Re: Is xt_owner's owner_mt() racy with sock_orphan()? [worse with new TYPESAFE_BY_RCU file lifetime?]

[Phil Sutter <phil@nwl.cc>]

On Wed, Dec 06, 2023 at 10:02:04PM +0100, Jann Horn wrote:
> On Wed, Dec 6, 2023 at 9:42 PM Phil Sutter <phil@nwl.cc> wrote:
> >
> > On Wed, Dec 06, 2023 at 05:28:44PM +0100, Jann Horn wrote:
> > > On Tue, Dec 5, 2023 at 10:40 PM Phil Sutter <phil@nwl.cc> wrote:
> > > > On Tue, Dec 05, 2023 at 06:08:29PM +0100, Jann Horn wrote:
> > > > > On Tue, Dec 5, 2023 at 5:40 PM Jann Horn <jannh@google.com> wrote:
> > > > > >
> > > > > > Hi!
> > > > > >
> > > > > > I think this code is racy, but testing that seems like a pain...
> > > > > >
> > > > > > owner_mt() in xt_owner runs in context of a NF_INET_LOCAL_OUT or
> > > > > > NF_INET_POST_ROUTING hook. It first checks that sk->sk_socket is
> > > > > > non-NULL, then checks that sk->sk_socket->file is non-NULL, then
> > > > > > accesses the ->f_cred of that file.
> > > > > >
> > > > > > I don't see anything that protects this against a concurrent
> > > > > > sock_orphan(), which NULLs out the sk->sk_socket pointer, if we're in
> > > > >
> > > > > Ah, and all the other users of ->sk_socket in net/netfilter/ do it
> > > > > under the sk_callback_lock... so I guess the fix would be to add the
> > > > > same in owner_mt?
> > > >
> > > > Sounds reasonable, although I wonder how likely a socket is to
> > > > orphan while netfilter is processing a packet it just sent.
> > > >
> > > > How about the attached patch? Not sure what hash to put into a Fixes:
> > > > tag given this is a day 1 bug and ipt_owner/ip6t_owner predate git.
> > >
> > > Looks mostly reasonable to me; though I guess it's a bit weird to have
> > > two separate bailout paths for checking whether sk->sk_socket is NULL,
> > > where the first check can race, and the second check uses different
> > > logic for determining the return value; I don't know whether that
> > > actually matters semantically. But I'm not sure how to make it look
> > > nicer either.
> >
> > I find the code pretty confusing since it combines three matches (socket
> > UID, socket GID and socket existence) via binary ops. The second bail
> > disregards socket existence bits, I assumed it was deliberate and thus
> > decided to leave the first part as-is.
> >
> > > I guess you could add a READ_ONCE() around the first read to signal
> > > that that's a potentially racy read, but I don't feel strongly about
> > > that.
> >
> > Is this just annotation or do you see a practical effect of using
> > READ_ONCE() there?
> 
> I mostly just meant that as an annotation. My understanding is that in
> theory, racy reads can cause the compiler to do some terrible things
> to your code (https://lore.kernel.org/all/CAG48ez2nFks+yN1Kp4TZisso+rjvv_4UW0FTo8iFUd4Qyq1qDw@mail.gmail.com/),

Thanks for the pointer, this was an educational read!

> but that's almost certainly not going to happen here.

At least it's not a switch on a value in user-controlled memory. ;)

> (Well, I guess doing a READ_ONCE() at one side without doing
> WRITE_ONCE() on the other side is also unclean...)

For the annotation aspect it won't matter. Though since it will merely
improve reliability of that check in the given corner-case which is an
unreliable situation in the first place, I'd just leave it alone and
hope for the code to be replaced by the one in nft_meta.c eventually.

Thanks, Phil