[XFS] Use the right offset when ensuring a delayed allocate conversion has covered the offset originally requested. Can cause data corruption when multiple processes are performing writeout on different areas of the same file. Quite difficult to hit though.

SGI Modid: xfs-linux:xfs-kern:22377a

Signed-off-by: Nathan Scott <nathans@sgi.com>
Signed-off-by: Christoph Hellwig <hch@sgi.com>
.

1 files changed, 39 insertions, 29 deletions

diff --git a/fs/xfs/linux-2.6/xfs_aops.c b/fs/xfs/linux-2.6/xfs_aops.c
index 1edd3b694332..9278e9aba9ba 100644
--- a/fs/xfs/linux-2.6/xfs_aops.c
+++ b/fs/xfs/linux-2.6/xfs_aops.c
@@ -558,7 +558,8 @@ xfs_submit_page(
 	int			i;
 
 	BUG_ON(PageWriteback(page));
-	set_page_writeback(page);
+	if (bh_count)
+		set_page_writeback(page);
 	if (clear_dirty)
 		clear_page_dirty(page);
 	unlock_page(page);
@@ -578,9 +579,6 @@ xfs_submit_page(
 
 		if (probed_page && clear_dirty)
 			wbc->nr_to_write--;	/* Wrote an "extra" page */
-	} else {
-		end_page_writeback(page);
-		wbc->pages_skipped++;	/* We didn't write this page */
 	}
 }
 
@@ -602,21 +600,26 @@ xfs_convert_page(
 {
 	struct buffer_head	*bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 	xfs_iomap_t		*mp = iomapp, *tmp;
-	unsigned long		end, offset;
-	pgoff_t			end_index;
-	int			i = 0, index = 0;
+	unsigned long		offset, end_offset;
+	int			index = 0;
 	int			bbits = inode->i_blkbits;
+	int			len, page_dirty;
 
-	end_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
-	if (page->index < end_index) {
-		end = PAGE_CACHE_SIZE;
-	} else {
-		end = i_size_read(inode) & (PAGE_CACHE_SIZE-1);
-	}
+	end_offset = (i_size_read(inode) & (PAGE_CACHE_SIZE - 1));
+
+	/*
+	 * page_dirty is initially a count of buffers on the page before
+	 * EOF and is decrememted as we move each into a cleanable state.
+	 */
+	len = 1 << inode->i_blkbits;
+	end_offset = max(end_offset, PAGE_CACHE_SIZE);
+	end_offset = roundup(end_offset, len);
+	page_dirty = end_offset / len;
+
+	offset = 0;
 	bh = head = page_buffers(page);
 	do {
-		offset = i << bbits;
-		if (offset >= end)
+		if (offset >= end_offset)
 			break;
 		if (!(PageUptodate(page) || buffer_uptodate(bh)))
 			continue;
@@ -625,6 +628,7 @@ xfs_convert_page(
 			if (startio) {
 				lock_buffer(bh);
 				bh_arr[index++] = bh;
+				page_dirty--;
 			}
 			continue;
 		}
@@ -657,10 +661,11 @@ xfs_convert_page(
 			unlock_buffer(bh);
 			mark_buffer_dirty(bh);
 		}
-	} while (i++, (bh = bh->b_this_page) != head);
+		page_dirty--;
+	} while (offset += len, (bh = bh->b_this_page) != head);
 
-	if (startio) {
-		xfs_submit_page(page, wbc, bh_arr, index, 1, index == i);
+	if (startio && index) {
+		xfs_submit_page(page, wbc, bh_arr, index, 1, !page_dirty);
 	} else {
 		unlock_page(page);
 	}
@@ -743,19 +748,22 @@ xfs_page_state_convert(
 		}
 	}
 
-	offset = (loff_t)page->index << PAGE_CACHE_SHIFT;
 	end_offset = min_t(unsigned long long,
-			offset + PAGE_CACHE_SIZE, i_size_read(inode));
-
-	bh = head = page_buffers(page);
-	iomp = NULL;
+			(loff_t)(page->index + 1) << PAGE_CACHE_SHIFT, offset);
+	offset = (loff_t)page->index << PAGE_CACHE_SHIFT;
 
 	/*
-	 * page_dirty is initially a count of buffers on the page and
-	 * is decrememted as we move each into a cleanable state.
+	 * page_dirty is initially a count of buffers on the page before
+	 * EOF and is decrememted as we move each into a cleanable state.
 	 */
-	len = bh->b_size;
-	page_dirty = PAGE_CACHE_SIZE / len;
+	len = 1 << inode->i_blkbits;
+	p_offset = max(p_offset, PAGE_CACHE_SIZE);
+	p_offset = roundup(p_offset, len);
+	page_dirty = p_offset / len;
+
+	iomp = NULL;
+	p_offset = 0;
+	bh = head = page_buffers(page);
 
 	do {
 		if (offset >= end_offset)
@@ -877,8 +885,10 @@ xfs_page_state_convert(
 	if (uptodate && bh == head)
 		SetPageUptodate(page);
 
-	if (startio)
-		xfs_submit_page(page, wbc, bh_arr, cnt, 0, 1);
+	if (startio) {
+		WARN_ON(page_dirty);
+		xfs_submit_page(page, wbc, bh_arr, cnt, 0, !page_dirty);
+	}
 
 	if (iomp) {
 		offset = (iomp->iomap_offset + iomp->iomap_bsize - 1) >>