Re: [PATCH V6 net-next 3/4] net: hns3: dump more reg info based on ras mod

[Jijie Shao]

on 2024/3/29 23:15, Jakub Kicinski wrote:
> On Fri, 29 Mar 2024 18:34:02 +0800 Jijie Shao wrote:
> > > These seem to be duplicating standard stats from rtnl_link_stats64,
> > > ethtool_pause_stats, ethtool_eth_mac_stats, etc.
> > >
> > > You can add device specific stats, but please don't duplicate
> > > stats for which we have standard APIs.
> > Yeah, but these are not duplicate stats for ethtool or debugfs.
> Can you say more? I mean there are APIs to expose MIB counters.
> Perhaps your driver doesn't implement those APIs today.
> But (1) it should, and (2) once it does it will be a duplicate.

Sorry for the wrong reply before, these stats are already included
in the ethtool -S stats.

According to the suggestions provided by the chip, the statistics help
analyze the cause of the MAC-related abnormal interrupt.

> > Generally, driver will reset to restore the normal state.
> > After the reset, many registers are cleared. Therefore,
> > it is difficult to analyze the reason of RAS.
> Perhaps I'm missing the significance of the reset when it comes
> to counters reported via standard APIs. Are rtnl_link_stats64
> going to behave differently across a reset than these debug entries?
1. These statistics are the same as rtnl_link_stats64. However, these are not updated in real time.
   They are updated only when users query them or driver updates them every 5 minutes.
   However, these are cleared after the reset, which makes debugging difficult.
2. Currently, only a few MIB statistics are required, not all.
3. Are you suggesting that we use rtnl_link_stats64 to provide MIB statistics?

> > We wang to add this information only when RAS is occurring, And
> > these information will help to analyze the reason of RAS.
> >
> > these information does not appear in any new API.
> >
> > Therefore, we hope that we can add this information to
> > reduce the difficulty of analyzing certain issues.