On 03/08/2016 11:10 AM, Michal Hocko wrote:
> On Tue 08-03-16 10:52:15, Vlastimil Babka wrote:
>> On 03/08/2016 10:46 AM, Michal Hocko wrote:
> [...]
>>>>> @@ -3294,6 +3289,18 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned
>>>>> int order,
>>>>> did_some_progress > 0, no_progress_loops))
>>>>> goto retry;
>>>>>
>>>>> + /*
>>>>> + * !costly allocations are really important and we have to make sure
>>>>> + * the compaction wasn't deferred or didn't bail out early due to locks
>>>>> + * contention before we go OOM.
>>>>> + */
>>>>> + if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
>>>>> + if (compact_result <= COMPACT_CONTINUE)
>>>>
>>>> Same here.
>>>> I was going to say that this didn't have effect on Sergey's test, but
>>>> turns out it did :)
>>>
>>> This should work as expected because compact_result is unsigned long
>>> and so this is the unsigned arithmetic. I can make
>>> #define COMPACT_NONE -1UL
>>>
>>> to make the intention more obvious if you prefer, though.
>>
>> Well, what wasn't obvious to me is actually that here (unlike in the
>> test above) it was actually intended that COMPACT_NONE doesn't result in
>> a retry. But it makes sense, otherwise we would retry endlessly if
>> reclaim couldn't form a higher-order page, right.
>
> Yeah, that was the whole point. An alternative would be moving the test
> into should_compact_retry(order, compact_result, contended_compaction)
> which would be CONFIG_COMPACTION specific so we can get rid of the
> COMPACT_NONE altogether. Something like the following. We would lose the
> always initialized compact_result but this would matter only for
> order==0 and we check for that. Even gcc doesn't complain.
Yeah I like this version better, you can add my Acked-By.
Thanks.
> A more important question is whether the criteria I have chosen are
> reasonable and reasonably independent on the particular implementation
> of the compaction. I still cannot convince myself about the convergence
> here. Is it possible that the compaction would keep returning
> compact_result <= COMPACT_CONTINUE while not making any progress at all?
Theoretically, if reclaim/compaction suitability decisions and
allocation attempts didn't match the watermark checks, including the
alloc_flags and classzone_idx parameters. Possible scenarios:
- reclaim thinks compaction has enough to proceed, but compaction thinks
otherwise and returns COMPACT_SKIPPED
- compaction thinks it succeeded and returns COMPACT_PARTIAL, but
allocation attempt fails
- and perhaps some other combinations
> Sure we can see a case where somebody is stealing the compacted blocks
> but that is very same with the order-0 where parallel mem eaters will
> piggy back on the reclaimer and there is no upper boundary as well well.
Yep.
