On Fri, 20 Oct 2023, Baolin Wang wrote:
> When doing compaction, I found the lru_add_drain() is an obvious hotspot
> when migrating pages. The distribution of this hotspot is as follows:
> - 18.75% compact_zone
> - 17.39% migrate_pages
> - 13.79% migrate_pages_batch
> - 11.66% migrate_folio_move
> - 7.02% lru_add_drain
> + 7.02% lru_add_drain_cpu
> + 3.00% move_to_new_folio
> 1.23% rmap_walk
> + 1.92% migrate_folio_unmap
> + 3.20% migrate_pages_sync
> + 0.90% isolate_migratepages
>
> The lru_add_drain() was added by commit c3096e6782b7 ("mm/migrate:
> __unmap_and_move() push good newpage to LRU") to drain the newpage to LRU
> immediately, to help to build up the correct newpage->mlock_count in
> remove_migration_ptes() for mlocked pages. However, if there are no mlocked
> pages are migrating, then we can avoid this lru drain operation, especailly
> for the heavy concurrent scenarios.
>
> So we can record the source pages' mlocked status in migrate_folio_unmap(),
> and only drain the lru list when the mlocked status is set in migrate_folio_move().
> In addition, the page was already isolated from lru when migrating, so checking
> the mlocked status is stable by folio_test_mlocked() in migrate_folio_unmap().
>
> After this patch, I can see the hotpot of the lru_add_drain() is gone:
> - 9.41% migrate_pages_batch
> - 6.15% migrate_folio_move
> - 3.64% move_to_new_folio
> + 1.80% migrate_folio_extra
> + 1.70% buffer_migrate_folio
> + 1.41% rmap_walk
> + 0.62% folio_add_lru
> + 3.07% migrate_folio_unmap
>
> Meanwhile, the compaction latency shows some improvements when running
> thpscale:
> base patched
> Amean fault-both-1 1131.22 ( 0.00%) 1112.55 * 1.65%*
> Amean fault-both-3 2489.75 ( 0.00%) 2324.15 * 6.65%*
> Amean fault-both-5 3257.37 ( 0.00%) 3183.18 * 2.28%*
> Amean fault-both-7 4257.99 ( 0.00%) 4079.04 * 4.20%*
> Amean fault-both-12 6614.02 ( 0.00%) 6075.60 * 8.14%*
> Amean fault-both-18 10607.78 ( 0.00%) 8978.86 * 15.36%*
> Amean fault-both-24 14911.65 ( 0.00%) 11619.55 * 22.08%*
> Amean fault-both-30 14954.67 ( 0.00%) 14925.66 * 0.19%*
> Amean fault-both-32 16654.87 ( 0.00%) 15580.31 * 6.45%*
>
Seems a sensible change with good results (I'll conceal how little of
the stats I understand, I expect everyone else understands them: in my
naivety, I'm mainly curious why rmap_walk's 1.23% didn't get a + on it).
> Signed-off-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
> ---
> Chages from v1:
> - Use separate flags in __migrate_folio_record() to avoid to pack flags
> in each call site per Ying.
> ---
> mm/migrate.c | 47 +++++++++++++++++++++++++++++++++++------------
> 1 file changed, 35 insertions(+), 12 deletions(-)
>
> diff --git a/mm/migrate.c b/mm/migrate.c
> index 125194f5af0f..fac96139dbba 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -1027,22 +1027,39 @@ union migration_ptr {
> struct anon_vma *anon_vma;
> struct address_space *mapping;
> };
> +
> +enum {
> + PAGE_WAS_MAPPED = 1 << 0,
> + PAGE_WAS_MLOCKED = 1 << 1,
> +};
> +
I was whispering to myself "I bet someone will suggest BIT()";
and indeed that someone has turned out to be Huang, Ying.
> static void __migrate_folio_record(struct folio *dst,
> - unsigned long page_was_mapped,
> + unsigned int page_was_mapped,
> + unsigned int page_was_mlocked,
> struct anon_vma *anon_vma)
> {
> union migration_ptr ptr = { .anon_vma = anon_vma };
> + unsigned long page_flags = 0;
Huang, Ying preferred a different name, me too: old_page_state?
> +
> + if (page_was_mapped)
> + page_flags |= PAGE_WAS_MAPPED;
> + if (page_was_mlocked)
> + page_flags |= PAGE_WAS_MLOCKED;
What's annoying me about the patch is all this mix of page_was_mapped and
page_was_mlocked variables, then the old_page_state bits. Can't it be
done with PAGE_WAS_ bits in old_page_state throughout, without any
page_was_mapped and page_was_mlocked variables?
Hugh
