Re: [PATCH net-next v17 12/14] net: ethtool: tsinfo: Add support for reading tsinfo for a specific hwtstamp provider

[Jacob Keller <jacob.e.keller@intel.com>]

On 7/26/2024 12:04 PM, Kory Maincent wrote:
> Hello Jacob,
> 
> Thanks a lot for your full review! 
> 
> On Wed, 17 Jul 2024 10:35:20 -0700
> Jacob Keller <jacob.e.keller@intel.com> wrote:
> 
>> On 7/9/2024 6:53 AM, Kory Maincent wrote:
>>  [...]  
>>
>> Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
>>
>> One thing which applies more broadly to the whole series, but I see the
>> focus right now is on selecting between NETDEV and PHYLIB.
>>
>> For ice (E800 series) hardware, the timestamps are captured by the PHY,
>> but its not managed by phylib, its managed by firmware. In our case we
>> would obviously report NETDEV in this case. The hardware only has one
>> timestamp point and the fact that it happens at the PHY layer is not
>> relevant since you can't select or change it.
>>
>> There are some future plans in the work for hardware based on the ixgbe
>> driver which could timestamp at either the MAC or PHY (with varying
>> trade-offs in precision vs what can be timestamped), and (perhaps
>> unfortunately), the PHY would likely not manageable by phylib.
>>
>> There is also the possibility of something like DMA or completion
>> timestamps which are distinct from MAC timestamps. But again can have
>> varying trade offs.
> 
> As we already discussed in older version of this patch series the
> hwtstamp qualifier will be used to select between IEEE 1588 timestamp or DMA
> timestamp. See patch 8 :
> +/*
> + * Possible type of htstamp provider. Mainly "precise" the default one
> + * is for IEEE 1588 quality and "approx" is for NICs DMA point.
> + */
> 
> We could add other enumeration values in the future if needed, to manage new
> cases.
> 
> Just figured out there is a NIT in the doc. h*w*tstamp.
> 

Ah, perfect, thanks for the clarification!

>> I'm hopeful this work can be extended somehow to enable selection
>> between the different mechanisms, even when the kernel device being
>> represented is the same netdev.
> 
> Another nice features would be the support for simultaneous hardware timestamp
> but I sadly won't be able to work on this.
> > Regards,

Yes this would be useful, though I think we're somewhat limited by the
API that returns to userspace currently.