Re: [regression] TCP_MD5SIG on established sockets
From: Eric Dumazet
Date: Tue Jun 30 2020 - 22:30:58 EST
On Tue, Jun 30, 2020 at 7:23 PM Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Tue, Jun 30, 2020 at 07:17:46PM -0700, Eric Dumazet wrote:
> >
> > The main issue of the prior code was the double read of key->keylen in
> > tcp_md5_hash_key(), not that few bytes could change under us.
> >
> > I used smp_rmb() to ease backports, since old kernels had no
> > READ_ONCE()/WRITE_ONCE(), but ACCESS_ONCE() instead.
>
> If it's the double-read that you're protecting against, you should
> just use barrier() and the comment should say so too.
I made this clear in the changelog, do we want comments all over the places ?
Do not get me wrong, we had this bug for years and suddenly this is a
big deal...
MD5 keys are read with RCU protection, and tcp_md5_do_add()
might update in-place a prior key.
Normally, typical RCU updates would allocate a new piece
of memory. In this case only key->key and key->keylen might
be updated, and we do not care if an incoming packet could
see the old key, the new one, or some intermediate value,
since changing the key on a live flow is known to be problematic
anyway.
We only want to make sure that in the case key->keylen
is changed, cpus in tcp_md5_hash_key() wont try to use
uninitialized data, or crash because key->keylen was
read twice to feed sg_init_one() and ahash_request_set_crypt()