Re: [PATCH] btrfs: Defer setting new inode mode until after do_set_acl succeeds

[David Sterba <dsterba@suse.cz>]

On Thu, Mar 21, 2019 at 11:27:03AM +0200, Nikolay Borisov wrote:
> Currently a reference to inode->i_mode is passed directly to
> posix_acl_update_mode when setting an ACL which results in the inode's
> mode always being changed. In case of errors (e.g. in do_set_acl or
> even starting a transaction) the old mode needs to be re-assigned to
> ->i_mode. This mode recovery is done only in case do_set_acl fails,
> which leads to buggy behavior in case btrfs_start_transaction fails.
> 
> Fix it by simply setting the new mode to a temporary variable which is
> assigned to inode->i_mode's only when do_set_acl succeeds. This covers
> both failure cases explained above.
> 
> Fixes: db0f220e98eb ("btrfs: start transaction in btrfs_set_acl")
> Signed-off-by: Nikolay Borisov <nborisov@suse.com>
> ---
>  fs/btrfs/acl.c | 10 ++++------
>  1 file changed, 4 insertions(+), 6 deletions(-)
> 
> diff --git a/fs/btrfs/acl.c b/fs/btrfs/acl.c
> index b722866e1442..a0cfd2049ea5 100644
> --- a/fs/btrfs/acl.c
> +++ b/fs/btrfs/acl.c
> @@ -99,7 +99,6 @@ static int do_set_acl(struct btrfs_trans_handle *trans, struct inode *inode,
>  	}
>  
>  	ret = btrfs_setxattr(trans, inode, name, value, size, 0);
> -
>  out:
>  	kfree(value);
>  
> @@ -112,12 +111,12 @@ static int do_set_acl(struct btrfs_trans_handle *trans, struct inode *inode,
>  int btrfs_set_acl(struct inode *inode, struct posix_acl *acl, int type)
>  {
>  	int ret;
> -	umode_t old_mode = inode->i_mode;
> +	umode_t mode;
>  	struct btrfs_trans_handle *trans;
>  	struct btrfs_root *root = BTRFS_I(inode)->root;
>  
>  	if (type == ACL_TYPE_ACCESS && acl) {
> -		ret = posix_acl_update_mode(inode, &inode->i_mode, &acl);
> +		ret = posix_acl_update_mode(inode, &mode, &acl);
>  		if (ret)
>  			return ret;
>  	}
> @@ -127,9 +126,8 @@ int btrfs_set_acl(struct inode *inode, struct posix_acl *acl, int type)
>  		return PTR_ERR(trans);
>  
>  	ret = do_set_acl(trans, inode, acl, type);
> -	if (ret) {
> -		inode->i_mode = old_mode;
> -	} else {
> +	if (!ret) {
> +		inode->i_mode = mode;

I like this flow much better, ie. update i_mod only if it's correct,
otherwise it's a no-op. Then we don't have to revert the change on each
possible error.

The patch that causes it is still in misc-next so the hash is not stable
and we can't use it for reference. So I suggest to fold the single
i_mode after transaction to "btrfs: start transaction in btrfs_set_acl"
and your patch can rework the i_mode handling.

We'll see when you see V2 if this could work or just go with a single
patch.