From 5eddd76ec2fd1988f0a3450fde9730b10dd22992 Mon Sep 17 00:00:00 2001 From: Alexandre Cassen Date: Wed, 19 Feb 2025 12:20:37 +0200 Subject: [PATCH 1/2] xfrm: fix tunnel mode TX datapath in packet offload mode Packets that match the output xfrm policy are delivered to the netstack. In IPsec packet mode for tunnel mode, the HW is responsible for building the hard header and outer IP header. In such a situation, the inner header may refer to a network that is not directly reachable by the host, resulting in a failed neighbor resolution. The packet is then dropped. xfrm policy defines the netdevice to use for xmit so we can send packets directly to it. Makes direct xmit exclusive to tunnel mode, since some rules may apply in transport mode. Fixes: f8a70afafc17 ("xfrm: add TX datapath support for IPsec packet offload mode") Signed-off-by: Alexandre Cassen Signed-off-by: Leon Romanovsky Signed-off-by: Steffen Klassert --- net/xfrm/xfrm_output.c | 41 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 41 insertions(+) diff --git a/net/xfrm/xfrm_output.c b/net/xfrm/xfrm_output.c index f7abd42c077d..e5b3e343a5ec 100644 --- a/net/xfrm/xfrm_output.c +++ b/net/xfrm/xfrm_output.c @@ -612,6 +612,40 @@ int xfrm_output_resume(struct sock *sk, struct sk_buff *skb, int err) } EXPORT_SYMBOL_GPL(xfrm_output_resume); +static int xfrm_dev_direct_output(struct sock *sk, struct xfrm_state *x, + struct sk_buff *skb) +{ + struct dst_entry *dst = skb_dst(skb); + struct net *net = xs_net(x); + int err; + + dst = skb_dst_pop(skb); + if (!dst) { + XFRM_INC_STATS(net, LINUX_MIB_XFRMOUTERROR); + kfree_skb(skb); + return -EHOSTUNREACH; + } + skb_dst_set(skb, dst); + nf_reset_ct(skb); + + err = skb_dst(skb)->ops->local_out(net, sk, skb); + if (unlikely(err != 1)) { + kfree_skb(skb); + return err; + } + + /* In transport mode, network destination is + * directly reachable, while in tunnel mode, + * inner packet network may not be. In packet + * offload type, HW is responsible for hard + * header packet mangling so directly xmit skb + * to netdevice. + */ + skb->dev = x->xso.dev; + __skb_push(skb, skb->dev->hard_header_len); + return dev_queue_xmit(skb); +} + static int xfrm_output2(struct net *net, struct sock *sk, struct sk_buff *skb) { return xfrm_output_resume(sk, skb, 1); @@ -735,6 +769,13 @@ int xfrm_output(struct sock *sk, struct sk_buff *skb) return -EHOSTUNREACH; } + /* Exclusive direct xmit for tunnel mode, as + * some filtering or matching rules may apply + * in transport mode. + */ + if (x->props.mode == XFRM_MODE_TUNNEL) + return xfrm_dev_direct_output(sk, x, skb); + return xfrm_output_resume(sk, skb, 0); } From 0aae2867aa6067f73d066bc98385e23c8454a1d7 Mon Sep 17 00:00:00 2001 From: Cosmin Ratiu Date: Wed, 19 Feb 2025 12:52:48 +0200 Subject: [PATCH 2/2] xfrm_output: Force software GSO only in tunnel mode The cited commit fixed a software GSO bug with VXLAN + IPSec in tunnel mode. Unfortunately, it is slightly broader than necessary, as it also severely affects performance for Geneve + IPSec transport mode over a device capable of both HW GSO and IPSec crypto offload. In this case, xfrm_output unnecessarily triggers software GSO instead of letting the HW do it. In simple iperf3 tests over Geneve + IPSec transport mode over a back-2-back pair of NICs with MTU 1500, the performance was observed to be up to 6x worse when doing software GSO compared to leaving it to the hardware. This commit makes xfrm_output only trigger software GSO in crypto offload cases for already encapsulated packets in tunnel mode, as not doing so would then cause the inner tunnel skb->inner_networking_header to be overwritten and break software GSO for that packet later if the device turns out to not be capable of HW GSO. Taking a closer look at the conditions for the original bug, to better understand the reasons for this change: - vxlan_build_skb -> iptunnel_handle_offloads sets inner_protocol and inner network header. - then, udp_tunnel_xmit_skb -> ip_tunnel_xmit adds outer transport and network headers. - later in the xmit path, xfrm_output -> xfrm_outer_mode_output -> xfrm4_prepare_output -> xfrm4_tunnel_encap_add overwrites the inner network header with the one set in ip_tunnel_xmit before adding the second outer header. - __dev_queue_xmit -> validate_xmit_skb checks whether GSO segmentation needs to happen based on dev features. In the original bug, the hw couldn't segment the packets, so skb_gso_segment was invoked. - deep in the .gso_segment callback machinery, __skb_udp_tunnel_segment tries to use the wrong inner network header, expecting the one set in iptunnel_handle_offloads but getting the one set by xfrm instead. - a bit later, ipv6_gso_segment accesses the wrong memory based on that wrong inner network header. With the new change, the original bug (or similar ones) cannot happen again, as xfrm will now trigger software GSO before applying a tunnel. This concern doesn't exist in packet offload mode, when the HW adds encapsulation headers. For the non-offloaded packets (crypto in SW), software GSO is still done unconditionally in the else branch. Reviewed-by: Dragos Tatulea Reviewed-by: Yael Chemla Reviewed-by: Leon Romanovsky Fixes: a204aef9fd77 ("xfrm: call xfrm_output_gso when inner_protocol is set in xfrm_output") Signed-off-by: Cosmin Ratiu Signed-off-by: Steffen Klassert --- net/xfrm/xfrm_output.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/net/xfrm/xfrm_output.c b/net/xfrm/xfrm_output.c index e5b3e343a5ec..3cabc87978dd 100644 --- a/net/xfrm/xfrm_output.c +++ b/net/xfrm/xfrm_output.c @@ -799,7 +799,7 @@ int xfrm_output(struct sock *sk, struct sk_buff *skb) skb->encapsulation = 1; if (skb_is_gso(skb)) { - if (skb->inner_protocol) + if (skb->inner_protocol && x->props.mode == XFRM_MODE_TUNNEL) return xfrm_output_gso(net, sk, skb); skb_shinfo(skb)->gso_type |= SKB_GSO_ESP;