Re: [PATCH v1] mm,hwpoison: set PG_hwpoison for busy hugetlb pages

[Miaohe Lin <linmiaohe@huawei.com>]

On 2022/5/12 12:32, HORIGUCHI NAOYA(堀口 直也) wrote:
> On Wed, May 11, 2022 at 11:35:55AM -0700, Mike Kravetz wrote:
>> On 5/11/22 08:19, Naoya Horiguchi wrote:
>>> From: Naoya Horiguchi <naoya.horiguchi@nec.com>
>>>
>>> If memory_failure() fails to grab page refcount on a hugetlb page
>>> because it's busy, it returns without setting PG_hwpoison on it.
>>> This not only loses a chance of error containment, but breaks the rule
>>> that action_result() should be called only when memory_failure() do
>>> any of handling work (even if that's just setting PG_hwpoison).
>>> This inconsistency could harm code maintainability.
>>>
>>> So set PG_hwpoison and call hugetlb_set_page_hwpoison() for such a case.
>>>
>>> Fixes: 405ce051236c ("mm/hwpoison: fix race between hugetlb free/demotion and memory_failure_hugetlb()")
>>> Signed-off-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
>>> ---
>>>  include/linux/mm.h  | 1 +
>>>  mm/memory-failure.c | 8 ++++----
>>>  2 files changed, 5 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/include/linux/mm.h b/include/linux/mm.h
>>> index d446e834a3e5..04de0c3e4f9f 100644
>>> --- a/include/linux/mm.h
>>> +++ b/include/linux/mm.h
>>> @@ -3187,6 +3187,7 @@ enum mf_flags {
>>>  	MF_MUST_KILL = 1 << 2,
>>>  	MF_SOFT_OFFLINE = 1 << 3,
>>>  	MF_UNPOISON = 1 << 4,
>>> +	MF_NO_RETRY = 1 << 5,
>>>  };
>>>  extern int memory_failure(unsigned long pfn, int flags);
>>>  extern void memory_failure_queue(unsigned long pfn, int flags);
>>> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
>>> index 6a28d020a4da..e3269b991016 100644
>>> --- a/mm/memory-failure.c
>>> +++ b/mm/memory-failure.c
>>> @@ -1526,7 +1526,8 @@ int __get_huge_page_for_hwpoison(unsigned long pfn, int flags)
>>>  			count_increased = true;
>>>  	} else {
>>>  		ret = -EBUSY;
>>> -		goto out;
>>> +		if (!(flags & MF_NO_RETRY))
>>> +			goto out;
>>>  	}
>>
>> Hi Naoya,
>>
>> We are in the else block because !HPageFreed() and !HPageMigratable().
>> IIUC, this likely means the page is isolated.  One common reason for isolation
>> is migration.  So, the page could be isolated and on a list for migration.
> 
> Yes, and I also detected this issue by testing race between hugepage allocation
> and memory_failure(). 
> 
>>
>> I took a quick look at the hugetlb migration code and did not see any checks
>> for PageHWPoison after a hugetlb page is isolated.  I could have missed
>> something?  If there are no checks, we will read the PageHWPoison page
>> in kernel mode while copying to the migration target.
> 
> Yes, that could happen.  This patch does not affect ongoing hugepage migration.
> But after the migration source hugepage is freed, the PG_hwpoison should work
> to prevent reusing.
> 
>>
>> Is this an issue?  Is is something we need to be concerned with?  Memory
>> errors can happen at any time, and gracefully handling them is best effort.
> 
> Right, so doing nothing for this case could be OK if doing something causes
> some issues or makes code too complicated.  The motivation of this patch is
> that now I think memory_failure() should do something (at least setting
> PG_hwpoison) unless the page is already hwpoisoned or rejected by
> hwpoison_filter(), because of the effect after free as mentioned above.
> 
> This is also expected in other case too. For example, slab is a unhandlable
> type of page, but we do set PG_hwpoison.  This flag should not affect any of
> ongoing slab-related process, but that's OK because it becomes effective
> after the slab page is freed.
> 
> So this patch is intended to align to the behavior.  Allowing hugepage
> migration to do something good using PG_hwpoison seems to me an unsolved
> separate issue.

I tend to agree with Naoya. And could we try to do it better? IMHO, we could do a
get_page_unless_zero here to ensure that hugetlb page migration will fail due to
this extra page reference and thus preventing the page content from being accessed.
Does this work? Or am I miss something?

Thanks!

> 
> Thanks,
> Naoya Horiguchi
>