On 9/29/2021 7:38 AM, Andrew Rybchenko wrote:
> On 9/28/21 8:10 PM, Stephen Hemminger wrote:
>> On Tue, 28 Sep 2021 19:48:54 +0300
>> Andrew Rybchenko <andrew.rybche...@oktetlabs.ru> wrote:
>>
>>> Some drivers do not provide per-queue statistics. So, there is no point
>>> to have these misleading zeros in xstats.
>>>
>>> Fixes: f30e69b41f94 ("ethdev: add device flag to bypass auto-filled queue
>>> xstats")
>>> Cc: sta...@dpdk.org
>>>
>>> Signed-off-by: Andrew Rybchenko <andrew.rybche...@oktetlabs.ru>
>> Really?
>> It is useful to have zeros rather than random data there.
>
> I guess there is a misunderstanding here. Auto-filling xstats is
> an addition of per-queue basic statistics to xstats by ethdev
> layer. It makes sense to do it if and only if there is some
> sensible data there.
>
> There is a related deprecation notice saying that per-queue
> stats should be removed from basic stats since per-queue
> stats should be provided by xstats API natively.
>
> Basically RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS means
> that corresponding driver is not ready vs the deprecation notice.
> So, I want to clean it up to see not yet ready drivers only.
>
As you said, 'RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS' flag set by driver means,
driver is not ready on representing queue stats in xstats and ethdev layer is
filling it automatically from basic stats.
First we should wait for drivers to implement it, later clean queue stats from
basic stats and remove the flag.
I am not sure if we can remove the deprecation notice in this release, but agree
to add a deadline for the drivers, which can be 22.11.
