Re: [net-next 08/15] net/mlx5e: CT: Preparation for offloading +trk+new ct rules

On 1/22/2021 4:16 AM, Marcelo Ricardo Leitner wrote:

On Fri, Jan 22, 2021 at 02:18:34AM +0100, Pablo Neira Ayuso wrote:

quoted

Hi Oz,

On Wed, Jan 20, 2021 at 06:09:48PM +0200, Oz Shlomo wrote:

quoted

On 1/14/2021 11:50 PM, Marcelo Ricardo Leitner wrote:

quoted

Thoughts?

I wonder if we should develop a generic mechanism to optimize CT software
for a use case that is faulty by design.
This has limited value for software as it would only reduce the conntrack
table size (packet classification is still required).
However, this feature may have a big impact on hardware offload.
Normally hardware offload relies on software to handle new connections.
Causing all new connections to be processed by software.
With this patch the hardware may autonomously set the +new connection state
for the relevant connections.

Could you fix this issue with unidirectional flows by checking for
IPS_CONFIRMED status bit? The idea is to hardware offload the entry
after the first packet goes through software successfully. Then, there
is no need to wait for the established state that requires to see
traffic in both directions.

That's an interesting idea. This way, basically all that needs to be
changed is tcf_ct_flow_table_process_conn() to handle this new
condition for UDP packets and on tcf_ct_act().

Will act_ct need to maintain a port list and classify the packet to realize whether the udp packet 
is part of a unidirection or biderectional udp connection?

It has a small performance penaulty if compared to the original
solution, as now the first packet(s) goes to sw, but looks like a good
compromise between supporting a (from what I could understand)
somewhat lazy flow design (as I still think these didn't need to go
through conntrack), an uniform system behavior (with and without
offload, with mlx5 or another driver) and a more generic approach.
Other situations that rely on unidirectional UDP flows will benefit
from it as well.

The hardware offload perspective is a bit different.
With this approach the system will offload a rule per connection instead of offloading one mega-flow 
rule on dst udp port.
This will increase the hardware scale requirements in terms of number of offloaded rules.
In addition, a counter will need to be instantiated per rule and the software will need to manage 
the aging of these connections.

We hoped that the hardware can fully offload this scenario, avoiding the need for sw processing at all.

This way I even think it doesn't need to be configurable right now.
It will be easier to add a knob to switch back to the old behavior if
needed later on, if anything.

   Marcelo