[PATCH 0/2] Better support for THP in page cache

[PATCH 0/2] Better support for THP in page cache

[Matthew Wilcox]

This is the first step towards representing THPs more efficiently in
the page cache.  The next step is to insert one head page instead of
HPAGE_PMD_NR.  This passes a Trinity run, but I'm sure there's a
better test-case out there for THP.

Matthew Wilcox (2):
  mm: Remove redundant test from find_get_pages_contig
  page cache: Store only head pages in i_pages

 include/linux/pagemap.h |   9 ++++
 mm/filemap.c            | 106 +++++++++++-----------------------------
 mm/khugepaged.c         |   4 +-
 mm/shmem.c              |   2 +-
 mm/swap_state.c         |   2 +-
 5 files changed, 42 insertions(+), 81 deletions(-)

-- 
2.19.1

[PATCH 1/2] mm: Remove redundant test from find_get_pages_contig

[Matthew Wilcox]

After we establish a reference on the page, we check the pointer continues
to be in the correct position in i_pages.  There's no need to check the
page->mapping or page->index afterwards; if those can change after we've
got the reference, they can change after we return the page to the caller.

Signed-off-by: Matthew Wilcox <willy@infradead.org>
---
 mm/filemap.c | 10 ----------
 1 file changed, 10 deletions(-)

diff --git a/mm/filemap.c b/mm/filemap.c
index 81adec8ee02cc..538531590ef2d 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1776,16 +1776,6 @@ unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
 		if (unlikely(page != xas_reload(&xas)))
 			goto put_page;
 
-		/*
-		 * must check mapping and index after taking the ref.
-		 * otherwise we can get both false positives and false
-		 * negatives, which is just confusing to the caller.
-		 */
-		if (!page->mapping || page_to_pgoff(page) != xas.xa_index) {
-			put_page(page);
-			break;
-		}
-
 		pages[ret] = page;
 		if (++ret == nr_pages)
 			break;
-- 
2.19.1

[PATCH 2/2] page cache: Store only head pages in i_pages

[Matthew Wilcox]

Transparent Huge Pages are currently stored in i_pages as pointers to
consecutive subpages.  This patch changes that to storing consecutive
pointers to the head page in preparation for storing huge pages more
efficiently in i_pages.

Large parts of this are "inspired" by Kirill's patch
https://lore.kernel.org/lkml/20170126115819.58875-2-kirill.shutemov@linux.intel.com/

Signed-off-by: Matthew Wilcox <willy@infradead.org>
---
 include/linux/pagemap.h |  9 ++++
 mm/filemap.c            | 96 +++++++++++++----------------------------
 mm/khugepaged.c         |  4 +-
 mm/shmem.c              |  2 +-
 mm/swap_state.c         |  2 +-
 5 files changed, 42 insertions(+), 71 deletions(-)

diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
index 226f96f0dee06..41bf976574e74 100644
--- a/include/linux/pagemap.h
+++ b/include/linux/pagemap.h
@@ -345,6 +345,15 @@ static inline struct page *grab_cache_page_nowait(struct address_space *mapping,
 			mapping_gfp_mask(mapping));
 }
 
+static inline struct page *find_subpage(struct page *page, pgoff_t offset)
+{
+	VM_BUG_ON_PAGE(PageTail(page), page);
+	VM_BUG_ON_PAGE(page->index > offset, page);
+	VM_BUG_ON_PAGE(page->index + (1 << compound_order(page)) <= offset,
+			page);
+	return page - page->index + offset;
+}
+
 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset);
 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset);
 unsigned find_get_entries(struct address_space *mapping, pgoff_t start,
diff --git a/mm/filemap.c b/mm/filemap.c
index 538531590ef2d..d7274591381ac 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1407,7 +1407,7 @@ EXPORT_SYMBOL(page_cache_prev_miss);
 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
 {
 	XA_STATE(xas, &mapping->i_pages, offset);
-	struct page *head, *page;
+	struct page *page;
 
 	rcu_read_lock();
 repeat:
@@ -1422,25 +1422,19 @@ struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
 	if (!page || xa_is_value(page))
 		goto out;
 
-	head = compound_head(page);
-	if (!page_cache_get_speculative(head))
-		goto repeat;
-
-	/* The page was split under us? */
-	if (compound_head(page) != head) {
-		put_page(head);
+	if (!page_cache_get_speculative(page))
 		goto repeat;
-	}
 
 	/*
-	 * Has the page moved?
+	 * Has the page moved or been split?
 	 * This is part of the lockless pagecache protocol. See
 	 * include/linux/pagemap.h for details.
 	 */
 	if (unlikely(page != xas_reload(&xas))) {
-		put_page(head);
+		put_page(page);
 		goto repeat;
 	}
+	page = find_subpage(page, offset);
 out:
 	rcu_read_unlock();
 
@@ -1611,7 +1605,6 @@ unsigned find_get_entries(struct address_space *mapping,
 
 	rcu_read_lock();
 	xas_for_each(&xas, page, ULONG_MAX) {
-		struct page *head;
 		if (xas_retry(&xas, page))
 			continue;
 		/*
@@ -1622,17 +1615,13 @@ unsigned find_get_entries(struct address_space *mapping,
 		if (xa_is_value(page))
 			goto export;
 
-		head = compound_head(page);
-		if (!page_cache_get_speculative(head))
+		if (!page_cache_get_speculative(page))
 			goto retry;
 
-		/* The page was split under us? */
-		if (compound_head(page) != head)
-			goto put_page;
-
-		/* Has the page moved? */
+		/* Has the page moved or been split? */
 		if (unlikely(page != xas_reload(&xas)))
 			goto put_page;
+		page = find_subpage(page, xas.xa_index);
 
 export:
 		indices[ret] = xas.xa_index;
@@ -1641,7 +1630,7 @@ unsigned find_get_entries(struct address_space *mapping,
 			break;
 		continue;
 put_page:
-		put_page(head);
+		put_page(page);
 retry:
 		xas_reset(&xas);
 	}
@@ -1683,33 +1672,27 @@ unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start,
 
 	rcu_read_lock();
 	xas_for_each(&xas, page, end) {
-		struct page *head;
 		if (xas_retry(&xas, page))
 			continue;
 		/* Skip over shadow, swap and DAX entries */
 		if (xa_is_value(page))
 			continue;
 
-		head = compound_head(page);
-		if (!page_cache_get_speculative(head))
+		if (!page_cache_get_speculative(page))
 			goto retry;
 
-		/* The page was split under us? */
-		if (compound_head(page) != head)
-			goto put_page;
-
-		/* Has the page moved? */
+		/* Has the page moved or been split? */
 		if (unlikely(page != xas_reload(&xas)))
 			goto put_page;
 
-		pages[ret] = page;
+		pages[ret] = find_subpage(page, xas.xa_index);
 		if (++ret == nr_pages) {
 			*start = page->index + 1;
 			goto out;
 		}
 		continue;
 put_page:
-		put_page(head);
+		put_page(page);
 retry:
 		xas_reset(&xas);
 	}
@@ -1754,7 +1737,6 @@ unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
 
 	rcu_read_lock();
 	for (page = xas_load(&xas); page; page = xas_next(&xas)) {
-		struct page *head;
 		if (xas_retry(&xas, page))
 			continue;
 		/*
@@ -1764,24 +1746,19 @@ unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
 		if (xa_is_value(page))
 			break;
 
-		head = compound_head(page);
-		if (!page_cache_get_speculative(head))
+		if (!page_cache_get_speculative(page))
 			goto retry;
 
-		/* The page was split under us? */
-		if (compound_head(page) != head)
-			goto put_page;
-
-		/* Has the page moved? */
+		/* Has the page moved or been split? */
 		if (unlikely(page != xas_reload(&xas)))
 			goto put_page;
 
-		pages[ret] = page;
+		pages[ret] = find_subpage(page, xas.xa_index);
 		if (++ret == nr_pages)
 			break;
 		continue;
 put_page:
-		put_page(head);
+		put_page(page);
 retry:
 		xas_reset(&xas);
 	}
@@ -1815,7 +1792,6 @@ unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
 
 	rcu_read_lock();
 	xas_for_each_marked(&xas, page, end, tag) {
-		struct page *head;
 		if (xas_retry(&xas, page))
 			continue;
 		/*
@@ -1826,26 +1802,21 @@ unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index,
 		if (xa_is_value(page))
 			continue;
 
-		head = compound_head(page);
-		if (!page_cache_get_speculative(head))
+		if (!page_cache_get_speculative(page))
 			goto retry;
 
-		/* The page was split under us? */
-		if (compound_head(page) != head)
-			goto put_page;
-
-		/* Has the page moved? */
+		/* Has the page moved or been split? */
 		if (unlikely(page != xas_reload(&xas)))
 			goto put_page;
 
-		pages[ret] = page;
+		pages[ret] = find_subpage(page, xas.xa_index);
 		if (++ret == nr_pages) {
 			*index = page->index + 1;
 			goto out;
 		}
 		continue;
 put_page:
-		put_page(head);
+		put_page(page);
 retry:
 		xas_reset(&xas);
 	}
@@ -1892,7 +1863,6 @@ unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
 
 	rcu_read_lock();
 	xas_for_each_marked(&xas, page, ULONG_MAX, tag) {
-		struct page *head;
 		if (xas_retry(&xas, page))
 			continue;
 		/*
@@ -1903,17 +1873,13 @@ unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
 		if (xa_is_value(page))
 			goto export;
 
-		head = compound_head(page);
-		if (!page_cache_get_speculative(head))
+		if (!page_cache_get_speculative(page))
 			goto retry;
 
-		/* The page was split under us? */
-		if (compound_head(page) != head)
-			goto put_page;
-
-		/* Has the page moved? */
+		/* Has the page moved or been split? */
 		if (unlikely(page != xas_reload(&xas)))
 			goto put_page;
+		page = find_subpage(page, xas.xa_index);
 
 export:
 		indices[ret] = xas.xa_index;
@@ -1922,7 +1888,7 @@ unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
 			break;
 		continue;
 put_page:
-		put_page(head);
+		put_page(page);
 retry:
 		xas_reset(&xas);
 	}
@@ -2533,7 +2499,7 @@ void filemap_map_pages(struct vm_fault *vmf,
 	pgoff_t last_pgoff = start_pgoff;
 	unsigned long max_idx;
 	XA_STATE(xas, &mapping->i_pages, start_pgoff);
-	struct page *head, *page;
+	struct page *page;
 
 	rcu_read_lock();
 	xas_for_each(&xas, page, end_pgoff) {
@@ -2542,17 +2508,13 @@ void filemap_map_pages(struct vm_fault *vmf,
 		if (xa_is_value(page))
 			goto next;
 
-		head = compound_head(page);
-		if (!page_cache_get_speculative(head))
+		if (!page_cache_get_speculative(page))
 			goto next;
 
-		/* The page was split under us? */
-		if (compound_head(page) != head)
-			goto skip;
-
-		/* Has the page moved? */
+		/* Has the page moved or been split? */
 		if (unlikely(page != xas_reload(&xas)))
 			goto skip;
+		page = find_subpage(page, xas.xa_index);
 
 		if (!PageUptodate(page) ||
 				PageReadahead(page) ||
diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index c13625c1ad5e5..7d6a1319dd42e 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -1363,7 +1363,7 @@ static void collapse_shmem(struct mm_struct *mm,
 				result = SCAN_FAIL;
 				break;
 			}
-			xas_store(&xas, new_page + (index % HPAGE_PMD_NR));
+			xas_store(&xas, new_page);
 			nr_none++;
 			continue;
 		}
@@ -1431,7 +1431,7 @@ static void collapse_shmem(struct mm_struct *mm,
 		list_add_tail(&page->lru, &pagelist);
 
 		/* Finally, replace with the new page. */
-		xas_store(&xas, new_page + (index % HPAGE_PMD_NR));
+		xas_store(&xas, new_page);
 		continue;
 out_lru:
 		xas_unlock_irq(&xas);
diff --git a/mm/shmem.c b/mm/shmem.c
index ea26d7a0342d7..c2ba84cbb0c0e 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -603,7 +603,7 @@ static int shmem_add_to_page_cache(struct page *page,
 		if (xas_error(&xas))
 			goto unlock;
 next:
-		xas_store(&xas, page + i);
+		xas_store(&xas, page);
 		if (++i < nr) {
 			xas_next(&xas);
 			goto next;
diff --git a/mm/swap_state.c b/mm/swap_state.c
index fd2f21e1c60ae..dcf9e466d2945 100644
--- a/mm/swap_state.c
+++ b/mm/swap_state.c
@@ -132,7 +132,7 @@ int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp)
 		for (i = 0; i < nr; i++) {
 			VM_BUG_ON_PAGE(xas.xa_index != idx + i, page);
 			set_page_private(page + i, entry.val + i);
-			xas_store(&xas, page + i);
+			xas_store(&xas, page);
 			xas_next(&xas);
 		}
 		address_space->nrpages += nr;
-- 
2.19.1

Re: [PATCH 1/2] mm: Remove redundant test from find_get_pages_contig

[Kirill A. Shutemov]

On Thu, Nov 22, 2018 at 01:32:23PM -0800, Matthew Wilcox wrote:
> After we establish a reference on the page, we check the pointer continues
> to be in the correct position in i_pages.  There's no need to check the
> page->mapping or page->index afterwards; if those can change after we've
> got the reference, they can change after we return the page to the caller.

Hm. IIRC, page->mapping can be set to NULL due truncation, but what about
index? When it can be changed? Truncation doesn't touch it.

-- 
 Kirill A. Shutemov

Re: [PATCH 2/2] page cache: Store only head pages in i_pages

[Kirill A. Shutemov]

On Thu, Nov 22, 2018 at 01:32:24PM -0800, Matthew Wilcox wrote:
> Transparent Huge Pages are currently stored in i_pages as pointers to
> consecutive subpages.  This patch changes that to storing consecutive
> pointers to the head page in preparation for storing huge pages more
> efficiently in i_pages.

I probably miss something, I don't see how it wouldn't break
split_huge_page().

I don't see what would replace head pages in i_pages with
formerly-tail-pages?

Hm?

-- 
 Kirill A. Shutemov

Re: [PATCH 2/2] page cache: Store only head pages in i_pages

[Matthew Wilcox]

On Fri, Nov 23, 2018 at 01:56:44PM +0300, Kirill A. Shutemov wrote:
> On Thu, Nov 22, 2018 at 01:32:24PM -0800, Matthew Wilcox wrote:
> > Transparent Huge Pages are currently stored in i_pages as pointers to
> > consecutive subpages.  This patch changes that to storing consecutive
> > pointers to the head page in preparation for storing huge pages more
> > efficiently in i_pages.
> 
> I probably miss something, I don't see how it wouldn't break
> split_huge_page().
> 
> I don't see what would replace head pages in i_pages with
> formerly-tail-pages?

You're quite right.  Where's your test-suite?  ;-)

I think this should do the job:

+++ b/mm/huge_memory.c
@@ -2464,6 +2464,9 @@ static void __split_huge_page(struct page *page, struct list_head *list,
                        if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
                                shmem_uncharge(head->mapping->host, 1);
                        put_page(head + i);
+               } else if (!PageAnon(page)) {
+                       __xa_store(&head->mapping->i_pages, head[i].index,
+                                       head + i, 0);
                }
        }
 

Having looked at this area, I think there was actually a bug in the patch
you wrote that I'm cribbing from.  You inserted the tail pages before
calling __split_huge_page_tail(), so a racing lookup would have found
a tail page before it got transformed into a non-tail page.

Re: [PATCH 2/2] page cache: Store only head pages in i_pages

[Kirill A. Shutemov]

On Fri, Nov 23, 2018 at 09:19:00AM -0800, Matthew Wilcox wrote:
> On Fri, Nov 23, 2018 at 01:56:44PM +0300, Kirill A. Shutemov wrote:
> > On Thu, Nov 22, 2018 at 01:32:24PM -0800, Matthew Wilcox wrote:
> > > Transparent Huge Pages are currently stored in i_pages as pointers to
> > > consecutive subpages.  This patch changes that to storing consecutive
> > > pointers to the head page in preparation for storing huge pages more
> > > efficiently in i_pages.
> > 
> > I probably miss something, I don't see how it wouldn't break
> > split_huge_page().
> > 
> > I don't see what would replace head pages in i_pages with
> > formerly-tail-pages?
> 
> You're quite right.  Where's your test-suite?  ;-)

Yeah-yeah...

> I think this should do the job:
> 
> +++ b/mm/huge_memory.c
> @@ -2464,6 +2464,9 @@ static void __split_huge_page(struct page *page, struct list_head *list,
>                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
>                                 shmem_uncharge(head->mapping->host, 1);
>                         put_page(head + i);
> +               } else if (!PageAnon(page)) {
> +                       __xa_store(&head->mapping->i_pages, head[i].index,
> +                                       head + i, 0);
>                 }
>         }

Looks good to me. But I still need to look into the rest of the patch.

> Having looked at this area, I think there was actually a bug in the patch
> you wrote that I'm cribbing from.  You inserted the tail pages before
> calling __split_huge_page_tail(), so a racing lookup would have found
> a tail page before it got transformed into a non-tail page.

I don't think so.

The page still has refcount==0 and any lookup of the page suppose to fail
due to !page_cache_get_speculative() or block on tree lock.

-- 
 Kirill A. Shutemov

Re: [PATCH 1/2] mm: Remove redundant test from find_get_pages_contig

[Matthew Wilcox]

On Fri, Nov 23, 2018 at 01:47:32PM +0300, Kirill A. Shutemov wrote:
> On Thu, Nov 22, 2018 at 01:32:23PM -0800, Matthew Wilcox wrote:
> > After we establish a reference on the page, we check the pointer continues
> > to be in the correct position in i_pages.  There's no need to check the
> > page->mapping or page->index afterwards; if those can change after we've
> > got the reference, they can change after we return the page to the caller.
> 
> Hm. IIRC, page->mapping can be set to NULL due truncation, but what about
> index? When it can be changed? Truncation doesn't touch it.

I think index can only be changed after the refcount has hit zero and
the page is safely out of the pagecache.  I agree that page->mapping can
be set to NULL after the call to xas_reload() ... but then it can also
happen after the check, so the check isn't really buying us anything
that the xas_reload() call doesn't already check.