From: Steffen Klassert <steffen.klass...@secunet.com> Date: Mon, 11 Jan 2021 09:43:22 +0100
> On Mon, Jan 11, 2021 at 11:02:42AM +0900, Dongseok Yi wrote: >> On 2021-01-08 22:35, Steffen Klassert wrote: >>> On Fri, Jan 08, 2021 at 09:52:28PM +0900, Dongseok Yi wrote: >>>> It is a workaround patch. >>>> >>>> UDP/IP header of UDP GROed frag_skbs are not updated even after NAT >>>> forwarding. Only the header of head_skb from ip_finish_output_gso -> >>>> skb_gso_segment is updated but following frag_skbs are not updated. >>>> >>>> A call path skb_mac_gso_segment -> inet_gso_segment -> >>>> udp4_ufo_fragment -> __udp_gso_segment -> __udp_gso_segment_list >>>> does not try to update any UDP/IP header of the segment list. >>>> >>>> It might make sense because each skb of frag_skbs is converted to a >>>> list of regular packets. Header update with checksum calculation may >>>> be not needed for UDP GROed frag_skbs. >>>> >>>> But UDP GRO frag_list is started from udp_gro_receive, we don't know >>>> whether the skb will be NAT forwarded at that time. For workaround, >>>> try to get sock always when call udp4_gro_receive -> udp_gro_receive >>>> to check if the skb is for local. >>>> >>>> I'm still not sure if UDP GRO frag_list is really designed for local >>>> session only. Can kernel support NAT forward for UDP GRO frag_list? >>>> What am I missing? >>> >>> The initial idea when I implemented this was to have a fast >>> forwarding path for UDP. So forwarding is a usecase, but NAT >>> is a problem, indeed. A quick fix could be to segment the >>> skb before it gets NAT forwarded. Alternatively we could >>> check for a header change in __udp_gso_segment_list and >>> update the header of the frag_skbs accordingly in that case. >> >> Thank you for explaining. >> Can I think of it as a known issue? > > No, it was not known before you reported it. > >> I think we should have a fix >> because NAT can be triggered by user. Can I check the current status? >> Already planning a patch or a new patch should be written? > > We have to do a new patch to fix that issue. If you want do > do so, go ahead. This patch is incorrect. I do NAT UDP GRO Fraglists via nftables (both with and without flow offload) with no issues since March'20. Packet loss rates are always +/- 0, so I can say it works properly. I can share any details / dump any runtime data if needed. Thanks, Al
