14/07/2021 17:00, Jan Viktorin: > > > On Tue, 13 Jul 2021 12:26:35 +0300 > > > Andrew Rybchenko <andrew.rybche...@oktetlabs.ru> wrote: > > > > >>> > > This matters for the bonding case as well, doesn't it?. > > > > >>> > > It is not desirable to accidently omit a packet that was > > > > >>> > > received by primary ingress logic instead of being > > > > >>> > > redirected into the dedicated queue. > > > > >>> > > > > > > >>> > > Are there any chances that for mlx5 it would be possible > > > > >>> > > to insert flow rules before calling rte_eth_dev_start? > > > > >>> > > Anyway, the behaviour should be specified and documented > > > > >>> > > in DPDK more precisely to avoid such uncertainty in the > > > > >>> > > future. > > > > >>> > I agree the documentation should be fixed. > > > > >>> > > > > >>> +1 > > > > > > > > Cc Thomas and Ferruh since ethdev documentation should be > > > > clarified. > > > > > > > > It looks like there is no consensus if the patch is a right > > > > direction or wrong. For me it looks wrong taking all above > > > > arguments in to account (mainly necessity to be able to insert > > > > flows before pushing start button which enables the traffic if HW > > > > supports it). > > > > > > > > So, I'm applying first two patches and hold on this one. > > Andrew, I believe that it would be helpful to start some new thread > otherwise we would get lost here :). It seems that we will have few > more fixes for the bonding driver. Do you prefer an entirely new > patchset or v2 of this topic? Or any other advise how to proceed?
This thread is about 3 things: - bonding issue - ethdev doc - mlx5 design That's too much topics to address in one thread :) You may restart the discussion with a doc update if the stop/start requirement is not clear.