On Mon, Oct 26, 2015 at 11:44 AM, Hannes Frederic Sowa <han...@stressinduktion.org> wrote: > > > On Mon, Oct 26, 2015, at 15:19, Tom Herbert wrote: >> > We already concluded that drivers do have this problem and not the stack >> > above ip6_fragment. The places I am aware of I fixed in this patch. Also >> > IPv4 to me seems unaffected, albeit one can certainly clean up the logic >> > in net-next. >> > >> I don't understand why checksum for IP fragments is a driver problem. >> When fragments are sent to driver they should never have >> CHECKSUM_PARTIAL set (or maybe that is what you are seeing?). > > Because either the drivers or the hardware does not correctly iterate > over the extension headers to fetch the final nexthdr field which is > used to compute the checksum. This is different from IPv4. > > I can only guess e.g. from the e1000e driver: > > case cpu_to_be16(ETH_P_IPV6): > /* XXX not handling all IPV6 headers */ > if (ipv6_hdr(skb)->nexthdr == IPPROTO_TCP) > cmd_len |= E1000_TXD_CMD_TCP; > break; > Yes, but in the case of a fragment that code should never be hit since ip_summed shouldn't be CHECKSUM_PARTIAL for a fragment (maybe after the fix in ip_output). For other cases of extension headers the e1000e is broken since it apparently does call skb_checksum_help for protocols it doesn't understand (the /* XXX not handling all IPV6 headers */ comment is worrisome!)
Tom -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html