aboutsummaryrefslogtreecommitdiffstats
path: root/fs/btrfs/relocation.c
diff options
context:
space:
mode:
Diffstat (limited to 'fs/btrfs/relocation.c')
-rw-r--r--fs/btrfs/relocation.c1975
1 files changed, 1217 insertions, 758 deletions
diff --git a/fs/btrfs/relocation.c b/fs/btrfs/relocation.c
index 0b23942cbc0d..b37d723b9d4a 100644
--- a/fs/btrfs/relocation.c
+++ b/fs/btrfs/relocation.c
@@ -21,6 +21,7 @@
21#include <linux/writeback.h> 21#include <linux/writeback.h>
22#include <linux/blkdev.h> 22#include <linux/blkdev.h>
23#include <linux/rbtree.h> 23#include <linux/rbtree.h>
24#include <linux/slab.h>
24#include "ctree.h" 25#include "ctree.h"
25#include "disk-io.h" 26#include "disk-io.h"
26#include "transaction.h" 27#include "transaction.h"
@@ -43,8 +44,12 @@ struct tree_entry {
43struct backref_node { 44struct backref_node {
44 struct rb_node rb_node; 45 struct rb_node rb_node;
45 u64 bytenr; 46 u64 bytenr;
46 /* objectid tree block owner */ 47
48 u64 new_bytenr;
49 /* objectid of tree block owner, can be not uptodate */
47 u64 owner; 50 u64 owner;
51 /* link to pending, changed or detached list */
52 struct list_head list;
48 /* list of upper level blocks reference this block */ 53 /* list of upper level blocks reference this block */
49 struct list_head upper; 54 struct list_head upper;
50 /* list of child blocks in the cache */ 55 /* list of child blocks in the cache */
@@ -55,9 +60,9 @@ struct backref_node {
55 struct extent_buffer *eb; 60 struct extent_buffer *eb;
56 /* level of tree block */ 61 /* level of tree block */
57 unsigned int level:8; 62 unsigned int level:8;
58 /* 1 if the block is root of old snapshot */ 63 /* is the block in non-reference counted tree */
59 unsigned int old_root:1; 64 unsigned int cowonly:1;
60 /* 1 if no child blocks in the cache */ 65 /* 1 if no child node in the cache */
61 unsigned int lowest:1; 66 unsigned int lowest:1;
62 /* is the extent buffer locked */ 67 /* is the extent buffer locked */
63 unsigned int locked:1; 68 unsigned int locked:1;
@@ -65,6 +70,16 @@ struct backref_node {
65 unsigned int processed:1; 70 unsigned int processed:1;
66 /* have backrefs of this block been checked */ 71 /* have backrefs of this block been checked */
67 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;
68}; 83};
69 84
70/* 85/*
@@ -73,7 +88,6 @@ struct backref_node {
73struct backref_edge { 88struct backref_edge {
74 struct list_head list[2]; 89 struct list_head list[2];
75 struct backref_node *node[2]; 90 struct backref_node *node[2];
76 u64 blockptr;
77}; 91};
78 92
79#define LOWER 0 93#define LOWER 0
@@ -82,9 +96,25 @@ struct backref_edge {
82struct backref_cache { 96struct backref_cache {
83 /* red black tree of all backref nodes in the cache */ 97 /* red black tree of all backref nodes in the cache */
84 struct rb_root rb_root; 98 struct rb_root rb_root;
85 /* 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 */
86 struct list_head pending[BTRFS_MAX_LEVEL]; 106 struct list_head pending[BTRFS_MAX_LEVEL];
87 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;
88}; 118};
89 119
90/* 120/*
@@ -112,15 +142,6 @@ struct tree_block {
112 unsigned int key_ready:1; 142 unsigned int key_ready:1;
113}; 143};
114 144
115/* inode vector */
116#define INODEVEC_SIZE 16
117
118struct inodevec {
119 struct list_head list;
120 struct inode *inode[INODEVEC_SIZE];
121 int nr;
122};
123
124#define MAX_EXTENTS 128 145#define MAX_EXTENTS 128
125 146
126struct file_extent_cluster { 147struct file_extent_cluster {
@@ -137,36 +158,43 @@ struct reloc_control {
137 struct btrfs_root *extent_root; 158 struct btrfs_root *extent_root;
138 /* inode for moving data */ 159 /* inode for moving data */
139 struct inode *data_inode; 160 struct inode *data_inode;
140 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;
141 /* tree blocks have been processed */ 167 /* tree blocks have been processed */
142 struct extent_io_tree processed_blocks; 168 struct extent_io_tree processed_blocks;
143 /* map start of tree root to corresponding reloc tree */ 169 /* map start of tree root to corresponding reloc tree */
144 struct mapping_tree reloc_root_tree; 170 struct mapping_tree reloc_root_tree;
145 /* list of reloc trees */ 171 /* list of reloc trees */
146 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
147 u64 search_start; 178 u64 search_start;
148 u64 extents_found; 179 u64 extents_found;
149 u64 extents_skipped; 180
150 int stage; 181 int block_rsv_retries;
151 int create_reloc_root; 182
183 unsigned int stage:8;
184 unsigned int create_reloc_tree:1;
185 unsigned int merge_reloc_tree:1;
152 unsigned int found_file_extent:1; 186 unsigned int found_file_extent:1;
153 unsigned int found_old_snapshot:1; 187 unsigned int commit_transaction:1;
154}; 188};
155 189
156/* stages of data relocation */ 190/* stages of data relocation */
157#define MOVE_DATA_EXTENTS 0 191#define MOVE_DATA_EXTENTS 0
158#define UPDATE_DATA_PTRS 1 192#define UPDATE_DATA_PTRS 1
159 193
160/* 194static void remove_backref_node(struct backref_cache *cache,
161 * merge reloc tree to corresponding fs tree in worker threads 195 struct backref_node *node);
162 */ 196static void __mark_block_processed(struct reloc_control *rc,
163struct async_merge { 197 struct backref_node *node);
164 struct btrfs_work work;
165 struct reloc_control *rc;
166 struct btrfs_root *root;
167 struct completion *done;
168 atomic_t *num_pending;
169};
170 198
171static void mapping_tree_init(struct mapping_tree *tree) 199static void mapping_tree_init(struct mapping_tree *tree)
172{ 200{
@@ -180,15 +208,80 @@ static void backref_cache_init(struct backref_cache *cache)
180 cache->rb_root = RB_ROOT; 208 cache->rb_root = RB_ROOT;
181 for (i = 0; i < BTRFS_MAX_LEVEL; i++) 209 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
182 INIT_LIST_HEAD(&cache->pending[i]); 210 INIT_LIST_HEAD(&cache->pending[i]);
183 spin_lock_init(&cache->lock); 211 INIT_LIST_HEAD(&cache->changed);
212 INIT_LIST_HEAD(&cache->detached);
213 INIT_LIST_HEAD(&cache->leaves);
214}
215
216static void backref_cache_cleanup(struct backref_cache *cache)
217{
218 struct backref_node *node;
219 int i;
220
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;
184} 276}
185 277
186static void backref_node_init(struct backref_node *node) 278static void free_backref_edge(struct backref_cache *cache,
279 struct backref_edge *edge)
187{ 280{
188 memset(node, 0, sizeof(*node)); 281 if (edge) {
189 INIT_LIST_HEAD(&node->upper); 282 cache->nr_edges--;
190 INIT_LIST_HEAD(&node->lower); 283 kfree(edge);
191 RB_CLEAR_NODE(&node->rb_node); 284 }
192} 285}
193 286
194static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr, 287static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr,
@@ -249,6 +342,7 @@ static struct backref_node *walk_up_backref(struct backref_node *node,
249 edges[idx++] = edge; 342 edges[idx++] = edge;
250 node = edge->node[UPPER]; 343 node = edge->node[UPPER];
251 } 344 }
345 BUG_ON(node->detached);
252 *index = idx; 346 *index = idx;
253 return node; 347 return node;
254} 348}
@@ -280,13 +374,18 @@ static struct backref_node *walk_down_backref(struct backref_edge *edges[],
280 return NULL; 374 return NULL;
281} 375}
282 376
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
283static void drop_node_buffer(struct backref_node *node) 385static void drop_node_buffer(struct backref_node *node)
284{ 386{
285 if (node->eb) { 387 if (node->eb) {
286 if (node->locked) { 388 unlock_node_buffer(node);
287 btrfs_tree_unlock(node->eb);
288 node->locked = 0;
289 }
290 free_extent_buffer(node->eb); 389 free_extent_buffer(node->eb);
291 node->eb = NULL; 390 node->eb = NULL;
292 } 391 }
@@ -295,14 +394,14 @@ static void drop_node_buffer(struct backref_node *node)
295static void drop_backref_node(struct backref_cache *tree, 394static void drop_backref_node(struct backref_cache *tree,
296 struct backref_node *node) 395 struct backref_node *node)
297{ 396{
298 BUG_ON(!node->lowest);
299 BUG_ON(!list_empty(&node->upper)); 397 BUG_ON(!list_empty(&node->upper));
300 398
301 drop_node_buffer(node); 399 drop_node_buffer(node);
400 list_del(&node->list);
302 list_del(&node->lower); 401 list_del(&node->lower);
303 402 if (!RB_EMPTY_NODE(&node->rb_node))
304 rb_erase(&node->rb_node, &tree->rb_root); 403 rb_erase(&node->rb_node, &tree->rb_root);
305 kfree(node); 404 free_backref_node(tree, node);
306} 405}
307 406
308/* 407/*
@@ -317,27 +416,121 @@ static void remove_backref_node(struct backref_cache *cache,
317 if (!node) 416 if (!node)
318 return; 417 return;
319 418
320 BUG_ON(!node->lowest); 419 BUG_ON(!node->lowest && !node->detached);
321 while (!list_empty(&node->upper)) { 420 while (!list_empty(&node->upper)) {
322 edge = list_entry(node->upper.next, struct backref_edge, 421 edge = list_entry(node->upper.next, struct backref_edge,
323 list[LOWER]); 422 list[LOWER]);
324 upper = edge->node[UPPER]; 423 upper = edge->node[UPPER];
325 list_del(&edge->list[LOWER]); 424 list_del(&edge->list[LOWER]);
326 list_del(&edge->list[UPPER]); 425 list_del(&edge->list[UPPER]);
327 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 }
328 /* 435 /*
329 * add the node to pending list if no other 436 * add the node to leaf node list if no other
330 * child block cached. 437 * child block cached.
331 */ 438 */
332 if (list_empty(&upper->lower)) { 439 if (list_empty(&upper->lower)) {
333 list_add_tail(&upper->lower, 440 list_add_tail(&upper->lower, &cache->leaves);
334 &cache->pending[upper->level]);
335 upper->lowest = 1; 441 upper->lowest = 1;
336 } 442 }
337 } 443 }
444
338 drop_backref_node(cache, node); 445 drop_backref_node(cache, node);
339} 446}
340 447
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
341/* 534/*
342 * find reloc tree by address of tree root 535 * find reloc tree by address of tree root
343 */ 536 */
@@ -452,11 +645,12 @@ int find_inline_backref(struct extent_buffer *leaf, int slot,
452 * for all upper level blocks that directly/indirectly reference the 645 * for all upper level blocks that directly/indirectly reference the
453 * block are also cached. 646 * block are also cached.
454 */ 647 */
455static struct backref_node *build_backref_tree(struct reloc_control *rc, 648static noinline_for_stack
456 struct backref_cache *cache, 649struct backref_node *build_backref_tree(struct reloc_control *rc,
457 struct btrfs_key *node_key, 650 struct btrfs_key *node_key,
458 int level, u64 bytenr) 651 int level, u64 bytenr)
459{ 652{
653 struct backref_cache *cache = &rc->backref_cache;
460 struct btrfs_path *path1; 654 struct btrfs_path *path1;
461 struct btrfs_path *path2; 655 struct btrfs_path *path2;
462 struct extent_buffer *eb; 656 struct extent_buffer *eb;
@@ -472,6 +666,8 @@ static struct backref_node *build_backref_tree(struct reloc_control *rc,
472 unsigned long end; 666 unsigned long end;
473 unsigned long ptr; 667 unsigned long ptr;
474 LIST_HEAD(list); 668 LIST_HEAD(list);
669 LIST_HEAD(useless);
670 int cowonly;
475 int ret; 671 int ret;
476 int err = 0; 672 int err = 0;
477 673
@@ -482,15 +678,13 @@ static struct backref_node *build_backref_tree(struct reloc_control *rc,
482 goto out; 678 goto out;
483 } 679 }
484 680
485 node = kmalloc(sizeof(*node), GFP_NOFS); 681 node = alloc_backref_node(cache);
486 if (!node) { 682 if (!node) {
487 err = -ENOMEM; 683 err = -ENOMEM;
488 goto out; 684 goto out;
489 } 685 }
490 686
491 backref_node_init(node);
492 node->bytenr = bytenr; 687 node->bytenr = bytenr;
493 node->owner = 0;
494 node->level = level; 688 node->level = level;
495 node->lowest = 1; 689 node->lowest = 1;
496 cur = node; 690 cur = node;
@@ -586,17 +780,21 @@ again:
586#ifdef BTRFS_COMPAT_EXTENT_TREE_V0 780#ifdef BTRFS_COMPAT_EXTENT_TREE_V0
587 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY || 781 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY ||
588 key.type == BTRFS_EXTENT_REF_V0_KEY) { 782 key.type == BTRFS_EXTENT_REF_V0_KEY) {
589 if (key.objectid == key.offset && 783 if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
590 key.type == BTRFS_EXTENT_REF_V0_KEY) {
591 struct btrfs_extent_ref_v0 *ref0; 784 struct btrfs_extent_ref_v0 *ref0;
592 ref0 = btrfs_item_ptr(eb, path1->slots[0], 785 ref0 = btrfs_item_ptr(eb, path1->slots[0],
593 struct btrfs_extent_ref_v0); 786 struct btrfs_extent_ref_v0);
594 root = find_tree_root(rc, eb, ref0); 787 if (key.objectid == key.offset) {
595 if (root) 788 root = find_tree_root(rc, eb, ref0);
596 cur->root = root; 789 if (root && !should_ignore_root(root))
597 else 790 cur->root = root;
598 cur->old_root = 1; 791 else
599 break; 792 list_add(&cur->list, &useless);
793 break;
794 }
795 if (is_cowonly_root(btrfs_ref_root_v0(eb,
796 ref0)))
797 cur->cowonly = 1;
600 } 798 }
601#else 799#else
602 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 800 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
@@ -613,22 +811,20 @@ again:
613 break; 811 break;
614 } 812 }
615 813
616 edge = kzalloc(sizeof(*edge), GFP_NOFS); 814 edge = alloc_backref_edge(cache);
617 if (!edge) { 815 if (!edge) {
618 err = -ENOMEM; 816 err = -ENOMEM;
619 goto out; 817 goto out;
620 } 818 }
621 rb_node = tree_search(&cache->rb_root, key.offset); 819 rb_node = tree_search(&cache->rb_root, key.offset);
622 if (!rb_node) { 820 if (!rb_node) {
623 upper = kmalloc(sizeof(*upper), GFP_NOFS); 821 upper = alloc_backref_node(cache);
624 if (!upper) { 822 if (!upper) {
625 kfree(edge); 823 free_backref_edge(cache, edge);
626 err = -ENOMEM; 824 err = -ENOMEM;
627 goto out; 825 goto out;
628 } 826 }
629 backref_node_init(upper);
630 upper->bytenr = key.offset; 827 upper->bytenr = key.offset;
631 upper->owner = 0;
632 upper->level = cur->level + 1; 828 upper->level = cur->level + 1;
633 /* 829 /*
634 * backrefs for the upper level block isn't 830 * backrefs for the upper level block isn't
@@ -638,11 +834,12 @@ again:
638 } else { 834 } else {
639 upper = rb_entry(rb_node, struct backref_node, 835 upper = rb_entry(rb_node, struct backref_node,
640 rb_node); 836 rb_node);
837 BUG_ON(!upper->checked);
641 INIT_LIST_HEAD(&edge->list[UPPER]); 838 INIT_LIST_HEAD(&edge->list[UPPER]);
642 } 839 }
643 list_add(&edge->list[LOWER], &cur->upper); 840 list_add_tail(&edge->list[LOWER], &cur->upper);
644 edge->node[UPPER] = upper;
645 edge->node[LOWER] = cur; 841 edge->node[LOWER] = cur;
842 edge->node[UPPER] = upper;
646 843
647 goto next; 844 goto next;
648 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 845 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
@@ -656,11 +853,17 @@ again:
656 goto out; 853 goto out;
657 } 854 }
658 855
856 if (!root->ref_cows)
857 cur->cowonly = 1;
858
659 if (btrfs_root_level(&root->root_item) == cur->level) { 859 if (btrfs_root_level(&root->root_item) == cur->level) {
660 /* tree root */ 860 /* tree root */
661 BUG_ON(btrfs_root_bytenr(&root->root_item) != 861 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
662 cur->bytenr); 862 cur->bytenr);
663 cur->root = root; 863 if (should_ignore_root(root))
864 list_add(&cur->list, &useless);
865 else
866 cur->root = root;
664 break; 867 break;
665 } 868 }
666 869
@@ -691,11 +894,14 @@ again:
691 if (!path2->nodes[level]) { 894 if (!path2->nodes[level]) {
692 BUG_ON(btrfs_root_bytenr(&root->root_item) != 895 BUG_ON(btrfs_root_bytenr(&root->root_item) !=
693 lower->bytenr); 896 lower->bytenr);
694 lower->root = root; 897 if (should_ignore_root(root))
898 list_add(&lower->list, &useless);
899 else
900 lower->root = root;
695 break; 901 break;
696 } 902 }
697 903
698 edge = kzalloc(sizeof(*edge), GFP_NOFS); 904 edge = alloc_backref_edge(cache);
699 if (!edge) { 905 if (!edge) {
700 err = -ENOMEM; 906 err = -ENOMEM;
701 goto out; 907 goto out;
@@ -704,16 +910,17 @@ again:
704 eb = path2->nodes[level]; 910 eb = path2->nodes[level];
705 rb_node = tree_search(&cache->rb_root, eb->start); 911 rb_node = tree_search(&cache->rb_root, eb->start);
706 if (!rb_node) { 912 if (!rb_node) {
707 upper = kmalloc(sizeof(*upper), GFP_NOFS); 913 upper = alloc_backref_node(cache);
708 if (!upper) { 914 if (!upper) {
709 kfree(edge); 915 free_backref_edge(cache, edge);
710 err = -ENOMEM; 916 err = -ENOMEM;
711 goto out; 917 goto out;
712 } 918 }
713 backref_node_init(upper);
714 upper->bytenr = eb->start; 919 upper->bytenr = eb->start;
715 upper->owner = btrfs_header_owner(eb); 920 upper->owner = btrfs_header_owner(eb);
716 upper->level = lower->level + 1; 921 upper->level = lower->level + 1;
922 if (!root->ref_cows)
923 upper->cowonly = 1;
717 924
718 /* 925 /*
719 * if we know the block isn't shared 926 * if we know the block isn't shared
@@ -743,10 +950,12 @@ again:
743 rb_node); 950 rb_node);
744 BUG_ON(!upper->checked); 951 BUG_ON(!upper->checked);
745 INIT_LIST_HEAD(&edge->list[UPPER]); 952 INIT_LIST_HEAD(&edge->list[UPPER]);
953 if (!upper->owner)
954 upper->owner = btrfs_header_owner(eb);
746 } 955 }
747 list_add_tail(&edge->list[LOWER], &lower->upper); 956 list_add_tail(&edge->list[LOWER], &lower->upper);
748 edge->node[UPPER] = upper;
749 edge->node[LOWER] = lower; 957 edge->node[LOWER] = lower;
958 edge->node[UPPER] = upper;
750 959
751 if (rb_node) 960 if (rb_node)
752 break; 961 break;
@@ -784,8 +993,13 @@ next:
784 * into the cache. 993 * into the cache.
785 */ 994 */
786 BUG_ON(!node->checked); 995 BUG_ON(!node->checked);
787 rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node); 996 cowonly = node->cowonly;
788 BUG_ON(rb_node); 997 if (!cowonly) {
998 rb_node = tree_insert(&cache->rb_root, node->bytenr,
999 &node->rb_node);
1000 BUG_ON(rb_node);
1001 list_add_tail(&node->lower, &cache->leaves);
1002 }
789 1003
790 list_for_each_entry(edge, &node->upper, list[LOWER]) 1004 list_for_each_entry(edge, &node->upper, list[LOWER])
791 list_add_tail(&edge->list[UPPER], &list); 1005 list_add_tail(&edge->list[UPPER], &list);
@@ -794,6 +1008,14 @@ next:
794 edge = list_entry(list.next, struct backref_edge, list[UPPER]); 1008 edge = list_entry(list.next, struct backref_edge, list[UPPER]);
795 list_del_init(&edge->list[UPPER]); 1009 list_del_init(&edge->list[UPPER]);
796 upper = edge->node[UPPER]; 1010 upper = edge->node[UPPER];
1011 if (upper->detached) {
1012 list_del(&edge->list[LOWER]);
1013 lower = edge->node[LOWER];
1014 free_backref_edge(cache, edge);
1015 if (list_empty(&lower->upper))
1016 list_add(&lower->list, &useless);
1017 continue;
1018 }
797 1019
798 if (!RB_EMPTY_NODE(&upper->rb_node)) { 1020 if (!RB_EMPTY_NODE(&upper->rb_node)) {
799 if (upper->lowest) { 1021 if (upper->lowest) {
@@ -806,25 +1028,69 @@ next:
806 } 1028 }
807 1029
808 BUG_ON(!upper->checked); 1030 BUG_ON(!upper->checked);
809 rb_node = tree_insert(&cache->rb_root, upper->bytenr, 1031 BUG_ON(cowonly != upper->cowonly);
810 &upper->rb_node); 1032 if (!cowonly) {
811 BUG_ON(rb_node); 1033 rb_node = tree_insert(&cache->rb_root, upper->bytenr,
1034 &upper->rb_node);
1035 BUG_ON(rb_node);
1036 }
812 1037
813 list_add_tail(&edge->list[UPPER], &upper->lower); 1038 list_add_tail(&edge->list[UPPER], &upper->lower);
814 1039
815 list_for_each_entry(edge, &upper->upper, list[LOWER]) 1040 list_for_each_entry(edge, &upper->upper, list[LOWER])
816 list_add_tail(&edge->list[UPPER], &list); 1041 list_add_tail(&edge->list[UPPER], &list);
817 } 1042 }
1043 /*
1044 * process useless backref nodes. backref nodes for tree leaves
1045 * are deleted from the cache. backref nodes for upper level
1046 * tree blocks are left in the cache to avoid unnecessary backref
1047 * lookup.
1048 */
1049 while (!list_empty(&useless)) {
1050 upper = list_entry(useless.next, struct backref_node, list);
1051 list_del_init(&upper->list);
1052 BUG_ON(!list_empty(&upper->upper));
1053 if (upper == node)
1054 node = NULL;
1055 if (upper->lowest) {
1056 list_del_init(&upper->lower);
1057 upper->lowest = 0;
1058 }
1059 while (!list_empty(&upper->lower)) {
1060 edge = list_entry(upper->lower.next,
1061 struct backref_edge, list[UPPER]);
1062 list_del(&edge->list[UPPER]);
1063 list_del(&edge->list[LOWER]);
1064 lower = edge->node[LOWER];
1065 free_backref_edge(cache, edge);
1066
1067 if (list_empty(&lower->upper))
1068 list_add(&lower->list, &useless);
1069 }
1070 __mark_block_processed(rc, upper);
1071 if (upper->level > 0) {
1072 list_add(&upper->list, &cache->detached);
1073 upper->detached = 1;
1074 } else {
1075 rb_erase(&upper->rb_node, &cache->rb_root);
1076 free_backref_node(cache, upper);
1077 }
1078 }
818out: 1079out:
819 btrfs_free_path(path1); 1080 btrfs_free_path(path1);
820 btrfs_free_path(path2); 1081 btrfs_free_path(path2);
821 if (err) { 1082 if (err) {
822 INIT_LIST_HEAD(&list); 1083 while (!list_empty(&useless)) {
1084 lower = list_entry(useless.next,
1085 struct backref_node, upper);
1086 list_del_init(&lower->upper);
1087 }
823 upper = node; 1088 upper = node;
1089 INIT_LIST_HEAD(&list);
824 while (upper) { 1090 while (upper) {
825 if (RB_EMPTY_NODE(&upper->rb_node)) { 1091 if (RB_EMPTY_NODE(&upper->rb_node)) {
826 list_splice_tail(&upper->upper, &list); 1092 list_splice_tail(&upper->upper, &list);
827 kfree(upper); 1093 free_backref_node(cache, upper);
828 } 1094 }
829 1095
830 if (list_empty(&list)) 1096 if (list_empty(&list))
@@ -832,15 +1098,104 @@ out:
832 1098
833 edge = list_entry(list.next, struct backref_edge, 1099 edge = list_entry(list.next, struct backref_edge,
834 list[LOWER]); 1100 list[LOWER]);
1101 list_del(&edge->list[LOWER]);
835 upper = edge->node[UPPER]; 1102 upper = edge->node[UPPER];
836 kfree(edge); 1103 free_backref_edge(cache, edge);
837 } 1104 }
838 return ERR_PTR(err); 1105 return ERR_PTR(err);
839 } 1106 }
1107 BUG_ON(node && node->detached);
840 return node; 1108 return node;
841} 1109}
842 1110
843/* 1111/*
1112 * helper to add backref node for the newly created snapshot.
1113 * the backref node is created by cloning backref node that
1114 * corresponds to root of source tree
1115 */
1116static int clone_backref_node(struct btrfs_trans_handle *trans,
1117 struct reloc_control *rc,
1118 struct btrfs_root *src,
1119 struct btrfs_root *dest)
1120{
1121 struct btrfs_root *reloc_root = src->reloc_root;
1122 struct backref_cache *cache = &rc->backref_cache;
1123 struct backref_node *node = NULL;
1124 struct backref_node *new_node;
1125 struct backref_edge *edge;
1126 struct backref_edge *new_edge;
1127 struct rb_node *rb_node;
1128
1129 if (cache->last_trans > 0)
1130 update_backref_cache(trans, cache);
1131
1132 rb_node = tree_search(&cache->rb_root, src->commit_root->start);
1133 if (rb_node) {
1134 node = rb_entry(rb_node, struct backref_node, rb_node);
1135 if (node->detached)
1136 node = NULL;
1137 else
1138 BUG_ON(node->new_bytenr != reloc_root->node->start);
1139 }
1140
1141 if (!node) {
1142 rb_node = tree_search(&cache->rb_root,
1143 reloc_root->commit_root->start);
1144 if (rb_node) {
1145 node = rb_entry(rb_node, struct backref_node,
1146 rb_node);
1147 BUG_ON(node->detached);
1148 }
1149 }
1150
1151 if (!node)
1152 return 0;
1153
1154 new_node = alloc_backref_node(cache);
1155 if (!new_node)
1156 return -ENOMEM;
1157
1158 new_node->bytenr = dest->node->start;
1159 new_node->level = node->level;
1160 new_node->lowest = node->lowest;
1161 new_node->root = dest;
1162
1163 if (!node->lowest) {
1164 list_for_each_entry(edge, &node->lower, list[UPPER]) {
1165 new_edge = alloc_backref_edge(cache);
1166 if (!new_edge)
1167 goto fail;
1168
1169 new_edge->node[UPPER] = new_node;
1170 new_edge->node[LOWER] = edge->node[LOWER];
1171 list_add_tail(&new_edge->list[UPPER],
1172 &new_node->lower);
1173 }
1174 }
1175
1176 rb_node = tree_insert(&cache->rb_root, new_node->bytenr,
1177 &new_node->rb_node);
1178 BUG_ON(rb_node);
1179
1180 if (!new_node->lowest) {
1181 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
1182 list_add_tail(&new_edge->list[LOWER],
1183 &new_edge->node[LOWER]->upper);
1184 }
1185 }
1186 return 0;
1187fail:
1188 while (!list_empty(&new_node->lower)) {
1189 new_edge = list_entry(new_node->lower.next,
1190 struct backref_edge, list[UPPER]);
1191 list_del(&new_edge->list[UPPER]);
1192 free_backref_edge(cache, new_edge);
1193 }
1194 free_backref_node(cache, new_node);
1195 return -ENOMEM;
1196}
1197
1198/*
844 * helper to add 'address of tree root -> reloc tree' mapping 1199 * helper to add 'address of tree root -> reloc tree' mapping
845 */ 1200 */
846static int __add_reloc_root(struct btrfs_root *root) 1201static int __add_reloc_root(struct btrfs_root *root)
@@ -900,12 +1255,8 @@ static int __update_reloc_root(struct btrfs_root *root, int del)
900 return 0; 1255 return 0;
901} 1256}
902 1257
903/* 1258static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
904 * create reloc tree for a given fs tree. reloc tree is just a 1259 struct btrfs_root *root, u64 objectid)
905 * snapshot of the fs tree with special root objectid.
906 */
907int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
908 struct btrfs_root *root)
909{ 1260{
910 struct btrfs_root *reloc_root; 1261 struct btrfs_root *reloc_root;
911 struct extent_buffer *eb; 1262 struct extent_buffer *eb;
@@ -913,36 +1264,45 @@ int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
913 struct btrfs_key root_key; 1264 struct btrfs_key root_key;
914 int ret; 1265 int ret;
915 1266
916 if (root->reloc_root) {
917 reloc_root = root->reloc_root;
918 reloc_root->last_trans = trans->transid;
919 return 0;
920 }
921
922 if (!root->fs_info->reloc_ctl ||
923 !root->fs_info->reloc_ctl->create_reloc_root ||
924 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
925 return 0;
926
927 root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 1267 root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
928 BUG_ON(!root_item); 1268 BUG_ON(!root_item);
929 1269
930 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 1270 root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
931 root_key.type = BTRFS_ROOT_ITEM_KEY; 1271 root_key.type = BTRFS_ROOT_ITEM_KEY;
932 root_key.offset = root->root_key.objectid; 1272 root_key.offset = objectid;
933 1273
934 ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 1274 if (root->root_key.objectid == objectid) {
935 BTRFS_TREE_RELOC_OBJECTID); 1275 /* called by btrfs_init_reloc_root */
936 BUG_ON(ret); 1276 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
1277 BTRFS_TREE_RELOC_OBJECTID);
1278 BUG_ON(ret);
1279
1280 btrfs_set_root_last_snapshot(&root->root_item,
1281 trans->transid - 1);
1282 } else {
1283 /*
1284 * called by btrfs_reloc_post_snapshot_hook.
1285 * the source tree is a reloc tree, all tree blocks
1286 * modified after it was created have RELOC flag
1287 * set in their headers. so it's OK to not update
1288 * the 'last_snapshot'.
1289 */
1290 ret = btrfs_copy_root(trans, root, root->node, &eb,
1291 BTRFS_TREE_RELOC_OBJECTID);
1292 BUG_ON(ret);
1293 }
937 1294
938 btrfs_set_root_last_snapshot(&root->root_item, trans->transid - 1);
939 memcpy(root_item, &root->root_item, sizeof(*root_item)); 1295 memcpy(root_item, &root->root_item, sizeof(*root_item));
940 btrfs_set_root_refs(root_item, 1);
941 btrfs_set_root_bytenr(root_item, eb->start); 1296 btrfs_set_root_bytenr(root_item, eb->start);
942 btrfs_set_root_level(root_item, btrfs_header_level(eb)); 1297 btrfs_set_root_level(root_item, btrfs_header_level(eb));
943 btrfs_set_root_generation(root_item, trans->transid); 1298 btrfs_set_root_generation(root_item, trans->transid);
944 memset(&root_item->drop_progress, 0, sizeof(struct btrfs_disk_key)); 1299
945 root_item->drop_level = 0; 1300 if (root->root_key.objectid == objectid) {
1301 btrfs_set_root_refs(root_item, 0);
1302 memset(&root_item->drop_progress, 0,
1303 sizeof(struct btrfs_disk_key));
1304 root_item->drop_level = 0;
1305 }
946 1306
947 btrfs_tree_unlock(eb); 1307 btrfs_tree_unlock(eb);
948 free_extent_buffer(eb); 1308 free_extent_buffer(eb);
@@ -956,6 +1316,37 @@ int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
956 &root_key); 1316 &root_key);
957 BUG_ON(IS_ERR(reloc_root)); 1317 BUG_ON(IS_ERR(reloc_root));
958 reloc_root->last_trans = trans->transid; 1318 reloc_root->last_trans = trans->transid;
1319 return reloc_root;
1320}
1321
1322/*
1323 * create reloc tree for a given fs tree. reloc tree is just a
1324 * snapshot of the fs tree with special root objectid.
1325 */
1326int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
1327 struct btrfs_root *root)
1328{
1329 struct btrfs_root *reloc_root;
1330 struct reloc_control *rc = root->fs_info->reloc_ctl;
1331 int clear_rsv = 0;
1332
1333 if (root->reloc_root) {
1334 reloc_root = root->reloc_root;
1335 reloc_root->last_trans = trans->transid;
1336 return 0;
1337 }
1338
1339 if (!rc || !rc->create_reloc_tree ||
1340 root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1341 return 0;
1342
1343 if (!trans->block_rsv) {
1344 trans->block_rsv = rc->block_rsv;
1345 clear_rsv = 1;
1346 }
1347 reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
1348 if (clear_rsv)
1349 trans->block_rsv = NULL;
959 1350
960 __add_reloc_root(reloc_root); 1351 __add_reloc_root(reloc_root);
961 root->reloc_root = reloc_root; 1352 root->reloc_root = reloc_root;
@@ -979,7 +1370,8 @@ int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
979 reloc_root = root->reloc_root; 1370 reloc_root = root->reloc_root;
980 root_item = &reloc_root->root_item; 1371 root_item = &reloc_root->root_item;
981 1372
982 if (btrfs_root_refs(root_item) == 0) { 1373 if (root->fs_info->reloc_ctl->merge_reloc_tree &&
1374 btrfs_root_refs(root_item) == 0) {
983 root->reloc_root = NULL; 1375 root->reloc_root = NULL;
984 del = 1; 1376 del = 1;
985 } 1377 }
@@ -1101,8 +1493,7 @@ static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
1101 goto out; 1493 goto out;
1102 } 1494 }
1103 1495
1104 if (new_bytenr) 1496 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1105 *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1106 ret = 0; 1497 ret = 0;
1107out: 1498out:
1108 btrfs_free_path(path); 1499 btrfs_free_path(path);
@@ -1113,19 +1504,18 @@ out:
1113 * update file extent items in the tree leaf to point to 1504 * update file extent items in the tree leaf to point to
1114 * the new locations. 1505 * the new locations.
1115 */ 1506 */
1116static int replace_file_extents(struct btrfs_trans_handle *trans, 1507static noinline_for_stack
1117 struct reloc_control *rc, 1508int replace_file_extents(struct btrfs_trans_handle *trans,
1118 struct btrfs_root *root, 1509 struct reloc_control *rc,
1119 struct extent_buffer *leaf, 1510 struct btrfs_root *root,
1120 struct list_head *inode_list) 1511 struct extent_buffer *leaf)
1121{ 1512{
1122 struct btrfs_key key; 1513 struct btrfs_key key;
1123 struct btrfs_file_extent_item *fi; 1514 struct btrfs_file_extent_item *fi;
1124 struct inode *inode = NULL; 1515 struct inode *inode = NULL;
1125 struct inodevec *ivec = NULL;
1126 u64 parent; 1516 u64 parent;
1127 u64 bytenr; 1517 u64 bytenr;
1128 u64 new_bytenr; 1518 u64 new_bytenr = 0;
1129 u64 num_bytes; 1519 u64 num_bytes;
1130 u64 end; 1520 u64 end;
1131 u32 nritems; 1521 u32 nritems;
@@ -1165,21 +1555,12 @@ static int replace_file_extents(struct btrfs_trans_handle *trans,
1165 * to complete and drop the extent cache 1555 * to complete and drop the extent cache
1166 */ 1556 */
1167 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 1557 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1168 if (!ivec || ivec->nr == INODEVEC_SIZE) {
1169 ivec = kmalloc(sizeof(*ivec), GFP_NOFS);
1170 BUG_ON(!ivec);
1171 ivec->nr = 0;
1172 list_add_tail(&ivec->list, inode_list);
1173 }
1174 if (first) { 1558 if (first) {
1175 inode = find_next_inode(root, key.objectid); 1559 inode = find_next_inode(root, key.objectid);
1176 if (inode)
1177 ivec->inode[ivec->nr++] = inode;
1178 first = 0; 1560 first = 0;
1179 } else if (inode && inode->i_ino < key.objectid) { 1561 } else if (inode && inode->i_ino < key.objectid) {
1562 btrfs_add_delayed_iput(inode);
1180 inode = find_next_inode(root, key.objectid); 1563 inode = find_next_inode(root, key.objectid);
1181 if (inode)
1182 ivec->inode[ivec->nr++] = inode;
1183 } 1564 }
1184 if (inode && inode->i_ino == key.objectid) { 1565 if (inode && inode->i_ino == key.objectid) {
1185 end = key.offset + 1566 end = key.offset +
@@ -1203,8 +1584,10 @@ static int replace_file_extents(struct btrfs_trans_handle *trans,
1203 1584
1204 ret = get_new_location(rc->data_inode, &new_bytenr, 1585 ret = get_new_location(rc->data_inode, &new_bytenr,
1205 bytenr, num_bytes); 1586 bytenr, num_bytes);
1206 if (ret > 0) 1587 if (ret > 0) {
1588 WARN_ON(1);
1207 continue; 1589 continue;
1590 }
1208 BUG_ON(ret < 0); 1591 BUG_ON(ret < 0);
1209 1592
1210 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 1593 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
@@ -1224,6 +1607,8 @@ static int replace_file_extents(struct btrfs_trans_handle *trans,
1224 } 1607 }
1225 if (dirty) 1608 if (dirty)
1226 btrfs_mark_buffer_dirty(leaf); 1609 btrfs_mark_buffer_dirty(leaf);
1610 if (inode)
1611 btrfs_add_delayed_iput(inode);
1227 return 0; 1612 return 0;
1228} 1613}
1229 1614
@@ -1247,11 +1632,11 @@ int memcmp_node_keys(struct extent_buffer *eb, int slot,
1247 * if no block got replaced, 0 is returned. if there are other 1632 * if no block got replaced, 0 is returned. if there are other
1248 * errors, a negative error number is returned. 1633 * errors, a negative error number is returned.
1249 */ 1634 */
1250static int replace_path(struct btrfs_trans_handle *trans, 1635static noinline_for_stack
1251 struct btrfs_root *dest, struct btrfs_root *src, 1636int replace_path(struct btrfs_trans_handle *trans,
1252 struct btrfs_path *path, struct btrfs_key *next_key, 1637 struct btrfs_root *dest, struct btrfs_root *src,
1253 struct extent_buffer **leaf, 1638 struct btrfs_path *path, struct btrfs_key *next_key,
1254 int lowest_level, int max_level) 1639 int lowest_level, int max_level)
1255{ 1640{
1256 struct extent_buffer *eb; 1641 struct extent_buffer *eb;
1257 struct extent_buffer *parent; 1642 struct extent_buffer *parent;
@@ -1262,16 +1647,16 @@ static int replace_path(struct btrfs_trans_handle *trans,
1262 u64 new_ptr_gen; 1647 u64 new_ptr_gen;
1263 u64 last_snapshot; 1648 u64 last_snapshot;
1264 u32 blocksize; 1649 u32 blocksize;
1650 int cow = 0;
1265 int level; 1651 int level;
1266 int ret; 1652 int ret;
1267 int slot; 1653 int slot;
1268 1654
1269 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 1655 BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1270 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 1656 BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1271 BUG_ON(lowest_level > 1 && leaf);
1272 1657
1273 last_snapshot = btrfs_root_last_snapshot(&src->root_item); 1658 last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1274 1659again:
1275 slot = path->slots[lowest_level]; 1660 slot = path->slots[lowest_level];
1276 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 1661 btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1277 1662
@@ -1285,8 +1670,10 @@ static int replace_path(struct btrfs_trans_handle *trans,
1285 return 0; 1670 return 0;
1286 } 1671 }
1287 1672
1288 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 1673 if (cow) {
1289 BUG_ON(ret); 1674 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb);
1675 BUG_ON(ret);
1676 }
1290 btrfs_set_lock_blocking(eb); 1677 btrfs_set_lock_blocking(eb);
1291 1678
1292 if (next_key) { 1679 if (next_key) {
@@ -1330,7 +1717,7 @@ static int replace_path(struct btrfs_trans_handle *trans,
1330 1717
1331 if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 1718 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1332 memcmp_node_keys(parent, slot, path, level)) { 1719 memcmp_node_keys(parent, slot, path, level)) {
1333 if (level <= lowest_level && !leaf) { 1720 if (level <= lowest_level) {
1334 ret = 0; 1721 ret = 0;
1335 break; 1722 break;
1336 } 1723 }
@@ -1338,16 +1725,12 @@ static int replace_path(struct btrfs_trans_handle *trans,
1338 eb = read_tree_block(dest, old_bytenr, blocksize, 1725 eb = read_tree_block(dest, old_bytenr, blocksize,
1339 old_ptr_gen); 1726 old_ptr_gen);
1340 btrfs_tree_lock(eb); 1727 btrfs_tree_lock(eb);
1341 ret = btrfs_cow_block(trans, dest, eb, parent, 1728 if (cow) {
1342 slot, &eb); 1729 ret = btrfs_cow_block(trans, dest, eb, parent,
1343 BUG_ON(ret); 1730 slot, &eb);
1344 btrfs_set_lock_blocking(eb); 1731 BUG_ON(ret);
1345
1346 if (level <= lowest_level) {
1347 *leaf = eb;
1348 ret = 0;
1349 break;
1350 } 1732 }
1733 btrfs_set_lock_blocking(eb);
1351 1734
1352 btrfs_tree_unlock(parent); 1735 btrfs_tree_unlock(parent);
1353 free_extent_buffer(parent); 1736 free_extent_buffer(parent);
@@ -1356,6 +1739,13 @@ static int replace_path(struct btrfs_trans_handle *trans,
1356 continue; 1739 continue;
1357 } 1740 }
1358 1741
1742 if (!cow) {
1743 btrfs_tree_unlock(parent);
1744 free_extent_buffer(parent);
1745 cow = 1;
1746 goto again;
1747 }
1748
1359 btrfs_node_key_to_cpu(path->nodes[level], &key, 1749 btrfs_node_key_to_cpu(path->nodes[level], &key,
1360 path->slots[level]); 1750 path->slots[level]);
1361 btrfs_release_path(src, path); 1751 btrfs_release_path(src, path);
@@ -1561,20 +1951,6 @@ static int invalidate_extent_cache(struct btrfs_root *root,
1561 return 0; 1951 return 0;
1562} 1952}
1563 1953
1564static void put_inodes(struct list_head *list)
1565{
1566 struct inodevec *ivec;
1567 while (!list_empty(list)) {
1568 ivec = list_entry(list->next, struct inodevec, list);
1569 list_del(&ivec->list);
1570 while (ivec->nr > 0) {
1571 ivec->nr--;
1572 iput(ivec->inode[ivec->nr]);
1573 }
1574 kfree(ivec);
1575 }
1576}
1577
1578static int find_next_key(struct btrfs_path *path, int level, 1954static int find_next_key(struct btrfs_path *path, int level,
1579 struct btrfs_key *key) 1955 struct btrfs_key *key)
1580 1956
@@ -1607,13 +1983,14 @@ static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1607 struct btrfs_root *reloc_root; 1983 struct btrfs_root *reloc_root;
1608 struct btrfs_root_item *root_item; 1984 struct btrfs_root_item *root_item;
1609 struct btrfs_path *path; 1985 struct btrfs_path *path;
1610 struct extent_buffer *leaf = NULL; 1986 struct extent_buffer *leaf;
1611 unsigned long nr; 1987 unsigned long nr;
1612 int level; 1988 int level;
1613 int max_level; 1989 int max_level;
1614 int replaced = 0; 1990 int replaced = 0;
1615 int ret; 1991 int ret;
1616 int err = 0; 1992 int err = 0;
1993 u32 min_reserved;
1617 1994
1618 path = btrfs_alloc_path(); 1995 path = btrfs_alloc_path();
1619 if (!path) 1996 if (!path)
@@ -1647,34 +2024,23 @@ static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1647 btrfs_unlock_up_safe(path, 0); 2024 btrfs_unlock_up_safe(path, 0);
1648 } 2025 }
1649 2026
1650 if (level == 0 && rc->stage == UPDATE_DATA_PTRS) { 2027 min_reserved = root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1651 trans = btrfs_start_transaction(root, 1); 2028 memset(&next_key, 0, sizeof(next_key));
1652 2029
1653 leaf = path->nodes[0]; 2030 while (1) {
1654 btrfs_item_key_to_cpu(leaf, &key, 0); 2031 trans = btrfs_start_transaction(root, 0);
1655 btrfs_release_path(reloc_root, path); 2032 trans->block_rsv = rc->block_rsv;
1656 2033
1657 ret = btrfs_search_slot(trans, root, &key, path, 0, 1); 2034 ret = btrfs_block_rsv_check(trans, root, rc->block_rsv,
1658 if (ret < 0) { 2035 min_reserved, 0);
1659 err = ret; 2036 if (ret) {
1660 goto out; 2037 BUG_ON(ret != -EAGAIN);
2038 ret = btrfs_commit_transaction(trans, root);
2039 BUG_ON(ret);
2040 continue;
1661 } 2041 }
1662 2042
1663 leaf = path->nodes[0];
1664 btrfs_unlock_up_safe(path, 1);
1665 ret = replace_file_extents(trans, rc, root, leaf,
1666 &inode_list);
1667 if (ret < 0)
1668 err = ret;
1669 goto out;
1670 }
1671
1672 memset(&next_key, 0, sizeof(next_key));
1673
1674 while (1) {
1675 leaf = NULL;
1676 replaced = 0; 2043 replaced = 0;
1677 trans = btrfs_start_transaction(root, 1);
1678 max_level = level; 2044 max_level = level;
1679 2045
1680 ret = walk_down_reloc_tree(reloc_root, path, &level); 2046 ret = walk_down_reloc_tree(reloc_root, path, &level);
@@ -1688,14 +2054,9 @@ static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1688 if (!find_next_key(path, level, &key) && 2054 if (!find_next_key(path, level, &key) &&
1689 btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 2055 btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1690 ret = 0; 2056 ret = 0;
1691 } else if (level == 1 && rc->stage == UPDATE_DATA_PTRS) {
1692 ret = replace_path(trans, root, reloc_root,
1693 path, &next_key, &leaf,
1694 level, max_level);
1695 } else { 2057 } else {
1696 ret = replace_path(trans, root, reloc_root, 2058 ret = replace_path(trans, root, reloc_root, path,
1697 path, &next_key, NULL, 2059 &next_key, level, max_level);
1698 level, max_level);
1699 } 2060 }
1700 if (ret < 0) { 2061 if (ret < 0) {
1701 err = ret; 2062 err = ret;
@@ -1707,16 +2068,6 @@ static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1707 btrfs_node_key_to_cpu(path->nodes[level], &key, 2068 btrfs_node_key_to_cpu(path->nodes[level], &key,
1708 path->slots[level]); 2069 path->slots[level]);
1709 replaced = 1; 2070 replaced = 1;
1710 } else if (leaf) {
1711 /*
1712 * no block got replaced, try replacing file extents
1713 */
1714 btrfs_item_key_to_cpu(leaf, &key, 0);
1715 ret = replace_file_extents(trans, rc, root, leaf,
1716 &inode_list);
1717 btrfs_tree_unlock(leaf);
1718 free_extent_buffer(leaf);
1719 BUG_ON(ret < 0);
1720 } 2071 }
1721 2072
1722 ret = walk_up_reloc_tree(reloc_root, path, &level); 2073 ret = walk_up_reloc_tree(reloc_root, path, &level);
@@ -1733,15 +2084,10 @@ static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1733 root_item->drop_level = level; 2084 root_item->drop_level = level;
1734 2085
1735 nr = trans->blocks_used; 2086 nr = trans->blocks_used;
1736 btrfs_end_transaction(trans, root); 2087 btrfs_end_transaction_throttle(trans, root);
1737 2088
1738 btrfs_btree_balance_dirty(root, nr); 2089 btrfs_btree_balance_dirty(root, nr);
1739 2090
1740 /*
1741 * put inodes outside transaction, otherwise we may deadlock.
1742 */
1743 put_inodes(&inode_list);
1744
1745 if (replaced && rc->stage == UPDATE_DATA_PTRS) 2091 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1746 invalidate_extent_cache(root, &key, &next_key); 2092 invalidate_extent_cache(root, &key, &next_key);
1747 } 2093 }
@@ -1764,87 +2110,125 @@ out:
1764 sizeof(root_item->drop_progress)); 2110 sizeof(root_item->drop_progress));
1765 root_item->drop_level = 0; 2111 root_item->drop_level = 0;
1766 btrfs_set_root_refs(root_item, 0); 2112 btrfs_set_root_refs(root_item, 0);
2113 btrfs_update_reloc_root(trans, root);
1767 } 2114 }
1768 2115
1769 nr = trans->blocks_used; 2116 nr = trans->blocks_used;
1770 btrfs_end_transaction(trans, root); 2117 btrfs_end_transaction_throttle(trans, root);
1771 2118
1772 btrfs_btree_balance_dirty(root, nr); 2119 btrfs_btree_balance_dirty(root, nr);
1773 2120
1774 put_inodes(&inode_list);
1775
1776 if (replaced && rc->stage == UPDATE_DATA_PTRS) 2121 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1777 invalidate_extent_cache(root, &key, &next_key); 2122 invalidate_extent_cache(root, &key, &next_key);
1778 2123
1779 return err; 2124 return err;
1780} 2125}
1781 2126
1782/* 2127static noinline_for_stack
1783 * callback for the work threads. 2128int prepare_to_merge(struct reloc_control *rc, int err)
1784 * this function merges reloc tree with corresponding fs tree,
1785 * and then drops the reloc tree.
1786 */
1787static void merge_func(struct btrfs_work *work)
1788{ 2129{
1789 struct btrfs_trans_handle *trans; 2130 struct btrfs_root *root = rc->extent_root;
1790 struct btrfs_root *root;
1791 struct btrfs_root *reloc_root; 2131 struct btrfs_root *reloc_root;
1792 struct async_merge *async; 2132 struct btrfs_trans_handle *trans;
2133 LIST_HEAD(reloc_roots);
2134 u64 num_bytes = 0;
2135 int ret;
2136 int retries = 0;
2137
2138 mutex_lock(&root->fs_info->trans_mutex);
2139 rc->merging_rsv_size += root->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
2140 rc->merging_rsv_size += rc->nodes_relocated * 2;
2141 mutex_unlock(&root->fs_info->trans_mutex);
2142again:
2143 if (!err) {
2144 num_bytes = rc->merging_rsv_size;
2145 ret = btrfs_block_rsv_add(NULL, root, rc->block_rsv,
2146 num_bytes, &retries);
2147 if (ret)
2148 err = ret;
2149 }
2150
2151 trans = btrfs_join_transaction(rc->extent_root, 1);
2152
2153 if (!err) {
2154 if (num_bytes != rc->merging_rsv_size) {
2155 btrfs_end_transaction(trans, rc->extent_root);
2156 btrfs_block_rsv_release(rc->extent_root,
2157 rc->block_rsv, num_bytes);
2158 retries = 0;
2159 goto again;
2160 }
2161 }
1793 2162
1794 async = container_of(work, struct async_merge, work); 2163 rc->merge_reloc_tree = 1;
1795 reloc_root = async->root; 2164
2165 while (!list_empty(&rc->reloc_roots)) {
2166 reloc_root = list_entry(rc->reloc_roots.next,
2167 struct btrfs_root, root_list);
2168 list_del_init(&reloc_root->root_list);
1796 2169
1797 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1798 root = read_fs_root(reloc_root->fs_info, 2170 root = read_fs_root(reloc_root->fs_info,
1799 reloc_root->root_key.offset); 2171 reloc_root->root_key.offset);
1800 BUG_ON(IS_ERR(root)); 2172 BUG_ON(IS_ERR(root));
1801 BUG_ON(root->reloc_root != reloc_root); 2173 BUG_ON(root->reloc_root != reloc_root);
1802 2174
1803 merge_reloc_root(async->rc, root); 2175 /*
1804 2176 * set reference count to 1, so btrfs_recover_relocation
1805 trans = btrfs_start_transaction(root, 1); 2177 * knows it should resumes merging
2178 */
2179 if (!err)
2180 btrfs_set_root_refs(&reloc_root->root_item, 1);
1806 btrfs_update_reloc_root(trans, root); 2181 btrfs_update_reloc_root(trans, root);
1807 btrfs_end_transaction(trans, root);
1808 }
1809 2182
1810 btrfs_drop_snapshot(reloc_root, 0); 2183 list_add(&reloc_root->root_list, &reloc_roots);
2184 }
1811 2185
1812 if (atomic_dec_and_test(async->num_pending)) 2186 list_splice(&reloc_roots, &rc->reloc_roots);
1813 complete(async->done);
1814 2187
1815 kfree(async); 2188 if (!err)
2189 btrfs_commit_transaction(trans, rc->extent_root);
2190 else
2191 btrfs_end_transaction(trans, rc->extent_root);
2192 return err;
1816} 2193}
1817 2194
1818static int merge_reloc_roots(struct reloc_control *rc) 2195static noinline_for_stack
2196int merge_reloc_roots(struct reloc_control *rc)
1819{ 2197{
1820 struct async_merge *async;
1821 struct btrfs_root *root; 2198 struct btrfs_root *root;
1822 struct completion done; 2199 struct btrfs_root *reloc_root;
1823 atomic_t num_pending; 2200 LIST_HEAD(reloc_roots);
2201 int found = 0;
2202 int ret;
2203again:
2204 root = rc->extent_root;
2205 mutex_lock(&root->fs_info->trans_mutex);
2206 list_splice_init(&rc->reloc_roots, &reloc_roots);
2207 mutex_unlock(&root->fs_info->trans_mutex);
1824 2208
1825 init_completion(&done); 2209 while (!list_empty(&reloc_roots)) {
1826 atomic_set(&num_pending, 1); 2210 found = 1;
2211 reloc_root = list_entry(reloc_roots.next,
2212 struct btrfs_root, root_list);
1827 2213
1828 while (!list_empty(&rc->reloc_roots)) { 2214 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1829 root = list_entry(rc->reloc_roots.next, 2215 root = read_fs_root(reloc_root->fs_info,
1830 struct btrfs_root, root_list); 2216 reloc_root->root_key.offset);
1831 list_del_init(&root->root_list); 2217 BUG_ON(IS_ERR(root));
2218 BUG_ON(root->reloc_root != reloc_root);
1832 2219
1833 async = kmalloc(sizeof(*async), GFP_NOFS); 2220 ret = merge_reloc_root(rc, root);
1834 BUG_ON(!async); 2221 BUG_ON(ret);
1835 async->work.func = merge_func; 2222 } else {
1836 async->work.flags = 0; 2223 list_del_init(&reloc_root->root_list);
1837 async->rc = rc; 2224 }
1838 async->root = root; 2225 btrfs_drop_snapshot(reloc_root, rc->block_rsv, 0);
1839 async->done = &done;
1840 async->num_pending = &num_pending;
1841 atomic_inc(&num_pending);
1842 btrfs_queue_worker(&rc->workers, &async->work);
1843 } 2226 }
1844 2227
1845 if (!atomic_dec_and_test(&num_pending)) 2228 if (found) {
1846 wait_for_completion(&done); 2229 found = 0;
1847 2230 goto again;
2231 }
1848 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 2232 BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
1849 return 0; 2233 return 0;
1850} 2234}
@@ -1875,119 +2259,169 @@ static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
1875 return btrfs_record_root_in_trans(trans, root); 2259 return btrfs_record_root_in_trans(trans, root);
1876} 2260}
1877 2261
1878/* 2262static noinline_for_stack
1879 * select one tree from trees that references the block. 2263struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
1880 * for blocks in refernce counted trees, we preper reloc tree. 2264 struct reloc_control *rc,
1881 * if no reloc tree found and reloc_only is true, NULL is returned. 2265 struct backref_node *node,
1882 */ 2266 struct backref_edge *edges[], int *nr)
1883static struct btrfs_root *__select_one_root(struct btrfs_trans_handle *trans,
1884 struct backref_node *node,
1885 struct backref_edge *edges[],
1886 int *nr, int reloc_only)
1887{ 2267{
1888 struct backref_node *next; 2268 struct backref_node *next;
1889 struct btrfs_root *root; 2269 struct btrfs_root *root;
1890 int index; 2270 int index = 0;
1891 int loop = 0; 2271
1892again:
1893 index = 0;
1894 next = node; 2272 next = node;
1895 while (1) { 2273 while (1) {
1896 cond_resched(); 2274 cond_resched();
1897 next = walk_up_backref(next, edges, &index); 2275 next = walk_up_backref(next, edges, &index);
1898 root = next->root; 2276 root = next->root;
1899 if (!root) { 2277 BUG_ON(!root);
1900 BUG_ON(!node->old_root); 2278 BUG_ON(!root->ref_cows);
1901 goto skip;
1902 }
1903
1904 /* no other choice for non-refernce counted tree */
1905 if (!root->ref_cows) {
1906 BUG_ON(reloc_only);
1907 break;
1908 }
1909 2279
1910 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 2280 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
1911 record_reloc_root_in_trans(trans, root); 2281 record_reloc_root_in_trans(trans, root);
1912 break; 2282 break;
1913 } 2283 }
1914 2284
1915 if (loop) { 2285 btrfs_record_root_in_trans(trans, root);
1916 btrfs_record_root_in_trans(trans, root); 2286 root = root->reloc_root;
2287
2288 if (next->new_bytenr != root->node->start) {
2289 BUG_ON(next->new_bytenr);
2290 BUG_ON(!list_empty(&next->list));
2291 next->new_bytenr = root->node->start;
2292 next->root = root;
2293 list_add_tail(&next->list,
2294 &rc->backref_cache.changed);
2295 __mark_block_processed(rc, next);
1917 break; 2296 break;
1918 } 2297 }
1919 2298
1920 if (reloc_only || next != node) { 2299 WARN_ON(1);
1921 if (!root->reloc_root)
1922 btrfs_record_root_in_trans(trans, root);
1923 root = root->reloc_root;
1924 /*
1925 * if the reloc tree was created in current
1926 * transation, there is no node in backref tree
1927 * corresponds to the root of the reloc tree.
1928 */
1929 if (btrfs_root_last_snapshot(&root->root_item) ==
1930 trans->transid - 1)
1931 break;
1932 }
1933skip:
1934 root = NULL; 2300 root = NULL;
1935 next = walk_down_backref(edges, &index); 2301 next = walk_down_backref(edges, &index);
1936 if (!next || next->level <= node->level) 2302 if (!next || next->level <= node->level)
1937 break; 2303 break;
1938 } 2304 }
2305 if (!root)
2306 return NULL;
1939 2307
1940 if (!root && !loop && !reloc_only) { 2308 *nr = index;
1941 loop = 1; 2309 next = node;
1942 goto again; 2310 /* setup backref node path for btrfs_reloc_cow_block */
2311 while (1) {
2312 rc->backref_cache.path[next->level] = next;
2313 if (--index < 0)
2314 break;
2315 next = edges[index]->node[UPPER];
1943 } 2316 }
1944
1945 if (root)
1946 *nr = index;
1947 else
1948 *nr = 0;
1949
1950 return root; 2317 return root;
1951} 2318}
1952 2319
2320/*
2321 * select a tree root for relocation. return NULL if the block
2322 * is reference counted. we should use do_relocation() in this
2323 * case. return a tree root pointer if the block isn't reference
2324 * counted. return -ENOENT if the block is root of reloc tree.
2325 */
1953static noinline_for_stack 2326static noinline_for_stack
1954struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans, 2327struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans,
1955 struct backref_node *node) 2328 struct backref_node *node)
1956{ 2329{
2330 struct backref_node *next;
2331 struct btrfs_root *root;
2332 struct btrfs_root *fs_root = NULL;
1957 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 2333 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
1958 int nr; 2334 int index = 0;
1959 return __select_one_root(trans, node, edges, &nr, 0); 2335
2336 next = node;
2337 while (1) {
2338 cond_resched();
2339 next = walk_up_backref(next, edges, &index);
2340 root = next->root;
2341 BUG_ON(!root);
2342
2343 /* no other choice for non-refernce counted tree */
2344 if (!root->ref_cows)
2345 return root;
2346
2347 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2348 fs_root = root;
2349
2350 if (next != node)
2351 return NULL;
2352
2353 next = walk_down_backref(edges, &index);
2354 if (!next || next->level <= node->level)
2355 break;
2356 }
2357
2358 if (!fs_root)
2359 return ERR_PTR(-ENOENT);
2360 return fs_root;
1960} 2361}
1961 2362
1962static noinline_for_stack 2363static noinline_for_stack
1963struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 2364u64 calcu_metadata_size(struct reloc_control *rc,
1964 struct backref_node *node, 2365 struct backref_node *node, int reserve)
1965 struct backref_edge *edges[], int *nr)
1966{ 2366{
1967 return __select_one_root(trans, node, edges, nr, 1); 2367 struct backref_node *next = node;
2368 struct backref_edge *edge;
2369 struct backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2370 u64 num_bytes = 0;
2371 int index = 0;
2372
2373 BUG_ON(reserve && node->processed);
2374
2375 while (next) {
2376 cond_resched();
2377 while (1) {
2378 if (next->processed && (reserve || next != node))
2379 break;
2380
2381 num_bytes += btrfs_level_size(rc->extent_root,
2382 next->level);
2383
2384 if (list_empty(&next->upper))
2385 break;
2386
2387 edge = list_entry(next->upper.next,
2388 struct backref_edge, list[LOWER]);
2389 edges[index++] = edge;
2390 next = edge->node[UPPER];
2391 }
2392 next = walk_down_backref(edges, &index);
2393 }
2394 return num_bytes;
1968} 2395}
1969 2396
1970static void grab_path_buffers(struct btrfs_path *path, 2397static int reserve_metadata_space(struct btrfs_trans_handle *trans,
1971 struct backref_node *node, 2398 struct reloc_control *rc,
1972 struct backref_edge *edges[], int nr) 2399 struct backref_node *node)
1973{ 2400{
1974 int i = 0; 2401 struct btrfs_root *root = rc->extent_root;
1975 while (1) { 2402 u64 num_bytes;
1976 drop_node_buffer(node); 2403 int ret;
1977 node->eb = path->nodes[node->level]; 2404
1978 BUG_ON(!node->eb); 2405 num_bytes = calcu_metadata_size(rc, node, 1) * 2;
1979 if (path->locks[node->level])
1980 node->locked = 1;
1981 path->nodes[node->level] = NULL;
1982 path->locks[node->level] = 0;
1983
1984 if (i >= nr)
1985 break;
1986 2406
1987 edges[i]->blockptr = node->eb->start; 2407 trans->block_rsv = rc->block_rsv;
1988 node = edges[i]->node[UPPER]; 2408 ret = btrfs_block_rsv_add(trans, root, rc->block_rsv, num_bytes,
1989 i++; 2409 &rc->block_rsv_retries);
2410 if (ret) {
2411 if (ret == -EAGAIN)
2412 rc->commit_transaction = 1;
2413 return ret;
1990 } 2414 }
2415
2416 rc->block_rsv_retries = 0;
2417 return 0;
2418}
2419
2420static void release_metadata_space(struct reloc_control *rc,
2421 struct backref_node *node)
2422{
2423 u64 num_bytes = calcu_metadata_size(rc, node, 0) * 2;
2424 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, num_bytes);
1991} 2425}
1992 2426
1993/* 2427/*
@@ -1998,6 +2432,7 @@ static void grab_path_buffers(struct btrfs_path *path,
1998 * in that case this function just updates pointers. 2432 * in that case this function just updates pointers.
1999 */ 2433 */
2000static int do_relocation(struct btrfs_trans_handle *trans, 2434static int do_relocation(struct btrfs_trans_handle *trans,
2435 struct reloc_control *rc,
2001 struct backref_node *node, 2436 struct backref_node *node,
2002 struct btrfs_key *key, 2437 struct btrfs_key *key,
2003 struct btrfs_path *path, int lowest) 2438 struct btrfs_path *path, int lowest)
@@ -2018,18 +2453,25 @@ static int do_relocation(struct btrfs_trans_handle *trans,
2018 BUG_ON(lowest && node->eb); 2453 BUG_ON(lowest && node->eb);
2019 2454
2020 path->lowest_level = node->level + 1; 2455 path->lowest_level = node->level + 1;
2456 rc->backref_cache.path[node->level] = node;
2021 list_for_each_entry(edge, &node->upper, list[LOWER]) { 2457 list_for_each_entry(edge, &node->upper, list[LOWER]) {
2022 cond_resched(); 2458 cond_resched();
2023 if (node->eb && node->eb->start == edge->blockptr)
2024 continue;
2025 2459
2026 upper = edge->node[UPPER]; 2460 upper = edge->node[UPPER];
2027 root = select_reloc_root(trans, upper, edges, &nr); 2461 root = select_reloc_root(trans, rc, upper, edges, &nr);
2028 if (!root) 2462 BUG_ON(!root);
2029 continue; 2463
2030 2464 if (upper->eb && !upper->locked) {
2031 if (upper->eb && !upper->locked) 2465 if (!lowest) {
2466 ret = btrfs_bin_search(upper->eb, key,
2467 upper->level, &slot);
2468 BUG_ON(ret);
2469 bytenr = btrfs_node_blockptr(upper->eb, slot);
2470 if (node->eb->start == bytenr)
2471 goto next;
2472 }
2032 drop_node_buffer(upper); 2473 drop_node_buffer(upper);
2474 }
2033 2475
2034 if (!upper->eb) { 2476 if (!upper->eb) {
2035 ret = btrfs_search_slot(trans, root, key, path, 0, 1); 2477 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
@@ -2039,11 +2481,17 @@ static int do_relocation(struct btrfs_trans_handle *trans,
2039 } 2481 }
2040 BUG_ON(ret > 0); 2482 BUG_ON(ret > 0);
2041 2483
2042 slot = path->slots[upper->level]; 2484 if (!upper->eb) {
2485 upper->eb = path->nodes[upper->level];
2486 path->nodes[upper->level] = NULL;
2487 } else {
2488 BUG_ON(upper->eb != path->nodes[upper->level]);
2489 }
2043 2490
2044 btrfs_unlock_up_safe(path, upper->level + 1); 2491 upper->locked = 1;
2045 grab_path_buffers(path, upper, edges, nr); 2492 path->locks[upper->level] = 0;
2046 2493
2494 slot = path->slots[upper->level];
2047 btrfs_release_path(NULL, path); 2495 btrfs_release_path(NULL, path);
2048 } else { 2496 } else {
2049 ret = btrfs_bin_search(upper->eb, key, upper->level, 2497 ret = btrfs_bin_search(upper->eb, key, upper->level,
@@ -2052,14 +2500,11 @@ static int do_relocation(struct btrfs_trans_handle *trans,
2052 } 2500 }
2053 2501
2054 bytenr = btrfs_node_blockptr(upper->eb, slot); 2502 bytenr = btrfs_node_blockptr(upper->eb, slot);
2055 if (!lowest) { 2503 if (lowest) {
2056 if (node->eb->start == bytenr) { 2504 BUG_ON(bytenr != node->bytenr);
2057 btrfs_tree_unlock(upper->eb);
2058 upper->locked = 0;
2059 continue;
2060 }
2061 } else { 2505 } else {
2062 BUG_ON(node->bytenr != bytenr); 2506 if (node->eb->start == bytenr)
2507 goto next;
2063 } 2508 }
2064 2509
2065 blocksize = btrfs_level_size(root, node->level); 2510 blocksize = btrfs_level_size(root, node->level);
@@ -2071,13 +2516,13 @@ static int do_relocation(struct btrfs_trans_handle *trans,
2071 if (!node->eb) { 2516 if (!node->eb) {
2072 ret = btrfs_cow_block(trans, root, eb, upper->eb, 2517 ret = btrfs_cow_block(trans, root, eb, upper->eb,
2073 slot, &eb); 2518 slot, &eb);
2519 btrfs_tree_unlock(eb);
2520 free_extent_buffer(eb);
2074 if (ret < 0) { 2521 if (ret < 0) {
2075 err = ret; 2522 err = ret;
2076 break; 2523 goto next;
2077 } 2524 }
2078 btrfs_set_lock_blocking(eb); 2525 BUG_ON(node->eb != eb);
2079 node->eb = eb;
2080 node->locked = 1;
2081 } else { 2526 } else {
2082 btrfs_set_node_blockptr(upper->eb, slot, 2527 btrfs_set_node_blockptr(upper->eb, slot,
2083 node->eb->start); 2528 node->eb->start);
@@ -2095,67 +2540,80 @@ static int do_relocation(struct btrfs_trans_handle *trans,
2095 ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 2540 ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2096 BUG_ON(ret); 2541 BUG_ON(ret);
2097 } 2542 }
2098 if (!lowest) { 2543next:
2099 btrfs_tree_unlock(upper->eb); 2544 if (!upper->pending)
2100 upper->locked = 0; 2545 drop_node_buffer(upper);
2101 } 2546 else
2547 unlock_node_buffer(upper);
2548 if (err)
2549 break;
2102 } 2550 }
2551
2552 if (!err && node->pending) {
2553 drop_node_buffer(node);
2554 list_move_tail(&node->list, &rc->backref_cache.changed);
2555 node->pending = 0;
2556 }
2557
2103 path->lowest_level = 0; 2558 path->lowest_level = 0;
2559 BUG_ON(err == -ENOSPC);
2104 return err; 2560 return err;
2105} 2561}
2106 2562
2107static int link_to_upper(struct btrfs_trans_handle *trans, 2563static int link_to_upper(struct btrfs_trans_handle *trans,
2564 struct reloc_control *rc,
2108 struct backref_node *node, 2565 struct backref_node *node,
2109 struct btrfs_path *path) 2566 struct btrfs_path *path)
2110{ 2567{
2111 struct btrfs_key key; 2568 struct btrfs_key key;
2112 if (!node->eb || list_empty(&node->upper))
2113 return 0;
2114 2569
2115 btrfs_node_key_to_cpu(node->eb, &key, 0); 2570 btrfs_node_key_to_cpu(node->eb, &key, 0);
2116 return do_relocation(trans, node, &key, path, 0); 2571 return do_relocation(trans, rc, node, &key, path, 0);
2117} 2572}
2118 2573
2119static int finish_pending_nodes(struct btrfs_trans_handle *trans, 2574static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2120 struct backref_cache *cache, 2575 struct reloc_control *rc,
2121 struct btrfs_path *path) 2576 struct btrfs_path *path, int err)
2122{ 2577{
2578 LIST_HEAD(list);
2579 struct backref_cache *cache = &rc->backref_cache;
2123 struct backref_node *node; 2580 struct backref_node *node;
2124 int level; 2581 int level;
2125 int ret; 2582 int ret;
2126 int err = 0;
2127 2583
2128 for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 2584 for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2129 while (!list_empty(&cache->pending[level])) { 2585 while (!list_empty(&cache->pending[level])) {
2130 node = list_entry(cache->pending[level].next, 2586 node = list_entry(cache->pending[level].next,
2131 struct backref_node, lower); 2587 struct backref_node, list);
2132 BUG_ON(node->level != level); 2588 list_move_tail(&node->list, &list);
2589 BUG_ON(!node->pending);
2133 2590
2134 ret = link_to_upper(trans, node, path); 2591 if (!err) {
2135 if (ret < 0) 2592 ret = link_to_upper(trans, rc, node, path);
2136 err = ret; 2593 if (ret < 0)
2137 /* 2594 err = ret;
2138 * this remove the node from the pending list and 2595 }
2139 * may add some other nodes to the level + 1
2140 * pending list
2141 */
2142 remove_backref_node(cache, node);
2143 } 2596 }
2597 list_splice_init(&list, &cache->pending[level]);
2144 } 2598 }
2145 BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root));
2146 return err; 2599 return err;
2147} 2600}
2148 2601
2149static void mark_block_processed(struct reloc_control *rc, 2602static void mark_block_processed(struct reloc_control *rc,
2150 struct backref_node *node) 2603 u64 bytenr, u32 blocksize)
2604{
2605 set_extent_bits(&rc->processed_blocks, bytenr, bytenr + blocksize - 1,
2606 EXTENT_DIRTY, GFP_NOFS);
2607}
2608
2609static void __mark_block_processed(struct reloc_control *rc,
2610 struct backref_node *node)
2151{ 2611{
2152 u32 blocksize; 2612 u32 blocksize;
2153 if (node->level == 0 || 2613 if (node->level == 0 ||
2154 in_block_group(node->bytenr, rc->block_group)) { 2614 in_block_group(node->bytenr, rc->block_group)) {
2155 blocksize = btrfs_level_size(rc->extent_root, node->level); 2615 blocksize = btrfs_level_size(rc->extent_root, node->level);
2156 set_extent_bits(&rc->processed_blocks, node->bytenr, 2616 mark_block_processed(rc, node->bytenr, blocksize);
2157 node->bytenr + blocksize - 1, EXTENT_DIRTY,
2158 GFP_NOFS);
2159 } 2617 }
2160 node->processed = 1; 2618 node->processed = 1;
2161} 2619}
@@ -2178,7 +2636,7 @@ static void update_processed_blocks(struct reloc_control *rc,
2178 if (next->processed) 2636 if (next->processed)
2179 break; 2637 break;
2180 2638
2181 mark_block_processed(rc, next); 2639 __mark_block_processed(rc, next);
2182 2640
2183 if (list_empty(&next->upper)) 2641 if (list_empty(&next->upper))
2184 break; 2642 break;
@@ -2201,138 +2659,6 @@ static int tree_block_processed(u64 bytenr, u32 blocksize,
2201 return 0; 2659 return 0;
2202} 2660}
2203 2661
2204/*
2205 * check if there are any file extent pointers in the leaf point to
2206 * data require processing
2207 */
2208static int check_file_extents(struct reloc_control *rc,
2209 u64 bytenr, u32 blocksize, u64 ptr_gen)
2210{
2211 struct btrfs_key found_key;
2212 struct btrfs_file_extent_item *fi;
2213 struct extent_buffer *leaf;
2214 u32 nritems;
2215 int i;
2216 int ret = 0;
2217
2218 leaf = read_tree_block(rc->extent_root, bytenr, blocksize, ptr_gen);
2219
2220 nritems = btrfs_header_nritems(leaf);
2221 for (i = 0; i < nritems; i++) {
2222 cond_resched();
2223 btrfs_item_key_to_cpu(leaf, &found_key, i);
2224 if (found_key.type != BTRFS_EXTENT_DATA_KEY)
2225 continue;
2226 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
2227 if (btrfs_file_extent_type(leaf, fi) ==
2228 BTRFS_FILE_EXTENT_INLINE)
2229 continue;
2230 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
2231 if (bytenr == 0)
2232 continue;
2233 if (in_block_group(bytenr, rc->block_group)) {
2234 ret = 1;
2235 break;
2236 }
2237 }
2238 free_extent_buffer(leaf);
2239 return ret;
2240}
2241
2242/*
2243 * scan child blocks of a given block to find blocks require processing
2244 */
2245static int add_child_blocks(struct btrfs_trans_handle *trans,
2246 struct reloc_control *rc,
2247 struct backref_node *node,
2248 struct rb_root *blocks)
2249{
2250 struct tree_block *block;
2251 struct rb_node *rb_node;
2252 u64 bytenr;
2253 u64 ptr_gen;
2254 u32 blocksize;
2255 u32 nritems;
2256 int i;
2257 int err = 0;
2258
2259 nritems = btrfs_header_nritems(node->eb);
2260 blocksize = btrfs_level_size(rc->extent_root, node->level - 1);
2261 for (i = 0; i < nritems; i++) {
2262 cond_resched();
2263 bytenr = btrfs_node_blockptr(node->eb, i);
2264 ptr_gen = btrfs_node_ptr_generation(node->eb, i);
2265 if (ptr_gen == trans->transid)
2266 continue;
2267 if (!in_block_group(bytenr, rc->block_group) &&
2268 (node->level > 1 || rc->stage == MOVE_DATA_EXTENTS))
2269 continue;
2270 if (tree_block_processed(bytenr, blocksize, rc))
2271 continue;
2272
2273 readahead_tree_block(rc->extent_root,
2274 bytenr, blocksize, ptr_gen);
2275 }
2276
2277 for (i = 0; i < nritems; i++) {
2278 cond_resched();
2279 bytenr = btrfs_node_blockptr(node->eb, i);
2280 ptr_gen = btrfs_node_ptr_generation(node->eb, i);
2281 if (ptr_gen == trans->transid)
2282 continue;
2283 if (!in_block_group(bytenr, rc->block_group) &&
2284 (node->level > 1 || rc->stage == MOVE_DATA_EXTENTS))
2285 continue;
2286 if (tree_block_processed(bytenr, blocksize, rc))
2287 continue;
2288 if (!in_block_group(bytenr, rc->block_group) &&
2289 !check_file_extents(rc, bytenr, blocksize, ptr_gen))
2290 continue;
2291
2292 block = kmalloc(sizeof(*block), GFP_NOFS);
2293 if (!block) {
2294 err = -ENOMEM;
2295 break;
2296 }
2297 block->bytenr = bytenr;
2298 btrfs_node_key_to_cpu(node->eb, &block->key, i);
2299 block->level = node->level - 1;
2300 block->key_ready = 1;
2301 rb_node = tree_insert(blocks, block->bytenr, &block->rb_node);
2302 BUG_ON(rb_node);
2303 }
2304 if (err)
2305 free_block_list(blocks);
2306 return err;
2307}
2308
2309/*
2310 * find adjacent blocks require processing
2311 */
2312static noinline_for_stack
2313int add_adjacent_blocks(struct btrfs_trans_handle *trans,
2314 struct reloc_control *rc,
2315 struct backref_cache *cache,
2316 struct rb_root *blocks, int level,
2317 struct backref_node **upper)
2318{
2319 struct backref_node *node;
2320 int ret = 0;
2321
2322 WARN_ON(!list_empty(&cache->pending[level]));
2323
2324 if (list_empty(&cache->pending[level + 1]))
2325 return 1;
2326
2327 node = list_entry(cache->pending[level + 1].next,
2328 struct backref_node, lower);
2329 if (node->eb)
2330 ret = add_child_blocks(trans, rc, node, blocks);
2331
2332 *upper = node;
2333 return ret;
2334}
2335
2336static int get_tree_block_key(struct reloc_control *rc, 2662static int get_tree_block_key(struct reloc_control *rc,
2337 struct tree_block *block) 2663 struct tree_block *block)
2338{ 2664{
@@ -2370,40 +2696,53 @@ static int relocate_tree_block(struct btrfs_trans_handle *trans,
2370 struct btrfs_path *path) 2696 struct btrfs_path *path)
2371{ 2697{
2372 struct btrfs_root *root; 2698 struct btrfs_root *root;
2373 int ret; 2699 int release = 0;
2700 int ret = 0;
2374 2701
2702 if (!node)
2703 return 0;
2704
2705 BUG_ON(node->processed);
2375 root = select_one_root(trans, node); 2706 root = select_one_root(trans, node);
2376 if (unlikely(!root)) { 2707 if (root == ERR_PTR(-ENOENT)) {
2377 rc->found_old_snapshot = 1;
2378 update_processed_blocks(rc, node); 2708 update_processed_blocks(rc, node);
2379 return 0; 2709 goto out;
2380 } 2710 }
2381 2711
2382 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 2712 if (!root || root->ref_cows) {
2383 ret = do_relocation(trans, node, key, path, 1); 2713 ret = reserve_metadata_space(trans, rc, node);
2384 if (ret < 0) 2714 if (ret)
2385 goto out;
2386 if (node->level == 0 && rc->stage == UPDATE_DATA_PTRS) {
2387 ret = replace_file_extents(trans, rc, root,
2388 node->eb, NULL);
2389 if (ret < 0)
2390 goto out;
2391 }
2392 drop_node_buffer(node);
2393 } else if (!root->ref_cows) {
2394 path->lowest_level = node->level;
2395 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2396 btrfs_release_path(root, path);
2397 if (ret < 0)
2398 goto out; 2715 goto out;
2399 } else if (root != node->root) { 2716 release = 1;
2400 WARN_ON(node->level > 0 || rc->stage != UPDATE_DATA_PTRS);
2401 } 2717 }
2402 2718
2403 update_processed_blocks(rc, node); 2719 if (root) {
2404 ret = 0; 2720 if (root->ref_cows) {
2721 BUG_ON(node->new_bytenr);
2722 BUG_ON(!list_empty(&node->list));
2723 btrfs_record_root_in_trans(trans, root);
2724 root = root->reloc_root;
2725 node->new_bytenr = root->node->start;
2726 node->root = root;
2727 list_add_tail(&node->list, &rc->backref_cache.changed);
2728 } else {
2729 path->lowest_level = node->level;
2730 ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2731 btrfs_release_path(root, path);
2732 if (ret > 0)
2733 ret = 0;
2734 }
2735 if (!ret)
2736 update_processed_blocks(rc, node);
2737 } else {
2738 ret = do_relocation(trans, rc, node, key, path, 1);
2739 }
2405out: 2740out:
2406 drop_node_buffer(node); 2741 if (ret || node->level == 0 || node->cowonly) {
2742 if (release)
2743 release_metadata_space(rc, node);
2744 remove_backref_node(&rc->backref_cache, node);
2745 }
2407 return ret; 2746 return ret;
2408} 2747}
2409 2748
@@ -2414,12 +2753,10 @@ static noinline_for_stack
2414int relocate_tree_blocks(struct btrfs_trans_handle *trans, 2753int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2415 struct reloc_control *rc, struct rb_root *blocks) 2754 struct reloc_control *rc, struct rb_root *blocks)
2416{ 2755{
2417 struct backref_cache *cache;
2418 struct backref_node *node; 2756 struct backref_node *node;
2419 struct btrfs_path *path; 2757 struct btrfs_path *path;
2420 struct tree_block *block; 2758 struct tree_block *block;
2421 struct rb_node *rb_node; 2759 struct rb_node *rb_node;
2422 int level = -1;
2423 int ret; 2760 int ret;
2424 int err = 0; 2761 int err = 0;
2425 2762
@@ -2427,21 +2764,9 @@ int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2427 if (!path) 2764 if (!path)
2428 return -ENOMEM; 2765 return -ENOMEM;
2429 2766
2430 cache = kmalloc(sizeof(*cache), GFP_NOFS);
2431 if (!cache) {
2432 btrfs_free_path(path);
2433 return -ENOMEM;
2434 }
2435
2436 backref_cache_init(cache);
2437
2438 rb_node = rb_first(blocks); 2767 rb_node = rb_first(blocks);
2439 while (rb_node) { 2768 while (rb_node) {
2440 block = rb_entry(rb_node, struct tree_block, rb_node); 2769 block = rb_entry(rb_node, struct tree_block, rb_node);
2441 if (level == -1)
2442 level = block->level;
2443 else
2444 BUG_ON(level != block->level);
2445 if (!block->key_ready) 2770 if (!block->key_ready)
2446 reada_tree_block(rc, block); 2771 reada_tree_block(rc, block);
2447 rb_node = rb_next(rb_node); 2772 rb_node = rb_next(rb_node);
@@ -2459,7 +2784,7 @@ int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2459 while (rb_node) { 2784 while (rb_node) {
2460 block = rb_entry(rb_node, struct tree_block, rb_node); 2785 block = rb_entry(rb_node, struct tree_block, rb_node);
2461 2786
2462 node = build_backref_tree(rc, cache, &block->key, 2787 node = build_backref_tree(rc, &block->key,
2463 block->level, block->bytenr); 2788 block->level, block->bytenr);
2464 if (IS_ERR(node)) { 2789 if (IS_ERR(node)) {
2465 err = PTR_ERR(node); 2790 err = PTR_ERR(node);
@@ -2469,79 +2794,62 @@ int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2469 ret = relocate_tree_block(trans, rc, node, &block->key, 2794 ret = relocate_tree_block(trans, rc, node, &block->key,
2470 path); 2795 path);
2471 if (ret < 0) { 2796 if (ret < 0) {
2472 err = ret; 2797 if (ret != -EAGAIN || rb_node == rb_first(blocks))
2798 err = ret;
2473 goto out; 2799 goto out;
2474 } 2800 }
2475 remove_backref_node(cache, node);
2476 rb_node = rb_next(rb_node); 2801 rb_node = rb_next(rb_node);
2477 } 2802 }
2478 2803out:
2479 if (level > 0)
2480 goto out;
2481
2482 free_block_list(blocks); 2804 free_block_list(blocks);
2805 err = finish_pending_nodes(trans, rc, path, err);
2483 2806
2484 /* 2807 btrfs_free_path(path);
2485 * now backrefs of some upper level tree blocks have been cached, 2808 return err;
2486 * try relocating blocks referenced by these upper level blocks. 2809}
2487 */
2488 while (1) {
2489 struct backref_node *upper = NULL;
2490 if (trans->transaction->in_commit ||
2491 trans->transaction->delayed_refs.flushing)
2492 break;
2493 2810
2494 ret = add_adjacent_blocks(trans, rc, cache, blocks, level, 2811static noinline_for_stack
2495 &upper); 2812int prealloc_file_extent_cluster(struct inode *inode,
2496 if (ret < 0) 2813 struct file_extent_cluster *cluster)
2497 err = ret; 2814{
2498 if (ret != 0) 2815 u64 alloc_hint = 0;
2499 break; 2816 u64 start;
2817 u64 end;
2818 u64 offset = BTRFS_I(inode)->index_cnt;
2819 u64 num_bytes;
2820 int nr = 0;
2821 int ret = 0;
2500 2822
2501 rb_node = rb_first(blocks); 2823 BUG_ON(cluster->start != cluster->boundary[0]);
2502 while (rb_node) { 2824 mutex_lock(&inode->i_mutex);
2503 block = rb_entry(rb_node, struct tree_block, rb_node);
2504 if (trans->transaction->in_commit ||
2505 trans->transaction->delayed_refs.flushing)
2506 goto out;
2507 BUG_ON(!block->key_ready);
2508 node = build_backref_tree(rc, cache, &block->key,
2509 level, block->bytenr);
2510 if (IS_ERR(node)) {
2511 err = PTR_ERR(node);
2512 goto out;
2513 }
2514 2825
2515 ret = relocate_tree_block(trans, rc, node, 2826 ret = btrfs_check_data_free_space(inode, cluster->end +
2516 &block->key, path); 2827 1 - cluster->start);
2517 if (ret < 0) { 2828 if (ret)
2518 err = ret; 2829 goto out;
2519 goto out;
2520 }
2521 remove_backref_node(cache, node);
2522 rb_node = rb_next(rb_node);
2523 }
2524 free_block_list(blocks);
2525 2830
2526 if (upper) { 2831 while (nr < cluster->nr) {
2527 ret = link_to_upper(trans, upper, path); 2832 start = cluster->boundary[nr] - offset;
2528 if (ret < 0) { 2833 if (nr + 1 < cluster->nr)
2529 err = ret; 2834 end = cluster->boundary[nr + 1] - 1 - offset;
2530 break; 2835 else
2531 } 2836 end = cluster->end - offset;
2532 remove_backref_node(cache, upper); 2837
2533 } 2838 lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2839 num_bytes = end + 1 - start;
2840 ret = btrfs_prealloc_file_range(inode, 0, start,
2841 num_bytes, num_bytes,
2842 end + 1, &alloc_hint);
2843 unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS);
2844 if (ret)
2845 break;
2846 nr++;
2534 } 2847 }
2848 btrfs_free_reserved_data_space(inode, cluster->end +
2849 1 - cluster->start);
2535out: 2850out:
2536 free_block_list(blocks); 2851 mutex_unlock(&inode->i_mutex);
2537 2852 return ret;
2538 ret = finish_pending_nodes(trans, cache, path);
2539 if (ret < 0)
2540 err = ret;
2541
2542 kfree(cache);
2543 btrfs_free_path(path);
2544 return err;
2545} 2853}
2546 2854
2547static noinline_for_stack 2855static noinline_for_stack
@@ -2587,7 +2895,6 @@ static int relocate_file_extent_cluster(struct inode *inode,
2587 u64 offset = BTRFS_I(inode)->index_cnt; 2895 u64 offset = BTRFS_I(inode)->index_cnt;
2588 unsigned long index; 2896 unsigned long index;
2589 unsigned long last_index; 2897 unsigned long last_index;
2590 unsigned int dirty_page = 0;
2591 struct page *page; 2898 struct page *page;
2592 struct file_ra_state *ra; 2899 struct file_ra_state *ra;
2593 int nr = 0; 2900 int nr = 0;
@@ -2600,21 +2907,24 @@ static int relocate_file_extent_cluster(struct inode *inode,
2600 if (!ra) 2907 if (!ra)
2601 return -ENOMEM; 2908 return -ENOMEM;
2602 2909
2603 index = (cluster->start - offset) >> PAGE_CACHE_SHIFT; 2910 ret = prealloc_file_extent_cluster(inode, cluster);
2604 last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT; 2911 if (ret)
2912 goto out;
2605 2913
2606 mutex_lock(&inode->i_mutex); 2914 file_ra_state_init(ra, inode->i_mapping);
2607 2915
2608 i_size_write(inode, cluster->end + 1 - offset);
2609 ret = setup_extent_mapping(inode, cluster->start - offset, 2916 ret = setup_extent_mapping(inode, cluster->start - offset,
2610 cluster->end - offset, cluster->start); 2917 cluster->end - offset, cluster->start);
2611 if (ret) 2918 if (ret)
2612 goto out_unlock; 2919 goto out;
2613
2614 file_ra_state_init(ra, inode->i_mapping);
2615 2920
2616 WARN_ON(cluster->start != cluster->boundary[0]); 2921 index = (cluster->start - offset) >> PAGE_CACHE_SHIFT;
2922 last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT;
2617 while (index <= last_index) { 2923 while (index <= last_index) {
2924 ret = btrfs_delalloc_reserve_metadata(inode, PAGE_CACHE_SIZE);
2925 if (ret)
2926 goto out;
2927
2618 page = find_lock_page(inode->i_mapping, index); 2928 page = find_lock_page(inode->i_mapping, index);
2619 if (!page) { 2929 if (!page) {
2620 page_cache_sync_readahead(inode->i_mapping, 2930 page_cache_sync_readahead(inode->i_mapping,
@@ -2622,8 +2932,10 @@ static int relocate_file_extent_cluster(struct inode *inode,
2622 last_index + 1 - index); 2932 last_index + 1 - index);
2623 page = grab_cache_page(inode->i_mapping, index); 2933 page = grab_cache_page(inode->i_mapping, index);
2624 if (!page) { 2934 if (!page) {
2935 btrfs_delalloc_release_metadata(inode,
2936 PAGE_CACHE_SIZE);
2625 ret = -ENOMEM; 2937 ret = -ENOMEM;
2626 goto out_unlock; 2938 goto out;
2627 } 2939 }
2628 } 2940 }
2629 2941
@@ -2639,8 +2951,10 @@ static int relocate_file_extent_cluster(struct inode *inode,
2639 if (!PageUptodate(page)) { 2951 if (!PageUptodate(page)) {
2640 unlock_page(page); 2952 unlock_page(page);
2641 page_cache_release(page); 2953 page_cache_release(page);
2954 btrfs_delalloc_release_metadata(inode,
2955 PAGE_CACHE_SIZE);
2642 ret = -EIO; 2956 ret = -EIO;
2643 goto out_unlock; 2957 goto out;
2644 } 2958 }
2645 } 2959 }
2646 2960
@@ -2659,10 +2973,9 @@ static int relocate_file_extent_cluster(struct inode *inode,
2659 EXTENT_BOUNDARY, GFP_NOFS); 2973 EXTENT_BOUNDARY, GFP_NOFS);
2660 nr++; 2974 nr++;
2661 } 2975 }
2662 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
2663 2976
2977 btrfs_set_extent_delalloc(inode, page_start, page_end, NULL);
2664 set_page_dirty(page); 2978 set_page_dirty(page);
2665 dirty_page++;
2666 2979
2667 unlock_extent(&BTRFS_I(inode)->io_tree, 2980 unlock_extent(&BTRFS_I(inode)->io_tree,
2668 page_start, page_end, GFP_NOFS); 2981 page_start, page_end, GFP_NOFS);
@@ -2670,20 +2983,11 @@ static int relocate_file_extent_cluster(struct inode *inode,
2670 page_cache_release(page); 2983 page_cache_release(page);
2671 2984
2672 index++; 2985 index++;
2673 if (nr < cluster->nr && 2986 balance_dirty_pages_ratelimited(inode->i_mapping);
2674 page_end + 1 + offset == cluster->boundary[nr]) { 2987 btrfs_throttle(BTRFS_I(inode)->root);
2675 balance_dirty_pages_ratelimited_nr(inode->i_mapping,
2676 dirty_page);
2677 dirty_page = 0;
2678 }
2679 }
2680 if (dirty_page) {
2681 balance_dirty_pages_ratelimited_nr(inode->i_mapping,
2682 dirty_page);
2683 } 2988 }
2684 WARN_ON(nr != cluster->nr); 2989 WARN_ON(nr != cluster->nr);
2685out_unlock: 2990out:
2686 mutex_unlock(&inode->i_mutex);
2687 kfree(ra); 2991 kfree(ra);
2688 return ret; 2992 return ret;
2689} 2993}
@@ -2869,9 +3173,6 @@ out:
2869static int block_use_full_backref(struct reloc_control *rc, 3173static int block_use_full_backref(struct reloc_control *rc,
2870 struct extent_buffer *eb) 3174 struct extent_buffer *eb)
2871{ 3175{
2872 struct btrfs_path *path;
2873 struct btrfs_extent_item *ei;
2874 struct btrfs_key key;
2875 u64 flags; 3176 u64 flags;
2876 int ret; 3177 int ret;
2877 3178
@@ -2879,28 +3180,14 @@ static int block_use_full_backref(struct reloc_control *rc,
2879 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV) 3180 btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV)
2880 return 1; 3181 return 1;
2881 3182
2882 path = btrfs_alloc_path(); 3183 ret = btrfs_lookup_extent_info(NULL, rc->extent_root,
2883 BUG_ON(!path); 3184 eb->start, eb->len, NULL, &flags);
2884
2885 key.objectid = eb->start;
2886 key.type = BTRFS_EXTENT_ITEM_KEY;
2887 key.offset = eb->len;
2888
2889 path->search_commit_root = 1;
2890 path->skip_locking = 1;
2891 ret = btrfs_search_slot(NULL, rc->extent_root,
2892 &key, path, 0, 0);
2893 BUG_ON(ret); 3185 BUG_ON(ret);
2894 3186
2895 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2896 struct btrfs_extent_item);
2897 flags = btrfs_extent_flags(path->nodes[0], ei);
2898 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2899 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) 3187 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
2900 ret = 1; 3188 ret = 1;
2901 else 3189 else
2902 ret = 0; 3190 ret = 0;
2903 btrfs_free_path(path);
2904 return ret; 3191 return ret;
2905} 3192}
2906 3193
@@ -3073,22 +3360,10 @@ int add_data_references(struct reloc_control *rc,
3073 struct btrfs_extent_inline_ref *iref; 3360 struct btrfs_extent_inline_ref *iref;
3074 unsigned long ptr; 3361 unsigned long ptr;
3075 unsigned long end; 3362 unsigned long end;
3076 u32 blocksize; 3363 u32 blocksize = btrfs_level_size(rc->extent_root, 0);
3077 int ret; 3364 int ret;
3078 int err = 0; 3365 int err = 0;
3079 3366
3080 ret = get_new_location(rc->data_inode, NULL, extent_key->objectid,
3081 extent_key->offset);
3082 BUG_ON(ret < 0);
3083 if (ret > 0) {
3084 /* the relocated data is fragmented */
3085 rc->extents_skipped++;
3086 btrfs_release_path(rc->extent_root, path);
3087 return 0;
3088 }
3089
3090 blocksize = btrfs_level_size(rc->extent_root, 0);
3091
3092 eb = path->nodes[0]; 3367 eb = path->nodes[0];
3093 ptr = btrfs_item_ptr_offset(eb, path->slots[0]); 3368 ptr = btrfs_item_ptr_offset(eb, path->slots[0]);
3094 end = ptr + btrfs_item_size_nr(eb, path->slots[0]); 3369 end = ptr + btrfs_item_size_nr(eb, path->slots[0]);
@@ -3169,7 +3444,8 @@ int add_data_references(struct reloc_control *rc,
3169 */ 3444 */
3170static noinline_for_stack 3445static noinline_for_stack
3171int find_next_extent(struct btrfs_trans_handle *trans, 3446int find_next_extent(struct btrfs_trans_handle *trans,
3172 struct reloc_control *rc, struct btrfs_path *path) 3447 struct reloc_control *rc, struct btrfs_path *path,
3448 struct btrfs_key *extent_key)
3173{ 3449{
3174 struct btrfs_key key; 3450 struct btrfs_key key;
3175 struct extent_buffer *leaf; 3451 struct extent_buffer *leaf;
@@ -3224,6 +3500,7 @@ next:
3224 rc->search_start = end + 1; 3500 rc->search_start = end + 1;
3225 } else { 3501 } else {
3226 rc->search_start = key.objectid + key.offset; 3502 rc->search_start = key.objectid + key.offset;
3503 memcpy(extent_key, &key, sizeof(key));
3227 return 0; 3504 return 0;
3228 } 3505 }
3229 } 3506 }
@@ -3261,12 +3538,49 @@ static int check_extent_flags(u64 flags)
3261 return 0; 3538 return 0;
3262} 3539}
3263 3540
3541static noinline_for_stack
3542int prepare_to_relocate(struct reloc_control *rc)
3543{
3544 struct btrfs_trans_handle *trans;
3545 int ret;
3546
3547 rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root);
3548 if (!rc->block_rsv)
3549 return -ENOMEM;
3550
3551 /*
3552 * reserve some space for creating reloc trees.
3553 * btrfs_init_reloc_root will use them when there
3554 * is no reservation in transaction handle.
3555 */
3556 ret = btrfs_block_rsv_add(NULL, rc->extent_root, rc->block_rsv,
3557 rc->extent_root->nodesize * 256,
3558 &rc->block_rsv_retries);
3559 if (ret)
3560 return ret;
3561
3562 rc->block_rsv->refill_used = 1;
3563 btrfs_add_durable_block_rsv(rc->extent_root->fs_info, rc->block_rsv);
3564
3565 memset(&rc->cluster, 0, sizeof(rc->cluster));
3566 rc->search_start = rc->block_group->key.objectid;
3567 rc->extents_found = 0;
3568 rc->nodes_relocated = 0;
3569 rc->merging_rsv_size = 0;
3570 rc->block_rsv_retries = 0;
3571
3572 rc->create_reloc_tree = 1;
3573 set_reloc_control(rc);
3574
3575 trans = btrfs_join_transaction(rc->extent_root, 1);
3576 btrfs_commit_transaction(trans, rc->extent_root);
3577 return 0;
3578}
3264 3579
3265static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 3580static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3266{ 3581{
3267 struct rb_root blocks = RB_ROOT; 3582 struct rb_root blocks = RB_ROOT;
3268 struct btrfs_key key; 3583 struct btrfs_key key;
3269 struct file_extent_cluster *cluster;
3270 struct btrfs_trans_handle *trans = NULL; 3584 struct btrfs_trans_handle *trans = NULL;
3271 struct btrfs_path *path; 3585 struct btrfs_path *path;
3272 struct btrfs_extent_item *ei; 3586 struct btrfs_extent_item *ei;
@@ -3276,33 +3590,25 @@ static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3276 int ret; 3590 int ret;
3277 int err = 0; 3591 int err = 0;
3278 3592
3279 cluster = kzalloc(sizeof(*cluster), GFP_NOFS);
3280 if (!cluster)
3281 return -ENOMEM;
3282
3283 path = btrfs_alloc_path(); 3593 path = btrfs_alloc_path();
3284 if (!path) { 3594 if (!path)
3285 kfree(cluster);
3286 return -ENOMEM; 3595 return -ENOMEM;
3287 }
3288
3289 rc->extents_found = 0;
3290 rc->extents_skipped = 0;
3291
3292 rc->search_start = rc->block_group->key.objectid;
3293 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3294 GFP_NOFS);
3295
3296 rc->create_reloc_root = 1;
3297 set_reloc_control(rc);
3298 3596
3299 trans = btrfs_start_transaction(rc->extent_root, 1); 3597 ret = prepare_to_relocate(rc);
3300 btrfs_commit_transaction(trans, rc->extent_root); 3598 if (ret) {
3599 err = ret;
3600 goto out_free;
3601 }
3301 3602
3302 while (1) { 3603 while (1) {
3303 trans = btrfs_start_transaction(rc->extent_root, 1); 3604 trans = btrfs_start_transaction(rc->extent_root, 0);
3605
3606 if (update_backref_cache(trans, &rc->backref_cache)) {
3607 btrfs_end_transaction(trans, rc->extent_root);
3608 continue;
3609 }
3304 3610
3305 ret = find_next_extent(trans, rc, path); 3611 ret = find_next_extent(trans, rc, path, &key);
3306 if (ret < 0) 3612 if (ret < 0)
3307 err = ret; 3613 err = ret;
3308 if (ret != 0) 3614 if (ret != 0)
@@ -3312,9 +3618,7 @@ static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3312 3618
3313 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 3619 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3314 struct btrfs_extent_item); 3620 struct btrfs_extent_item);
3315 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 3621 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3316 item_size = btrfs_item_size_nr(path->nodes[0],
3317 path->slots[0]);
3318 if (item_size >= sizeof(*ei)) { 3622 if (item_size >= sizeof(*ei)) {
3319 flags = btrfs_extent_flags(path->nodes[0], ei); 3623 flags = btrfs_extent_flags(path->nodes[0], ei);
3320 ret = check_extent_flags(flags); 3624 ret = check_extent_flags(flags);
@@ -3355,73 +3659,100 @@ static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3355 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 3659 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3356 ret = add_tree_block(rc, &key, path, &blocks); 3660 ret = add_tree_block(rc, &key, path, &blocks);
3357 } else if (rc->stage == UPDATE_DATA_PTRS && 3661 } else if (rc->stage == UPDATE_DATA_PTRS &&
3358 (flags & BTRFS_EXTENT_FLAG_DATA)) { 3662 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3359 ret = add_data_references(rc, &key, path, &blocks); 3663 ret = add_data_references(rc, &key, path, &blocks);
3360 } else { 3664 } else {
3361 btrfs_release_path(rc->extent_root, path); 3665 btrfs_release_path(rc->extent_root, path);
3362 ret = 0; 3666 ret = 0;
3363 } 3667 }
3364 if (ret < 0) { 3668 if (ret < 0) {
3365 err = 0; 3669 err = ret;
3366 break; 3670 break;
3367 } 3671 }
3368 3672
3369 if (!RB_EMPTY_ROOT(&blocks)) { 3673 if (!RB_EMPTY_ROOT(&blocks)) {
3370 ret = relocate_tree_blocks(trans, rc, &blocks); 3674 ret = relocate_tree_blocks(trans, rc, &blocks);
3371 if (ret < 0) { 3675 if (ret < 0) {
3676 if (ret != -EAGAIN) {
3677 err = ret;
3678 break;
3679 }
3680 rc->extents_found--;
3681 rc->search_start = key.objectid;
3682 }
3683 }
3684
3685 ret = btrfs_block_rsv_check(trans, rc->extent_root,
3686 rc->block_rsv, 0, 5);
3687 if (ret < 0) {
3688 if (ret != -EAGAIN) {
3372 err = ret; 3689 err = ret;
3690 WARN_ON(1);
3373 break; 3691 break;
3374 } 3692 }
3693 rc->commit_transaction = 1;
3375 } 3694 }
3376 3695
3377 nr = trans->blocks_used; 3696 if (rc->commit_transaction) {
3378 btrfs_end_transaction(trans, rc->extent_root); 3697 rc->commit_transaction = 0;
3698 ret = btrfs_commit_transaction(trans, rc->extent_root);
3699 BUG_ON(ret);
3700 } else {
3701 nr = trans->blocks_used;
3702 btrfs_end_transaction_throttle(trans, rc->extent_root);
3703 btrfs_btree_balance_dirty(rc->extent_root, nr);
3704 }
3379 trans = NULL; 3705 trans = NULL;
3380 btrfs_btree_balance_dirty(rc->extent_root, nr);
3381 3706
3382 if (rc->stage == MOVE_DATA_EXTENTS && 3707 if (rc->stage == MOVE_DATA_EXTENTS &&
3383 (flags & BTRFS_EXTENT_FLAG_DATA)) { 3708 (flags & BTRFS_EXTENT_FLAG_DATA)) {
3384 rc->found_file_extent = 1; 3709 rc->found_file_extent = 1;
3385 ret = relocate_data_extent(rc->data_inode, 3710 ret = relocate_data_extent(rc->data_inode,
3386 &key, cluster); 3711 &key, &rc->cluster);
3387 if (ret < 0) { 3712 if (ret < 0) {
3388 err = ret; 3713 err = ret;
3389 break; 3714 break;
3390 } 3715 }
3391 } 3716 }
3392 } 3717 }
3393 btrfs_free_path(path); 3718
3719 btrfs_release_path(rc->extent_root, path);
3720 clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY,
3721 GFP_NOFS);
3394 3722
3395 if (trans) { 3723 if (trans) {
3396 nr = trans->blocks_used; 3724 nr = trans->blocks_used;
3397 btrfs_end_transaction(trans, rc->extent_root); 3725 btrfs_end_transaction_throttle(trans, rc->extent_root);
3398 btrfs_btree_balance_dirty(rc->extent_root, nr); 3726 btrfs_btree_balance_dirty(rc->extent_root, nr);
3399 } 3727 }
3400 3728
3401 if (!err) { 3729 if (!err) {
3402 ret = relocate_file_extent_cluster(rc->data_inode, cluster); 3730 ret = relocate_file_extent_cluster(rc->data_inode,
3731 &rc->cluster);
3403 if (ret < 0) 3732 if (ret < 0)
3404 err = ret; 3733 err = ret;
3405 } 3734 }
3406 3735
3407 kfree(cluster); 3736 rc->create_reloc_tree = 0;
3737 set_reloc_control(rc);
3408 3738
3409 rc->create_reloc_root = 0; 3739 backref_cache_cleanup(&rc->backref_cache);
3410 smp_mb(); 3740 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3411 3741
3412 if (rc->extents_found > 0) { 3742 err = prepare_to_merge(rc, err);
3413 trans = btrfs_start_transaction(rc->extent_root, 1);
3414 btrfs_commit_transaction(trans, rc->extent_root);
3415 }
3416 3743
3417 merge_reloc_roots(rc); 3744 merge_reloc_roots(rc);
3418 3745
3746 rc->merge_reloc_tree = 0;
3419 unset_reloc_control(rc); 3747 unset_reloc_control(rc);
3748 btrfs_block_rsv_release(rc->extent_root, rc->block_rsv, (u64)-1);
3420 3749
3421 /* get rid of pinned extents */ 3750 /* get rid of pinned extents */
3422 trans = btrfs_start_transaction(rc->extent_root, 1); 3751 trans = btrfs_join_transaction(rc->extent_root, 1);
3423 btrfs_commit_transaction(trans, rc->extent_root); 3752 btrfs_commit_transaction(trans, rc->extent_root);
3424 3753out_free:
3754 btrfs_free_block_rsv(rc->extent_root, rc->block_rsv);
3755 btrfs_free_path(path);
3425 return err; 3756 return err;
3426} 3757}
3427 3758
@@ -3447,7 +3778,8 @@ static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3447 btrfs_set_inode_generation(leaf, item, 1); 3778 btrfs_set_inode_generation(leaf, item, 1);
3448 btrfs_set_inode_size(leaf, item, 0); 3779 btrfs_set_inode_size(leaf, item, 0);
3449 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 3780 btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3450 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS); 3781 btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3782 BTRFS_INODE_PREALLOC);
3451 btrfs_mark_buffer_dirty(leaf); 3783 btrfs_mark_buffer_dirty(leaf);
3452 btrfs_release_path(root, path); 3784 btrfs_release_path(root, path);
3453out: 3785out:
@@ -3459,8 +3791,9 @@ out:
3459 * helper to create inode for data relocation. 3791 * helper to create inode for data relocation.
3460 * the inode is in data relocation tree and its link count is 0 3792 * the inode is in data relocation tree and its link count is 0
3461 */ 3793 */
3462static struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 3794static noinline_for_stack
3463 struct btrfs_block_group_cache *group) 3795struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3796 struct btrfs_block_group_cache *group)
3464{ 3797{
3465 struct inode *inode = NULL; 3798 struct inode *inode = NULL;
3466 struct btrfs_trans_handle *trans; 3799 struct btrfs_trans_handle *trans;
@@ -3474,8 +3807,9 @@ static struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3474 if (IS_ERR(root)) 3807 if (IS_ERR(root))
3475 return ERR_CAST(root); 3808 return ERR_CAST(root);
3476 3809
3477 trans = btrfs_start_transaction(root, 1); 3810 trans = btrfs_start_transaction(root, 6);
3478 BUG_ON(!trans); 3811 if (IS_ERR(trans))
3812 return ERR_CAST(trans);
3479 3813
3480 err = btrfs_find_free_objectid(trans, root, objectid, &objectid); 3814 err = btrfs_find_free_objectid(trans, root, objectid, &objectid);
3481 if (err) 3815 if (err)
@@ -3495,7 +3829,6 @@ static struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3495out: 3829out:
3496 nr = trans->blocks_used; 3830 nr = trans->blocks_used;
3497 btrfs_end_transaction(trans, root); 3831 btrfs_end_transaction(trans, root);
3498
3499 btrfs_btree_balance_dirty(root, nr); 3832 btrfs_btree_balance_dirty(root, nr);
3500 if (err) { 3833 if (err) {
3501 if (inode) 3834 if (inode)
@@ -3505,6 +3838,21 @@ out:
3505 return inode; 3838 return inode;
3506} 3839}
3507 3840
3841static struct reloc_control *alloc_reloc_control(void)
3842{
3843 struct reloc_control *rc;
3844
3845 rc = kzalloc(sizeof(*rc), GFP_NOFS);
3846 if (!rc)
3847 return NULL;
3848
3849 INIT_LIST_HEAD(&rc->reloc_roots);
3850 backref_cache_init(&rc->backref_cache);
3851 mapping_tree_init(&rc->reloc_root_tree);
3852 extent_io_tree_init(&rc->processed_blocks, NULL, GFP_NOFS);
3853 return rc;
3854}
3855
3508/* 3856/*
3509 * function to relocate all extents in a block group. 3857 * function to relocate all extents in a block group.
3510 */ 3858 */
@@ -3513,24 +3861,26 @@ int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3513 struct btrfs_fs_info *fs_info = extent_root->fs_info; 3861 struct btrfs_fs_info *fs_info = extent_root->fs_info;
3514 struct reloc_control *rc; 3862 struct reloc_control *rc;
3515 int ret; 3863 int ret;
3864 int rw = 0;
3516 int err = 0; 3865 int err = 0;
3517 3866
3518 rc = kzalloc(sizeof(*rc), GFP_NOFS); 3867 rc = alloc_reloc_control();
3519 if (!rc) 3868 if (!rc)
3520 return -ENOMEM; 3869 return -ENOMEM;
3521 3870
3522 mapping_tree_init(&rc->reloc_root_tree); 3871 rc->extent_root = extent_root;
3523 extent_io_tree_init(&rc->processed_blocks, NULL, GFP_NOFS);
3524 INIT_LIST_HEAD(&rc->reloc_roots);
3525 3872
3526 rc->block_group = btrfs_lookup_block_group(fs_info, group_start); 3873 rc->block_group = btrfs_lookup_block_group(fs_info, group_start);
3527 BUG_ON(!rc->block_group); 3874 BUG_ON(!rc->block_group);
3528 3875
3529 btrfs_init_workers(&rc->workers, "relocate", 3876 if (!rc->block_group->ro) {
3530 fs_info->thread_pool_size, NULL); 3877 ret = btrfs_set_block_group_ro(extent_root, rc->block_group);
3531 3878 if (ret) {
3532 rc->extent_root = extent_root; 3879 err = ret;
3533 btrfs_prepare_block_group_relocation(extent_root, rc->block_group); 3880 goto out;
3881 }
3882 rw = 1;
3883 }
3534 3884
3535 rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 3885 rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
3536 if (IS_ERR(rc->data_inode)) { 3886 if (IS_ERR(rc->data_inode)) {
@@ -3547,9 +3897,6 @@ int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3547 btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0); 3897 btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0);
3548 3898
3549 while (1) { 3899 while (1) {
3550 rc->extents_found = 0;
3551 rc->extents_skipped = 0;
3552
3553 mutex_lock(&fs_info->cleaner_mutex); 3900 mutex_lock(&fs_info->cleaner_mutex);
3554 3901
3555 btrfs_clean_old_snapshots(fs_info->tree_root); 3902 btrfs_clean_old_snapshots(fs_info->tree_root);
@@ -3558,7 +3905,7 @@ int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3558 mutex_unlock(&fs_info->cleaner_mutex); 3905 mutex_unlock(&fs_info->cleaner_mutex);
3559 if (ret < 0) { 3906 if (ret < 0) {
3560 err = ret; 3907 err = ret;
3561 break; 3908 goto out;
3562 } 3909 }
3563 3910
3564 if (rc->extents_found == 0) 3911 if (rc->extents_found == 0)
@@ -3572,18 +3919,6 @@ int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3572 invalidate_mapping_pages(rc->data_inode->i_mapping, 3919 invalidate_mapping_pages(rc->data_inode->i_mapping,
3573 0, -1); 3920 0, -1);
3574 rc->stage = UPDATE_DATA_PTRS; 3921 rc->stage = UPDATE_DATA_PTRS;
3575 } else if (rc->stage == UPDATE_DATA_PTRS &&
3576 rc->extents_skipped >= rc->extents_found) {
3577 iput(rc->data_inode);
3578 rc->data_inode = create_reloc_inode(fs_info,
3579 rc->block_group);
3580 if (IS_ERR(rc->data_inode)) {
3581 err = PTR_ERR(rc->data_inode);
3582 rc->data_inode = NULL;
3583 break;
3584 }
3585 rc->stage = MOVE_DATA_EXTENTS;
3586 rc->found_file_extent = 0;
3587 } 3922 }
3588 } 3923 }
3589 3924
@@ -3596,8 +3931,9 @@ int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start)
3596 WARN_ON(rc->block_group->reserved > 0); 3931 WARN_ON(rc->block_group->reserved > 0);
3597 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0); 3932 WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0);
3598out: 3933out:
3934 if (err && rw)
3935 btrfs_set_block_group_rw(extent_root, rc->block_group);
3599 iput(rc->data_inode); 3936 iput(rc->data_inode);
3600 btrfs_stop_workers(&rc->workers);
3601 btrfs_put_block_group(rc->block_group); 3937 btrfs_put_block_group(rc->block_group);
3602 kfree(rc); 3938 kfree(rc);
3603 return err; 3939 return err;
@@ -3608,7 +3944,7 @@ static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
3608 struct btrfs_trans_handle *trans; 3944 struct btrfs_trans_handle *trans;
3609 int ret; 3945 int ret;
3610 3946
3611 trans = btrfs_start_transaction(root->fs_info->tree_root, 1); 3947 trans = btrfs_start_transaction(root->fs_info->tree_root, 0);
3612 3948
3613 memset(&root->root_item.drop_progress, 0, 3949 memset(&root->root_item.drop_progress, 0,
3614 sizeof(root->root_item.drop_progress)); 3950 sizeof(root->root_item.drop_progress));
@@ -3701,20 +4037,20 @@ int btrfs_recover_relocation(struct btrfs_root *root)
3701 if (list_empty(&reloc_roots)) 4037 if (list_empty(&reloc_roots))
3702 goto out; 4038 goto out;
3703 4039
3704 rc = kzalloc(sizeof(*rc), GFP_NOFS); 4040 rc = alloc_reloc_control();
3705 if (!rc) { 4041 if (!rc) {
3706 err = -ENOMEM; 4042 err = -ENOMEM;
3707 goto out; 4043 goto out;
3708 } 4044 }
3709 4045
3710 mapping_tree_init(&rc->reloc_root_tree);
3711 INIT_LIST_HEAD(&rc->reloc_roots);
3712 btrfs_init_workers(&rc->workers, "relocate",
3713 root->fs_info->thread_pool_size, NULL);
3714 rc->extent_root = root->fs_info->extent_root; 4046 rc->extent_root = root->fs_info->extent_root;
3715 4047
3716 set_reloc_control(rc); 4048 set_reloc_control(rc);
3717 4049
4050 trans = btrfs_join_transaction(rc->extent_root, 1);
4051
4052 rc->merge_reloc_tree = 1;
4053
3718 while (!list_empty(&reloc_roots)) { 4054 while (!list_empty(&reloc_roots)) {
3719 reloc_root = list_entry(reloc_roots.next, 4055 reloc_root = list_entry(reloc_roots.next,
3720 struct btrfs_root, root_list); 4056 struct btrfs_root, root_list);
@@ -3734,20 +4070,16 @@ int btrfs_recover_relocation(struct btrfs_root *root)
3734 fs_root->reloc_root = reloc_root; 4070 fs_root->reloc_root = reloc_root;
3735 } 4071 }
3736 4072
3737 trans = btrfs_start_transaction(rc->extent_root, 1);
3738 btrfs_commit_transaction(trans, rc->extent_root); 4073 btrfs_commit_transaction(trans, rc->extent_root);
3739 4074
3740 merge_reloc_roots(rc); 4075 merge_reloc_roots(rc);
3741 4076
3742 unset_reloc_control(rc); 4077 unset_reloc_control(rc);
3743 4078
3744 trans = btrfs_start_transaction(rc->extent_root, 1); 4079 trans = btrfs_join_transaction(rc->extent_root, 1);
3745 btrfs_commit_transaction(trans, rc->extent_root); 4080 btrfs_commit_transaction(trans, rc->extent_root);
3746out: 4081out:
3747 if (rc) { 4082 kfree(rc);
3748 btrfs_stop_workers(&rc->workers);
3749 kfree(rc);
3750 }
3751 while (!list_empty(&reloc_roots)) { 4083 while (!list_empty(&reloc_roots)) {
3752 reloc_root = list_entry(reloc_roots.next, 4084 reloc_root = list_entry(reloc_roots.next,
3753 struct btrfs_root, root_list); 4085 struct btrfs_root, root_list);
@@ -3813,3 +4145,130 @@ int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len)
3813 btrfs_put_ordered_extent(ordered); 4145 btrfs_put_ordered_extent(ordered);
3814 return 0; 4146 return 0;
3815} 4147}
4148
4149void btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
4150 struct btrfs_root *root, struct extent_buffer *buf,
4151 struct extent_buffer *cow)
4152{
4153 struct reloc_control *rc;
4154 struct backref_node *node;
4155 int first_cow = 0;
4156 int level;
4157 int ret;
4158
4159 rc = root->fs_info->reloc_ctl;
4160 if (!rc)
4161 return;
4162
4163 BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
4164 root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
4165
4166 level = btrfs_header_level(buf);
4167 if (btrfs_header_generation(buf) <=
4168 btrfs_root_last_snapshot(&root->root_item))
4169 first_cow = 1;
4170
4171 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
4172 rc->create_reloc_tree) {
4173 WARN_ON(!first_cow && level == 0);
4174
4175 node = rc->backref_cache.path[level];
4176 BUG_ON(node->bytenr != buf->start &&
4177 node->new_bytenr != buf->start);
4178
4179 drop_node_buffer(node);
4180 extent_buffer_get(cow);
4181 node->eb = cow;
4182 node->new_bytenr = cow->start;
4183
4184 if (!node->pending) {
4185 list_move_tail(&node->list,
4186 &rc->backref_cache.pending[level]);
4187 node->pending = 1;
4188 }
4189
4190 if (first_cow)
4191 __mark_block_processed(rc, node);
4192
4193 if (first_cow && level > 0)
4194 rc->nodes_relocated += buf->len;
4195 }
4196
4197 if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS) {
4198 ret = replace_file_extents(trans, rc, root, cow);
4199 BUG_ON(ret);
4200 }
4201}
4202
4203/*
4204 * called before creating snapshot. it calculates metadata reservation
4205 * requried for relocating tree blocks in the snapshot
4206 */
4207void btrfs_reloc_pre_snapshot(struct btrfs_trans_handle *trans,
4208 struct btrfs_pending_snapshot *pending,
4209 u64 *bytes_to_reserve)
4210{
4211 struct btrfs_root *root;
4212 struct reloc_control *rc;
4213
4214 root = pending->root;
4215 if (!root->reloc_root)
4216 return;
4217
4218 rc = root->fs_info->reloc_ctl;
4219 if (!rc->merge_reloc_tree)
4220 return;
4221
4222 root = root->reloc_root;
4223 BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4224 /*
4225 * relocation is in the stage of merging trees. the space
4226 * used by merging a reloc tree is twice the size of
4227 * relocated tree nodes in the worst case. half for cowing
4228 * the reloc tree, half for cowing the fs tree. the space
4229 * used by cowing the reloc tree will be freed after the
4230 * tree is dropped. if we create snapshot, cowing the fs
4231 * tree may use more space than it frees. so we need
4232 * reserve extra space.
4233 */
4234 *bytes_to_reserve += rc->nodes_relocated;
4235}
4236
4237/*
4238 * called after snapshot is created. migrate block reservation
4239 * and create reloc root for the newly created snapshot
4240 */
4241void btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4242 struct btrfs_pending_snapshot *pending)
4243{
4244 struct btrfs_root *root = pending->root;
4245 struct btrfs_root *reloc_root;
4246 struct btrfs_root *new_root;
4247 struct reloc_control *rc;
4248 int ret;
4249
4250 if (!root->reloc_root)
4251 return;
4252
4253 rc = root->fs_info->reloc_ctl;
4254 rc->merging_rsv_size += rc->nodes_relocated;
4255
4256 if (rc->merge_reloc_tree) {
4257 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4258 rc->block_rsv,
4259 rc->nodes_relocated);
4260 BUG_ON(ret);
4261 }
4262
4263 new_root = pending->snap;
4264 reloc_root = create_reloc_root(trans, root->reloc_root,
4265 new_root->root_key.objectid);
4266
4267 __add_reloc_root(reloc_root);
4268 new_root->reloc_root = reloc_root;
4269
4270 if (rc->create_reloc_tree) {
4271 ret = clone_backref_node(trans, rc, root, reloc_root);
4272 BUG_ON(ret);
4273 }
4274}