On 17/08/2022 18:02, Borislav Petkov wrote:

On Wed, Aug 17, 2022 at 05:28:34PM -0300, Guilherme G. Piccoli wrote:

quoted

My understanding is the same as yours, i.e., this is not possible to
collect from vmcore, it requires register reading. But again: if you
kdump your machine today, you won't collect this information, patch
changed nothing in that regard.

Why won't you be able to collect it? You can certainly access dmesg in
the vmcore and see those errors logged there.

Sorry for the confusion, let me try to be a bit more clear:

(1) if we kdump but we *didn't run* s10_edac_dberr_handler() before
kdump, the information is lost, since s10_edac_dberr_handler() performs
register readings. That information is not contained inside the vmcore.

(2) If for some reason the function s10_edac_dberr_handler() *was
executed prior to kdump*, of course the registers information would be
on dmesg, easy to collect in the vmcore.

Makes sense?

quoted

The one thing it changes is that you'd skip the altera register dump if
kdump is set AND you managed to also set "crash_kexec_post_notifiers".

What your patch changes is, it prevents s10_edac_dberr_handler() from
logging potentially important fatal hw errors when kdump is loaded.

Agreed. If kdump is loaded, we cannot log that information (modulo that
we do not collect it today by default on kdump as well).
The other part of story (the reason of the patch) is that we plan to
start running this panic notifier a bit earlier, being able to collect
such edac logs with pstore, for example.

If Dinh is fine with that, I'll take the patch. But it looks like a bad
idea to me.

I think we should seek what the majority of the folks consider the best,
in order to converge to some well-accepted solution. I'm completely OK
in dropping this one and rework with some other idea, or we can leave
this panic notifier as is, continue running that a bit later.

Tony / Petr (when back), suggestions are welcome =)
Cheers,


Guilherme