Re: [PATCH] x86/mce/AMD: Fix partial SMCA bank init when CPU 0 != thread 0

From: Jack Miller
Date: Wed Jun 28 2017 - 14:53:10 EST


On Wed, Jun 28, 2017 at 1:00 PM, Ghannam, Yazen <Yazen.Ghannam@xxxxxxx> wrote:
>> -----Original Message-----
>> From: themoken@xxxxxxxxx [mailto:themoken@xxxxxxxxx] On Behalf Of
>> Jack Miller
>> Sent: Wednesday, June 28, 2017 1:44 PM
>> To: Borislav Petkov <bp@xxxxxxx>
>> Cc: Jack Miller <jack@xxxxxxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx;
>> tglx@xxxxxxxxxxxxx; Ghannam, Yazen <Yazen.Ghannam@xxxxxxx>;
>> x86@xxxxxxxxxx
>> Subject: Re: [PATCH] x86/mce/AMD: Fix partial SMCA bank init when CPU 0 !=
>> thread 0
>>
>> On Wed, Jun 28, 2017 at 4:22 AM, Borislav Petkov <bp@xxxxxxx> wrote:
>> > On Tue, Jun 27, 2017 at 07:06:30PM -0500, Jack Miller wrote:
>> >> After a call to firmware SwitchBSP(),
>> >
>> > What is that and who does that?
>>
>> SwitchBSP() is part of the UEFI MPServices Protocol which I believe is an
>> extension but it is supported by all of the firmwares I've tested on.
>>
>> In this case, I'm using a bootloader to SwitchBSP() so that hardware thread 0
>> (and thus core 0) can be offlined on AMD hardware (cpu0_hotplug
>> unsupported). This is currently working by passing 'nomce' to the kernel, but
>> obviously I'd prefer not to disable it.
>>
>
> Which core are you using as the BSP with SwitchBSP()?

Core 4, hardware thread 8 overall. I am testing on a Ryzen 7 machine.

>
>> >
>> >> Linux can be booted with a thread
>> >> that isn't the first in the system. That thread automatically becomes
>> >> CPU 0.
>> >
>> > Btw, you should be seeing other explosions too as a lot of code
>> > assumes CPU 0 is the BSP.
>>
>> Actually, with 'nomce' or this patch applied the system seems to chug along
>> merrily, no further errors in dmesg, no further BUGs. Linux still gets all of the
>> topology correct (i.e. CPU 0's core/thread/siblings are correctly identified) so
>> really, aside from userspace programs doing naive stuff with CPU affinity (like
>> expecting even,odd CPUs to be SMT pairs), I think the overall result here is
>> that most threads are interchangeable... except when probing certain
>> features like these MCA types.
>>
>
> Do you see 23 banks named in the new BSP's /sys/devices/system/machinecheck/
> folder? You should see non-core banks like l3_cache, umc, etc.

With my patch applied, I see entries like l3_cache under hardware
thread 0's directory (it's shifted to CPU 1, so machinecheck1).
Without my patch, only machinecheck0 has anything interesting in it
(insn_fetch, l2_cache etc.) because the init failed on CPU 1.

Jack