On 9.12.20 г. 22:17 ч., Josef Bacik wrote:
While running btrfs/011 in a loop I would often ASSERT() while trying to add a new free space entry that already existed, or get an -EEXIST while adding a new block to the extent tree, which is another indication of double allocation.
This occurs because when we do the free space tree population, we create the new root and then populate the tree and commit the transaction. The problem is when you create a new root, the root node and commit root node are the same. This means that caching a block group before the transaction is committed can race with other operations modifying the free space tree, and thus you can get double adds and other sort of
FST creation happens during mount so what would initiate block group caching at that time, the race scenario should be better described in the change log. E.g. what those other operations might be, considering we are in mount ?
shenanigans. This is only a problem for the first transaction, once we've committed the transaction we created the free space tree in we're OK to use the free space tree to cache block groups.
Fix this by marking the fs_info as unsafe to load the free space tree, and fall back on the old slow method. We could be smarter than this, for example caching the block group while we're populating the free space tree, but since this is a serious problem I've opted for the simplest solution.
cc: stable@vger.kernel.org Fixes: a5ed91828518 ("Btrfs: implement the free space B-tree") Signed-off-by: Josef Bacik josef@toxicpanda.com
fs/btrfs/block-group.c | 11 ++++++++++- fs/btrfs/ctree.h | 3 +++ fs/btrfs/free-space-tree.c | 9 ++++++++- 3 files changed, 21 insertions(+), 2 deletions(-)
?<snip>
/* diff --git a/fs/btrfs/free-space-tree.c b/fs/btrfs/free-space-tree.c index e33a65bd9a0c..8fbda221f4b5 100644 --- a/fs/btrfs/free-space-tree.c +++ b/fs/btrfs/free-space-tree.c @@ -1150,6 +1150,7 @@ int btrfs_create_free_space_tree(struct btrfs_fs_info *fs_info) return PTR_ERR(trans); set_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
- set_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags); free_space_root = btrfs_create_tree(trans, BTRFS_FREE_SPACE_TREE_OBJECTID); if (IS_ERR(free_space_root)) {
@@ -1171,8 +1172,14 @@ int btrfs_create_free_space_tree(struct btrfs_fs_info *fs_info) btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE); btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE_VALID); clear_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
- ret = btrfs_commit_transaction(trans);
- return btrfs_commit_transaction(trans);
- /*
* Now that we've committed the transaction any reading of our commit* root will be safe, so we can caching from the free space tree now.*/- clear_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);
- return ret;
I guess you can't simply move the clearing of the BTRFS_FS_CREATING_FREE_SPACE_TREE after the commit since it blocks delayed refs running.
abort: clear_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
Shouldn't the new flag be cleared on abort ?