Re: [PATCH] acpi/nfit: badrange report spill over to clean range

[Jane Chu <jane.chu@oracle.com>]

On 7/15/2022 12:17 PM, Dan Williams wrote:
> [ add Tony ]
> 
> Jane Chu wrote:
>> On 7/14/2022 6:19 PM, Dan Williams wrote:
>>> Jane Chu wrote:
>>>> I meant to say there would be 8 calls to the nfit_handle_mce() callback,
>>>> one call for each poison with accurate address.
>>>>
>>>> Also, short ARS would find 2 poisons.
>>>>
>>>> I attached the console output, my annotation is prefixed with "<==".
>>>
>>> [29078.634817] {4}[Hardware Error]:   physical_address: 0x00000040a0602600		<== 2nd poison @ 0x600
>>> [29078.642200] {4}[Hardware Error]:   physical_address_mask: 0xffffffffffffff00
>>>
>>> Why is nfit_handle_mce() seeing a 4K address mask when the CPER record
>>> is seeing a 256-byte address mask?
>>
>> Good question!  One would think both GHES reporting and
>> nfit_handle_mce() are consuming the same mce record...
>> Who might know?
> 
> Did some grepping...
> 
> Have a look at: apei_mce_report_mem_error()
> 
> "The call is coming from inside the house!"
> 
> Luckily we do not need to contact a BIOS engineer to get this fixed.

Great, thank you!
Just put together a quick fix for review after I tested it.

> 
>>> Sigh, is this "firmware-first" causing the kernel to get bad information
>>> via the native mechanisms >
>>> I would expect that if this test was truly worried about minimizing BIOS
>>> latency it would disable firmware-first error reporting. I wonder if
>>> that fixes the observed problem?
>>
>> Could you elaborate on firmware-first error please?  What are the
>> possible consequences disabling it? and how to disable it?
> 
> With my Linux kernel developer hat on, firmware-first error handling is
> really only useful for supporting legacy operating systems that do not
> have native machine check handling, or for platforms that have bugs that
> would otherwise cause OS native error handling to fail. Otherwise, for
> modern Linux, firmware-first error handling is pure overhead and a
> source of bugs.
> 
> In this case the bug is in the Linux code that translates the ACPI event
> back into an MCE record.

Thanks!

-jane