Re: [PATCH RESEND 0/8] hugetlb: add demote/split page functionality

[Hillf Danton <hdanton@sina.com>]

On Wed, 8 Sep 2021 14:00:19 -0700 Mike Kravetz wrote:
>On 9/7/21 1:50 AM, Hillf Danton wrote:
>> On Mon, 6 Sep 2021 16:40:28 +0200 Vlastimil Babka wrote:
>>> On 9/2/21 20:17, Mike Kravetz wrote:
>>>>
>>>> Here is some very high level information from a long stall that was
>>>> interrupted.  This was an order 9 allocation from alloc_buddy_huge_page().
>>>>
>>>> 55269.530564] __alloc_pages_slowpath: jiffies 47329325 tries 609673 cpu_tries 1   node 0 FAIL
>>>> [55269.539893]     r_tries 25       c_tries 609647   reclaim 47325161 compact 607     
>>>>
>>>> Yes, in __alloc_pages_slowpath for 47329325 jiffies before being interrupted.
>>>> should_reclaim_retry returned true 25 times and should_compact_retry returned
>>>> true 609647 times.
>>>> Almost all time (47325161 jiffies) spent in __alloc_pages_direct_reclaim, and
>>>> 607 jiffies spent in __alloc_pages_direct_compact.
>>>>
>>>> Looks like both
>>>> reclaim retries > MAX_RECLAIM_RETRIES
>>>> and
>>>> compaction retries > MAX_COMPACT_RETRIES
>>>>
>>> Yeah AFAICS that's only possible with the scenario I suspected. I guess
>>> we should put a limit on compact retries (maybe some multiple of
>>> MAX_COMPACT_RETRIES) even if it thinks that reclaim could help, while
>>> clearly it doesn't (i.e. because somebody else is stealing the page like
>>> in your test case).
>> 
>> And/or clamp reclaim retries for costly orders
>> 
>> 	reclaim retries = MAX_RECLAIM_RETRIES - order;
>> 
>> to pull down the chance for stall as low as possible.
>
>Thanks, and sorry for not replying quickly.  I only get back to this as
>time allows.
>
>We could clamp the number of compaction and reclaim retries in
>__alloc_pages_slowpath as suggested.  However, I noticed that a single
>reclaim call could take a bunch of time.  As a result, I instrumented
>shrink_node to see what might be happening.  Here is some information
>from a long stall.  Note that I only dump stats when jiffies > 100000.
>
>[ 8136.874706] shrink_node: 507654 total jiffies,  3557110 tries
>[ 8136.881130]              130596341 reclaimed, 32 nr_to_reclaim
>[ 8136.887643]              compaction_suitable results:
>[ 8136.893276]     idx COMPACT_SKIPPED, 3557109
>[ 8672.399839] shrink_node: 522076 total jiffies,  3466228 tries
>[ 8672.406268]              124427720 reclaimed, 32 nr_to_reclaim
>[ 8672.412782]              compaction_suitable results:
>[ 8672.418421]     idx COMPACT_SKIPPED, 3466227
>[ 8908.099592] __alloc_pages_slowpath: jiffies 2939938  tries 17068 cpu_tries 1   node 0 success
>[ 8908.109120]     r_tries 11       c_tries 17056    reclaim 2939865  compact 9
>
>In this case, clamping the number of retries from should_compact_retry
>and should_reclaim_retry could help.  Mostly because we will not be
>calling back into the reclaim code?  Notice the long amount of time spent
>in shrink_node.  The 'tries' in shrink_node come about from that:
>
>	if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
>				    sc))
>		goto again;
>
>compaction_suitable results is the values returned from calls to
>should_continue_reclaim -> compaction_suitable.
>
>Trying to think if there might be an intelligent way to quit early.

Given the downgrade of costly order to zero on the kswapd side, what you
found suggests the need to bridge the gap between sc->nr_to_reclaim and
compact_gap(sc->order) for direct reclaims.

If nr_to_reclaim is the primary target for direct reclaim, one option is
ask kswapd to do costly order reclaims, with stall handled by r_tries and
c_tries in the slowpath.

+++ x/mm/vmscan.c
@@ -3220,7 +3220,11 @@ again:
 
 	if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
 				    sc))
-		goto again;
+		if (!current_is_kswapd() && sc->nr_reclaimed >=
+					    sc->nr_to_reclaim)
+			/* job done */ ;
+		else
+			goto again;
 
 	/*
 	 * Kswapd gives up on balancing particular nodes after too