Btrfs: Fix oops and use after free during space balancing
authorChris Mason <chris.mason@oracle.com>
Thu, 4 Jun 2009 19:34:51 +0000 (15:34 -0400)
committerChris Mason <chris.mason@oracle.com>
Thu, 4 Jun 2009 19:41:27 +0000 (15:41 -0400)
The btrfs allocator uses list_for_each to walk the available block
groups when searching for free blocks.  It starts off with a hint
to help find the best block group for a given allocation.

The hint is resolved into a block group, but we don't properly check
to make sure the block group we find isn't in the middle of being
freed due to filesystem shrinking or balancing.  If it is being
freed, the list pointers in it are bogus and can't be trusted.  But,
the code happily goes along and uses them in the list_for_each loop,
leading to all kinds of fun.

The fix used here is to check to make sure the block group we find really
is on the list before we use it.  list_del_init is used when removing
it from the list, so we can do a proper check.

The allocation clustering code has a similar bug where it will trust
the block group in the current free space cluster.  If our allocation
flags have changed (going from single spindle dup to raid1 for example)
because the drives in the FS have changed, we're not allowed to use
the old block group any more.

The fix used here is to check the current cluster against the
current allocation flags.

Signed-off-by: Chris Mason <chris.mason@oracle.com>
fs/btrfs/extent-tree.c

index 3e2c7c7..35af933 100644 (file)
@@ -2622,7 +2622,18 @@ static noinline int find_free_extent(struct btrfs_trans_handle *trans,
                                                       search_start);
                if (block_group && block_group_bits(block_group, data)) {
                        down_read(&space_info->groups_sem);
-                       goto have_block_group;
+                       if (list_empty(&block_group->list) ||
+                           block_group->ro) {
+                               /*
+                                * someone is removing this block group,
+                                * we can't jump into the have_block_group
+                                * target because our list pointers are not
+                                * valid
+                                */
+                               btrfs_put_block_group(block_group);
+                               up_read(&space_info->groups_sem);
+                       } else
+                               goto have_block_group;
                } else if (block_group) {
                        btrfs_put_block_group(block_group);
                }
@@ -2656,6 +2667,13 @@ have_block_group:
                         * people trying to start a new cluster
                         */
                        spin_lock(&last_ptr->refill_lock);
+                       if (last_ptr->block_group &&
+                           (last_ptr->block_group->ro ||
+                           !block_group_bits(last_ptr->block_group, data))) {
+                               offset = 0;
+                               goto refill_cluster;
+                       }
+
                        offset = btrfs_alloc_from_cluster(block_group, last_ptr,
                                                 num_bytes, search_start);
                        if (offset) {
@@ -2681,10 +2699,17 @@ have_block_group:
 
                                last_ptr_loop = 1;
                                search_start = block_group->key.objectid;
+                               /*
+                                * we know this block group is properly
+                                * in the list because
+                                * btrfs_remove_block_group, drops the
+                                * cluster before it removes the block
+                                * group from the list
+                                */
                                goto have_block_group;
                        }
                        spin_unlock(&last_ptr->lock);
-
+refill_cluster:
                        /*
                         * this cluster didn't work out, free it and
                         * start over
@@ -5968,6 +5993,7 @@ int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
 {
        struct btrfs_path *path;
        struct btrfs_block_group_cache *block_group;
+       struct btrfs_free_cluster *cluster;
        struct btrfs_key key;
        int ret;
 
@@ -5979,6 +6005,21 @@ int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
 
        memcpy(&key, &block_group->key, sizeof(key));
 
+       /* make sure this block group isn't part of an allocation cluster */
+       cluster = &root->fs_info->data_alloc_cluster;
+       spin_lock(&cluster->refill_lock);
+       btrfs_return_cluster_to_free_space(block_group, cluster);
+       spin_unlock(&cluster->refill_lock);
+
+       /*
+        * make sure this block group isn't part of a metadata
+        * allocation cluster
+        */
+       cluster = &root->fs_info->meta_alloc_cluster;
+       spin_lock(&cluster->refill_lock);
+       btrfs_return_cluster_to_free_space(block_group, cluster);
+       spin_unlock(&cluster->refill_lock);
+
        path = btrfs_alloc_path();
        BUG_ON(!path);
 
@@ -5988,7 +6029,11 @@ int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
        spin_unlock(&root->fs_info->block_group_cache_lock);
        btrfs_remove_free_space_cache(block_group);
        down_write(&block_group->space_info->groups_sem);
-       list_del(&block_group->list);
+       /*
+        * we must use list_del_init so people can check to see if they
+        * are still on the list after taking the semaphore
+        */
+       list_del_init(&block_group->list);
        up_write(&block_group->space_info->groups_sem);
 
        spin_lock(&block_group->space_info->lock);