Would it be too much to ask for those expending time on this debate, to expend a little extra time doing some code review and testing this patch?
Or, like while the flames are being composed, merely sending data through it? This particular patch improves ipv4 by over 10% especially when used with the sfq and sfqred qdiscs, fixes encapsulation for ipsec (both ipv4 and ipv6), and yes, more than doubles ipv6 performance on the ag71xx. Although simple in outline, it is pretty invasive into core bits of the kernel. I wish it wasn't, but the alternatives (have the driver rebuffer misaligned packets) were far worse for performance, and the final patch isn't all that big. I agree with felix's suggestion to not let __packed leak out of the kernel headers, and will respin the patch for that. I'm dubious that the check alignment and branch is actually all that useful in the ipv6 checksum code, too, on an arch with a small cache. And I'm still trying to track down the last causes of traps in the routing path, and several other ipv6 and ipv4 related bugs. I'm pretty sure that the traps in the routing path have caused many a tcp reset for me over the last year, under load. http://www.bufferbloat.net/projects/cerowrt/issues/371 http://www.bufferbloat.net/projects/cerowrt/issues http://www.bufferbloat.net/projects/cerowrt/issues/360 -- Dave Täht SKYPE: davetaht US Tel: 1-239-829-5608 http://www.bufferbloat.net _______________________________________________ openwrt-devel mailing list openwrt-devel@lists.openwrt.org https://lists.openwrt.org/mailman/listinfo/openwrt-devel
