On Thu, 23 Feb 2017 09:02:39 -0600
Eric Blake <ebl...@redhat.com> wrote:

> On 02/20/2017 08:40 AM, Greg Kurz wrote:
> > All operations dealing with extended attributes are vulnerable to symlink
> > attacks because they use path-based syscalls which can traverse symbolic
> > links while walking through the dirname part of the path.
> > 
> > The solution is to introduce helpers based on opendir_nofollow(). This
> > calls for "at" versions of the extended attribute syscalls, which don't
> > exist unfortunately. This patch implement them by simulating the "at"
> > behavior with fchdir(). Since the current working directory is process
> > wide, and we don't want to confuse another thread in QEMU, all the work
> > is done in a separate process.  
> 
> Can you emulate *at using /proc/fd/nnn/xyz?  Coreutils was one of the
> early adopters of the power of *at functions, and found that emulation
> of *at via procfs was a LOT more efficient than emulation via fchdir
> (although both emulations still exist in gnulib, since procfs is not
> universal).
> 

Yeah, Stefan suggested this on irc. I had also found a tentative patchset to
implement genuine f*xattrat() calls in the kernel 3 yrs ago, that never got
merged. The author, Florian Weimer, also told me /proc was the way to go.

It looks like we have a consensus :)

Attachment: pgpM17aNYwCrm.pgp
Description: OpenPGP digital signature

Reply via email to