lkml.org 
[lkml]   [2020]   [May]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] perf/x86/intel/uncore: Fix oops when counting IMC uncore events on some TGL
From
Date


On 5/27/2020 11:17 AM, David Laight wrote:
> From: Liang, Kan
>> Sent: 27 May 2020 16:01
>> On 5/27/2020 10:51 AM, David Laight wrote:
>>> From: Liang, Kan
>>>> Sent: 27 May 2020 15:47
>>>> On 5/27/2020 8:59 AM, David Laight wrote:
>>>>> From: kan.liang@linux.intel.com
>>>>>> Sent: 27 May 2020 13:31
>>>>>>
>>>>>> From: Kan Liang <kan.liang@linux.intel.com>
>>>>>>
>>>>>> When counting IMC uncore events on some TGL machines, an oops will be
>>>>>> triggered.
>>>>>> [ 393.101262] BUG: unable to handle page fault for address:
>>>>>> ffffb45200e15858
>>>>>> [ 393.101269] #PF: supervisor read access in kernel mode
>>>>>> [ 393.101271] #PF: error_code(0x0000) - not-present page
>>>>>>
>>>>>> Current perf uncore driver still use the IMC MAP SIZE inherited from
>>>>>> SNB, which is 0x6000.
>>>>>> However, the offset of IMC uncore counters for some TGL machines is
>>>>>> larger than 0x6000, e.g. 0xd8a0.
>>>>>>
>>>>>> Enlarge the IMC MAP SIZE for TGL to 0xe000.
>>>>>
>>>>> Replacing one 'random' constant with a different one
>>>>> doesn't seem like a proper fix.
>>>>>
>>>>> Surely the actual bounds of the 'memory' area are properly
>>>>> defined somewhere.
>>>>> Or at least should come from a table.
>>>>>
>>>>> You also need to verify that the offsets are within the mapped area.
>>>>> An unexpected offset shouldn't try to access an invalid address.
>>>>
>>>> Thanks for the review.
>>>>
>>>> I agree that we should add a check before mapping the area to prevent
>>>> the issue happens again.
>>>>
>>>> I think the check should be a generic check for all platforms which try
>>>> to map an area, not just for TGL. I will submit a separate patch for the
>>>> check.
>>>
>>> You need a check that the actual access is withing the mapped area.
>>> So instead of getting an OOPS you get a error.
>>>
>>> This is after you've mapped it.
>>
>> Sure. Will add a WARN_ONCE() before the actual access.
>
> No that will still panic some systems.
> pr_warn() is all you need.
>

If we print a warning for each access, there will be too many warnings.
I think I will use pr_warn_once instead.

Thanks,
Kan

\
 
 \ /
  Last update: 2020-05-27 18:03    [W:0.045 / U:1.316 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site