[XFS] Fix memory corruption with small buffer reads

When we have multiple buffers in a single page for a blocksize == pagesize
filesystem we might overwrite the page contents if two callers hit it
shortly after each other. To prevent that we need to keep the page locked
until I/O is completed and the page marked uptodate.

Thanks to Eric Sandeen for triaging this bug and finding a reproducible
testcase and Dave Chinner for additional advice.

This should fix kernel.org bz #10421.
Tested-by: Eric Sandeen <sandeen@sandeen.net>

SGI-PV: 981813
SGI-Modid: xfs-linux-melb:xfs-kern:31173a
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>

fs/xfs/linux-2.6/xfs_buf.c

@@ -387,6 +387,8 @@ _xfs_buf_lookup_pages(
 		if (unlikely(page == NULL)) {
 			if (flags & XBF_READ_AHEAD) {
 				bp->b_page_count = i;
+				for (i = 0; i < bp->b_page_count; i++)
+					unlock_page(bp->b_pages[i]);
 				return -ENOMEM;
 			}
 
@@ -416,17 +418,24 @@ _xfs_buf_lookup_pages(
 		ASSERT(!PagePrivate(page));
 		if (!PageUptodate(page)) {
 			page_count--;
-			if (blocksize < PAGE_CACHE_SIZE && !PagePrivate(page)) {
+			if (blocksize >= PAGE_CACHE_SIZE) {
+				if (flags & XBF_READ)
+					bp->b_flags |= _XBF_PAGE_LOCKED;
+			} else if (!PagePrivate(page)) {
 				if (test_page_region(page, offset, nbytes))
 					page_count++;
 			}
 		}
 
-		unlock_page(page);
 		bp->b_pages[i] = page;
 		offset = 0;
 	}
 
+	if (!(bp->b_flags & _XBF_PAGE_LOCKED)) {
+		for (i = 0; i < bp->b_page_count; i++)
+			unlock_page(bp->b_pages[i]);
+	}
+
 	if (page_count == bp->b_page_count)
 		bp->b_flags |= XBF_DONE;
 
@@ -746,6 +755,7 @@ xfs_buf_associate_memory(
 	bp->b_count_desired = len;
 	bp->b_buffer_length = buflen;
 	bp->b_flags |= XBF_MAPPED;
+	bp->b_flags &= ~_XBF_PAGE_LOCKED;
 
 	return 0;
 }
@@ -1093,8 +1103,10 @@ _xfs_buf_ioend(
 	xfs_buf_t		*bp,
 	int			schedule)
 {
-	if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
+	if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
+		bp->b_flags &= ~_XBF_PAGE_LOCKED;
 		xfs_buf_ioend(bp, schedule);
+	}
 }
 
 STATIC void
@@ -1125,6 +1137,9 @@ xfs_buf_bio_end_io(
 
 		if (--bvec >= bio->bi_io_vec)
 			prefetchw(&bvec->bv_page->flags);
+
+		if (bp->b_flags & _XBF_PAGE_LOCKED)
+			unlock_page(page);
 	} while (bvec >= bio->bi_io_vec);
 
 	_xfs_buf_ioend(bp, 1);
@@ -1163,7 +1178,8 @@ _xfs_buf_ioapply(
 	 * filesystem block size is not smaller than the page size.
 	 */
 	if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
-	    (bp->b_flags & XBF_READ) &&
+	    ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
+	      (XBF_READ|_XBF_PAGE_LOCKED)) &&
 	    (blocksize >= PAGE_CACHE_SIZE)) {
 		bio = bio_alloc(GFP_NOIO, 1);
 

fs/xfs/linux-2.6/xfs_buf.h

@@ -66,6 +66,25 @@ typedef enum {
 	_XBF_PAGES = (1 << 18),	    /* backed by refcounted pages	   */
 	_XBF_RUN_QUEUES = (1 << 19),/* run block device task queue	   */
 	_XBF_DELWRI_Q = (1 << 21),   /* buffer on delwri queue		   */
+
+	/*
+	 * Special flag for supporting metadata blocks smaller than a FSB.
+	 *
+	 * In this case we can have multiple xfs_buf_t on a single page and
+	 * need to lock out concurrent xfs_buf_t readers as they only
+	 * serialise access to the buffer.
+	 *
+	 * If the FSB size >= PAGE_CACHE_SIZE case, we have no serialisation
+	 * between reads of the page. Hence we can have one thread read the
+	 * page and modify it, but then race with another thread that thinks
+	 * the page is not up-to-date and hence reads it again.
+	 *
+	 * The result is that the first modifcation to the page is lost.
+	 * This sort of AGF/AGI reading race can happen when unlinking inodes
+	 * that require truncation and results in the AGI unlinked list
+	 * modifications being lost.
+	 */
+	_XBF_PAGE_LOCKED = (1 << 22),
 } xfs_buf_flags_t;
 
 typedef enum {