Re: [PATCH net-next 2/2] net: dsa: mv88e6xxx: implement egress tbf qdisc for 6393x family

Hi Sunil,

On 6/12/23 08:34, Sunil Kovvuri wrote:

On Fri, Jun 9, 2023 at 11:08 PM Alexis Lothoré <alexis.lothore@bootlin.com
<mailto:alexis.lothore@bootlin.com>> wrote:

    On 6/9/23 19:16, Andrew Lunn wrote:
    >> Yes, I can do that (or maybe -EINVAL to match Vladimir's comment ?). I think
    >> it's worth mentioning that I encountered an issue regarding those values
    during
    >> tests: I use tc program to set the tbf, and I observed that tc does not even
    >> reach kernel to set the qdisc if we pass no burst/latency value OR if we
    set it
    >> to 0. So tc enforces right on userspace side non-zero value for those
    >> parameters, and I have passed random values and ignored them on kernel side.
    >
    > That is not good. Please take a look around and see if any other
    > driver offloads TBF, and what they do with burst.
    >
    >> Checking available doc about tc-tbf makes me feel like that indeed a TBF
    qdisc
    >> command without burst or latency value makes no sense, except my use case can
    >> not have such values. That's what I struggled a bit to find a proper qdisc to
    >> match hardware cap. I may fallback to a custom netlink program to improve
    testing.
    >
    > We don't really want a custom application, since we want users to use
    > TC to set this up.
    >
    > Looking at the 6390 datasheet, Queue Counter Registers, mode 8 gives
    > the number of egress buffers for a port. You could validate that the
    > switch has at least the requested number of buffers assigned to the
    > port? There is quite a bit you can configure, so maybe there is a way
    > to influence the number of buffers, so you can actually implement the
    > burst parameter?

    Thanks for the pointers. I will check the egress buffers configuration and see
    if I can come up with something better


For setting up simple per-port ratelimit, instead of TBF isn't "egress matchall"
suitable here ?

I guess you are suggesting matchall + policer ? At first glance, I see no
obvious elements showing if one or another is more relevant. From user point of
view, controls are pretty much the same (rate + burst at least), but it looks
like policer is more of a pass/drop action, contrary to TBF which has some delay
notions, so it would solve the latency/limit absence of control. I am not sure
how it would look like on kernel side and how it would behave (how is managed
the filter, how can the policer be offloaded). I see some port_policer_add/del
callbacks in DSA, I will take a look at that as well and check differences with
TBF. Thanks for the suggestion.

Alexis

Thanks,
Sunil. 

-- 
Alexis Lothoré, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com