Hi Boris,
On 24/08/18 13:01, Borislav Petkov wrote:
> On Fri, Aug 24, 2018 at 10:48:24AM +0100, James Morse wrote:
>> so edac_raw_mc_handle_error() has no clue where the error happened. (I
>> haven't
>> read what it does with this information yet).
>
> See edac_inc_ce_error(), for example - it uses the layers which are not
> negative (-1) to increment the error counts of the respective layer. It
> all depends on what granularity of the hardware part you're reporting
> the error for: is it a DIMM rank, a whole DIMM or for a channel which
> can span multiple DIMM ranks. And so on...
>
> Look at some of the drivers and how they're doing that layering. It all
> depends on whether you can get the precise info from the hw.
Hmmm, in this example we need the information from firmware, as that is where
ghes-edac gets its information from.
We already count the module/device/dimms in the smbios table, memory is
described as 'EDAC_MC_LAYER_ALL_MEM' with num_dimms. I think all we're missing
is which dimm in ghes_edac_report_mem_error(). We have the handle, we just need
a number between 1 and num_dimms.
If it turns out firmware doesn't populate the handles in its cper records, then
we can keep e->enable_per_layer_report false when calling
edac_raw_mc_handle_error().
(I suggest we ignore 'card', and just do this for the device/dimms).
>> Naively I thought we could generate some index during
>> ghes_edac_count_dimms(),
>> and use this as e->${whichever}_layer. I hoped there would be something we
>> could
>> already use as the index, but I can't spot it, so this will be more than the
>> one-liner I was hoping for!
>
> If you can get that info from the hardware and injecting an error into
> a DIMM gives you the correct DIMM number so that we can increment the
> proper counter, then you're golden. I don't think that works reliably on
> x86, though, therefore the lumping together.
... 'correct DIMM number' ...
Does x86 have another source of memory-topology information it needs to
correlate smbios with?
For arm there is nothing else describing the memory-topology, so as long as we
can correlate the smbios table and ghes:cper records through the handles, we can
get this working for all systems.
Thanks,
James
