Btrfs: alloc_mutex latency reduction
[pandora-kernel.git] / fs / btrfs / extent-tree.c
index 41a6346..febc629 100644 (file)
@@ -25,6 +25,7 @@
 #include "print-tree.h"
 #include "transaction.h"
 #include "volumes.h"
+#include "locking.h"
 
 #define BLOCK_GROUP_DATA     EXTENT_WRITEBACK
 #define BLOCK_GROUP_METADATA EXTENT_UPTODATE
@@ -36,7 +37,28 @@ static int finish_current_insert(struct btrfs_trans_handle *trans, struct
                                 btrfs_root *extent_root);
 static int del_pending_extents(struct btrfs_trans_handle *trans, struct
                               btrfs_root *extent_root);
+static struct btrfs_block_group_cache *
+__btrfs_find_block_group(struct btrfs_root *root,
+                        struct btrfs_block_group_cache *hint,
+                        u64 search_start, int data, int owner);
 
+void maybe_lock_mutex(struct btrfs_root *root)
+{
+       if (root != root->fs_info->extent_root &&
+           root != root->fs_info->chunk_root &&
+           root != root->fs_info->dev_root) {
+               mutex_lock(&root->fs_info->alloc_mutex);
+       }
+}
+
+void maybe_unlock_mutex(struct btrfs_root *root)
+{
+       if (root != root->fs_info->extent_root &&
+           root != root->fs_info->chunk_root &&
+           root != root->fs_info->dev_root) {
+               mutex_unlock(&root->fs_info->alloc_mutex);
+       }
+}
 
 static int cache_block_group(struct btrfs_root *root,
                             struct btrfs_block_group_cache *block_group)
@@ -66,6 +88,12 @@ static int cache_block_group(struct btrfs_root *root,
                return -ENOMEM;
 
        path->reada = 2;
+       /*
+        * we get into deadlocks with paths held by callers of this function.
+        * since the alloc_mutex is protecting things right now, just
+        * skip the locking here
+        */
+       path->skip_locking = 1;
        first_free = block_group->key.objectid;
        key.objectid = block_group->key.objectid;
        key.offset = 0;
@@ -217,6 +245,7 @@ static int noinline find_search_start(struct btrfs_root *root,
        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;
 
@@ -318,10 +347,10 @@ static int block_group_state_bits(u64 flags)
        return bits;
 }
 
-struct btrfs_block_group_cache *btrfs_find_block_group(struct btrfs_root *root,
-                                                struct btrfs_block_group_cache
-                                                *hint, u64 search_start,
-                                                int data, int owner)
+static struct btrfs_block_group_cache *
+__btrfs_find_block_group(struct btrfs_root *root,
+                        struct btrfs_block_group_cache *hint,
+                        u64 search_start, int data, int owner)
 {
        struct btrfs_block_group_cache *cache;
        struct extent_io_tree *block_group_cache;
@@ -350,19 +379,25 @@ struct btrfs_block_group_cache *btrfs_find_block_group(struct btrfs_root *root,
                struct btrfs_block_group_cache *shint;
                shint = btrfs_lookup_first_block_group(info, search_start);
                if (shint && block_group_bits(shint, data) && !shint->ro) {
+                       spin_lock(&shint->lock);
                        used = btrfs_block_group_used(&shint->item);
                        if (used + shint->pinned <
                            div_factor(shint->key.offset, factor)) {
+                               spin_unlock(&shint->lock);
                                return shint;
                        }
+                       spin_unlock(&shint->lock);
                }
        }
        if (hint && !hint->ro && block_group_bits(hint, data)) {
+               spin_lock(&hint->lock);
                used = btrfs_block_group_used(&hint->item);
                if (used + hint->pinned <
                    div_factor(hint->key.offset, factor)) {
+                       spin_unlock(&hint->lock);
                        return hint;
                }
+               spin_unlock(&hint->lock);
                last = hint->key.objectid + hint->key.offset;
        } else {
                if (hint)
@@ -384,6 +419,7 @@ again:
                }
 
                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);
 
@@ -391,9 +427,11 @@ again:
                        free_check = div_factor(cache->key.offset, factor);
                        if (used + cache->pinned < free_check) {
                                found_group = cache;
+                               spin_unlock(&cache->lock);
                                goto found;
                        }
                }
+               spin_unlock(&cache->lock);
                cond_resched();
        }
        if (!wrapped) {
@@ -411,6 +449,16 @@ found:
        return found_group;
 }
 
+struct btrfs_block_group_cache *btrfs_find_block_group(struct btrfs_root *root,
+                                                struct btrfs_block_group_cache
+                                                *hint, u64 search_start,
+                                                int data, int owner)
+{
+
+       struct btrfs_block_group_cache *ret;
+       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)
 {
@@ -646,7 +694,7 @@ out:
        return ret;
 }
 
-int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
+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,
@@ -696,6 +744,22 @@ int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
        return 0;
 }
 
+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)
+{
+       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);
+       return ret;
+}
+
 int btrfs_extent_post_op(struct btrfs_trans_handle *trans,
                         struct btrfs_root *root)
 {
@@ -760,6 +824,10 @@ u32 btrfs_count_snapshots_in_path(struct btrfs_root *root,
        struct btrfs_extent_ref *ref_item;
        int level = -1;
 
+       /* FIXME, needs locking */
+       BUG();
+
+       mutex_lock(&root->fs_info->alloc_mutex);
        path = btrfs_alloc_path();
 again:
        if (level == -1)
@@ -854,33 +922,9 @@ again:
 
 out:
        btrfs_free_path(path);
+       mutex_unlock(&root->fs_info->alloc_mutex);
        return total_count;
 }
-int btrfs_inc_root_ref(struct btrfs_trans_handle *trans,
-                      struct btrfs_root *root, u64 owner_objectid)
-{
-       u64 generation;
-       u64 key_objectid;
-       u64 level;
-       u32 nritems;
-       struct btrfs_disk_key disk_key;
-
-       level = btrfs_header_level(root->node);
-       generation = trans->transid;
-       nritems = btrfs_header_nritems(root->node);
-       if (nritems > 0) {
-               if (level == 0)
-                       btrfs_item_key(root->node, &disk_key, 0);
-               else
-                       btrfs_node_key(root->node, &disk_key, 0);
-               key_objectid = btrfs_disk_key_objectid(&disk_key);
-       } else {
-               key_objectid = 0;
-       }
-       return btrfs_inc_extent_ref(trans, root, root->node->start,
-                                   root->node->len, owner_objectid,
-                                   generation, level, key_objectid);
-}
 
 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
                  struct extent_buffer *buf)
@@ -900,6 +944,7 @@ int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
        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);
@@ -913,24 +958,32 @@ int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
                        disk_bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
                        if (disk_bytenr == 0)
                                continue;
-                       ret = btrfs_inc_extent_ref(trans, root, disk_bytenr,
+
+                       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);
-                       ret = btrfs_inc_extent_ref(trans, root, bytenr,
+
+                       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;
                        }
                }
@@ -1019,6 +1072,7 @@ int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
        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);
@@ -1045,6 +1099,7 @@ int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
                                  BLOCK_GROUP_DIRTY, GFP_NOFS);
        }
        btrfs_free_path(path);
+       mutex_unlock(&root->fs_info->alloc_mutex);
        return werr;
 }
 
@@ -1162,26 +1217,29 @@ static int do_chunk_alloc(struct btrfs_trans_handle *trans,
                space_info->force_alloc = 0;
        }
        if (space_info->full)
-               return 0;
+               goto out;
 
        thresh = div_factor(space_info->total_bytes, 6);
        if (!force &&
           (space_info->bytes_used + space_info->bytes_pinned + alloc_bytes) <
            thresh)
-               return 0;
+               goto out;
 
+       mutex_lock(&extent_root->fs_info->chunk_mutex);
        ret = btrfs_alloc_chunk(trans, extent_root, &start, &num_bytes, flags);
        if (ret == -ENOSPC) {
 printk("space info full %Lu\n", flags);
                space_info->full = 1;
-               return 0;
+               goto out_unlock;
        }
        BUG_ON(ret);
 
        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;
 }
 
@@ -1198,6 +1256,7 @@ static int update_block_group(struct btrfs_trans_handle *trans,
        u64 start;
        u64 end;
 
+       WARN_ON(!mutex_is_locked(&root->fs_info->alloc_mutex));
        while(total) {
                cache = btrfs_lookup_block_group(info, bytenr);
                if (!cache) {
@@ -1210,21 +1269,25 @@ static int update_block_group(struct btrfs_trans_handle *trans,
                set_extent_bits(&info->block_group_cache, start, end,
                                BLOCK_GROUP_DIRTY, GFP_NOFS);
 
+               spin_lock(&cache->lock);
                old_val = btrfs_block_group_used(&cache->item);
                num_bytes = min(total, cache->key.offset - byte_in_group);
                if (alloc) {
                        old_val += num_bytes;
                        cache->space_info->bytes_used += num_bytes;
+                       btrfs_set_block_group_used(&cache->item, old_val);
+                       spin_unlock(&cache->lock);
                } else {
                        old_val -= num_bytes;
                        cache->space_info->bytes_used -= num_bytes;
+                       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);
                        }
                }
-               btrfs_set_block_group_used(&cache->item, old_val);
                total -= num_bytes;
                bytenr += num_bytes;
        }
@@ -1253,6 +1316,7 @@ static int update_pinned_extents(struct btrfs_root *root,
        struct btrfs_block_group_cache *cache;
        struct btrfs_fs_info *fs_info = root->fs_info;
 
+       WARN_ON(!mutex_is_locked(&root->fs_info->alloc_mutex));
        if (pin) {
                set_extent_dirty(&fs_info->pinned_extents,
                                bytenr, bytenr + num - 1, GFP_NOFS);
@@ -1272,14 +1336,18 @@ static int update_pinned_extents(struct btrfs_root *root,
                }
                if (pin) {
                        if (cache) {
+                               spin_lock(&cache->lock);
                                cache->pinned += len;
                                cache->space_info->bytes_pinned += len;
+                               spin_unlock(&cache->lock);
                        }
                        fs_info->total_pinned += len;
                } else {
                        if (cache) {
+                               spin_lock(&cache->lock);
                                cache->pinned -= len;
                                cache->space_info->bytes_pinned -= len;
+                               spin_unlock(&cache->lock);
                        }
                        fs_info->total_pinned -= len;
                }
@@ -1318,6 +1386,7 @@ int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
        struct extent_io_tree *free_space_cache;
        free_space_cache = &root->fs_info->free_space_cache;
 
+       mutex_lock(&root->fs_info->alloc_mutex);
        while(1) {
                ret = find_first_extent_bit(unpin, 0, &start, &end,
                                            EXTENT_DIRTY);
@@ -1326,7 +1395,13 @@ int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
                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);
+               if (need_resched()) {
+                       mutex_unlock(&root->fs_info->alloc_mutex);
+                       cond_resched();
+                       mutex_lock(&root->fs_info->alloc_mutex);
+               }
        }
+       mutex_unlock(&root->fs_info->alloc_mutex);
        return 0;
 }
 
@@ -1345,6 +1420,7 @@ static int finish_current_insert(struct btrfs_trans_handle *trans,
        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();
@@ -1361,20 +1437,39 @@ static int finish_current_insert(struct btrfs_trans_handle *trans,
                                        &extent_item, sizeof(extent_item));
                clear_extent_bits(&info->extent_ins, start, end, EXTENT_LOCKED,
                                  GFP_NOFS);
-               eb = read_tree_block(extent_root, ins.objectid, ins.offset,
-                                    trans->transid);
+
+               eb = btrfs_find_tree_block(extent_root, ins.objectid,
+                                          ins.offset);
+
+               if (!btrfs_buffer_uptodate(eb, trans->transid)) {
+                       mutex_unlock(&extent_root->fs_info->alloc_mutex);
+                       btrfs_read_buffer(eb, trans->transid);
+                       mutex_lock(&extent_root->fs_info->alloc_mutex);
+               }
+
+               btrfs_tree_lock(eb);
                level = btrfs_header_level(eb);
                if (level == 0) {
                        btrfs_item_key(eb, &first, 0);
                } else {
                        btrfs_node_key(eb, &first, 0);
                }
+               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);
-               free_extent_buffer(eb);
+               if (need_resched()) {
+                       mutex_unlock(&extent_root->fs_info->alloc_mutex);
+                       cond_resched();
+                       mutex_lock(&extent_root->fs_info->alloc_mutex);
+               }
        }
        btrfs_free_path(path);
        return 0;
@@ -1384,12 +1479,14 @@ static int pin_down_bytes(struct btrfs_root *root, u64 bytenr, u32 num_bytes,
                          int pending)
 {
        int err = 0;
-       struct extent_buffer *buf;
 
+       WARN_ON(!mutex_is_locked(&root->fs_info->alloc_mutex));
        if (!pending) {
+               struct extent_buffer *buf;
                buf = btrfs_find_tree_block(root, bytenr, num_bytes);
                if (buf) {
-                       if (btrfs_buffer_uptodate(buf, 0)) {
+                       if (btrfs_try_tree_lock(buf) &&
+                           btrfs_buffer_uptodate(buf, 0)) {
                                u64 transid =
                                    root->fs_info->running_transaction->transid;
                                u64 header_transid =
@@ -1398,9 +1495,11 @@ static int pin_down_bytes(struct btrfs_root *root, u64 bytenr, u32 num_bytes,
                                    !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);
                        }
                        free_extent_buffer(buf);
                }
@@ -1435,6 +1534,7 @@ static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
        struct btrfs_extent_item *ei;
        u32 refs;
 
+       WARN_ON(!mutex_is_locked(&root->fs_info->alloc_mutex));
        key.objectid = bytenr;
        btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
        key.offset = num_bytes;
@@ -1526,9 +1626,11 @@ static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
                }
 
                /* block accounting for super block */
+               spin_lock_irq(&info->delalloc_lock);
                super_used = btrfs_super_bytes_used(&info->super_copy);
                btrfs_set_super_bytes_used(&info->super_copy,
                                           super_used - num_bytes);
+               spin_unlock_irq(&info->delalloc_lock);
 
                /* block accounting for root item */
                root_used = btrfs_root_used(&root->root_item);
@@ -1562,6 +1664,7 @@ static int del_pending_extents(struct btrfs_trans_handle *trans, struct
        struct extent_io_tree *pending_del;
        struct extent_io_tree *pinned_extents;
 
+       WARN_ON(!mutex_is_locked(&extent_root->fs_info->alloc_mutex));
        pending_del = &extent_root->fs_info->pending_del;
        pinned_extents = &extent_root->fs_info->pinned_extents;
 
@@ -1570,15 +1673,28 @@ static int del_pending_extents(struct btrfs_trans_handle *trans, struct
                                            EXTENT_LOCKED);
                if (ret)
                        break;
-               update_pinned_extents(extent_root, start, end + 1 - start, 1);
                clear_extent_bits(pending_del, start, end, EXTENT_LOCKED,
                                  GFP_NOFS);
-               ret = __free_extent(trans, extent_root,
-                                    start, end + 1 - start,
-                                    extent_root->root_key.objectid,
-                                    0, 0, 0, 0, 0);
+               if (!test_range_bit(&extent_root->fs_info->extent_ins,
+                                   start, end, EXTENT_LOCKED, 0)) {
+                       update_pinned_extents(extent_root, start,
+                                             end + 1 - start, 1);
+                       ret = __free_extent(trans, extent_root,
+                                            start, end + 1 - start,
+                                            extent_root->root_key.objectid,
+                                            0, 0, 0, 0, 0);
+               } else {
+                       clear_extent_bits(&extent_root->fs_info->extent_ins,
+                                         start, end, EXTENT_LOCKED, GFP_NOFS);
+               }
                if (ret)
                        err = ret;
+
+               if (need_resched()) {
+                       mutex_unlock(&extent_root->fs_info->alloc_mutex);
+                       cond_resched();
+                       mutex_lock(&extent_root->fs_info->alloc_mutex);
+               }
        }
        return err;
 }
@@ -1586,10 +1702,11 @@ static int del_pending_extents(struct btrfs_trans_handle *trans, struct
 /*
  * remove an extent from the root, returns 0 on success
  */
-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)
+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 *extent_root = root->fs_info->extent_root;
        int pending_ret;
@@ -1606,10 +1723,28 @@ int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
        ret = __free_extent(trans, root, bytenr, num_bytes, 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);
        return ret ? ret : pending_ret;
 }
 
+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)
+{
+       int ret;
+
+       maybe_lock_mutex(root);
+       ret = __btrfs_free_extent(trans, root, bytenr, num_bytes,
+                                 root_objectid, ref_generation,
+                                 owner_objectid, owner_offset, pin);
+       maybe_unlock_mutex(root);
+       return ret;
+}
+
 static u64 stripe_align(struct btrfs_root *root, u64 val)
 {
        u64 mask = ((u64)root->stripesize - 1);
@@ -1814,34 +1949,17 @@ error:
        return ret;
 }
 
-/*
- * finds a free extent and does all the dirty work required for allocation
- * returns the key for the extent through ins, and a tree buffer for
- * the first block of the extent through buf.
- *
- * returns 0 if everything worked, non-zero otherwise.
- */
-int btrfs_alloc_extent(struct btrfs_trans_handle *trans,
-                      struct btrfs_root *root,
-                      u64 num_bytes, u64 min_alloc_size,
-                      u64 root_objectid, u64 ref_generation,
-                      u64 owner, u64 owner_offset,
-                      u64 empty_size, u64 hint_byte,
-                      u64 search_end, struct btrfs_key *ins, u64 data)
+static int __btrfs_reserve_extent(struct btrfs_trans_handle *trans,
+                                 struct btrfs_root *root,
+                                 u64 num_bytes, u64 min_alloc_size,
+                                 u64 empty_size, u64 hint_byte,
+                                 u64 search_end, struct btrfs_key *ins,
+                                 u64 data)
 {
        int ret;
-       int pending_ret;
-       u64 super_used;
-       u64 root_used;
        u64 search_start = 0;
        u64 alloc_profile;
-       u32 sizes[2];
        struct btrfs_fs_info *info = root->fs_info;
-       struct btrfs_root *extent_root = info->extent_root;
-       struct btrfs_extent_item *extent_item;
-       struct btrfs_extent_ref *ref;
-       struct btrfs_path *path;
-       struct btrfs_key keys[2];
 
        if (data) {
                alloc_profile = info->avail_data_alloc_bits &
@@ -1891,23 +2009,59 @@ again:
        }
        if (ret) {
                printk("allocation failed flags %Lu\n", data);
+               BUG();
        }
-       BUG_ON(ret);
-       if (ret)
-               return ret;
+       clear_extent_dirty(&root->fs_info->free_space_cache,
+                          ins->objectid, ins->objectid + ins->offset - 1,
+                          GFP_NOFS);
+       return 0;
+}
+
+int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
+                                 struct btrfs_root *root,
+                                 u64 num_bytes, u64 min_alloc_size,
+                                 u64 empty_size, u64 hint_byte,
+                                 u64 search_end, struct btrfs_key *ins,
+                                 u64 data)
+{
+       int ret;
+       maybe_lock_mutex(root);
+       ret = __btrfs_reserve_extent(trans, root, num_bytes, min_alloc_size,
+                                    empty_size, hint_byte, search_end, ins,
+                                    data);
+       maybe_unlock_mutex(root);
+       return ret;
+}
+
+static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
+                                        struct btrfs_root *root,
+                                        u64 root_objectid, u64 ref_generation,
+                                        u64 owner, u64 owner_offset,
+                                        struct btrfs_key *ins)
+{
+       int ret;
+       int pending_ret;
+       u64 super_used;
+       u64 root_used;
+       u64 num_bytes = ins->offset;
+       u32 sizes[2];
+       struct btrfs_fs_info *info = root->fs_info;
+       struct btrfs_root *extent_root = info->extent_root;
+       struct btrfs_extent_item *extent_item;
+       struct btrfs_extent_ref *ref;
+       struct btrfs_path *path;
+       struct btrfs_key keys[2];
 
        /* block accounting for super block */
+       spin_lock_irq(&info->delalloc_lock);
        super_used = btrfs_super_bytes_used(&info->super_copy);
        btrfs_set_super_bytes_used(&info->super_copy, super_used + num_bytes);
+       spin_unlock_irq(&info->delalloc_lock);
 
        /* block accounting for root item */
        root_used = btrfs_root_used(&root->root_item);
        btrfs_set_root_used(&root->root_item, root_used + num_bytes);
 
-       clear_extent_dirty(&root->fs_info->free_space_cache,
-                          ins->objectid, ins->objectid + ins->offset - 1,
-                          GFP_NOFS);
-
        if (root == extent_root) {
                set_extent_bits(&root->fs_info->extent_ins, ins->objectid,
                                ins->objectid + ins->offset - 1,
@@ -1915,10 +2069,6 @@ again:
                goto update_block;
        }
 
-       WARN_ON(trans->alloc_exclude_nr);
-       trans->alloc_exclude_start = ins->objectid;
-       trans->alloc_exclude_nr = ins->offset;
-
        memcpy(&keys[0], ins, sizeof(*ins));
        keys[1].offset = hash_extent_ref(root_objectid, ref_generation,
                                         owner, owner_offset);
@@ -1953,11 +2103,11 @@ again:
        finish_current_insert(trans, extent_root);
        pending_ret = del_pending_extents(trans, extent_root);
 
-       if (ret) {
-               return ret;
-       }
+       if (ret)
+               goto out;
        if (pending_ret) {
-               return pending_ret;
+               ret = pending_ret;
+               goto out;
        }
 
 update_block:
@@ -1967,36 +2117,60 @@ update_block:
                       ins->objectid, ins->offset);
                BUG();
        }
-       return 0;
+out:
+       return ret;
 }
 
+int btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
+                               struct btrfs_root *root,
+                               u64 root_objectid, u64 ref_generation,
+                               u64 owner, u64 owner_offset,
+                               struct btrfs_key *ins)
+{
+       int ret;
+       maybe_lock_mutex(root);
+       ret = __btrfs_alloc_reserved_extent(trans, root, root_objectid,
+                                           ref_generation, owner,
+                                           owner_offset, ins);
+       maybe_unlock_mutex(root);
+       return ret;
+}
 /*
- * helper function to allocate a block for a given tree
- * returns the tree buffer or NULL.
+ * finds a free extent and does all the dirty work required for allocation
+ * returns the key for the extent through ins, and a tree buffer for
+ * the first block of the extent through buf.
+ *
+ * returns 0 if everything worked, non-zero otherwise.
  */
-struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
-                                            struct btrfs_root *root,
-                                            u32 blocksize,
-                                            u64 root_objectid, u64 hint,
-                                            u64 empty_size)
+int btrfs_alloc_extent(struct btrfs_trans_handle *trans,
+                      struct btrfs_root *root,
+                      u64 num_bytes, u64 min_alloc_size,
+                      u64 root_objectid, u64 ref_generation,
+                      u64 owner, u64 owner_offset,
+                      u64 empty_size, u64 hint_byte,
+                      u64 search_end, struct btrfs_key *ins, u64 data)
 {
-       u64 ref_generation;
+       int ret;
 
-       if (root->ref_cows)
-               ref_generation = trans->transid;
-       else
-               ref_generation = 0;
+       maybe_lock_mutex(root);
 
+       ret = __btrfs_reserve_extent(trans, root, num_bytes,
+                                    min_alloc_size, empty_size, hint_byte,
+                                    search_end, ins, data);
+       BUG_ON(ret);
+       ret = __btrfs_alloc_reserved_extent(trans, root, root_objectid,
+                                           ref_generation, owner,
+                                           owner_offset, ins);
+       BUG_ON(ret);
 
-       return __btrfs_alloc_free_block(trans, root, blocksize, root_objectid,
-                                       ref_generation, 0, 0, hint, empty_size);
+       maybe_unlock_mutex(root);
+       return ret;
 }
-
 /*
  * helper function to allocate a block for a given tree
  * returns the tree buffer or NULL.
  */
-struct extent_buffer *__btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
+struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
                                             struct btrfs_root *root,
                                             u32 blocksize,
                                             u64 root_objectid,
@@ -2026,6 +2200,7 @@ struct extent_buffer *__btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
                return ERR_PTR(-ENOMEM);
        }
        btrfs_set_header_generation(buf, trans->transid);
+       btrfs_tree_lock(buf);
        clean_tree_block(trans, root, buf);
        btrfs_set_buffer_uptodate(buf);
 
@@ -2036,8 +2211,6 @@ struct extent_buffer *__btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
 
        set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
                         buf->start + buf->len - 1, GFP_NOFS);
-       if (!btrfs_test_opt(root, SSD))
-               btrfs_set_buffer_defrag(buf);
        trans->blocks_used++;
        return buf;
 }
@@ -2059,8 +2232,11 @@ static int noinline drop_leaf_ref(struct btrfs_trans_handle *trans,
        leaf_owner = btrfs_header_owner(leaf);
        leaf_generation = btrfs_header_generation(leaf);
 
+       mutex_unlock(&root->fs_info->alloc_mutex);
+
        for (i = 0; i < nritems; i++) {
                u64 disk_bytenr;
+               cond_resched();
 
                btrfs_item_key_to_cpu(leaf, &key, i);
                if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
@@ -2076,12 +2252,17 @@ static int noinline drop_leaf_ref(struct btrfs_trans_handle *trans,
                disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
                if (disk_bytenr == 0)
                        continue;
-               ret = btrfs_free_extent(trans, root, disk_bytenr,
+
+               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,
                                key.objectid, key.offset, 0);
+               mutex_unlock(&root->fs_info->alloc_mutex);
                BUG_ON(ret);
        }
+
+       mutex_lock(&root->fs_info->alloc_mutex);
        return 0;
 }
 
@@ -2121,17 +2302,30 @@ static void noinline reada_walk_down(struct btrfs_root *root,
                                continue;
                        }
                }
-               mutex_unlock(&root->fs_info->fs_mutex);
                ret = readahead_tree_block(root, bytenr, blocksize,
                                           btrfs_node_ptr_generation(node, i));
                last = bytenr + blocksize;
                cond_resched();
-               mutex_lock(&root->fs_info->fs_mutex);
                if (ret)
                        break;
        }
 }
 
+/*
+ * we want to avoid as much random IO as we can with the alloc mutex
+ * held, so drop the lock and do the lookup, then do it again with the
+ * lock held.
+ */
+int drop_snap_lookup_refcount(struct btrfs_root *root, u64 start, u64 len,
+                             u32 *refs)
+{
+       mutex_unlock(&root->fs_info->alloc_mutex);
+       lookup_extent_ref(NULL, root, start, len, refs);
+       cond_resched();
+       mutex_lock(&root->fs_info->alloc_mutex);
+       return lookup_extent_ref(NULL, root, start, len, refs);
+}
+
 /*
  * helper function for drop_snapshot, this walks down the tree dropping ref
  * counts as it goes.
@@ -2151,10 +2345,11 @@ static int noinline walk_down_tree(struct btrfs_trans_handle *trans,
        int ret;
        u32 refs;
 
+       mutex_lock(&root->fs_info->alloc_mutex);
+
        WARN_ON(*level < 0);
        WARN_ON(*level >= BTRFS_MAX_LEVEL);
-       ret = lookup_extent_ref(trans, root,
-                               path->nodes[*level]->start,
+       ret = drop_snap_lookup_refcount(root, path->nodes[*level]->start,
                                path->nodes[*level]->len, &refs);
        BUG_ON(ret);
        if (refs > 1)
@@ -2182,14 +2377,15 @@ static int noinline walk_down_tree(struct btrfs_trans_handle *trans,
                bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
                ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
                blocksize = btrfs_level_size(root, *level - 1);
-               ret = lookup_extent_ref(trans, root, bytenr, blocksize, &refs);
+
+               ret = drop_snap_lookup_refcount(root, bytenr, blocksize, &refs);
                BUG_ON(ret);
                if (refs != 1) {
                        parent = path->nodes[*level];
                        root_owner = btrfs_header_owner(parent);
                        root_gen = btrfs_header_generation(parent);
                        path->slots[*level]++;
-                       ret = btrfs_free_extent(trans, root, bytenr,
+                       ret = __btrfs_free_extent(trans, root, bytenr,
                                                blocksize, root_owner,
                                                root_gen, 0, 0, 1);
                        BUG_ON(ret);
@@ -2198,16 +2394,19 @@ static int noinline walk_down_tree(struct btrfs_trans_handle *trans,
                next = btrfs_find_tree_block(root, bytenr, blocksize);
                if (!next || !btrfs_buffer_uptodate(next, ptr_gen)) {
                        free_extent_buffer(next);
-                       reada_walk_down(root, cur, path->slots[*level]);
+                       mutex_unlock(&root->fs_info->alloc_mutex);
+
+                       if (path->slots[*level] == 0)
+                               reada_walk_down(root, cur, path->slots[*level]);
 
-                       mutex_unlock(&root->fs_info->fs_mutex);
                        next = read_tree_block(root, bytenr, blocksize,
                                               ptr_gen);
-                       mutex_lock(&root->fs_info->fs_mutex);
+                       cond_resched();
+                       mutex_lock(&root->fs_info->alloc_mutex);
 
                        /* we've dropped the lock, double check */
-                       ret = lookup_extent_ref(trans, root, bytenr,
-                                               blocksize, &refs);
+                       ret = lookup_extent_ref(NULL, root, bytenr, blocksize,
+                                               &refs);
                        BUG_ON(ret);
                        if (refs != 1) {
                                parent = path->nodes[*level];
@@ -2216,7 +2415,7 @@ static int noinline walk_down_tree(struct btrfs_trans_handle *trans,
 
                                path->slots[*level]++;
                                free_extent_buffer(next);
-                               ret = btrfs_free_extent(trans, root, bytenr,
+                               ret = __btrfs_free_extent(trans, root, bytenr,
                                                        blocksize,
                                                        root_owner,
                                                        root_gen, 0, 0, 1);
@@ -2244,13 +2443,15 @@ out:
        }
 
        root_gen = btrfs_header_generation(parent);
-       ret = btrfs_free_extent(trans, root, path->nodes[*level]->start,
+       ret = __btrfs_free_extent(trans, root, path->nodes[*level]->start,
                                path->nodes[*level]->len,
                                root_owner, root_gen, 0, 0, 1);
        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;
 }
 
@@ -2324,6 +2525,7 @@ int btrfs_drop_snapshot(struct btrfs_trans_handle *trans, struct btrfs_root
        int orig_level;
        struct btrfs_root_item *root_item = &root->root_item;
 
+       WARN_ON(!mutex_is_locked(&root->fs_info->drop_mutex));
        path = btrfs_alloc_path();
        BUG_ON(!path);
 
@@ -2350,6 +2552,16 @@ int btrfs_drop_snapshot(struct btrfs_trans_handle *trans, struct btrfs_root
                btrfs_node_key(node, &found_key, path->slots[level]);
                WARN_ON(memcmp(&found_key, &root_item->drop_progress,
                               sizeof(found_key)));
+               /*
+                * unlock our path, this is safe because only this
+                * function is allowed to delete this snapshot
+                */
+               for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
+                       if (path->nodes[i] && path->locks[i]) {
+                               path->locks[i] = 0;
+                               btrfs_tree_unlock(path->nodes[i]);
+                       }
+               }
        }
        while(1) {
                wret = walk_down_tree(trans, root, path, &level);
@@ -2363,8 +2575,10 @@ int btrfs_drop_snapshot(struct btrfs_trans_handle *trans, struct btrfs_root
                        break;
                if (wret < 0)
                        ret = wret;
-               ret = -EAGAIN;
-               break;
+               if (trans->transaction->in_commit) {
+                       ret = -EAGAIN;
+                       break;
+               }
        }
        for (i = 0; i <= orig_level; i++) {
                if (path->nodes[i]) {
@@ -2383,6 +2597,8 @@ int btrfs_free_block_groups(struct btrfs_fs_info *info)
        u64 end;
        u64 ptr;
        int ret;
+
+       mutex_lock(&info->alloc_mutex);
        while(1) {
                ret = find_first_extent_bit(&info->block_group_cache, 0,
                                            &start, &end, (unsigned int)-1);
@@ -2402,6 +2618,7 @@ int btrfs_free_block_groups(struct btrfs_fs_info *info)
                clear_extent_dirty(&info->free_space_cache, start,
                                   end, GFP_NOFS);
        }
+       mutex_unlock(&info->alloc_mutex);
        return 0;
 }
 
@@ -2484,7 +2701,6 @@ out_unlock:
        kfree(ra);
        trans = btrfs_start_transaction(BTRFS_I(inode)->root, 1);
        if (trans) {
-               btrfs_add_ordered_inode(inode);
                btrfs_end_transaction(trans, BTRFS_I(inode)->root);
                mark_inode_dirty(inode);
        }
@@ -2522,7 +2738,6 @@ static int find_root_for_ref(struct btrfs_root *root,
        u64 root_search_start = BTRFS_FS_TREE_OBJECTID;
        u64 found_bytenr;
        int ret;
-       int i;
 
        root_location.offset = (u64)-1;
        root_location.type = BTRFS_ROOT_ITEM_KEY;
@@ -2546,12 +2761,6 @@ static int find_root_for_ref(struct btrfs_root *root,
                                found_bytenr = path->nodes[level]->start;
                }
 
-               for (i = level; i < BTRFS_MAX_LEVEL; i++) {
-                       if (!path->nodes[i])
-                               break;
-                       free_extent_buffer(path->nodes[i]);
-                       path->nodes[i] = NULL;
-               }
                btrfs_release_path(cur_root, path);
 
                if (found_bytenr == bytenr) {
@@ -2600,6 +2809,8 @@ static int noinline relocate_one_reference(struct btrfs_root *extent_root,
        int ret;
        int level;
 
+       WARN_ON(!mutex_is_locked(&extent_root->fs_info->alloc_mutex));
+
        ref = btrfs_item_ptr(path->nodes[0], path->slots[0],
                             struct btrfs_extent_ref);
        ref_root = btrfs_ref_root(path->nodes[0], ref);
@@ -2618,6 +2829,7 @@ static int noinline relocate_one_reference(struct btrfs_root *extent_root,
        found_root = btrfs_read_fs_root_no_name(extent_root->fs_info,
                                                &root_location);
        BUG_ON(!found_root);
+       mutex_unlock(&extent_root->fs_info->alloc_mutex);
 
        if (ref_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
                found_key.objectid = ref_objectid;
@@ -2644,7 +2856,6 @@ static int noinline relocate_one_reference(struct btrfs_root *extent_root,
                    *last_file_root == ref_root)
                        goto out;
 
-               mutex_unlock(&extent_root->fs_info->fs_mutex);
                inode = btrfs_iget_locked(extent_root->fs_info->sb,
                                          ref_objectid, found_root);
                if (inode->i_state & I_NEW) {
@@ -2660,24 +2871,23 @@ static int noinline relocate_one_reference(struct btrfs_root *extent_root,
                /* this can happen if the reference is not against
                 * the latest version of the tree root
                 */
-               if (is_bad_inode(inode)) {
-                       mutex_lock(&extent_root->fs_info->fs_mutex);
+               if (is_bad_inode(inode))
                        goto out;
-               }
+
                *last_file_objectid = inode->i_ino;
                *last_file_root = found_root->root_key.objectid;
                *last_file_offset = ref_offset;
 
                relocate_inode_pages(inode, ref_offset, extent_key->offset);
                iput(inode);
-               mutex_lock(&extent_root->fs_info->fs_mutex);
        } else {
                struct btrfs_trans_handle *trans;
                struct extent_buffer *eb;
-               int i;
+               int needs_lock = 0;
 
                eb = read_tree_block(found_root, extent_key->objectid,
                                     extent_key->offset, 0);
+               btrfs_tree_lock(eb);
                level = btrfs_header_level(eb);
 
                if (level == 0)
@@ -2685,6 +2895,7 @@ static int noinline relocate_one_reference(struct btrfs_root *extent_root,
                else
                        btrfs_node_key_to_cpu(eb, &found_key, 0);
 
+               btrfs_tree_unlock(eb);
                free_extent_buffer(eb);
 
                ret = find_root_for_ref(extent_root, path, &found_key,
@@ -2694,26 +2905,40 @@ static int noinline relocate_one_reference(struct btrfs_root *extent_root,
                if (ret)
                        goto out;
 
+               /*
+                * right here almost anything could happen to our key,
+                * but that's ok.  The cow below will either relocate it
+                * or someone else will have relocated it.  Either way,
+                * it is in a different spot than it was before and
+                * we're happy.
+                */
+
                trans = btrfs_start_transaction(found_root, 1);
 
+               if (found_root == extent_root->fs_info->extent_root ||
+                   found_root == extent_root->fs_info->chunk_root ||
+                   found_root == extent_root->fs_info->dev_root) {
+                       needs_lock = 1;
+                       mutex_lock(&extent_root->fs_info->alloc_mutex);
+               }
+
                path->lowest_level = level;
                path->reada = 2;
                ret = btrfs_search_slot(trans, found_root, &found_key, path,
                                        0, 1);
                path->lowest_level = 0;
-               for (i = level; i < BTRFS_MAX_LEVEL; i++) {
-                       if (!path->nodes[i])
-                               break;
-                       free_extent_buffer(path->nodes[i]);
-                       path->nodes[i] = NULL;
-               }
                btrfs_release_path(found_root, path);
+
                if (found_root == found_root->fs_info->extent_root)
                        btrfs_extent_post_op(trans, found_root);
+               if (needs_lock)
+                       mutex_unlock(&extent_root->fs_info->alloc_mutex);
+
                btrfs_end_transaction(trans, found_root);
-       }
 
+       }
 out:
+       mutex_lock(&extent_root->fs_info->alloc_mutex);
        return 0;
 }
 
@@ -2853,9 +3078,15 @@ int __alloc_chunk_for_shrink(struct btrfs_root *root,
        u64 new_alloc_flags;
        u64 calc;
 
+       spin_lock(&shrink_block_group->lock);
        if (btrfs_block_group_used(&shrink_block_group->item) > 0) {
+               spin_unlock(&shrink_block_group->lock);
+               mutex_unlock(&root->fs_info->alloc_mutex);
 
                trans = btrfs_start_transaction(root, 1);
+               mutex_lock(&root->fs_info->alloc_mutex);
+               spin_lock(&shrink_block_group->lock);
+
                new_alloc_flags = update_block_group_flags(root,
                                                   shrink_block_group->flags);
                if (new_alloc_flags != shrink_block_group->flags) {
@@ -2864,10 +3095,16 @@ int __alloc_chunk_for_shrink(struct btrfs_root *root,
                } else {
                        calc = shrink_block_group->key.offset;
                }
+               spin_unlock(&shrink_block_group->lock);
+
                do_chunk_alloc(trans, root->fs_info->extent_root,
                               calc + 2 * 1024 * 1024, new_alloc_flags, force);
+
+               mutex_unlock(&root->fs_info->alloc_mutex);
                btrfs_end_transaction(trans, root);
-       }
+               mutex_lock(&root->fs_info->alloc_mutex);
+       } else
+               spin_unlock(&shrink_block_group->lock);
        return 0;
 }
 
@@ -2888,6 +3125,7 @@ int btrfs_shrink_extent_tree(struct btrfs_root *root, u64 shrink_start)
        int ret;
        int progress;
 
+       mutex_lock(&root->fs_info->alloc_mutex);
        shrink_block_group = btrfs_lookup_block_group(root->fs_info,
                                                      shrink_start);
        BUG_ON(!shrink_block_group);
@@ -2942,9 +3180,9 @@ again:
                if (ret < 0)
                        goto out;
 
+next:
                leaf = path->nodes[0];
                nritems = btrfs_header_nritems(leaf);
-next:
                if (path->slots[0] >= nritems) {
                        ret = btrfs_next_leaf(root, path);
                        if (ret < 0)
@@ -2964,9 +3202,7 @@ next:
 
                if (progress && need_resched()) {
                        memcpy(&key, &found_key, sizeof(key));
-                       mutex_unlock(&root->fs_info->fs_mutex);
                        cond_resched();
-                       mutex_lock(&root->fs_info->fs_mutex);
                        btrfs_release_path(root, path);
                        btrfs_search_slot(NULL, root, &key, path, 0, 0);
                        progress = 0;
@@ -2996,15 +3232,15 @@ next:
                printk("btrfs relocate found %llu last extent was %llu\n",
                       (unsigned long long)total_found,
                       (unsigned long long)found_key.objectid);
+               mutex_unlock(&root->fs_info->alloc_mutex);
                trans = btrfs_start_transaction(tree_root, 1);
                btrfs_commit_transaction(trans, tree_root);
 
-               mutex_unlock(&root->fs_info->fs_mutex);
                btrfs_clean_old_snapshots(tree_root);
-               mutex_lock(&root->fs_info->fs_mutex);
 
                trans = btrfs_start_transaction(tree_root, 1);
                btrfs_commit_transaction(trans, tree_root);
+               mutex_lock(&root->fs_info->alloc_mutex);
                goto again;
        }
 
@@ -3012,7 +3248,11 @@ next:
         * we've freed all the extents, now remove the block
         * group item from the tree
         */
+       mutex_unlock(&root->fs_info->alloc_mutex);
+
        trans = btrfs_start_transaction(root, 1);
+
+       mutex_lock(&root->fs_info->alloc_mutex);
        memcpy(&key, &shrink_block_group->key, sizeof(key));
 
        ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
@@ -3034,8 +3274,12 @@ next:
        kfree(shrink_block_group);
 
        btrfs_del_item(trans, root, path);
+       btrfs_release_path(root, path);
+       mutex_unlock(&root->fs_info->alloc_mutex);
        btrfs_commit_transaction(trans, root);
 
+       mutex_lock(&root->fs_info->alloc_mutex);
+
        /* the code to unpin extents might set a few bits in the free
         * space cache for this range again
         */
@@ -3044,20 +3288,22 @@ next:
                           (unsigned int)-1, GFP_NOFS);
 out:
        btrfs_free_path(path);
+       mutex_unlock(&root->fs_info->alloc_mutex);
        return ret;
 }
 
 int find_first_block_group(struct btrfs_root *root, struct btrfs_path *path,
                           struct btrfs_key *key)
 {
-       int ret;
+       int ret = 0;
        struct btrfs_key found_key;
        struct extent_buffer *leaf;
        int slot;
 
        ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
        if (ret < 0)
-               return ret;
+               goto out;
+
        while(1) {
                slot = path->slots[0];
                leaf = path->nodes[0];
@@ -3066,18 +3312,20 @@ int find_first_block_group(struct btrfs_root *root, struct btrfs_path *path,
                        if (ret == 0)
                                continue;
                        if (ret < 0)
-                               goto error;
+                               goto out;
                        break;
                }
                btrfs_item_key_to_cpu(leaf, &found_key, slot);
 
                if (found_key.objectid >= key->objectid &&
-                   found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY)
-                       return 0;
+                   found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
+                       ret = 0;
+                       goto out;
+               }
                path->slots[0]++;
        }
        ret = -ENOENT;
-error:
+out:
        return ret;
 }
 
@@ -3103,6 +3351,7 @@ int btrfs_read_block_groups(struct btrfs_root *root)
        if (!path)
                return -ENOMEM;
 
+       mutex_lock(&root->fs_info->alloc_mutex);
        while(1) {
                ret = find_first_block_group(root, path, &key);
                if (ret > 0) {
@@ -3120,6 +3369,7 @@ int btrfs_read_block_groups(struct btrfs_root *root)
                        break;
                }
 
+               spin_lock_init(&cache->lock);
                read_extent_buffer(leaf, &cache->item,
                                   btrfs_item_ptr_offset(leaf, path->slots[0]),
                                   sizeof(cache->item));
@@ -3147,10 +3397,12 @@ int btrfs_read_block_groups(struct btrfs_root *root)
                /* use EXTENT_LOCKED to prevent merging */
                set_extent_bits(block_group_cache, found_key.objectid,
                                found_key.objectid + found_key.offset - 1,
-                               bit | EXTENT_LOCKED, GFP_NOFS);
+                               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;
@@ -3158,6 +3410,7 @@ int btrfs_read_block_groups(struct btrfs_root *root)
        ret = 0;
 error:
        btrfs_free_path(path);
+       mutex_unlock(&root->fs_info->alloc_mutex);
        return ret;
 }
 
@@ -3172,6 +3425,7 @@ int btrfs_make_block_group(struct btrfs_trans_handle *trans,
        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;
 
@@ -3179,6 +3433,7 @@ int btrfs_make_block_group(struct btrfs_trans_handle *trans,
        BUG_ON(!cache);
        cache->key.objectid = chunk_offset;
        cache->key.offset = size;
+       spin_lock_init(&cache->lock);
        btrfs_set_key_type(&cache->key, BTRFS_BLOCK_GROUP_ITEM_KEY);
 
        btrfs_set_block_group_used(&cache->item, bytes_used);
@@ -3193,10 +3448,13 @@ int btrfs_make_block_group(struct btrfs_trans_handle *trans,
        bit = block_group_state_bits(type);
        set_extent_bits(block_group_cache, chunk_offset,
                        chunk_offset + size - 1,
-                       bit | EXTENT_LOCKED, GFP_NOFS);
-
+                       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_insert_item(trans, extent_root, &cache->key, &cache->item,
                                sizeof(cache->item));
        BUG_ON(ret);
@@ -3205,5 +3463,6 @@ int btrfs_make_block_group(struct btrfs_trans_handle *trans,
        ret = del_pending_extents(trans, extent_root);
        BUG_ON(ret);
        set_avail_alloc_bits(extent_root->fs_info, type);
+
        return 0;
 }