Virtualization - Nov 2021 - [RFC PATCH 0/3] Use pageblock_order for cma and alloc_contig_range alignment.

On 11/29/21 23:08, Zi Yan wrote:
> On 23 Nov 2021, at 12:32, Vlastimil Babka wrote:
> 
>> On 11/23/21 17:35, Zi Yan wrote:
>>> On 19 Nov 2021, at 10:15, Zi Yan wrote:
>>>>>> From what my understanding, cma required alignment of
>>>>>> max(MAX_ORDER - 1, pageblock_order), because when
MIGRATE_CMA was introduced,
>>>>>> __free_one_page() does not prevent merging two
different pageblocks, when
>>>>>> MAX_ORDER - 1 > pageblock_order. But current
__free_one_page() implementation
>>>>>> does prevent that.
>>>>>
>>>>> But it does prevent that only for isolated pageblock, not
CMA, and yout
>>>>> patchset doesn't seem to expand that to CMA? Or am I
missing something.
>>>>
>>>> Yeah, you are right. Originally, I thought preventing merging
isolated pageblock
>>>> with other types of pageblocks is sufficient, since MIGRATE_CMA
is always
>>>> converted from MIGRATE_ISOLATE. But that is not true. I will
rework the code.
>>>> Thanks for pointing this out.
>>>>
>>>
>>> I find that two pageblocks with different migratetypes, like
MIGRATE_RECLAIMABLE
>>> and MIGRATE_MOVABLE can be merged into a single free page after I
checked
>>> __free_one_page() in detail and printed pageblock information
during buddy page
>>> merging.
>>
>> Yes, that can happen.
>>
>> I am not sure what consequence it will cause. Do you have any idea?
>>
>> For MIGRATE_RECLAIMABLE or MIGRATE_MOVABLE or even MIGRATE_UNMOVABLE
it's
>> absolutely fine. As long as these pageblocks are fully free (and they
are if
>> it's a single free page spanning 2 pageblocks), they can be of any
of these
>> type, as they can be reused as needed without causing fragmentation.
>>
>> But in case of MIGRATE_CMA and MIGRATE_ISOLATE, uncontrolled merging
would
>> break the specifics of those types. That's why the code is careful
for
>> MIGRATE_ISOLATE, and MIGRATE_CMA was until now done in MAX_ORDER
granularity.
> 
> Thanks for the explanation. Basically migratetypes that can fall back to
each
> other can be merged into a single free page, right?
Yes.
> How about MIGRATE_HIGHATOMIC? It should not be merged with other
migratetypes
> from my understanding.
Hmm it shouldn't minimally because it has an accounting that would become
broken. So it should prevent merging or make sure the reservations are with
MAX_ORDER granularity, but seems that neither is true? CCing Mel.
> --
> Best Regards,
> Yan, Zi
>

On Tue, Nov 30, 2021 at 10:11:36AM +0100, Vlastimil Babka
wrote:
> >>> I find that two pageblocks with different migratetypes, like
MIGRATE_RECLAIMABLE
> >>> and MIGRATE_MOVABLE can be merged into a single free page
after I checked
> >>> __free_one_page() in detail and printed pageblock information
during buddy page
> >>> merging.
> >>
> >> Yes, that can happen.
> >>
> >> I am not sure what consequence it will cause. Do you have any
idea?
> >>
> >> For MIGRATE_RECLAIMABLE or MIGRATE_MOVABLE or even
MIGRATE_UNMOVABLE it's
> >> absolutely fine. As long as these pageblocks are fully free (and
they are if
> >> it's a single free page spanning 2 pageblocks), they can be of
any of these
> >> type, as they can be reused as needed without causing
fragmentation.
> >>
> >> But in case of MIGRATE_CMA and MIGRATE_ISOLATE, uncontrolled
merging would
> >> break the specifics of those types. That's why the code is
careful for
> >> MIGRATE_ISOLATE, and MIGRATE_CMA was until now done in MAX_ORDER
granularity.
> > 
> > Thanks for the explanation. Basically migratetypes that can fall back
to each
> > other can be merged into a single free page, right?
> 
> Yes.
> 
> > How about MIGRATE_HIGHATOMIC? It should not be merged with other
migratetypes
> > from my understanding.
> 
> Hmm it shouldn't minimally because it has an accounting that would
become
> broken. So it should prevent merging or make sure the reservations are with
> MAX_ORDER granularity, but seems that neither is true? CCing Mel.
> 
MIGRATE_HIGHATOMIC pageblocks can have pages allocated of different
types, particularly UNMOVABLE and potentially RECLAIMABLE. The
reserving or releasing MIGRATE_HIGHATOMIC pageblocks should be done with
reserve_highatomic_pageblock and unreserve_highatomic_pageblock to get
the accounting right.

However, there does not appear to be any special protection against a
page in a highatomic pageblock getting merged with a buddy of another
pageblock type. The pageblock would still have the right setting but on
allocation, the pages could split to the wrong free list and be lost
until the pages belonging to MIGRATE_HIGHATOMIC were freed again.

Not sure how much of a problem that is in practice, it's been a while
since I've heard of high-order atomic allocation failures.

-- 
Mel Gorman
SUSE Labs