Date:   Tue, 06 Sep 2022 09:17:39 +0200
From:   Michael Walle <michael@...le.cc>
To:     Vladimir Oltean <vladimir.oltean@....com>
Cc:     netdev@...r.kernel.org, "David S. Miller" <davem@...emloft.net>,
        Eric Dumazet <edumazet@...gle.com>,
        Jakub Kicinski <kuba@...nel.org>,
        Paolo Abeni <pabeni@...hat.com>,
        Xiaoliang Yang <xiaoliang.yang_1@....com>,
        Claudiu Manoil <claudiu.manoil@....com>,
        Alexandre Belloni <alexandre.belloni@...tlin.com>,
        UNGLinuxDriver@...rochip.com, Andrew Lunn <andrew@...n.ch>,
        Vivien Didelot <vivien.didelot@...il.com>,
        Florian Fainelli <f.fainelli@...il.com>,
        Vinicius Costa Gomes <vinicius.gomes@...el.com>,
        Maxim Kochetkov <fido_max@...ox.ru>,
        Colin Foster <colin.foster@...advantage.com>,
        Richie Pearn <richard.pearn@....com>,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH v2 net 1/3] net: dsa: felix: tc-taprio intervals smaller
 than MTU should send at least one packet

Am 2022-09-06 02:11, schrieb Vladimir Oltean:
> On Tue, Sep 06, 2022 at 12:53:20AM +0200, Michael Walle wrote:
>> I haven't looked at the overall code, but the solution described
>> above sounds good.
>> 
>> FWIW, I don't think such a schedule, where exactly one frame
>> can be sent, is very likely in the wild though. Imagine a piece
>> of software is generating one frame per cycle. It might happen
>> that during one (hardware) cycle there is no frame ready (because
>> it is software and it jitters), but then in the next cycle, there
>> are now two frames ready. In that case you'll always lag one frame
>> behind and you'll never recover from it.
>> 
>> Either I'd make sure I can send at two frames in one cycle, or
>> my software would only send a frame every other cycle.
> 
> A 10 us interval is a 10 us interval, it shouldn't matter if you slice
> it up as one 1250B frame, or two 500B frames, or four 200B frames, etc.
> Except with the Microchip hardware implementation, it does. In v1, we
> were slicing the 10 us interval in half for useful traffic and half for
> the guard band. So we could fit more small packets in 5 us. In v2, at
> your proposal, we are slicing it in 33 ns for the useful traffic, and
> 10 us - 33 ns for the guard band. This indeed allows for a single
> packet, be it big or small. It's how the hardware works; without any
> other input data point, a slicing point needs to be put somewhere.
> Somehow it's just as arbitrary in v2 as where it was in v1, just
> optimized for a different metric which you're now saying is less 
> practical.

I actually checked the code before writing and saw that one could
change the guard band by setting the MTU of the interface. I though,
"ah ok, then there is no issue". After sleeping, I noticed that you'd
restrict the size of all the frames on the interface. Doh ;)

-michael

> By the way, I was a fool in last year's discussion on guard bands for
> saying that there isn't any way for the user to control per-tc MTU.
> IEEE 802.1Qbv, later standardized as IEEE 802.1Q clause 8.6.8.4
> Enhancements for scheduled traffic, does contain a queueMaxSDUTable
> structure with queueMaxSDU elements. I guess I have no choice except to
> add this to the tc-taprio UAPI in a net-next patch, because as I've
> explained above, even though I've solved the port hanging issue, this
> hardware needs more fine tuning to obtain a differentiation between 
> many
> small packets vs few large packets per interval.

Powered by blists - more mailing lists