Re: [PATCH 3/3] mce: acpi/apei: trace: Enable ghes memory error trace event

[Mauro Carvalho Chehab <m.chehab@samsung.com>]

Em Tue, 13 Aug 2013 22:25:58 +0530
"Naveen N. Rao" <naveen.n.rao@linux.vnet.ibm.com> escreveu:

(sorry for a late answer, I had to do a small travel yesterday)

> On 08/13/2013 05:51 PM, Mauro Carvalho Chehab wrote:
> > Em Tue, 13 Aug 2013 17:06:14 +0530
> > "Naveen N. Rao" <naveen.n.rao@linux.vnet.ibm.com> escreveu:
> >
> >> On 08/12/2013 11:26 PM, Borislav Petkov wrote:
> >>> On Mon, Aug 12, 2013 at 02:25:57PM -0300, Mauro Carvalho Chehab wrote:
> >>>> Userspace still needs the EDAC sysfs, in order to identify how the
> >>>> memory is organized, and do the proper memory labels association.
> >>>>
> >>>> What edac_ghes does is to fill those sysfs nodes, and to call the
> >>>> existing tracing to report errors.
> >>
> >> I suppose you're referring to the entries under /sys/devices/system/edac/mc?
> >
> > Yes.
> >
> >>
> >> I'm not sure I understand how this helps. ghes_edac seems to just be
> >> populating this based on dmi, which if I'm not mistaken, can be obtained
> >> in userspace (mcelog as an example).
> >>
> >> Also, on my system, all DIMMs are being reported under mc0. I doubt if
> >> the labels there are accurate.
> >
> > Yes, this is the current status of ghes_edac, where BIOS doesn't provide any
> > reliable way to associate a given APEI report to a physical DIMM slot label.
> >
> > The plan is to add more logic there as BIOSes start to provide some reliable
> > way to do such association. I discussed this subject with a few vendors
> > while I was working at Red Hat.
> 
> Hmm... is there anything specific in the APEI report that could help? 

I didn't see anything at APEI spec that would allow to describe how the
memory is organized. So, it is hard for the ghes_edac driver to discover
how many memory controllers, channels and slots are available. This data
is needed, in order to allow userspace to pass the labels for each DIMM,
or for the Kernel to auto-discover.

> More importantly, is there a need to do this in-kernel rather than in 
> user-space?

Yes, due to 2 aspects:

On a critical error, the machine will die. The EDAC core will print the
error at dmesg, but no other record to be latter parsed will be available;

With hot pluggable memories, dynamic channel rerouting, memory poisoning
and other funny things, it could not be possible to point to a DIMM, 
if the parsing is done on a latter time.

Regards,
Mauro