ -181,15 +208,80 @@ static void backref_cache_init(struct backref_cache *cache)
181 cache->rb_root = RB_ROOT; 208 cache->rb_root = RB_ROOT;
182 for (i = 0; i < BTRFS_MAX_LEVEL; i++) 209 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
183 INIT_LIST_HEAD(&cache->pending[i]); 210 INIT_LIST_HEAD(&cache->pending[i]);
184 spin_lock_init(&cache->lock); 211 INIT_LIST_HEAD(&cache->changed);
212 INIT_LIST_HEAD(&cache->detached);
213 INIT_LIST_HEAD(&cache->leaves);
185} 214}
186 215
187static void backref_node_init(struct backref_node *node) 216static void backref_cache_cleanup(struct backref_cache *cache)
188{ 217{
189 memset(node, 0, sizeof(*node)); 218 struct backref_node *node;
190 INIT_LIST_HEAD(&node->upper); 219 int i;
191 INIT_LIST_HEAD(&node->lower); 220
192 RB_CLEAR_NODE(&node->rb_node); 221 while (!list_empty(&cache->detached)) {
222 node = list_entry(cache->detached.next,
223 struct backref_node, list);
224 remove_backref_node(cache, node);
225 }
226
227 while (!list_empty(&cache->leaves)) {
228 node = list_entry(cache->leaves.next,
229 struct backref_node, lower);
230 remove_backref_node(cache, node);
231 }
232
233 cache->last_trans = 0;
234
235 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
236 BUG_ON(!list_empty(&cache->pending[i]));
237 BUG_ON(!list_empty(&cache->changed));
238 BUG_ON(!list_empty(&cache->detached));
239 BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
240 BUG_ON(cache->nr_nodes);
241 BUG_ON(cache->nr_edges);
242}
243
244static struct backref_node *alloc_backref_node(struct backref_cache *cache)
245{
246 struct backref_node *node;
247
248 node = kzalloc(sizeof(*node), GFP_NOFS);
249 if (node) {
250 INIT_LIST_HEAD(&node->list);
251 INIT_LIST_HEAD(&node->upper);
252 INIT_LIST_HEAD(&node->lower);
253 RB_CLEAR_NODE(&node->rb_node);
254 cache->nr_nodes++;
255 }
256 return node;
257}
258
259static void free_backref_node(struct backref_cache *cache,
260 struct backref_node *node)
261{
262 if (node) {
263 cache->nr_nodes--;
264 kfree(node);
265 }
266}
267
268static struct backref_edge *alloc_backref_edge(struct backref_cache *cache)
269{
270 struct backref_edge *edge;
271
272 edge = kzalloc(sizeof(*edge), GFP_NOFS);
273 if (edge)
274 cache->nr_edges++;
275 return edge;
276}
277
278static void free_backref_edge(struct backref_cache *cache,
279 struct backref_edge *edge)
280{
281 if (edge) {
282 cache->nr_edges--;
283 kfree(edge);
284 }
193} 285}
194 286
195static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr, 287static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
@@ -250,6 +342,7 @@ static struct backref_node *walk_up_backref(struct backref_node *node,
250 edges[idx++] = edge; 342 edges[idx++] = edge;
251 node = edge->node[UPPER]; 343 node = edge->node[UPPER];
252 } 344 }
345 BUG_ON(node->detached);
253 *index = idx; 346 *index = idx;
254 return node; 347 return node;
255} 348}
@@ -281,13 +374,18 @@ static struct backref_node *walk_down_backref(struct backref_edge *edges[],
281 return NULL; 374 return NULL;
282} 375}
283 376
377static void unlock_node_buffer(struct backref_node *node)
378{
379 if (node->locked) {
380 btrfs_tree_unlock(node->eb);
381 node->locked = 0;
382 }
383}
384
284static void drop_node_buffer(struct backref_node *node) 385static void drop_node_buffer(struct backref_node *node)
285{ 386{
286 if (node->eb) { 387 if (node->eb) {
287 if (node->locked) { 388 unlock_node_buffer(node);
288 btrfs_tree_unlock(node->eb);
289 node->locked = 0;
290 }
291 free_extent_buffer(node->eb); 389 free_extent_buffer(node->eb);
292 node->eb = NULL; 390 node->eb = NULL;
293 } 391 }
@@ -296,14 +394,14 @@ static void drop_node_buffer(struct backref_node *node)
296static void drop_backref_node(struct backref_cache *tree, 394static void drop_backref_node(struct backref_cache *tree,
297 struct backref_node *node) 395 struct backref_node *node)
298{ 396{
299 BUG_ON(!node->lowest);
300 BUG_ON(!list_empty(&node->upper)); 397 BUG_ON(!list_empty(&node->upper));
301 398
302 drop_node_buffer(node); 399 drop_node_buffer(node);
400 list_del(&node->list);
303 list_del(&node->lower); 401 list_del(&node->lower);
304 402 if (!RB_EMPTY_NODE(&node->rb_node))
305 rb_erase(&node->rb_node, &tree->rb_root); 403 rb_erase(&node->rb_node, &tree->rb_root);
306 kfree(node); 404 free_backref_node(tree, node);
307} 405}
308 406
309/* 407/*
@@ -318,27 +416,121 @@ static void remove_backref_node(struct backref_cache *cache,
318 if (!node) 416 if (!node)
319 return; 417 return;
320 418
321 BUG_ON(!node->lowest); 419 BUG_ON(!node->lowest && !node->detached);
322 while (!list_empty(&node->upper)) { 420 while (!list_empty(&node->upper)) {
323 edge = list_entry(node->upper.next, struct backref_edge, 421 edge = list_entry(node->upper.next, struct backref_edge,
324 list[LOWER]); 422 list[LOWER]);
325 upper = edge->node[UPPER]; 423 upper = edge->node[UPPER];
326 list_del(&edge->list[LOWER]); 424 list_del(&edge->list[LOWER]);
327 list_del(&edge->list[UPPER]); 425 list_del(&edge->list[UPPER]);
328 kfree(edge); 426 free_backref_edge(cache, edge);
427
428 if (RB_EMPTY_NODE(&upper->rb_node)) {
429 BUG_ON(!list_empty(&node->upper));
430 drop_backref_node(cache, node);
431 node = upper;
432 node->lowest = 1;
433 continue;
434 }
329 /* 435 /*
330 * add the node to pending list if no other 436 * add the node to leaf node list if no other
331 * child block cached. 437 * child block cached.
332 */ 438 */
333 if (list_empty(&upper->lower)) { 439 if (list_empty(&upper->lower)) {
334 list_add_tail(&upper->lower, 440 list_add_tail(&upper->lower, &cache->leaves);
335 &cache->pending[upper->level]);
336 upper->lowest = 1; 441 upper->lowest = 1;
337 } 442 }
338 } 443 }
444
339 drop_backref_node(cache, node); 445 drop_backref_node(cache, node);
340} 446}
341 447
448static void update_backref_node(struct backref_cache *cache,
449 struct backref_node *node, u64 bytenr)
450{
451 struct rb_node *rb_node;
452 rb_erase(&node->rb_node, &cache->rb_root);
453 node->bytenr = bytenr;
454 rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node);
455 BUG_ON(rb_node);
456}
457
458/*
459 * update backref cache after a transaction commit
460 */
461static int update_backref_cache(struct btrfs_trans_handle *trans,
462 struct backref_cache *cache)
463{
464 struct backref_node *node;
465 int level = 0;
466
467 if (cache->last_trans == 0) {
468 cache->last_trans = trans->transid;
469 return 0;
470 }
471
472 if (cache->last_trans == trans->transid)
473 return 0;
474
475 /*
476 * detached nodes are used to avoid unnecessary backref
477 * lookup. transaction commit changes the extent tree.
478 * so the detached nodes are no longer useful.
479 */
480 while (!list_empty(&cache->detached)) {
481 node = list_entry(cache->detached.next,
482 struct backref_node, list);
483 remove_backref_node(cache, node);
484 }
485
486 while (!list_empty(&cache->changed)) {
487 node = list_entry(cache->changed.next,
488 struct backref_node, list);
489 list_del_init(&node->list);
490 BUG_ON(node->pending);
491 update_backref_node(cache, node, node->new_bytenr);
492 }
493
494 /*
495 * some nodes can be left in the pending list if there were
496 * errors during processing the pending nodes.
497 */
498 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
499 list_for_each_entry(node, &cache->pending[level], list) {
500 BUG_ON(!node->pending);
501 if (node->bytenr == node->new_bytenr)
502 continue;
503 update_backref_node(cache, node, node->new_bytenr);
504 }
505 }
506
507 cache->last_trans = 0;
508 return 1;
509}
510
511static int should_ignore_root(struct btrfs_root *root)
512{
513 struct btrfs_root *reloc_root;
514
515 if (!root->ref_cows)
516 return 0;
517
518 reloc_root = root->reloc_root;
519 if (!reloc_root)
520 return 0;
521
522 if (btrfs_root_last_snapshot(&reloc_root->root_item) ==
523 root->fs_info->running_transaction->transid - 1)
524 return 0;
525 /*
526 * if there is reloc tree and it was created in previous
527 * transaction backref lookup can find the reloc tree,
528 * so backref node for the fs tree root is useless for
529 * relocation.
530 */
531 return 1;
532}
533
342/* 534/*
343 * find reloc tree by address of tree root 535 * find reloc tree by address of tree root
344 */ 536 */
@@ -453,11 +645,12 @@ int find_inline_backref(struct extent_buffer *leaf, int slot,
453 * for all upper level blocks that directly/indirectly reference the 645 * for all upper level blocks that directly/indirectly reference the
454 * block are also cached. 646 * block are also cached.
455 */ 647 */
456static struct backref_node *build_backref_tree(struct reloc_control *rc, 648static noinline_for_stack
457 struct backref_cache *cache, 649struct backref_node *build_backref_tree(struct reloc_control *rc,
458 struct btrfs_key *node_key, 650 struct btrfs_key *node_key,
459 int level, u64 bytenr) 651 int level, u64 bytenr)
460{ 652{
653 struct backref_cache *cache = &rc->backref_cache;
461 struct btrfs_path *path1; 654 struct btrfs_path *path1;
462 struct btrfs_path *path2; 655 struct btrfs_path *path2;
463 struct extent_buffer *eb; 656 struct extent_buffer *eb;
@@ -473,6 +666,8 @@ static struct backref_node *build_backref_tree(struct reloc_control *rc,
473 unsigned long end; 666 unsigned long end;
474 unsigned long ptr; 667 unsigned long ptr;
475 LIST_HEAD(list); 668 LIST_HEAD(list);
669 LIST_HEAD(useless);
670 int cowonly;
476 int ret; 671 int ret;
477 int err = 0; 672 int err = 0;
478 673
@@ -483,15 +678,13 @@ static struct backref_node *build_backref_tree(struct reloc_control *rc,
483 goto out; 678 goto out;
484 } 679 }
485 680
486 node = kmalloc(sizeof(*node), GFP_NOFS); 681 node = alloc_backref_node(cache);
487 if (!node) { 682 if (!node) {
488 err = -ENOMEM; 683 err = -ENOMEM;
489 goto out; 684 goto out;
490 } 685 }
491 686
492 backref_node_init(node);
493 node->bytenr = bytenr; 687 node->bytenr = bytenr;
494 node->owner = 0;
495 node->level = level; 688 node->level = level;
496 node->lowest = 1; 689 node->lowest = 1;
497 cur = node; 690 cur = node;
@@ -587,17 +780,20 @@ again:
587#ifdef BTRFS_COMPAT_EXTENT_TREE_V0 780#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
588 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY || 781 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
589 key.type == BTRFS_EXTENT_REF_V0_KEY) { 782 key.type == BTRFS_EXTENT_REF_V0_KEY) {
590 if (key.objectid == key.offset && 783 if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
591 key.type == BTRFS_EXTENT_REF_V0_KEY) {
592 struct btrfs_extent_ref_v0 *ref0; 784 struct btrfs_extent_ref_v0 *ref0;
593 ref0 = btrfs_item_ptr(eb, path1->slots[0], 785 ref0 = btrfs_item_ptr(eb, path1->slots[0],
594 struct btrfs_extent_ref_v0); 786 struct btrfs_extent_ref_v0);
595 root = find_tree_root(rc, eb, ref0); 787 root = find_tree_root(rc, eb, ref0);
596 if (root) 788 if (!root->ref_cows)
597 cur->root = root; 789 cur->cowonly = 1;
598 else 790 if (key.objectid == key.offset) {
599 cur->old_root = 1; 791 if (root && !should_ignore_root(root))
600 break; 792 cur->root = root;
793 else
794 list_add(&cur->list, &useless);
795 break;
796 }
601 } 797 }
602#else 798#else
603 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 799 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
@@ -614,22 +810,20 @@ again:
614 break; 810 break;
615 } 811 }
616 812
617 edge = kzalloc(sizeof(*edge), GFP_NOFS); 813 edge = alloc_backref_edge(cache);
618 if (!edge) { 814 if (!edge) {
619 err = -ENOMEM; 815 err = -ENOMEM;
620 goto out; 816 goto out;
621 } 817 }
622 rb_node = tree_search(&cache->rb_root, key.offset); 818 rb_node = tree_search(&cache->rb_root, key.offset);
623 if (!rb_node) { 819 if (!rb_node) {
624 upper = kmalloc(sizeof(*upper), GFP_NOFS); 820 upper = alloc_backref_node(cache);
625 if (!upper) { 821 if (!upper) {
626 kfree(edge); 822 free_backref_edge(cache, edge);
627 err = -ENOMEM; 823 err = -ENOMEM;
628 goto out; 824 goto out;
629 } 825 }
630 backref_node_init(upper);
631 upper->bytenr = key.offset; 826 upper->bytenr = key.offset;
632 upper->owner = 0;
633 upper->level = cur->level + 1; 827 upper->level = cur->level + 1;
634 /* 828 /*
635 * backrefs for the upper level block isn't 829 * backrefs for the upper level block isn't
@@ -639,11 +833,12 @@ again:
639 } else { 833 } else {
640 upper = rb_entry(rb_node, struct backref_node, 834 upper = rb_entry(rb_node, struct backref_node,
641 rb_node); 835 rb_node);
836 BUG_ON(!upper->checked);
642 INIT_LIST_HEAD(&edge->list[UPPER]); 837 INIT_LIST_HEAD(&edge->list[UPPER]);
643 } 838 }
644 list_add(&edge->list[LOWER], &cur->upper); 839 list_add_tail(&edge->list[LOWER], &cur->upper);
645 edge->node[UPPER] = upper;
646 edge->node[LOWER] = cur; 840 edge->node[LOWER] = cur;
841 edge->node[UPPER] = upper;
647 842
648 goto next; 843 goto next;
649 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 844 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
@@ -657,11 +852,17 @@ again:
657 goto out; 852 goto out;
658 } 853 }
659 854
855 if (!root->ref_cows)
856 cur->cowonly = 1;
857
660 if (btrfs_root_level(&root->root_item) == cur->level) { 858 if (btrfs_root_level(&root->root_item) == cur->level) {
661 /* tree root */ 859 /* tree root */
662 BUG_ON(btrfs_root_bytenr(&root->root_item) != 860 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
663 cur->bytenr); 861 cur->bytenr);
664 cur->root = root; 862 if (should_ignore_root(root))
863 list_add(&cur->list, &useless);
864 else
865 cur->root = root;
665 break; 866 break;
666 } 867 }
667 868
@@ -692,11 +893,14 @@ again:
692 if (!path2->nodes[level]) { 893 if (!path2->nodes[level]) {
693 BUG_ON(btrfs_root_bytenr(&root->root_item) != 894 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
694 lower->bytenr); 895 lower->bytenr);
695 lower->root = root; 896 if (should_ignore_root(root))
897 list_add(&lower->list, &useless);
898 else
899 lower->root = root;
696 break; 900 break;
697 } 901 }
698 902
699 edge = kzalloc(sizeof(*edge), GFP_NOFS); 903 edge = alloc_backref_edge(cache);
700 if (!edge) { 904 if (!edge) {
701 err = -ENOMEM; 905 err = -ENOMEM;
702 goto out; 906 goto out;
@@ -705,16 +909,17 @@ again:
705 eb = path2->nodes[level]; 909 eb = path2->nodes[level];
706 rb_node = tree_search(&cache->rb_root, eb->start); 910 rb_node = tree_search(&cache->rb_root, eb->start);
707 if (!rb_node) { 911 if (!rb_node) {
708 upper = kmalloc(sizeof(*upper), GFP_NOFS); 912 upper = alloc_backref_node(cache);
709 if (!upper) { 913 if (!upper) {
710 kfree(edge); 914 free_backref_edge(cache, edge);
711 err = -ENOMEM; 915 err = -ENOMEM;
712 goto out; 916 goto out;
713 } 917 }
714 backref_node_init(upper);
715 upper->bytenr = eb->start; 918 upper->bytenr = eb->start;
716 upper->owner = btrfs_header_owner(eb); 919 upper->owner = btrfs_header_owner(eb);
717 upper->level = lower->level + 1; 920 upper->level = lower->level + 1;
921 if (!root->ref_cows)
922 upper->cowonly = 1;
718 923
719 /* 924 /*
720 * if we know the block isn't shared 925 * if we know the block isn't shared
@@ -744,10 +949,12 @@ again:
744 rb_node); 949 rb_node);
745 BUG_ON(!upper->checked); 950 BUG_ON(!upper->checked);
746 INIT_LIST_HEAD(&edge->list[UPPER]); 951 INIT_LIST_HEAD(&edge->list[UPPER]);
952 if (!upper->owner)
953 upper->owner = btrfs_header_owner(eb);
747 } 954 }
748 list_add_tail(&edge->list[LOWER], &lower->upper); 955 list_add_tail(&edge->list[LOWER], &lower->upper);
749 edge->node[UPPER] = upper;
750 edge->node[LOWER] = lower; 956 edge->node[LOWER] = lower;
957 edge->node[UPPER] = upper;
751 958
752 if (rb_node) 959 if (rb_node)
753 break; 960 break;
@@ -785,8 +992,13 @@ next:
785 * into the cache. 992 * into the cache.
786 */ 993 */
787 BUG_ON(!node->checked); 994 BUG_ON(!node->checked);
788 rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node); 995 cowonly = node->cowonly;
789 BUG_ON(rb_node); 996 if (!cowonly) {
997 rb_node = tree_insert(&cache->rb_root, node->bytenr,
998 &node->rb_node);
999 BUG_ON(rb_node);
1000 list_add_tail(&node->lower, &cache->leaves);
1001 }
790 1002
791 list_for_each_entry(edge, &node->upper, list[LOWER]) 1003 list_for_each_entry(edge, &node->upper, list[LOWER])
792 list_add_tail(&edge->list[UPPER], &list); 1004 list_add_tail(&edge->list[UPPER], &list);
@@ -795,6 +1007,14 @@ next:
795 edge = list_entry(list.next, struct backref_edge, list[UPPER]); 1007 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
796 list_del_init(&edge->list[UPPER]); 1008 list_del_init(&edge->list[UPPER]);
797 upper = edge->node[UPPER]; 1009 upper = edge->node[UPPER];
1010 if (upper->detached) {
1011 list_del(&edge->list[LOWER]);
1012 lower = edge->node[LOWER];
1013 free_backref_edge(cache, edge);
1014 if (list_empty(&lower->upper))
1015 list_add(&lower->list, &useless);
1016 continue;
1017 }
798 1018
799 if (!RB_EMPTY_NODE(&upper->rb_node)) { 1019 if (!RB_EMPTY_NODE(&upper->rb_node)) {
800 if (upper->lowest) { 1020 if (upper->lowest) {
@@ -807,25 +1027,69 @@ next:
807 } 1027 }
808 1028
809 BUG_ON(!upper->checked); 1029 BUG_ON(!upper->checked);
810 rb_node = tree_insert(&cache->rb_root, upper->bytenr, 1030 BUG_ON(cowonly != upper->cowonly);
811 &upper->rb_node); 1031 if (!cowonly) {
812 BUG_ON(rb_node); 1032 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1033 &upper->rb_node);
1034 BUG_ON(rb_node);
1035 }
813 1036
814 list_add_tail(&edge->list[UPPER], &upper->lower); 1037 list_add_tail(&edge->list[UPPER], &upper->lower);
815 1038
816 list_for_each_entry(edge, &upper->upper, list[LOWER]) 1039 list_for_each_entry(edge, &upper->upper, list[LOWER])
817 list_add_tail(&edge->list[UPPER], &list); 1040 list_add_tail(&edge->list[UPPER], &list);
818 } 1041 }
1042 /*
1043 * process useless backref nodes. backref nodes for tree leaves
1044 * are deleted from the cache. backref nodes for upper level
1045 * tree blocks are left in the cache to avoid unnecessary backref
1046 * lookup.
1047 */
1048 while (!list_empty(&useless)) {
1049 upper = list_entry(useless.next, struct backref_node, list);
1050 list_del_init(&upper->list);
1051 BUG_ON(!list_empty(&upper->upper));
1052 if (upper == node)
1053 node = NULL;
1054 if (upper->lowest) {
1055 list_del_init(&upper->lower);
1056 upper->lowest = 0;
1057 }
1058 while (!list_empty(&upper->lower)) {
1059 edge = list_entry(upper->lower.next,
1060 struct backref_edge, list[UPPER]);
1061 list_del(&edge->list[UPPER]);
1062 list_del(&edge->list[LOWER]);
1063 lower = edge->node[LOWER];
1064 free_backref_edge(cache, edge);
1065
1066 if (list_empty(&lower->upper))
1067 list_add(&lower->list, &useless);
1068 }
1069 __mark_block_processed(rc, upper);
1070 if (upper->level > 0) {
1071 list_add(&upper->list, &cache->detached);
1072 upper->detached = 1;
1073 } else {
1074 rb_erase(&upper->rb_node, &cache->rb_root);
1075 free_backref_node(cache, upper);
1076 }
1077 }
819out: 1078out:
820 btrfs_free_path(path1); 1079 btrfs_free_path(path1);
821 btrfs_free_path(path2); 1080 btrfs_free_path(path2);
822 if (err) { 1081 if (err) {
823 INIT_LIST_HEAD(&list); 1082 while (!list_empty(&useless)) {
1083 lower = list_entry(useless.next,
1084 struct backref_node, upper);
1085 list_del_init(&lower->upper);
1086 }
824 upper = node; 1087 upper = node;
1088 INIT_LIST_HEAD(&list);
825 while (upper) { 1089 while (upper) {
826 if (RB_EMPTY_NODE(&upper->rb_node)) { 1090 if (RB_EMPTY_NODE(&upper->rb_node)) {
827 list_splice_tail(&upper->upper, &list); 1091 list_splice_tail(&upper->upper, &list);
828 kfree(upper); 1092 free_backref_node(cache, upper);
829 } 1093 }
830 1094
831 if (list_empty(&list)) 1095 if (list_empty(&list))
@@ -833,15 +1097,104 @@ out:
833 1097
834 edge = list_entry(list.next, struct backref_edge, 1098 edge = list_entry(list.next, struct backref_edge,
835 list[LOWER]); 1099 list[LOWER]);
1100 list_del(&edge->list[LOWER]);
836 upper = edge->node[UPPER]; 1101 upper = edge->node[UPPER];
837 kfree(edge); 1102 free_backref_edge(cache, edge);
838 } 1103 }
839 return ERR_PTR(err); 1104 return ERR_PTR(err);
840 } 1105 }
1106 BUG_ON(node && node->detached);
841 return node; 1107 return node;
842} 1108}
843 1109
844/* 1110/*
1111 * helper to add backref node for the newly created snapshot.
1112 * the backref node is created by cloning backref node that
1113 * corresponds to root of source tree
1114 */
1115static int clone_backref_node(struct btrfs_trans_handle *trans,
1116 struct reloc_control *rc,
1117 struct btrfs_root *src,
1118 struct btrfs_root *dest)
1119{
1120 struct btrfs_root *reloc_root = src->reloc_root;
1121 struct backref_cache *cache = &rc->backref_cache;
1122 struct backref_node *node = NULL;
1123 struct backref_node *new_node;
1124 struct backref_edge *edge;
1125 struct backref_edge *new_edge;
1126 struct rb_node *rb_node;
1127
1128 if (cache->last_trans > 0)
1129 update_backref_cache(trans, cache);
1130
1131 rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1132 if (rb_node) {
1133 node = rb_entry(rb_node, struct backref_node, rb_node);
1134 if (node->detached)
1135 node = NULL;
1136 else
1137 BUG_ON(node->new_bytenr != reloc_root->node->start);
1138 }
1139
1140 if (!node) {
1141 rb_node = tree_search(&cache->rb_root,
1142 reloc_root->commit_root->start);
1143 if (rb_node) {
1144 node = rb_entry(rb_node, struct backref_node,
1145 rb_node);
1146 BUG_ON(node->detached);
1147 }
1148 }
1149
1150 if (!node)
1151 return 0;
1152
1153 new_node = alloc_backref_node(cache);
1154 if (!new_node)
1155 return -ENOMEM;
1156
1157 new_node->bytenr = dest->node->start;
1158 new_node->level = node->level;
1159 new_node->lowest = node->lowest;
1160 new_node->root = dest;
1161
1162 if (!node->lowest) {
1163 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1164 new_edge = alloc_backref_edge(cache);
1165 if (!new_edge)
1166 goto fail;
1167
1168 new_edge->node[UPPER] = new_node;
1169 new_edge->node[LOWER] = edge->node[LOWER];
1170 list_add_tail(&new_edge->list[UPPER],
1171 &new_node->lower);
1172 }
1173 }
1174
1175 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1176 &new_node->rb_node);
1177 BUG_ON(rb_node);
1178
1179 if (!new_node->lowest) {
1180 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1181 list_add_tail(&new_edge->list[LOWER],
1182 &new_edge->node[LOWER]->upper);
1183 }
1184 }
1185 return 0;
1186fail:
1187 while (!list_empty(&new_node->lower)) {
1188 new_edge = list_entry(new_node->lower.next,
1189 struct backref_edge, list[UPPER]);
1190 list_del(&new_edge->list[UPPER]);
1191 free_backref_edge(cache, new_edge);
1192 }
1193 free_backref_node(cache, new_node);
1194 return -ENOMEM;
1195}
1196
1197/*
845 * helper to add 'address of tree root -> reloc tree' mapping 1198 * helper to add 'address of tree root -> reloc tree' mapping
846 */ 1199 */
847static int __add_reloc_root(struct btrfs_root *root) 1200static int __add_reloc_root(struct btrfs_root *root)
@@ -901,12 +1254,8 @@ static int __update_reloc_root(struct btrfs_root *root, int del)
901 return 0; 1254 return 0;
902} 1255}
903 1256
904/* 1257static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
905 * create reloc tree for a given fs tree. reloc tree is just a 1258 struct btrfs_root *root, u64 objectid)
906 * snapshot of the fs tree with special root objectid.
907 */
908int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
909 struct btrfs_root *root)
910{ 1259{
911 struct btrfs_root *reloc_root; 1260 struct btrfs_root *reloc_root;
912 struct extent_buffer *eb; 1261 struct extent_buffer *eb;
@@ -914,36 +1263,45 @@ int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
914 struct btrfs_key root_key; 1263 struct btrfs_key root_key;
915 int ret; 1264 int ret;
916 1265
917 if (root->reloc_root) {
918 reloc_root = root->reloc_root;
919 reloc_root->last_trans = trans->transid;
920 return 0;
921 }
922
923 if (!root->fs_info->reloc_ctl ||
924 !root->fs_info->reloc_ctl->create_reloc_root ||
925 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
926 return 0;
927
928 root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 1266 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
929 BUG_ON(!root_item); 1267 BUG_ON(!root_item);
930 1268
931 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 1269 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
932 root_key.type = BTRFS_ROOT_ITEM_KEY; 1270 root_key.type = BTRFS_ROOT_ITEM_KEY;
933 root_key.offset = root->root_key.objectid; 1271 root_key.offset = objectid;
934 1272
935 ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 1273 if (root->root_key.objectid == objectid) {
936 BTRFS_TREE_RELOC_OBJECTID); 1274 /* called by btrfs_init_reloc_root */
937 BUG_ON(ret); 1275 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1276 BTRFS_TREE_RELOC_OBJECTID);
1277 BUG_ON(ret);
1278
1279 btrfs_set_root_last_snapshot(&root->root_item,
1280 trans->transid - 1);
1281 } else {
1282 /*
1283 * called by btrfs_reloc_post_snapshot_hook.
1284 * the source tree is a reloc tree, all tree blocks
1285 * modified after it was created have RELOC flag
1286 * set in their headers. so it's OK to not update
1287 * the 'last_snapshot'.
1288 */
1289 ret = btrfs_copy_root(trans, root, root->node, &eb,
1290 BTRFS_TREE_RELOC_OBJECTID);
1291 BUG_ON(ret);
1292 }
938 1293
939 btrfs_set_root_last_snapshot(&root->root_item, trans->transid - 1);
940 memcpy(root_item, &root->root_item, sizeof(*root_item)); 1294 memcpy(root_item, &root->root_item, sizeof(*root_item));
941 btrfs_set_root_refs(root_item, 1);
942 btrfs_set_root_bytenr(root_item, eb->start); 1295 btrfs_set_root_bytenr(root_item, eb->start);
943 btrfs_set_root_level(root_item, btrfs_header_level(eb)); 1296 btrfs_set_root_level(root_item, btrfs_header_level(eb));
944 btrfs_set_root_generation(root_item, trans->transid); 1297 btrfs_set_root_generation(root_item, trans->transid);
945 memset(&root_item->drop_progress, 0, sizeof(struct btrfs_disk_key)); 1298
946 root_item->drop_level = 0; 1299 if (root->root_key.objectid == objectid) {
1300 btrfs_set_root_refs(root_item, 0);
1301 memset(&root_item->drop_progress, 0,
1302 sizeof(struct btrfs_disk_key));
1303 root_item->drop_level = 0;
1304 }
947 1305
948 btrfs_tree_unlock(eb); 1306 btrfs_tree_unlock(eb);
949 free_extent_buffer(eb); 1307 free_extent_buffer(eb);
@@ -957,6 +1315,37 @@ int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
957 &root_key); 1315 &root_key);
958 BUG_ON(IS_ERR(reloc_root)); 1316 BUG_ON(IS_ERR(reloc_root));
959 reloc_root->last_trans = trans->transid; 1317 reloc_root->last_trans = trans->transid;
1318 return reloc_root;
1319}
1320
1321/*
1322 * create reloc tree for a given fs tree. reloc tree is just a
1323 * snapshot of the fs tree with special root objectid.
1324 */
1325int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1326 struct btrfs_root *root)
1327{
1328 struct btrfs_root *reloc_root;
1329 struct reloc_control *rc = root->fs_info->reloc_ctl;
1330 int clear_rsv = 0;
1331
1332 if (root->reloc_root) {
1333 reloc_root = root->reloc_root;
1334 reloc_root->last_trans = trans->transid;
1335 return 0;
1336 }
1337
1338 if (!rc || !rc->create_reloc_tree ||
1339 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1340 return 0;
1341
1342 if (!trans->block_rsv) {
1343 trans->block_rsv = rc->block_rsv;
1344 clear_rsv = 1;
1345 }
1346 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1347 if (clear_rsv)
1348 trans->block_rsv = NULL;
960 1349
961 __add_reloc_root(reloc_root); 1350 __add_reloc_root(reloc_root);
962 root->reloc_root = reloc_root; 1351 root->reloc_root = reloc_root;
@@ -980,7 +1369,8 @@ int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
980 reloc_root = root->reloc_root; 1369 reloc_root = root->reloc_root;
981 root_item = &reloc_root->root_item; 1370 root_item = &reloc_root->root_item;
982 1371
983 if (btrfs_root_refs(root_item) == 0) { 1372 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1373 btrfs_root_refs(root_item) == 0) {
984 root->reloc_root = NULL; 1374 root->reloc_root = NULL;
985 del = 1; 1375 del = 1;
986 } 1376 }
@@ -1102,8 +1492,7 @@ static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1102 goto out; 1492 goto out;
1103 } 1493 }
1104 1494
1105 if (new_bytenr) 1495 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1106 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1107 ret = 0; 1496 ret = 0;
1108out: 1497out:
1109 btrfs_free_path(path); 1498 btrfs_free_path(path);
@@ -1114,19 +1503,18 @@ out:
1114 * update file extent items in the tree leaf to point to 1503 * update file extent items in the tree leaf to point to
1115 * the new locations. 1504 * the new locations.
1116 */ 1505 */
1117static int replace_file_extents(struct btrfs_trans_handle *trans, 1506static noinline_for_stack
1118 struct reloc_control *rc, 1507int replace_file_extents(struct btrfs_trans_handle *trans,
1119 struct btrfs_root *root, 1508 struct reloc_control *rc,
1120 struct extent_buffer *leaf, 1509 struct btrfs_root *root,
1121 struct list_head *inode_list) 1510 struct extent_buffer *leaf)
1122{ 1511{
1123 struct btrfs_key key; 1512 struct btrfs_key key;
1124 struct btrfs_file_extent_item *fi; 1513 struct btrfs_file_extent_item *fi;
1125 struct inode *inode = NULL; 1514 struct inode *inode = NULL;
1126 struct inodevec *ivec = NULL;
1127 u64 parent; 1515 u64 parent;
1128 u64 bytenr; 1516 u64 bytenr;
1129 u64 new_bytenr; 1517 u64 new_bytenr = 0;
1130 u64 num_bytes; 1518 u64 num_bytes;
1131 u64 end; 1519 u64 end;
1132 u32 nritems; 1520 u32 nritems;
@@ -1166,21 +1554,12 @@ static int replace_file_extents(struct btrfs_trans_handle *trans,
1166 * to complete and drop the extent cache 1554 * to complete and drop the extent cache
1167 */ 1555 */
1168 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 1556 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1169 if (!ivec || ivec->nr == INODEVEC_SIZE) {
1170 ivec = kmalloc(sizeof(*ivec), GFP_NOFS);
1171 BUG_ON(!ivec);
1172 ivec->nr = 0;
1173 list_add_tail(&ivec->list, inode_list);
1174 }
1175 if (first) { 1557 if (first) {
1176 inode = find_next_inode(root, key.objectid); 1558 inode = find_next_inode(root, key.objectid);
1177 if (inode)
1178 ivec->inode[ivec->nr++] = inode;
1179 first = 0; 1559 first = 0;
1180 } else if (inode && inode->i_ino < key.objectid) { 1560 } else if (inode && inode->i_ino < key.objectid) {
1561 btrfs_add_delayed_iput(inode);
1181 inode = find_next_inode(root, key.objectid); 1562 inode = find_next_inode(root, key.objectid);
1182 if (inode)
1183 ivec->inode[ivec->nr++] = inode;
1184 } 1563 }
1185 if (inode && inode->i_ino == key.objectid) { 1564 if (inode && inode->i_ino == key.objectid) {
1186 end = key.offset + 1565 end = key.offset +
@@ -1204,8 +1583,10 @@ static int replace_file_extents(struct btrfs_trans_handle *trans,
1204 1583
1205 ret = get_new_location(rc->data_inode, &new_bytenr, 1584 ret = get_new_location(rc->data_inode, &new_bytenr,
1206 bytenr, num_bytes); 1585 bytenr, num_bytes);
1207 if (ret > 0) 1586 if (ret > 0) {
1587 WARN_ON(1);
1208 continue; 1588 continue;
1589 }
1209 BUG_ON(ret < 0); 1590 BUG_ON(ret < 0);
1210 1591
1211 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 1592 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
@@ -1225,6 +1606,8 @@ static int replace_file_extents(struct btrfs_trans_handle *trans,
1225 } 1606 }
1226 if (dirty) 1607 if (dirty)
1227 btrfs_mark_buffer_dirty(leaf); 1608 btrfs_mark_buffer_dirty(leaf);
1609 if (inode)
1610 btrfs_add_delayed_iput(inode);
1228 return 0; 1611 return 0;
1229} 1612}
1230 1613
@@ -1248,11 +1631,11 @@ int memcmp_node_keys(struct extent_buffer *eb, int slot,
1248 * if no block got replaced, 0 is returned. if there are other 1631 * if no block got replaced, 0 is returned. if there are other
1249 * errors, a negative error number is returned. 1632 * errors, a negative error number is returned.
1250 */ 1633 */
1251static int replace_path(struct btrfs_trans_handle *trans, 1634static noinline_for_stack
1252 struct btrfs_root *dest, struct btrfs_root *src, 1635int replace_path(struct btrfs_trans_handle *trans,
1253 struct btrfs_path *path, struct btrfs_key *next_key, 1636 struct btrfs_root *dest, struct btrfs_root *src,
1254 struct extent_buffer **leaf, 1637 struct btrfs_path *path, struct btrfs_key *next_key,
1255 int lowest_level, int max_level) 1638 int lowest_level, int max_level)
1256{ 1639{
1257 struct extent_buffer *eb; 1640 struct extent_buffer *eb;
1258 struct extent_buffer *parent; 1641 struct extent_buffer *parent;
@@ -1263,16 +1646,16 @@ static int replace_path(struct btrfs_trans_handle *trans,
1263 u64 new_ptr_gen; 1646 u64 new_ptr_gen;
1264 u64 last_snapshot; 1647 u64 last_snapshot;
1265 u32 blocksize; 1648 u32 blocksize;
1649 int cow = 0;
1266 int level; 1650 int level;
1267 int ret; 1651 int ret;
1268 int slot; 1652 int slot;
1269 1653
1270 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 1654 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1271 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 1655 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1272 BUG_ON(lowest_level > 1 && leaf);
1273 1656
1274 last_snapshot = btrfs_root_last_snapshot(&src->root_item); 1657 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1275 1658again:
1276 slot = path->slots[lowest_level]; 1659 slot = path->slots[lowest_level];
1277 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 1660 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1278 1661
@@ -1286,8 +1669,10 @@ static int replace_path(struct btrfs_trans_handle *trans,
1286 return 0; 1669 return 0;
1287 } 1670 }
1288 1671
1289 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 1672 if (cow) {
1290 BUG_ON(ret); 1673 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1674 BUG_ON(ret);
1675 }
1291 btrfs_set_lock_blocking(eb); 1676 btrfs_set_lock_blocking(eb);
1292 1677
1293 if (next_key) { 1678 if (next_key) {
@@ -1331,7 +1716,7 @@ static int replace_path(struct btrfs_trans_handle *trans,
1331 1716
1332 if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 1717 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1333 memcmp_node_keys(parent, slot, path, level)) { 1718 memcmp_node_keys(parent, slot, path, level)) {
1334 if (level <= lowest_level && !leaf) { 1719 if (level <= lowest_level) {
1335 ret = 0; 1720 ret = 0;
1336 break; 1721 break;
1337 } 1722 }
@@ -1339,16 +1724,12 @@ static int replace_path(struct btrfs_trans_handle *trans,
1339 eb = read_tree_block(dest, old_bytenr, blocksize, 1724 eb = read_tree_block(dest, old_bytenr, blocksize,
1340 old_ptr_gen); 1725 old_ptr_gen);
1341 btrfs_tree_lock(eb); 1726 btrfs_tree_lock(eb);
1342 ret = btrfs_cow_block(trans, dest, eb, parent, 1727 if (cow) {
1343 slot, &eb); 1728 ret = btrfs_cow_block(trans, dest, eb, parent,
1344 BUG_ON(ret); 1729 slot, &eb);
1345 btrfs_set_lock_blocking(eb); 1730 BUG_ON(ret);
1346
1347 if (level <= lowest_level) {
1348 *leaf = eb;
1349 ret = 0;
1350 break;
1351 } 1731 }
1732 btrfs_set_lock_blocking(eb);
1352 1733
1353 btrfs_tree_unlock(parent); 1734 btrfs_tree_unlock(parent);
1354 free_extent_buffer(parent); 1735 free_extent_buffer(parent);
@@ -1357,6 +1738,13 @@ static int replace_path(struct btrfs_trans_handle *trans,
1357 continue; 1738 continue;
1358 } 1739 }
1359 1740
1741 if (!cow) {
1742 btrfs_tree_unlock(parent);
1743 free_extent_buffer(parent);
1744 cow = 1;
1745 goto again;
1746 }
1747
1360 btrfs_node_key_to_cpu(path->nodes[level], &key, 1748 btrfs_node_key_to_cpu(path->nodes[level], &key,
1361 path->slots[level]); 1749 path->slots[level]);
1362 btrfs_release_path(src, path); 1750 btrfs_release_path(src, path);
@@ -1562,20 +1950,6 @@ static int invalidate_extent_cache(struct btrfs_root *root,
1562 return 0; 1950 return 0;
1563} 1951}
1564 1952
1565static void put_inodes(struct list_head *list)
1566{
1567 struct inodevec *ivec;
1568 while (!list_empty(list)) {
1569 ivec = list_entry(list->next, struct inodevec, list);
1570 list_del(&ivec->list);
1571 while (ivec->nr > 0) {
1572 ivec->nr--;
1573 iput(ivec->inode[ivec->nr]);
1574 }
1575 kfree(ivec);
1576 }
1577}
1578
1579static int find_next_key(struct btrfs_path *path, int level, 1953static int find_next_key(struct btrfs_path *path, int level,
1580 struct btrfs_key *key) 1954 struct btrfs_key *key)
1581 1955
@@ -1608,13 +1982,14 @@ static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1608 struct btrfs_root *reloc_root; 1982 struct btrfs_root *reloc_root;
1609 struct btrfs_root_item *root_item; 1983 struct btrfs_root_item *root_item;
1610 struct btrfs_path *path; 1984 struct btrfs_path *path;
1611 struct extent_buffer *leaf = NULL; 1985 struct extent_buffer *leaf;
1612 unsigned long nr; 1986 unsigned long nr;
1613 int level; 1987 int level;
1614 int max_level; 1988 int max_level;
1615 int replaced = 0; 1989 int replaced = 0;
1616 int ret; 1990 int ret;
1617 int err = 0; 1991 int err = 0;
1992 u32 min_reserved;
1618 1993
1619 path = btrfs_alloc_path(); 1994 path = btrfs_alloc_path();
1620 if (!path) 1995 if (!path)
@@ -1648,34 +2023,23 @@ static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1648 btrfs_unlock_up_safe(path, 0); 2023 btrfs_unlock_up_safe(path, 0);
1649 } 2024 }
1650 2025
1651 if (level == 0 && rc->stage == UPDATE_DATA_PTRS) { 2026 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1652 trans = btrfs_start_transaction(root, 0); 2027 memset(&next_key, 0, sizeof(next_key));
1653 2028
1654 leaf = path->nodes[0]; 2029 while (1) {
1655 btrfs_item_key_to_cpu(leaf, &key, 0); 2030 trans = btrfs_start_transaction(root, 0);
1656 btrfs_release_path(reloc_root, path); 2031 trans->block_rsv = rc->block_rsv;
1657 2032
1658 ret = btrfs_search_slot(trans, root, &key, path, 0, 1); 2033 ret = btrfs_block_rsv_check(trans, root, rc->block_rsv,
1659 if (ret < 0) { 2034 min_reserved, 0);
1660 err = ret; 2035 if (ret) {
1661 goto out; 2036 BUG_ON(ret != -EAGAIN);
2037 ret = btrfs_commit_transaction(trans, root);
2038 BUG_ON(ret);
2039 continue;
1662 } 2040 }
1663 2041
1664 leaf = path->nodes[0];
1665 btrfs_unlock_up_safe(path, 1);
1666 ret = replace_file_extents(trans, rc, root, leaf,
1667 &inode_list);
1668 if (ret < 0)
1669 err = ret;
1670 goto out;
1671 }
1672
1673 memset(&next_key, 0, sizeof(next_key));
1674
1675 while (1) {
1676 leaf = NULL;
1677 replaced = 0; 2042 replaced = 0;
1678 trans = btrfs_start_transaction(root, 0);
1679 max_level = level; 2043 max_level = level;
1680 2044
1681 ret = walk_down_reloc_tree(reloc_root, path, &level); 2045 ret = walk_down_reloc_tree(reloc_root, path, &level);
@@ -1689,14 +2053,9 @@ static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1689 if (!find_next_key(path, level, &key) && 2053 if (!find_next_key(path, level, &key) &&
1690 btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 2054 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1691 ret = 0; 2055 ret = 0;
1692 } else if (level == 1 && rc->stage == UPDATE_DATA_PTRS) {
1693 ret = replace_path(trans, root, reloc_root,
1694 path, &next_key, &leaf,
1695 level, max_level);
1696 } else { 2056 } else {
1697 ret = replace_path(trans, root, reloc_root, 2057 ret = replace_path(trans, root, reloc_root, path,
1698 path, &next_key, NULL, 2058 &next_key, level, max_level);
1699 level, max_level);
1700 } 2059 }
1701 if (ret < 0) { 2060 if (ret < 0) {
1702 err = ret; 2061 err = ret;
@@ -1708,16 +2067,6 @@ static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1708 btrfs_node_key_to_cpu(path->nodes[level], &key, 2067 btrfs_node_key_to_cpu(path->nodes[level], &key,
1709 path->slots[level]); 2068 path->slots[level]);
1710 replaced = 1; 2069 replaced = 1;
1711 } else if (leaf) {
1712 /*
1713 * no block got replaced, try replacing file extents
1714 */
1715 btrfs_item_key_to_cpu(leaf, &key, 0);
1716 ret = replace_file_extents(trans, rc, root, leaf,
1717 &inode_list);
1718 btrfs_tree_unlock(leaf);
1719 free_extent_buffer(leaf);
1720 BUG_ON(ret < 0);
1721 } 2070 }
1722 2071
1723 ret = walk_up_reloc_tree(reloc_root, path, &level); 2072 ret = walk_up_reloc_tree(reloc_root, path, &level);
@@ -1734,15 +2083,10 @@ static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1734 root_item->drop_level = level; 2083 root_item->drop_level = level;
1735 2084
1736 nr = trans->blocks_used; 2085 nr = trans->blocks_used;
1737 btrfs_end_transaction(trans, root); 2086 btrfs_end_transaction_throttle(trans, root);
1738 2087
1739 btrfs_btree_balance_dirty(root, nr); 2088 btrfs_btree_balance_dirty(root, nr);
1740 2089
1741 /*
1742 * put inodes outside transaction, otherwise we may deadlock.
1743 */
1744 put_inodes(&inode_list);
1745
1746 if (replaced && rc->stage == UPDATE_DATA_PTRS) 2090 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1747 invalidate_extent_cache(root, &key, &next_key); 2091 invalidate_extent_cache(root, &key, &next_key);
1748 } 2092 }
@@ -1765,87 +2109,125 @@ out:
1765 sizeof(root_item->drop_progress)); 2109 sizeof(root_item->drop_progress));
1766 root_item->drop_level = 0; 2110 root_item->drop_level = 0;
1767 btrfs_set_root_refs(root_item, 0); 2111 btrfs_set_root_refs(root_item, 0);
2112 btrfs_update_reloc_root(trans, root);
1768 } 2113 }
1769 2114
1770 nr = trans->blocks_used; 2115 nr = trans->blocks_used;
1771 btrfs_end_transaction(trans, root); 2116 btrfs_end_transaction_throttle(trans, root);
1772 2117
1773 btrfs_btree_balance_dirty(root, nr); 2118 btrfs_btree_balance_dirty(root, nr);
1774 2119
1775 put_inodes(&inode_list);
1776
1777 if (replaced && rc->stage == UPDATE_DATA_PTRS) 2120 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1778 invalidate_extent_cache(root, &key, &next_key); 2121 invalidate_extent_cache(root, &key, &next_key);
1779 2122
1780 return err; 2123 return err;
1781} 2124}
1782 2125
1783/* 2126static noinline_for_stack
1784 * callback for the work threads. 2127int prepare_to_merge(struct reloc_control *rc, int err)
1785 * this function merges reloc tree with corresponding fs tree,
1786 * and then drops the reloc tree.
1787 */
1788static void merge_func(struct btrfs_work *work)
1789{ 2128{
1790 struct btrfs_trans_handle *trans; 2129 struct btrfs_root *root = rc->extent_root;
1791 struct btrfs_root *root;
1792 struct btrfs_root *reloc_root; 2130 struct btrfs_root *reloc_root;
1793 struct async_merge *async; 2131 struct btrfs_trans_handle *trans;
2132 LIST_HEAD(reloc_roots);
2133 u64 num_bytes = 0;
2134 int ret;
2135 int retries = 0;
1794 2136
1795 async = container_of(work, struct async_merge, work); 2137 mutex_lock(&root->fs_info->trans_mutex);
1796 reloc_root = async->root; 2138 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2139 rc->merging_rsv_size += rc->nodes_relocated * 2;
2140 mutex_unlock(&root->fs_info->trans_mutex);
2141again:
2142 if (!err) {
2143 num_bytes = rc->merging_rsv_size;
2144 ret = btrfs_block_rsv_add(NULL, root, rc->block_rsv,
2145 num_bytes, &retries);
2146 if (ret)
2147 err = ret;
2148 }
2149
2150 trans = btrfs_join_transaction(rc->extent_root, 1);
2151
2152 if (!err) {
2153 if (num_bytes != rc->merging_rsv_size) {
2154 btrfs_end_transaction(trans, rc->extent_root);
2155 btrfs_block_rsv_release(rc->extent_root,
2156 rc->block_rsv, num_bytes);
2157 retries = 0;
2158 goto again;
2159 }
2160 }
2161
2162 rc->merge_reloc_tree = 1;
2163
2164 while (!list_empty(&rc->reloc_roots)) {
2165 reloc_root = list_entry(rc->reloc_roots.next,
2166 struct btrfs_root, root_list);
2167 list_del_init(&reloc_root->root_list);
1797 2168
1798 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1799 root = read_fs_root(reloc_root->fs_info, 2169 root = read_fs_root(reloc_root->fs_info,
1800 reloc_root->root_key.offset); 2170 reloc_root->root_key.offset);
1801 BUG_ON(IS_ERR(root)); 2171 BUG_ON(IS_ERR(root));
1802 BUG_ON(root->reloc_root != reloc_root); 2172 BUG_ON(root->reloc_root != reloc_root);
1803 2173
1804 merge_reloc_root(async->rc, root); 2174 /*
1805 2175 * set reference count to 1, so btrfs_recover_relocation
1806 trans = btrfs_start_transaction(root, 0); 2176 * knows it should resumes merging
2177 */
2178 if (!err)
2179 btrfs_set_root_refs(&reloc_root->root_item, 1);
1807 btrfs_update_reloc_root(trans, root); 2180 btrfs_update_reloc_root(trans, root);
1808 btrfs_end_transaction(trans, root);
1809 }
1810 2181
1811 btrfs_drop_snapshot(reloc_root, 0); 2182 list_add(&reloc_root->root_list, &reloc_roots);
2183 }
1812 2184
1813 if (atomic_dec_and_test(async->num_pending)) 2185 list_splice(&reloc_roots, &rc->reloc_roots);
1814 complete(async->done);
1815 2186
1816 kfree(async); 2187 if (!err)
2188 btrfs_commit_transaction(trans, rc->extent_root);
2189 else
2190 btrfs_end_transaction(trans, rc->extent_root);
2191 return err;
1817} 2192}
1818 2193
1819static int merge_reloc_roots(struct reloc_control *rc) 2194static noinline_for_stack
2195int merge_reloc_roots(struct reloc_control *rc)
1820{ 2196{
1821 struct async_merge *async;
1822 struct btrfs_root *root; 2197 struct btrfs_root *root;
1823 struct completion done; 2198 struct btrfs_root *reloc_root;
1824 atomic_t num_pending; 2199 LIST_HEAD(reloc_roots);
2200 int found = 0;
2201 int ret;
2202again:
2203 root = rc->extent_root;
2204 mutex_lock(&root->fs_info->trans_mutex);
2205 list_splice_init(&rc->reloc_roots, &reloc_roots);
2206 mutex_unlock(&root->fs_info->trans_mutex);
1825 2207
1826 init_completion(&done); 2208 while (!list_empty(&reloc_roots)) {
1827 atomic_set(&num_pending, 1); 2209 found = 1;
2210 reloc_root = list_entry(reloc_roots.next,
2211 struct btrfs_root, root_list);
1828 2212
1829 while (!list_empty(&rc->reloc_roots)) { 2213 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1830 root = list_entry(rc->reloc_roots.next, 2214 root = read_fs_root(reloc_root->fs_info,
1831 struct btrfs_root, root_list); 2215 reloc_root->root_key.offset);
1832 list_del_init(&root->root_list); 2216 BUG_ON(IS_ERR(root));
2217 BUG_ON(root->reloc_root != reloc_root);
1833 2218
1834 async = kmalloc(sizeof(*async), GFP_NOFS); 2219 ret = merge_reloc_root(rc, root);
1835 BUG_ON(!async); 2220 BUG_ON(ret);
1836 async->work.func = merge_func; 2221 } else {
1837 async->work.flags = 0; 2222 list_del_init(&reloc_root->root_list);
1838 async->rc = rc; 2223 }
1839 async->root = root; 2224 btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0);
1840 async->done = &done;
1841 async->num_pending = &num_pending;
1842 atomic_inc(&num_pending);
1843 btrfs_queue_worker(&rc->workers, &async->work);
1844 } 2225 }
1845 2226
1846 if (!atomic_dec_and_test(&num_pending)) 2227 if (found) {
1847 wait_for_completion(&done); 2228 found = 0;
1848 2229 goto again;
2230 }
1849 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 2231 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
1850 return 0; 2232 return 0;
1851} 2233}
@@ -1876,119 +2258,169 @@ static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
1876 return btrfs_record_root_in_trans(trans, root); 2258 return btrfs_record_root_in_trans(trans, root);
1877} 2259}
1878 2260
1879/* 2261static noinline_for_stack
1880 * select one tree from trees that references the block. 2262struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
1881 * for blocks in refernce counted trees, we preper reloc tree. 2263 struct reloc_control *rc,
1882 * if no reloc tree found and reloc_only is true, NULL is returned. 2264 struct backref_node *node,
1883 */ 2265 struct backref_edge *edges[], int *nr)
1884static struct btrfs_root *__select_one_root(struct btrfs_trans_handle *trans,
1885 struct backref_node *node,
1886 struct backref_edge *edges[],
1887 int *nr, int reloc_only)
1888{ 2266{
1889 struct backref_node *next; 2267 struct backref_node *next;
1890 struct btrfs_root *root; 2268 struct btrfs_root *root;
1891 int index; 2269 int index = 0;
1892 int loop = 0; 2270
1893again:
1894 index = 0;
1895 next = node; 2271 next = node;
1896 while (1) { 2272 while (1) {
1897 cond_resched(); 2273 cond_resched();
1898 next = walk_up_backref(next, edges, &index); 2274 next = walk_up_backref(next, edges, &index);
1899 root = next->root; 2275 root = next->root;
1900 if (!root) { 2276 BUG_ON(!root);
1901 BUG_ON(!node->old_root); 2277 BUG_ON(!root->ref_cows);
1902 goto skip;
1903 }
1904
1905 /* no other choice for non-refernce counted tree */
1906 if (!root->ref_cows) {
1907 BUG_ON(reloc_only);
1908 break;
1909 }
1910 2278
1911 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 2279 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
1912 record_reloc_root_in_trans(trans, root); 2280 record_reloc_root_in_trans(trans, root);
1913 break; 2281 break;
1914 } 2282 }
1915 2283
1916 if (loop) { 2284 btrfs_record_root_in_trans(trans, root);
1917 btrfs_record_root_in_trans(trans, root); 2285 root = root->reloc_root;
2286
2287 if (next->new_bytenr != root->node->start) {
2288 BUG_ON(next->new_bytenr);
2289 BUG_ON(!list_empty(&next->list));
2290 next->new_bytenr = root->node->start;
2291 next->root = root;
2292 list_add_tail(&next->list,
2293 &rc->backref_cache.changed);
2294 __mark_block_processed(rc, next);
1918 break; 2295 break;
1919 } 2296 }
1920 2297
1921 if (reloc_only || next != node) { 2298 WARN_ON(1);
1922 if (!root->reloc_root)
1923 btrfs_record_root_in_trans(trans, root);
1924 root = root->reloc_root;
1925 /*
1926 * if the reloc tree was created in current
1927 * transation, there is no node in backref tree
1928 * corresponds to the root of the reloc tree.
1929 */
1930 if (btrfs_root_last_snapshot(&root->root_item) ==
1931 trans->transid - 1)
1932 break;
1933 }
1934skip:
1935 root = NULL; 2299 root = NULL;
1936 next = walk_down_backref(edges, &index); 2300 next = walk_down_backref(edges, &index);
1937 if (!next || next->level <= node->level) 2301 if (!next || next->level <= node->level)
1938 break; 2302 break;
1939 } 2303 }
2304 if (!root)
2305 return NULL;
1940 2306
1941 if (!root && !loop && !reloc_only) { 2307 *nr = index;
1942 loop = 1; 2308 next = node;
1943 goto again; 2309 /* setup backref node path for btrfs_reloc_cow_block */
2310 while (1) {
2311 rc->backref_cache.path[next->level] = next;
2312 if (--index < 0)
2313 break;
2314 next = edges[index]->node[UPPER];
1944 } 2315 }
1945
1946 if (root)
1947 *nr = index;
1948 else
1949 *nr = 0;
1950
1951 return root; 2316 return root;
1952} 2317}
1953 2318
2319/*
2320 * select a tree root for relocation. return NULL if the block
2321 * is reference counted. we should use do_relocation() in this
2322 * case. return a tree root pointer if the block isn't reference
2323 * counted. return -ENOENT if the block is root of reloc tree.
2324 */
1954static noinline_for_stack 2325static noinline_for_stack
1955struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans, 2326struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
1956 struct backref_node *node) 2327 struct backref_node *node)
1957{ 2328{
2329 struct backref_node *next;
2330 struct btrfs_root *root;
2331 struct btrfs_root *fs_root = NULL;
1958 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 2332 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
1959 int nr; 2333 int index = 0;
1960 return __select_one_root(trans, node, edges, &nr, 0); 2334
2335 next = node;
2336 while (1) {
2337 cond_resched();
2338 next = walk_up_backref(next, edges, &index);
2339 root = next->root;
2340 BUG_ON(!root);
2341
2342 /* no other choice for non-refernce counted tree */
2343 if (!root->ref_cows)
2344 return root;
2345
2346 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2347 fs_root = root;
2348
2349 if (next != node)
2350 return NULL;
2351
2352 next = walk_down_backref(edges, &index);
2353 if (!next || next->level <= node->level)
2354 break;
2355 }
2356
2357 if (!fs_root)
2358 return ERR_PTR(-ENOENT);
2359 return fs_root;
1961} 2360}
1962 2361
1963static noinline_for_stack 2362static noinline_for_stack
1964struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 2363u64 calcu_metadata_size(struct reloc_control *rc,
1965 struct backref_node *node, 2364 struct backref_node *node, int reserve)
1966 struct backref_edge *edges[], int *nr)
1967{ 2365{
1968 return __select_one_root(trans, node, edges, nr, 1); 2366 struct backref_node *next = node;
2367 struct backref_edge *edge;
2368 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2369 u64 num_bytes = 0;
2370 int index = 0;
2371
2372 BUG_ON(reserve && node->processed);
2373
2374 while (next) {
2375 cond_resched();
2376 while (1) {
2377 if (next->processed && (reserve || next != node))
2378 break;
2379
2380 num_bytes += btrfs_level_size(rc->extent_root,
2381 next->level);
2382
2383 if (list_empty(&next->upper))
2384 break;
2385
2386 edge = list_entry(next->upper.next,
2387 struct backref_edge, list[LOWER]);
2388 edges[index++] = edge;
2389 next = edge->node[UPPER];
2390 }
2391 next = walk_down_backref(edges, &index);
2392 }
2393 return num_bytes;
1969} 2394}
1970 2395
1971static void grab_path_buffers(struct btrfs_path *path, 2396static int reserve_metadata_space(struct btrfs_trans_handle *trans,
1972 struct backref_node *node, 2397 struct reloc_control *rc,
1973 struct backref_edge *edges[], int nr) 2398 struct backref_node *node)
1974{ 2399{
1975 int i = 0; 2400 struct btrfs_root *root = rc->extent_root;
1976 while (1) { 2401 u64 num_bytes;
1977 drop_node_buffer(node); 2402 int ret;
1978 node->eb = path->nodes[node->level];
1979 BUG_ON(!node->eb);
1980 if (path->locks[node->level])
1981 node->locked = 1;
1982 path->nodes[node->level] = NULL;
1983 path->locks[node->level] = 0;
1984
1985 if (i >= nr)
1986 break;
1987 2403
1988 edges[i]->blockptr = node->eb->start; 2404 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
1989 node = edges[i]->node[UPPER]; 2405
1990 i++; 2406 trans->block_rsv = rc->block_rsv;
2407 ret = btrfs_block_rsv_add(trans, root, rc->block_rsv, num_bytes,
2408 &rc->block_rsv_retries);
2409 if (ret) {
2410 if (ret == -EAGAIN)
2411 rc->commit_transaction = 1;
2412 return ret;
1991 } 2413 }
2414
2415 rc->block_rsv_retries = 0;
2416 return 0;
2417}
2418
2419static void release_metadata_space(struct reloc_control *rc,
2420 struct backref_node *node)
2421{
2422 u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2423 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
1992} 2424}
1993 2425
1994/* 2426/*
@@ -1999,6 +2431,7 @@ static void grab_path_buffers(struct btrfs_path *path,
1999 * in that case this function just updates pointers. 2431 * in that case this function just updates pointers.
2000 */ 2432 */
2001static int do_relocation(struct btrfs_trans_handle *trans, 2433static int do_relocation(struct btrfs_trans_handle *trans,
2434 struct reloc_control *rc,
2002 struct backref_node *node, 2435 struct backref_node *node,
2003 struct btrfs_key *key, 2436 struct btrfs_key *key,
2004 struct btrfs_path *path, int lowest) 2437 struct btrfs_path *path, int lowest)
@@ -2019,18 +2452,25 @@ static int do_relocation(struct btrfs_trans_handle *trans,
2019 BUG_ON(lowest && node->eb); 2452 BUG_ON(lowest && node->eb);
2020 2453
2021 path->lowest_level = node->level + 1; 2454 path->lowest_level = node->level + 1;
2455 rc->backref_cache.path[node->level] = node;
2022 list_for_each_entry(edge, &node->upper, list[LOWER]) { 2456 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2023 cond_resched(); 2457 cond_resched();
2024 if (node->eb && node->eb->start == edge->blockptr)
2025 continue;
2026 2458
2027 upper = edge->node[UPPER]; 2459 upper = edge->node[UPPER];
2028 root = select_reloc_root(trans, upper, edges, &nr); 2460 root = select_reloc_root(trans, rc, upper, edges, &nr);
2029 if (!root) 2461 BUG_ON(!root);
2030 continue; 2462
2031 2463 if (upper->eb && !upper->locked) {
2032 if (upper->eb && !upper->locked) 2464 if (!lowest) {
2465 ret = btrfs_bin_search(upper->eb, key,
2466 upper->level, &slot);
2467 BUG_ON(ret);
2468 bytenr = btrfs_node_blockptr(upper->eb, slot);
2469 if (node->eb->start == bytenr)
2470 goto next;
2471 }
2033 drop_node_buffer(upper); 2472 drop_node_buffer(upper);
2473 }
2034 2474
2035 if (!upper->eb) { 2475 if (!upper->eb) {
2036 ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2476 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
@@ -2040,11 +2480,17 @@ static int do_relocation(struct btrfs_trans_handle *trans,
2040 } 2480 }
2041 BUG_ON(ret > 0); 2481 BUG_ON(ret > 0);
2042 2482
2043 slot = path->slots[upper->level]; 2483 if (!upper->eb) {
2484 upper->eb = path->nodes[upper->level];
2485 path->nodes[upper->level] = NULL;
2486 } else {
2487 BUG_ON(upper->eb != path->nodes[upper->level]);
2488 }
2044 2489
2045 btrfs_unlock_up_safe(path, upper->level + 1); 2490 upper->locked = 1;
2046 grab_path_buffers(path, upper, edges, nr); 2491 path->locks[upper->level] = 0;
2047 2492
2493 slot = path->slots[upper->level];
2048 btrfs_release_path(NULL, path); 2494 btrfs_release_path(NULL, path);
2049 } else { 2495 } else {
2050 ret = btrfs_bin_search(upper->eb, key, upper->level, 2496 ret = btrfs_bin_search(upper->eb, key, upper->level,
@@ -2053,14 +2499,11 @@ static int do_relocation(struct btrfs_trans_handle *trans,
2053 } 2499 }
2054 2500
2055 bytenr = btrfs_node_blockptr(upper->eb, slot); 2501 bytenr = btrfs_node_blockptr(upper->eb, slot);
2056 if (!lowest) { 2502 if (lowest) {
2057 if (node->eb->start == bytenr) { 2503 BUG_ON(bytenr != node->bytenr);
2058 btrfs_tree_unlock(upper->eb);
2059 upper->locked = 0;
2060 continue;
2061 }
2062 } else { 2504 } else {
2063 BUG_ON(node->bytenr != bytenr); 2505 if (node->eb->start == bytenr)
2506 goto next;
2064 } 2507 }
2065 2508
2066 blocksize = btrfs_level_size(root, node->level); 2509 blocksize = btrfs_level_size(root, node->level);
@@ -2072,13 +2515,13 @@ static int do_relocation(struct btrfs_trans_handle *trans,
2072 if (!node->eb) { 2515 if (!node->eb) {
2073 ret = btrfs_cow_block(trans, root, eb, upper->eb, 2516 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2074 slot, &eb); 2517 slot, &eb);
2518 btrfs_tree_unlock(eb);
2519 free_extent_buffer(eb);
2075 if (ret < 0) { 2520 if (ret < 0) {
2076 err = ret; 2521 err = ret;
2077 break; 2522 goto next;
2078 } 2523 }
2079 btrfs_set_lock_blocking(eb); 2524 BUG_ON(node->eb != eb);
2080 node->eb = eb;
2081 node->locked = 1;
2082 } else { 2525 } else {
2083 btrfs_set_node_blockptr(upper->eb, slot, 2526 btrfs_set_node_blockptr(upper->eb, slot,
2084 node->eb->start); 2527 node->eb->start);
@@ -2096,67 +2539,80 @@ static int do_relocation(struct btrfs_trans_handle *trans,
2096 ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 2539 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2097 BUG_ON(ret); 2540 BUG_ON(ret);
2098 } 2541 }
2099 if (!lowest) { 2542next:
2100 btrfs_tree_unlock(upper->eb); 2543 if (!upper->pending)
2101 upper->locked = 0; 2544 drop_node_buffer(upper);
2102 } 2545 else
2546 unlock_node_buffer(upper);
2547 if (err)
2548 break;
2103 } 2549 }
2550
2551 if (!err && node->pending) {
2552 drop_node_buffer(node);
2553 list_move_tail(&node->list, &rc->backref_cache.changed);
2554 node->pending = 0;
2555 }
2556
2104 path->lowest_level = 0; 2557 path->lowest_level = 0;
2558 BUG_ON(err == -ENOSPC);
2105 return err; 2559 return err;
2106} 2560}
2107 2561
2108static int link_to_upper(struct btrfs_trans_handle *trans, 2562static int link_to_upper(struct btrfs_trans_handle *trans,
2563 struct reloc_control *rc,
2109 struct backref_node *node, 2564 struct backref_node *node,
2110 struct btrfs_path *path) 2565 struct btrfs_path *path)
2111{ 2566{
2112 struct btrfs_key key; 2567 struct btrfs_key key;
2113 if (!node->eb || list_empty(&node->upper))
2114 return 0;
2115 2568
2116 btrfs_node_key_to_cpu(node->eb, &key, 0); 2569 btrfs_node_key_to_cpu(node->eb, &key, 0);
2117 return do_relocation(trans, node, &key, path, 0); 2570 return do_relocation(trans, rc, node, &key, path, 0);
2118} 2571}
2119 2572
2120static int finish_pending_nodes(struct btrfs_trans_handle *trans, 2573static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2121 struct backref_cache *cache, 2574 struct reloc_control *rc,
2122 struct btrfs_path *path) 2575 struct btrfs_path *path, int err)
2123{ 2576{
2577 LIST_HEAD(list);
2578 struct backref_cache *cache = &rc->backref_cache;
2124 struct backref_node *node; 2579 struct backref_node *node;
2125 int level; 2580 int level;
2126 int ret; 2581 int ret;
2127 int err = 0;
2128 2582
2129 for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 2583 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2130 while (!list_empty(&cache->pending[level])) { 2584 while (!list_empty(&cache->pending[level])) {
2131 node = list_entry(cache->pending[level].next, 2585 node = list_entry(cache->pending[level].next,
2132 struct backref_node, lower); 2586 struct backref_node, list);
2133 BUG_ON(node->level != level); 2587 list_move_tail(&node->list, &list);
2588 BUG_ON(!node->pending);
2134 2589
2135 ret = link_to_upper(trans, node, path); 2590 if (!err) {
2136 if (ret < 0) 2591 ret = link_to_upper(trans, rc, node, path);
2137 err = ret; 2592 if (ret < 0)
2138 /* 2593 err = ret;
2139 * this remove the node from the pending list and 2594 }
2140 * may add some other nodes to the level + 1
2141 * pending list
2142 */
2143 remove_backref_node(cache, node);
2144 } 2595 }
2596 list_splice_init(&list, &cache->pending[level]);
2145 } 2597 }
2146 BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
2147 return err; 2598 return err;
2148} 2599}
2149 2600
2150static void mark_block_processed(struct reloc_control *rc, 2601static void mark_block_processed(struct reloc_control *rc,
2151 struct backref_node *node) 2602 u64 bytenr, u32 blocksize)
2603{
2604 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2605 EXTENT_DIRTY, GFP_NOFS);
2606}
2607
2608static void __mark_block_processed(struct reloc_control *rc,
2609 struct backref_node *node)
2152{ 2610{
2153 u32 blocksize; 2611 u32 blocksize;
2154 if (node->level == 0 || 2612 if (node->level == 0 ||
2155 in_block_group(node->bytenr, rc->block_group)) { 2613 in_block_group(node->bytenr, rc->block_group)) {
2156 blocksize = btrfs_level_size(rc->extent_root, node->level); 2614 blocksize = btrfs_level_size(rc->extent_root, node->level);
2157 set_extent_bits(&rc->processed_blocks, node->bytenr, 2615 mark_block_processed(rc, node->bytenr, blocksize);
2158 node->bytenr + blocksize - 1, EXTENT_DIRTY,
2159 GFP_NOFS);
2160 } 2616 }
2161 node->processed = 1; 2617 node->processed = 1;
2162} 2618}
@@ -2179,7 +2635,7 @@ static void update_processed_blocks(struct reloc_control *rc,
2179 if (next->processed) 2635 if (next->processed)
2180 break; 2636 break;
2181 2637
2182 mark_block_processed(rc, next); 2638 __mark_block_processed(rc, next);
2183 2639
2184 if (list_empty(&next->upper)) 2640 if (list_empty(&next->upper))
2185 break; 2641 break;
@@ -2202,138 +2658,6 @@ static int tree_block_processed(u64 bytenr, u32 blocksize,
2202 return 0; 2658 return 0;
2203} 2659}
2204 2660
2205/*
2206 * check if there are any file extent pointers in the leaf point to
2207 * data require processing
2208 */
2209static int check_file_extents(struct reloc_control *rc,
2210 u64 bytenr, u32 blocksize, u64 ptr_gen)
2211{
2212 struct btrfs_key found_key;
2213 struct btrfs_file_extent_item *fi;
2214 struct extent_buffer *leaf;
2215 u32 nritems;
2216 int i;
2217 int ret = 0;
2218
2219 leaf = read_tree_block(rc->extent_root, bytenr, blocksize, ptr_gen);
2220
2221 nritems = btrfs_header_nritems(leaf);
2222 for (i = 0; i < nritems; i++) {
2223 cond_resched();
2224 btrfs_item_key_to_cpu(leaf, &found_key, i);
2225 if (found_key.type != BTRFS_EXTENT_DATA_KEY)
2226 continue;
2227 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
2228 if (btrfs_file_extent_type(leaf, fi) ==
2229 BTRFS_FILE_EXTENT_INLINE)
2230 continue;
2231 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
2232 if (bytenr == 0)
2233 continue;
2234 if (in_block_group(bytenr, rc->block_group)) {
2235 ret = 1;
2236 break;
2237 }
2238 }
2239 free_extent_buffer(leaf);
2240 return ret;
2241}
2242
2243/*
2244 * scan child blocks of a given block to find blocks require processing
2245 */
2246static int add_child_blocks(struct btrfs_trans_handle *trans,
2247 struct reloc_control *rc,
2248 struct backref_node *node,
2249 struct rb_root *blocks)
2250{
2251 struct tree_block *block;
2252 struct rb_node *rb_node;
2253 u64 bytenr;
2254 u64 ptr_gen;
2255 u32 blocksize;
2256 u32 nritems;
2257 int i;
2258 int err = 0;
2259
2260 nritems = btrfs_header_nritems(node->eb);
2261 blocksize = btrfs_level_size(rc->extent_root, node->level - 1);
2262 for (i = 0; i < nritems; i++) {
2263 cond_resched();
2264 bytenr = btrfs_node_blockptr(node->eb, i);
2265 ptr_gen = btrfs_node_ptr_generation(node->eb, i);
2266 if (ptr_gen == trans->transid)
2267 continue;
2268 if (!in_block_group(bytenr, rc->block_group) &&
2269 (node->level > 1 || rc->stage == MOVE_DATA_EXTENTS))
2270 continue;
2271 if (tree_block_processed(bytenr, blocksize, rc))
2272 continue;
2273
2274 readahead_tree_block(rc->extent_root,
2275 bytenr, blocksize, ptr_gen);
2276 }
2277
2278 for (i = 0; i < nritems; i++) {
2279 cond_resched();
2280 bytenr = btrfs_node_blockptr(node->eb, i);
2281 ptr_gen = btrfs_node_ptr_generation(node->eb, i);
2282 if (ptr_gen == trans->transid)
2283 continue;
2284 if (!in_block_group(bytenr, rc->block_group) &&
2285 (node->level > 1 || rc->stage == MOVE_DATA_EXTENTS))
2286 continue;
2287 if (tree_block_processed(bytenr, blocksize, rc))
2288 continue;
2289 if (!in_block_group(bytenr, rc->block_group) &&
2290 !check_file_extents(rc, bytenr, blocksize, ptr_gen))
2291 continue;
2292
2293 block = kmalloc(sizeof(*block), GFP_NOFS);
2294 if (!block) {
2295 err = -ENOMEM;
2296 break;
2297 }
2298 block->bytenr = bytenr;
2299 btrfs_node_key_to_cpu(node->eb, &block->key, i);
2300 block->level = node->level - 1;
2301 block->key_ready = 1;
2302 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
2303 BUG_ON(rb_node);
2304 }
2305 if (err)
2306 free_block_list(blocks);
2307 return err;
2308}
2309
2310/*
2311 * find adjacent blocks require processing
2312 */
2313static noinline_for_stack
2314int add_adjacent_blocks(struct btrfs_trans_handle *trans,
2315 struct reloc_control *rc,
2316 struct backref_cache *cache,
2317 struct rb_root *blocks, int level,
2318 struct backref_node **upper)
2319{
2320 struct backref_node *node;
2321 int ret = 0;
2322
2323 WARN_ON(!list_empty(&cache->pending[level]));
2324
2325 if (list_empty(&cache->pending[level + 1]))
2326 return 1;
2327
2328 node = list_entry(cache->pending[level + 1].next,
2329 struct backref_node, lower);
2330 if (node->eb)
2331 ret = add_child_blocks(trans, rc, node, blocks);
2332
2333 *upper = node;
2334 return ret;
2335}
2336
2337static int get_tree_block_key(struct reloc_control *rc, 2661static int get_tree_block_key(struct reloc_control *rc,
2338 struct tree_block *block) 2662 struct tree_block *block)
2339{ 2663{
@@ -2371,40 +2695,53 @@ static int relocate_tree_block(struct btrfs_trans_handle *trans,
2371 struct btrfs_path *path) 2695 struct btrfs_path *path)
2372{ 2696{
2373 struct btrfs_root *root; 2697 struct btrfs_root *root;
2374 int ret; 2698 int release = 0;
2699 int ret = 0;
2700
2701 if (!node)
2702 return 0;
2375 2703
2704 BUG_ON(node->processed);
2376 root = select_one_root(trans, node); 2705 root = select_one_root(trans, node);
2377 if (unlikely(!root)) { 2706 if (root == ERR_PTR(-ENOENT)) {
2378 rc->found_old_snapshot = 1;
2379 update_processed_blocks(rc, node); 2707 update_processed_blocks(rc, node);
2380 return 0; 2708 goto out;
2381 } 2709 }
2382 2710
2383 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 2711 if (!root || root->ref_cows) {
2384 ret = do_relocation(trans, node, key, path, 1); 2712 ret = reserve_metadata_space(trans, rc, node);
2385 if (ret < 0) 2713 if (ret)
2386 goto out;
2387 if (node->level == 0 && rc->stage == UPDATE_DATA_PTRS) {
2388 ret = replace_file_extents(trans, rc, root,
2389 node->eb, NULL);
2390 if (ret < 0)
2391 goto out;
2392 }
2393 drop_node_buffer(node);
2394 } else if (!root->ref_cows) {
2395 path->lowest_level = node->level;
2396 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2397 btrfs_release_path(root, path);
2398 if (ret < 0)
2399 goto out; 2714 goto out;
2400 } else if (root != node->root) { 2715 release = 1;
2401 WARN_ON(node->level > 0 || rc->stage != UPDATE_DATA_PTRS);
2402 } 2716 }
2403 2717
2404 update_processed_blocks(rc, node); 2718 if (root) {
2405 ret = 0; 2719 if (root->ref_cows) {
2720 BUG_ON(node->new_bytenr);
2721 BUG_ON(!list_empty(&node->list));
2722 btrfs_record_root_in_trans(trans, root);
2723 root = root->reloc_root;
2724 node->new_bytenr = root->node->start;
2725 node->root = root;
2726 list_add_tail(&node->list, &rc->backref_cache.changed);
2727 } else {
2728 path->lowest_level = node->level;
2729 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2730 btrfs_release_path(root, path);
2731 if (ret > 0)
2732 ret = 0;
2733 }
2734 if (!ret)
2735 update_processed_blocks(rc, node);
2736 } else {
2737 ret = do_relocation(trans, rc, node, key, path, 1);
2738 }
2406out: 2739out:
2407 drop_node_buffer(node); 2740 if (ret || node->level == 0 || node->cowonly) {
2741 if (release)
2742 release_metadata_space(rc, node);
2743 remove_backref_node(&rc->backref_cache, node);
2744 }
2408 return ret; 2745 return ret;
2409} 2746}
2410 2747
@@ -2415,12 +2752,10 @@ static noinline_for_stack
2415int relocate_tree_blocks(struct btrfs_trans_handle *trans, 2752int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2416 struct reloc_control *rc, struct rb_root *blocks) 2753 struct reloc_control *rc, struct rb_root *blocks)
2417{ 2754{
2418 struct backref_cache *cache;
2419 struct backref_node *node; 2755 struct backref_node *node;
2420 struct btrfs_path *path; 2756 struct btrfs_path *path;
2421 struct tree_block *block; 2757 struct tree_block *block;
2422 struct rb_node *rb_node; 2758 struct rb_node *rb_node;
2423 int level = -1;
2424 int ret; 2759 int ret;
2425 int err = 0; 2760 int err = 0;
2426 2761
@@ -2428,21 +2763,9 @@ int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2428 if (!path) 2763 if (!path)
2429 return -ENOMEM; 2764 return -ENOMEM;
2430 2765
2431 cache = kmalloc(sizeof(*cache), GFP_NOFS);
2432 if (!cache) {
2433 btrfs_free_path(path);
2434 return -ENOMEM;
2435 }
2436
2437 backref_cache_init(cache);
2438
2439 rb_node = rb_first(blocks); 2766 rb_node = rb_first(blocks);
2440 while (rb_node) { 2767 while (rb_node) {
2441 block = rb_entry(rb_node, struct tree_block, rb_node); 2768 block = rb_entry(rb_node, struct tree_block, rb_node);
2442 if (level == -1)
2443 level = block->level;
2444 else
2445 BUG_ON(level != block->level);
2446 if (!block->key_ready) 2769 if (!block->key_ready)
2447 reada_tree_block(rc, block); 2770 reada_tree_block(rc, block);
2448 rb_node = rb_next(rb_node); 2771 rb_node = rb_next(rb_node);
@@ -2460,7 +2783,7 @@ int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2460 while (rb_node) { 2783 while (rb_node) {
2461 block = rb_entry(rb_node, struct tree_block, rb_node); 2784 block = rb_entry(rb_node, struct tree_block, rb_node);
2462 2785
2463 node = build_backref_tree(rc, cache, &block->key, 2786 node = build_backref_tree(rc, &block->key,
2464 block->level, block->bytenr); 2787 block->level, block->bytenr);
2465 if (IS_ERR(node)) { 2788 if (IS_ERR(node)) {
2466 err = PTR_ERR(node); 2789 err = PTR_ERR(node);
@@ -2470,77 +2793,16 @@ int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2470 ret = relocate_tree_block(trans, rc, node, &block->key, 2793 ret = relocate_tree_block(trans, rc, node, &block->key,
2471 path); 2794 path);
2472 if (ret < 0) { 2795 if (ret < 0) {
2473 err = ret; 2796 if (ret != -EAGAIN || rb_node == rb_first(blocks))
2797 err = ret;
2474 goto out; 2798 goto out;
2475 } 2799 }
2476 remove_backref_node(cache, node);
2477 rb_node = rb_next(rb_node); 2800 rb_node = rb_next(rb_node);
2478 } 2801 }
2479
2480 if (level > 0)
2481 goto out;
2482
2483 free_block_list(blocks);
2484
2485 /*
2486 * now backrefs of some upper level tree blocks have been cached,
2487 * try relocating blocks referenced by these upper level blocks.
2488 */
2489 while (1) {
2490 struct backref_node *upper = NULL;
2491 if (trans->transaction->in_commit ||
2492 trans->transaction->delayed_refs.flushing)
2493 break;
2494
2495 ret = add_adjacent_blocks(trans, rc, cache, blocks, level,
2496 &upper);
2497 if (ret < 0)
2498 err = ret;
2499 if (ret != 0)
2500 break;
2501
2502 rb_node = rb_first(blocks);
2503 while (rb_node) {
2504 block = rb_entry(rb_node, struct tree_block, rb_node);
2505 if (trans->transaction->in_commit ||
2506 trans->transaction->delayed_refs.flushing)
2507 goto out;
2508 BUG_ON(!block->key_ready);
2509 node = build_backref_tree(rc, cache, &block->key,
2510 level, block->bytenr);
2511 if (IS_ERR(node)) {
2512 err = PTR_ERR(node);
2513 goto out;
2514 }
2515
2516 ret = relocate_tree_block(trans, rc, node,
2517 &block->key, path);
2518 if (ret < 0) {
2519 err = ret;
2520 goto out;
2521 }
2522 remove_backref_node(cache, node);
2523 rb_node = rb_next(rb_node);
2524 }
2525 free_block_list(blocks);
2526
2527 if (upper) {
2528 ret = link_to_upper(trans, upper, path);
2529 if (ret < 0) {
2530 err = ret;
2531 break;
2532 }
2533 remove_backref_node(cache, upper);
2534 }
2535 }
2536out: 2802out:
2537 free_block_list(blocks); 2803 free_block_list(blocks);
2804 err = finish_pending_nodes(trans, rc, path, err);
2538 2805
2539 ret = finish_pending_nodes(trans, cache, path);
2540 if (ret < 0)
2541 err = ret;
2542
2543 kfree(cache);
2544 btrfs_free_path(path); 2806 btrfs_free_path(path);
2545 return err; 2807 return err;
2546} 2808}
@@ -2910,9 +3172,6 @@ out:
2910static int block_use_full_backref(struct reloc_control *rc, 3172static int block_use_full_backref(struct reloc_control *rc,
2911 struct extent_buffer *eb) 3173 struct extent_buffer *eb)
2912{ 3174{
2913 struct btrfs_path *path;
2914 struct btrfs_extent_item *ei;
2915 struct btrfs_key key;
2916 u64 flags; 3175 u64 flags;
2917 int ret; 3176 int ret;
2918 3177
@@ -2920,28 +3179,14 @@ static int block_use_full_backref(struct reloc_control *rc,
2920 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV) 3179 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
2921 return 1; 3180 return 1;
2922 3181
2923 path = btrfs_alloc_path(); 3182 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
2924 BUG_ON(!path); 3183 eb->start, eb->len, NULL, &flags);
2925
2926 key.objectid = eb->start;
2927 key.type = BTRFS_EXTENT_ITEM_KEY;
2928 key.offset = eb->len;
2929
2930 path->search_commit_root = 1;
2931 path->skip_locking = 1;
2932 ret = btrfs_search_slot(NULL, rc->extent_root,
2933 &key, path, 0, 0);
2934 BUG_ON(ret); 3184 BUG_ON(ret);
2935 3185
2936 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2937 struct btrfs_extent_item);
2938 flags = btrfs_extent_flags(path->nodes[0], ei);
2939 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2940 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) 3186 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
2941 ret = 1; 3187 ret = 1;
2942 else 3188 else
2943 ret = 0; 3189 ret = 0;
2944 btrfs_free_path(path);
2945 return ret; 3190 return ret;
2946} 3191}
2947 3192
@@ -3114,22 +3359,10 @@ int add_data_references(struct reloc_control *rc,
3114 struct btrfs_extent_inline_ref *iref; 3359 struct btrfs_extent_inline_ref *iref;
3115 unsigned long ptr; 3360 unsigned long ptr;
3116 unsigned long end; 3361 unsigned long end;
3117 u32 blocksize; 3362 u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3118 int ret; 3363 int ret;
3119 int err = 0; 3364 int err = 0;
3120 3365
3121 ret = get_new_location(rc->data_inode, NULL, extent_key->objectid,
3122 extent_key->offset);
3123 BUG_ON(ret < 0);
3124 if (ret > 0) {
3125 /* the relocated data is fragmented */
3126 rc->extents_skipped++;
3127 btrfs_release_path(rc->extent_root, path);
3128 return 0;
3129 }
3130
3131 blocksize = btrfs_level_size(rc->extent_root, 0);
3132
3133 eb = path->nodes[0]; 3366 eb = path->nodes[0];
3134 ptr = btrfs_item_ptr_offset(eb, path->slots[0]); 3367 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3135 end = ptr + btrfs_item_size_nr(eb, path->slots[0]); 3368 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
@@ -3210,7 +3443,8 @@ int add_data_references(struct reloc_control *rc,
3210 */ 3443 */
3211static noinline_for_stack 3444static noinline_for_stack
3212int find_next_extent(struct btrfs_trans_handle *trans, 3445int find_next_extent(struct btrfs_trans_handle *trans,
3213 struct reloc_control *rc, struct btrfs_path *path) 3446 struct reloc_control *rc, struct btrfs_path *path,
3447 struct btrfs_key *extent_key)
3214{ 3448{
3215 struct btrfs_key key; 3449 struct btrfs_key key;
3216 struct extent_buffer *leaf; 3450 struct extent_buffer *leaf;
@@ -3265,6 +3499,7 @@ next:
3265 rc->search_start = end + 1; 3499 rc->search_start = end + 1;
3266 } else { 3500 } else {
3267 rc->search_start = key.objectid + key.offset; 3501 rc->search_start = key.objectid + key.offset;
3502 memcpy(extent_key, &key, sizeof(key));
3268 return 0; 3503 return 0;
3269 } 3504 }
3270 } 3505 }
@@ -3302,12 +3537,49 @@ static int check_extent_flags(u64 flags)
3302 return 0; 3537 return 0;
3303} 3538}
3304 3539
3540static noinline_for_stack
3541int prepare_to_relocate(struct reloc_control *rc)
3542{
3543 struct btrfs_trans_handle *trans;
3544 int ret;
3545
3546 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root);
3547 if (!rc->block_rsv)
3548 return -ENOMEM;
3549
3550 /*
3551 * reserve some space for creating reloc trees.
3552 * btrfs_init_reloc_root will use them when there
3553 * is no reservation in transaction handle.
3554 */
3555 ret = btrfs_block_rsv_add(NULL, rc->extent_root, rc->block_rsv,
3556 rc->extent_root->nodesize * 256,
3557 &rc->block_rsv_retries);
3558 if (ret)
3559 return ret;
3560
3561 rc->block_rsv->refill_used = 1;
3562 btrfs_add_durable_block_rsv(rc->extent_root->fs_info, rc->block_rsv);
3563
3564 memset(&rc->cluster, 0, sizeof(rc->cluster));
3565 rc->search_start = rc->block_group->key.objectid;
3566 rc->extents_found = 0;
3567 rc->nodes_relocated = 0;
3568 rc->merging_rsv_size = 0;
3569 rc->block_rsv_retries = 0;
3570
3571 rc->create_reloc_tree = 1;
3572 set_reloc_control(rc);
3573
3574 trans = btrfs_join_transaction(rc->extent_root, 1);
3575 btrfs_commit_transaction(trans, rc->extent_root);
3576 return 0;
3577}
3305 3578
3306static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 3579static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3307{ 3580{
3308 struct rb_root blocks = RB_ROOT; 3581 struct rb_root blocks = RB_ROOT;
3309 struct btrfs_key key; 3582 struct btrfs_key key;
3310 struct file_extent_cluster *cluster;
3311 struct btrfs_trans_handle *trans = NULL; 3583 struct btrfs_trans_handle *trans = NULL;
3312 struct btrfs_path *path; 3584 struct btrfs_path *path;
3313 struct btrfs_extent_item *ei; 3585 struct btrfs_extent_item *ei;
@@ -3317,33 +3589,25 @@ static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3317 int ret; 3589 int ret;
3318 int err = 0; 3590 int err = 0;
3319 3591
3320 cluster = kzalloc(sizeof(*cluster), GFP_NOFS);
3321 if (!cluster)
3322 return -ENOMEM;
3323
3324 path = btrfs_alloc_path(); 3592 path = btrfs_alloc_path();
3325 if (!path) { 3593 if (!path)
3326 kfree(cluster);
3327 return -ENOMEM; 3594 return -ENOMEM;
3328 }
3329
3330 rc->extents_found = 0;
3331 rc->extents_skipped = 0;
3332
3333 rc->search_start = rc->block_group->key.objectid;
3334 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3335 GFP_NOFS);
3336
3337 rc->create_reloc_root = 1;
3338 set_reloc_control(rc);
3339 3595
3340 trans = btrfs_join_transaction(rc->extent_root, 1); 3596 ret = prepare_to_relocate(rc);
3341 btrfs_commit_transaction(trans, rc->extent_root); 3597 if (ret) {
3598 err = ret;
3599 goto out_free;
3600 }
3342 3601
3343 while (1) { 3602 while (1) {
3344 trans = btrfs_start_transaction(rc->extent_root, 0); 3603 trans = btrfs_start_transaction(rc->extent_root, 0);
3345 3604
3346 ret = find_next_extent(trans, rc, path); 3605 if (update_backref_cache(trans, &rc->backref_cache)) {
3606 btrfs_end_transaction(trans, rc->extent_root);
3607 continue;
3608 }
3609
3610 ret = find_next_extent(trans, rc, path, &key);
3347 if (ret < 0) 3611 if (ret < 0)
3348 err = ret; 3612 err = ret;
3349 if (ret != 0) 3613 if (ret != 0)
@@ -3353,9 +3617,7 @@ static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3353 3617
3354 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 3618 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3355 struct btrfs_extent_item); 3619 struct btrfs_extent_item);
3356 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 3620 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3357 item_size = btrfs_item_size_nr(path->nodes[0],
3358 path->slots[0]);
3359 if (item_size >= sizeof(*ei)) { 3621 if (item_size >= sizeof(*ei)) {
3360 flags = btrfs_extent_flags(path->nodes[0], ei); 3622 flags = btrfs_extent_flags(path->nodes[0], ei);
3361 ret = check_extent_flags(flags); 3623 ret = check_extent_flags(flags);
@@ -3396,73 +3658,100 @@ static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3396 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 3658 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3397 ret = add_tree_block(rc, &key, path, &blocks); 3659 ret = add_tree_block(rc, &key, path, &blocks);
3398 } else if (rc->stage == UPDATE_DATA_PTRS && 3660 } else if (rc->stage == UPDATE_DATA_PTRS &&
3399 (flags & BTRFS_EXTENT_FLAG_DATA)) { 3661 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3400 ret = add_data_references(rc, &key, path, &blocks); 3662 ret = add_data_references(rc, &key, path, &blocks);
3401 } else { 3663 } else {
3402 btrfs_release_path(rc->extent_root, path); 3664 btrfs_release_path(rc->extent_root, path);
3403 ret = 0; 3665 ret = 0;
3404 } 3666 }
3405 if (ret < 0) { 3667 if (ret < 0) {
3406 err = 0; 3668 err = ret;
3407 break; 3669 break;
3408 } 3670 }
3409 3671
3410 if (!RB_EMPTY_ROOT(&blocks)) { 3672 if (!RB_EMPTY_ROOT(&blocks)) {
3411 ret = relocate_tree_blocks(trans, rc, &blocks); 3673 ret = relocate_tree_blocks(trans, rc, &blocks);
3412 if (ret < 0) { 3674 if (ret < 0) {
3675 if (ret != -EAGAIN) {
3676 err = ret;
3677 break;
3678 }
3679 rc->extents_found--;
3680 rc->search_start = key.objectid;
3681 }
3682 }
3683
3684 ret = btrfs_block_rsv_check(trans, rc->extent_root,
3685 rc->block_rsv, 0, 5);
3686 if (ret < 0) {
3687 if (ret != -EAGAIN) {
3413 err = ret; 3688 err = ret;
3689 WARN_ON(1);
3414 break; 3690 break;
3415 } 3691 }
3692 rc->commit_transaction = 1;
3416 } 3693 }
3417 3694
3418 nr = trans->blocks_used; 3695 if (rc->commit_transaction) {
3419 btrfs_end_transaction(trans, rc->extent_root); 3696 rc->commit_transaction = 0;
3697 ret = btrfs_commit_transaction(trans, rc->extent_root);
3698 BUG_ON(ret);
3699 } else {
3700 nr = trans->blocks_used;
3701 btrfs_end_transaction_throttle(trans, rc->extent_root);
3702 btrfs_btree_balance_dirty(rc->extent_root, nr);
3703 }
3420 trans = NULL; 3704 trans = NULL;
3421 btrfs_btree_balance_dirty(rc->extent_root, nr);
3422 3705
3423 if (rc->stage == MOVE_DATA_EXTENTS && 3706 if (rc->stage == MOVE_DATA_EXTENTS &&
3424 (flags & BTRFS_EXTENT_FLAG_DATA)) { 3707 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3425 rc->found_file_extent = 1; 3708 rc->found_file_extent = 1;
3426 ret = relocate_data_extent(rc->data_inode, 3709 ret = relocate_data_extent(rc->data_inode,
3427 &key, cluster); 3710 &key, &rc->cluster);
3428 if (ret < 0) { 3711 if (ret < 0) {
3429 err = ret; 3712 err = ret;
3430 break; 3713 break;
3431 } 3714 }
3432 } 3715 }
3433 } 3716 }
3434 btrfs_free_path(path); 3717
3718 btrfs_release_path(rc->extent_root, path);
3719 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3720 GFP_NOFS);
3435 3721
3436 if (trans) { 3722 if (trans) {
3437 nr = trans->blocks_used; 3723 nr = trans->blocks_used;
3438 btrfs_end_transaction(trans, rc->extent_root); 3724 btrfs_end_transaction_throttle(trans, rc->extent_root);
3439 btrfs_btree_balance_dirty(rc->extent_root, nr); 3725 btrfs_btree_balance_dirty(rc->extent_root, nr);
3440 } 3726 }
3441 3727
3442 if (!err) { 3728 if (!err) {
3443 ret = relocate_file_extent_cluster(rc->data_inode, cluster); 3729 ret = relocate_file_extent_cluster(rc->data_inode,
3730 &rc->cluster);
3444 if (ret < 0) 3731 if (ret < 0)
3445 err = ret; 3732 err = ret;
3446 } 3733 }
3447 3734
3448 kfree(cluster); 3735 rc->create_reloc_tree = 0;
3736 set_reloc_control(rc);
3449 3737
3450 rc->create_reloc_root = 0; 3738 backref_cache_cleanup(&rc->backref_cache);
3451 smp_mb(); 3739 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3452 3740
3453 if (rc->extents_found > 0) { 3741 err = prepare_to_merge(rc, err);
3454 trans = btrfs_join_transaction(rc->extent_root, 1);
3455 btrfs_commit_transaction(trans, rc->extent_root);
3456 }
3457 3742
3458 merge_reloc_roots(rc); 3743 merge_reloc_roots(rc);
3459 3744
3745 rc->merge_reloc_tree = 0;
3460 unset_reloc_control(rc); 3746 unset_reloc_control(rc);
3747 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3461 3748
3462 /* get rid of pinned extents */ 3749 /* get rid of pinned extents */
3463 trans = btrfs_join_transaction(rc->extent_root, 1); 3750 trans = btrfs_join_transaction(rc->extent_root, 1);
3464 btrfs_commit_transaction(trans, rc->extent_root); 3751 btrfs_commit_transaction(trans, rc->extent_root);
3465 3752out_free:
3753 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
3754 btrfs_free_path(path);
3466 return err; 3755 return err;
3467} 3756}
3468 3757
@@ -3488,7 +3777,8 @@ static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3488 btrfs_set_inode_generation(leaf, item, 1); 3777 btrfs_set_inode_generation(leaf, item, 1);
3489 btrfs_set_inode_size(leaf, item, 0); 3778 btrfs_set_inode_size(leaf, item, 0);
3490 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 3779 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3491 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS); 3780 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3781 BTRFS_INODE_PREALLOC);
3492 btrfs_mark_buffer_dirty(leaf); 3782 btrfs_mark_buffer_dirty(leaf);
3493 btrfs_release_path(root, path); 3783 btrfs_release_path(root, path);
3494out: 3784out:
@@ -3500,8 +3790,9 @@ out:
3500 * helper to create inode for data relocation. 3790 * helper to create inode for data relocation.
3501 * the inode is in data relocation tree and its link count is 0 3791 * the inode is in data relocation tree and its link count is 0
3502 */ 3792 */
3503static struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 3793static noinline_for_stack
3504 struct btrfs_block_group_cache *group) 3794struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3795 struct btrfs_block_group_cache *group)
3505{ 3796{
3506 struct inode *inode = NULL; 3797 struct inode *inode = NULL;
3507 struct btrfs_trans_handle *trans; 3798 struct btrfs_trans_handle *trans;
@@ -3516,7 +3807,8 @@ static struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3516 return ERR_CAST(root); 3807 return ERR_CAST(root);
3517 3808
3518 trans = btrfs_start_transaction(root, 6); 3809 trans = btrfs_start_transaction(root, 6);
3519 BUG_ON(!trans); 3810 if (IS_ERR(trans))
3811 return ERR_CAST(trans);
3520 3812
3521 err = btrfs_find_free_objectid(trans, root, objectid, &objectid); 3813 err = btrfs_find_free_objectid(trans, root, objectid, &objectid);
3522 if (err) 3814 if (err)
@@ -3536,7 +3828,6 @@ static struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3536out: 3828out:
3537 nr = trans->blocks_used; 3829 nr = trans->blocks_used;
3538 btrfs_end_transaction(trans, root); 3830 btrfs_end_transaction(trans, root);
3539
3540 btrfs_btree_balance_dirty(root, nr); 3831 btrfs_btree_balance_dirty(root, nr);
3541 if (err) { 3832 if (err) {
3542 if (inode) 3833 if (inode)
@@ -3546,6 +3837,21 @@ out:
3546 return inode; 3837 return inode;
3547} 3838}
3548 3839
3840static struct reloc_control *alloc_reloc_control(void)
3841{
3842 struct reloc_control *rc;
3843
3844 rc = kzalloc(sizeof(*rc), GFP_NOFS);
3845 if (!rc)
3846 return NULL;
3847
3848 INIT_LIST_HEAD(&rc->reloc_roots);
3849 backref_cache_init(&rc->backref_cache);
3850 mapping_tree_init(&rc->reloc_root_tree);
3851 extent_io_tree_init(&rc->processed_blocks, NULL, GFP_NOFS);
3852 return rc;
3853}
3854
3549/* 3855/*
3550 * function to relocate all extents in a block group. 3856 * function to relocate all extents in a block group.
3551 */ 3857 */
@@ -3557,15 +3863,12 @@ int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3557 int rw = 0; 3863 int rw = 0;
3558 int err = 0; 3864 int err = 0;
3559 3865
3560 rc = kzalloc(sizeof(*rc), GFP_NOFS); 3866 rc = alloc_reloc_control();
3561 if (!rc) 3867 if (!rc)
3562 return -ENOMEM; 3868 return -ENOMEM;
3563 3869
3564 mapping_tree_init(&rc->reloc_root_tree);
3565 extent_io_tree_init(&rc->processed_blocks, NULL, GFP_NOFS);
3566 INIT_LIST_HEAD(&rc->reloc_roots);
3567
3568 rc->extent_root = extent_root; 3870 rc->extent_root = extent_root;
3871
3569 rc->block_group = btrfs_lookup_block_group(fs_info, group_start); 3872 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
3570 BUG_ON(!rc->block_group); 3873 BUG_ON(!rc->block_group);
3571 3874
@@ -3578,9 +3881,6 @@ int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3578 rw = 1; 3881 rw = 1;
3579 } 3882 }
3580 3883
3581 btrfs_init_workers(&rc->workers, "relocate",
3582 fs_info->thread_pool_size, NULL);
3583
3584 rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 3884 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
3585 if (IS_ERR(rc->data_inode)) { 3885 if (IS_ERR(rc->data_inode)) {
3586 err = PTR_ERR(rc->data_inode); 3886 err = PTR_ERR(rc->data_inode);
@@ -3596,9 +3896,6 @@ int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3596 btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0); 3896 btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0);
3597 3897
3598 while (1) { 3898 while (1) {
3599 rc->extents_found = 0;
3600 rc->extents_skipped = 0;
3601
3602 mutex_lock(&fs_info->cleaner_mutex); 3899 mutex_lock(&fs_info->cleaner_mutex);
3603 3900
3604 btrfs_clean_old_snapshots(fs_info->tree_root); 3901 btrfs_clean_old_snapshots(fs_info->tree_root);
@@ -3607,7 +3904,7 @@ int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3607 mutex_unlock(&fs_info->cleaner_mutex); 3904 mutex_unlock(&fs_info->cleaner_mutex);
3608 if (ret < 0) { 3905 if (ret < 0) {
3609 err = ret; 3906 err = ret;
3610 break; 3907 goto out;
3611 } 3908 }
3612 3909
3613 if (rc->extents_found == 0) 3910 if (rc->extents_found == 0)
@@ -3621,18 +3918,6 @@ int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3621 invalidate_mapping_pages(rc->data_inode->i_mapping, 3918 invalidate_mapping_pages(rc->data_inode->i_mapping,
3622 0, -1); 3919 0, -1);
3623 rc->stage = UPDATE_DATA_PTRS; 3920 rc->stage = UPDATE_DATA_PTRS;
3624 } else if (rc->stage == UPDATE_DATA_PTRS &&
3625 rc->extents_skipped >= rc->extents_found) {
3626 iput(rc->data_inode);
3627 rc->data_inode = create_reloc_inode(fs_info,
3628 rc->block_group);
3629 if (IS_ERR(rc->data_inode)) {
3630 err = PTR_ERR(rc->data_inode);
3631 rc->data_inode = NULL;
3632 break;
3633 }
3634 rc->stage = MOVE_DATA_EXTENTS;
3635 rc->found_file_extent = 0;
3636 } 3921 }
3637 } 3922 }
3638 3923
@@ -3648,7 +3933,6 @@ out:
3648 if (err && rw) 3933 if (err && rw)
3649 btrfs_set_block_group_rw(extent_root, rc->block_group); 3934 btrfs_set_block_group_rw(extent_root, rc->block_group);
3650 iput(rc->data_inode); 3935 iput(rc->data_inode);
3651 btrfs_stop_workers(&rc->workers);
3652 btrfs_put_block_group(rc->block_group); 3936 btrfs_put_block_group(rc->block_group);
3653 kfree(rc); 3937 kfree(rc);
3654 return err; 3938 return err;
@@ -3752,20 +4036,20 @@ int btrfs_recover_relocation(struct btrfs_root *root)
3752 if (list_empty(&reloc_roots)) 4036 if (list_empty(&reloc_roots))
3753 goto out; 4037 goto out;
3754 4038
3755 rc = kzalloc(sizeof(*rc), GFP_NOFS); 4039 rc = alloc_reloc_control();
3756 if (!rc) { 4040 if (!rc) {
3757 err = -ENOMEM; 4041 err = -ENOMEM;
3758 goto out; 4042 goto out;
3759 } 4043 }
3760 4044
3761 mapping_tree_init(&rc->reloc_root_tree);
3762 INIT_LIST_HEAD(&rc->reloc_roots);
3763 btrfs_init_workers(&rc->workers, "relocate",
3764 root->fs_info->thread_pool_size, NULL);
3765 rc->extent_root = root->fs_info->extent_root; 4045 rc->extent_root = root->fs_info->extent_root;
3766 4046
3767 set_reloc_control(rc); 4047 set_reloc_control(rc);
3768 4048
4049 trans = btrfs_join_transaction(rc->extent_root, 1);
4050
4051 rc->merge_reloc_tree = 1;
4052
3769 while (!list_empty(&reloc_roots)) { 4053 while (!list_empty(&reloc_roots)) {
3770 reloc_root = list_entry(reloc_roots.next, 4054 reloc_root = list_entry(reloc_roots.next,
3771 struct btrfs_root, root_list); 4055 struct btrfs_root, root_list);
@@ -3785,20 +4069,16 @@ int btrfs_recover_relocation(struct btrfs_root *root)
3785 fs_root->reloc_root = reloc_root; 4069 fs_root->reloc_root = reloc_root;
3786 } 4070 }
3787 4071
3788 trans = btrfs_start_transaction(rc->extent_root, 1);
3789 btrfs_commit_transaction(trans, rc->extent_root); 4072 btrfs_commit_transaction(trans, rc->extent_root);
3790 4073
3791 merge_reloc_roots(rc); 4074 merge_reloc_roots(rc);
3792 4075
3793 unset_reloc_control(rc); 4076 unset_reloc_control(rc);
3794 4077
3795 trans = btrfs_start_transaction(rc->extent_root, 1); 4078 trans = btrfs_join_transaction(rc->extent_root, 1);
3796 btrfs_commit_transaction(trans, rc->extent_root); 4079 btrfs_commit_transaction(trans, rc->extent_root);
3797out: 4080out:
3798 if (rc) { 4081 kfree(rc);
3799 btrfs_stop_workers(&rc->workers);
3800 kfree(rc);
3801 }
3802 while (!list_empty(&reloc_roots)) { 4082 while (!list_empty(&reloc_roots)) {
3803 reloc_root = list_entry(reloc_roots.next, 4083 reloc_root = list_entry(reloc_roots.next,
3804 struct btrfs_root, root_list); 4084 struct btrfs_root, root_list);
@@ -3864,3 +4144,130 @@ int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
3864 btrfs_put_ordered_extent(ordered); 4144 btrfs_put_ordered_extent(ordered);
3865 return 0; 4145 return 0;
3866} 4146}
4147
4148void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4149 struct btrfs_root *root, struct extent_buffer *buf,
4150 struct extent_buffer *cow)
4151{
4152 struct reloc_control *rc;
4153 struct backref_node *node;
4154 int first_cow = 0;
4155 int level;
4156 int ret;
4157
4158 rc = root->fs_info->reloc_ctl;
4159 if (!rc)
4160 return;
4161
4162 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4163 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4164
4165 level = btrfs_header_level(buf);
4166 if (btrfs_header_generation(buf) <=
4167 btrfs_root_last_snapshot(&root->root_item))
4168 first_cow = 1;
4169
4170 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4171 rc->create_reloc_tree) {
4172 WARN_ON(!first_cow && level == 0);
4173
4174 node = rc->backref_cache.path[level];
4175 BUG_ON(node->bytenr != buf->start &&
4176 node->new_bytenr != buf->start);
4177
4178 drop_node_buffer(node);
4179 extent_buffer_get(cow);
4180 node->eb = cow;
4181 node->new_bytenr = cow->start;
4182
4183 if (!node->pending) {
4184 list_move_tail(&node->list,
4185 &rc->backref_cache.pending[level]);
4186 node->pending = 1;
4187 }
4188
4189 if (first_cow)
4190 __mark_block_processed(rc, node);
4191
4192 if (first_cow && level > 0)
4193 rc->nodes_relocated += buf->len;
4194 }
4195
4196 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4197 ret = replace_file_extents(trans, rc, root, cow);
4198 BUG_ON(ret);
4199 }
4200}
4201
4202/*
4203 * called before creating snapshot. it calculates metadata reservation
4204 * requried for relocating tree blocks in the snapshot
4205 */
4206void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4207 struct btrfs_pending_snapshot *pending,
4208 u64 *bytes_to_reserve)
4209{
4210 struct btrfs_root *root;
4211 struct reloc_control *rc;
4212
4213 root = pending->root;
4214 if (!root->reloc_root)
4215 return;
4216
4217 rc = root->fs_info->reloc_ctl;
4218 if (!rc->merge_reloc_tree)
4219 return;
4220
4221 root = root->reloc_root;
4222 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4223 /*
4224 * relocation is in the stage of merging trees. the space
4225 * used by merging a reloc tree is twice the size of
4226 * relocated tree nodes in the worst case. half for cowing
4227 * the reloc tree, half for cowing the fs tree. the space
4228 * used by cowing the reloc tree will be freed after the
4229 * tree is dropped. if we create snapshot, cowing the fs
4230 * tree may use more space than it frees. so we need
4231 * reserve extra space.
4232 */
4233 *bytes_to_reserve += rc->nodes_relocated;
4234}
4235
4236/*
4237 * called after snapshot is created. migrate block reservation
4238 * and create reloc root for the newly created snapshot
4239 */
4240void btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4241 struct btrfs_pending_snapshot *pending)
4242{
4243 struct btrfs_root *root = pending->root;
4244 struct btrfs_root *reloc_root;
4245 struct btrfs_root *new_root;
4246 struct reloc_control *rc;
4247 int ret;
4248
4249 if (!root->reloc_root)
4250 return;
4251
4252 rc = root->fs_info->reloc_ctl;
4253 rc->merging_rsv_size += rc->nodes_relocated;
4254
4255 if (rc->merge_reloc_tree) {
4256 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4257 rc->block_rsv,
4258 rc->nodes_relocated);
4259 BUG_ON(ret);
4260 }
4261
4262 new_root = pending->snap;
4263 reloc_root = create_reloc_root(trans, root->reloc_root,
4264 new_root->root_key.objectid);
4265
4266 __add_reloc_root(reloc_root);
4267 new_root->reloc_root = reloc_root;
4268
4269 if (rc->create_reloc_tree) {
4270 ret = clone_backref_node(trans, rc, root, reloc_root);
4271 BUG_ON(ret);
4272 }
4273}
diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
index cfe7f588ef05..c346d320173a 100644
--- a/fs/btrfs/transaction.c
+++ b/fs/btrfs/transaction.c
@@ -853,6 +853,7 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
853 goto fail; 853 goto fail;
854 } 854 }
855 855
856 btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
856 btrfs_orphan_pre_snapshot(trans, pending, &to_reserve); 857 btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
857 858
858 if (to_reserve > 0) { 859 if (to_reserve > 0) {
@@ -924,6 +925,7 @@ static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
924 pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key); 925 pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
925 BUG_ON(IS_ERR(pending->snap)); 926 BUG_ON(IS_ERR(pending->snap));
926 927
928 btrfs_reloc_post_snapshot(trans, pending);
927 btrfs_orphan_post_snapshot(trans, pending); 929 btrfs_orphan_post_snapshot(trans, pending);
928fail: 930fail:
929 kfree(new_root_item); 931 kfree(new_root_item);
@@ -1213,9 +1215,9 @@ int btrfs_clean_old_snapshots(struct btrfs_root *root)
1213 1215
1214 if (btrfs_header_backref_rev(root->node) < 1216 if (btrfs_header_backref_rev(root->node) <
1215 BTRFS_MIXED_BACKREF_REV) 1217 BTRFS_MIXED_BACKREF_REV)
1216 btrfs_drop_snapshot(root, 0); 1218 btrfs_drop_snapshot(root, NULL, 0);
1217 else 1219 else
1218 btrfs_drop_snapshot(root, 1); 1220 btrfs_drop_snapshot(root, NULL, 1);
1219 } 1221 }
1220 return 0; 1222 return 0;
1221} 1223}
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-11 14:45:02 -0500