On 4/14/2023 7:14 AM, Mel Gorman wrote:
A bug was reported by Yuanxi Liu where allocating 1G pages at runtime is
taking an excessive amount of time for large amounts of memory. Further
testing allocating huge pages that the cost is linear i.e. if allocating
1G pages in batches of 10 then the time to allocate nr_hugepages from
10->20->30->etc increases linearly even though 10 pages are allocated at
each step. Profiles indicated that much of the time is spent checking the
validity within already existing huge pages and then attempting a migration
that fails after isolating the range, draining pages and a whole lot of
other useless work.
Commit eb14d4eefdc4 ("mm,page_alloc: drop unnecessary checks from
pfn_range_valid_contig") removed two checks, one which ignored huge pages
for contiguous allocations as huge pages can sometimes migrate. While
there may be value on migrating a 2M page to satisfy a 1G allocation, it's
potentially expensive if the 1G allocation fails and it's pointless to
try moving a 1G page for a new 1G allocation or scan the tail pages for
valid PFNs.
Reintroduce the PageHuge check and assume any contiguous region with
hugetlbfs pages is unsuitable for a new 1G allocation.
The hpagealloc test allocates huge pages in batches and reports the
average latency per page over time. This test happens just after boot when
fragmentation is not an issue. Units are in milliseconds.
hpagealloc
6.3.0-rc6 6.3.0-rc6 6.3.0-rc6
vanilla hugeallocrevert-v1r1 hugeallocsimple-v1r2
Min Latency 26.42 ( 0.00%) 5.07 ( 80.82%) 18.94 ( 28.30%)
1st-qrtle Latency 356.61 ( 0.00%) 5.34 ( 98.50%) 19.85 ( 94.43%)
2nd-qrtle Latency 697.26 ( 0.00%) 5.47 ( 99.22%) 20.44 ( 97.07%)
3rd-qrtle Latency 972.94 ( 0.00%) 5.50 ( 99.43%) 20.81 ( 97.86%)
Max-1 Latency 26.42 ( 0.00%) 5.07 ( 80.82%) 18.94 ( 28.30%)
Max-5 Latency 82.14 ( 0.00%) 5.11 ( 93.78%) 19.31 ( 76.49%)
Max-10 Latency 150.54 ( 0.00%) 5.20 ( 96.55%) 19.43 ( 87.09%)
Max-90 Latency 1164.45 ( 0.00%) 5.53 ( 99.52%) 20.97 ( 98.20%)
Max-95 Latency 1223.06 ( 0.00%) 5.55 ( 99.55%) 21.06 ( 98.28%)
Max-99 Latency 1278.67 ( 0.00%) 5.57 ( 99.56%) 22.56 ( 98.24%)
Max Latency 1310.90 ( 0.00%) 8.06 ( 99.39%) 26.62 ( 97.97%)
Amean Latency 678.36 ( 0.00%) 5.44 * 99.20%* 20.44 * 96.99%*
6.3.0-rc6 6.3.0-rc6 6.3.0-rc6
vanilla revert-v1 hugeallocfix-v2
Duration User 0.28 0.27 0.30
Duration System 808.66 17.77 35.99
Duration Elapsed 830.87 18.08 36.33
The vanilla kernel is poor, taking up to 1.3 second to allocate a huge page
and almost 10 minutes in total to run the test. Reverting the problematic
commit reduces it to 8ms at worst and the patch takes 26ms. This patch
fixes the main issue with skipping huge pages but leaves the page_count()
out because a page with an elevated count potentially can migrate.
A while ago I submitted this patch set that should significantly improve
the chances of a 2MB Huge Page being successfully migrated:
https://lore.kernel.org/linux-mm/20220921223639.1152392-1-opendmb@xxxxxxxxx/
Unfortunately, it is collecting dust and needs to be updated to support
Folios, but I would be curious how it affects the performance of this test.
BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=217022
Fixes: eb14d4eefdc4 ("mm,page_alloc: drop unnecessary checks from pfn_range_valid_contig")
Reported-by: Yuanxi Liu <y.liu@xxxxxxxxxxx>
Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
---
mm/page_alloc.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 7136c36c5d01..b47f520c3051 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -9450,6 +9450,9 @@ static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
if (PageReserved(page))
return false;
+
+ if (PageHuge(page))
+ return false;
}
return true;
}
I am OK with this patch too, but I could resubmit my patch with Mike's
suggested variant and Folio support if anyone wants to try it instead of
this approach.
Regards,
Doug
