On (01/29/16 10:54), Eric Dumazet wrote: > > Why ipv6 stack itself does not trigger the issue ? > > Maybe the driver itself does not properly align IP headers on sparc ? > > Make sure NET_IP_ALIGN is 2 on your build. > > Note that x86 does not care, but a driver should always align Ethernet > > header to NET_IP_ALIGN, so that IP headers are aligned to 4 bytes > > boundaries.
Consolidating a few responses: NET_IP_ALIGN is 2, of course, or else I would have seen errors about unaligned access in many more places. The issue is being triggered by ipv6 stack (I'm running iperf over ipv6) so I'm not sure I understand the first question. I tried out the suggested patch of setting page_offset to NET_IP_ALIGN in ixgbe_main.c - and yes, it kills the unaligned errors, so that's a Good Thing. (though I see davem just responded to the patch) However, the patch itself is not arch specific, and the unaligned address issue should have impacted other archs (silently) as well - even for ipv4, It would also impact other logic in __skb_flow_dissect: e.g., the access to iph->saddr for ipv4 frags would be unaligned in __skb_flow_dissect? (And Ias I mentioned in my original mail, the comments in __skb_get_poff seem to indicate that others have tripped up on this before - and yes, casting is not a good thing but seems like that's what they did to paper over the problem) > Hmmm.... it seems that flow dissector can "support L2 GRE", leading to > unaligned accesses since a header is 14 bytes (not multiple of 4) : > IP stacks in linux assume IP headers are always aligned to 4 bytes, > so it means having a GRE header like this would align trap on some > arches. I'm not sure I see how the above commit could have been impacting me, I dont have any l2 gre invovled (simple p2p Ethernet II + ipv6 + tcp, no vlans, gre or other exotic stuff). --Sowmini
