From: Filipe Manana Date: Wed, 6 Aug 2025 11:11:32 +0000 (+0100) Subject: btrfs: avoid load/store tearing races when checking if an inode was logged X-Git-Url: https://www.infradead.org/git/?a=commitdiff_plain;h=986bf6ed44dff7fbae7b43a0882757ee7f5ba21b;p=users%2Fjedix%2Flinux-maple.git btrfs: avoid load/store tearing races when checking if an inode was logged At inode_logged() we do a couple lockless checks for ->logged_trans, and these are generally safe except the second one in case we get a load or store tearing due to a concurrent call updating ->logged_trans (either at btrfs_log_inode() or later at inode_logged()). In the first case it's safe to compare to the current transaction ID since once ->logged_trans is set the current transaction, we never set it to a lower value. In the second case, where we check if it's greater than zero, we are prone to load/store tearing races, since we can have a concurrent task updating to the current transaction ID with store tearing for example, instead of updating with a single 64 bits write, to update with two 32 bits writes or four 16 bits writes. In that case the reading side at inode_logged() could see a positive value that does not match the current transaction and then return a false negative. Fix this by doing the second check while holding the inode's spinlock, add some comments about it too. Also add the data_race() annotation to the first check to avoid any reports from KCSAN (or similar tools) and comment about it. Fixes: 0f8ce49821de ("btrfs: avoid inode logging during rename and link when possible") Reviewed-by: Boris Burkov Signed-off-by: Filipe Manana Signed-off-by: David Sterba --- diff --git a/fs/btrfs/tree-log.c b/fs/btrfs/tree-log.c index 2a3d9e996e729..7d5d90845ca98 100644 --- a/fs/btrfs/tree-log.c +++ b/fs/btrfs/tree-log.c @@ -3382,15 +3382,32 @@ static int inode_logged(const struct btrfs_trans_handle *trans, struct btrfs_key key; int ret; - if (inode->logged_trans == trans->transid) + /* + * Quick lockless call, since once ->logged_trans is set to the current + * transaction, we never set it to a lower value anywhere else. + */ + if (data_race(inode->logged_trans) == trans->transid) return 1; /* - * If logged_trans is not 0, then we know the inode logged was not logged - * in this transaction, so we can return false right away. + * If logged_trans is not 0 and not trans->transid, then we know the + * inode was not logged in this transaction, so we can return false + * right away. We take the lock to avoid a race caused by load/store + * tearing with a concurrent btrfs_log_inode() call or a concurrent task + * in this function further below - an update to trans->transid can be + * teared into two 32 bits updates for example, in which case we could + * see a positive value that is not trans->transid and assume the inode + * was not logged when it was. */ - if (inode->logged_trans > 0) + spin_lock(&inode->lock); + if (inode->logged_trans == trans->transid) { + spin_unlock(&inode->lock); + return 1; + } else if (inode->logged_trans > 0) { + spin_unlock(&inode->lock); return 0; + } + spin_unlock(&inode->lock); /* * If no log tree was created for this root in this transaction, then