On Mon, 28 Nov 2022 11:12:38 +0000 Ferruh Yigit <ferruh.yi...@amd.com> wrote:
> On 11/26/2022 5:34 PM, Stephen Hemminger wrote: > > On Fri, 25 Nov 2022 19:16:00 -0800 > > Rushil Gupta <rush...@google.com> wrote: > > > >>>> > >>>> The driver should be filling in the per-queue stats as well. > >>>> q_ipackets, q_opackets, q_ibytes, q_obytes, q_errors > >>> > >>> > >>> Hi Stephen, > >>> > >>> Queue stats moved to xstats, and there is a long term goal to move all > >>> queue stats from basic stats to xstats for all PMDs, and remove interim > >>> 'RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS' flag. > >>> > >>> That is why request to new PMDs is to not introduce queue stats in basic > >>> stats, but in xstats. > > > > Agree that xstats are better but: > > > > * the current checked in version of GVE does not have driver op for xstats > > > > * if driver fills in the q_XXX[] stats then eth_dev_get_xstats_basic will be > > able to use it. > > > > That is an option, but it will require driver to add > 'RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS' flag what we want to get rid of in > long term. > That is why we are requesting new driver to add the xstats > implementation instead of adding queue support in basic stats. It is up > to PMD to provide xstats implementation if they want queue stats. Agreed, having xstats in driver is the best way. Doing what virtio does now would be a good example.
