Date:   Fri, 24 Feb 2023 08:57:30 +0100
From:   Sabrina Dubroca <sd@queasysnail.net>
To:     Hangyu Hua <hbh25y@gmail.com>
Cc:     borisp@nvidia.com, john.fastabend@gmail.com, kuba@kernel.org,
        davem@davemloft.net, edumazet@google.com, pabeni@redhat.com,
        netdev@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] net: tls: fix possible info leak in
 tls_set_device_offload()
Message-ID: <Y/ht6gQL+u6fj3dG@hog>
References: <20230223090508.443157-1-hbh25y@gmail.com>
 <Y/dK6OoNpYswIqrD@hog>
 <310391ea-7c71-395e-5dcb-b0a983e6fc93@gmail.com>
 <04c4d6ee-f893-5248-26cf-2c6d1c9b3aa5@gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <04c4d6ee-f893-5248-26cf-2c6d1c9b3aa5@gmail.com>
Precedence: bulk

2023-02-24, 11:33:29 +0800, Hangyu Hua wrote:
> On 24/2/2023 11:07, Hangyu Hua wrote:
> > On 23/2/2023 19:15, Sabrina Dubroca wrote:
> > > 2023-02-23, 17:05:08 +0800, Hangyu Hua wrote:
> > > > After tls_set_device_offload() fails, we enter tls_set_sw_offload(). But
> > > > tls_set_sw_offload can't set cctx->iv and cctx->rec_seq to NULL
> > > > if it fails
> > > > before kmalloc cctx->iv. This may cause info leak when we call
> > > > do_tls_getsockopt_conf().
> > > 
> > > Is there really an issue here?
> > > 
> > > If both tls_set_device_offload and tls_set_sw_offload fail,
> > > do_tls_setsockopt_conf will clear crypto_{send,recv} from the context.
> > > Then the TLS_CRYPTO_INFO_READY in do_tls_getsockopt_conf will fail, so
> > > we won't try to access iv or rec_seq.
> > > 
> > 
> > My bad. I forget memzero_explicit. Then this is harmless. But I still
> > think it is better to set them to NULL like tls_set_sw_offload's error
> > path because we don't know there are another way to do this(I will
> > change the commit log). What do you think?

Yes, I guess for consistency between functions it would be ok.

> Like a rare case, there is a race condition between
> do_tls_getsockopt_conf and do_tls_setsockopt_conf while the previous
> condition is met. TLS_CRYPTO_INFO_READY(crypto_info) is not
> protected by lock_sock in do_tls_getsockopt_conf. It's just too
> difficult to satisfy both conditions at the same time.

Ugh, thanks for noticing this. We should move the lock_sock in
getsockopt before TLS_CRYPTO_INFO_READY. Do you want to write that
patch?

Thanks.

-- 
Sabrina