[ovs-dev] [RFC PATCH] netdev-dpdk: Expose per rxq/txq basic statistics.

Kevin Traynor ktraynor at redhat.com
Tue Nov 23 15:40:13 UTC 2021


On 23/11/2021 15:13, David Marchand wrote:
> On Thu, Nov 18, 2021 at 4:31 PM Kevin Traynor <ktraynor at redhat.com> wrote:
>>> @@ -1583,6 +1584,13 @@ netdev_dpdk_get_xstat_name(struct netdev_dpdk *dev, uint64_t id)
>>>        return dev->rte_xstats_names[id].name;
>>>    }
>>>
>>> +/* We filter out everything except per rxq/txq basic stats, and dropped,
>>> + * error and management counters.
>>> + * Note: rx_qX_errors is handled by the _errors$ pattern.
>>> + */
>>> +#define DPDK_STATS_REGEX_FILTER \
>>> +    "(^(rx_q|tx_q)[0-9]*_(packets|bytes)$|_errors$|_dropped$|_management_)"
>>> +
>>
>> iirc, there was some pmds that were not conforming correctly with this
>> rx_q syntax, but I *think* they were fixed a few releases ago in DPDK.
>> Just mentioning because if there was still some, they could be fixed or
>> the regex could be expanded to capture them.
> 
> In theory, those per q stats are part of the "basic" stats handled at
> the ethdev level.
> There should be no issue for them.
> 
> 
> I remember an issue with incorrect stats, but it was a conflict
> between OVS and i40e stats.
> Is this what you had in mind?
> 
> 

No, I had a thought that some driver for maybe Broadcom or Mellanox was 
reporting say 'rxq' instead of 'rx_q' and it was noticed and fixed while 
some other work on xstats. But it's hazy and I can't find it now, so 
best just to ignore unless someone hits it.



More information about the dev mailing list