lkml.org 
[lkml]   [2022]   [Aug]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH v5] x86/mce: retrieve poison range from hardware
Date
On 8/1/2022 2:20 PM, Dan Williams wrote:
> Jane Chu wrote:
>> On 8/1/2022 9:44 AM, Dan Williams wrote:
>>> Jane Chu wrote:
>>>> With Commit 7917f9cdb503 ("acpi/nfit: rely on mce->misc to determine
>>>> poison granularity") that changed nfit_handle_mce() callback to report
>>>> badrange according to 1ULL << MCI_MISC_ADDR_LSB(mce->misc), it's been
>>>> discovered that the mce->misc LSB field is 0x1000 bytes, hence injecting
>>>> 2 back-to-back poisons and the driver ends up logging 8 badblocks,
>>>> because 0x1000 bytes is 8 512-byte.
>>>>
>>>> Dan Williams noticed that apei_mce_report_mem_error() hardcode
>>>> the LSB field to PAGE_SHIFT instead of consulting the input
>>>> struct cper_sec_mem_err record. So change to rely on hardware whenever
>>>> support is available.
>>>>
>>>> Link: https://lore.kernel.org/r/7ed50fd8-521e-cade-77b1-738b8bfb8502@oracle.com
>>>>
>>>> Reviewed-by: Dan Williams <dan.j.williams@intel.com>
>>>> Signed-off-by: Jane Chu <jane.chu@oracle.com>
>>>> ---
>>>> arch/x86/kernel/cpu/mce/apei.c | 14 +++++++++++++-
>>>> 1 file changed, 13 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/arch/x86/kernel/cpu/mce/apei.c b/arch/x86/kernel/cpu/mce/apei.c
>>>> index 717192915f28..2c7ea0ba9dd7 100644
>>>> --- a/arch/x86/kernel/cpu/mce/apei.c
>>>> +++ b/arch/x86/kernel/cpu/mce/apei.c
>>>> @@ -29,15 +29,27 @@
>>>> void apei_mce_report_mem_error(int severity, struct cper_sec_mem_err *mem_err)
>>>> {
>>>> struct mce m;
>>>> + int lsb = PAGE_SHIFT;
>>>>
>>>> if (!(mem_err->validation_bits & CPER_MEM_VALID_PA))
>>>> return;
>>>>
>>>> + /*
>>>> + * Even if the ->validation_bits are set for address mask,
>>>> + * to be extra safe, check and reject an error radius '0',
>>>> + * and fallback to the default page size.
>>>> + */
>>>> + if (mem_err->validation_bits & CPER_MEM_VALID_PA_MASK) {
>>>> + lsb = __ffs64(mem_err->physical_addr_mask);
>>>> + if (lsb == 1)
>>>
>>> This was the reason I recommended hweight64 and min_not_zero() as
>>> hweight64 does not have the undefined behavior. However, an even better
>>> option is to just do:
>>>
>>> find_first_bit(&mem_err->physical_addr_mask, PAGE_SHIFT)
>>>
>>> ...as that trims the result to the PAGE_SHIFT max and handles the zero
>>> case.
>>
>> Thanks Dan! However it looks like find_first_bit() could call into
>> __ffs(x) which has the same limitation as __ffs64(x), as Tony pointed out.
>
> Not quite, no. __ffs() behavior is *undefined* if the input is zero.
> find_first_bit() is *defined* and returns @size is the input is zero.
> Which is the behavior this wants to default to PAGE_SHIFT in the absence
> of any smaller granularity information.
>

You're right, because of this line -
return val ? __ffs(val) : size;

Thanks!
-jane


\
 
 \ /
  Last update: 2022-08-02 00:02    [W:2.961 / U:0.132 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site