Re: [PATCH v5 1/1] mm/memory-failure: disable unpoison once hw error happens

[David Hildenbrand <david@redhat.com>]

On 15.06.22 10:15, HORIGUCHI NAOYA(堀口 直也) wrote:
> On Wed, Jun 15, 2022 at 10:00:05AM +0800, zhenwei pi wrote:
>> Currently unpoison_memory(unsigned long pfn) is designed for soft
>> poison(hwpoison-inject) only. Since 17fae1294ad9d, the KPTE gets
>> cleared on a x86 platform once hardware memory corrupts.
>>
>> Unpoisoning a hardware corrupted page puts page back buddy only,
>> the kernel has a chance to access the page with *NOT PRESENT* KPTE.
>> This leads BUG during accessing on the corrupted KPTE.
>>
>> Suggested by David&Naoya, disable unpoison mechanism when a real HW error
>> happens to avoid BUG like this:
> ...
> 
>>
>> Fixes: 847ce401df392 ("HWPOISON: Add unpoisoning support")
>> Fixes: 17fae1294ad9d ("x86/{mce,mm}: Unmap the entire page if the whole page is affected and poisoned")
>> Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
>> Cc: David Hildenbrand <david@redhat.com>
>> Signed-off-by: zhenwei pi <pizhenwei@bytedance.com>
> 
> Cc to stable?
> I think that the current approach seems predictable to me than earlier versions,
> so I can agree with sending this to stable a little more confidently.
> 
>> ---
>>  Documentation/vm/hwpoison.rst |  3 ++-
>>  drivers/base/memory.c         |  2 +-
>>  include/linux/mm.h            |  1 +
>>  mm/hwpoison-inject.c          |  2 +-
>>  mm/madvise.c                  |  2 +-
>>  mm/memory-failure.c           | 12 ++++++++++++
>>  6 files changed, 18 insertions(+), 4 deletions(-)
>>
> 
> ...
> 
>> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
>> index b85661cbdc4a..385b5e99bfc1 100644
>> --- a/mm/memory-failure.c
>> +++ b/mm/memory-failure.c
>> @@ -69,6 +69,8 @@ int sysctl_memory_failure_recovery __read_mostly = 1;
>>  
>>  atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
>>  
>> +static bool hw_memory_failure;
> 
> Could you set the initial value explicitly?  Using a default value is good,
> but doing as the surrounding code do is better for consistency.  And this
> variable can be updated only once, so adding __read_mostly macro is also fine.

No strong opinion. __read_mostly makes sense, but I assume we don't
really care about performance that much when dealing with HW errors.

With or without changes around this initialization

Acked-by: David Hildenbrand <david@redhat.com>

-- 
Thanks,

David / dhildenb