Re: [PATCH v2 1/2] x86/mce/AMD: Redo use of SMCA MCA_DE{STAT,ADDR} registers

[Borislav Petkov]

On Wed, Apr 05, 2017 at 07:29:48PM +0000, Ghannam, Yazen wrote:
> If it's set, then I expect a Deferred error in MCA_STATUS since any Correctable
> Errors will be overwritten. Multiple bank types can generate Deferred errors
> so there may also be cases where for some types a valid Uncorrectable error
> happens and overwrites the Deferred error before we can handle it. In which
> case we lose the Deferred error if we don't check MCA_DESTAT.

So if we have an UE, wouldn't that raise an #MC? I guess in such cases
we should concentrate only on the deferred errors and let the #MC
handler deal with them. As we do now.

> If it's not set, then it's possible to have a valid Correctable error in MCA_STATUS
> while the valid Deferred error is in MCA_DESTAT.

What's logging the CE? We probably should log it too before something
overwrites it.

Anyway, ok, I think I know what needs to happen now:

amd_deferred_error_interrupt:

	if (__log_error_deferred(bank))
		return;

This one read MC?_STATUS and does the logging for when the deferred
error is in the normal MSRs. It returns true if it succeeded. It reads
and hands down both MC?_STATUS and MC?_ADDR to __log_error() so that it
doesn't have to read MC?_STATUS twice.

If __log_error_deferred() has read a different type of error, we still
log it? I'm not sure about this. I guess we can ignore that case for
now.

Then:

	if (mca_flags.smca)
		__log_error_deferred_smca(bank));

which handles the SMCA case. It too reads MSR_AMD64_SMCA_MCx_DESTAT
and MSR_AMD64_SMCA_MCx_DEADDR and hands them down to __log_error() for
logging.

For the __log_error() call in amd_threshold_interrupt(), you define a
log_error() wrapper which reads the default MSRs and hands them down to
__log_error().

So __log_error() always gets STATUS and ADDR MSR values and it doesn't
need to read them from the MSRs but only log them.

-- 
Regards/Gruss,
    Boris.

Good mailing practices for 400: avoid top-posting and trim the reply.