Re: [dpdk-dev] [PATCH v2 2/2] ethdev: add new offload flag to keep CRC

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 3/29/2018 8:56 AM, Shahaf Shuler wrote:
> Thursday, March 29, 2018 10:43 AM, Thomas Monjalon:
>> 29/03/2018 07:38, Shahaf Shuler:
>>> Wednesday, March 21, 2018 9:48 PM, Ferruh Yigit:
>>>> DEV_RX_OFFLOAD_KEEP_CRC offload flag added.
>>>>
>>>> DEV_RX_OFFLOAD_CRC_STRIP flag will remain one more release but
>>>> default behavior in PMDs is to strip the CRC independent from this flag.
>>>>
>>>> Until DEV_RX_OFFLOAD_CRC_STRIP flag is removed:
>>>> - Setting both KEEP_CRC & CRC_STRIP is INVALID
>>>> - Setting only CRC_STRIP PMD should strip the CRC
>>>> - Setting only KEEP_CRC PMD should keep the CRC
>>>> - Not setting both PMD should strip the CRC
>>>
>>> We cannot have such default behavior, it may break existing applications.
>>>
>>> The API of ethdev offloads says (even though it is not well documented) :
>> DEV_RX_OFFLOAD_CRC_STRIP (emphasis the STRIP).
>>> meaning, if set -> STRIP, if not set -> don't strip.  I am aware to at least one
>> application which wants to have the CRC, and for that purpose it naturally
>> don't set the offload flag.
>>>
>>> The fact some PMDs lack the ability to toggle the CRC stripping should be
>> exposed in the "limitations" section of their related guide.
>>>
>>> Up to here, this is the behavior as defined by the API.
>>>
>>> Now, we want to change it, and I think it makes sense. However I think we
>> should take similar approach to what we did with the ethdev offloads API:
>>> 1. at first stage a new offload flag is exposed DEV_RX_OFFLOAD_KEEP_CRC
>> and implemented on the PMDs.
>>
>> This is what is proposed above, except that we change the default
>> behaviour.
>> If we introduce a new flag which is the contrary of the old one, we need to
>> choose which one is the default, so which one has no effect.
> 
> The default behavior should remain as long as we don't deprecate the DEV_RX_OFFLOAD_CRC_STRIP flag.
> 
> Logic should be :
> Until DEV_RX_OFFLOAD_CRC_STRIP flag is removed:
>  - Setting both KEEP_CRC & CRC_STRIP is INVALID - enforced by ethdev. 
>  - Setting only CRC_STRIP PMD should strip the CRC
>  - Setting only KEEP_CRC PMD should keep the CRC
>  - Not setting both PMD should *not* strip the CRC 

Hi Shahaf,

I think we have two options,

1- This is v1 of this patch and also what you suggest:
v18.05:
- Send deprecation notice

v18.08:
- Add KEEP_CRC flag
- Change the meaning not setting CRC_STRIP to "strip the CRC"

v18.11:
- Remove the CRC_STRIP flag completely

I think this is more proper but takes more time.


2- Based on all the reality that all devices are doing CRC strip already:

v18.05:
- Add KEEP_CRC flag
- Change the meaning not setting CRC_STRIP to "strip the CRC"

v18.08:
- Remove the CRC_STRIP flag completely


With option two since not setting both KEEP_CRC & CRC_STRIP will mean "strip the
CRC", it won't require a change in PMDs, only we can pay attention to get new
PMDs according plan.

This can be more problematic for the case that application ask for keeping CRC,
because the way to say this was not setting CRC_STRIP, it changed to setting
KEEP_CRC without notification. This can be less problem since many PMD already
doesn't support keep crc.

Are you aware of any use case that is broken with option 2?


> 
>>
>>> 2. there is a conversion function on ethdev. Which for example converts
>> ~DEV_RX_OFFLOAD_CRC_STRIP -> DEV_RX_OFFLOAD_KEEP_CRC for the
>> PMDs.
>>> 3. deprecation of DEV_RX_OFFLOAD_CRC_STRIP for applications and
>> remove of the conversion functions.
>>
>>
>>
>