Re: [PATCH v2 5/8] mm/isolation: change pageblock isolation logic to fix freepage counting bugs

From: Joonsoo Kim
Date: Thu Aug 07 2014 - 09:35:34 EST


2014-08-07 22:04 GMT+09:00 Vlastimil Babka <vbabka@xxxxxxx>:
> On 08/07/2014 02:26 PM, Joonsoo Kim wrote:
>>
>> 2014-08-07 17:53 GMT+09:00 Vlastimil Babka <vbabka@xxxxxxx>:
>>>
>>> Ah, right. I thought that everything going to pcp lists would be through
>>>
>>> freeing which would already observe the isolate migratetype and skip
>>> pcplist. I forgot about the direct filling of pcplists from buddy list.
>>> You're right that we don't want extra hooks there.
>>>
>>> Still, couldn't this be solved in a simpler way via another pcplist drain
>>> after the pages are moved from normal to isolate buddy list? Should be
>>> even
>>> faster because instead of disable - drain - enable (5 all-cpu kicks,
>>> since
>>> each pageset_update does 2 kicks) you have drain - drain (2 kicks). While
>>> it's true that pageset_update is single-zone operation, I guess we would
>>> easily benefit from having a single-zone drain operation as well.
>>
>>
>> I hope so, but, it's not possible. Consider following situation.
>>
>> Page A: on pcplist of CPU2 and it is on isolate pageblock.
>>
>> CPU 1 CPU 2
>> drain pcplist
>> wait IPI finished move A to normal buddy list
>> finish IPI
>> A is moved to pcplist by allocation request
>>
>> move doesn't catch A,
>> because it is on pcplist.
>>
>> drain pcplist
>> wait IPI finished move A to normal buddy list
>> finish IPI
>> A is moved to pcplist by allocation request
>>
>> repeat!!
>>
>> It could happen infinitely, though, low possibility.
>
>
> Hm I see. Not a correctness issue, but still a failure to isolate. Probably
> not impossible with enough CPU's and considering the fact that after
> pcplists are drained, the next allocation request will try to refill them.
> And during the drain, the pages are added to the beginning of the free_list
> AFAICS, so they will be in the first refill batch.

I think that it is correctness issue. When page A is moved to normal buddy
list, merge could happen and freepage counting would be incorrect.

> OK, another attempt for alternative solution proposal :) It's not that I
> would think disabling pcp would be so bad, just want to be sure there is no
> better alternative.

Yeah, welcome any comment. :)

> What if the drain operation had a flag telling it to recheck pageblock
> migratetype and don't assume it's on the correct pcplist. Then the problem
> would go away I think? Would it be possible to do without affecting the
> normal drain-pcplist-when-full path? So that the cost is only applied to
> isolation, but lower cost than pcplist disabling.
>
> Actually I look that free_pcppages_bulk() doesn't consider migratetype of
> the pcplist, but uses get_freepage_migratetype(page). So the pcplist drain
> could first scan the pcplists and rewrite the freepage_migratetype according
> to pageblock_migratetype. Then the free_pcppages_bulk() operation would be
> unchanged for normal operation.
>
> Or is this too clumsy? We could be also smart and have an alternative to
> free_pcppages_bulk() which would omit the round-robin stuff (not needed for
> this kind of drain), and have a pfn range to limit its operation to pages
> that we are isolating.
> Hm I guess with this approach some pages might still escape us if they were
> moving between normal buddy list and pcplist through rmqueue_bulk() and
> free_pcppages_bulk() (and not through our drain) at the wrong moments, but I
> guess that would require a really specific workload (alternating between
> burst of allocations and deallocations) and consistently unlucky timing.
>

Yes, it has similar problem as I mentioned above.

Page A: on pcplist of CPU2 and it is on isolate pageblock.

CPU 1 CPU 2
A is on normal buddy list
drain pcplist
wait IPI finished
finish IPI
A is moved to pcplist by allocation request
move doesn't catch A,
because it is on pcplist.
move A to normal buddy list by free request

drain pcplist
wait IPI finished
finish IPI
A is moved to pcplist by allocation request
move doesn't catch A,
because it is on pcplist.
move A to normal buddy list by free request

repeat!!

Although it is really corner case, I would like to choose error-free
approach something like pcplist disable. :)

Thanks.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/