Btrfs: alloc_mutex latency reduction
[pandora-kernel.git] / fs / btrfs / extent-tree.c
index dc3c03c..febc629 100644 (file)
@@ -245,6 +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;
 
@@ -318,7 +319,7 @@ no_cache:
                cache = btrfs_lookup_first_block_group(root->fs_info, last);
        }
        cache_miss = 0;
-       cache = __btrfs_find_block_group(root, cache, last, data, 0);
+       cache = btrfs_find_block_group(root, cache, last, data, 0);
        if (!cache)
                goto no_cache;
        *cache_ret = cache;
@@ -378,19 +379,25 @@ __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)
@@ -412,6 +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);
 
@@ -419,9 +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) {
@@ -446,9 +456,7 @@ struct btrfs_block_group_cache *btrfs_find_block_group(struct btrfs_root *root,
 {
 
        struct btrfs_block_group_cache *ret;
-       mutex_lock(&root->fs_info->alloc_mutex);
        ret = __btrfs_find_block_group(root, hint, search_start, data, owner);
-       mutex_unlock(&root->fs_info->alloc_mutex);
        return ret;
 }
 static u64 hash_extent_ref(u64 root_objectid, u64 ref_generation,
@@ -933,10 +941,10 @@ int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
        if (!root->ref_cows)
                return 0;
 
-       mutex_lock(&root->fs_info->alloc_mutex);
        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);
@@ -950,29 +958,36 @@ 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;
+
+                       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;
                        }
                }
        }
-       mutex_unlock(&root->fs_info->alloc_mutex);
        return 0;
 fail:
        WARN_ON(1);
@@ -1003,7 +1018,6 @@ fail:
                }
        }
 #endif
-       mutex_unlock(&root->fs_info->alloc_mutex);
        return ret;
 }
 
@@ -1242,6 +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) {
@@ -1254,21 +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;
        }
@@ -1297,6 +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);
@@ -1316,14 +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;
                }
@@ -1371,6 +1395,11 @@ 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;
@@ -1391,6 +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();
@@ -1407,8 +1437,16 @@ 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) {
@@ -1427,6 +1465,11 @@ static int finish_current_insert(struct btrfs_trans_handle *trans,
                                          0, level,
                                          btrfs_disk_key_objectid(&first));
                BUG_ON(err);
+               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;
@@ -1437,11 +1480,12 @@ static int pin_down_bytes(struct btrfs_root *root, u64 bytenr, u32 num_bytes,
 {
        int err = 0;
 
+       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_try_tree_lock(buf) &&
+                       if (btrfs_try_tree_lock(buf) &&
                            btrfs_buffer_uptodate(buf, 0)) {
                                u64 transid =
                                    root->fs_info->running_transaction->transid;
@@ -1490,6 +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;
@@ -1619,6 +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;
 
@@ -1627,15 +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;
 }
@@ -1664,6 +1723,8 @@ static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
        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;
 }
@@ -1753,12 +1814,12 @@ static int noinline find_free_extent(struct btrfs_trans_handle *trans,
                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,
+               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,
+               block_group = btrfs_find_block_group(root,
                                                     trans->block_group,
                                                     search_start, data, 1);
        }
@@ -1880,7 +1941,7 @@ enospc:
        }
        block_group = btrfs_lookup_first_block_group(info, search_start);
        cond_resched();
-       block_group = __btrfs_find_block_group(root, block_group,
+       block_group = btrfs_find_block_group(root, block_group,
                                             search_start, data, 0);
        goto check_failed;
 
@@ -1888,36 +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];
-
-       maybe_lock_mutex(root);
 
        if (data) {
                alloc_profile = info->avail_data_alloc_bits &
@@ -1967,11 +2009,48 @@ again:
        }
        if (ret) {
                printk("allocation failed flags %Lu\n", data);
-       }
-       if (ret) {
                BUG();
-               goto out;
        }
+       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);
@@ -1983,10 +2062,6 @@ again:
        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,
@@ -1994,10 +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);
@@ -2047,6 +2118,51 @@ update_block:
                BUG();
        }
 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;
+}
+/*
+ * 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)
+{
+       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);
+       BUG_ON(ret);
+       ret = __btrfs_alloc_reserved_extent(trans, root, root_objectid,
+                                           ref_generation, owner,
+                                           owner_offset, ins);
+       BUG_ON(ret);
+
        maybe_unlock_mutex(root);
        return ret;
 }
@@ -2095,8 +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;
 }
@@ -2118,8 +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)
@@ -2135,12 +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;
+
+               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;
 }
 
@@ -2283,8 +2405,8 @@ static int noinline walk_down_tree(struct btrfs_trans_handle *trans,
                        mutex_lock(&root->fs_info->alloc_mutex);
 
                        /* we've dropped the lock, double check */
-                       ret = drop_snap_lookup_refcount(root, bytenr,
-                                               blocksize, &refs);
+                       ret = lookup_extent_ref(NULL, root, bytenr, blocksize,
+                                               &refs);
                        BUG_ON(ret);
                        if (refs != 1) {
                                parent = path->nodes[*level];
@@ -2430,6 +2552,10 @@ 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;
@@ -2575,7 +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);
        }
@@ -2613,7 +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;
@@ -2637,12 +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) {
@@ -2691,6 +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);
@@ -2709,6 +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;
@@ -2750,9 +2871,9 @@ 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)) {
+               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;
@@ -2762,7 +2883,7 @@ static int noinline relocate_one_reference(struct btrfs_root *extent_root,
        } 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);
@@ -2784,26 +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;
 }
 
@@ -2943,9 +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) {
@@ -2954,10 +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;
 }
 
@@ -3033,9 +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)
@@ -3085,6 +3232,7 @@ 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);
 
@@ -3092,6 +3240,7 @@ next:
 
                trans = btrfs_start_transaction(tree_root, 1);
                btrfs_commit_transaction(trans, tree_root);
+               mutex_lock(&root->fs_info->alloc_mutex);
                goto again;
        }
 
@@ -3099,7 +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);
@@ -3121,8 +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
         */
@@ -3212,6 +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));
@@ -3239,10 +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;
@@ -3265,6 +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;
 
@@ -3272,6 +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);
@@ -3286,10 +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);