Re: [PATCH net-next 0/5] mlxsw: Add support for RED qevent "mark"

On Sat, Jan 23, 2021 at 11:55:27AM -0800, Jakub Kicinski wrote:

On Sat, 23 Jan 2021 17:28:02 +0200 Ido Schimmel wrote:

quoted

quoted

Thanks for the explanation. I feel more and more convinced now that
we should have TC_ACT_TRAP_MIRROR and the devlink trap should only 
be on/off :S Current model of "if ACT_TRAP consult devlink for trap
configuration" is impossible to model in SW since it doesn't have a
equivalent of devlink traps. Or we need that equivalent..  

Wait, the current model is not "if ACT_TRAP consult devlink for trap
configuration". 'ecn_mark' action is always 'trap' ('mirror' in v2) and
can't be changed. Such packets can always be sent to the CPU, but the
decision of whether to send them or not is based on the presence of tc
filters attached to RED's 'mark' qevent with TC_ACT_TRAP
(TC_ACT_TRAP_MIRROR in v2).

I see, missed that, but I think my point conceptually stands, right?
Part of forwarding behavior was (in v1) only expressed in control 
plane (devlink) not dataplane (tc).

I don't think so? The action was set to 'trap' in both devlink and tc.

quoted

I believe that with the proposed changes in v2 it should be perfectly
clear that ECN marked packets are forwarded in hardware and a copy is
sent to the CPU.

Yup, sounds good.

Thanks!