Re: [RFC] vfs: skip extra attributes check on removal for symlinks

[Luis R. Rodriguez]

On Mon, May 07, 2018 at 05:30:55PM -0700, Darrick J. Wong wrote:
> On Tue, May 01, 2018 at 05:45:12PM +0000, Luis R. Rodriguez wrote:
> > On Tue, May 01, 2018 at 10:23:19AM -0700, Darrick J. Wong wrote:
> > > On Thu, Apr 26, 2018 at 04:46:39PM -0700, Luis R. Rodriguez wrote:
> > > > Linux filesystems cannot set extra file attributes (stx_attributes as per
> > > > statx(2)) on a symbolic link. To set extra file attributes you issue
> > > > ioctl(2) with FS_IOC_SETFLAGS, *all* ioctl(2) calls on a symbolic link
> > > > yield EBADF.
> > > > 
> > > > This is because ioctl(2) tries to obtain struct fd from the symbolic link
> > > > file descriptor passed using fdget(), fdget() in turn always returns no
> > > > file set when a file descriptor is open with O_PATH. As per symlink(2)
> > > > O_PATH and O_NOFOLLOW must *always* be used when you want to get the file
> > > > descriptor of a symbolic link, and this holds true for Linux, as such extra
> > > > file attributes cannot possibly be set on symbolic links on Linux.
> > > > 
> > > > Filesystems repair utilities should be updated to detect this as
> > > > corruption and correct this, however, the VFS *does* respect these
> > > > extra attributes on symlinks for removal.
> > > > 
> > > > Since we cannot set these attributes we should special-case the
> > > > immutable/append on delete for symlinks, this would be consistent with
> > > > what we *do* allow on Linux for all filesystems.
> > > 
> > > Ah, ok, so the problem here is that you can't rm an "immutable" symlink
> > > nor can you clear the immutable flag on such a beast, so therefore
> > > ignore the immutable (and append) flags if we're trying to delete a
> > > symlink?
> > 
> > Yup.
> > 
> > > I think we ought to teach the xfs inode verifier to check for
> > > immutable/append symlinks and return error so that we don't end up with
> > > such things in core in the first place,
> > 
> > Agreed. But note that one way to end up with these things is through
> > corruption. Once a user finds this the first signs they'll run into
> > (unless they have the awesome new online scrubber) is they cannot delete
> > some odd file and not know why. And then they'll see they cannot change
> > or remove either the immutable or append flag. The immutable attribute
> > is known to not let you delete the file, but its less known that the
> > append attribute implies the same. Folks would scratch their heads.
> > 
> > Since one cannot *set* these attributes on symlinks on Linux, other than
> > ignoring such attributes perhaps we should warn about it? As the only
> > way you could end up with that is if your filesystem got corrupted.
> > 
> > But without filesystems having a fix for that merged users can't do anything.
> > 
> > > and fix xfs_repair to zap such things.
> > 
> > This is what my patch for xfs_repair does, which is pending review.
> > 
> > But note that special files are handled differently, I explain the logic on the
> > RFC commit log, which is why I suggested splitting up adding a fix for special
> > files as a separate patch.
> > 
> > > That said, for the filesystems that aren't going to check their inodes,
> > > I guess this is a (hackish) way to avoid presenting undeletable gunk in
> > > the fs to the user...
> > 
> > That's why its RFC. What is the right thing to do here?
> 
> Ideally, none of the filesystems should ever feed garbage to the vfs,
> but it's not all that obvious exactly what things the vfs will trip over.
> And knowing that a lot of the filesystems do minimal checking if any, I
> guess I'll just say that...
> 
> ...XFS (and probably ext4) should catch immutable symlinks in the inode
> verifiers so that xfs_iget returns -EFSCORRUPTED.  For the rest of the
> filesystems, it's probably fine to let them delete the symlink since the
> user wants to kill it anyway.

Alrighty, I have this implemented now.

> > My own logic here was since we cannot possibly allow extended attributes on
> > symlinks is to not use them then as well, in this case for delete.
> > 
> > > (Were it up to me I'd make a common vfs_check_inode() to reject
> > > struct inode containing garbage that the vfs won't deal with,
> > 
> > There certainly are cases that we could come up with for the VFS where
> > if such things are found, regardless of the filesystem, we're very sure
> > its a corruption. This is just one example, as you note.
> > 
> > So it is correct that there are two things here:
> > 
> > 1) Do we respect the attribute for symlink on delete
> > 2) Do we warn to users of the fact that the inode is very likely corrupted
> >    regardless of the filesystem?
> 
> But I suppose we could WARN_ON_ONCE to state that we're allowing
> deletion of an immutable symlink that we couldn't possibly have set.

Yes, I think that's better than to allow filesystems to do things even
though the VFS in this case *knows* better.

> Anyway, couple this patch with a second one to fix the xfs verifier and
> I'll be happy.

Groovy, thanks, let's not forget the xfs_repair respective fix :) let me know
if you have any feedback on that.

  Luis