On Thu, 14 Oct 2021 12:44:37 +0000 Alvin Šipraga wrote:

On 10/13/21 5:13 PM, Jakub Kicinski wrote:

quoted

On Wed, 13 Oct 2021 08:33:36 +0000 Alvin Šipraga wrote:  

quoted

I implement the dsa_switch_ops callback .get_ethtool_stats, using an
existing function rtl8366_get_ethtool_stats in the switch helper library
rtl8366.c. It was my understanding that this is the correct way to
expose counters within the DSA framework - please correct me if that is
wrong.  

It's the legacy way, today we have a unified API for reporting those
stats so user space SW doesn't have to maintain a myriad string matches
to get to basic IEEE stats across vendors. Driver authors have a truly
incredible ability to invent their own names for standard stats. It
appears that your pick of names is also unique :)

It should be trivial to plumb the relevant ethtool_ops thru to
dsa_switch_ops if relevant dsa ops don't exist.

You should also populate correct stats in dsa_switch_ops::get_stats64
(see the large comment above the definition of struct
rtnl_link_stats64 for mapping). A word of warning there, tho, that
callback runs in an atomic context so if your driver needs to block it
has to read the stats periodically from a async work.  

OK, so just to clarify:

- get_ethtool_stats is deprecated - do not use

It can still be used, but standardized interfaces should be preferred
whenever possible, especially when appropriate uAPI already exists.

- get_eth_{phy,mac,ctrl,rmon}_stats is the new API - add DSA plumbing 
and use this

Yup.

- get_stats64 orthogonal to ethtool stats but still important - use also 
this

Yes, users should be able to depend on basic interface stats (packets,
bytes, crc errors) to be correct.

For stats64 I will need to poll asynchronously - do you have any 
suggestion for how frequently I should do that? I see one DSA driver 
doing it every 3 seconds, for example.

3 sec seems fine.