On 27.01.22 22:16, Yang Shi wrote:
> On Wed, Jan 26, 2022 at 10:54 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>>
>>>>> Just page lock or elevated page refcount could serialize against THP
>>>>> split AFAIK.
>>>>>
>>>>>>
>>>>>> But yeah, using the mapcount of a page that is not even mapped
>>>>>> (migration entry) is clearly wrong.
>>>>>>
>>>>>> To summarize: reading the mapcount on an unlocked page will easily
>>>>>> return a wrong result and the result should not be relied upon. reading
>>>>>> the mapcount of a migration entry is dangerous and certainly wrong.
>>>>>
>>>>> Depends on your usecase. Some just want to get a snapshot, just like
>>>>> smaps, they don't care.
>>>>
>>>> Right, but as discussed, even the snapshot might be slightly wrong. That
>>>> might be just fine for smaps (and I would have enjoyed a comment in the
>>>> code stating that :) ).
>>>
>>> I think that is documented already, see Documentation/filesystems/proc.rst:
>>>
>>> Note: reading /proc/PID/maps or /proc/PID/smaps is inherently racy (consistent
>>> output can be achieved only in the single read call).
>>
>> Right, but I think there is a difference between
>>
>> * Atomic values that change immediately afterwards ("this value used to
>> be true at one point in time")
>> * Values that are unstable because we cannot read them atomically ("this
>> value never used to be true")
>>
>> I'd assume with the documented race we actually talk about the first
>> point, but I might be just wrong.
>>
>>>
>>> Of course, if the extra note is preferred in the code, I could try to
>>> add some in a separate patch.
>>
>> When staring at the (original) code I would have hoped to find something
>> like:
>>
>> /*
>> * We use page_mapcount() to get a snapshot of the mapcount. Without
>> * holding the page lock this snapshot can be slightly wrong as we
>> * cannot always read the mapcount atomically. As long we hold the PT
>> * lock, the page cannot get unmapped and it's at safe to call
>> * page_mapcount().
>> */
>>
>> With the addition of
>>
>> "... For unmapped pages (e.g., migration entries) we cannot guarantee
>> that, so treat the mapcount as being 1."
>
> It seems a little bit confusing to me, it is not safe to call with PTL
> held either, right? I'd like to rephrase the note to:
The implication that could have been spelled out is that only a mapped
page can get unmapped. (I know, there are some weird migration entries
nowadays ...)
/*
* We use page_mapcount() to get a snapshot of the mapcount. Without
* holding the page lock this snapshot can be slightly wrong as we
* cannot always read the mapcount atomically. As long we hold the PT
* lock, a mapped page cannot get unmapped and it's at safe to call
* page_mapcount(). Especially for migration entries, it's not safe to
* call page_mapcount(), so we treat the mapcount as being 1.
*/
--
Thanks,
David / dhildenb
