Re: [PATCH 4/6] x86-mce: Add spinlocks to prevent duplicated MCP and CMCI reports.

[Borislav Petkov <bp@alien8.de>]

On Fri, Jul 18, 2014 at 02:23:04PM -0700, Tony Luck wrote:
> On Thu, Jul 17, 2014 at 3:50 AM, Borislav Petkov <bp@alien8.de> wrote:
> > Well, maybe it is about time we tracked shared banks.
> 
> For cpus that support CMCI and the MCi_CTL2 registers we do track
> sharing. Only one cpu gets to be the "owner" of a bank that supports
> CMCI (the first one to find it and set bit 30 in the CTL2 register).
> 
> The test_bit()  at the top of the loop in machine_check_poll() makes
> sure only the owner of a bank actually looks at it.
> 
>         for (i = 0; i < mca_cfg.banks; i++) {
>                 if (!mce_banks[i].ctl || !test_bit(i, *b))
>                         continue;
> 
> If we don't have CMCI, then we don't have the CTL2 registers, and
> so have no way to find out which banks are shared.

Ah, so Havard's corrected explanation was this:

"I don't think we got the description right here. I think the real issue
here was machine check polls happening on multiple CPUs with shared
banks, all reporting the same MCEs. This is very reproducible when
booting with mce=no_cmci, since all CPUs will handle all banks, and
there's AFAICT no good way to identify shared banks without enabling
CMCI."

Remind me, why would one boot with mce=no_cmci at all, on a CMCI
machine?

> I'd be surprised if it was a problem in practice. If we have  CMCI,
> then we limit the banks that we look at (and if we see a high rate
> of interrupts, then we turn off interrupts an poll).
>
> If we don't have CMCI, then we are polling at a pretty low rate
> (current code adjusts the rate higher if we are finding errors to
> log, but we don't let that rate rise forever ... cap is ~ 1HZ).

Right, it would be interesting to see how a huuge machine (4 sockets
with lotsa memory) behaves under a CMCI storm...

-- 
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--