Re: [PATCH 0/3] vfs: plug some holes involving LAST_BIND symlinks and file bind mounts (try #5)

[ebiederm@xmission.com (Eric W. Biederman)]

Jeff Layton <jlayton@redhat.com> writes:

> There are a few situations where a lookup can end up returning a dentry
> without revalidating it, and without checking whether the calling
> process has permissions to access it. Two situations identified so far
> are:
>
> 1) LAST_BIND symlinks (such as those under /proc/<pid>)
>
> 2) file bind mounts
>
> This patchset is intended to fix this by forcing revalidation of the
> returned dentries at appropriate locations.
>
> In the case of LAST_BIND symlinks it also adds a check to verify that
> the target of the symlink is accessible by the current process by
> walking mounts and dentries back up to the root and checking permission
> on each inode.
>
> This set fixes the reproducers I have (including the reproducer that
> Pavel provided for the permissions bypass). It's still pretty rough
> though and I expect that it'll need revision. At this point, I'm mainly
> looking to get these questions answered:
>
> 1) what should we do if these dentries are found to be invalid? Is it ok
> to d_invalidate them? Or is that likely to break something (particularly
> in the case of file bind mounts)?

The normal sequence in do_revalidate should be safe.  In practice what we
should see is d_drop().  If we access the dentries via another path today
we already go through d_revalidate.  It is only the reference count on
the dentry that keeps them alive and working.  The cases I have looked
at for distributed filesystems have to call d_drop themselves so I don't
know if it would add anything if the vfs called d_revalidate.  Especially
since FS_REVAL_DOT doesn't have that logic.

> 2) I'm using FS_REVAL_DOT as an indicator of whether to force a
> d_revalidate. I think that it's appropriate to key off of that flag, but
> we may want to rename it (maybe FS_FORCE_D_REVAL ?).

Perhaps FS_ALWAYS_REVAL.    I don't think it makes much of
a difference either way.  I expect a rename should come after we fix
nfsv4 so there is a chance at pushing the fixes back to stable.

> 3) is check_path_accessible racy? It seems to work, but something
> doesn't seem quite right with this approach. Is this defeatable somehow?
> Could a rename of one of the intermediate path components cause
> problems?

check_path_accessible seems pretty horrible.   If a process is running
inside of a subdirectory it doesn't have permissions to access, say
a chroot, /proc/self/fd/XXX becomes completely unusable.

Eric