Re: [Qemu-devel] [PATCH] 9pfs: fix vulnerability in openat_dir() and local_unlinkat_common()

[Greg Kurz <groug@kaod.org>]

[-- Attachment #1: Type: text/plain, Size: 1486 bytes --]

On Fri, 3 Mar 2017 12:14:10 -0600
Eric Blake <eblake@redhat.com> wrote:

> On 03/03/2017 11:25 AM, Greg Kurz wrote:
> > We should pass O_NOFOLLOW otherwise openat() will follow symlinks and make
> > QEMU vulnerable.
> > 
> > O_PATH was used as an optimization: the fd returned by openat_dir() is only
> > passed to openat() actually, so we don't really need to reach the underlying
> > filesystem.
> > 
> > O_NOFOLLOW | O_PATH isn't an option: if name is a symlink, openat() will
> > return a fd, forcing us to do some other syscall to detect we have a
> > symlink. Also, O_PATH doesn't exist in glibc 2.13 and older.  
> 
> But the very next use of openat(fd, ) should fail with EBADF if fd is
> not a directory, so you don't need any extra syscalls.  I agree that we
> _need_ O_NOFOLLOW, but I'm not yet convinced that we must avoid O_PATH
> where it works.
> 

You may have a point indeed.

> I'm in the middle of writing a test program to probe kernel behavior and
> demonstrate (at least to myself) whether there are scenarios where
> O_PATH makes it possible to open something where omitting it did not,
> while at the same time validating that O_NOFOLLOW doesn't cause problems
> if a symlink-fd is returned instead of a directory fd, based on our
> subsequent use of that fd in a *at call.
> 

I must leave right now, but please share the result of your experiment.

Thanks for your support on helping to fix 9p, Eric :)

Cheers.

--
Greg

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 181 bytes --]