Btrfs: Don't drop extent_map cache during releasepage on the btree inode

The btree inode should only have a single extent_map in the cache,
it doesn't make sense to ever drop it.

Signed-off-by: Chris Mason <chris.mason@oracle.com>

fs/btrfs/disk-io.c

@@ -76,13 +76,12 @@ struct extent_map *btree_get_extent(struct inode *inode, struct page *page,
 	struct extent_map *em;
 	int ret;
 
-again:
 	spin_lock(&em_tree->lock);
 	em = lookup_extent_mapping(em_tree, start, len);
 	spin_unlock(&em_tree->lock);
-	if (em) {
+	if (em)
 		goto out;
-	}
+
 	em = alloc_extent_map(GFP_NOFS);
 	if (!em) {
 		em = ERR_PTR(-ENOMEM);
@@ -95,15 +94,21 @@ again:
 
 	spin_lock(&em_tree->lock);
 	ret = add_extent_mapping(em_tree, em);
-	spin_unlock(&em_tree->lock);
-
 	if (ret == -EEXIST) {
 		free_extent_map(em);
-		em = NULL;
-		goto again;
+		em = lookup_extent_mapping(em_tree, start, len);
+		if (em)
+			ret = 0;
+		else
+			ret = -EIO;
 	} else if (ret) {
-		em = ERR_PTR(ret);
+		free_extent_map(em);
+		em = NULL;
 	}
+	spin_unlock(&em_tree->lock);
+
+	if (ret)
+		em = ERR_PTR(ret);
 out:
 	return em;
 }
@@ -496,7 +501,7 @@ static int btree_releasepage(struct page *page, gfp_t gfp_flags)
 	}
 	tree = &BTRFS_I(page->mapping->host)->io_tree;
 	map = &BTRFS_I(page->mapping->host)->extent_tree;
-	ret = try_release_extent_mapping(map, tree, page, gfp_flags);
+	ret = try_release_extent_state(map, tree, page, gfp_flags);
 	if (ret == 1) {
 		invalidate_extent_lru(tree, page_offset(page), PAGE_CACHE_SIZE);
 		ClearPagePrivate(page);

fs/btrfs/extent_io.c

@@ -2463,6 +2463,31 @@ err:
 }
 EXPORT_SYMBOL(extent_prepare_write);
 
+/*
+ * a helper for releasepage, this tests for areas of the page that
+ * are locked or under IO and drops the related state bits if it is safe
+ * to drop the page.
+ */
+int try_release_extent_state(struct extent_map_tree *map,
+			     struct extent_io_tree *tree, struct page *page,
+			     gfp_t mask)
+{
+	u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
+	u64 end = start + PAGE_CACHE_SIZE - 1;
+	int ret = 1;
+
+	if (test_range_bit(tree, start, end, EXTENT_IOBITS, 0))
+		ret = 0;
+	else {
+		if ((mask & GFP_NOFS) == GFP_NOFS)
+			mask = GFP_NOFS;
+		clear_extent_bit(tree, start, end, EXTENT_UPTODATE,
+				 1, 1, mask);
+	}
+	return ret;
+}
+EXPORT_SYMBOL(try_release_extent_state);
+
 /*
  * a helper for releasepage.  As long as there are no locked extents
  * in the range corresponding to the page, both state records and extent
@@ -2475,8 +2500,7 @@ int try_release_extent_mapping(struct extent_map_tree *map,
 	struct extent_map *em;
 	u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
 	u64 end = start + PAGE_CACHE_SIZE - 1;
-	u64 orig_start = start;
-	int ret = 1;
+
 	if ((mask & __GFP_WAIT) &&
 	    page->mapping->host->i_size > 16 * 1024 * 1024) {
 		u64 len;
@@ -2507,15 +2531,7 @@ int try_release_extent_mapping(struct extent_map_tree *map,
 			free_extent_map(em);
 		}
 	}
-	if (test_range_bit(tree, orig_start, end, EXTENT_IOBITS, 0))
-		ret = 0;
-	else {
-		if ((mask & GFP_NOFS) == GFP_NOFS)
-			mask = GFP_NOFS;
-		clear_extent_bit(tree, orig_start, end, EXTENT_UPTODATE,
-				 1, 1, mask);
-	}
-	return ret;
+	return try_release_extent_state(map, tree, page, mask);
 }
 EXPORT_SYMBOL(try_release_extent_mapping);
 

fs/btrfs/extent_io.h

@@ -110,6 +110,9 @@ void extent_io_tree_empty_lru(struct extent_io_tree *tree);
 int try_release_extent_mapping(struct extent_map_tree *map,
 			       struct extent_io_tree *tree, struct page *page,
 			       gfp_t mask);
+int try_release_extent_state(struct extent_map_tree *map,
+			     struct extent_io_tree *tree, struct page *page,
+			     gfp_t mask);
 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask);
 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask);
 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,