Re: [RFC PATCH v3 3/3] mm/compaction: enhance compaction finish condition

[Vlastimil Babka <vbabka@suse.cz>]

On 02/02/2015 02:23 PM, Joonsoo Kim wrote:
> 2015-02-02 19:20 GMT+09:00 Vlastimil Babka <vbabka@suse.cz>:
>> On 02/02/2015 08:15 AM, Joonsoo Kim wrote:
>>
>> So I've realized that this problaby won't always work as intended :/ Because we
>> still differ from what page allocator does.
>> Consider we compact for UNMOVABLE allocation. First we try RECLAIMABLE fallback.
>> Turns out we could fallback, but not steal, hence we skip it due to
>> only_stealable == true. So we try MOVABLE, and turns out we can steal, so we
>> finish compaction.
>> Then the allocation attempt follows, and it will fallback to RECLAIMABLE,
>> without extra stealing. The compaction decision for MOVABLE was moot.
>> Is it a big problem? Probably not, the compaction will still perform some extra
>> anti-fragmentation on average, but we should consider it.
> 
> Hello,
> 
> First of all, thanks for quick review. :)
> 
> Hmm... I don't get it. Is this case possible in current implementation?
> can_steal_fallback() decides whether steal is possible or not, based
> on freepage order
> and start_migratetype. If fallback freepage is on RECLAIMABLE and
> MOVABLE type and
> they are same order, can_steal could be true for both or false for
> neither. If order is
> different, compaction decision would be recognized by
> __rmqueue_fallback() since it
> try to find freepage from high order to low order.

Ah, right, I got confused into thinking that the result of can_steal depends on
how many freepages it found within the pageblock to steal. Sorry about the noise.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>