Re: [RFC PATCH 2/3] mm, compaction: use MIN_COMPACT_COSTLY_PRIORITY everywhere for costly orders

[Mike Kravetz <mike.kravetz@oracle.com>]

On 7/31/19 5:06 AM, Vlastimil Babka wrote:
> On 7/24/19 7:50 PM, Mike Kravetz wrote:
>> For PAGE_ALLOC_COSTLY_ORDER allocations, MIN_COMPACT_COSTLY_PRIORITY is
>> minimum (highest priority).  Other places in the compaction code key off
>> of MIN_COMPACT_PRIORITY.  Costly order allocations will never get to
>> MIN_COMPACT_PRIORITY.  Therefore, some conditions will never be met for
>> costly order allocations.
>>
>> This was observed when hugetlb allocations could stall for minutes or
>> hours when should_compact_retry() would return true more often then it
>> should.  Specifically, this was in the case where compact_result was
>> COMPACT_DEFERRED and COMPACT_PARTIAL_SKIPPED and no progress was being
>> made.
> 
> Hmm, the point of MIN_COMPACT_COSTLY_PRIORITY was that costly
> allocations will not reach the priority where compaction becomes too
> expensive. With your patch, they still don't reach that priority value,
> but are allowed to be thorough anyway, even sooner. That just seems like
> a wrong way to fix the problem.

Thanks Vlastimil, here is why I took the approach I did.

I instrumented some of the long stalls.  Here is one common example:
should_compact_retry returned true 5000000 consecutive times.  However,
the variable compaction_retries is zero.  We never get to the code that
increments the compaction_retries count because compaction_made_progress
is false and compaction_withdrawn is true.  As suggested earlier, I noted
why compaction_withdrawn is true.  Of the 5000000 calls,
4921875 were COMPACT_DEFERRED
78125 were COMPACT_PARTIAL_SKIPPED
Note that 5000000/64(1 << COMPACT_MAX_DEFER_SHIFT) == 78125

I then started looking into why COMPACT_DEFERRED and COMPACT_PARTIAL_SKIPPED
were being set/returned so often.
COMPACT_DEFERRED is set/returned in try_to_compact_pages.  Specifically,
		if (prio > MIN_COMPACT_PRIORITY
					&& compaction_deferred(zone, order)) {
			rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
			continue;
		}
COMPACT_PARTIAL_SKIPPED is set/returned in __compact_finished. Specifically,
	if (compact_scanners_met(cc)) {
		/* Let the next compaction start anew. */
		reset_cached_positions(cc->zone);

		/* ... */

		if (cc->direct_compaction)
			cc->zone->compact_blockskip_flush = true;

		if (cc->whole_zone)
			return COMPACT_COMPLETE;
		else
			return COMPACT_PARTIAL_SKIPPED;
	}

In both cases, compact_priority being MIN_COMPACT_COSTLY_PRIORITY and not
being able to go to MIN_COMPACT_PRIORITY caused the 'compaction_withdrawn'
result to be set/returned.

I do not know the subtleties of the compaction code, but it seems like
retrying in this manner does not make sense.

>                                 If should_compact_retry() returns
> misleading results for costly allocations, then that should be fixed
> instead?
> 
> Alternatively, you might want to say that hugetlb allocations are not
> like other random costly allocations, because the admin setting
> nr_hugepages is prepared to take the cost (I thought that was indicated
> by the __GFP_RETRY_MAYFAIL flag, but seeing all the other users of it,
> I'm not sure anymore).

The example above, resulted in a stall of a little over 5 minutes.  However,
I have seen them last for hours.  Sure, the caller (admin for hugetlbfs)
knows there may be high costs.  But, I think minutes/hours to try and allocate
a single huge page is too much.  We should fail sooner that that.

>                        In that case should_compact_retry() could take
> __GFP_RETRY_MAYFAIL into account and allow MIN_COMPACT_PRIORITY even for
> costly allocations.

I'll put something like this together to test.
-- 
Mike Kravetz