[PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[Xin Long]

Currently, udp v6 socket can not process v4 packets with UDP GRO, as
udp_encap_needed_key is not increased when udp_tunnel_encap_enable()
is called for v6 socket.

This patchset is to increase it and remove the unnecessary code in
bareudp in Patch 1/2, and improve rxrpc encap_enable by calling
udp_tunnel_encap_enable().

Xin Long (2):
  udp: call udp_encap_enable for v6 sockets when enabling encap
  rxrpc: call udp_tunnel_encap_enable in rxrpc_open_socket

 drivers/net/bareudp.c    | 6 ------
 include/net/udp.h        | 1 +
 include/net/udp_tunnel.h | 3 +--
 net/ipv4/udp.c           | 6 ++++++
 net/ipv6/udp.c           | 4 +++-
 net/rxrpc/local_object.c | 6 +-----
 6 files changed, 12 insertions(+), 14 deletions(-)

-- 
2.1.0

[PATCHv4 net-next 1/2] udp: call udp_encap_enable for v6 sockets when enabling encap

[Xin Long]

When enabling encap for a ipv6 socket without udp_encap_needed_key
increased, UDP GRO won't work for v4 mapped v6 address packets as
sk will be NULL in udp4_gro_receive().

This patch is to enable it by increasing udp_encap_needed_key for
v6 sockets in udp_tunnel_encap_enable(), and correspondingly
decrease udp_encap_needed_key in udpv6_destroy_sock().

v1->v2:
  - add udp_encap_disable() and export it.
v2->v3:
  - add the change for rxrpc and bareudp into one patch, as Alex
    suggested.
v3->v4:
  - move rxrpc part to another patch.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
---
 drivers/net/bareudp.c    | 6 ------
 include/net/udp.h        | 1 +
 include/net/udp_tunnel.h | 3 +--
 net/ipv4/udp.c           | 6 ++++++
 net/ipv6/udp.c           | 4 +++-
 5 files changed, 11 insertions(+), 9 deletions(-)

diff --git a/drivers/net/bareudp.c b/drivers/net/bareudp.c
index 1b8f597..7511bca 100644
--- a/drivers/net/bareudp.c
+++ b/drivers/net/bareudp.c
@@ -240,12 +240,6 @@ static int bareudp_socket_create(struct bareudp_dev *bareudp, __be16 port)
 	tunnel_cfg.encap_destroy = NULL;
 	setup_udp_tunnel_sock(bareudp->net, sock, &tunnel_cfg);
 
-	/* As the setup_udp_tunnel_sock does not call udp_encap_enable if the
-	 * socket type is v6 an explicit call to udp_encap_enable is needed.
-	 */
-	if (sock->sk->sk_family == AF_INET6)
-		udp_encap_enable();
-
 	rcu_assign_pointer(bareudp->sock, sock);
 	return 0;
 }
diff --git a/include/net/udp.h b/include/net/udp.h
index 877832b..1e7b6cd 100644
--- a/include/net/udp.h
+++ b/include/net/udp.h
@@ -467,6 +467,7 @@ void udp_init(void);
 
 DECLARE_STATIC_KEY_FALSE(udp_encap_needed_key);
 void udp_encap_enable(void);
+void udp_encap_disable(void);
 #if IS_ENABLED(CONFIG_IPV6)
 DECLARE_STATIC_KEY_FALSE(udpv6_encap_needed_key);
 void udpv6_encap_enable(void);
diff --git a/include/net/udp_tunnel.h b/include/net/udp_tunnel.h
index 282d10e..afc7ce7 100644
--- a/include/net/udp_tunnel.h
+++ b/include/net/udp_tunnel.h
@@ -181,9 +181,8 @@ static inline void udp_tunnel_encap_enable(struct socket *sock)
 #if IS_ENABLED(CONFIG_IPV6)
 	if (sock->sk->sk_family == PF_INET6)
 		ipv6_stub->udpv6_encap_enable();
-	else
 #endif
-		udp_encap_enable();
+	udp_encap_enable();
 }
 
 #define UDP_TUNNEL_NIC_MAX_TABLES	4
diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index 69ea765..48208fb 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -596,6 +596,12 @@ void udp_encap_enable(void)
 }
 EXPORT_SYMBOL(udp_encap_enable);
 
+void udp_encap_disable(void)
+{
+	static_branch_dec(&udp_encap_needed_key);
+}
+EXPORT_SYMBOL(udp_encap_disable);
+
 /* Handler for tunnels with arbitrary destination ports: no socket lookup, go
  * through error handlers in encapsulations looking for a match.
  */
diff --git a/net/ipv6/udp.c b/net/ipv6/udp.c
index b9f3dfd..d754292 100644
--- a/net/ipv6/udp.c
+++ b/net/ipv6/udp.c
@@ -1608,8 +1608,10 @@ void udpv6_destroy_sock(struct sock *sk)
 			if (encap_destroy)
 				encap_destroy(sk);
 		}
-		if (up->encap_enabled)
+		if (up->encap_enabled) {
 			static_branch_dec(&udpv6_encap_needed_key);
+			udp_encap_disable();
+		}
 	}
 
 	inet6_destroy_sock(sk);
-- 
2.1.0

[PATCHv4 net-next 2/2] rxrpc: call udp_tunnel_encap_enable in rxrpc_open_socket

[Xin Long]

When doing encap_enable/increasing encap_needed_key, up->encap_enabled
is not set in rxrpc_open_socket(), and it will cause encap_needed_key
not being decreased in udpv6_destroy_sock().

This patch is to improve it by just calling udp_tunnel_encap_enable()
where it increases both UDP and UDPv6 encap_needed_key and sets
up->encap_enabled.

Signed-off-by: Xin Long <lucien.xin@gmail.com>
---
 net/rxrpc/local_object.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/net/rxrpc/local_object.c b/net/rxrpc/local_object.c
index 8c28810..93e05d2 100644
--- a/net/rxrpc/local_object.c
+++ b/net/rxrpc/local_object.c
@@ -135,11 +135,7 @@ static int rxrpc_open_socket(struct rxrpc_local *local, struct net *net)
 	udp_sk(usk)->gro_receive = NULL;
 	udp_sk(usk)->gro_complete = NULL;
 
-	udp_encap_enable();
-#if IS_ENABLED(CONFIG_AF_RXRPC_IPV6)
-	if (local->srx.transport.family == AF_INET6)
-		udpv6_encap_enable();
-#endif
+	udp_tunnel_encap_enable(local->socket);
 	usk->sk_error_report = rxrpc_error_report;
 
 	/* if a local address was supplied then bind it */
-- 
2.1.0

Re: [PATCHv4 net-next 2/2] rxrpc: call udp_tunnel_encap_enable in rxrpc_open_socket

[David Howells]

Xin Long <lucien.xin@gmail.com> wrote:

> -	udp_encap_enable();
> -#if IS_ENABLED(CONFIG_AF_RXRPC_IPV6)
> -	if (local->srx.transport.family == AF_INET6)
> -		udpv6_encap_enable();
> -#endif
> +	udp_tunnel_encap_enable(local->socket);

You need this too:

	--- a/net/rxrpc/local_object.c
	+++ b/net/rxrpc/local_object.c
	@@ -16,6 +16,7 @@
	 #include <linux/hashtable.h>
	 #include <net/sock.h>
	 #include <net/udp.h>
	+#include <net/udp_tunnel.h>
	 #include <net/af_rxrpc.h>
	 #include "ar-internal.h"

With that, it seems to work still:

	Acked-and-tested-by: David Howells <dhowells@redhat.com>

David

Re: [PATCHv4 net-next 1/2] udp: call udp_encap_enable for v6 sockets when enabling encap

[Willem de Bruijn]

On Tue, Jan 26, 2021 at 5:59 PM Xin Long <lucien.xin@gmail.com> wrote:
>
> When enabling encap for a ipv6 socket without udp_encap_needed_key
> increased, UDP GRO won't work for v4 mapped v6 address packets as
> sk will be NULL in udp4_gro_receive().
>
> This patch is to enable it by increasing udp_encap_needed_key for
> v6 sockets in udp_tunnel_encap_enable(), and correspondingly
> decrease udp_encap_needed_key in udpv6_destroy_sock().
>
> v1->v2:
>   - add udp_encap_disable() and export it.
> v2->v3:
>   - add the change for rxrpc and bareudp into one patch, as Alex
>     suggested.
> v3->v4:
>   - move rxrpc part to another patch.
>
> Signed-off-by: Xin Long <lucien.xin@gmail.com>

Acked-by: Willem de Bruijn <willemb@google.com>

Re: [PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[Xin Long]

Hi, Jakub,

I saw the state of this patchset is still new, should I repost it?

On Tue, Jan 26, 2021 at 1:10 PM Xin Long <lucien.xin@gmail.com> wrote:
>
> Currently, udp v6 socket can not process v4 packets with UDP GRO, as
> udp_encap_needed_key is not increased when udp_tunnel_encap_enable()
> is called for v6 socket.
>
> This patchset is to increase it and remove the unnecessary code in
> bareudp in Patch 1/2, and improve rxrpc encap_enable by calling
> udp_tunnel_encap_enable().
>
> Xin Long (2):
>   udp: call udp_encap_enable for v6 sockets when enabling encap
>   rxrpc: call udp_tunnel_encap_enable in rxrpc_open_socket
>
>  drivers/net/bareudp.c    | 6 ------
>  include/net/udp.h        | 1 +
>  include/net/udp_tunnel.h | 3 +--
>  net/ipv4/udp.c           | 6 ++++++
>  net/ipv6/udp.c           | 4 +++-
>  net/rxrpc/local_object.c | 6 +-----
>  6 files changed, 12 insertions(+), 14 deletions(-)
>
> --
> 2.1.0
>

Re: [PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[David Howells]

Xin Long <lucien.xin@gmail.com> wrote:

> I saw the state of this patchset is still new, should I repost it?

It needs a fix in patch 2 (see my response to that patch).

Thanks,
David

Re: [PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[Xin Long]

On Wed, Feb 3, 2021 at 4:00 PM David Howells <dhowells@redhat.com> wrote:
>
> Xin Long <lucien.xin@gmail.com> wrote:
>
> > I saw the state of this patchset is still new, should I repost it?
>
> It needs a fix in patch 2 (see my response to that patch).
>
Sorry, my mistake, I forgot to enable rxrpc when building kernel.
Will repost, Thank you.

BTW, I'm also thinking to use udp_sock_create(), the only problem I can
see is it may not do bind() in rxrpc_open_socket(), is that true? or we
can actually bind to some address when a local address is not supplied?

Re: [PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[David Howells]

Xin Long <lucien.xin@gmail.com> wrote:

> BTW, I'm also thinking to use udp_sock_create(), the only problem I can
> see is it may not do bind() in rxrpc_open_socket(), is that true? or we
> can actually bind to some address when a local address is not supplied?

If a local address isn't explicitly bound to the AF_RXRPC socket, binding the
UDP socket to a random local port is fine.  In fact, sometimes I want to
explicitly bind an rxrpc server socket to a random port.  See fs/afs/rxrpc.c
function afs_open_socket():

	/* bind the callback manager's address to make this a server socket */
	memset(&srx, 0, sizeof(srx));
	srx.srx_family			= AF_RXRPC;
	srx.srx_service			= CM_SERVICE;
	srx.transport_type		= SOCK_DGRAM;
	srx.transport_len		= sizeof(srx.transport.sin6);
	srx.transport.sin6.sin6_family	= AF_INET6;
	srx.transport.sin6.sin6_port	= htons(AFS_CM_PORT);
	...
	ret = kernel_bind(socket, (struct sockaddr *) &srx, sizeof(srx));
	if (ret == -EADDRINUSE) {
		srx.transport.sin6.sin6_port = 0;

		^^^ That's hoping to get a random port bound.

		ret = kernel_bind(socket, (struct sockaddr *) &srx, sizeof(srx));
	}
	if (ret < 0)
		goto error_2;

The client cache manager server socket here is used to receive notifications
back from the fileserver.  There's a standard port (7001) for the service, but
if that's in use, we can use any other port.  The fileserver grabs the source
port from incoming RPC requests - and then uses that when sending 3rd-party
change notifications back.

If you could arrange for a random port to be assigned in such a case (and
indicated back to the caller), that would be awesome.  Possibly I just don't
need to actually use bind in this case.

David

Re: [PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[Xin Long]

[-- Attachment #1: Type: text/plain, Size: 2755 bytes --]

On Wed, Feb 3, 2021 at 5:14 PM David Howells <dhowells@redhat.com> wrote:
>
> Xin Long <lucien.xin@gmail.com> wrote:
>
> > BTW, I'm also thinking to use udp_sock_create(), the only problem I can
> > see is it may not do bind() in rxrpc_open_socket(), is that true? or we
> > can actually bind to some address when a local address is not supplied?
>
> If a local address isn't explicitly bound to the AF_RXRPC socket, binding the
> UDP socket to a random local port is fine.  In fact, sometimes I want to
> explicitly bind an rxrpc server socket to a random port.  See fs/afs/rxrpc.c
> function afs_open_socket():
>
>         /* bind the callback manager's address to make this a server socket */
>         memset(&srx, 0, sizeof(srx));
>         srx.srx_family                  = AF_RXRPC;
>         srx.srx_service                 = CM_SERVICE;
>         srx.transport_type              = SOCK_DGRAM;
>         srx.transport_len               = sizeof(srx.transport.sin6);
>         srx.transport.sin6.sin6_family  = AF_INET6;
>         srx.transport.sin6.sin6_port    = htons(AFS_CM_PORT);
>         ...
>         ret = kernel_bind(socket, (struct sockaddr *) &srx, sizeof(srx));
>         if (ret == -EADDRINUSE) {
>                 srx.transport.sin6.sin6_port = 0;
>
>                 ^^^ That's hoping to get a random port bound.
>
>                 ret = kernel_bind(socket, (struct sockaddr *) &srx, sizeof(srx));
>         }
>         if (ret < 0)
>                 goto error_2;
>
> The client cache manager server socket here is used to receive notifications
> back from the fileserver.  There's a standard port (7001) for the service, but
> if that's in use, we can use any other port.  The fileserver grabs the source
> port from incoming RPC requests - and then uses that when sending 3rd-party
> change notifications back.
>
> If you could arrange for a random port to be assigned in such a case (and
> indicated back to the caller), that would be awesome.  Possibly I just don't
> need to actually use bind in this case.
>
The patch is attached (based on this patch):

+       udp_conf.family = srx->transport.family;
+       if (udp_conf.family == AF_INET) {
+               udp_conf.local_ip = srx->transport.sin.sin_addr;
+               udp_conf.local_udp_port = srx->transport.sin.sin_port;
+       } else {
+               udp_conf.local_ip6 = srx->transport.sin6.sin6_addr;
+               udp_conf.local_udp_port = srx->transport.sin6.sin6_port;
+       }
+       ret = udp_sock_create(net, &udp_conf, &local->socket);

I think this will work well. When the socket is not bound,
srx->transport.sin.sin(6)_addr/sin(6)_port are zero. It'll
bind to a random port in udp_sock_create().

BTW: do you have any testing for this?

Thanks.

[-- Attachment #2: 0001-rxrpc-use-udp-tunnel-APIs-instead-of-open-code-in-rx.patch --]
[-- Type: application/octet-stream, Size: 4027 bytes --]

From 5748d6312ad8025f6e347a0b067b204320b1c770 Mon Sep 17 00:00:00 2001
Message-Id: <5748d6312ad8025f6e347a0b067b204320b1c770.1612359043.git.lucien.xin@gmail.com>
From: Xin Long <lucien.xin@gmail.com>
Date: Wed, 3 Feb 2021 08:25:28 -0500
Subject: [PATCH net-next] rxrpc: use udp tunnel APIs instead of open code in
 rxrpc_open_socket

Signed-off-by: Xin Long <lucien.xin@gmail.com>
---
 net/rxrpc/local_object.c | 69 ++++++++++++++--------------------------
 1 file changed, 24 insertions(+), 45 deletions(-)

diff --git a/net/rxrpc/local_object.c b/net/rxrpc/local_object.c
index 33b49367d575..546fd237a649 100644
--- a/net/rxrpc/local_object.c
+++ b/net/rxrpc/local_object.c
@@ -107,54 +107,42 @@ static struct rxrpc_local *rxrpc_alloc_local(struct rxrpc_net *rxnet,
  */
 static int rxrpc_open_socket(struct rxrpc_local *local, struct net *net)
 {
+	struct udp_tunnel_sock_cfg tuncfg = {NULL};
+	struct sockaddr_rxrpc *srx = &local->srx;
+	struct udp_port_cfg udp_conf = {0};
 	struct sock *usk;
 	int ret;
 
 	_enter("%p{%d,%d}",
-	       local, local->srx.transport_type, local->srx.transport.family);
-
-	/* create a socket to represent the local endpoint */
-	ret = sock_create_kern(net, local->srx.transport.family,
-			       local->srx.transport_type, 0, &local->socket);
+	       local, srx->transport_type, srx->transport.family);
+
+	udp_conf.family = srx->transport.family;
+	if (udp_conf.family == AF_INET) {
+		udp_conf.local_ip = srx->transport.sin.sin_addr;
+		udp_conf.local_udp_port = srx->transport.sin.sin_port;
+	} else {
+		udp_conf.local_ip6 = srx->transport.sin6.sin6_addr;
+		udp_conf.local_udp_port = srx->transport.sin6.sin6_port;
+	}
+	ret = udp_sock_create(net, &udp_conf, &local->socket);
 	if (ret < 0) {
 		_leave(" = %d [socket]", ret);
 		return ret;
 	}
 
+	tuncfg.encap_type = UDP_ENCAP_RXRPC;
+	tuncfg.encap_rcv = rxrpc_input_packet;
+	tuncfg.sk_user_data = local;
+	setup_udp_tunnel_sock(net, local->socket, &tuncfg);
+
 	/* set the socket up */
 	usk = local->socket->sk;
-	inet_sk(usk)->mc_loop = 0;
-
-	/* Enable CHECKSUM_UNNECESSARY to CHECKSUM_COMPLETE conversion */
-	inet_inc_convert_csum(usk);
-
-	rcu_assign_sk_user_data(usk, local);
-
-	udp_sk(usk)->encap_type = UDP_ENCAP_RXRPC;
-	udp_sk(usk)->encap_rcv = rxrpc_input_packet;
-	udp_sk(usk)->encap_destroy = NULL;
-	udp_sk(usk)->gro_receive = NULL;
-	udp_sk(usk)->gro_complete = NULL;
-
-	udp_tunnel_encap_enable(local->socket);
 	usk->sk_error_report = rxrpc_error_report;
 
-	/* if a local address was supplied then bind it */
-	if (local->srx.transport_len > sizeof(sa_family_t)) {
-		_debug("bind");
-		ret = kernel_bind(local->socket,
-				  (struct sockaddr *)&local->srx.transport,
-				  local->srx.transport_len);
-		if (ret < 0) {
-			_debug("bind failed %d", ret);
-			goto error;
-		}
-	}
-
-	switch (local->srx.transport.family) {
+	switch (srx->transport.family) {
 	case AF_INET6:
 		/* we want to receive ICMPv6 errors */
-		ip6_sock_set_recverr(local->socket->sk);
+		ip6_sock_set_recverr(usk);
 
 		/* Fall through and set IPv4 options too otherwise we don't get
 		 * errors from IPv4 packets sent through the IPv6 socket.
@@ -162,13 +150,13 @@ static int rxrpc_open_socket(struct rxrpc_local *local, struct net *net)
 		fallthrough;
 	case AF_INET:
 		/* we want to receive ICMP errors */
-		ip_sock_set_recverr(local->socket->sk);
+		ip_sock_set_recverr(usk);
 
 		/* we want to set the don't fragment bit */
-		ip_sock_set_mtu_discover(local->socket->sk, IP_PMTUDISC_DO);
+		ip_sock_set_mtu_discover(usk, IP_PMTUDISC_DO);
 
 		/* We want receive timestamps. */
-		sock_enable_timestamps(local->socket->sk);
+		sock_enable_timestamps(usk);
 		break;
 
 	default:
@@ -177,15 +165,6 @@ static int rxrpc_open_socket(struct rxrpc_local *local, struct net *net)
 
 	_leave(" = 0");
 	return 0;
-
-error:
-	kernel_sock_shutdown(local->socket, SHUT_RDWR);
-	local->socket->sk->sk_user_data = NULL;
-	sock_release(local->socket);
-	local->socket = NULL;
-
-	_leave(" = %d", ret);
-	return ret;
 }
 
 /*
-- 
2.18.1

Re: [PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[David Howells]

Xin Long <lucien.xin@gmail.com> wrote:

> BTW: do you have any testing for this?

If you're using a distro like a recent-ish Fedora or, I think, Debian, you
should be able to install a kafs-client package.  If that works, start the
afs.mount service with systemctl and then look in /afs.  You should see
directories corresponding to a bunch of places that you can try accessing.  I
recommend you try "ls /afs/openafs.org".

If you don't have that available, if you have the keyutils package installed,
you can try:

	mount -t afs %openafs.org:root.cell /mnt

then do "ls /mnt".

David

Re: [PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[David Howells]

Xin Long <lucien.xin@gmail.com> wrote:

> > If you could arrange for a random port to be assigned in such a case (and
> > indicated back to the caller), that would be awesome.  Possibly I just don't
> > need to actually use bind in this case.
> >
> The patch is attached (based on this patch):

Initial testing seems to show that it works.  I'll poke at it some more
tomorrow.

David

Re: [PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[David Howells]

Xin Long <lucien.xin@gmail.com> wrote:

> Subject: [PATCH net-next] rxrpc: use udp tunnel APIs instead of open code in
>  rxrpc_open_socket
> 
> Signed-off-by: Xin Long <lucien.xin@gmail.com>

You can add "Acked-by: David Howells <dhowells@redhat.com>" if you want.

David

Re: [PATCHv4 net-next 0/2] net: enable udp v6 sockets receiving v4 packets with UDP GRO

[Xin Long]

On Fri, Feb 5, 2021 at 5:14 PM David Howells <dhowells@redhat.com> wrote:
>
> Xin Long <lucien.xin@gmail.com> wrote:
>
> > Subject: [PATCH net-next] rxrpc: use udp tunnel APIs instead of open code in
> >  rxrpc_open_socket
> >
> > Signed-off-by: Xin Long <lucien.xin@gmail.com>
>
> You can add "Acked-by: David Howells <dhowells@redhat.com>" if you want.
>
OK, Thank you so much!