[dpdk-dev] [PATCH 2/2] net/vhost: support mrg-rxbuf disabling

[Maxime Coquelin]

On 6/26/19 1:18 PM, Matan Azrad wrote:
> 
> 
> From: Maxime Coquelin
>> On 6/26/19 9:50 AM, Matan Azrad wrote:
>>> Hi Maxim
>>>
>>> Any response here?
>>>
>>> Besides that,
>>>
>>> Regarding the TSO and this patch:
>>> I think we shouldn't be so strict to not take them for this version:
>>> 1. The later time was a technical issue with the mailer - a mistake.
>>> 2. The patches don't change any default and makes sense - will not hurt
>> anyone.
>>>
>>> So I think we can do it beyond the letter of the law.
>>>
>>>    From: Maxime Coquelin
>>>    > Sent: Thursday, June 20, 2019 10:19 AM
>>>    > To: Matan Azrad <matan at mellanox.com>; Noa Ezra
>>>    <noae at mellanox.com>
>>>    > Cc: dev at dpdk.org
>>>    > Subject: Re: [PATCH 2/2] net/vhost: support mrg-rxbuf disabling
>>>    >
>>>    >
>>>    >
>>>    > On 6/20/19 8:52 AM, Matan Azrad wrote:
>>>    > > Hi all
>>>    > >
>>>    > >> -----Original Message-----
>>>    > >> From: Noa Ezra
>>>    > >> Sent: Thursday, June 20, 2019 8:58 AM
>>>    > >> To: Maxime Coquelin <maxime.coquelin at redhat.com>
>>>    > >> Cc: Matan Azrad <matan at mellanox.com>; dev at dpdk.org
>>>    > >> Subject: RE: [PATCH 2/2] net/vhost: support mrg-rxbuf disabling
>>>    > >>
>>>    > >> Hi Maxime,
>>>    > >> Thanks for your comment, please see below.
>>>    > >>
>>>    > >>> -----Original Message-----
>>>    > >>> From: Maxime Coquelin [mailto:maxime.coquelin at redhat.com]
>>>    > >>> Sent: Wednesday, June 19, 2019 12:10 PM
>>>    > >>> To: Noa Ezra <noae at mellanox.com>
>>>    > >>> Cc: Matan Azrad <matan at mellanox.com>; dev at dpdk.org
>>>    > >>> Subject: Re: [PATCH 2/2] net/vhost: support mrg-rxbuf disabling
>>>    > >>>
>>>    > >>> Hi Noa,
>>>    > >>>
>>>    > >>> On 6/19/19 8:13 AM, Noa Ezra wrote:
>>>    > >>>> Rx mergeable buffers is a virtio feature that allows chaining of
>>>    > >>>> multiple virtio descriptors to handle large packet size.
>>>    > >>>> This behavior is supported and enabled by default, however in
>>>    > >>>> case the user knows that rx mergeable buffers are not needed, he
>>>    > >>>> can disable the feature.
>>>    > >>>> The user should also set mrg_rxbuf=off in virtual machine's xml.
>>>    > >>>
>>>    > >>> I'm not sure to understand why it is needed, as the vhost-user
>>>    > >>> library supports the feature, it's better to let it being advertised.
>>>    > >>>
>>>    > >>> As you say, it is up to the user to disable it in the VM's XML.
>>>    > >>> Done this way, the feature won't be negotiated.
>>>    > >>>
>>>    > >> I agree with you, I'll remove this patch from the series.
>>>    > >
>>>    > > Are you sure that no performance impact exists for redundant
>>>    > > merg-rx-buf
>>>    > configuration here?
>>>    >
>>>    > I'm not sure to understand what you mean, could you please elaborate?
>>>    >
>>>    I guess that if this feature is enabled and the feature actually are not used
>>>    (no packets are scattered or merged) it will hurt the performance.
>>
>> Well, latest performance measurements does not show a big impact now on
>> enabling mergeable buffers feature unconditionaly.
>   
> Did you test small packets \ big?

64B packets, in non-vector mode on Virtio PMD side.

> 
>>>    So if one of the sides doesn't want to use it because of performance, it
>> may
>>>    want to disable it.
>>
>> And even if there is an impact, the way to disable it is through
>> Libvirt/Qemu.
> 
> Not sure, as TSO application may decide to not do it in spite of it is configured in Qemu.
>   
>>>    > > What if the second side want it and the current side no?
>>>    >
>>>    > The feature won't be negotiated, assuming it has been disabled in
>> QEMU
>>>    > cmdline (or via libvirt).
>>>    > > It may be that the vhost PMD user may want to disable it to save
>>>    > performance from some reasons, no?
>>>    > >
>>>    >
>>>    > Then this user should disable it at QEMU level.
>>>    >
>>>    So the vhost PMD is not one of the sides to decide?
>>>    If so, why do we need the APIs to configure the features?
>>
>> Are you talking about the rte_vhost_driver_set_features() and related
>> APIs?
>   
> Yes
> 
>> This is used for example by the external backends that support features
>> specific to the backend type (e.g. crypto), or also used by OVS-DPDK, to
>> disable TSO. So these usages are for functional reasons, not tuning.
> 
> Exactly, applications (like OVS) may decide to disable features because a lot of reasons.
> 
>>>    Looks like also the qemu is configured with the feature the VM\host sides
>>>    may decide in some cases to disable it.
>>
>> For functional reasons, I agree. So I that's why I agree with your tso
>> patch as the application has to support it, but that's not the case of
>> the mergeable buffers features.
> 
> Performance reasons are not good enough?

No, that's not what I mean.
I mean that the application should be able to disable a feature when it
does not meet the functional requirement.

For performance tuning, the qemu way is available, and enough.

> 
>> Tiwei, what's your opinion on this?
>>
>>>    > Regards,
>>>    > Maxime
>>>