[PATCH v3 0/3] introduce maximum Rx buffer size

[Ferruh Yigit]

On 11/2/2023 4:51 PM, Morten Brørup wrote:
>> From: Ferruh Yigit [mailto:ferruh.yigit at amd.com]
>> Sent: Thursday, 2 November 2023 17.24
>>
>> On 11/2/2023 1:59 AM, lihuisong (C) wrote:
>>>
>>> 在 2023/11/2 0:08, Stephen Hemminger 写道:
>>>> On Wed, 1 Nov 2023 10:36:07 +0800
>>>> "lihuisong (C)" <lihuisong at huawei.com> wrote:
>>>>
>>>>>> Do we need to report this size? It's a common feature for all
>> PMDs.
>>>>>> It would make sense then to have max_rx_bufsize set to 16K by
>> default
>>>>>> in ethdev, and PMD could then raise/lower based on hardware.
>>>>> It is not appropriate to set to 16K by default in ethdev layer.
>>>>> Because I don't see any check for the upper bound in some driver,
>> like
>>>>> axgbe, enetc and so on.
>>>>> I'm not sure if they have no upper bound.
>>>>> And some driver's maximum buffer size is "16384(16K) - 128"
>>>>> So it's better to set to UINT32_MAX by default.
>>>>> what do you think?
>>>> The goal is always giving application a working upper bound, and
>>>> enforcing
>>>> that as much as possible in ethdev layer. It doesnt matter which
>> pattern
>>>> does that.  Fortunately, telling application an incorrect answer is
>>>> not fatal.
>>>> If over estimated, application pool would be wasting space.
>>>> If under estimated, application will get more fragmented packets.
>>> I know what you mean.
>>> If we set UINT32_MAX, it just means that driver don't report this
>> upper
>>> bound.
>>> This is also a very common way of handling. And it has no effect on
>> the
>>> drivers that doesn't report this value.
>>> On the contrary, if we set a default value (like 16K) in ethdev, user
>>> may be misunderstood and confused by that, right?
>>> After all, this isn't the real upper bound of all drivers. And this
>>> fixed default value may affect the behavior of some driver that I
>> didn't
>>> find their upper bound.
>>> So I'd like to keep it as UINT32_MAX.
>>>
>>
>>
>> Hi Stephen, Morten,
>>
>> I saw scattered Rx mentioned, there may be some misalignment,
>> the purpose of the patch is not to enable application to set as big as
>> possible mbuf size, so that application can escape from parsing
>> multi-segment mbufs.
>> Indeed application can provide a large mbuf anyway, to have same
>> result,
>> without knowing this information.
>>
>> Main motivation is other way around, device may have restriction on
>> buffer size that a single descriptor can address, independent from
>> scattered Rx used, if mbuf size is bigger than this device limit, each
>> mbuf will have some unused space.
>> Patch has intention to inform this max per mbuf/descriptor buffer size,
>> so that application doesn't allocate bigger mbuf and waste memory.
> 
> Good point!
> 
> Let's categorize this patch series as a memory optimization for applications that support jumbo frames, but are trying to avoid (or reduce) scattered RX. :-)
> 

It is a memory optimization patch, but again nothing to do with jumbo
frames or scattered Rx.