Message ID | 20190625103359.31102-1-ssuryaextr@gmail.com |
---|---|
State | Superseded |
Delegated to: | David Miller |
Headers | show |
Series | [net] vrf: reset rt_iif for recirculated mcast out pkts | expand |
On 6/25/19 4:33 AM, Stephen Suryaputra wrote: > @@ -363,10 +376,20 @@ int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb) > #endif > ) { > struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC); > - if (newskb) > + if (newskb) { > + /* Reset rt_iif so that inet_iif() will return > + * skb->dev->ifIndex which is the VRF device for > + * socket lookup. Setting this to VRF ifindex > + * causes ipi_ifindex in in_pktinfo to be > + * overwritten, see ipv4_pktinfo_prepare(). > + */ > + if (netif_is_l3_slave(dev)) seems like the rt_iif is a problem for recirculated mcast packets in general, not just ones tied to a VRF. > + ip_mc_reset_rt_iif(net, rt, newskb); > + > NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, > net, sk, newskb, NULL, newskb->dev, > ip_mc_finish_output); > + } > } > > /* Multicasts with ttl 0 must not go beyond the host */
From: Stephen Suryaputra <ssuryaextr@gmail.com> Date: Tue, 25 Jun 2019 06:33:59 -0400 > Multicast egress packets has skb_rtable(skb)->rt_iif set to the oif. > Depending on the socket, these packets might be recirculated back as > input and raw sockets that are opened for them are bound to the VRF. But > since skb_rtable(skb) is set and its rt_iif is non-zero, inet_iif() > function returns rt_iif instead of skb_iif (the VRF netdev). Hence, the > socket lookup fails. > > Signed-off-by: Stephen Suryaputra <ssuryaextr@gmail.com> David A., please review. Thank you.
On 6/25/19 2:22 PM, David Ahern wrote: > On 6/25/19 4:33 AM, Stephen Suryaputra wrote: >> @@ -363,10 +376,20 @@ int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb) >> #endif >> ) { >> struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC); >> - if (newskb) >> + if (newskb) { >> + /* Reset rt_iif so that inet_iif() will return >> + * skb->dev->ifIndex which is the VRF device for >> + * socket lookup. Setting this to VRF ifindex >> + * causes ipi_ifindex in in_pktinfo to be >> + * overwritten, see ipv4_pktinfo_prepare(). >> + */ >> + if (netif_is_l3_slave(dev)) > > seems like the rt_iif is a problem for recirculated mcast packets in > general, not just ones tied to a VRF. > >> + ip_mc_reset_rt_iif(net, rt, newskb); >> + >> NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, >> net, sk, newskb, NULL, newskb->dev, >> ip_mc_finish_output); >> + } >> } >> >> /* Multicasts with ttl 0 must not go beyond the host */ Also, wouldn't this problem apply to broadcast packets as well if they get recirculated back up the stack? Maybe then the reset_rt_iif needs to be done in ip_mc_finish_output.
On Tue, Jun 25, 2019 at 4:22 PM David Ahern <dsahern@gmail.com> wrote: > > On 6/25/19 4:33 AM, Stephen Suryaputra wrote: > > @@ -363,10 +376,20 @@ int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb) > > #endif > > ) { > > struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC); > > - if (newskb) > > + if (newskb) { > > + /* Reset rt_iif so that inet_iif() will return > > + * skb->dev->ifIndex which is the VRF device for > > + * socket lookup. Setting this to VRF ifindex > > + * causes ipi_ifindex in in_pktinfo to be > > + * overwritten, see ipv4_pktinfo_prepare(). > > + */ > > + if (netif_is_l3_slave(dev)) > > seems like the rt_iif is a problem for recirculated mcast packets in > general, not just ones tied to a VRF. It seems so to me too but I wonder why this hasn't been seen... Can I get more feedbacks on this? If there is an agreement to fix this generally, I will remove the if clause and respin the patch with more accurate changelog. > > > + ip_mc_reset_rt_iif(net, rt, newskb); > > + > > NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, > > net, sk, newskb, NULL, newskb->dev, > > ip_mc_finish_output); > > + } > > } > > > > /* Multicasts with ttl 0 must not go beyond the host */ Thanks.
On 6/25/19 2:42 PM, Stephen Suryaputra wrote: > On Tue, Jun 25, 2019 at 4:22 PM David Ahern <dsahern@gmail.com> wrote: >> >> On 6/25/19 4:33 AM, Stephen Suryaputra wrote: >>> @@ -363,10 +376,20 @@ int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb) >>> #endif >>> ) { >>> struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC); >>> - if (newskb) >>> + if (newskb) { >>> + /* Reset rt_iif so that inet_iif() will return >>> + * skb->dev->ifIndex which is the VRF device for >>> + * socket lookup. Setting this to VRF ifindex >>> + * causes ipi_ifindex in in_pktinfo to be >>> + * overwritten, see ipv4_pktinfo_prepare(). >>> + */ >>> + if (netif_is_l3_slave(dev)) >> >> seems like the rt_iif is a problem for recirculated mcast packets in >> general, not just ones tied to a VRF. > > It seems so to me too but I wonder why this hasn't been seen... > Can I get more feedbacks on this? If there is an agreement to fix this > generally, I will remove the if clause and respin the patch with more > accurate changelog. rt_iif is used to save the original oif during a route lookup so if packets are delivered locally apps can know the real oif and not the loopback/VRF device which is just a trick for local traffic. The raw socket lookup was recently changed to handle local traffic with raw sockets bound to a device. e.g., ping was recently changed to use SO_BINDTODEVICE vs IP_PKTINFO and revealed that the socket lookup was not considering rt_iif when doing the lookup for 'ping -I <dev> <dev addr>' The mcast use case seems to get hung up on rt_iif being set when packets are recirculated for local delivery which is slightly different use case than local traffic to local addresses.
diff --git a/include/net/route.h b/include/net/route.h index 065b47754f05..55ff71ffb796 100644 --- a/include/net/route.h +++ b/include/net/route.h @@ -221,6 +221,7 @@ void ip_rt_get_source(u8 *src, struct sk_buff *skb, struct rtable *rt); struct rtable *rt_dst_alloc(struct net_device *dev, unsigned int flags, u16 type, bool nopolicy, bool noxfrm, bool will_cache); +struct rtable *rt_dst_clone(struct net_device *dev, struct rtable *rt); struct in_ifaddr; void fib_add_ifaddr(struct in_ifaddr *); diff --git a/net/ipv4/ip_output.c b/net/ipv4/ip_output.c index 16f9159234a2..a5e240bad3ce 100644 --- a/net/ipv4/ip_output.c +++ b/net/ipv4/ip_output.c @@ -329,6 +329,19 @@ static int ip_mc_finish_output(struct net *net, struct sock *sk, return dev_loopback_xmit(net, sk, skb); } +static void ip_mc_reset_rt_iif(struct net *net, struct rtable *rt, + struct sk_buff *newskb) +{ + struct rtable *new_rt; + + new_rt = rt_dst_clone(net->loopback_dev, rt); + if (new_rt) { + new_rt->rt_iif = 0; + skb_dst_drop(newskb); + skb_dst_set(newskb, &new_rt->dst); + } +} + int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb) { struct rtable *rt = skb_rtable(skb); @@ -363,10 +376,20 @@ int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb) #endif ) { struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC); - if (newskb) + if (newskb) { + /* Reset rt_iif so that inet_iif() will return + * skb->dev->ifIndex which is the VRF device for + * socket lookup. Setting this to VRF ifindex + * causes ipi_ifindex in in_pktinfo to be + * overwritten, see ipv4_pktinfo_prepare(). + */ + if (netif_is_l3_slave(dev)) + ip_mc_reset_rt_iif(net, rt, newskb); + NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, net, sk, newskb, NULL, newskb->dev, ip_mc_finish_output); + } } /* Multicasts with ttl 0 must not go beyond the host */ diff --git a/net/ipv4/route.c b/net/ipv4/route.c index 6cb7cff22db9..8ea0735a6754 100644 --- a/net/ipv4/route.c +++ b/net/ipv4/route.c @@ -1647,6 +1647,39 @@ struct rtable *rt_dst_alloc(struct net_device *dev, } EXPORT_SYMBOL(rt_dst_alloc); +struct rtable *rt_dst_clone(struct net_device *dev, struct rtable *rt) +{ + struct rtable *new_rt; + + new_rt = dst_alloc(&ipv4_dst_ops, dev, 1, DST_OBSOLETE_FORCE_CHK, + rt->dst.flags); + + if (new_rt) { + new_rt->rt_genid = rt_genid_ipv4(dev_net(dev)); + new_rt->rt_flags = rt->rt_flags; + new_rt->rt_type = rt->rt_type; + new_rt->rt_is_input = rt->rt_is_input; + new_rt->rt_iif = rt->rt_iif; + new_rt->rt_pmtu = rt->rt_pmtu; + new_rt->rt_mtu_locked = rt->rt_mtu_locked; + new_rt->rt_gw_family = rt->rt_gw_family; + if (rt->rt_gw_family == AF_INET) + new_rt->rt_gw4 = rt->rt_gw4; + else if (rt->rt_gw_family == AF_INET6) + new_rt->rt_gw6 = rt->rt_gw6; + INIT_LIST_HEAD(&new_rt->rt_uncached); + + new_rt->dst.flags |= DST_HOST; + new_rt->dst.input = rt->dst.input; + new_rt->dst.output = rt->dst.output; + new_rt->dst.error = rt->dst.error; + new_rt->dst.lastuse = jiffies; + new_rt->dst.lwtstate = lwtstate_get(rt->dst.lwtstate); + } + return new_rt; +} +EXPORT_SYMBOL(rt_dst_clone); + /* called in rcu_read_lock() section */ int ip_mc_validate_source(struct sk_buff *skb, __be32 daddr, __be32 saddr, u8 tos, struct net_device *dev,
Multicast egress packets has skb_rtable(skb)->rt_iif set to the oif. Depending on the socket, these packets might be recirculated back as input and raw sockets that are opened for them are bound to the VRF. But since skb_rtable(skb) is set and its rt_iif is non-zero, inet_iif() function returns rt_iif instead of skb_iif (the VRF netdev). Hence, the socket lookup fails. Signed-off-by: Stephen Suryaputra <ssuryaextr@gmail.com> --- include/net/route.h | 1 + net/ipv4/ip_output.c | 25 ++++++++++++++++++++++++- net/ipv4/route.c | 33 +++++++++++++++++++++++++++++++++ 3 files changed, 58 insertions(+), 1 deletion(-)