| Message ID | 1443145960-20514-1-git-send-email-wen.gang.wang@oracle.com |
|---|---|
| State | Changes Requested, archived |
| Delegated to: | David Miller |
| Headers | show |
Hi, Any comment on this patch? thanks, wengang 在 2015年09月25日 09:52, Wengang Wang 写道: > This is the v2, comparing the v1, the changes is: > * for loopback outbound device, it continue skipping cached route; > for others, it goes through the cached route. > > For multicast, we should find valid route(thus get the meaniful pmtu) for > the packet on the socket which is not bound to a device(sk_bound_dev_if > being 0) too. > > From man page of socket(7) > > SO_BINDTODEVICE > Bind this socket to a particular device like “eth0”, as > specified in the passed interface name. If the name is an > empty string or the option length is zero, the socket > device binding is removed. The passed option is a > variable-length null-terminated interface name string with > the maximum size of IFNAMSIZ. If a socket is bound to an > interface, only packets received from that particular > interface are processed by the socket. Note that this works > only for some socket types, particularly AF_INET sockets. > It is not supported for packet sockets (use normal bind(2) > there). > > The man page doesn't say when socket not bound packets won't be routed. > > A problem is hit that all multicast packets dropped by kernel(from sender > host). The lower layer is IPoIB with MTU being 7000. And I was sending 4096 > length multicast packets. Inside IPoIB the first send is dropped because > is exeeding the internal packet size limitation mcast_mtu which is 2044. > So IPoIB calls ip_rt_update_pmtu (indirectly) trying to set path mtu. A > correct route is configured for the multicast, so the setting of pmtu > cucceeded and the next multicast packet(to the same target) is expected > to succeed(it would be well fragmented accroding to the pmtu I just set). > But actually the second and later multicast packets got dropped too. And > the reason is that the neighor looking up(fib_lookup) is skipped because of > the socket is not bound to device(sk_bound_dev_if being 0). After applied > the patch I proposed here, it works fine. > > Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com> > --- > net/ipv4/route.c | 6 +++++- > 1 file changed, 5 insertions(+), 1 deletion(-) > > diff --git a/net/ipv4/route.c b/net/ipv4/route.c > index 5f4a556..c0534c2 100644 > --- a/net/ipv4/route.c > +++ b/net/ipv4/route.c > @@ -2097,7 +2097,10 @@ struct rtable *__ip_route_output_key(struct net *net, struct flowi4 *fl4) > */ > > fl4->flowi4_oif = dev_out->ifindex; > - goto make_route; > + if (dev_out->flags & IFF_LOOPBACK) > + goto make_route; > + else > + goto lookup; > } > > if (!(fl4->flowi4_flags & FLOWI_FLAG_ANYSRC)) { > @@ -2153,6 +2156,7 @@ struct rtable *__ip_route_output_key(struct net *net, struct flowi4 *fl4) > goto make_route; > } > > +lookup: > if (fib_lookup(net, fl4, &res, 0)) { > res.fi = NULL; > res.table = NULL; -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
From: Wengang Wang <wen.gang.wang@oracle.com> Date: Thu, 8 Oct 2015 11:31:05 +0800 > Any comment on this patch? I'm still thinking about it, so if nobody else chimes in with a review that convinces me one way or another you will just have to patiently wait for me. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
From: Wengang Wang <wen.gang.wang@oracle.com> Date: Fri, 25 Sep 2015 09:52:40 +0800 > diff --git a/net/ipv4/route.c b/net/ipv4/route.c > index 5f4a556..c0534c2 100644 > --- a/net/ipv4/route.c > +++ b/net/ipv4/route.c > @@ -2097,7 +2097,10 @@ struct rtable *__ip_route_output_key(struct net *net, struct flowi4 *fl4) > */ > > fl4->flowi4_oif = dev_out->ifindex; > - goto make_route; > + if (dev_out->flags & IFF_LOOPBACK) > + goto make_route; > + else > + goto lookup; > } This is still broken. By definition invoking fib_lookup() and depending upon it finding something in this path is going to break things for somebody, somewhere. Before your change, if we lacked a multicast route, the user would still get a functioning path. Furthermore, most of the other "goto make_route" cases in this function suffer from the same exact problem you're trying to solve. Therefore, special casing one instance makes no sense at all. I want you to, instead of making potentially lethal semantic changes here, fix the real problem instead. That is, I want you to fix how we do not cache routes we create merely because we lack a fib_info. Thanks. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/net/ipv4/route.c b/net/ipv4/route.c index 5f4a556..c0534c2 100644 --- a/net/ipv4/route.c +++ b/net/ipv4/route.c @@ -2097,7 +2097,10 @@ struct rtable *__ip_route_output_key(struct net *net, struct flowi4 *fl4) */ fl4->flowi4_oif = dev_out->ifindex; - goto make_route; + if (dev_out->flags & IFF_LOOPBACK) + goto make_route; + else + goto lookup; } if (!(fl4->flowi4_flags & FLOWI_FLAG_ANYSRC)) { @@ -2153,6 +2156,7 @@ struct rtable *__ip_route_output_key(struct net *net, struct flowi4 *fl4) goto make_route; } +lookup: if (fib_lookup(net, fl4, &res, 0)) { res.fi = NULL; res.table = NULL;
This is the v2, comparing the v1, the changes is: * for loopback outbound device, it continue skipping cached route; for others, it goes through the cached route. For multicast, we should find valid route(thus get the meaniful pmtu) for the packet on the socket which is not bound to a device(sk_bound_dev_if being 0) too. From man page of socket(7) SO_BINDTODEVICE Bind this socket to a particular device like “eth0”, as specified in the passed interface name. If the name is an empty string or the option length is zero, the socket device binding is removed. The passed option is a variable-length null-terminated interface name string with the maximum size of IFNAMSIZ. If a socket is bound to an interface, only packets received from that particular interface are processed by the socket. Note that this works only for some socket types, particularly AF_INET sockets. It is not supported for packet sockets (use normal bind(2) there). The man page doesn't say when socket not bound packets won't be routed. A problem is hit that all multicast packets dropped by kernel(from sender host). The lower layer is IPoIB with MTU being 7000. And I was sending 4096 length multicast packets. Inside IPoIB the first send is dropped because is exeeding the internal packet size limitation mcast_mtu which is 2044. So IPoIB calls ip_rt_update_pmtu (indirectly) trying to set path mtu. A correct route is configured for the multicast, so the setting of pmtu cucceeded and the next multicast packet(to the same target) is expected to succeed(it would be well fragmented accroding to the pmtu I just set). But actually the second and later multicast packets got dropped too. And the reason is that the neighor looking up(fib_lookup) is skipped because of the socket is not bound to device(sk_bound_dev_if being 0). After applied the patch I proposed here, it works fine. Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com> --- net/ipv4/route.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)