Re: [PATCH net-next v2 0/5] net: sparx5: add tc matchall port policer offload

[Paolo Abeni <pabeni@redhat.com>]

On 11/5/26 10:39, Daniel Machon wrote:
>>> This series adds offload of tc matchall filters with a police action to
>>> the ingress port policers of Sparx5 and lan969x, which are both served
>>> by the sparx5 driver.
>>>
>>> In hardware, each ingress port has four single-rate leaky bucket
>>> policers. A policer drops the frames that exceed its rate and burst.
>>> The four policers of a port share the pass and drop statistics counters
>>> of the port.
>>>
>>> The driver uses one policer per port, because with more than one, the
>>> statistics reported to tc could not be attributed to the right policer.
>>> This matches lan966x and ocelot. The policer meters each frame without
>>> preamble and inter-frame gap, so the configured rate is a data rate, as
>>> tc police expects and as on lan966x and ocelot.
>>>
>>> Since the policer meters all ingress frames on the port, only filters on
>>> chain 0 with protocol all are offloaded, and filters on shared blocks
>>> are rejected. A police action that is already offloaded as a port
>>> policer on another port is rejected too, since in software it is a
>>> single bucket for all of its users. Only drop is supported as the exceed
>>> action, and the police overhead parameter is not supported.
>>>
>>> Patch #1 drops a redundant prefix from the existing POL_UPD_INT_CFG
>>> register macros.
>>>
>>> Patch #2 adds the registers used by the port policer.
>>>
>>> Patch #3 tracks whether the ingress block of a port is shared, and adds
>>> the per-port policer state.
>>>
>>> Patch #4 adds the hardware programming of the port policer.
>>>
>>> Patch #5 adds the tc matchall offload.
>>>
>>> Signed-off-by: Daniel Machon <daniel.machon@microchip.com>
>> I understand the following will be very frustrating, but for unknown
>> reason AI review did not kick on this series.
> 
> Yes, I noticed that. I thought you maybe had some magic button you could push,
> in order to trigger the review manually.

Unfortunately, I personally don't have it. AI review is sponsored by meta,
only Jakub has direct access to that piece of the infra, and he is AFK for a bit.

>> The changes looks sane to me, and addressing points raised on v1, but
>> I think nowadays we need an explicit AI feedback.
> 
> Ack. I think the missing sashiko-nipa check is not isolated to this patchset
> alone. Just FYI.
Indeed I hit more patches/series without AI review. In some cases I can run
the model locally (for single patch, due to many constraints).

Thanks,

Paolo