question: should_compact_retry limit

question: should_compact_retry limit

[Mike Kravetz]

While looking at some really long hugetlb page allocation times, I noticed
instances where should_compact_retry() was returning true more often that
I expected.  In one allocation attempt, it returned true 765668 times in a
row.  To me, this was unexpected because of the following:

#define MAX_COMPACT_RETRIES 16
int max_retries = MAX_COMPACT_RETRIES;

However, if should_compact_retry() returns true via the following path we
do not increase the retry count.

	/*
	 * make sure the compaction wasn't deferred or didn't bail out early
	 * due to locks contention before we declare that we should give up.
	 * But do not retry if the given zonelist is not suitable for
	 * compaction.
	 */
	if (compaction_withdrawn(compact_result)) {
		ret = compaction_zonelist_suitable(ac, order, alloc_flags);
		goto out;
	}

Just curious, is this intentional?
-- 
Mike Kravetz

Re: question: should_compact_retry limit

[Vlastimil Babka]

On 6/5/19 1:30 AM, Mike Kravetz wrote:
> While looking at some really long hugetlb page allocation times, I noticed
> instances where should_compact_retry() was returning true more often that
> I expected.  In one allocation attempt, it returned true 765668 times in a
> row.  To me, this was unexpected because of the following:
> 
> #define MAX_COMPACT_RETRIES 16
> int max_retries = MAX_COMPACT_RETRIES;
> 
> However, if should_compact_retry() returns true via the following path we
> do not increase the retry count.
> 
> 	/*
> 	 * make sure the compaction wasn't deferred or didn't bail out early
> 	 * due to locks contention before we declare that we should give up.
> 	 * But do not retry if the given zonelist is not suitable for
> 	 * compaction.
> 	 */
> 	if (compaction_withdrawn(compact_result)) {
> 		ret = compaction_zonelist_suitable(ac, order, alloc_flags);
> 		goto out;
> 	}
> 
> Just curious, is this intentional?

Hmm I guess we didn't expect compaction_withdrawn() to be so
consistently returned. Do you know what value of compact_result is there
in your test?

Re: question: should_compact_retry limit

[Mike Kravetz]

On 6/5/19 12:58 AM, Vlastimil Babka wrote:
> On 6/5/19 1:30 AM, Mike Kravetz wrote:
>> While looking at some really long hugetlb page allocation times, I noticed
>> instances where should_compact_retry() was returning true more often that
>> I expected.  In one allocation attempt, it returned true 765668 times in a
>> row.  To me, this was unexpected because of the following:
>>
>> #define MAX_COMPACT_RETRIES 16
>> int max_retries = MAX_COMPACT_RETRIES;
>>
>> However, if should_compact_retry() returns true via the following path we
>> do not increase the retry count.
>>
>> 	/*
>> 	 * make sure the compaction wasn't deferred or didn't bail out early
>> 	 * due to locks contention before we declare that we should give up.
>> 	 * But do not retry if the given zonelist is not suitable for
>> 	 * compaction.
>> 	 */
>> 	if (compaction_withdrawn(compact_result)) {
>> 		ret = compaction_zonelist_suitable(ac, order, alloc_flags);
>> 		goto out;
>> 	}
>>
>> Just curious, is this intentional?
> 
> Hmm I guess we didn't expect compaction_withdrawn() to be so
> consistently returned. Do you know what value of compact_result is there
> in your test?

Added some instrumentation to record values and ran test,

557904 Total

549186 COMPACT_DEFERRED
  8718 COMPACT_PARTIAL_SKIPPED

Do note that this is not my biggest problem with these allocations.  That is
should_continue_reclaim returning true more often that in should.  Still
trying to get more info on that.  This was just something curious I also
discovered.
-- 
Mike Kravetz

Re: question: should_compact_retry limit

[Vlastimil Babka]

On 6/5/19 6:05 PM, Mike Kravetz wrote:
> On 6/5/19 12:58 AM, Vlastimil Babka wrote:
>> On 6/5/19 1:30 AM, Mike Kravetz wrote:
>> Hmm I guess we didn't expect compaction_withdrawn() to be so
>> consistently returned. Do you know what value of compact_result is there
>> in your test?
> 
> Added some instrumentation to record values and ran test,
> 
> 557904 Total
> 
> 549186 COMPACT_DEFERRED

Retrying mindlessly with compaction deferred sounds definitely wrong,
I'll try to look at it. Thanks.

>   8718 COMPACT_PARTIAL_SKIPPED
> 
> Do note that this is not my biggest problem with these allocations.  That is
> should_continue_reclaim returning true more often that in should.  Still
> trying to get more info on that.  This was just something curious I also
> discovered.
>