Message-ID: <b7e1cb8b-b6b1-c0fa-3864-4036750f3164@ti.com>
Date:   Mon, 30 Dec 2019 12:03:46 -0500
From:   Murali Karicheri <m-karicheri2@...com>
To:     Vinicius Costa Gomes <vinicius.gomes@...el.com>,
        Ivan Khoronzhuk <ivan.khoronzhuk@...aro.org>
CC:     Po Liu <po.liu@....com>,
        Andre Guedes <andre.guedes@...ux.intel.com>,
        "alexandru.ardelean@...log.com" <alexandru.ardelean@...log.com>,
        "allison@...utok.net" <allison@...utok.net>,
        "andrew@...n.ch" <andrew@...n.ch>,
        "ayal@...lanox.com" <ayal@...lanox.com>,
        "davem@...emloft.net" <davem@...emloft.net>,
        "f.fainelli@...il.com" <f.fainelli@...il.com>,
        "gregkh@...uxfoundation.org" <gregkh@...uxfoundation.org>,
        "hauke.mehrtens@...el.com" <hauke.mehrtens@...el.com>,
        "hkallweit1@...il.com" <hkallweit1@...il.com>,
        "jiri@...lanox.com" <jiri@...lanox.com>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "pablo@...filter.org" <pablo@...filter.org>,
        "saeedm@...lanox.com" <saeedm@...lanox.com>,
        "tglx@...utronix.de" <tglx@...utronix.de>,
        Vladimir Oltean <vladimir.oltean@....com>,
        "simon.horman@...ronome.com" <simon.horman@...ronome.com>,
        Claudiu Manoil <claudiu.manoil@....com>,
        Alexandru Marginean <alexandru.marginean@....com>,
        Xiaoliang Yang <xiaoliang.yang_1@....com>,
        Roy Zang <roy.zang@....com>, Mingkai Hu <mingkai.hu@....com>,
        Jerry Huang <jerry.huang@....com>, Leo Li <leoyang.li@....com>
Subject: Re: [EXT] Re: [v1,net-next, 1/2] ethtool: add setting frame
 preemption of traffic classes

Hi Vinicius,

On 12/18/2019 08:54 PM, Vinicius Costa Gomes wrote:
> Hi Ivan,
> 
> Ivan Khoronzhuk <ivan.khoronzhuk@...aro.org> writes:
> 
>>>>> Quoting Po Liu (2019-11-27 01:59:18)
>>>>>> IEEE Std 802.1Qbu standard defined the frame preemption of port
>>>>>> traffic classes. This patch introduce a method to set traffic classes
>>>>>> preemption. Add a parameter 'preemption' in struct
>>>>>> ethtool_link_settings. The value will be translated to a binary, each
>>>>>> bit represent a traffic class. Bit "1" means preemptable traffic
>>>>>> class. Bit "0" means express traffic class.  MSB represent high number
>>>>>> traffic class.
>>>>>>
>>>>>> If hardware support the frame preemption, driver could set the
>>>>>> ethernet device with hw_features and features with NETIF_F_PREEMPTION
>>>>>> when initializing the port driver.
>>>>>>
>>>>>> User can check the feature 'tx-preemption' by command 'ethtool -k
>>>>>> devname'. If hareware set preemption feature. The property would be a
>>>>>> fixed value 'on' if hardware support the frame preemption.
>>>>>> Feature would show a fixed value 'off' if hardware don't support the
>>>>>> frame preemption.
>>>
>>> Having some knobs in ethtool to enable when/how Frame Preemption is
>>> advertised on the wire makes sense. I also agree that it should be "on"
>>> by default.
>>>
>>>>>>
>>>>>> Command 'ethtool devname' and 'ethtool -s devname preemption N'
>>>>>> would show/set which traffic classes are frame preemptable.
>>>>>>
>>>>>> Port driver would implement the frame preemption in the function
>>>>>> get_link_ksettings() and set_link_ksettings() in the struct ethtool_ops.
>>>>>
>>>>> In an early RFC series [1], we proposed a way to support frame preemption. I'm
>>>>> not sure if you have considered it before implementing this other proposal
>>>>> based on ethtool interface so I thought it would be a good idea to bring that up
>>>>> to your attention, just in case.
>>>>
>>>> Sorry, I didn't notice the RFC proposal. Using ethtool set the
>>>> preemption just thinking about 8021Qbu as standalone. And not limit to
>>>> the taprio if user won't set 802.1Qbv.
>>>
>>> I see your point of using frame-preemption "standalone", I have two
>>> ideas:
>>>
>>> 1. add support in taprio to be configured without any schedule in the
>>> "full offload" mode. In practice, allowing taprio to work somewhat
>>> similar to (mqprio + frame-preemption), changes in the code should de
>>> fairly small;
>>
>> +
>>
>> And if follow mqprio settings logic then preemption also can be enabled
>> immediately while configuring taprio first time, and similarly new ADMIN
>> can't change it and can be set w/o preemption option afterwards.
>>
>> So that following is correct:
>>
>> OPER
>> $ tc qdisc add dev IFACE parent root handle 100 taprio \
>>        base-time 10000000 \
>>        num_tc 3 \
>>        map 2 2 1 0 2 2 2 2 2 2 2 2 2 2 2 2 \
>>        queues 1@0 1@1 2@2 \
>>        preemption 0 1 1 1
>>        flags 1
>>
>> then
>> ADMIN
>> $ tc qdisc add dev IFACE parent root handle 100 taprio \
>>        base-time 12000000 \
>>        num_tc 3 \
>>        map 2 2 1 0 2 2 2 2 2 2 2 2 2 2 2 2 \
>>        queues 1@0 1@1 2@2 \
>>        preemption 0 1 1 1
>>        sched-entry S 01 300000 \
>>        sched-entry S 02 300000 \
>>        flags 1
>>
>> then
>> ADMIN
>> $ tc qdisc add dev IFACE parent root handle 100 taprio \
>>        base-time 13000000 \
>>        sched-entry S 01 300000 \
>>        sched-entry S 02 300000 \
>>        flags 1
>>
>> BUT:
>>
>> 1) The question is only should it be in this way? I mean preemption to be
>> enabled immediately? Also should include other parameters like
>> fragment size.
> 
> We can decide what things are allowed/useful here. For example, it might
> make sense to allow "preemption" to be changed. We can extend taprio to
> support changing the fragment size, if that makes sense.
> 
>>
>> 2) What if I want to use frame preemption with another "transmission selection
>> algorithm"? Say another one "time sensitive" - CBS? How is it going to be
>> stacked?
> 
> I am not seeing any (conceptual*) problems when plugging a cbs (for
> example) qdisc into one of taprio children. Or, are you talking about a
> more general problem?
> 
> * here I am considering that support for taprio without an schedule is
>   added.
> 
>>
>> In this case ethtool looks better, allowing this "MAC level" feature, to be
>> configured separately.
> 
> My only issue with using ethtool is that then we would have two
> different interfaces for "complementary" features. And it would make
> things even harder to configure and debug. The fact that one talks about
> traffic classes and the other transmission queues doesn't make me more
> comfortable as well.
> 
> On the other hand, as there isn't a way to implement frame preemption in
> software, I agree that it makes it kind of awkward to have it in the tc
> subsystem.
Absolutely. I think frame pre-emption feature flag, per queue express/
pre-empt state, frag size, timers (hold/release) to be configured
independently (perhaps through ethtool) and then taprio should check
this with the lower device and then allow supporting additional Gate
operations such as Hold/release if supported by underlying device.

What do you think? Why to abuse tc for this?

Thanks and regards,

Murali
> 
> At this point, I am slightly in favor of the taprio approach (yes, I am
> biased :-), but I can be convinced otherwise. I will be only a little
> sad if we choose to go with ethtool for now, and then add support up in
> the stack, something similar to "ethtool -N" and "tc-flower".
> 

>>
>>>
>>> 2. extend mqprio to support frame-preemption;
>>>
>>>>
>>>> As some feedback  also want to set the MAC merge minimal fragment size
>>>> and get some more information of 802.3br.
>>>
>>> The minimal fragment size, I guess, also makes sense to be kept in
>>> ethtool. That is we have a sane default, and allow the user to change
>>> this setting for special cases.
>>>
>>>>
>>>>>
>>>>> In that initial proposal, Frame Preemption feature is configured via taprio qdisc.
>>>>> For example:
>>>>>
>>>>> $ tc qdisc add dev IFACE parent root handle 100 taprio \
>>>>>        num_tc 3 \
>>>>>        map 2 2 1 0 2 2 2 2 2 2 2 2 2 2 2 2 \
>>>>>        queues 1@0 1@1 2@2 \
>>>>>        preemption 0 1 1 1 \
>>>>>        base-time 10000000 \
>>>>>        sched-entry S 01 300000 \
>>>>>        sched-entry S 02 300000 \
>>>>>        sched-entry S 04 400000 \
>>>>>        clockid CLOCK_TAI
>>>>>
>>>>> It also aligns with the gate control operations Set-And-Hold-MAC and Set-And-
>>>>> Release-MAC that can be set via 'sched-entry' (see Table 8.7 from
>>>>> 802.1Q-2018 for further details.
>>>>
>>>> I am curious about Set-And-Hold-Mac via 'sched-entry'. Actually, it
>>>> could be understand as guardband by hardware preemption. MAC should
>>>> auto calculate the nano seconds before  express entry slot start to
>>>> break to two fragments. Set-And-Hold-MAC should minimal larger than
>>>> the fragment-size oct times.
>>>
>>> Another interesting point. My first idea is that when the schedule is
>>> offloaded to the driver and the driver detects that the "entry" width is
>>> smaller than the fragment side, the driver could reject that schedule
>>> with a nice error message.
>>
>> Looks ok, if entry command is RELEASE or SET only, but not HOLD, and
>> only if it contains express queues. And if for entry is expectable to have
>> interval shorter, the entry has to be marked as HOLD then.
>>
>> But not every offload is able to support mac/hold per sched (there is
>> no HOLD/RELEASE commands in this case). For this case seems like here can
>> be 2 cases:
> 
> Yeah, the hw I have in hand also doesn't support the HOLD/RELEASE
> commands.
> 
>>
>> 1) there is no “gate close” event for the preemptible traffic
>> 2) there is "gate close" event for the preemptable traffic
>>
>> And both can have the following impact, if assume the main reason to
>> this guard check is to guarantee the express queue cannot be blocked while
>> this "close to short" interval opening ofc:
>>
>> If a preemption fragment is started before "express" frame, then interval
>> should allow to complete preemption fragment and has to have enough time
>> to insert express frame. So here situation when maximum packet size per
>> each queue can have place.
>>
>> In case of TI am65 this queue MTU is configurable per queue (for similar
>> reasons and couple more (packet fill feature for instance)) and can be
>> used for guard check also, but not clear where it should be. Seems like
>> it should be done using ethtool also, but can be needed for taprio
>> interface....
> 
> For now, at least for the hardware I am working on, something like this
> is configurable, but I wasn't planning on exposing it, using the maximum
> ethernet frame size seemed a good default.
> 
>>
>>>>
>>>>>
>>>>> Please share your thoughts on this.
>>>>
>>>> I am good to see there is frame preemption proposal. Each way is ok
>>>> for me but ethtool is more flexible. I've seen the RFC the code. The
>>>> hardware offload is in the mainline, but preemption is not yet, I
>>>> don't know why. Could you post it again?
>>>
>>> It's not mainline because this kind of stuff will not be accepted
>>> upstream without in-tree users. And you are the first one to propose
>>> such a thing :-)
>>>
>>> It's just now that I have something that supports frame-preemption, the
>>> code I have is approaching RFC-like quality. I will send another RFC
>>> this week hopefully, and we can see how things look in practice.
>>>
>>>
>>> Cheers,
>>> --
>>> Vinicius
>>
>> -- 
>> Regards,
>> Ivan Khoronzhuk
> 
> Cheers,
> --
> Vinicius
> 

-- 
Murali Karicheri
Texas Instruments

Powered by blists - more mailing lists