1 files changed, 67 insertions, 46 deletions

diff --git a/mm/filemap.c b/mm/filemap.c
index da7a35d83de7..61b441b191ad 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -101,7 +101,7 @@
  *    ->tree_lock		(page_remove_rmap->set_page_dirty)
  *    bdi.wb->list_lock		(page_remove_rmap->set_page_dirty)
  *    ->inode->i_lock		(page_remove_rmap->set_page_dirty)
- *    ->memcg->move_lock	(page_remove_rmap->mem_cgroup_begin_page_stat)
+ *    ->memcg->move_lock	(page_remove_rmap->lock_page_memcg)
  *    bdi.wb->list_lock		(zap_pte_range->set_page_dirty)
  *    ->inode->i_lock		(zap_pte_range->set_page_dirty)
  *    ->private_lock		(zap_pte_range->__set_page_dirty_buffers)
@@ -176,11 +176,9 @@ static void page_cache_tree_delete(struct address_space *mapping,
 /*
  * Delete a page from the page cache and free it. Caller has to make
  * sure the page is locked and that nobody else uses it - or that usage
- * is safe.  The caller must hold the mapping's tree_lock and
- * mem_cgroup_begin_page_stat().
+ * is safe.  The caller must hold the mapping's tree_lock.
  */
-void __delete_from_page_cache(struct page *page, void *shadow,
-			      struct mem_cgroup *memcg)
+void __delete_from_page_cache(struct page *page, void *shadow)
 {
 	struct address_space *mapping = page->mapping;
 
@@ -239,8 +237,7 @@ void __delete_from_page_cache(struct page *page, void *shadow,
 	 * anyway will be cleared before returning page into buddy allocator.
 	 */
 	if (WARN_ON_ONCE(PageDirty(page)))
-		account_page_cleaned(page, mapping, memcg,
-				     inode_to_wb(mapping->host));
+		account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
 }
 
 /**
@@ -254,7 +251,6 @@ void __delete_from_page_cache(struct page *page, void *shadow,
 void delete_from_page_cache(struct page *page)
 {
 	struct address_space *mapping = page->mapping;
-	struct mem_cgroup *memcg;
 	unsigned long flags;
 
 	void (*freepage)(struct page *);
@@ -263,11 +259,9 @@ void delete_from_page_cache(struct page *page)
 
 	freepage = mapping->a_ops->freepage;
 
-	memcg = mem_cgroup_begin_page_stat(page);
 	spin_lock_irqsave(&mapping->tree_lock, flags);
-	__delete_from_page_cache(page, NULL, memcg);
+	__delete_from_page_cache(page, NULL);
 	spin_unlock_irqrestore(&mapping->tree_lock, flags);
-	mem_cgroup_end_page_stat(memcg);
 
 	if (freepage)
 		freepage(page);
@@ -551,7 +545,6 @@ int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
 	if (!error) {
 		struct address_space *mapping = old->mapping;
 		void (*freepage)(struct page *);
-		struct mem_cgroup *memcg;
 		unsigned long flags;
 
 		pgoff_t offset = old->index;
@@ -561,9 +554,8 @@ int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
 		new->mapping = mapping;
 		new->index = offset;
 
-		memcg = mem_cgroup_begin_page_stat(old);
 		spin_lock_irqsave(&mapping->tree_lock, flags);
-		__delete_from_page_cache(old, NULL, memcg);
+		__delete_from_page_cache(old, NULL);
 		error = radix_tree_insert(&mapping->page_tree, offset, new);
 		BUG_ON(error);
 		mapping->nrpages++;
@@ -576,8 +568,7 @@ int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
 		if (PageSwapBacked(new))
 			__inc_zone_page_state(new, NR_SHMEM);
 		spin_unlock_irqrestore(&mapping->tree_lock, flags);
-		mem_cgroup_end_page_stat(memcg);
-		mem_cgroup_replace_page(old, new);
+		mem_cgroup_migrate(old, new);
 		radix_tree_preload_end();
 		if (freepage)
 			freepage(old);
@@ -1668,6 +1659,15 @@ find_page:
 					index, last_index - index);
 		}
 		if (!PageUptodate(page)) {
+			/*
+			 * See comment in do_read_cache_page on why
+			 * wait_on_page_locked is used to avoid unnecessarily
+			 * serialisations and why it's safe.
+			 */
+			wait_on_page_locked_killable(page);
+			if (PageUptodate(page))
+				goto page_ok;
+
 			if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
 					!mapping->a_ops->is_partially_uptodate)
 				goto page_not_up_to_date;
@@ -2303,7 +2303,7 @@ static struct page *wait_on_page_read(struct page *page)
 	return page;
 }
 
-static struct page *__read_cache_page(struct address_space *mapping,
+static struct page *do_read_cache_page(struct address_space *mapping,
 				pgoff_t index,
 				int (*filler)(void *, struct page *),
 				void *data,
@@ -2325,53 +2325,74 @@ repeat:
 			/* Presumably ENOMEM for radix tree node */
 			return ERR_PTR(err);
 		}
+
+filler:
 		err = filler(data, page);
 		if (err < 0) {
 			page_cache_release(page);
-			page = ERR_PTR(err);
-		} else {
-			page = wait_on_page_read(page);
+			return ERR_PTR(err);
 		}
-	}
-	return page;
-}
-
-static struct page *do_read_cache_page(struct address_space *mapping,
-				pgoff_t index,
-				int (*filler)(void *, struct page *),
-				void *data,
-				gfp_t gfp)
 
-{
-	struct page *page;
-	int err;
+		page = wait_on_page_read(page);
+		if (IS_ERR(page))
+			return page;
+		goto out;
+	}
+	if (PageUptodate(page))
+		goto out;
 
-retry:
-	page = __read_cache_page(mapping, index, filler, data, gfp);
-	if (IS_ERR(page))
-		return page;
+	/*
+	 * Page is not up to date and may be locked due one of the following
+	 * case a: Page is being filled and the page lock is held
+	 * case b: Read/write error clearing the page uptodate status
+	 * case c: Truncation in progress (page locked)
+	 * case d: Reclaim in progress
+	 *
+	 * Case a, the page will be up to date when the page is unlocked.
+	 *    There is no need to serialise on the page lock here as the page
+	 *    is pinned so the lock gives no additional protection. Even if the
+	 *    the page is truncated, the data is still valid if PageUptodate as
+	 *    it's a race vs truncate race.
+	 * Case b, the page will not be up to date
+	 * Case c, the page may be truncated but in itself, the data may still
+	 *    be valid after IO completes as it's a read vs truncate race. The
+	 *    operation must restart if the page is not uptodate on unlock but
+	 *    otherwise serialising on page lock to stabilise the mapping gives
+	 *    no additional guarantees to the caller as the page lock is
+	 *    released before return.
+	 * Case d, similar to truncation. If reclaim holds the page lock, it
+	 *    will be a race with remove_mapping that determines if the mapping
+	 *    is valid on unlock but otherwise the data is valid and there is
+	 *    no need to serialise with page lock.
+	 *
+	 * As the page lock gives no additional guarantee, we optimistically
+	 * wait on the page to be unlocked and check if it's up to date and
+	 * use the page if it is. Otherwise, the page lock is required to
+	 * distinguish between the different cases. The motivation is that we
+	 * avoid spurious serialisations and wakeups when multiple processes
+	 * wait on the same page for IO to complete.
+	 */
+	wait_on_page_locked(page);
 	if (PageUptodate(page))
 		goto out;
 
+	/* Distinguish between all the cases under the safety of the lock */
 	lock_page(page);
+
+	/* Case c or d, restart the operation */
 	if (!page->mapping) {
 		unlock_page(page);
 		page_cache_release(page);
-		goto retry;
+		goto repeat;
 	}
+
+	/* Someone else locked and filled the page in a very small window */
 	if (PageUptodate(page)) {
 		unlock_page(page);
 		goto out;
 	}
-	err = filler(data, page);
-	if (err < 0) {
-		page_cache_release(page);
-		return ERR_PTR(err);
-	} else {
-		page = wait_on_page_read(page);
-		if (IS_ERR(page))
-			return page;
-	}
+	goto filler;
+
 out:
 	mark_page_accessed(page);
 	return page;