On 5/14/16, 5:49 AM, Jiri Pirko wrote:
> Fri, May 13, 2016 at 08:47:48PM CEST, ro...@cumulusnetworks.com wrote:
>> On 5/12/16, 11:03 PM, Jiri Pirko wrote:
>>> Thu, May 12, 2016 at 11:10:08PM CEST, ro...@cumulusnetworks.com wrote:
>>>> On 5/12/16, 4:48 AM, Jiri Pirko wrote:
>>>>> From: Nogah Frankel <nog...@mellanox.com>
>>>>>
>>>>> Till now we had a ndo statistics function that returned SW statistics.
>>>>> We want to change the "basic" statistics to return HW statistics if
>>>>> available.
>>>>> In this case we need to expose a new ndo to return the SW statistics.
>>>>> Add a new ndo declaration to get SW statistics
>>>>> Add a function that gets SW statistics if a competible ndo exist
>>>>>
>>>>> Signed-off-by: Nogah Frankel <nog...@mellanox.com>
>>>>> Reviewed-by: Ido Schimmel <ido...@mellanox.com>
>>>>> Signed-off-by: Jiri Pirko <j...@mellanox.com>
>>>>> ---
>>>>>
>>>> To me netdev stats is combined 'SW + HW' stats for that netdev.
>>>> ndo_get_stats64 callback into the drivers does the magic of adding HW stats
>>>> to SW (netdev) stats and returning (see enic_get_stats). HW stats is
>>>> available for netdevs
>>>> that are offloaded or are backed by hardware. SW stats is the stats that
>>>> the driver maintains
>>>> (logical or physical). HW stats is queried and added to the SW stats.
>>> I'm not sure I follow. HW stats already contain SW stats. Because on
>>> slow path every packet that is not offloaded and goes through kernel is
>>> counted into HW stats as well (because it goes through HW port).
>> yes, correct... we don't want to double count those. But since these stats
>> are
>> generally queried from hw, I am calling them HW stats.
>> you will not really maintain a software counter for this. But, the driver
>> can maintain its own
>> counters for rx and tx errors etc and I call these SW stats. They are
>> counted at the driver.
>>
>>> If you
>>> do HW stats + SW stats, what you get makes no sense. Am I missing something?
>> If you go by my definition of HW and SW stats above, on a ndo_get_stats64()
>> call,
>> you will add the SW counters + HW counters and return. In my definition, the
>> pkts
>> that was rx'ed or tx'ed successfully are always in the HW count.
>>
>>> Btw, looking at enic_get_stats, looks exactly what we introduce for
>>> mlxsw in this patchset.
>> In enic_get_stats, the ones counted in software are the ones taken from
>> 'enic->'
>> net_stats->rx_over_errors = enic->rq_truncated_pkts;
>> net_stats->rx_crc_errors = enic->rq_bad_fcs;
>>
>>> With this patchset, we only allow user to se the actual stats for
>>> slow-path aka SW stats.
>> hmm...ok. But i am not sure how many will use this new attribute.
>> When you do 'ip -s link show' you really want all counters on that port
>> hardware or software does not matter at that point.
>>
>> My suggestion to move this to ethtool like attribute is because that is an
>> existing
>> way to break down your stats which ever way you want. And the best part is
>> it can be
>> customized (say rx_pkts_cpu_saw)
> I bevieve that ethtool is really not a place to expose sw stats. Does
> not make sense.
2 things:
- i was surprised you don't want your ndo_get_stats64 to be a unified view of
HW and SW stats
- by bringing up ethtool like stats (IFLA_STATS_LINK_HW_EXTENDED) I am just
saying
it has always been a way to breakdown stats. If you don't want to show explicit
SW stats there,
there is always a way to show HW only stats....and now you know the delta
between the unified stats
and the HW only stats is your SW stats.
