mm: hugetlb: free the 2nd vmemmap page associated with each HugeTLB page

Patch series "Free the 2nd vmemmap page associated with each HugeTLB
page", v7.

This series can minimize the overhead of struct page for 2MB HugeTLB
pages significantly.  It further reduces the overhead of struct page by
12.5% for a 2MB HugeTLB compared to the previous approach, which means
2GB per 1TB HugeTLB.  It is a nice gain.  Comments and reviews are
welcome.  Thanks.

The main implementation and details can refer to the commit log of patch
1.  In this series, I have changed the following four helpers, the
following table shows the impact of the overhead of those helpers.

	+------------------+-----------------------+
	|       APIs       | head page | tail page |
	+------------------+-----------+-----------+
	|    PageHead()    |     Y     |     N     |
	+------------------+-----------+-----------+
	|    PageTail()    |     Y     |     N     |
	+------------------+-----------+-----------+
	|  PageCompound()  |     N     |     N     |
	+------------------+-----------+-----------+
	|  compound_head() |     Y     |     N     |
	+------------------+-----------+-----------+

	Y: Overhead is increased.
	N: Overhead is _NOT_ increased.

It shows that the overhead of those helpers on a tail page don't change
between "hugetlb_free_vmemmap=on" and "hugetlb_free_vmemmap=off".  But the
overhead on a head page will be increased when "hugetlb_free_vmemmap=on"
(except PageCompound()).  So I believe that Matthew Wilcox's folio series
will help with this.

The users of PageHead() and PageTail() are much less than compound_head()
and most users of PageTail() are VM_BUG_ON(), so I have done some tests
about the overhead of compound_head() on head pages.

I have tested the overhead of calling compound_head() on a head page,
which is 2.11ns (Measure the call time of 10 million times
compound_head(), and then average).

For a head page whose address is not aligned with PAGE_SIZE or a
non-compound page, the overhead of compound_head() is 2.54ns which is
increased by 20%.  For a head page whose address is aligned with
PAGE_SIZE, the overhead of compound_head() is 2.97ns which is increased by
40%.  Most pages are the former.  I do not think the overhead is
significant since the overhead of compound_head() itself is low.

This patch (of 5):

This patch minimizes the overhead of struct page for 2MB HugeTLB pages
significantly.  It further reduces the overhead of struct page by 12.5%
for a 2MB HugeTLB compared to the previous approach, which means 2GB per
1TB HugeTLB (2MB type).

After the feature of "Free sonme vmemmap pages of HugeTLB page" is
enabled, the mapping of the vmemmap addresses associated with a 2MB
HugeTLB page becomes the figure below.

     HugeTLB                    struct pages(8 pages)         page frame(8 pages)
 +-----------+ ---virt_to_page---> +-----------+   mapping to   +-----------+---> PG_head
 |           |                     |     0     | -------------> |     0     |
 |           |                     +-----------+                +-----------+
 |           |                     |     1     | -------------> |     1     |
 |           |                     +-----------+                +-----------+
 |           |                     |     2     | ----------------^ ^ ^ ^ ^ ^
 |           |                     +-----------+                   | | | | |
 |           |                     |     3     | ------------------+ | | | |
 |           |                     +-----------+                     | | | |
 |           |                     |     4     | --------------------+ | | |
 |    2MB    |                     +-----------+                       | | |
 |           |                     |     5     | ----------------------+ | |
 |           |                     +-----------+                         | |
 |           |                     |     6     | ------------------------+ |
 |           |                     +-----------+                           |
 |           |                     |     7     | --------------------------+
 |           |                     +-----------+
 |           |
 |           |
 |           |
 +-----------+

As we can see, the 2nd vmemmap page frame (indexed by 1) is reused and
remaped. However, the 2nd vmemmap page frame is also can be freed to
the buddy allocator, then we can change the mapping from the figure
above to the figure below.

    HugeTLB                    struct pages(8 pages)         page frame(8 pages)
 +-----------+ ---virt_to_page---> +-----------+   mapping to   +-----------+---> PG_head
 |           |                     |     0     | -------------> |     0     |
 |           |                     +-----------+                +-----------+
 |           |                     |     1     | ---------------^ ^ ^ ^ ^ ^ ^
 |           |                     +-----------+                  | | | | | |
 |           |                     |     2     | -----------------+ | | | | |
 |           |                     +-----------+                    | | | | |
 |           |                     |     3     | -------------------+ | | | |
 |           |                     +-----------+                      | | | |
 |           |                     |     4     | ---------------------+ | | |
 |    2MB    |                     +-----------+                        | | |
 |           |                     |     5     | -----------------------+ | |
 |           |                     +-----------+                          | |
 |           |                     |     6     | -------------------------+ |
 |           |                     +-----------+                            |
 |           |                     |     7     | ---------------------------+
 |           |                     +-----------+
 |           |
 |           |
 |           |
 +-----------+

After we do this, all tail vmemmap pages (1-7) are mapped to the head
vmemmap page frame (0).  In other words, there are more than one page
struct with PG_head associated with each HugeTLB page.  We __know__ that
there is only one head page struct, the tail page structs with PG_head are
fake head page structs.  We need an approach to distinguish between those
two different types of page structs so that compound_head(), PageHead()
and PageTail() can work properly if the parameter is the tail page struct
but with PG_head.

The following code snippet describes how to distinguish between real and
fake head page struct.

	if (test_bit(PG_head, &page->flags)) {
		unsigned long head = READ_ONCE(page[1].compound_head);

		if (head & 1) {
			if (head == (unsigned long)page + 1)
				==> head page struct
			else
				==> tail page struct
		} else
			==> head page struct
	}

We can safely access the field of the @page[1] with PG_head because the
@page is a compound page composed with at least two contiguous pages.

[songmuchun@bytedance.com: restore lost comment changes]

Link: https://lkml.kernel.org/r/20211101031651.75851-1-songmuchun@bytedance.com
Link: https://lkml.kernel.org/r/20211101031651.75851-2-songmuchun@bytedance.com
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Reviewed-by: Barry Song <song.bao.hua@hisilicon.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Michal Hocko <mhocko@suse.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Chen Huang <chenhuang5@huawei.com>
Cc: Bodeddula Balasubramaniam <bodeddub@amazon.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Cc: Fam Zheng <fam.zheng@bytedance.com>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 74 insertions, 4 deletions

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index 1c3b6e5c8bfd..111e453f23d2 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -190,13 +190,69 @@ enum pageflags {
 
 #ifndef __GENERATING_BOUNDS_H
 
+#ifdef CONFIG_HUGETLB_PAGE_FREE_VMEMMAP
+extern bool hugetlb_free_vmemmap_enabled;
+
+/*
+ * If the feature of freeing some vmemmap pages associated with each HugeTLB
+ * page is enabled, the head vmemmap page frame is reused and all of the tail
+ * vmemmap addresses map to the head vmemmap page frame (furture details can
+ * refer to the figure at the head of the mm/hugetlb_vmemmap.c).  In other
+ * words, there are more than one page struct with PG_head associated with each
+ * HugeTLB page.  We __know__ that there is only one head page struct, the tail
+ * page structs with PG_head are fake head page structs.  We need an approach
+ * to distinguish between those two different types of page structs so that
+ * compound_head() can return the real head page struct when the parameter is
+ * the tail page struct but with PG_head.
+ *
+ * The page_fixed_fake_head() returns the real head page struct if the @page is
+ * fake page head, otherwise, returns @page which can either be a true page
+ * head or tail.
+ */
+static __always_inline const struct page *page_fixed_fake_head(const struct page *page)
+{
+	if (!hugetlb_free_vmemmap_enabled)
+		return page;
+
+	/*
+	 * Only addresses aligned with PAGE_SIZE of struct page may be fake head
+	 * struct page. The alignment check aims to avoid access the fields (
+	 * e.g. compound_head) of the @page[1]. It can avoid touch a (possibly)
+	 * cold cacheline in some cases.
+	 */
+	if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
+	    test_bit(PG_head, &page->flags)) {
+		/*
+		 * We can safely access the field of the @page[1] with PG_head
+		 * because the @page is a compound page composed with at least
+		 * two contiguous pages.
+		 */
+		unsigned long head = READ_ONCE(page[1].compound_head);
+
+		if (likely(head & 1))
+			return (const struct page *)(head - 1);
+	}
+	return page;
+}
+#else
+static inline const struct page *page_fixed_fake_head(const struct page *page)
+{
+	return page;
+}
+#endif
+
+static __always_inline int page_is_fake_head(struct page *page)
+{
+	return page_fixed_fake_head(page) != page;
+}
+
 static inline unsigned long _compound_head(const struct page *page)
 {
 	unsigned long head = READ_ONCE(page->compound_head);
 
 	if (unlikely(head & 1))
 		return head - 1;
-	return (unsigned long)page;
+	return (unsigned long)page_fixed_fake_head(page);
 }
 
 #define compound_head(page)	((typeof(page))_compound_head(page))
@@ -231,12 +287,13 @@ static inline unsigned long _compound_head(const struct page *page)
 
 static __always_inline int PageTail(struct page *page)
 {
-	return READ_ONCE(page->compound_head) & 1;
+	return READ_ONCE(page->compound_head) & 1 || page_is_fake_head(page);
 }
 
 static __always_inline int PageCompound(struct page *page)
 {
-	return test_bit(PG_head, &page->flags) || PageTail(page);
+	return test_bit(PG_head, &page->flags) ||
+	       READ_ONCE(page->compound_head) & 1;
 }
 
 #define	PAGE_POISON_PATTERN	-1l
@@ -695,7 +752,20 @@ static inline bool test_set_page_writeback(struct page *page)
 	return set_page_writeback(page);
 }
 
-__PAGEFLAG(Head, head, PF_ANY) CLEARPAGEFLAG(Head, head, PF_ANY)
+static __always_inline bool folio_test_head(struct folio *folio)
+{
+	return test_bit(PG_head, folio_flags(folio, FOLIO_PF_ANY));
+}
+
+static __always_inline int PageHead(struct page *page)
+{
+	PF_POISONED_CHECK(page);
+	return test_bit(PG_head, &page->flags) && !page_is_fake_head(page);
+}
+
+__SETPAGEFLAG(Head, head, PF_ANY)
+__CLEARPAGEFLAG(Head, head, PF_ANY)
+CLEARPAGEFLAG(Head, head, PF_ANY)
 
 /**
  * folio_test_large() - Does this folio contain more than one page?