Btrfs: fix sleep with spinlock held during unmount
[linux-2.6-block.git] / fs / btrfs / extent-tree.c
index c0bb6b9ac4c6b7af43430927826e670ef5cdcd11..fe4e11b31a43898a64aec09db50667e488cb853e 100644 (file)
 #include "locking.h"
 #include "ref-cache.h"
 
-#define BLOCK_GROUP_DATA     EXTENT_WRITEBACK
-#define BLOCK_GROUP_METADATA EXTENT_UPTODATE
-#define BLOCK_GROUP_SYSTEM   EXTENT_NEW
+#define PENDING_EXTENT_INSERT 0
+#define PENDING_EXTENT_DELETE 1
+#define PENDING_BACKREF_UPDATE 2
 
-#define BLOCK_GROUP_DIRTY EXTENT_DIRTY
+struct pending_extent_op {
+       int type;
+       u64 bytenr;
+       u64 num_bytes;
+       u64 parent;
+       u64 orig_parent;
+       u64 generation;
+       u64 orig_generation;
+       int level;
+};
 
 static int finish_current_insert(struct btrfs_trans_handle *trans, struct
                                 btrfs_root *extent_root);
@@ -62,17 +71,136 @@ void maybe_unlock_mutex(struct btrfs_root *root)
        }
 }
 
+static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
+{
+       return (cache->flags & bits) == bits;
+}
+
+/*
+ * this adds the block group to the fs_info rb tree for the block group
+ * cache
+ */
+int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
+                               struct btrfs_block_group_cache *block_group)
+{
+       struct rb_node **p;
+       struct rb_node *parent = NULL;
+       struct btrfs_block_group_cache *cache;
+
+       spin_lock(&info->block_group_cache_lock);
+       p = &info->block_group_cache_tree.rb_node;
+
+       while (*p) {
+               parent = *p;
+               cache = rb_entry(parent, struct btrfs_block_group_cache,
+                                cache_node);
+               if (block_group->key.objectid < cache->key.objectid) {
+                       p = &(*p)->rb_left;
+               } else if (block_group->key.objectid > cache->key.objectid) {
+                       p = &(*p)->rb_right;
+               } else {
+                       spin_unlock(&info->block_group_cache_lock);
+                       return -EEXIST;
+               }
+       }
+
+       rb_link_node(&block_group->cache_node, parent, p);
+       rb_insert_color(&block_group->cache_node,
+                       &info->block_group_cache_tree);
+       spin_unlock(&info->block_group_cache_lock);
+
+       return 0;
+}
+
+/*
+ * This will return the block group at or after bytenr if contains is 0, else
+ * it will return the block group that contains the bytenr
+ */
+static struct btrfs_block_group_cache *
+block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
+                             int contains)
+{
+       struct btrfs_block_group_cache *cache, *ret = NULL;
+       struct rb_node *n;
+       u64 end, start;
+
+       spin_lock(&info->block_group_cache_lock);
+       n = info->block_group_cache_tree.rb_node;
+
+       while (n) {
+               cache = rb_entry(n, struct btrfs_block_group_cache,
+                                cache_node);
+               end = cache->key.objectid + cache->key.offset - 1;
+               start = cache->key.objectid;
+
+               if (bytenr < start) {
+                       if (!contains && (!ret || start < ret->key.objectid))
+                               ret = cache;
+                       n = n->rb_left;
+               } else if (bytenr > start) {
+                       if (contains && bytenr <= end) {
+                               ret = cache;
+                               break;
+                       }
+                       n = n->rb_right;
+               } else {
+                       ret = cache;
+                       break;
+               }
+       }
+       spin_unlock(&info->block_group_cache_lock);
+
+       return ret;
+}
+
+/*
+ * this is only called by cache_block_group, since we could have freed extents
+ * we need to check the pinned_extents for any extents that can't be used yet
+ * since their free space will be released as soon as the transaction commits.
+ */
+static int add_new_free_space(struct btrfs_block_group_cache *block_group,
+                             struct btrfs_fs_info *info, u64 start, u64 end)
+{
+       u64 extent_start, extent_end, size;
+       int ret;
+
+       while (start < end) {
+               ret = find_first_extent_bit(&info->pinned_extents, start,
+                                           &extent_start, &extent_end,
+                                           EXTENT_DIRTY);
+               if (ret)
+                       break;
+
+               if (extent_start == start) {
+                       start = extent_end + 1;
+               } else if (extent_start > start && extent_start < end) {
+                       size = extent_start - start;
+                       ret = btrfs_add_free_space(block_group, start, size);
+                       BUG_ON(ret);
+                       start = extent_end + 1;
+               } else {
+                       break;
+               }
+       }
+
+       if (start < end) {
+               size = end - start;
+               ret = btrfs_add_free_space(block_group, start, size);
+               BUG_ON(ret);
+       }
+
+       return 0;
+}
+
 static int cache_block_group(struct btrfs_root *root,
                             struct btrfs_block_group_cache *block_group)
 {
        struct btrfs_path *path;
-       int ret;
+       int ret = 0;
        struct btrfs_key key;
        struct extent_buffer *leaf;
-       struct extent_io_tree *free_space_cache;
        int slot;
        u64 last = 0;
-       u64 hole_size;
        u64 first_free;
        int found = 0;
 
@@ -80,7 +208,6 @@ static int cache_block_group(struct btrfs_root *root,
                return 0;
 
        root = root->fs_info->extent_root;
-       free_space_cache = &root->fs_info->free_space_cache;
 
        if (block_group->cached)
                return 0;
@@ -96,16 +223,17 @@ static int cache_block_group(struct btrfs_root *root,
         * skip the locking here
         */
        path->skip_locking = 1;
-       first_free = block_group->key.objectid;
+       first_free = max_t(u64, block_group->key.objectid,
+                          BTRFS_SUPER_INFO_OFFSET + BTRFS_SUPER_INFO_SIZE);
        key.objectid = block_group->key.objectid;
        key.offset = 0;
        btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
        ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
        if (ret < 0)
-               return ret;
+               goto err;
        ret = btrfs_previous_item(root, path, 0, BTRFS_EXTENT_ITEM_KEY);
        if (ret < 0)
-               return ret;
+               goto err;
        if (ret == 0) {
                leaf = path->nodes[0];
                btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
@@ -119,32 +247,28 @@ static int cache_block_group(struct btrfs_root *root,
                        ret = btrfs_next_leaf(root, path);
                        if (ret < 0)
                                goto err;
-                       if (ret == 0) {
+                       if (ret == 0)
                                continue;
-                       } else {
+                       else
                                break;
-                       }
                }
                btrfs_item_key_to_cpu(leaf, &key, slot);
-               if (key.objectid < block_group->key.objectid) {
+               if (key.objectid < block_group->key.objectid)
                        goto next;
-               }
+
                if (key.objectid >= block_group->key.objectid +
-                   block_group->key.offset) {
+                   block_group->key.offset)
                        break;
-               }
 
                if (btrfs_key_type(&key) == BTRFS_EXTENT_ITEM_KEY) {
                        if (!found) {
                                last = first_free;
                                found = 1;
                        }
-                       if (key.objectid > last) {
-                               hole_size = key.objectid - last;
-                               set_extent_dirty(free_space_cache, last,
-                                                last + hole_size - 1,
-                                                GFP_NOFS);
-                       }
+
+                       add_new_free_space(block_group, root->fs_info, last,
+                                          key.objectid);
+
                        last = key.objectid + key.offset;
                }
 next:
@@ -153,179 +277,91 @@ next:
 
        if (!found)
                last = first_free;
-       if (block_group->key.objectid +
-           block_group->key.offset > last) {
-               hole_size = block_group->key.objectid +
-                       block_group->key.offset - last;
-               set_extent_dirty(free_space_cache, last,
-                                last + hole_size - 1, GFP_NOFS);
-       }
+
+       add_new_free_space(block_group, root->fs_info, last,
+                          block_group->key.objectid +
+                          block_group->key.offset);
+
        block_group->cached = 1;
+       ret = 0;
 err:
        btrfs_free_path(path);
-       return 0;
+       return ret;
 }
 
+/*
+ * return the block group that starts at or after bytenr
+ */
 struct btrfs_block_group_cache *btrfs_lookup_first_block_group(struct
                                                       btrfs_fs_info *info,
                                                         u64 bytenr)
 {
-       struct extent_io_tree *block_group_cache;
-       struct btrfs_block_group_cache *block_group = NULL;
-       u64 ptr;
-       u64 start;
-       u64 end;
-       int ret;
+       struct btrfs_block_group_cache *cache;
 
-       bytenr = max_t(u64, bytenr,
-                      BTRFS_SUPER_INFO_OFFSET + BTRFS_SUPER_INFO_SIZE);
-       block_group_cache = &info->block_group_cache;
-       ret = find_first_extent_bit(block_group_cache,
-                                   bytenr, &start, &end,
-                                   BLOCK_GROUP_DATA | BLOCK_GROUP_METADATA |
-                                   BLOCK_GROUP_SYSTEM);
-       if (ret) {
-               return NULL;
-       }
-       ret = get_state_private(block_group_cache, start, &ptr);
-       if (ret)
-               return NULL;
+       cache = block_group_cache_tree_search(info, bytenr, 0);
 
-       block_group = (struct btrfs_block_group_cache *)(unsigned long)ptr;
-       return block_group;
+       return cache;
 }
 
+/*
+ * return the block group that contains teh given bytenr
+ */
 struct btrfs_block_group_cache *btrfs_lookup_block_group(struct
                                                         btrfs_fs_info *info,
                                                         u64 bytenr)
 {
-       struct extent_io_tree *block_group_cache;
-       struct btrfs_block_group_cache *block_group = NULL;
-       u64 ptr;
-       u64 start;
-       u64 end;
-       int ret;
+       struct btrfs_block_group_cache *cache;
 
-       bytenr = max_t(u64, bytenr,
-                      BTRFS_SUPER_INFO_OFFSET + BTRFS_SUPER_INFO_SIZE);
-       block_group_cache = &info->block_group_cache;
-       ret = find_first_extent_bit(block_group_cache,
-                                   bytenr, &start, &end,
-                                   BLOCK_GROUP_DATA | BLOCK_GROUP_METADATA |
-                                   BLOCK_GROUP_SYSTEM);
-       if (ret) {
-               return NULL;
-       }
-       ret = get_state_private(block_group_cache, start, &ptr);
-       if (ret)
-               return NULL;
+       cache = block_group_cache_tree_search(info, bytenr, 1);
 
-       block_group = (struct btrfs_block_group_cache *)(unsigned long)ptr;
-       if (block_group->key.objectid <= bytenr && bytenr <
-           block_group->key.objectid + block_group->key.offset)
-               return block_group;
-       return NULL;
+       return cache;
 }
 
-static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
-{
-       return (cache->flags & bits) == bits;
-}
-
-static int noinline find_search_start(struct btrfs_root *root,
-                             struct btrfs_block_group_cache **cache_ret,
-                             u64 *start_ret, u64 num, int data)
+static int noinline find_free_space(struct btrfs_root *root,
+                                   struct btrfs_block_group_cache **cache_ret,
+                                   u64 *start_ret, u64 num, int data)
 {
        int ret;
        struct btrfs_block_group_cache *cache = *cache_ret;
-       struct extent_io_tree *free_space_cache;
-       struct extent_state *state;
+       struct btrfs_free_space *info = NULL;
        u64 last;
-       u64 start = 0;
-       u64 cache_miss = 0;
        u64 total_fs_bytes;
        u64 search_start = *start_ret;
-       int wrapped = 0;
 
        WARN_ON(!mutex_is_locked(&root->fs_info->alloc_mutex));
        total_fs_bytes = btrfs_super_total_bytes(&root->fs_info->super_copy);
-       free_space_cache = &root->fs_info->free_space_cache;
 
        if (!cache)
                goto out;
 
+       last = max(search_start, cache->key.objectid);
+
 again:
        ret = cache_block_group(root, cache);
-       if (ret) {
+       if (ret)
                goto out;
-       }
 
-       last = max(search_start, cache->key.objectid);
-       if (!block_group_bits(cache, data) || cache->ro)
+       if (cache->ro || !block_group_bits(cache, data))
                goto new_group;
 
-       spin_lock_irq(&free_space_cache->lock);
-       state = find_first_extent_bit_state(free_space_cache, last, EXTENT_DIRTY);
-       while(1) {
-               if (!state) {
-                       if (!cache_miss)
-                               cache_miss = last;
-                       spin_unlock_irq(&free_space_cache->lock);
-                       goto new_group;
-               }
-
-               start = max(last, state->start);
-               last = state->end + 1;
-               if (last - start < num) {
-                       do {
-                               state = extent_state_next(state);
-                       } while(state && !(state->state & EXTENT_DIRTY));
-                       continue;
-               }
-               spin_unlock_irq(&free_space_cache->lock);
-               if (cache->ro) {
-                       goto new_group;
-               }
-               if (start + num > cache->key.objectid + cache->key.offset)
-                       goto new_group;
-               if (!block_group_bits(cache, data)) {
-                       printk("block group bits don't match %Lu %d\n", cache->flags, data);
-               }
-               *start_ret = start;
+       info = btrfs_find_free_space(cache, last, num);
+       if (info) {
+               *start_ret = info->offset;
                return 0;
        }
-out:
-       cache = btrfs_lookup_block_group(root->fs_info, search_start);
-       if (!cache) {
-               printk("Unable to find block group for %Lu\n", search_start);
-               WARN_ON(1);
-       }
-       return -ENOSPC;
 
 new_group:
        last = cache->key.objectid + cache->key.offset;
-wrapped:
+
        cache = btrfs_lookup_first_block_group(root->fs_info, last);
-       if (!cache || cache->key.objectid >= total_fs_bytes) {
-no_cache:
-               if (!wrapped) {
-                       wrapped = 1;
-                       last = search_start;
-                       goto wrapped;
-               }
+       if (!cache || cache->key.objectid >= total_fs_bytes)
                goto out;
-       }
-       if (cache_miss && !cache->cached) {
-               cache_block_group(root, cache);
-               last = cache_miss;
-               cache = btrfs_lookup_first_block_group(root->fs_info, last);
-       }
-       cache_miss = 0;
-       cache = btrfs_find_block_group(root, cache, last, data, 0);
-       if (!cache)
-               goto no_cache;
+
        *cache_ret = cache;
        goto again;
+
+out:
+       return -ENOSPC;
 }
 
 static u64 div_factor(u64 num, int factor)
@@ -337,16 +373,19 @@ static u64 div_factor(u64 num, int factor)
        return num;
 }
 
-static int block_group_state_bits(u64 flags)
+static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
+                                                 u64 flags)
 {
-       int bits = 0;
-       if (flags & BTRFS_BLOCK_GROUP_DATA)
-               bits |= BLOCK_GROUP_DATA;
-       if (flags & BTRFS_BLOCK_GROUP_METADATA)
-               bits |= BLOCK_GROUP_METADATA;
-       if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
-               bits |= BLOCK_GROUP_SYSTEM;
-       return bits;
+       struct list_head *head = &info->space_info;
+       struct list_head *cur;
+       struct btrfs_space_info *found;
+       list_for_each(cur, head) {
+               found = list_entry(cur, struct btrfs_space_info, list);
+               if (found->flags == flags)
+                       return found;
+       }
+       return NULL;
+
 }
 
 static struct btrfs_block_group_cache *
@@ -355,28 +394,19 @@ __btrfs_find_block_group(struct btrfs_root *root,
                         u64 search_start, int data, int owner)
 {
        struct btrfs_block_group_cache *cache;
-       struct extent_io_tree *block_group_cache;
        struct btrfs_block_group_cache *found_group = NULL;
        struct btrfs_fs_info *info = root->fs_info;
+       struct btrfs_space_info *sinfo;
        u64 used;
        u64 last = 0;
-       u64 start;
-       u64 end;
        u64 free_check;
-       u64 ptr;
-       int bit;
-       int ret;
        int full_search = 0;
        int factor = 10;
        int wrapped = 0;
 
-       block_group_cache = &info->block_group_cache;
-
        if (data & BTRFS_BLOCK_GROUP_METADATA)
                factor = 9;
 
-       bit = block_group_state_bits(data);
-
        if (search_start) {
                struct btrfs_block_group_cache *shint;
                shint = btrfs_lookup_first_block_group(info, search_start);
@@ -407,20 +437,30 @@ __btrfs_find_block_group(struct btrfs_root *root,
                else
                        last = search_start;
        }
+       sinfo = __find_space_info(root->fs_info, data);
+       if (!sinfo)
+               goto found;
 again:
        while(1) {
-               ret = find_first_extent_bit(block_group_cache, last,
-                                           &start, &end, bit);
-               if (ret)
-                       break;
+               struct list_head *l;
 
-               ret = get_state_private(block_group_cache, start, &ptr);
-               if (ret) {
-                       last = end + 1;
-                       continue;
+               cache = NULL;
+
+               spin_lock(&sinfo->lock);
+               list_for_each(l, &sinfo->block_groups) {
+                       struct btrfs_block_group_cache *entry;
+                       entry = list_entry(l, struct btrfs_block_group_cache,
+                                          list);
+                       if ((entry->key.objectid >= last) &&
+                           (!cache || (entry->key.objectid <
+                                       cache->key.objectid)))
+                               cache = entry;
                }
+               spin_unlock(&sinfo->lock);
+
+               if (!cache)
+                       break;
 
-               cache = (struct btrfs_block_group_cache *)(unsigned long)ptr;
                spin_lock(&cache->lock);
                last = cache->key.objectid + cache->key.offset;
                used = btrfs_block_group_used(&cache->item);
@@ -461,48 +501,16 @@ struct btrfs_block_group_cache *btrfs_find_block_group(struct btrfs_root *root,
        ret = __btrfs_find_block_group(root, hint, search_start, data, owner);
        return ret;
 }
-static u64 hash_extent_ref(u64 root_objectid, u64 ref_generation,
-                          u64 owner, u64 owner_offset)
-{
-       u32 high_crc = ~(u32)0;
-       u32 low_crc = ~(u32)0;
-       __le64 lenum;
-       lenum = cpu_to_le64(root_objectid);
-       high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
-       lenum = cpu_to_le64(ref_generation);
-       low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
-       if (owner >= BTRFS_FIRST_FREE_OBJECTID) {
-               lenum = cpu_to_le64(owner);
-               low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
-               lenum = cpu_to_le64(owner_offset);
-               low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
-       }
-       return ((u64)high_crc << 32) | (u64)low_crc;
-}
-
-static int match_extent_ref(struct extent_buffer *leaf,
-                           struct btrfs_extent_ref *disk_ref,
-                           struct btrfs_extent_ref *cpu_ref)
-{
-       int ret;
-       int len;
-
-       if (cpu_ref->objectid)
-               len = sizeof(*cpu_ref);
-       else
-               len = 2 * sizeof(u64);
-       ret = memcmp_extent_buffer(leaf, cpu_ref, (unsigned long)disk_ref,
-                                  len);
-       return ret == 0;
-}
 
 /* simple helper to search for an existing extent at a given offset */
-int btrfs_lookup_extent(struct btrfs_root *root, struct btrfs_path *path,
-                       u64 start, u64 len)
+int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 {
        int ret;
        struct btrfs_key key;
+       struct btrfs_path *path;
 
+       path = btrfs_alloc_path();
+       BUG_ON(!path);
        maybe_lock_mutex(root);
        key.objectid = start;
        key.offset = len;
@@ -510,72 +518,7 @@ int btrfs_lookup_extent(struct btrfs_root *root, struct btrfs_path *path,
        ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
                                0, 0);
        maybe_unlock_mutex(root);
-       return ret;
-}
-
-static int noinline lookup_extent_backref(struct btrfs_trans_handle *trans,
-                                         struct btrfs_root *root,
-                                         struct btrfs_path *path, u64 bytenr,
-                                         u64 root_objectid,
-                                         u64 ref_generation, u64 owner,
-                                         u64 owner_offset, int del)
-{
-       u64 hash;
-       struct btrfs_key key;
-       struct btrfs_key found_key;
-       struct btrfs_extent_ref ref;
-       struct extent_buffer *leaf;
-       struct btrfs_extent_ref *disk_ref;
-       int ret;
-       int ret2;
-
-       btrfs_set_stack_ref_root(&ref, root_objectid);
-       btrfs_set_stack_ref_generation(&ref, ref_generation);
-       btrfs_set_stack_ref_objectid(&ref, owner);
-       btrfs_set_stack_ref_offset(&ref, owner_offset);
-
-       hash = hash_extent_ref(root_objectid, ref_generation, owner,
-                              owner_offset);
-       key.offset = hash;
-       key.objectid = bytenr;
-       key.type = BTRFS_EXTENT_REF_KEY;
-
-       while (1) {
-               ret = btrfs_search_slot(trans, root, &key, path,
-                                       del ? -1 : 0, del);
-               if (ret < 0)
-                       goto out;
-               leaf = path->nodes[0];
-               if (ret != 0) {
-                       u32 nritems = btrfs_header_nritems(leaf);
-                       if (path->slots[0] >= nritems) {
-                               ret2 = btrfs_next_leaf(root, path);
-                               if (ret2)
-                                       goto out;
-                               leaf = path->nodes[0];
-                       }
-                       btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
-                       if (found_key.objectid != bytenr ||
-                           found_key.type != BTRFS_EXTENT_REF_KEY)
-                               goto out;
-                       key.offset = found_key.offset;
-                       if (del) {
-                               btrfs_release_path(root, path);
-                               continue;
-                       }
-               }
-               disk_ref = btrfs_item_ptr(path->nodes[0],
-                                         path->slots[0],
-                                         struct btrfs_extent_ref);
-               if (match_extent_ref(path->nodes[0], disk_ref, &ref)) {
-                       ret = 0;
-                       goto out;
-               }
-               btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
-               key.offset = found_key.offset + 1;
-               btrfs_release_path(root, path);
-       }
-out:
+       btrfs_free_path(path);
        return ret;
 }
 
@@ -596,7 +539,7 @@ out:
  * File extents can be referenced by:
  *
  * - multiple snapshots, subvolumes, or different generations in one subvol
- * - different files inside a single subvolume (in theory, not implemented yet)
+ * - different files inside a single subvolume
  * - different offsets inside a file (bookend extents in file.c)
  *
  * The extent ref structure has fields for:
@@ -605,119 +548,284 @@ out:
  * - Generation number of the tree holding the reference
  * - objectid of the file holding the reference
  * - offset in the file corresponding to the key holding the reference
+ * - number of references holding by parent node (alway 1 for tree blocks)
+ *
+ * Btree leaf may hold multiple references to a file extent. In most cases,
+ * these references are from same file and the corresponding offsets inside
+ * the file are close together. So inode objectid and offset in file are
+ * just hints, they provide hints about where in the btree the references
+ * can be found and when we can stop searching.
  *
  * When a file extent is allocated the fields are filled in:
- *     (root_key.objectid, trans->transid, inode objectid, offset in file)
+ *     (root_key.objectid, trans->transid, inode objectid, offset in file, 1)
  *
  * When a leaf is cow'd new references are added for every file extent found
- * in the leaf.  It looks the same as the create case, but trans->transid
- * will be different when the block is cow'd.
+ * in the leaf.  It looks similar to the create case, but trans->transid will
+ * be different when the block is cow'd.
  *
- *     (root_key.objectid, trans->transid, inode objectid, offset in file)
+ *     (root_key.objectid, trans->transid, inode objectid, offset in file,
+ *      number of references in the leaf)
  *
- * When a file extent is removed either during snapshot deletion or file
- * truncation, the corresponding back reference is found
- * by searching for:
+ * Because inode objectid and offset in file are just hints, they are not
+ * used when backrefs are deleted. When a file extent is removed either
+ * during snapshot deletion or file truncation, we find the corresponding
+ * back back reference and check the following fields.
  *
- *     (btrfs_header_owner(leaf), btrfs_header_generation(leaf),
- *      inode objectid, offset in file)
+ *     (btrfs_header_owner(leaf), btrfs_header_generation(leaf))
  *
  * Btree extents can be referenced by:
  *
  * - Different subvolumes
  * - Different generations of the same subvolume
  *
- * Storing sufficient information for a full reverse mapping of a btree
- * block would require storing the lowest key of the block in the backref,
- * and it would require updating that lowest key either before write out or
- * every time it changed.  Instead, the objectid of the lowest key is stored
- * along with the level of the tree block.  This provides a hint
- * about where in the btree the block can be found.  Searches through the
- * btree only need to look for a pointer to that block, so they stop one
- * level higher than the level recorded in the backref.
- *
- * Some btrees do not do reference counting on their extents.  These
- * include the extent tree and the tree of tree roots.  Backrefs for these
- * trees always have a generation of zero.
- *
  * When a tree block is created, back references are inserted:
  *
- * (root->root_key.objectid, trans->transid or zero, level, lowest_key_objectid)
+ * (root->root_key.objectid, trans->transid, level, 0, 1)
  *
- * When a tree block is cow'd in a reference counted root,
- * new back references are added for all the blocks it points to.
- * These are of the form (trans->transid will have increased since creation):
+ * When a tree block is cow'd, new back references are added for all the
+ * blocks it points to. If the tree block isn't in reference counted root,
+ * the old back references are removed. These new back references are of
+ * the form (trans->transid will have increased since creation):
  *
- * (root->root_key.objectid, trans->transid, level, lowest_key_objectid)
+ * (root->root_key.objectid, trans->transid, level, 0, 1)
  *
- * Because the lowest_key_objectid and the level are just hints
- * they are not used when backrefs are deleted.  When a backref is deleted:
+ * When a backref is in deleting, the following fields are checked:
  *
  * if backref was for a tree root:
- *     root_objectid = root->root_key.objectid
+ *     (btrfs_header_owner(itself), btrfs_header_generation(itself))
  * else
- *     root_objectid = btrfs_header_owner(parent)
+ *     (btrfs_header_owner(parent), btrfs_header_generation(parent))
  *
- * (root_objectid, btrfs_header_generation(parent) or zero, 0, 0)
+ * Back Reference Key composing:
  *
- * Back Reference Key hashing:
- *
- * Back references have four fields, each 64 bits long.  Unfortunately,
- * This is hashed into a single 64 bit number and placed into the key offset.
- * The key objectid corresponds to the first byte in the extent, and the
- * key type is set to BTRFS_EXTENT_REF_KEY
+ * The key objectid corresponds to the first byte in the extent, the key
+ * type is set to BTRFS_EXTENT_REF_KEY, and the key offset is the first
+ * byte of parent extent. If a extent is tree root, the key offset is set
+ * to the key objectid.
  */
-int btrfs_insert_extent_backref(struct btrfs_trans_handle *trans,
-                                struct btrfs_root *root,
-                                struct btrfs_path *path, u64 bytenr,
-                                u64 root_objectid, u64 ref_generation,
-                                u64 owner, u64 owner_offset)
+
+static int noinline lookup_extent_backref(struct btrfs_trans_handle *trans,
+                                         struct btrfs_root *root,
+                                         struct btrfs_path *path, u64 bytenr,
+                                         u64 parent, u64 ref_root,
+                                         u64 ref_generation, int del)
 {
-       u64 hash;
        struct btrfs_key key;
-       struct btrfs_extent_ref ref;
-       struct btrfs_extent_ref *disk_ref;
+       struct btrfs_extent_ref *ref;
+       struct extent_buffer *leaf;
        int ret;
 
-       btrfs_set_stack_ref_root(&ref, root_objectid);
-       btrfs_set_stack_ref_generation(&ref, ref_generation);
-       btrfs_set_stack_ref_objectid(&ref, owner);
-       btrfs_set_stack_ref_offset(&ref, owner_offset);
+       key.objectid = bytenr;
+       key.type = BTRFS_EXTENT_REF_KEY;
+       key.offset = parent;
+
+       ret = btrfs_search_slot(trans, root, &key, path, del ? -1 : 0, 1);
+       if (ret < 0)
+               goto out;
+       if (ret > 0) {
+               ret = -ENOENT;
+               goto out;
+       }
+
+       leaf = path->nodes[0];
+       ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_ref);
+       if (btrfs_ref_root(leaf, ref) != ref_root ||
+           btrfs_ref_generation(leaf, ref) != ref_generation) {
+               ret = -EIO;
+               WARN_ON(1);
+               goto out;
+       }
+       ret = 0;
+out:
+       return ret;
+}
+
+static int noinline insert_extent_backref(struct btrfs_trans_handle *trans,
+                                         struct btrfs_root *root,
+                                         struct btrfs_path *path,
+                                         u64 bytenr, u64 parent,
+                                         u64 ref_root, u64 ref_generation,
+                                         u64 owner_objectid, u64 owner_offset)
+{
+       struct btrfs_key key;
+       struct extent_buffer *leaf;
+       struct btrfs_extent_ref *ref;
+       u32 num_refs;
+       int ret;
 
-       hash = hash_extent_ref(root_objectid, ref_generation, owner,
-                              owner_offset);
-       key.offset = hash;
        key.objectid = bytenr;
        key.type = BTRFS_EXTENT_REF_KEY;
+       key.offset = parent;
 
-       ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(ref));
-       while (ret == -EEXIST) {
-               disk_ref = btrfs_item_ptr(path->nodes[0], path->slots[0],
-                                         struct btrfs_extent_ref);
-               if (match_extent_ref(path->nodes[0], disk_ref, &ref))
+       ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*ref));
+       if (ret == 0) {
+               leaf = path->nodes[0];
+               ref = btrfs_item_ptr(leaf, path->slots[0],
+                                    struct btrfs_extent_ref);
+               btrfs_set_ref_root(leaf, ref, ref_root);
+               btrfs_set_ref_generation(leaf, ref, ref_generation);
+               btrfs_set_ref_objectid(leaf, ref, owner_objectid);
+               btrfs_set_ref_offset(leaf, ref, owner_offset);
+               btrfs_set_ref_num_refs(leaf, ref, 1);
+       } else if (ret == -EEXIST) {
+               u64 existing_owner;
+               BUG_ON(owner_objectid < BTRFS_FIRST_FREE_OBJECTID);
+               leaf = path->nodes[0];
+               ref = btrfs_item_ptr(leaf, path->slots[0],
+                                    struct btrfs_extent_ref);
+               if (btrfs_ref_root(leaf, ref) != ref_root ||
+                   btrfs_ref_generation(leaf, ref) != ref_generation) {
+                       ret = -EIO;
+                       WARN_ON(1);
                        goto out;
-               key.offset++;
-               btrfs_release_path(root, path);
-               ret = btrfs_insert_empty_item(trans, root, path, &key,
-                                             sizeof(ref));
-       }
-       if (ret)
+               }
+
+               num_refs = btrfs_ref_num_refs(leaf, ref);
+               BUG_ON(num_refs == 0);
+               btrfs_set_ref_num_refs(leaf, ref, num_refs + 1);
+
+               existing_owner = btrfs_ref_objectid(leaf, ref);
+               if (existing_owner == owner_objectid &&
+                   btrfs_ref_offset(leaf, ref) > owner_offset) {
+                       btrfs_set_ref_offset(leaf, ref, owner_offset);
+               } else if (existing_owner != owner_objectid &&
+                          existing_owner != BTRFS_MULTIPLE_OBJECTIDS) {
+                       btrfs_set_ref_objectid(leaf, ref,
+                                       BTRFS_MULTIPLE_OBJECTIDS);
+                       btrfs_set_ref_offset(leaf, ref, 0);
+               }
+               ret = 0;
+       } else {
                goto out;
-       disk_ref = btrfs_item_ptr(path->nodes[0], path->slots[0],
-                                 struct btrfs_extent_ref);
-       write_extent_buffer(path->nodes[0], &ref, (unsigned long)disk_ref,
-                           sizeof(ref));
+       }
        btrfs_mark_buffer_dirty(path->nodes[0]);
 out:
        btrfs_release_path(root, path);
        return ret;
 }
 
+static int noinline remove_extent_backref(struct btrfs_trans_handle *trans,
+                                         struct btrfs_root *root,
+                                         struct btrfs_path *path)
+{
+       struct extent_buffer *leaf;
+       struct btrfs_extent_ref *ref;
+       u32 num_refs;
+       int ret = 0;
+
+       leaf = path->nodes[0];
+       ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_ref);
+       num_refs = btrfs_ref_num_refs(leaf, ref);
+       BUG_ON(num_refs == 0);
+       num_refs -= 1;
+       if (num_refs == 0) {
+               ret = btrfs_del_item(trans, root, path);
+       } else {
+               btrfs_set_ref_num_refs(leaf, ref, num_refs);
+               btrfs_mark_buffer_dirty(leaf);
+       }
+       btrfs_release_path(root, path);
+       return ret;
+}
+
+static int __btrfs_update_extent_ref(struct btrfs_trans_handle *trans,
+                                    struct btrfs_root *root, u64 bytenr,
+                                    u64 orig_parent, u64 parent,
+                                    u64 orig_root, u64 ref_root,
+                                    u64 orig_generation, u64 ref_generation,
+                                    u64 owner_objectid, u64 owner_offset)
+{
+       int ret;
+       struct btrfs_root *extent_root = root->fs_info->extent_root;
+       struct btrfs_path *path;
+
+       if (root == root->fs_info->extent_root) {
+               struct pending_extent_op *extent_op;
+               u64 num_bytes;
+
+               BUG_ON(owner_objectid >= BTRFS_MAX_LEVEL);
+               num_bytes = btrfs_level_size(root, (int)owner_objectid);
+               if (test_range_bit(&root->fs_info->extent_ins, bytenr,
+                               bytenr + num_bytes - 1, EXTENT_LOCKED, 0)) {
+                       u64 priv;
+                       ret = get_state_private(&root->fs_info->extent_ins,
+                                               bytenr, &priv);
+                       BUG_ON(ret);
+                       extent_op = (struct pending_extent_op *)
+                                                       (unsigned long)priv;
+                       BUG_ON(extent_op->parent != orig_parent);
+                       BUG_ON(extent_op->generation != orig_generation);
+                       extent_op->parent = parent;
+                       extent_op->generation = ref_generation;
+               } else {
+                       extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
+                       BUG_ON(!extent_op);
+
+                       extent_op->type = PENDING_BACKREF_UPDATE;
+                       extent_op->bytenr = bytenr;
+                       extent_op->num_bytes = num_bytes;
+                       extent_op->parent = parent;
+                       extent_op->orig_parent = orig_parent;
+                       extent_op->generation = ref_generation;
+                       extent_op->orig_generation = orig_generation;
+                       extent_op->level = (int)owner_objectid;
+
+                       set_extent_bits(&root->fs_info->extent_ins,
+                                       bytenr, bytenr + num_bytes - 1,
+                                       EXTENT_LOCKED, GFP_NOFS);
+                       set_state_private(&root->fs_info->extent_ins,
+                                         bytenr, (unsigned long)extent_op);
+               }
+               return 0;
+       }
+
+       path = btrfs_alloc_path();
+       if (!path)
+               return -ENOMEM;
+       ret = lookup_extent_backref(trans, extent_root, path,
+                                   bytenr, orig_parent, orig_root,
+                                   orig_generation, 1);
+       if (ret)
+               goto out;
+       ret = remove_extent_backref(trans, extent_root, path);
+       if (ret)
+               goto out;
+       ret = insert_extent_backref(trans, extent_root, path, bytenr,
+                                   parent, ref_root, ref_generation,
+                                   owner_objectid, owner_offset);
+       BUG_ON(ret);
+       finish_current_insert(trans, extent_root);
+       del_pending_extents(trans, extent_root);
+out:
+       btrfs_free_path(path);
+       return ret;
+}
+
+int btrfs_update_extent_ref(struct btrfs_trans_handle *trans,
+                           struct btrfs_root *root, u64 bytenr,
+                           u64 orig_parent, u64 parent,
+                           u64 ref_root, u64 ref_generation,
+                           u64 owner_objectid, u64 owner_offset)
+{
+       int ret;
+       if (ref_root == BTRFS_TREE_LOG_OBJECTID &&
+           owner_objectid < BTRFS_FIRST_FREE_OBJECTID)
+               return 0;
+       maybe_lock_mutex(root);
+       ret = __btrfs_update_extent_ref(trans, root, bytenr, orig_parent,
+                                       parent, ref_root, ref_root,
+                                       ref_generation, ref_generation,
+                                       owner_objectid, owner_offset);
+       maybe_unlock_mutex(root);
+       return ret;
+}
+
 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
-                               struct btrfs_root *root,
-                               u64 bytenr, u64 num_bytes,
-                               u64 root_objectid, u64 ref_generation,
-                               u64 owner, u64 owner_offset)
+                                 struct btrfs_root *root, u64 bytenr,
+                                 u64 orig_parent, u64 parent,
+                                 u64 orig_root, u64 ref_root,
+                                 u64 orig_generation, u64 ref_generation,
+                                 u64 owner_objectid, u64 owner_offset)
 {
        struct btrfs_path *path;
        int ret;
@@ -726,24 +834,28 @@ static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
        struct btrfs_extent_item *item;
        u32 refs;
 
-       WARN_ON(num_bytes < root->sectorsize);
        path = btrfs_alloc_path();
        if (!path)
                return -ENOMEM;
 
        path->reada = 1;
        key.objectid = bytenr;
-       btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
-       key.offset = num_bytes;
+       key.type = BTRFS_EXTENT_ITEM_KEY;
+       key.offset = (u64)-1;
+
        ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key, path,
                                0, 1);
        if (ret < 0)
                return ret;
-       if (ret != 0) {
-               BUG();
-       }
-       BUG_ON(ret != 0);
+       BUG_ON(ret == 0 || path->slots[0] == 0);
+
+       path->slots[0]--;
        l = path->nodes[0];
+
+       btrfs_item_key_to_cpu(l, &key, path->slots[0]);
+       BUG_ON(key.objectid != bytenr);
+       BUG_ON(key.type != BTRFS_EXTENT_ITEM_KEY);
+
        item = btrfs_item_ptr(l, path->slots[0], struct btrfs_extent_item);
        refs = btrfs_extent_refs(l, item);
        btrfs_set_extent_refs(l, item, refs + 1);
@@ -752,9 +864,10 @@ static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
        btrfs_release_path(root->fs_info->extent_root, path);
 
        path->reada = 1;
-       ret = btrfs_insert_extent_backref(trans, root->fs_info->extent_root,
-                                         path, bytenr, root_objectid,
-                                         ref_generation, owner, owner_offset);
+       ret = insert_extent_backref(trans, root->fs_info->extent_root,
+                                   path, bytenr, parent,
+                                   ref_root, ref_generation,
+                                   owner_objectid, owner_offset);
        BUG_ON(ret);
        finish_current_insert(trans, root->fs_info->extent_root);
        del_pending_extents(trans, root->fs_info->extent_root);
@@ -764,18 +877,20 @@ static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
 }
 
 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
-                               struct btrfs_root *root,
-                               u64 bytenr, u64 num_bytes,
-                               u64 root_objectid, u64 ref_generation,
-                               u64 owner, u64 owner_offset)
+                        struct btrfs_root *root,
+                        u64 bytenr, u64 num_bytes, u64 parent,
+                        u64 ref_root, u64 ref_generation,
+                        u64 owner_objectid, u64 owner_offset)
 {
        int ret;
-
-       mutex_lock(&root->fs_info->alloc_mutex);
-       ret = __btrfs_inc_extent_ref(trans, root, bytenr, num_bytes,
-                                    root_objectid, ref_generation,
-                                    owner, owner_offset);
-       mutex_unlock(&root->fs_info->alloc_mutex);
+       if (ref_root == BTRFS_TREE_LOG_OBJECTID &&
+           owner_objectid < BTRFS_FIRST_FREE_OBJECTID)
+               return 0;
+       maybe_lock_mutex(root);
+       ret = __btrfs_inc_extent_ref(trans, root, bytenr, 0, parent,
+                                    0, ref_root, 0, ref_generation,
+                                    owner_objectid, owner_offset);
+       maybe_unlock_mutex(root);
        return ret;
 }
 
@@ -787,9 +902,9 @@ int btrfs_extent_post_op(struct btrfs_trans_handle *trans,
        return 0;
 }
 
-static int lookup_extent_ref(struct btrfs_trans_handle *trans,
-                            struct btrfs_root *root, u64 bytenr,
-                            u64 num_bytes, u32 *refs)
+int btrfs_lookup_extent_ref(struct btrfs_trans_handle *trans,
+                           struct btrfs_root *root, u64 bytenr,
+                           u64 num_bytes, u32 *refs)
 {
        struct btrfs_path *path;
        int ret;
@@ -820,7 +935,6 @@ out:
        return 0;
 }
 
-
 static int get_reference_status(struct btrfs_root *root, u64 bytenr,
                                u64 parent_gen, u64 ref_objectid,
                                u64 *min_generation, u32 *ref_count)
@@ -837,7 +951,7 @@ static int get_reference_status(struct btrfs_root *root, u64 bytenr,
        int ret;
 
        key.objectid = bytenr;
-       key.offset = 0;
+       key.offset = (u64)-1;
        key.type = BTRFS_EXTENT_ITEM_KEY;
 
        path = btrfs_alloc_path();
@@ -846,7 +960,10 @@ static int get_reference_status(struct btrfs_root *root, u64 bytenr,
        if (ret < 0)
                goto out;
        BUG_ON(ret == 0);
+       if (ret < 0 || path->slots[0] == 0)
+               goto out;
 
+       path->slots[0]--;
        leaf = path->nodes[0];
        btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
 
@@ -883,7 +1000,7 @@ static int get_reference_status(struct btrfs_root *root, u64 bytenr,
                                          struct btrfs_extent_ref);
                ref_generation = btrfs_ref_generation(leaf, ref_item);
                /*
-                * For (parent_gen > 0 && parent_gen > ref_gen):
+                * For (parent_gen > 0 && parent_gen > ref_generation):
                 *
                 * we reach here through the oldest root, therefore
                 * all other reference from same snapshot should have
@@ -893,8 +1010,7 @@ static int get_reference_status(struct btrfs_root *root, u64 bytenr,
                    (parent_gen > 0 && parent_gen > ref_generation) ||
                    (ref_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
                     ref_objectid != btrfs_ref_objectid(leaf, ref_item))) {
-                       if (ref_count)
-                               *ref_count = 2;
+                       *ref_count = 2;
                        break;
                }
 
@@ -994,80 +1110,29 @@ out:
        return ret;
 }
 
-int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
-                 struct extent_buffer *buf, int cache_ref)
+int btrfs_cache_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
+                   struct extent_buffer *buf, u32 nr_extents)
 {
-       u64 bytenr;
        u32 nritems;
        struct btrfs_key key;
        struct btrfs_file_extent_item *fi;
        int i;
        int level;
-       int ret;
-       int faili;
-       int nr_file_extents = 0;
+       int ret = 0;
 
        if (!root->ref_cows)
                return 0;
 
        level = btrfs_header_level(buf);
        nritems = btrfs_header_nritems(buf);
-       for (i = 0; i < nritems; i++) {
-               cond_resched();
-               if (level == 0) {
-                       u64 disk_bytenr;
-                       btrfs_item_key_to_cpu(buf, &key, i);
-                       if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
-                               continue;
-                       fi = btrfs_item_ptr(buf, i,
-                                           struct btrfs_file_extent_item);
-                       if (btrfs_file_extent_type(buf, fi) ==
-                           BTRFS_FILE_EXTENT_INLINE)
-                               continue;
-                       disk_bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
-                       if (disk_bytenr == 0)
-                               continue;
-
-                       if (buf != root->commit_root)
-                               nr_file_extents++;
-
-                       mutex_lock(&root->fs_info->alloc_mutex);
-                       ret = __btrfs_inc_extent_ref(trans, root, disk_bytenr,
-                                   btrfs_file_extent_disk_num_bytes(buf, fi),
-                                   root->root_key.objectid, trans->transid,
-                                   key.objectid, key.offset);
-                       mutex_unlock(&root->fs_info->alloc_mutex);
-                       if (ret) {
-                               faili = i;
-                               WARN_ON(1);
-                               goto fail;
-                       }
-               } else {
-                       bytenr = btrfs_node_blockptr(buf, i);
-                       btrfs_node_key_to_cpu(buf, &key, i);
 
-                       mutex_lock(&root->fs_info->alloc_mutex);
-                       ret = __btrfs_inc_extent_ref(trans, root, bytenr,
-                                          btrfs_level_size(root, level - 1),
-                                          root->root_key.objectid,
-                                          trans->transid,
-                                          level - 1, key.objectid);
-                       mutex_unlock(&root->fs_info->alloc_mutex);
-                       if (ret) {
-                               faili = i;
-                               WARN_ON(1);
-                               goto fail;
-                       }
-               }
-       }
-       /* cache orignal leaf block's references */
-       if (level == 0 && cache_ref && buf != root->commit_root) {
+       if (level == 0) {
                struct btrfs_leaf_ref *ref;
                struct btrfs_extent_info *info;
 
-               ref = btrfs_alloc_leaf_ref(root, nr_file_extents);
+               ref = btrfs_alloc_leaf_ref(root, nr_extents);
                if (!ref) {
-                       WARN_ON(1);
+                       ret = -ENOMEM;
                        goto out;
                }
 
@@ -1075,10 +1140,10 @@ int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
                ref->bytenr = buf->start;
                ref->owner = btrfs_header_owner(buf);
                ref->generation = btrfs_header_generation(buf);
-               ref->nritems = nr_file_extents;
+               ref->nritems = nr_extents;
                info = ref->extents;
 
-               for (i = 0; nr_file_extents > 0 && i < nritems; i++) {
+               for (i = 0; nr_extents > 0 && i < nritems; i++) {
                        u64 disk_bytenr;
                        btrfs_item_key_to_cpu(buf, &key, i);
                        if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
@@ -1106,13 +1171,52 @@ int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
                btrfs_free_leaf_ref(root, ref);
        }
 out:
-       return 0;
-fail:
-       WARN_ON(1);
-#if 0
-       for (i =0; i < faili; i++) {
+       return ret;
+}
+
+int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
+                 struct extent_buffer *orig_buf, struct extent_buffer *buf,
+                 u32 *nr_extents)
+{
+       u64 bytenr;
+       u64 ref_root;
+       u64 orig_root;
+       u64 ref_generation;
+       u64 orig_generation;
+       u32 nritems;
+       u32 nr_file_extents = 0;
+       struct btrfs_key key;
+       struct btrfs_file_extent_item *fi;
+       int i;
+       int level;
+       int ret = 0;
+       int faili = 0;
+       int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
+                           u64, u64, u64, u64, u64, u64, u64, u64, u64);
+
+       ref_root = btrfs_header_owner(buf);
+       ref_generation = btrfs_header_generation(buf);
+       orig_root = btrfs_header_owner(orig_buf);
+       orig_generation = btrfs_header_generation(orig_buf);
+
+       nritems = btrfs_header_nritems(buf);
+       level = btrfs_header_level(buf);
+
+       if (root->ref_cows) {
+               process_func = __btrfs_inc_extent_ref;
+       } else {
+               if (level == 0 &&
+                   root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
+                       goto out;
+               if (level != 0 &&
+                   root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID)
+                       goto out;
+               process_func = __btrfs_update_extent_ref;
+       }
+
+       for (i = 0; i < nritems; i++) {
+               cond_resched();
                if (level == 0) {
-                       u64 disk_bytenr;
                        btrfs_item_key_to_cpu(buf, &key, i);
                        if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
                                continue;
@@ -1121,24 +1225,131 @@ fail:
                        if (btrfs_file_extent_type(buf, fi) ==
                            BTRFS_FILE_EXTENT_INLINE)
                                continue;
-                       disk_bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
-                       if (disk_bytenr == 0)
+                       bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
+                       if (bytenr == 0)
                                continue;
-                       err = btrfs_free_extent(trans, root, disk_bytenr,
-                                   btrfs_file_extent_disk_num_bytes(buf,
-                                                                     fi), 0);
-                       BUG_ON(err);
+
+                       nr_file_extents++;
+
+                       maybe_lock_mutex(root);
+                       ret = process_func(trans, root, bytenr,
+                                          orig_buf->start, buf->start,
+                                          orig_root, ref_root,
+                                          orig_generation, ref_generation,
+                                          key.objectid, key.offset);
+                       maybe_unlock_mutex(root);
+
+                       if (ret) {
+                               faili = i;
+                               WARN_ON(1);
+                               goto fail;
+                       }
                } else {
                        bytenr = btrfs_node_blockptr(buf, i);
-                       err = btrfs_free_extent(trans, root, bytenr,
-                                       btrfs_level_size(root, level - 1), 0);
-                       BUG_ON(err);
+                       maybe_lock_mutex(root);
+                       ret = process_func(trans, root, bytenr,
+                                          orig_buf->start, buf->start,
+                                          orig_root, ref_root,
+                                          orig_generation, ref_generation,
+                                          level - 1, 0);
+                       maybe_unlock_mutex(root);
+                       if (ret) {
+                               faili = i;
+                               WARN_ON(1);
+                               goto fail;
+                       }
                }
        }
-#endif
+out:
+       if (nr_extents) {
+               if (level == 0)
+                       *nr_extents = nr_file_extents;
+               else
+                       *nr_extents = nritems;
+       }
+       return 0;
+fail:
+       WARN_ON(1);
        return ret;
 }
 
+int btrfs_update_ref(struct btrfs_trans_handle *trans,
+                    struct btrfs_root *root, struct extent_buffer *orig_buf,
+                    struct extent_buffer *buf, int start_slot, int nr)
+
+{
+       u64 bytenr;
+       u64 ref_root;
+       u64 orig_root;
+       u64 ref_generation;
+       u64 orig_generation;
+       struct btrfs_key key;
+       struct btrfs_file_extent_item *fi;
+       int i;
+       int ret;
+       int slot;
+       int level;
+
+       BUG_ON(start_slot < 0);
+       BUG_ON(start_slot + nr > btrfs_header_nritems(buf));
+
+       ref_root = btrfs_header_owner(buf);
+       ref_generation = btrfs_header_generation(buf);
+       orig_root = btrfs_header_owner(orig_buf);
+       orig_generation = btrfs_header_generation(orig_buf);
+       level = btrfs_header_level(buf);
+
+       if (!root->ref_cows) {
+               if (level == 0 &&
+                   root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
+                       return 0;
+               if (level != 0 &&
+                   root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID)
+                       return 0;
+       }
+
+       for (i = 0, slot = start_slot; i < nr; i++, slot++) {
+               cond_resched();
+               if (level == 0) {
+                       btrfs_item_key_to_cpu(buf, &key, slot);
+                       if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
+                               continue;
+                       fi = btrfs_item_ptr(buf, slot,
+                                           struct btrfs_file_extent_item);
+                       if (btrfs_file_extent_type(buf, fi) ==
+                           BTRFS_FILE_EXTENT_INLINE)
+                               continue;
+                       bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
+                       if (bytenr == 0)
+                               continue;
+                       maybe_lock_mutex(root);
+                       ret = __btrfs_update_extent_ref(trans, root, bytenr,
+                                           orig_buf->start, buf->start,
+                                           orig_root, ref_root,
+                                           orig_generation, ref_generation,
+                                           key.objectid, key.offset);
+                       maybe_unlock_mutex(root);
+                       if (ret)
+                               goto fail;
+               } else {
+                       bytenr = btrfs_node_blockptr(buf, slot);
+                       maybe_lock_mutex(root);
+                       ret = __btrfs_update_extent_ref(trans, root, bytenr,
+                                           orig_buf->start, buf->start,
+                                           orig_root, ref_root,
+                                           orig_generation, ref_generation,
+                                           level - 1, 0);
+                       maybe_unlock_mutex(root);
+                       if (ret)
+                               goto fail;
+               }
+       }
+       return 0;
+fail:
+       WARN_ON(1);
+       return -1;
+}
+
 static int write_one_cache_group(struct btrfs_trans_handle *trans,
                                 struct btrfs_root *root,
                                 struct btrfs_path *path,
@@ -1174,34 +1385,37 @@ fail:
 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
                                   struct btrfs_root *root)
 {
-       struct extent_io_tree *block_group_cache;
-       struct btrfs_block_group_cache *cache;
-       int ret;
+       struct btrfs_block_group_cache *cache, *entry;
+       struct rb_node *n;
        int err = 0;
        int werr = 0;
        struct btrfs_path *path;
        u64 last = 0;
-       u64 start;
-       u64 end;
-       u64 ptr;
 
-       block_group_cache = &root->fs_info->block_group_cache;
        path = btrfs_alloc_path();
        if (!path)
                return -ENOMEM;
 
        mutex_lock(&root->fs_info->alloc_mutex);
        while(1) {
-               ret = find_first_extent_bit(block_group_cache, last,
-                                           &start, &end, BLOCK_GROUP_DIRTY);
-               if (ret)
-                       break;
+               cache = NULL;
+               spin_lock(&root->fs_info->block_group_cache_lock);
+               for (n = rb_first(&root->fs_info->block_group_cache_tree);
+                    n; n = rb_next(n)) {
+                       entry = rb_entry(n, struct btrfs_block_group_cache,
+                                        cache_node);
+                       if (entry->dirty) {
+                               cache = entry;
+                               break;
+                       }
+               }
+               spin_unlock(&root->fs_info->block_group_cache_lock);
 
-               last = end + 1;
-               ret = get_state_private(block_group_cache, start, &ptr);
-               if (ret)
+               if (!cache)
                        break;
-               cache = (struct btrfs_block_group_cache *)(unsigned long)ptr;
+
+               last += cache->key.offset;
+
                err = write_one_cache_group(trans, root,
                                            path, cache);
                /*
@@ -1213,29 +1427,14 @@ int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
                        werr = err;
                        continue;
                }
-               clear_extent_bits(block_group_cache, start, end,
-                                 BLOCK_GROUP_DIRTY, GFP_NOFS);
+
+               cache->dirty = 0;
        }
        btrfs_free_path(path);
        mutex_unlock(&root->fs_info->alloc_mutex);
        return werr;
 }
 
-static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
-                                                 u64 flags)
-{
-       struct list_head *head = &info->space_info;
-       struct list_head *cur;
-       struct btrfs_space_info *found;
-       list_for_each(cur, head) {
-               found = list_entry(cur, struct btrfs_space_info, list);
-               if (found->flags == flags)
-                       return found;
-       }
-       return NULL;
-
-}
-
 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
                             u64 total_bytes, u64 bytes_used,
                             struct btrfs_space_info **space_info)
@@ -1255,6 +1454,8 @@ static int update_space_info(struct btrfs_fs_info *info, u64 flags,
                return -ENOMEM;
 
        list_add(&found->list, &info->space_info);
+       INIT_LIST_HEAD(&found->block_groups);
+       spin_lock_init(&found->lock);
        found->flags = flags;
        found->total_bytes = total_bytes;
        found->bytes_used = bytes_used;
@@ -1317,7 +1518,7 @@ static int do_chunk_alloc(struct btrfs_trans_handle *trans,
        u64 thresh;
        u64 start;
        u64 num_bytes;
-       int ret;
+       int ret = 0;
 
        flags = reduce_alloc_profile(extent_root, flags);
 
@@ -1354,10 +1555,11 @@ printk("space info full %Lu\n", flags);
        ret = btrfs_make_block_group(trans, extent_root, 0, flags,
                     BTRFS_FIRST_CHUNK_TREE_OBJECTID, start, num_bytes);
        BUG_ON(ret);
+
 out_unlock:
        mutex_unlock(&extent_root->fs_info->chunk_mutex);
 out:
-       return 0;
+       return ret;
 }
 
 static int update_block_group(struct btrfs_trans_handle *trans,
@@ -1370,8 +1572,6 @@ static int update_block_group(struct btrfs_trans_handle *trans,
        u64 total = num_bytes;
        u64 old_val;
        u64 byte_in_group;
-       u64 start;
-       u64 end;
 
        WARN_ON(!mutex_is_locked(&root->fs_info->alloc_mutex));
        while(total) {
@@ -1381,12 +1581,9 @@ static int update_block_group(struct btrfs_trans_handle *trans,
                }
                byte_in_group = bytenr - cache->key.objectid;
                WARN_ON(byte_in_group > cache->key.offset);
-               start = cache->key.objectid;
-               end = start + cache->key.offset - 1;
-               set_extent_bits(&info->block_group_cache, start, end,
-                               BLOCK_GROUP_DIRTY, GFP_NOFS);
 
                spin_lock(&cache->lock);
+               cache->dirty = 1;
                old_val = btrfs_block_group_used(&cache->item);
                num_bytes = min(total, cache->key.offset - byte_in_group);
                if (alloc) {
@@ -1400,9 +1597,11 @@ static int update_block_group(struct btrfs_trans_handle *trans,
                        btrfs_set_block_group_used(&cache->item, old_val);
                        spin_unlock(&cache->lock);
                        if (mark_free) {
-                               set_extent_dirty(&info->free_space_cache,
-                                                bytenr, bytenr + num_bytes - 1,
-                                                GFP_NOFS);
+                               int ret;
+                               ret = btrfs_add_free_space(cache, bytenr,
+                                                          num_bytes);
+                               if (ret)
+                                       return -1;
                        }
                }
                total -= num_bytes;
@@ -1413,16 +1612,13 @@ static int update_block_group(struct btrfs_trans_handle *trans,
 
 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
 {
-       u64 start;
-       u64 end;
-       int ret;
-       ret = find_first_extent_bit(&root->fs_info->block_group_cache,
-                                   search_start, &start, &end,
-                                   BLOCK_GROUP_DATA | BLOCK_GROUP_METADATA |
-                                   BLOCK_GROUP_SYSTEM);
-       if (ret)
+       struct btrfs_block_group_cache *cache;
+
+       cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
+       if (!cache)
                return 0;
-       return start;
+
+       return cache->key.objectid;
 }
 
 
@@ -1500,8 +1696,7 @@ int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
        u64 start;
        u64 end;
        int ret;
-       struct extent_io_tree *free_space_cache;
-       free_space_cache = &root->fs_info->free_space_cache;
+       struct btrfs_block_group_cache *cache;
 
        mutex_lock(&root->fs_info->alloc_mutex);
        while(1) {
@@ -1511,7 +1706,9 @@ int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
                        break;
                btrfs_update_pinned_extents(root, start, end + 1 - start, 0);
                clear_extent_dirty(unpin, start, end, GFP_NOFS);
-               set_extent_dirty(free_space_cache, start, end, GFP_NOFS);
+               cache = btrfs_lookup_block_group(root->fs_info, start);
+               if (cache->cached)
+                       btrfs_add_free_space(cache, start, end - start + 1);
                if (need_resched()) {
                        mutex_unlock(&root->fs_info->alloc_mutex);
                        cond_resched();
@@ -1527,19 +1724,18 @@ static int finish_current_insert(struct btrfs_trans_handle *trans,
 {
        u64 start;
        u64 end;
+       u64 priv;
        struct btrfs_fs_info *info = extent_root->fs_info;
-       struct extent_buffer *eb;
        struct btrfs_path *path;
-       struct btrfs_key ins;
-       struct btrfs_disk_key first;
+       struct btrfs_extent_ref *ref;
+       struct pending_extent_op *extent_op;
+       struct btrfs_key key;
        struct btrfs_extent_item extent_item;
        int ret;
-       int level;
        int err = 0;
 
        WARN_ON(!mutex_is_locked(&extent_root->fs_info->alloc_mutex));
        btrfs_set_stack_extent_refs(&extent_item, 1);
-       btrfs_set_key_type(&ins, BTRFS_EXTENT_ITEM_KEY);
        path = btrfs_alloc_path();
 
        while(1) {
@@ -1548,37 +1744,54 @@ static int finish_current_insert(struct btrfs_trans_handle *trans,
                if (ret)
                        break;
 
-               ins.objectid = start;
-               ins.offset = end + 1 - start;
-               err = btrfs_insert_item(trans, extent_root, &ins,
+               ret = get_state_private(&info->extent_ins, start, &priv);
+               BUG_ON(ret);
+               extent_op = (struct pending_extent_op *)(unsigned long)priv;
+
+               if (extent_op->type == PENDING_EXTENT_INSERT) {
+                       key.objectid = start;
+                       key.offset = end + 1 - start;
+                       key.type = BTRFS_EXTENT_ITEM_KEY;
+                       err = btrfs_insert_item(trans, extent_root, &key,
                                        &extent_item, sizeof(extent_item));
-               clear_extent_bits(&info->extent_ins, start, end, EXTENT_LOCKED,
-                                 GFP_NOFS);
+                       BUG_ON(err);
 
-               eb = btrfs_find_create_tree_block(extent_root, ins.objectid,
-                                          ins.offset);
+                       clear_extent_bits(&info->extent_ins, start, end,
+                                         EXTENT_LOCKED, GFP_NOFS);
 
-               if (!btrfs_buffer_uptodate(eb, trans->transid))
-                       btrfs_read_buffer(eb, trans->transid);
+                       err = insert_extent_backref(trans, extent_root, path,
+                                               start, extent_op->parent,
+                                               extent_root->root_key.objectid,
+                                               extent_op->generation,
+                                               extent_op->level, 0);
+                       BUG_ON(err);
+               } else if (extent_op->type == PENDING_BACKREF_UPDATE) {
+                       err = lookup_extent_backref(trans, extent_root, path,
+                                               start, extent_op->orig_parent,
+                                               extent_root->root_key.objectid,
+                                               extent_op->orig_generation, 0);
+                       BUG_ON(err);
 
-               btrfs_tree_lock(eb);
-               level = btrfs_header_level(eb);
-               if (level == 0) {
-                       btrfs_item_key(eb, &first, 0);
+                       clear_extent_bits(&info->extent_ins, start, end,
+                                         EXTENT_LOCKED, GFP_NOFS);
+
+                       key.objectid = start;
+                       key.offset = extent_op->parent;
+                       key.type = BTRFS_EXTENT_REF_KEY;
+                       err = btrfs_set_item_key_safe(trans, extent_root, path,
+                                                     &key);
+                       BUG_ON(err);
+                       ref = btrfs_item_ptr(path->nodes[0], path->slots[0],
+                                            struct btrfs_extent_ref);
+                       btrfs_set_ref_generation(path->nodes[0], ref,
+                                                extent_op->generation);
+                       btrfs_mark_buffer_dirty(path->nodes[0]);
+                       btrfs_release_path(extent_root, path);
                } else {
-                       btrfs_node_key(eb, &first, 0);
+                       BUG_ON(1);
                }
-               btrfs_tree_unlock(eb);
-               free_extent_buffer(eb);
-               /*
-                * the first key is just a hint, so the race we've created
-                * against reading it is fine
-                */
-               err = btrfs_insert_extent_backref(trans, extent_root, path,
-                                         start, extent_root->root_key.objectid,
-                                         0, level,
-                                         btrfs_disk_key_objectid(&first));
-               BUG_ON(err);
+               kfree(extent_op);
+
                if (need_resched()) {
                        mutex_unlock(&extent_root->fs_info->alloc_mutex);
                        cond_resched();
@@ -1589,52 +1802,44 @@ static int finish_current_insert(struct btrfs_trans_handle *trans,
        return 0;
 }
 
-static int pin_down_bytes(struct btrfs_root *root, u64 bytenr, u32 num_bytes,
-                         int is_data, int pending)
+static int pin_down_bytes(struct btrfs_trans_handle *trans,
+                         struct btrfs_root *root,
+                         u64 bytenr, u64 num_bytes, int is_data)
 {
        int err = 0;
+       struct extent_buffer *buf;
 
        WARN_ON(!mutex_is_locked(&root->fs_info->alloc_mutex));
-       if (!pending) {
-               struct extent_buffer *buf;
-
-               if (is_data)
-                       goto pinit;
-
-               buf = btrfs_find_tree_block(root, bytenr, num_bytes);
-               if (buf) {
-                       /* we can reuse a block if it hasn't been written
-                        * and it is from this transaction.  We can't
-                        * reuse anything from the tree log root because
-                        * it has tiny sub-transactions.
-                        */
-                       if (btrfs_buffer_uptodate(buf, 0) &&
-                           btrfs_try_tree_lock(buf)) {
-                               u64 transid =
-                                   root->fs_info->running_transaction->transid;
-                               u64 header_transid =
-                                       btrfs_header_generation(buf);
-                               if (btrfs_header_owner(buf) !=
-                                   BTRFS_TREE_LOG_OBJECTID &&
-                                   header_transid == transid &&
-                                   !btrfs_header_flag(buf,
-                                              BTRFS_HEADER_FLAG_WRITTEN)) {
-                                       clean_tree_block(NULL, root, buf);
-                                       btrfs_tree_unlock(buf);
-                                       free_extent_buffer(buf);
-                                       return 1;
-                               }
-                               btrfs_tree_unlock(buf);
-                       }
+       if (is_data)
+               goto pinit;
+
+       buf = btrfs_find_tree_block(root, bytenr, num_bytes);
+       if (!buf)
+               goto pinit;
+
+       /* we can reuse a block if it hasn't been written
+        * and it is from this transaction.  We can't
+        * reuse anything from the tree log root because
+        * it has tiny sub-transactions.
+        */
+       if (btrfs_buffer_uptodate(buf, 0) &&
+           btrfs_try_tree_lock(buf)) {
+               u64 header_owner = btrfs_header_owner(buf);
+               u64 header_transid = btrfs_header_generation(buf);
+               if (header_owner != BTRFS_TREE_LOG_OBJECTID &&
+                   header_transid == trans->transid &&
+                   !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
+                       clean_tree_block(NULL, root, buf);
+                       btrfs_tree_unlock(buf);
                        free_extent_buffer(buf);
+                       return 1;
                }
-pinit:
-               btrfs_update_pinned_extents(root, bytenr, num_bytes, 1);
-       } else {
-               set_extent_bits(&root->fs_info->pending_del,
-                               bytenr, bytenr + num_bytes - 1,
-                               EXTENT_LOCKED, GFP_NOFS);
+               btrfs_tree_unlock(buf);
        }
+       free_extent_buffer(buf);
+pinit:
+       btrfs_update_pinned_extents(root, bytenr, num_bytes, 1);
+
        BUG_ON(err < 0);
        return 0;
 }
@@ -1642,11 +1847,12 @@ pinit:
 /*
  * remove an extent from the root, returns 0 on success
  */
-static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
-                        *root, u64 bytenr, u64 num_bytes,
+static int __free_extent(struct btrfs_trans_handle *trans,
+                        struct btrfs_root *root,
+                        u64 bytenr, u64 num_bytes, u64 parent,
                         u64 root_objectid, u64 ref_generation,
-                        u64 owner_objectid, u64 owner_offset, int pin,
-                        int mark_free)
+                        u64 owner_objectid, u64 owner_offset,
+                        int pin, int mark_free)
 {
        struct btrfs_path *path;
        struct btrfs_key key;
@@ -1669,10 +1875,8 @@ static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
                return -ENOMEM;
 
        path->reada = 1;
-       ret = lookup_extent_backref(trans, extent_root, path,
-                                   bytenr, root_objectid,
-                                   ref_generation,
-                                   owner_objectid, owner_offset, 1);
+       ret = lookup_extent_backref(trans, extent_root, path, bytenr, parent,
+                                   root_objectid, ref_generation, 1);
        if (ret == 0) {
                struct btrfs_key found_key;
                extent_slot = path->slots[0];
@@ -1690,8 +1894,15 @@ static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
                        if (path->slots[0] - extent_slot > 5)
                                break;
                }
-               if (!found_extent)
-                       ret = btrfs_del_item(trans, extent_root, path);
+               if (!found_extent) {
+                       ret = remove_extent_backref(trans, extent_root, path);
+                       BUG_ON(ret);
+                       btrfs_release_path(extent_root, path);
+                       ret = btrfs_search_slot(trans, extent_root,
+                                               &key, path, -1, 1);
+                       BUG_ON(ret);
+                       extent_slot = path->slots[0];
+               }
        } else {
                btrfs_print_leaf(extent_root, path->nodes[0]);
                WARN_ON(1);
@@ -1700,14 +1911,6 @@ static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
                       root_objectid, ref_generation, owner_objectid,
                       owner_offset);
        }
-       if (!found_extent) {
-               btrfs_release_path(extent_root, path);
-               ret = btrfs_search_slot(trans, extent_root, &key, path, -1, 1);
-               if (ret < 0)
-                       return ret;
-               BUG_ON(ret);
-               extent_slot = path->slots[0];
-       }
 
        leaf = path->nodes[0];
        ei = btrfs_item_ptr(leaf, extent_slot,
@@ -1720,6 +1923,10 @@ static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
        btrfs_mark_buffer_dirty(leaf);
 
        if (refs == 0 && found_extent && path->slots[0] == extent_slot + 1) {
+               struct btrfs_extent_ref *ref;
+               ref = btrfs_item_ptr(leaf, path->slots[0],
+                                    struct btrfs_extent_ref);
+               BUG_ON(btrfs_ref_num_refs(leaf, ref) != 1);
                /* if the back ref and the extent are next to each other
                 * they get deleted below in one shot
                 */
@@ -1727,15 +1934,13 @@ static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
                num_to_del = 2;
        } else if (found_extent) {
                /* otherwise delete the extent back ref */
-               ret = btrfs_del_item(trans, extent_root, path);
+               ret = remove_extent_backref(trans, extent_root, path);
                BUG_ON(ret);
                /* if refs are 0, we need to setup the path for deletion */
                if (refs == 0) {
                        btrfs_release_path(extent_root, path);
                        ret = btrfs_search_slot(trans, extent_root, &key, path,
                                                -1, 1);
-                       if (ret < 0)
-                               return ret;
                        BUG_ON(ret);
                }
        }
@@ -1749,8 +1954,8 @@ static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 #endif
 
                if (pin) {
-                       ret = pin_down_bytes(root, bytenr, num_bytes,
-                            owner_objectid >= BTRFS_FIRST_FREE_OBJECTID, 0);
+                       ret = pin_down_bytes(trans, root, bytenr, num_bytes,
+                               owner_objectid >= BTRFS_FIRST_FREE_OBJECTID);
                        if (ret > 0)
                                mark_free = 1;
                        BUG_ON(ret < 0);
@@ -1769,9 +1974,7 @@ static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
                                           root_used - num_bytes);
                ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
                                      num_to_del);
-               if (ret) {
-                       return ret;
-               }
+               BUG_ON(ret);
                ret = update_block_group(trans, root, bytenr, num_bytes, 0,
                                         mark_free);
                BUG_ON(ret);
@@ -1810,33 +2013,61 @@ static int del_pending_extents(struct btrfs_trans_handle *trans, struct
 {
        int ret;
        int err = 0;
+       int mark_free = 0;
        u64 start;
        u64 end;
+       u64 priv;
        struct extent_io_tree *pending_del;
-       struct extent_io_tree *pinned_extents;
+       struct extent_io_tree *extent_ins;
+       struct pending_extent_op *extent_op;
 
        WARN_ON(!mutex_is_locked(&extent_root->fs_info->alloc_mutex));
+       extent_ins = &extent_root->fs_info->extent_ins;
        pending_del = &extent_root->fs_info->pending_del;
-       pinned_extents = &extent_root->fs_info->pinned_extents;
 
        while(1) {
                ret = find_first_extent_bit(pending_del, 0, &start, &end,
                                            EXTENT_LOCKED);
                if (ret)
                        break;
+
+               ret = get_state_private(pending_del, start, &priv);
+               BUG_ON(ret);
+               extent_op = (struct pending_extent_op *)(unsigned long)priv;
+
                clear_extent_bits(pending_del, start, end, EXTENT_LOCKED,
                                  GFP_NOFS);
-               if (!test_range_bit(&extent_root->fs_info->extent_ins,
-                                   start, end, EXTENT_LOCKED, 0)) {
-                       btrfs_update_pinned_extents(extent_root, start,
-                                             end + 1 - start, 1);
+
+               ret = pin_down_bytes(trans, extent_root, start,
+                                    end + 1 - start, 0);
+               mark_free = ret > 0;
+               if (!test_range_bit(extent_ins, start, end,
+                                   EXTENT_LOCKED, 0)) {
+free_extent:
                        ret = __free_extent(trans, extent_root,
-                                            start, end + 1 - start,
-                                            extent_root->root_key.objectid,
-                                            0, 0, 0, 0, 0);
+                                           start, end + 1 - start,
+                                           extent_op->orig_parent,
+                                           extent_root->root_key.objectid,
+                                           extent_op->orig_generation,
+                                           extent_op->level, 0, 0, mark_free);
+                       kfree(extent_op);
                } else {
-                       clear_extent_bits(&extent_root->fs_info->extent_ins,
-                                         start, end, EXTENT_LOCKED, GFP_NOFS);
+                       kfree(extent_op);
+                       ret = get_state_private(extent_ins, start, &priv);
+                       BUG_ON(ret);
+                       extent_op = (struct pending_extent_op *)
+                                                       (unsigned long)priv;
+
+                       clear_extent_bits(extent_ins, start, end,
+                                         EXTENT_LOCKED, GFP_NOFS);
+
+                       if (extent_op->type == PENDING_BACKREF_UPDATE)
+                               goto free_extent;
+
+                       ret = update_block_group(trans, extent_root, start,
+                                               end + 1 - start, 0, mark_free);
+                       BUG_ON(ret);
+                       kfree(extent_op);
                }
                if (ret)
                        err = ret;
@@ -1854,29 +2085,47 @@ static int del_pending_extents(struct btrfs_trans_handle *trans, struct
  * remove an extent from the root, returns 0 on success
  */
 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
-                              struct btrfs_root *root, u64 bytenr,
-                              u64 num_bytes, u64 root_objectid,
-                              u64 ref_generation, u64 owner_objectid,
-                              u64 owner_offset, int pin)
+                              struct btrfs_root *root,
+                              u64 bytenr, u64 num_bytes, u64 parent,
+                              u64 root_objectid, u64 ref_generation,
+                              u64 owner_objectid, u64 owner_offset, int pin)
 {
        struct btrfs_root *extent_root = root->fs_info->extent_root;
        int pending_ret;
        int ret;
 
        WARN_ON(num_bytes < root->sectorsize);
-       if (!root->ref_cows)
-               ref_generation = 0;
-
        if (root == extent_root) {
-               pin_down_bytes(root, bytenr, num_bytes, 0, 1);
+               struct pending_extent_op *extent_op;
+
+               extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
+               BUG_ON(!extent_op);
+
+               extent_op->type = PENDING_EXTENT_DELETE;
+               extent_op->bytenr = bytenr;
+               extent_op->num_bytes = num_bytes;
+               extent_op->parent = parent;
+               extent_op->orig_parent = parent;
+               extent_op->generation = ref_generation;
+               extent_op->orig_generation = ref_generation;
+               extent_op->level = (int)owner_objectid;
+
+               set_extent_bits(&root->fs_info->pending_del,
+                               bytenr, bytenr + num_bytes - 1,
+                               EXTENT_LOCKED, GFP_NOFS);
+               set_state_private(&root->fs_info->pending_del,
+                                 bytenr, (unsigned long)extent_op);
                return 0;
        }
        /* if metadata always pin */
        if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
                if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
+                       struct btrfs_block_group_cache *cache;
+
                        /* btrfs_free_reserved_extent */
-                       set_extent_dirty(&root->fs_info->free_space_cache,
-                                bytenr, bytenr + num_bytes - 1, GFP_NOFS);
+                       cache = btrfs_lookup_block_group(root->fs_info, bytenr);
+                       BUG_ON(!cache);
+                       btrfs_add_free_space(cache, bytenr, num_bytes);
                        return 0;
                }
                pin = 1;
@@ -1886,9 +2135,9 @@ static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
        if (ref_generation != trans->transid)
                pin = 1;
 
-       ret = __free_extent(trans, root, bytenr, num_bytes, root_objectid,
-                           ref_generation, owner_objectid, owner_offset,
-                           pin, pin == 0);
+       ret = __free_extent(trans, root, bytenr, num_bytes, parent,
+                           root_objectid, ref_generation, owner_objectid,
+                           owner_offset, pin, pin == 0);
 
        finish_current_insert(trans, root->fs_info->extent_root);
        pending_ret = del_pending_extents(trans, root->fs_info->extent_root);
@@ -1896,15 +2145,15 @@ static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
 }
 
 int btrfs_free_extent(struct btrfs_trans_handle *trans,
-                     struct btrfs_root *root, u64 bytenr,
-                     u64 num_bytes, u64 root_objectid,
-                     u64 ref_generation, u64 owner_objectid,
-                     u64 owner_offset, int pin)
+                     struct btrfs_root *root,
+                     u64 bytenr, u64 num_bytes, u64 parent,
+                     u64 root_objectid, u64 ref_generation,
+                     u64 owner_objectid, u64 owner_offset, int pin)
 {
        int ret;
 
        maybe_lock_mutex(root);
-       ret = __btrfs_free_extent(trans, root, bytenr, num_bytes,
+       ret = __btrfs_free_extent(trans, root, bytenr, num_bytes, parent,
                                  root_objectid, ref_generation,
                                  owner_objectid, owner_offset, pin);
        maybe_unlock_mutex(root);
@@ -1941,8 +2190,6 @@ static int noinline find_free_extent(struct btrfs_trans_handle *trans,
        u64 total_needed = num_bytes;
        u64 *last_ptr = NULL;
        struct btrfs_block_group_cache *block_group;
-       int full_scan = 0;
-       int wrapped = 0;
        int chunk_alloc_done = 0;
        int empty_cluster = 2 * 1024 * 1024;
        int allowed_chunk_alloc = 0;
@@ -1958,9 +2205,9 @@ static int noinline find_free_extent(struct btrfs_trans_handle *trans,
                empty_cluster = 256 * 1024;
        }
 
-       if ((data & BTRFS_BLOCK_GROUP_DATA) && btrfs_test_opt(root, SSD)) {
+       if ((data & BTRFS_BLOCK_GROUP_DATA) && btrfs_test_opt(root, SSD))
                last_ptr = &root->fs_info->last_data_alloc;
-       }
+
        if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
                last_ptr = &root->fs_info->last_log_alloc;
                if (!last_ptr == 0 && root->fs_info->last_alloc) {
@@ -1971,9 +2218,8 @@ static int noinline find_free_extent(struct btrfs_trans_handle *trans,
        if (last_ptr) {
                if (*last_ptr)
                        hint_byte = *last_ptr;
-               else {
+               else
                        empty_size += empty_cluster;
-               }
        }
 
        search_start = max(search_start, first_logical_byte(root, 0));
@@ -1982,145 +2228,172 @@ static int noinline find_free_extent(struct btrfs_trans_handle *trans,
        if (search_end == (u64)-1)
                search_end = btrfs_super_total_bytes(&info->super_copy);
 
-       if (hint_byte) {
-               block_group = btrfs_lookup_first_block_group(info, hint_byte);
-               if (!block_group)
-                       hint_byte = search_start;
-               block_group = btrfs_find_block_group(root, block_group,
-                                                    hint_byte, data, 1);
-               if (last_ptr && *last_ptr == 0 && block_group)
-                       hint_byte = block_group->key.objectid;
-       } else {
-               block_group = btrfs_find_block_group(root,
-                                                    trans->block_group,
-                                                    search_start, data, 1);
-       }
        search_start = max(search_start, hint_byte);
-
        total_needed += empty_size;
 
-check_failed:
-       if (!block_group) {
-               block_group = btrfs_lookup_first_block_group(info,
-                                                            search_start);
-               if (!block_group)
-                       block_group = btrfs_lookup_first_block_group(info,
-                                                      orig_search_start);
-       }
-       if (full_scan && !chunk_alloc_done) {
-               if (allowed_chunk_alloc) {
-                       do_chunk_alloc(trans, root,
-                                    num_bytes + 2 * 1024 * 1024, data, 1);
-                       allowed_chunk_alloc = 0;
-               } else if (block_group && block_group_bits(block_group, data)) {
-                       block_group->space_info->force_alloc = 1;
+new_group:
+       block_group = btrfs_lookup_block_group(info, search_start);
+
+       /*
+        * Ok this looks a little tricky, buts its really simple.  First if we
+        * didn't find a block group obviously we want to start over.
+        * Secondly, if the block group we found does not match the type we
+        * need, and we have a last_ptr and its not 0, chances are the last
+        * allocation we made was at the end of the block group, so lets go
+        * ahead and skip the looking through the rest of the block groups and
+        * start at the beginning.  This helps with metadata allocations,
+        * since you are likely to have a bunch of data block groups to search
+        * through first before you realize that you need to start over, so go
+        * ahead and start over and save the time.
+        */
+       if (!block_group || (!block_group_bits(block_group, data) &&
+                            last_ptr && *last_ptr)) {
+               if (search_start != orig_search_start) {
+                       if (last_ptr && *last_ptr)
+                               *last_ptr = 0;
+                       search_start = orig_search_start;
+                       goto new_group;
+               } else if (!chunk_alloc_done && allowed_chunk_alloc) {
+                       ret = do_chunk_alloc(trans, root,
+                                            num_bytes + 2 * 1024 * 1024,
+                                            data, 1);
+                       if (ret < 0) {
+                               struct btrfs_space_info *info;
+
+                               info = __find_space_info(root->fs_info, data);
+                               goto error;
+                       }
+                       BUG_ON(ret);
+                       chunk_alloc_done = 1;
+                       search_start = orig_search_start;
+                       goto new_group;
+               } else {
+                       ret = -ENOSPC;
+                       goto error;
                }
-               chunk_alloc_done = 1;
-       }
-       ret = find_search_start(root, &block_group, &search_start,
-                               total_needed, data);
-       if (ret == -ENOSPC && last_ptr && *last_ptr) {
-               *last_ptr = 0;
-               block_group = btrfs_lookup_first_block_group(info,
-                                                            orig_search_start);
-               search_start = orig_search_start;
-               ret = find_search_start(root, &block_group, &search_start,
-                                       total_needed, data);
        }
-       if (ret == -ENOSPC)
-               goto enospc;
-       if (ret)
-               goto error;
 
-       if (last_ptr && *last_ptr && search_start != *last_ptr) {
-               *last_ptr = 0;
-               if (!empty_size) {
-                       empty_size += empty_cluster;
-                       total_needed += empty_size;
+       /*
+        * this is going to seach through all of the existing block groups it
+        * can find, so if we don't find something we need to see if we can
+        * allocate what we need.
+        */
+       ret = find_free_space(root, &block_group, &search_start,
+                             total_needed, data);
+       if (ret == -ENOSPC) {
+               /*
+                * instead of allocating, start at the original search start
+                * and see if there is something to be found, if not then we
+                * allocate
+                */
+               if (search_start != orig_search_start) {
+                       if (last_ptr && *last_ptr) {
+                               *last_ptr = 0;
+                               total_needed += empty_cluster;
+                       }
+                       search_start = orig_search_start;
+                       goto new_group;
                }
-               block_group = btrfs_lookup_first_block_group(info,
-                                                      orig_search_start);
-               search_start = orig_search_start;
-               ret = find_search_start(root, &block_group,
-                                       &search_start, total_needed, data);
-               if (ret == -ENOSPC)
-                       goto enospc;
-               if (ret)
+
+               /*
+                * we've already allocated, we're pretty screwed
+                */
+               if (chunk_alloc_done) {
+                       goto error;
+               } else if (!allowed_chunk_alloc && block_group &&
+                          block_group_bits(block_group, data)) {
+                       block_group->space_info->force_alloc = 1;
+                       goto error;
+               } else if (!allowed_chunk_alloc) {
+                       goto error;
+               }
+
+               ret = do_chunk_alloc(trans, root, num_bytes + 2 * 1024 * 1024,
+                                    data, 1);
+               if (ret < 0)
                        goto error;
+
+               BUG_ON(ret);
+               chunk_alloc_done = 1;
+               if (block_group)
+                       search_start = block_group->key.objectid +
+                               block_group->key.offset;
+               else
+                       search_start = orig_search_start;
+               goto new_group;
        }
 
+       if (ret)
+               goto error;
+
        search_start = stripe_align(root, search_start);
        ins->objectid = search_start;
        ins->offset = num_bytes;
 
-       if (ins->objectid + num_bytes >= search_end)
-               goto enospc;
+       if (ins->objectid + num_bytes >= search_end) {
+               search_start = orig_search_start;
+               if (chunk_alloc_done) {
+                       ret = -ENOSPC;
+                       goto error;
+               }
+               goto new_group;
+       }
 
        if (ins->objectid + num_bytes >
            block_group->key.objectid + block_group->key.offset) {
+               if (search_start == orig_search_start && chunk_alloc_done) {
+                       ret = -ENOSPC;
+                       goto error;
+               }
                search_start = block_group->key.objectid +
                        block_group->key.offset;
                goto new_group;
        }
 
-       if (test_range_bit(&info->extent_ins, ins->objectid,
-                          ins->objectid + num_bytes -1, EXTENT_LOCKED, 0)) {
-               search_start = ins->objectid + num_bytes;
-               goto new_group;
-       }
-
-       if (test_range_bit(&info->pinned_extents, ins->objectid,
-                          ins->objectid + num_bytes -1, EXTENT_DIRTY, 0)) {
-               search_start = ins->objectid + num_bytes;
-               goto new_group;
-       }
-
        if (exclude_nr > 0 && (ins->objectid + num_bytes > exclude_start &&
            ins->objectid < exclude_start + exclude_nr)) {
                search_start = exclude_start + exclude_nr;
                goto new_group;
        }
 
-       if (!(data & BTRFS_BLOCK_GROUP_DATA)) {
-               block_group = btrfs_lookup_block_group(info, ins->objectid);
-               if (block_group)
-                       trans->block_group = block_group;
-       }
+       if (!(data & BTRFS_BLOCK_GROUP_DATA))
+               trans->block_group = block_group;
+
        ins->offset = num_bytes;
        if (last_ptr) {
                *last_ptr = ins->objectid + ins->offset;
                if (*last_ptr ==
-                   btrfs_super_total_bytes(&root->fs_info->super_copy)) {
+                   btrfs_super_total_bytes(&root->fs_info->super_copy))
                        *last_ptr = 0;
-               }
-       }
-       return 0;
-
-new_group:
-       if (search_start + num_bytes >= search_end) {
-enospc:
-               search_start = orig_search_start;
-               if (full_scan) {
-                       ret = -ENOSPC;
-                       goto error;
-               }
-               if (wrapped) {
-                       if (!full_scan)
-                               total_needed -= empty_size;
-                       full_scan = 1;
-               } else
-                       wrapped = 1;
        }
-       block_group = btrfs_lookup_first_block_group(info, search_start);
-       cond_resched();
-       block_group = btrfs_find_block_group(root, block_group,
-                                            search_start, data, 0);
-       goto check_failed;
 
+       ret = 0;
 error:
        return ret;
 }
 
+static void dump_space_info(struct btrfs_space_info *info, u64 bytes)
+{
+       struct btrfs_block_group_cache *cache;
+       struct list_head *l;
+
+       printk(KERN_INFO "space_info has %Lu free, is %sfull\n",
+              info->total_bytes - info->bytes_used - info->bytes_pinned,
+              (info->full) ? "" : "not ");
+
+       spin_lock(&info->lock);
+       list_for_each(l, &info->block_groups) {
+               cache = list_entry(l, struct btrfs_block_group_cache, list);
+               spin_lock(&cache->lock);
+               printk(KERN_INFO "block group %Lu has %Lu bytes, %Lu used "
+                      "%Lu pinned\n",
+                      cache->key.objectid, cache->key.offset,
+                      btrfs_block_group_used(&cache->item), cache->pinned);
+               btrfs_dump_free_space(cache, bytes);
+               spin_unlock(&cache->lock);
+       }
+       spin_unlock(&info->lock);
+}
 static int __btrfs_reserve_extent(struct btrfs_trans_handle *trans,
                                  struct btrfs_root *root,
                                  u64 num_bytes, u64 min_alloc_size,
@@ -2132,6 +2405,7 @@ static int __btrfs_reserve_extent(struct btrfs_trans_handle *trans,
        u64 search_start = 0;
        u64 alloc_profile;
        struct btrfs_fs_info *info = root->fs_info;
+       struct btrfs_block_group_cache *cache;
 
        if (data) {
                alloc_profile = info->avail_data_alloc_bits &
@@ -2159,11 +2433,9 @@ again:
                                     BTRFS_BLOCK_GROUP_METADATA |
                                     (info->metadata_alloc_profile &
                                      info->avail_metadata_alloc_bits), 0);
-                       BUG_ON(ret);
                }
                ret = do_chunk_alloc(trans, root->fs_info->extent_root,
                                     num_bytes + 2 * 1024 * 1024, data, 0);
-               BUG_ON(ret);
        }
 
        WARN_ON(num_bytes < root->sectorsize);
@@ -2174,26 +2446,44 @@ again:
 
        if (ret == -ENOSPC && num_bytes > min_alloc_size) {
                num_bytes = num_bytes >> 1;
+               num_bytes = num_bytes & ~(root->sectorsize - 1);
                num_bytes = max(num_bytes, min_alloc_size);
                do_chunk_alloc(trans, root->fs_info->extent_root,
                               num_bytes, data, 1);
                goto again;
        }
        if (ret) {
-               printk("allocation failed flags %Lu\n", data);
+               struct btrfs_space_info *sinfo;
+
+               sinfo = __find_space_info(root->fs_info, data);
+               printk("allocation failed flags %Lu, wanted %Lu\n",
+                      data, num_bytes);
+               dump_space_info(sinfo, num_bytes);
                BUG();
        }
-       clear_extent_dirty(&root->fs_info->free_space_cache,
-                          ins->objectid, ins->objectid + ins->offset - 1,
-                          GFP_NOFS);
-       return 0;
+       cache = btrfs_lookup_block_group(root->fs_info, ins->objectid);
+       if (!cache) {
+               printk(KERN_ERR "Unable to find block group for %Lu\n", ins->objectid);
+               return -ENOSPC;
+       }
+
+       ret = btrfs_remove_free_space(cache, ins->objectid, ins->offset);
+
+       return ret;
 }
 
 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
 {
+       struct btrfs_block_group_cache *cache;
+
        maybe_lock_mutex(root);
-       set_extent_dirty(&root->fs_info->free_space_cache,
-                        start, start + len - 1, GFP_NOFS);
+       cache = btrfs_lookup_block_group(root->fs_info, start);
+       if (!cache) {
+               printk(KERN_ERR "Unable to find block group for %Lu\n", start);
+               maybe_unlock_mutex(root);
+               return -ENOSPC;
+       }
+       btrfs_add_free_space(cache, start, len);
        maybe_unlock_mutex(root);
        return 0;
 }
@@ -2215,7 +2505,7 @@ int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
 }
 
 static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
-                                        struct btrfs_root *root,
+                                        struct btrfs_root *root, u64 parent,
                                         u64 root_objectid, u64 ref_generation,
                                         u64 owner, u64 owner_offset,
                                         struct btrfs_key *ins)
@@ -2233,6 +2523,9 @@ static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
        struct btrfs_path *path;
        struct btrfs_key keys[2];
 
+       if (parent == 0)
+               parent = ins->objectid;
+
        /* block accounting for super block */
        spin_lock_irq(&info->delalloc_lock);
        super_used = btrfs_super_bytes_used(&info->super_copy);
@@ -2244,17 +2537,32 @@ static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
        btrfs_set_root_used(&root->root_item, root_used + num_bytes);
 
        if (root == extent_root) {
+               struct pending_extent_op *extent_op;
+
+               extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
+               BUG_ON(!extent_op);
+
+               extent_op->type = PENDING_EXTENT_INSERT;
+               extent_op->bytenr = ins->objectid;
+               extent_op->num_bytes = ins->offset;
+               extent_op->parent = parent;
+               extent_op->orig_parent = 0;
+               extent_op->generation = ref_generation;
+               extent_op->orig_generation = 0;
+               extent_op->level = (int)owner;
+
                set_extent_bits(&root->fs_info->extent_ins, ins->objectid,
                                ins->objectid + ins->offset - 1,
                                EXTENT_LOCKED, GFP_NOFS);
+               set_state_private(&root->fs_info->extent_ins,
+                                 ins->objectid, (unsigned long)extent_op);
                goto update_block;
        }
 
        memcpy(&keys[0], ins, sizeof(*ins));
-       keys[1].offset = hash_extent_ref(root_objectid, ref_generation,
-                                        owner, owner_offset);
        keys[1].objectid = ins->objectid;
        keys[1].type = BTRFS_EXTENT_REF_KEY;
+       keys[1].offset = parent;
        sizes[0] = sizeof(*extent_item);
        sizes[1] = sizeof(*ref);
 
@@ -2263,8 +2571,8 @@ static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
 
        ret = btrfs_insert_empty_items(trans, extent_root, path, keys,
                                       sizes, 2);
-
        BUG_ON(ret);
+
        extent_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
                                     struct btrfs_extent_item);
        btrfs_set_extent_refs(path->nodes[0], extent_item, 1);
@@ -2275,6 +2583,7 @@ static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
        btrfs_set_ref_generation(path->nodes[0], ref, ref_generation);
        btrfs_set_ref_objectid(path->nodes[0], ref, owner);
        btrfs_set_ref_offset(path->nodes[0], ref, owner_offset);
+       btrfs_set_ref_num_refs(path->nodes[0], ref, 1);
 
        btrfs_mark_buffer_dirty(path->nodes[0]);
 
@@ -2303,16 +2612,19 @@ out:
 }
 
 int btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
-                               struct btrfs_root *root,
+                               struct btrfs_root *root, u64 parent,
                                u64 root_objectid, u64 ref_generation,
                                u64 owner, u64 owner_offset,
                                struct btrfs_key *ins)
 {
        int ret;
+
+       if (root_objectid == BTRFS_TREE_LOG_OBJECTID)
+               return 0;
        maybe_lock_mutex(root);
-       ret = __btrfs_alloc_reserved_extent(trans, root, root_objectid,
-                                           ref_generation, owner,
-                                           owner_offset, ins);
+       ret = __btrfs_alloc_reserved_extent(trans, root, parent,
+                                           root_objectid, ref_generation,
+                                           owner, owner_offset, ins);
        maybe_unlock_mutex(root);
        return ret;
 }
@@ -2323,7 +2635,7 @@ int btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
  * space cache bits as well
  */
 int btrfs_alloc_logged_extent(struct btrfs_trans_handle *trans,
-                               struct btrfs_root *root,
+                               struct btrfs_root *root, u64 parent,
                                u64 root_objectid, u64 ref_generation,
                                u64 owner, u64 owner_offset,
                                struct btrfs_key *ins)
@@ -2335,12 +2647,11 @@ int btrfs_alloc_logged_extent(struct btrfs_trans_handle *trans,
        block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
        cache_block_group(root, block_group);
 
-       clear_extent_dirty(&root->fs_info->free_space_cache,
-                          ins->objectid, ins->objectid + ins->offset - 1,
-                          GFP_NOFS);
-       ret = __btrfs_alloc_reserved_extent(trans, root, root_objectid,
-                                           ref_generation, owner,
-                                           owner_offset, ins);
+       ret = btrfs_remove_free_space(block_group, ins->objectid, ins->offset);
+       BUG_ON(ret);
+       ret = __btrfs_alloc_reserved_extent(trans, root, parent,
+                                           root_objectid, ref_generation,
+                                           owner, owner_offset, ins);
        maybe_unlock_mutex(root);
        return ret;
 }
@@ -2354,9 +2665,9 @@ int btrfs_alloc_logged_extent(struct btrfs_trans_handle *trans,
  */
 int btrfs_alloc_extent(struct btrfs_trans_handle *trans,
                       struct btrfs_root *root,
-                      u64 num_bytes, u64 min_alloc_size,
+                      u64 num_bytes, u64 parent, u64 min_alloc_size,
                       u64 root_objectid, u64 ref_generation,
-                      u64 owner, u64 owner_offset,
+                      u64 owner_objectid, u64 owner_offset,
                       u64 empty_size, u64 hint_byte,
                       u64 search_end, struct btrfs_key *ins, u64 data)
 {
@@ -2369,9 +2680,9 @@ int btrfs_alloc_extent(struct btrfs_trans_handle *trans,
                                     search_end, ins, data);
        BUG_ON(ret);
        if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
-               ret = __btrfs_alloc_reserved_extent(trans, root, root_objectid,
-                                                   ref_generation, owner,
-                                                   owner_offset, ins);
+               ret = __btrfs_alloc_reserved_extent(trans, root, parent,
+                                       root_objectid, ref_generation,
+                                       owner_objectid, owner_offset, ins);
                BUG_ON(ret);
 
        }
@@ -2409,10 +2720,9 @@ struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
  */
 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
                                             struct btrfs_root *root,
-                                            u32 blocksize,
+                                            u32 blocksize, u64 parent,
                                             u64 root_objectid,
                                             u64 ref_generation,
-                                            u64 first_objectid,
                                             int level,
                                             u64 hint,
                                             u64 empty_size)
@@ -2421,10 +2731,9 @@ struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
        int ret;
        struct extent_buffer *buf;
 
-       ret = btrfs_alloc_extent(trans, root, blocksize, blocksize,
-                                root_objectid, ref_generation,
-                                level, first_objectid, empty_size, hint,
-                                (u64)-1, &ins, 0);
+       ret = btrfs_alloc_extent(trans, root, blocksize, parent, blocksize,
+                                root_objectid, ref_generation, level, 0,
+                                empty_size, hint, (u64)-1, &ins, 0);
        if (ret) {
                BUG_ON(ret > 0);
                return ERR_PTR(ret);
@@ -2472,15 +2781,14 @@ int btrfs_drop_leaf_ref(struct btrfs_trans_handle *trans,
                mutex_lock(&root->fs_info->alloc_mutex);
                ret = __btrfs_free_extent(trans, root, disk_bytenr,
                                btrfs_file_extent_disk_num_bytes(leaf, fi),
-                               leaf_owner, leaf_generation,
+                               leaf->start, leaf_owner, leaf_generation,
                                key.objectid, key.offset, 0);
                mutex_unlock(&root->fs_info->alloc_mutex);
+               BUG_ON(ret);
 
                atomic_inc(&root->fs_info->throttle_gen);
                wake_up(&root->fs_info->transaction_throttle);
                cond_resched();
-
-               BUG_ON(ret);
        }
        return 0;
 }
@@ -2495,10 +2803,10 @@ static int noinline cache_drop_leaf_ref(struct btrfs_trans_handle *trans,
 
        for (i = 0; i < ref->nritems; i++) {
                mutex_lock(&root->fs_info->alloc_mutex);
-               ret = __btrfs_free_extent(trans, root,
-                                       info->bytenr, info->num_bytes,
-                                       ref->owner, ref->generation,
-                                       info->objectid, info->offset, 0);
+               ret = __btrfs_free_extent(trans, root, info->bytenr,
+                                         info->num_bytes, ref->bytenr,
+                                         ref->owner, ref->generation,
+                                         info->objectid, info->offset, 0);
                mutex_unlock(&root->fs_info->alloc_mutex);
 
                atomic_inc(&root->fs_info->throttle_gen);
@@ -2517,7 +2825,7 @@ int drop_snap_lookup_refcount(struct btrfs_root *root, u64 start, u64 len,
 {
        int ret;
 
-       ret = lookup_extent_ref(NULL, root, start, len, refs);
+       ret = btrfs_lookup_extent_ref(NULL, root, start, len, refs);
        BUG_ON(ret);
 
 #if 0 // some debugging code in case we see problems here
@@ -2613,8 +2921,8 @@ static int noinline walk_down_tree(struct btrfs_trans_handle *trans,
 
                        mutex_lock(&root->fs_info->alloc_mutex);
                        ret = __btrfs_free_extent(trans, root, bytenr,
-                                               blocksize, root_owner,
-                                               root_gen, 0, 0, 1);
+                                               blocksize, parent->start,
+                                               root_owner, root_gen, 0, 0, 1);
                        BUG_ON(ret);
                        mutex_unlock(&root->fs_info->alloc_mutex);
 
@@ -2631,8 +2939,6 @@ static int noinline walk_down_tree(struct btrfs_trans_handle *trans,
                 * So, we don't need to check it again
                 */
                if (*level == 1) {
-                       struct btrfs_key key;
-                       btrfs_node_key_to_cpu(cur, &key, path->slots[*level]);
                        ref = btrfs_lookup_leaf_ref(root, bytenr);
                        if (ref) {
                                ret = cache_drop_leaf_ref(trans, root, ref);
@@ -2691,12 +2997,13 @@ out:
 
        mutex_lock(&root->fs_info->alloc_mutex);
        ret = __btrfs_free_extent(trans, root, bytenr, blocksize,
-                                 root_owner, root_gen, 0, 0, 1);
+                                 parent->start, root_owner, root_gen,
+                                 0, 0, 1);
+       mutex_unlock(&root->fs_info->alloc_mutex);
        free_extent_buffer(path->nodes[*level]);
        path->nodes[*level] = NULL;
        *level += 1;
        BUG_ON(ret);
-       mutex_unlock(&root->fs_info->alloc_mutex);
 
        cond_resched();
        return 0;
@@ -2733,19 +3040,18 @@ static int noinline walk_up_tree(struct btrfs_trans_handle *trans,
                        root_item->drop_level = i;
                        return 0;
                } else {
-                       if (path->nodes[*level] == root->node) {
-                               root_owner = root->root_key.objectid;
-                               root_gen =
-                                  btrfs_header_generation(path->nodes[*level]);
-                       } else {
-                               struct extent_buffer *node;
-                               node = path->nodes[*level + 1];
-                               root_owner = btrfs_header_owner(node);
-                               root_gen = btrfs_header_generation(node);
-                       }
+                       struct extent_buffer *parent;
+                       if (path->nodes[*level] == root->node)
+                               parent = path->nodes[*level];
+                       else
+                               parent = path->nodes[*level + 1];
+
+                       root_owner = btrfs_header_owner(parent);
+                       root_gen = btrfs_header_generation(parent);
                        ret = btrfs_free_extent(trans, root,
                                                path->nodes[*level]->start,
                                                path->nodes[*level]->len,
+                                               parent->start,
                                                root_owner, root_gen, 0, 0, 1);
                        BUG_ON(ret);
                        free_extent_buffer(path->nodes[*level]);
@@ -2842,31 +3148,28 @@ out:
 
 int btrfs_free_block_groups(struct btrfs_fs_info *info)
 {
-       u64 start;
-       u64 end;
-       u64 ptr;
-       int ret;
+       struct btrfs_block_group_cache *block_group;
+       struct rb_node *n;
 
        mutex_lock(&info->alloc_mutex);
-       while(1) {
-               ret = find_first_extent_bit(&info->block_group_cache, 0,
-                                           &start, &end, (unsigned int)-1);
-               if (ret)
-                       break;
-               ret = get_state_private(&info->block_group_cache, start, &ptr);
-               if (!ret)
-                       kfree((void *)(unsigned long)ptr);
-               clear_extent_bits(&info->block_group_cache, start,
-                                 end, (unsigned int)-1, GFP_NOFS);
-       }
-       while(1) {
-               ret = find_first_extent_bit(&info->free_space_cache, 0,
-                                           &start, &end, EXTENT_DIRTY);
-               if (ret)
-                       break;
-               clear_extent_dirty(&info->free_space_cache, start,
-                                  end, GFP_NOFS);
+       spin_lock(&info->block_group_cache_lock);
+       while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
+               block_group = rb_entry(n, struct btrfs_block_group_cache,
+                                      cache_node);
+
+               spin_unlock(&info->block_group_cache_lock);
+               btrfs_remove_free_space_cache(block_group);
+               spin_lock(&info->block_group_cache_lock);
+
+               rb_erase(&block_group->cache_node,
+                        &info->block_group_cache_tree);
+
+               spin_lock(&block_group->space_info->lock);
+               list_del(&block_group->list);
+               spin_unlock(&block_group->space_info->lock);
+               kfree(block_group);
        }
+       spin_unlock(&info->block_group_cache_lock);
        mutex_unlock(&info->alloc_mutex);
        return 0;
 }
@@ -3385,7 +3688,6 @@ int btrfs_shrink_extent_tree(struct btrfs_root *root, u64 shrink_start)
        u64 total_found;
        u64 shrink_last_byte;
        struct btrfs_block_group_cache *shrink_block_group;
-       struct btrfs_fs_info *info = root->fs_info;
        struct btrfs_key key;
        struct btrfs_key found_key;
        struct extent_buffer *leaf;
@@ -3541,15 +3843,17 @@ next:
                goto out;
        }
 
-       clear_extent_bits(&info->block_group_cache, key.objectid,
-                         key.objectid + key.offset - 1,
-                         (unsigned int)-1, GFP_NOFS);
-
-
-       clear_extent_bits(&info->free_space_cache,
-                          key.objectid, key.objectid + key.offset - 1,
-                          (unsigned int)-1, GFP_NOFS);
+       spin_lock(&root->fs_info->block_group_cache_lock);
+       rb_erase(&shrink_block_group->cache_node,
+                &root->fs_info->block_group_cache_tree);
+       spin_unlock(&root->fs_info->block_group_cache_lock);
 
+       ret = btrfs_remove_free_space(shrink_block_group, key.objectid,
+                                     key.offset);
+       if (ret) {
+               btrfs_end_transaction(trans, root);
+               goto out;
+       }
        /*
        memset(shrink_block_group, 0, sizeof(*shrink_block_group));
        kfree(shrink_block_group);
@@ -3565,9 +3869,9 @@ next:
        /* the code to unpin extents might set a few bits in the free
         * space cache for this range again
         */
-       clear_extent_bits(&info->free_space_cache,
-                          key.objectid, key.objectid + key.offset - 1,
-                          (unsigned int)-1, GFP_NOFS);
+       /* XXX? */
+       ret = btrfs_remove_free_space(shrink_block_group, key.objectid,
+                                     key.offset);
 out:
        btrfs_free_path(path);
        mutex_unlock(&root->fs_info->alloc_mutex);
@@ -3615,16 +3919,13 @@ int btrfs_read_block_groups(struct btrfs_root *root)
 {
        struct btrfs_path *path;
        int ret;
-       int bit;
        struct btrfs_block_group_cache *cache;
        struct btrfs_fs_info *info = root->fs_info;
        struct btrfs_space_info *space_info;
-       struct extent_io_tree *block_group_cache;
        struct btrfs_key key;
        struct btrfs_key found_key;
        struct extent_buffer *leaf;
 
-       block_group_cache = &info->block_group_cache;
        root = info->extent_root;
        key.objectid = 0;
        key.offset = 0;
@@ -3652,6 +3953,7 @@ int btrfs_read_block_groups(struct btrfs_root *root)
                }
 
                spin_lock_init(&cache->lock);
+               INIT_LIST_HEAD(&cache->list);
                read_extent_buffer(leaf, &cache->item,
                                   btrfs_item_ptr_offset(leaf, path->slots[0]),
                                   sizeof(cache->item));
@@ -3660,31 +3962,19 @@ int btrfs_read_block_groups(struct btrfs_root *root)
                key.objectid = found_key.objectid + found_key.offset;
                btrfs_release_path(root, path);
                cache->flags = btrfs_block_group_flags(&cache->item);
-               bit = 0;
-               if (cache->flags & BTRFS_BLOCK_GROUP_DATA) {
-                       bit = BLOCK_GROUP_DATA;
-               } else if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
-                       bit = BLOCK_GROUP_SYSTEM;
-               } else if (cache->flags & BTRFS_BLOCK_GROUP_METADATA) {
-                       bit = BLOCK_GROUP_METADATA;
-               }
-               set_avail_alloc_bits(info, cache->flags);
 
                ret = update_space_info(info, cache->flags, found_key.offset,
                                        btrfs_block_group_used(&cache->item),
                                        &space_info);
                BUG_ON(ret);
                cache->space_info = space_info;
+               spin_lock(&space_info->lock);
+               list_add(&cache->list, &space_info->block_groups);
+               spin_unlock(&space_info->lock);
+
+               ret = btrfs_add_block_group_cache(root->fs_info, cache);
+               BUG_ON(ret);
 
-               /* use EXTENT_LOCKED to prevent merging */
-               set_extent_bits(block_group_cache, found_key.objectid,
-                               found_key.objectid + found_key.offset - 1,
-                               EXTENT_LOCKED, GFP_NOFS);
-               set_state_private(block_group_cache, found_key.objectid,
-                                 (unsigned long)cache);
-               set_extent_bits(block_group_cache, found_key.objectid,
-                               found_key.objectid + found_key.offset - 1,
-                               bit | EXTENT_LOCKED, GFP_NOFS);
                if (key.objectid >=
                    btrfs_super_total_bytes(&info->super_copy))
                        break;
@@ -3702,22 +3992,22 @@ int btrfs_make_block_group(struct btrfs_trans_handle *trans,
                           u64 size)
 {
        int ret;
-       int bit = 0;
        struct btrfs_root *extent_root;
        struct btrfs_block_group_cache *cache;
-       struct extent_io_tree *block_group_cache;
 
        WARN_ON(!mutex_is_locked(&root->fs_info->alloc_mutex));
        extent_root = root->fs_info->extent_root;
-       block_group_cache = &root->fs_info->block_group_cache;
 
        root->fs_info->last_trans_new_blockgroup = trans->transid;
 
        cache = kzalloc(sizeof(*cache), GFP_NOFS);
-       BUG_ON(!cache);
+       if (!cache)
+               return -ENOMEM;
+
        cache->key.objectid = chunk_offset;
        cache->key.offset = size;
        spin_lock_init(&cache->lock);
+       INIT_LIST_HEAD(&cache->list);
        btrfs_set_key_type(&cache->key, BTRFS_BLOCK_GROUP_ITEM_KEY);
 
        btrfs_set_block_group_used(&cache->item, bytes_used);
@@ -3728,16 +4018,12 @@ int btrfs_make_block_group(struct btrfs_trans_handle *trans,
        ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
                                &cache->space_info);
        BUG_ON(ret);
+       spin_lock(&cache->space_info->lock);
+       list_add(&cache->list, &cache->space_info->block_groups);
+       spin_unlock(&cache->space_info->lock);
 
-       bit = block_group_state_bits(type);
-       set_extent_bits(block_group_cache, chunk_offset,
-                       chunk_offset + size - 1,
-                       EXTENT_LOCKED, GFP_NOFS);
-       set_state_private(block_group_cache, chunk_offset,
-                         (unsigned long)cache);
-       set_extent_bits(block_group_cache, chunk_offset,
-                       chunk_offset + size - 1,
-                       bit | EXTENT_LOCKED, GFP_NOFS);
+       ret = btrfs_add_block_group_cache(root->fs_info, cache);
+       BUG_ON(ret);
 
        ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
                                sizeof(cache->item));