Re: Per queue stats

[Bruce Richardson <bruce.richardson@intel.com>]

On Tue, Oct 08, 2024 at 11:58:37AM +0200, Morten Brørup wrote:
> > From: Ferruh Yigit [mailto:ferruh.yigit@amd.com]
> > Sent: Sunday, 6 October 2024 22.51
> > 
> > On 10/4/2024 9:40 PM, Stephen Hemminger wrote:
> > > ...
> > >>>>> The stats queue mapping was a feature that was hinted at being
> > removed.
> > >>>>> It only exists because of HW limitations on Intel ixgbe NIC and
> > SW
> > >>>>> limitations from RTE_ETHDEV_QUEUE_STAT_CNTRS.
> > >>>>>
> > >>>>
> > >>>>
> > >>>> We have a plan to remove 'RTE_ETHDEV_QUEUE_STAT_CNTRS', by moving
> > queue
> > >>>> stats to xstats.
> > >>>>
> > >>>> But ixgbe limitation is there.
> > >>>>
> > >>>>> Perhaps there should be a generic SW emulation for this the
> > mapping?
> > >>>>>
> > >>>>
> > >>>> Ack, cc'ed Bruce.
> > >>>> But I am not sure ROI of the effort at this stage.
> > >>>
> > >>> Not sure what the specific ask for me is here. :-) Overall, I think
> > moving
> > >>> queue stats to xstats is the best way to go.
> > >>>
> > >>
> > >> cc'ed because of "generic SW emulation" comment.
> > >>
> > >> I was thinking if this mapping can be done transparent to the user
> > by
> > >> driver mapping queue <-> stats_register before reading stats, but
> > @Bruce
> > >> let me know this won't work because the stats tracking only happens
> > >> after the mapping.
> > >>
> > >> @Stephen, do you have something specific in your mind for SW
> > emulation
> > >> for mapping?
> > >
> > > Probably not worth doing mapping tables for SW drivers.
> > >
> > > My preference would be something like:
> > > 	1. Introduce new API's rte_ethdev_get_queue_stats() and
> > rte_ethdev_reset_queue_stats().
> > >            with callbacks in ethdev ops.
> > >            if driver does not support this, then do a shim in ethdev
> > that uses old stats fields.
> > >         2. Mark the stats mapping API's as deprecated and always
> > return error.
> > >            Making special case for ixgbe at this point really is not
> > worth maintaining.
> > >         3. After all drivers converted over, drop the fields in
> > ethdev stats for queues.
> > >            And fix the examples, etc. Yes this would be a big API
> > change.
> > >
> > > Would take a year to get done (ie 25.11).
> > >
> > 
> > The problem with existing a API, 'rte_eth_stats_get()', is it has queue
> > stats fixed sized array which limits us also for some cases causes
> > unnecessary memory consumption.
> > 
> > Our solution for now is move queue stats to xstats, as it is flexible
> > this solves the problem, but as you said in other thread xstats is free
> > form and may be hard for the application to use it.
> > 
> > One option is we can add some more data to the xstats, to help
> > application to consume it easier, like number of queues just before the
> > queue stats and define queue stats formatting syntax.
> > 
> > But if we will go with API option, instead of adding new queue stats
> > ('rte_ethdev_get_queue_stats()'), we can update the existing
> > 'rte_eth_stats_get()' to return dynamic number of queue stats.
> 
> Some users might not care about queue stats, so let's remove them entirely from rte_eth_stats_get(), and put them elsewhere.
> 
+1 for this. I think separate APIs for queue stats would make more sense.

/Bruce