[PATCH] c/r: Add AF_UNIX support (v3)

Oren Laadan orenl at cs.columbia.edu
Tue Jul 7 08:33:19 PDT 2009

Dan Smith wrote:
> OL> But there are two cases: if you are CAP_NET_ADMIN you are allowed
> OL> to go beyond that limit. So you need to add that test too.
> Okay, fair enough.
> OL> And in general, this helps to keep the checks - be it security,
> OL> resource limits, or whatever - in one place, instead of having to
> OL> duplicate code and, more importantly, risk getting out of sync
> OL> with the original checks (e.g., if sock_setsockopt changes).
> But sock_setsockopt() will also set the userlocks flag saying that
> you've specified the buffer size.  At the point at which I currently
> restore the buffers, we've already restored the value specified in the
> checkpoint stream, so we'd need to re-reset it as a special case after
> the call to sock_setsockopt().  Further, to get the "override" case,
> you have to call it with SO_RCVBUFFORCE which fails if you're not
> CAP_SYS_ADMIN.  So, do we try force, then if it fails try SO_RCVBUF,
> then if it fails actually fail?  Since sock_setsockopt() doubles the
> buffer size we get it, do we cut the value we want in half before
> passing it in?
> Doing all that seems like an abuse of sock_setsockopt() to me when the
> alternative is to check CAP_SYS_ADMIN and set the buffer size.
> OL> But we do care, because it is necessary to do the unlink() after
> OL> the bind(), like you do for listening sockets, for this scenario:
> OL> 	s = socket()
> OL> 	bind(s, pathname)
> OL> 	unlink(pathname)
> OL> 				<---- checkpoint/restart
> OL> 	r = socket()
> OL> 	bind(r, pathname)
> OL> The second bind() will succeed on the original system, but will
> OL> fail on the restarted system, unless you do that.
> Not if we don't actually call bind(s) in the unlinked case.  What I
> meant in my previous response is: if we're unlinked, then just fake
> the bind actions but don't actually do the bind()..unlink().  We
> already went over the case where we might unlink() a valid socket
> depending on the order, right?
> OL> BTW, I just looked again at the code, and I'm concerned about:
> OL> +		if (!un->linked) {
> OL> +			struct sockaddr_un *sun =
> OL> +				(struct sockaddr_un *)&h->laddr;
> OL> +			ret = sock_unix_unlink(sun->sun_path);
> OL> +		}
> OL> You need to verify that the address is not abstract, because I'm
> OL> not sure what sock_unix_unlink() would do given the empty
> OL> string. Usually this is filtered higher in the VFS (getname).
> Yep, but luckily that's gone with my recent changes to fake the bind()
> for unlinked sockets instead of actually doing the unlink() :)

Ahh.. and forgot to ask/mention: you do need to call sock_unix_unlink()
before attempting bind(), for the reasons we had discussed earlier
(consider same example as above, checkpoint/restart done before the
unlink(), then restart will otherwise fail).

So you still need to sanitize the file name for that case, no ?


More information about the Containers mailing list