Re: [PATCH v3 3/5] x86/microcode: Avoid any chance of MCE's during microcode update

[Ashok Raj]

On Wed, Aug 17, 2022 at 04:19:40PM +0200, Borislav Petkov wrote:
> On Wed, Aug 17, 2022 at 12:30:49PM +0000, Ashok Raj wrote:
> > You will find out when system returns after reboot and hopefully wasn't
> > promoted to a cold-boot which will loose MCE banks.
> 
> Not good enough!

I probably misread your question.. are you suggesting we add some WARN when
we initiate late_load? I thought you were asking if the HW must signal
something and OS should log when an MCE happens if MCIP=1


> 
> This should issue a warning in dmesg that a potential MCE while update
> is running would cause a lockup. That is if we don't disable MCE around
> it.
> 
> If we decide to disable MCE, it should say shutdown.

Ok, that clarifies it.. "IF we choose to set MCIP=1, we should tell users
that hell can break loose, get under the table" :-)

> 
> > Meaning deal with the effect of a really rare MCE. Rather than trying to
> > avoid it. Taking the MCE is more important than finishing the update,
> > and loosing what the error signaled was trying to convey.
> 
> Right now I'm inclined to not do anything and warn of a potential rare
> situation.

Encouraging.. So I'll drop that patch from the list next time around.

> 
> > > > Shutdown, shutdown.. There is only 1 MCE no matter how many CPUs you have.
> > > 
> > > Because all CPUs are executing the loop? Or how do you decide this?
> > 
> > Fatal errors signaled with PCC=1 in the MCAx.STATUS is *ALWAYS*
> 
> What does that have to do with
> 
> "There is only 1 MCE no matter how many CPUs you have."
> 
> ?
> 
> That's bullsh*t. Especially if the machine can do LMCE.

Well, not outlandish :)

LMCE is only for recoverable errors. When we have a fatal error, sometimes
the signalling and consumption of poison are going in different directions.
In order to minimize exposure of bad data from being consumed,
*ALL* Intel processors have always broadcast fatal errors. This is the
history behind why we broadcast.

BTW: This is all legacy behavior. Nothing should come as surprise.

LMCE is best effort. This is the current state.

Cheers,
Ashok