RE: [PATCH net-next 0/2] net: stmmac: add EthType Rx Frame steering

-----Original Message-----
From: Jakub Kicinski <kuba@kernel.org>
Sent: Friday, December 10, 2021 11:38 AM
To: Vladimir Oltean <olteanv@gmail.com>
Cc: Ong, Boon Leong <redacted>; David S . Miller
[off-list ref]; Giuseppe Cavallaro [off-list ref];
Alexandre Torgue [off-list ref]; Jose Abreu
[off-list ref]; Maxime Coquelin
[off-list ref]; alexandre.torgue@foss.st.com;
Kanzenbach, Kurt [off-list ref];
netdev@vger.kernel.org; linux-stm32@st-md-mailman.stormreply.com;
linux-arm-kernel@lists.infradead.org; Nambiar, Amritha
[off-list ref]
Subject: Re: [PATCH net-next 0/2] net: stmmac: add EthType Rx Frame
steering

On Fri, 10 Dec 2021 13:57:30 +0200 Vladimir Oltean wrote:

quoted

Is it the canonical approach to perform flow steering via tc-flower hw_tc,
as opposed to ethtool --config-nfc? My understanding from reading the
documentation is that tc-flower hw_tc only selects the hardware traffic
class for a packet, and that this has to do with prioritization
(although the concept in itself is a bit ill-defined as far as I
understand it, how does it relate to things like offloaded skbedit priority?).
But selecting a traffic class, in itself, doesn't (directly or
necessarily) select a ring per se, as ethtool does? Just like ethtool
doesn't select packet priority, just RX queue. When the RX queue
priority is configurable (see the "snps,priority" device tree property
in stmmac_mtl_setup) and more RX queues have the same priority, I'm not
sure what hw_tc is supposed to do in terms of RX queue selection?

You didn't mention the mqprio, but I think that's the piece that maps
TCs to queue pairs. You can have multiple queues in a TC.

Obviously that's still pretty weird what the flow rules should select
is an RSS context. mqprio is a qdisc, which means Tx, not Rx.

Adding Amritha who I believe added the concept of selecting Rx queues
via hw_tc. Can you comment?

So tc-mpqrio is the piece that is needed to set up the queue-groups. The offload
mode "hw 2" in mqprio will offload the TCs, the queue configurations and
bandwidth rate limits. The prio-tc map in mqprio will map a user priority to the
TC/queue-group. The priority to traffic class mapping and the user specified
queue ranges are used to configure the traffic class when the 'hw' option is set to 2.
Drivers can then configure queue-pairs based on the offsets and queue ranges
in mqprio.

The hw_tc option in tc-flower for ingress filter is used to direct Rx traffic to the
queue-group (configured via mqprio). Queue selection within the queue group can
be achieved using RSS.

I agree mqprio qdisc should be used to set up Tx queues only, but the limitation was the
absence of a single interface that could configure both Tx and Rx queue-groups/queue-sets
(ethtool did not support directing flows to a queue-group, but only a specific individual
queue, TC does not support Rx queue-group configuration either). The hw_tc in mqprio is a
range of class ids reserved to identify hardware traffic classes normally reported
via dev->num_tc. For Rx queue-group configuration, the gap is that the ingress/clsact qdisc
does not expose a set of virtual qdiscs similar to HW traffic classes in mqprio.
This was discussed in Slide 20 from Netdev 0x14 
(https://legacy.netdevconf.info/0x14/pub/slides/28/Application%20Device%20Queues%20for%20system-level%20network%20IO%20performance%20improvements.pdf)

-Amritha

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel