Re: [PATCH v3] vfs: new O_NODE open flag

[Andrew Lutomirski <luto@mit.edu>]

On Mon, Dec 7, 2009 at 9:13 AM, Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:
>> > That is *exactly* the problem, which is clearly what you are missing here.
>>
>> I don't think so, but maybe I'm wrong.  Could you describe your attack
>> scenario in detail then, please?
>
> First obvious attack: get an O_NODE handle to a device you have assigned
> to your ownership
>
>        while(1)
>                fchmod(fd, 0666);
>
> wait for device to unload, reload and be intended for another user
> Race udev to a real open. You have a similar problem with vhangup() and
> ttys.

Huh?  I would've thought that udev would (and already does?), on
device unload, chown to 0:0, then chmod to 0000, then unlink, in which
case that attack doesn't work.

If you mean that someone could have an O_NODE handle open, then the
device unregisters, then, before udev has a chance to do anything at
all, a new device w/ the same major/minor numbers appears, then the
O_NODE handle owner upgrades to a real fd, then we have worse worries:
the attacker could just open the device node on disk without O_NODE,
hard links, or any other funny tricks.  revoke() wouldn't fix that
either.

I'll admit that O_NODE scares me a bit from a security perspective,
but so do hard links and /proc/self/fd in general, and I'm not really
convinced that there are any new attacks here.

Would you be okay with a patch that prevented opening
/proc/self/fd/xxx on O_NODE handles?  I personally don't care about
O_NODE all that much, but I'd like a decent in-kernel AFS
implementation (and a decent revoke() implementation, and especially
the ability to revoke whole filesystems would be really nice too).

--Andy

>
> This cannot happen with the existing kernel because there cannot be an
> open handle when the original device unload occurs[1] and it cannot happen
> with vhangup because the hangup is basically a special case revoke()
> implementation for tty devices.
>
> O_NODE changes the whole lifetime semantics for inodes. It's not
> something you can do casually. pioctl() gets this right although for the
> same reason as path based chmod/chown/etc all get it right, O_NODE breaks
> it all horribly.
>
> Alan
> [1] If you think about it a wait for no references is the same barrier as
> a revoke but a blocking one.
>
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html