Btrfs: Keep extent mappings in ram until pending ordered extents are done

It was possible for stale mappings from disk to be used instead of the
new pending ordered extent.  This adds a flag to the extent map struct
to keep it pinned until the pending ordered extent is actually on disk.

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

1 files changed, 15 insertions, 12 deletions

diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index 3153b4fdd0b3..d4a63ae7ed1b 100644
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -2000,7 +2000,7 @@ static int __extent_writepage(struct page *page, struct writeback_control *wbc,
 	struct block_device *bdev;
 	int ret;
 	int nr = 0;
-	size_t page_offset = 0;
+	size_t pg_offset = 0;
 	size_t blocksize;
 	loff_t i_size = i_size_read(inode);
 	unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
@@ -2008,9 +2008,9 @@ static int __extent_writepage(struct page *page, struct writeback_control *wbc,
 	u64 delalloc_end;
 
 	WARN_ON(!PageLocked(page));
-	page_offset = i_size & (PAGE_CACHE_SIZE - 1);
+	pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
 	if (page->index > end_index ||
-	   (page->index == end_index && !page_offset)) {
+	   (page->index == end_index && !pg_offset)) {
 		page->mapping->a_ops->invalidatepage(page, 0);
 		unlock_page(page);
 		return 0;
@@ -2020,12 +2020,12 @@ static int __extent_writepage(struct page *page, struct writeback_control *wbc,
 		char *userpage;
 
 		userpage = kmap_atomic(page, KM_USER0);
-		memset(userpage + page_offset, 0,
-		       PAGE_CACHE_SIZE - page_offset);
+		memset(userpage + pg_offset, 0,
+		       PAGE_CACHE_SIZE - pg_offset);
 		kunmap_atomic(userpage, KM_USER0);
 		flush_dcache_page(page);
 	}
-	page_offset = 0;
+	pg_offset = 0;
 
 	set_page_extent_mapped(page);
 
@@ -2088,7 +2088,7 @@ static int __extent_writepage(struct page *page, struct writeback_control *wbc,
 			unlock_start = page_end + 1;
 			break;
 		}
-		em = epd->get_extent(inode, page, page_offset, cur,
+		em = epd->get_extent(inode, page, pg_offset, cur,
 				     end - cur + 1, 1);
 		if (IS_ERR(em) || !em) {
 			SetPageError(page);
@@ -2113,12 +2113,13 @@ static int __extent_writepage(struct page *page, struct writeback_control *wbc,
 
 			unlock_extent(tree, unlock_start, cur + iosize -1,
 				      GFP_NOFS);
+
 			if (tree->ops && tree->ops->writepage_end_io_hook)
 				tree->ops->writepage_end_io_hook(page, cur,
 							 cur + iosize - 1,
 							 NULL, 1);
 			cur = cur + iosize;
-			page_offset += iosize;
+			pg_offset += iosize;
 			unlock_start = cur;
 			continue;
 		}
@@ -2127,7 +2128,7 @@ static int __extent_writepage(struct page *page, struct writeback_control *wbc,
 		if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
 				   EXTENT_DIRTY, 0)) {
 			cur = cur + iosize;
-			page_offset += iosize;
+			pg_offset += iosize;
 			continue;
 		}
 		clear_extent_dirty(tree, cur, cur + iosize - 1, GFP_NOFS);
@@ -2141,6 +2142,7 @@ static int __extent_writepage(struct page *page, struct writeback_control *wbc,
 			SetPageError(page);
 		} else {
 			unsigned long max_nr = end_index + 1;
+
 			set_range_writeback(tree, cur, cur + iosize - 1);
 			if (!PageWriteback(page)) {
 				printk("warning page %lu not writeback, "
@@ -2150,14 +2152,14 @@ static int __extent_writepage(struct page *page, struct writeback_control *wbc,
 			}
 
 			ret = submit_extent_page(WRITE, tree, page, sector,
-						 iosize, page_offset, bdev,
+						 iosize, pg_offset, bdev,
 						 &epd->bio, max_nr,
 						 end_bio_extent_writepage, 0);
 			if (ret)
 				SetPageError(page);
 		}
 		cur = cur + iosize;
-		page_offset += iosize;
+		pg_offset += iosize;
 		nr++;
 	}
 done:
@@ -2579,7 +2581,8 @@ int try_release_extent_mapping(struct extent_map_tree *map,
 				spin_unlock(&map->lock);
 				break;
 			}
-			if (em->start != start) {
+			if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
+			    em->start != start) {
 				spin_unlock(&map->lock);
 				free_extent_map(em);
 				break;