Skip to contentSkip to conversation

mm/gup: speed up check_and_migrate_cma_pages() on huge page

PATCH Superseded Pingfan Liu 3 participants 4 messages Jun 26, 2019 – Jun 27, 2019

Subsystems (2)

v4 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

PATCH v4 1/1 mm/gup: speed up check_and_migrate_cma_pages() on huge page

Both hugetlb and thp locate on the same migration type of pageblock, since they are allocated from a free_list[]. Based on this fact, it is enough to check on a single subpage to decide the migration type of the whole huge page. By this way, it saves (2M/4K - 1) times loop for pmd_huge on x86, similar on other archs.

Furthermore, when executing isolate_huge_page(), it avoid taking global hugetlb_lock many times, and meanless remove/add to the local link list cma_page_list.

Cc Andrew Morton <[email protected]>
Cc Ira Weiny <[email protected]>
Cc Mike Rapoport <[email protected]>
Cc "Kirill A. Shutemov" <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc John Hubbard <[email protected]>
Cc "Aneesh Kumar K.V" <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Keith Busch <[email protected]>
Cc Mike Kravetz <[email protected]>

---
v3 -> v4: fix C language precedence issue

 mm/gup.c | 19 ++++++++++++-------
 1 file changed, 12 insertions(+), 7 deletions(-)
mm/gup.c +12 -7
index ddde097..ffca55b 100644
@@ -1342,19 +1342,22 @@ static long check_and_migrate_cma_pages(struct task_struct *tsk,
 	LIST_HEAD(cma_page_list);
 
 check_again:
-	for (i = 0; i < nr_pages; i++) {
+	for (i = 0; i < nr_pages;) {
+
+		struct page *head = compound_head(pages[i]);
+		long step = 1;
+
+		if (PageCompound(head))
+			step = (1 << compound_order(head)) - (pages[i] - head);
 		/*
 		 * If we get a page from the CMA zone, since we are going to
 		 * be pinning these entries, we might as well move them out
 		 * of the CMA zone if possible.
 		 */
-		if (is_migrate_cma_page(pages[i])) {
-
-			struct page *head = compound_head(pages[i]);
-
-			if (PageHuge(head)) {
+		if (is_migrate_cma_page(head)) {
+			if (PageHuge(head))
 				isolate_huge_page(head, &cma_page_list);
-			} else {
+			else {
 				if (!PageLRU(head) && drain_allow) {
 					lru_add_drain_all();
 					drain_allow = false;
@@ -1369,6 +1372,8 @@ static long check_and_migrate_cma_pages(struct task_struct *tsk,
 				}
 			}
 		}
+
+		i += step;
 	}
 
 	if (!list_empty(&cma_page_list)) {
Signature · 1 line
2.7.5
Reviewed-by on v4 1/1
#1 lore raw
Quoted 11 lines · Pingfan Liu

Both hugetlb and thp locate on the same migration type of pageblock, since they are allocated from a free_list[]. Based on this fact, it is enough to check on a single subpage to decide the migration type of the whole huge page. By this way, it saves (2M/4K - 1) times loop for pmd_huge on x86, similar on other archs.

Furthermore, when executing isolate_huge_page(), it avoid taking global hugetlb_lock many times, and meanless remove/add to the local link list cma_page_list.

Quoted 63 lines
Cc Andrew Morton <[email protected]>
Cc Ira Weiny <[email protected]>
Cc Mike Rapoport <[email protected]>
Cc "Kirill A. Shutemov" <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc John Hubbard <[email protected]>
Cc "Aneesh Kumar K.V" <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Keith Busch <[email protected]>
Cc Mike Kravetz <[email protected]>

---
v3 -> v4: fix C language precedence issue

 mm/gup.c | 19 ++++++++++++-------
 1 file changed, 12 insertions(+), 7 deletions(-)
mm/gup.c +12 -7
index ddde097..ffca55b 100644
@@ -1342,19 +1342,22 @@ static long check_and_migrate_cma_pages(struct task_struct *tsk,
 	LIST_HEAD(cma_page_list);
 
 check_again:
-	for (i = 0; i < nr_pages; i++) {
+	for (i = 0; i < nr_pages;) {
+
+		struct page *head = compound_head(pages[i]);
+		long step = 1;
+
+		if (PageCompound(head))
+			step = (1 << compound_order(head)) - (pages[i] - head);
 		/*
 		 * If we get a page from the CMA zone, since we are going to
 		 * be pinning these entries, we might as well move them out
 		 * of the CMA zone if possible.
 		 */
-		if (is_migrate_cma_page(pages[i])) {
-
-			struct page *head = compound_head(pages[i]);
-
-			if (PageHuge(head)) {
+		if (is_migrate_cma_page(head)) {
+			if (PageHuge(head))
 				isolate_huge_page(head, &cma_page_list);
-			} else {
+			else {
 				if (!PageLRU(head) && drain_allow) {
 					lru_add_drain_all();
 					drain_allow = false;
@@ -1369,6 +1372,8 @@ static long check_and_migrate_cma_pages(struct task_struct *tsk,
 				}
 			}
 		}
+
+		i += step;
 	}
 
 	if (!list_empty(&cma_page_list)) {
Signature · 1 line
2.7.5
maintainer
#2 lore raw
In reply to Pingfan Liu
Quoted 15 lines · On Wed, 26 Jun 2019 21:10:00 +0800 Pingfan Liu <[email protected]> wrote:

Both hugetlb and thp locate on the same migration type of pageblock, since they are allocated from a free_list[]. Based on this fact, it is enough to check on a single subpage to decide the migration type of the whole huge page. By this way, it saves (2M/4K - 1) times loop for pmd_huge on x86, similar on other archs.

Furthermore, when executing isolate_huge_page(), it avoid taking global hugetlb_lock many times, and meanless remove/add to the local link list cma_page_list.

...

mm/gup.c static long check_and_migrate_cma_pages(struct task_struct *tsk, · line 1350

Quoted 10 lines
mm/gup.c +6 -1
@@ -1342,19 +1342,22 @@ static long check_and_migrate_cma_pages(struct task_struct *tsk,
 	LIST_HEAD(cma_page_list);
 
 check_again:
-	for (i = 0; i < nr_pages; i++) {
+	for (i = 0; i < nr_pages;) {
+
+		struct page *head = compound_head(pages[i]);
+		long step = 1;
+
+		if (PageCompound(head))

I suspect this would work correctly if the PageCompound test was simply removed. Not that I'm really suggesting that it be removed - dunno.

mm/gup.c static long check_and_migrate_cma_pages(struct task_struct *tsk, · line 1351

Quoted 1 line
mm/gup.c +1 -0
+			step = (1 << compound_order(head)) - (pages[i] - head);

I don't understand this statement. Why does the position of this page in the pages[] array affect anything? There's an assumption about the contents of the skipped pages, I assume.

Could we please get a comment in here whcih fully explains the logic and any assumptions?

#3 lore raw
Quoted 20 lines · Andrew Morton
Quoted 17 lines · On Wed, 26 Jun 2019 21:10:00 +0800 Pingfan Liu <[email protected]> wrote:

Both hugetlb and thp locate on the same migration type of pageblock, since they are allocated from a free_list[]. Based on this fact, it is enough to check on a single subpage to decide the migration type of the whole huge page. By this way, it saves (2M/4K - 1) times loop for pmd_huge on x86, similar on other archs.

Furthermore, when executing isolate_huge_page(), it avoid taking global hugetlb_lock many times, and meanless remove/add to the local link list cma_page_list.

...

mm/gup.c +0 -0
@@ -1342,19 +1342,22 @@ static long check_and_migrate_cma_pages(struct task_struct *tsk,
      LIST_HEAD(cma_page_list);
 
Quoted 11 lines

mm/gup.c static long check_and_migrate_cma_pages(struct task_struct *tsk, · line 1350

Quoted 8 lines
mm/gup.c +6 -1
 check_again:
-     for (i = 0; i < nr_pages; i++) {
+     for (i = 0; i < nr_pages;) {
+
+             struct page *head = compound_head(pages[i]);
+             long step = 1;
+
+             if (PageCompound(head))

I suspect this would work correctly if the PageCompound test was simply removed. Not that I'm really suggesting that it be removed - dunno.

Yes, you are right. compound_order() can safely run on normal page, which means we can drop the check PageCompound().

Quoted 6 lines

mm/gup.c static long check_and_migrate_cma_pages(struct task_struct *tsk, · line 1351

Quoted 2 lines
mm/gup.c +1 -0
+                     step = (1 << compound_order(head)) - (pages[i] - head);

I don't understand this statement. Why does the position of this page in the pages[] array affect anything? There's an assumption about the contents of the skipped pages, I assume.

Because gup may start from a tail page.

Quoted 3 lines

Could we please get a comment in here whcih fully explains the logic and any assumptions?

Sure, I will.

Thanks,
Pingfan

Quoted 1 line