Intel igb 82576 (netmap mode) and RSS queue issuse with fragmented UDP packets.

[Slawa Olhovchenkov]

On Mon, Jun 01, 2015 at 07:53:51AM -0700, Adrian Chadd wrote:

> oh, hm, you asked the next question. Sorry, I haven't had coffee.
> 
> I vaguely remember testing this and discovering the /should/ be
> putting all the frames in a packet int he same queue, as long as the
> frames have a consistent set of fragment bits set. Ie, if you're doing
> UDP and all frames in a packet have the fragment bit set, then all
> fragments (first and the rest) go into the same queue. Frames in a
> flow that are fragmented for some and not others (eg TCP flows with a
> firewall/router doing explicit fragmentation, or a mix of small/large
> UDP packets) will end up going into different queues.
> 
> I can re-test this on ixgbe at some point, but IIRC that's how it behaved.
> 
> I don't have the RSS stuff done for chelsio, so I haven't done any
> experiments just yet.
> 
> Now, I also remember that with the badly setup flowdirector code in
> ixgbe it would mess that up, so the flowdir code was disabled in -HEAD
> and I believe now in -10.

I am expect that second and next fragments may distributed in
different queue. This is acceptable for me. And I think workaround of
this in driver/OS code may be very expensive (collect and save
5-tuple, expiration control of this hash and etc).
But I am very surprised that _first_ fragment distributed in different
queue. I.e. fragment with frag_offset==0. Only different with DF bit
clear and MF bit set. I am don't check case with frag_offset==0, DF
bit clear and MF bit clear. May be this is also distrebuted in
different queue.