[PATCH net] tcp: md5: fix IPv4-mapped support

[PATCH net] tcp: md5: fix IPv4-mapped support

[Eric Dumazet]

After the blamed commit, IPv4 SYN packets handled
by a dual stack IPv6 socket are dropped, even if
perfectly valid.

$ nstat | grep MD5
TcpExtTCPMD5Failure             5                  0.0

For a dual stack listener, an incoming IPv4 SYN packet
would call tcp_inbound_md5_hash() with @family == AF_INET,
while tp->af_specific is pointing to tcp_sock_ipv6_specific.

Only later when an IPv4-mapped child is created, tp->af_specific
is changed to tcp_sock_ipv6_mapped_specific.

Fixes: 7bbb765b7349 ("net/tcp: Merge TCP-MD5 inbound callbacks")
Reported-by: Brian Vazquez <brianvv@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Dmitry Safonov <dima@arista.com>
Cc: David Ahern <dsahern@kernel.org>
Cc: Leonard Crestez <cdleonard@gmail.com>
---
 net/ipv4/tcp.c | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 002a4a04efbe076ba603d7d42eb85e60d9bf4fb8..766881775abb795c884d048d51c361e805b91989 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -4459,9 +4459,18 @@ tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
 		return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
 	}
 
-	/* check the signature */
-	genhash = tp->af_specific->calc_md5_hash(newhash, hash_expected,
-						 NULL, skb);
+	/* Check the signature.
+	 * To support dual stack listeners, we need to handle
+	 * IPv4-mapped case.
+	 */
+	if (family == AF_INET)
+		genhash = tcp_v4_md5_hash_skb(newhash,
+					      hash_expected,
+					      NULL, skb);
+	else
+		genhash = tp->af_specific->calc_md5_hash(newhash,
+							 hash_expected,
+							 NULL, skb);
 
 	if (genhash || memcmp(hash_location, newhash, 16) != 0) {
 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
-- 
2.37.1.359.gd136c6c3e2-goog

Re: [PATCH net] tcp: md5: fix IPv4-mapped support

[David Ahern]

On 7/26/22 5:57 AM, Eric Dumazet wrote:
> After the blamed commit, IPv4 SYN packets handled
> by a dual stack IPv6 socket are dropped, even if
> perfectly valid.
> 
> $ nstat | grep MD5
> TcpExtTCPMD5Failure             5                  0.0
> 
> For a dual stack listener, an incoming IPv4 SYN packet
> would call tcp_inbound_md5_hash() with @family == AF_INET,
> while tp->af_specific is pointing to tcp_sock_ipv6_specific.
> 
> Only later when an IPv4-mapped child is created, tp->af_specific
> is changed to tcp_sock_ipv6_mapped_specific.
> 
> Fixes: 7bbb765b7349 ("net/tcp: Merge TCP-MD5 inbound callbacks")
> Reported-by: Brian Vazquez <brianvv@google.com>
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Cc: Dmitry Safonov <dima@arista.com>
> Cc: David Ahern <dsahern@kernel.org>
> Cc: Leonard Crestez <cdleonard@gmail.com>
> ---
>  net/ipv4/tcp.c | 15 ++++++++++++---
>  1 file changed, 12 insertions(+), 3 deletions(-)
> 
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index 002a4a04efbe076ba603d7d42eb85e60d9bf4fb8..766881775abb795c884d048d51c361e805b91989 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -4459,9 +4459,18 @@ tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
>  		return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
>  	}
>  
> -	/* check the signature */
> -	genhash = tp->af_specific->calc_md5_hash(newhash, hash_expected,
> -						 NULL, skb);
> +	/* Check the signature.
> +	 * To support dual stack listeners, we need to handle
> +	 * IPv4-mapped case.
> +	 */
> +	if (family == AF_INET)
> +		genhash = tcp_v4_md5_hash_skb(newhash,
> +					      hash_expected,
> +					      NULL, skb);
> +	else
> +		genhash = tp->af_specific->calc_md5_hash(newhash,
> +							 hash_expected,
> +							 NULL, skb);
>  
>  	if (genhash || memcmp(hash_location, newhash, 16) != 0) {
>  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);

We should get v4-mapped cases added to the fcnal-test.sh permutations.

Reviewed-by: David Ahern <dsahern@kernel.org>

Re: [PATCH net] tcp: md5: fix IPv4-mapped support

[Dmitry Safonov]

On 7/26/22 12:57, Eric Dumazet wrote:
> After the blamed commit, IPv4 SYN packets handled
> by a dual stack IPv6 socket are dropped, even if
> perfectly valid.
> 
> $ nstat | grep MD5
> TcpExtTCPMD5Failure             5                  0.0
> 
> For a dual stack listener, an incoming IPv4 SYN packet
> would call tcp_inbound_md5_hash() with @family == AF_INET,
> while tp->af_specific is pointing to tcp_sock_ipv6_specific.
> 
> Only later when an IPv4-mapped child is created, tp->af_specific
> is changed to tcp_sock_ipv6_mapped_specific.

Makes sense. Sorry, I didn't spot it at that moment.

> 
> Fixes: 7bbb765b7349 ("net/tcp: Merge TCP-MD5 inbound callbacks")
> Reported-by: Brian Vazquez <brianvv@google.com>
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Cc: Dmitry Safonov <dima@arista.com>
> Cc: David Ahern <dsahern@kernel.org>
> Cc: Leonard Crestez <cdleonard@gmail.com>

Reviewed-by: Dmitry Safonov <dima@arista.com>

Thanks,
           Dmitry

Re: [PATCH net] tcp: md5: fix IPv4-mapped support

[Eric Dumazet]

On Tue, Jul 26, 2022 at 5:06 PM David Ahern <dsahern@kernel.org> wrote:
>
> On 7/26/22 5:57 AM, Eric Dumazet wrote:
> > After the blamed commit, IPv4 SYN packets handled
> > by a dual stack IPv6 socket are dropped, even if
> > perfectly valid.
> >
> > $ nstat | grep MD5
> > TcpExtTCPMD5Failure             5                  0.0
> >
> > For a dual stack listener, an incoming IPv4 SYN packet
> > would call tcp_inbound_md5_hash() with @family == AF_INET,
> > while tp->af_specific is pointing to tcp_sock_ipv6_specific.
> >
> > Only later when an IPv4-mapped child is created, tp->af_specific
> > is changed to tcp_sock_ipv6_mapped_specific.
> >
> > Fixes: 7bbb765b7349 ("net/tcp: Merge TCP-MD5 inbound callbacks")
> > Reported-by: Brian Vazquez <brianvv@google.com>
> > Signed-off-by: Eric Dumazet <edumazet@google.com>
> > Cc: Dmitry Safonov <dima@arista.com>
> > Cc: David Ahern <dsahern@kernel.org>
> > Cc: Leonard Crestez <cdleonard@gmail.com>
> > ---
> >  net/ipv4/tcp.c | 15 ++++++++++++---
> >  1 file changed, 12 insertions(+), 3 deletions(-)
> >
> > diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> > index 002a4a04efbe076ba603d7d42eb85e60d9bf4fb8..766881775abb795c884d048d51c361e805b91989 100644
> > --- a/net/ipv4/tcp.c
> > +++ b/net/ipv4/tcp.c
> > @@ -4459,9 +4459,18 @@ tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
> >               return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
> >       }
> >
> > -     /* check the signature */
> > -     genhash = tp->af_specific->calc_md5_hash(newhash, hash_expected,
> > -                                              NULL, skb);
> > +     /* Check the signature.
> > +      * To support dual stack listeners, we need to handle
> > +      * IPv4-mapped case.
> > +      */
> > +     if (family == AF_INET)
> > +             genhash = tcp_v4_md5_hash_skb(newhash,
> > +                                           hash_expected,
> > +                                           NULL, skb);
> > +     else
> > +             genhash = tp->af_specific->calc_md5_hash(newhash,
> > +                                                      hash_expected,
> > +                                                      NULL, skb);
> >
> >       if (genhash || memcmp(hash_location, newhash, 16) != 0) {
> >               NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
>
> We should get v4-mapped cases added to the fcnal-test.sh permutations.
>
> Reviewed-by: David Ahern <dsahern@kernel.org>

Although fcnal-test.sh uses ~45 minutes currently :/
Maybe we should make it multi netns and multi threaded to speed up things.

And/or replace various "sleep 1" with more appropriate sync to make
this faster and not flaky in case of system load.

Re: [PATCH net] tcp: md5: fix IPv4-mapped support

[David Ahern]

On 7/26/22 11:57 AM, Eric Dumazet wrote:
> Although fcnal-test.sh uses ~45 minutes currently :/
> Maybe we should make it multi netns and multi threaded to speed up things.
> 
> And/or replace various "sleep 1" with more appropriate sync to make
> this faster and not flaky in case of system load.

There are currently 700+ permutations (800+ if Mike's vrf patch is only
fcnal-test). That's why the script takes a `-t TEST` argument - to only
run a subset.

nettest now has the capability for 1 command to run both client and
server in different namespaces. I have a branch that did the conversion
of fcnal-test.sh; validating the output to ensure no degradation in test
results (not just pass / fail but tests "fail" (negative tests) for the
right reason) took more time than I had. In the end it did not shorten
the test time by any significant margin so lost the motivation to wade
through the output on the before and after.

Re: [PATCH net] tcp: md5: fix IPv4-mapped support

[Leonard Crestez]

On 7/26/22 14:57, Eric Dumazet wrote:
> After the blamed commit, IPv4 SYN packets handled
> by a dual stack IPv6 socket are dropped, even if
> perfectly valid.
> 
> $ nstat | grep MD5
> TcpExtTCPMD5Failure             5                  0.0
> 
> For a dual stack listener, an incoming IPv4 SYN packet
> would call tcp_inbound_md5_hash() with @family == AF_INET,
> while tp->af_specific is pointing to tcp_sock_ipv6_specific.
> 
> Only later when an IPv4-mapped child is created, tp->af_specific
> is changed to tcp_sock_ipv6_mapped_specific.
> 
> Fixes: 7bbb765b7349 ("net/tcp: Merge TCP-MD5 inbound callbacks")
> Reported-by: Brian Vazquez <brianvv@google.com>
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Cc: Dmitry Safonov <dima@arista.com>
> Cc: David Ahern <dsahern@kernel.org>
> Cc: Leonard Crestez <cdleonard@gmail.com>

I had a test in this area for AO and MD5 but it was incorrect (it did 
not actually use an ipv4-mapped-ipv6 address for the ipv6 socket, it 
used an ipv6 wildcard address).

After fixing the test I can confirm that this patch does in fact fix 
something.

https://github.com/cdleonard/tcp-authopt-test/commit/662a6a7e1a818f4581fc0055e821bc1b4c8d04e8

Tested-by: Leonard Crestez <cdleonard@gmail.com>

> ---
>   net/ipv4/tcp.c | 15 ++++++++++++---
>   1 file changed, 12 insertions(+), 3 deletions(-)
> 
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index 002a4a04efbe076ba603d7d42eb85e60d9bf4fb8..766881775abb795c884d048d51c361e805b91989 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -4459,9 +4459,18 @@ tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
>   		return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
>   	}
>   
> -	/* check the signature */
> -	genhash = tp->af_specific->calc_md5_hash(newhash, hash_expected,
> -						 NULL, skb);
> +	/* Check the signature.
> +	 * To support dual stack listeners, we need to handle
> +	 * IPv4-mapped case.
> +	 */
> +	if (family == AF_INET)
> +		genhash = tcp_v4_md5_hash_skb(newhash,
> +					      hash_expected,
> +					      NULL, skb);
> +	else
> +		genhash = tp->af_specific->calc_md5_hash(newhash,
> +							 hash_expected,
> +							 NULL, skb);
>   
>   	if (genhash || memcmp(hash_location, newhash, 16) != 0) {
>   		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);

Re: [PATCH net] tcp: md5: fix IPv4-mapped support

[patchwork-bot+netdevbpf]

Hello:

This patch was applied to netdev/net.git (master)
by Jakub Kicinski <kuba@kernel.org>:

On Tue, 26 Jul 2022 11:57:43 +0000 you wrote:
> After the blamed commit, IPv4 SYN packets handled
> by a dual stack IPv6 socket are dropped, even if
> perfectly valid.
> 
> $ nstat | grep MD5
> TcpExtTCPMD5Failure             5                  0.0
> 
> [...]

Here is the summary with links:
  - [net] tcp: md5: fix IPv4-mapped support
    https://git.kernel.org/netdev/net/c/e62d2e110356

You are awesome, thank you!
-- 
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html