| Message ID | 1571660763-117936-1-git-send-email-linyunsheng@huawei.com |
|---|---|
| State | RFC |
| Delegated to: | David Miller |
| Headers | show |
| Series | [RFC] net: vlan: reverse 4 bytes of vlan header when setting initial MTU | expand |
From: Yunsheng Lin > Sent: 21 October 2019 13:26 > Currently the MTU of vlan netdevice is set to the same MTU > of the lower device, which requires the underlying device > to handle it as the comment has indicated: > > /* need 4 bytes for extra VLAN header info, > * hope the underlying device can handle it. > */ > new_dev->mtu = real_dev->mtu; > > Currently most of the physical netdevs seems to handle above > by reversing 2 * VLAN_HLEN for L2 packet len. s/reverse/reserve/g - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
On Mon, 21 Oct 2019 20:26:03 +0800 Yunsheng Lin <linyunsheng@huawei.com> wrote: > Currently the MTU of vlan netdevice is set to the same MTU > of the lower device, which requires the underlying device > to handle it as the comment has indicated: > > /* need 4 bytes for extra VLAN header info, > * hope the underlying device can handle it. > */ > new_dev->mtu = real_dev->mtu; > > Currently most of the physical netdevs seems to handle above > by reversing 2 * VLAN_HLEN for L2 packet len. > > But for vlan netdev over vxlan netdev case, the vxlan does not > seems to reverse the vlan header for vlan device, which may cause > performance degradation because vxlan may emit a packet that > exceed the MTU of the physical netdev, and cause the software > TSO to happen in ip_finish_output_gso(), software TSO call stack > as below: > > => ftrace_graph_call > => tcp_gso_segment > => tcp4_gso_segment > => inet_gso_segment > => skb_mac_gso_segment > => skb_udp_tunnel_segment > => udp4_ufo_fragment > => inet_gso_segment > => skb_mac_gso_segment > => __skb_gso_segment > => __ip_finish_output > => ip_output > => ip_local_out > => iptunnel_xmit > => udp_tunnel_xmit_skb > => vxlan_xmit_one > => vxlan_xmit > => dev_hard_start_xmit > => __dev_queue_xmit > => dev_queue_xmit > => vlan_dev_hard_start_xmit > => dev_hard_start_xmit > => __dev_queue_xmit > => dev_queue_xmit > => neigh_resolve_output > => ip_finish_output2 > => __ip_finish_output > => ip_output > => ip_local_out > => __ip_queue_xmit > => ip_queue_xmit > => __tcp_transmit_skb > => tcp_write_xmit > => __tcp_push_pending_frames > => tcp_push > => tcp_sendmsg_locked > => tcp_sendmsg > => inet_sendmsg > => sock_sendmsg > => sock_write_iter > => new_sync_write > => __vfs_write > => vfs_write > => ksys_write > => __arm64_sys_write > => el0_svc_common.constprop.0 > => el0_svc_handler > => el0_svc > > This patch set initial MTU of the vlan device to the MTU of the > lower device minus vlan header to handle the above case. > > Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> The MTU is visible to user space in many tools, and Linux (and BSD) have always treated VLAN header as not part of the MTU. You can't change that now.
On 2019/10/21 22:46, David Laight wrote: > From: Yunsheng Lin >> Sent: 21 October 2019 13:26 >> Currently the MTU of vlan netdevice is set to the same MTU >> of the lower device, which requires the underlying device >> to handle it as the comment has indicated: >> >> /* need 4 bytes for extra VLAN header info, >> * hope the underlying device can handle it. >> */ >> new_dev->mtu = real_dev->mtu; >> >> Currently most of the physical netdevs seems to handle above >> by reversing 2 * VLAN_HLEN for L2 packet len. > > s/reverse/reserve/g Thanks. > > - > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > Registration No: 1397386 (Wales) > > > . >
On 2019/10/22 7:27, Stephen Hemminger wrote: > On Mon, 21 Oct 2019 20:26:03 +0800 > Yunsheng Lin <linyunsheng@huawei.com> wrote: > >> Currently the MTU of vlan netdevice is set to the same MTU >> of the lower device, which requires the underlying device >> to handle it as the comment has indicated: >> >> /* need 4 bytes for extra VLAN header info, >> * hope the underlying device can handle it. >> */ >> new_dev->mtu = real_dev->mtu; >> >> Currently most of the physical netdevs seems to handle above >> by reversing 2 * VLAN_HLEN for L2 packet len. >> >> But for vlan netdev over vxlan netdev case, the vxlan does not >> seems to reverse the vlan header for vlan device, which may cause >> performance degradation because vxlan may emit a packet that >> exceed the MTU of the physical netdev, and cause the software >> TSO to happen in ip_finish_output_gso(), software TSO call stack >> as below: >> >> => ftrace_graph_call >> => tcp_gso_segment >> => tcp4_gso_segment >> => inet_gso_segment >> => skb_mac_gso_segment >> => skb_udp_tunnel_segment >> => udp4_ufo_fragment >> => inet_gso_segment >> => skb_mac_gso_segment >> => __skb_gso_segment >> => __ip_finish_output >> => ip_output >> => ip_local_out >> => iptunnel_xmit >> => udp_tunnel_xmit_skb >> => vxlan_xmit_one >> => vxlan_xmit >> => dev_hard_start_xmit >> => __dev_queue_xmit >> => dev_queue_xmit >> => vlan_dev_hard_start_xmit >> => dev_hard_start_xmit >> => __dev_queue_xmit >> => dev_queue_xmit >> => neigh_resolve_output >> => ip_finish_output2 >> => __ip_finish_output >> => ip_output >> => ip_local_out >> => __ip_queue_xmit >> => ip_queue_xmit >> => __tcp_transmit_skb >> => tcp_write_xmit >> => __tcp_push_pending_frames >> => tcp_push >> => tcp_sendmsg_locked >> => tcp_sendmsg >> => inet_sendmsg >> => sock_sendmsg >> => sock_write_iter >> => new_sync_write >> => __vfs_write >> => vfs_write >> => ksys_write >> => __arm64_sys_write >> => el0_svc_common.constprop.0 >> => el0_svc_handler >> => el0_svc >> >> This patch set initial MTU of the vlan device to the MTU of the >> lower device minus vlan header to handle the above case. >> >> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> > > The MTU is visible to user space in many tools, and Linux (and BSD) > have always treated VLAN header as not part of the MTU. You can't change > that now. Ok. Is there any other feasible way to bring back the performance gain in the vlan netdev over vxlan netdev case? Or we just leave it as it is, and expect user to manually configure the MTU of vlan netdev to the MTU of thelower device minus vlan header when the performace in the above case is a concern to user? Thanks. > > > . >
On 10/21/19 7:31 PM, Yunsheng Lin wrote: > Or we just leave it as it is, and expect user to manually configure the MTU > of vlan netdev to the MTU of thelower device minus vlan header when the > performace in the above case is a concern to user? > for now, I would think so. vlan on a vxlan device ... you are going after q-in-q'ish with vxlan? that can not be a common deployment.
diff --git a/net/8021q/vlan.c b/net/8021q/vlan.c index 54728d2..0c26b92 100644 --- a/net/8021q/vlan.c +++ b/net/8021q/vlan.c @@ -261,10 +261,9 @@ static int register_vlan_device(struct net_device *real_dev, u16 vlan_id) return -ENOBUFS; dev_net_set(new_dev, net); - /* need 4 bytes for extra VLAN header info, - * hope the underlying device can handle it. - */ - new_dev->mtu = real_dev->mtu; + new_dev->mtu = real_dev->mtu - VLAN_HLEN; + if (new_dev->mtu < ETH_MIN_MTU) + new_dev->mtu = ETH_MIN_MTU; vlan = vlan_dev_priv(new_dev); vlan->vlan_proto = htons(ETH_P_8021Q);
Currently the MTU of vlan netdevice is set to the same MTU of the lower device, which requires the underlying device to handle it as the comment has indicated: /* need 4 bytes for extra VLAN header info, * hope the underlying device can handle it. */ new_dev->mtu = real_dev->mtu; Currently most of the physical netdevs seems to handle above by reversing 2 * VLAN_HLEN for L2 packet len. But for vlan netdev over vxlan netdev case, the vxlan does not seems to reverse the vlan header for vlan device, which may cause performance degradation because vxlan may emit a packet that exceed the MTU of the physical netdev, and cause the software TSO to happen in ip_finish_output_gso(), software TSO call stack as below: => ftrace_graph_call => tcp_gso_segment => tcp4_gso_segment => inet_gso_segment => skb_mac_gso_segment => skb_udp_tunnel_segment => udp4_ufo_fragment => inet_gso_segment => skb_mac_gso_segment => __skb_gso_segment => __ip_finish_output => ip_output => ip_local_out => iptunnel_xmit => udp_tunnel_xmit_skb => vxlan_xmit_one => vxlan_xmit => dev_hard_start_xmit => __dev_queue_xmit => dev_queue_xmit => vlan_dev_hard_start_xmit => dev_hard_start_xmit => __dev_queue_xmit => dev_queue_xmit => neigh_resolve_output => ip_finish_output2 => __ip_finish_output => ip_output => ip_local_out => __ip_queue_xmit => ip_queue_xmit => __tcp_transmit_skb => tcp_write_xmit => __tcp_push_pending_frames => tcp_push => tcp_sendmsg_locked => tcp_sendmsg => inet_sendmsg => sock_sendmsg => sock_write_iter => new_sync_write => __vfs_write => vfs_write => ksys_write => __arm64_sys_write => el0_svc_common.constprop.0 => el0_svc_handler => el0_svc This patch set initial MTU of the vlan device to the MTU of the lower device minus vlan header to handle the above case. Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- net/8021q/vlan.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)