On Tue, 19 Mar 2019 17:18:08 +0000
Ferruh Yigit <ferruh.yi...@intel.com> wrote:
> On 3/14/2019 3:13 PM, David Marchand wrote:
> > Introduce a new api to retrieve per queue statistics from the drivers.
> > The api objectives:
> > - easily add some common per queue statistics and have it exposed
> > through the user xstats api while the user stats api is left untouched
> > - remove the limitations on the per queue statistics count (inherited
> > from ixgbe) and avoid recurrent bugs on stats array overflow
>
> The patch is adding two new dev_ops 'rxq_stats_get' & 'txq_stats_get', my
> concern is if it is overkill to have three dev_ops to get stats
> and I am feeling that is making xstat code more complex.
>
> Would it be simpler to add 'q_ierrors' & 'q_oerrors' to 'struct
> rte_eth_stats'?
>
> And perhaps we can do the 'fix rxq q_errors' patchset [1] after this change,
> so
> fix can be done with less changes, although it will push the fix into next
> release because of the ABI break.
> OR ethdev will be broken this release, because of max_mtu, since ABI is
> already
> broken perhaps we can squeeze this in.
>
> Overall I would like to get more comment on this, Andrew, Thomas?
>
> >
> > Signed-off-by: David Marchand <david.march...@redhat.com>
>
> <...>
>
My preference would be:
1. Make all DPDK drivers consistent in usage of current statistic values.
2. Propose an enhancement to have new ethdev statistics match some
pre-existing
standard like SNMP or other RFC.
3. Reduce custom (xstats) values by using #2. Leave it for driver specific
stuff.
I.e: don't modify existing API at all, make a new one.
PS: ethdev is one of those structures that needs to get removed/hidden from
application headers. It should be possible to add/remove stuff from ethdev
internals, device and bus
without breaking API/ABI.
