Re: [RFC] mm/vmstat: Add events for THP migration without split

From: John Hubbard
Date: Thu May 14 2020 - 14:29:26 EST


On 2020-05-11 21:22, Anshuman Khandual wrote:
Add the following new trace events which will help in validating migration
events involving PMD based THP pages.

1. THP_PMD_MIGRATION_ENTRY_SET
2. THP_PMD_MIGRATION_ENTRY_REMOVE

There are no clear method to confirm whether a THP migration happened with
out involving it's split. These trace events along with PGMIGRATE_SUCCESS
and PGMIGRATE_FAILURE will provide additional insights. After this change,



Hi Anshuman,

It's very nice to see this work, and I think that reporting a bit more
about THP migration stats is going to make development and performance
debugging a lot more efficient (and pleasant).


A single 2M THP (2K base page) when migrated

1. Without split

................
pgmigrate_success 1
pgmigrate_fail 0
................
thp_pmd_migration_entry_set 1
thp_pmd_migration_entry_remove 1
................


I do think we should decouple the trace event name(s) just a *little* more,
from the mechanisms used to migrate THPs. In other words, let's report
the number of THP migration successes, and name it accordingly--rather
than "set" and "remove", which are pretty low-level and furthermore
depend on today's exact code.

Maybe Zi Yan's recommended name is exactly right, in fact:

THP_PMD_MIGRATION_SUCCESS


thanks,
--
John Hubbard
NVIDIA


2. With split

................
pgmigrate_success 512
pgmigrate_fail 0
................
thp_pmd_migration_entry_set 0
thp_pmd_migration_entry_remove 0
................

pgmigrate_success as 1 instead of 512, provides a hint for possible THP
migration event. But then it gets mixed with normal page migrations over
time. These additional trace events provide required co-relation.

Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Cc: "Kirill A. Shutemov" <kirill.shutemov@xxxxxxxxxxxxxxx>
Cc: Zi Yan <ziy@xxxxxxxxxx>
Cc: linux-mm@xxxxxxxxx
Cc: linux-kernel@xxxxxxxxxxxxxxx
Signed-off-by: Anshuman Khandual <anshuman.khandual@xxxxxxx>
---
This is an indirect way for counting PMD migrations without split. Is there
a better method possible ? Just request for comments at the moment.

include/linux/vm_event_item.h | 4 ++++
mm/migrate.c | 1 +
mm/rmap.c | 1 +
mm/vmstat.c | 4 ++++
4 files changed, 10 insertions(+)

diff --git a/include/linux/vm_event_item.h b/include/linux/vm_event_item.h
index ffef0f279747..4b25102cf3ad 100644
--- a/include/linux/vm_event_item.h
+++ b/include/linux/vm_event_item.h
@@ -91,6 +91,10 @@ enum vm_event_item { PGPGIN, PGPGOUT, PSWPIN, PSWPOUT,
THP_ZERO_PAGE_ALLOC_FAILED,
THP_SWPOUT,
THP_SWPOUT_FALLBACK,
+#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
+ THP_PMD_MIGRATION_ENTRY_SET,
+ THP_PMD_MIGRATION_ENTRY_REMOVE,
+#endif
#endif
#ifdef CONFIG_MEMORY_BALLOON
BALLOON_INFLATE,
diff --git a/mm/migrate.c b/mm/migrate.c
index 7160c1556f79..8d50d55cbe97 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -228,6 +228,7 @@ static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
if (!pvmw.pte) {
VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
remove_migration_pmd(&pvmw, new);
+ count_vm_event(THP_PMD_MIGRATION_ENTRY_REMOVE);
continue;
}
#endif
diff --git a/mm/rmap.c b/mm/rmap.c
index f79a206b271a..3c1fe3f45cb5 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1418,6 +1418,7 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
set_pmd_migration_entry(&pvmw, page);
+ count_vm_event(THP_PMD_MIGRATION_ENTRY_SET);
continue;
}
#endif
diff --git a/mm/vmstat.c b/mm/vmstat.c
index 96d21a792b57..a5254b7ee531 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c
@@ -1274,6 +1274,10 @@ const char * const vmstat_text[] = {
"thp_zero_page_alloc_failed",
"thp_swpout",
"thp_swpout_fallback",
+#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
+ "thp_pmd_migration_entry_set",
+ "thp_pmd_migration_entry_remove",
+#endif
#endif
#ifdef CONFIG_MEMORY_BALLOON
"balloon_inflate",