[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20201105095344.0edecafa@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
Date: Thu, 5 Nov 2020 09:53:44 -0800
From: Jakub Kicinski <kuba@...nel.org>
To: Vinay Kumar Yadav <vinay.yadav@...lsio.com>
Cc: netdev@...r.kernel.org, davem@...emloft.net, borisp@...dia.com,
secdev@...lsio.com
Subject: Re: [PATCH net] net/tls: Fix kernel panic when socket is in TLS ULP
On Thu, 5 Nov 2020 23:20:13 +0530 Vinay Kumar Yadav wrote:
> On 11/5/2020 6:46 AM, Jakub Kicinski wrote:
> > On Tue, 3 Nov 2020 16:17:03 +0530 Vinay Kumar Yadav wrote:
> >> user can initialize tls ulp using setsockopt call on socket
> >> before listen() in case of tls-toe (TLS_HW_RECORD) and same
> >> setsockopt call on connected socket in case of kernel tls (TLS_SW).
> >> In presence of tls-toe devices, TLS ulp is initialized, tls context
> >> is allocated per listen socket and socket is listening at adapter
> >> as well as kernel tcp stack. now consider the scenario, connections
> >> are established in kernel stack.
> >> on every connection close which is established in kernel stack,
> >> it clears tls context which is created on listen socket causing
> >> kernel panic.
> >> Addressed the issue by setting child socket to base (non TLS ULP)
> >> when tls ulp is initialized on parent socket (listen socket).
> >>
> >> Fixes: 76f7164d02d4 ("net/tls: free ctx in sock destruct")
> >> Signed-off-by: Vinay Kumar Yadav <vinay.yadav@...lsio.com>
> >
> > We should prevent from the socket getting into LISTEN state in the
> > first place. Can we make a copy of proto_ops (like tls_sw_proto_ops)
> > and set listen to sock_no_listen?
>
> Once tls-toe (TLS_HW_RECORD) is configured on a socket, listen() call
> from user on same socket will create hash at two places.
What I'm saying is - disallow listen calls on sockets with tls-toe
installed on them. Is that not possible?
> tls_toe_hash() ---> ctx->sk_proto->hash(sk); dev->hash(dev, sk);
>
> when connection establishes, same sock is cloned in case of both
> (connection in adapter or kernel stack).
>
> Please suggest if we can handle it other way?
Powered by blists - more mailing lists