Btrfs: fix the number of transaction units needed to remove a block group

We were using only 1 transaction unit when attempting to delete an unused
block group but in reality we need 3 + N units, where N corresponds to the
number of stripes. We were accounting only for the addition of the orphan
item (for the block group's free space cache inode) but we were not
accounting that we need to delete one block group item from the extent
tree, one free space item from the tree of tree roots and N device extent
items from the device tree.

While one unit is not enough, it worked most of the time because for each
single unit we are too pessimistic and assume an entire tree path, with
the highest possible heigth (8), needs to be COWed with eventual node
splits at every possible level in the tree, so there was usually enough
reserved space for removing all the items and adding the orphan item.

However after adding the orphan item, writepages() can by called by the VM
subsystem against the btree inode when we are under memory pressure, which
causes writeback to start for the nodes we COWed before, this forces the
operation to remove the free space item to COW again some (or all of) the
same nodes (in the tree of tree roots). Even without writepages() being
called, we could fail with ENOSPC because these items are located in
multiple trees and one of them might have a higher heigth and require
node/leaf splits at many levels, exhausting all the reserved space before
removing all the items and adding the orphan.

In the kernel 4.0 release, commit 3d84be79 ("Btrfs: fix BUG_ON in
btrfs_orphan_add() when delete unused block group"), we attempted to fix
a BUG_ON due to ENOSPC when trying to add the orphan item by making the
cleaner kthread reserve one transaction unit before attempting to remove
the block group, but this was not enough. We had a couple user reports
still hitting the same BUG_ON after 4.0, like Stefan Priebe's report on
a 4.2-rc6 kernel for example:

    http://www.spinics.net/lists/linux-btrfs/msg46070.html

So fix this by reserving all the necessary units of metadata.
Reported-by: Stefan Priebe <s.priebe@profihost.ag>
Fixes: 3d84be79 ("Btrfs: fix BUG_ON in btrfs_orphan_add() when delete unused block group")
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>

fs/btrfs/ctree.h

@@ -3480,7 +3480,8 @@ int btrfs_make_block_group(struct btrfs_trans_handle *trans,
 			   u64 type, u64 chunk_objectid, u64 chunk_offset,
 			   u64 size);
 struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
-				struct btrfs_fs_info *fs_info);
+				struct btrfs_fs_info *fs_info,
+				const u64 chunk_offset);
 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
 			     struct btrfs_root *root, u64 group_start,
 			     struct extent_map *em);

fs/btrfs/extent-tree.c

@@ -10257,14 +10257,44 @@ int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
 }
 
 struct btrfs_trans_handle *
-btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info)
+btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
+				     const u64 chunk_offset)
 {
+	struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
+	struct extent_map *em;
+	struct map_lookup *map;
+	unsigned int num_items;
+
+	read_lock(&em_tree->lock);
+	em = lookup_extent_mapping(em_tree, chunk_offset, 1);
+	read_unlock(&em_tree->lock);
+	ASSERT(em && em->start == chunk_offset);
+
 	/*
+	 * We need to reserve 3 + N units from the metadata space info in order
+	 * to remove a block group (done at btrfs_remove_chunk() and at
+	 * btrfs_remove_block_group()), which are used for:
+	 *
 	 * 1 unit for adding the free space inode's orphan (located in the tree
 	 * of tree roots).
+	 * 1 unit for deleting the block group item (located in the extent
+	 * tree).
+	 * 1 unit for deleting the free space item (located in tree of tree
+	 * roots).
+	 * N units for deleting N device extent items corresponding to each
+	 * stripe (located in the device tree).
+	 *
+	 * In order to remove a block group we also need to reserve units in the
+	 * system space info in order to update the chunk tree (update one or
+	 * more device items and remove one chunk item), but this is done at
+	 * btrfs_remove_chunk() through a call to check_system_chunk().
 	 */
+	map = (struct map_lookup *)em->bdev;
+	num_items = 3 + map->num_stripes;
+	free_extent_map(em);
+
 	return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
-							   1, 1);
+							   num_items, 1);
 }
 
 /*
@@ -10333,7 +10363,8 @@ void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
 		 * Want to do this before we do anything else so we can recover
 		 * properly if we fail to join the transaction.
 		 */
-		trans = btrfs_start_trans_remove_block_group(fs_info);
+		trans = btrfs_start_trans_remove_block_group(fs_info,
+						     block_group->key.objectid);
 		if (IS_ERR(trans)) {
 			btrfs_dec_block_group_ro(root, block_group);
 			ret = PTR_ERR(trans);

fs/btrfs/volumes.c

@@ -2853,7 +2853,8 @@ static int btrfs_relocate_chunk(struct btrfs_root *root, u64 chunk_offset)
 	if (ret)
 		return ret;
 
-	trans = btrfs_start_trans_remove_block_group(root->fs_info);
+	trans = btrfs_start_trans_remove_block_group(root->fs_info,
+						     chunk_offset);
 	if (IS_ERR(trans)) {
 		ret = PTR_ERR(trans);
 		btrfs_std_error(root->fs_info, ret, NULL);