Merge branch 'xfs-dio-fix-4.6' into for-next

9 files changed, 143 insertions, 208 deletions

diff --git a/fs/dax.c b/fs/dax.c
index 4fd6b0c5c6b5..e38b2c589b54 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -267,8 +267,13 @@ ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
 	if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 		inode_unlock(inode);
 
-	if ((retval > 0) && end_io)
-		end_io(iocb, pos, retval, bh.b_private);
+	if (end_io) {
+		int err;
+
+		err = end_io(iocb, pos, retval, bh.b_private);
+		if (err)
+			retval = err;
+	}
 
 	if (!(flags & DIO_SKIP_DIO_COUNT))
 		inode_dio_end(inode);
diff --git a/fs/direct-io.c b/fs/direct-io.c
index 1b2f7ffc8b84..9c6f885cc518 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -253,8 +253,13 @@ static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret,
 	if (ret == 0)
 		ret = transferred;
 
-	if (dio->end_io && dio->result)
-		dio->end_io(dio->iocb, offset, transferred, dio->private);
+	if (dio->end_io) {
+		int err;
+
+		err = dio->end_io(dio->iocb, offset, ret, dio->private);
+		if (err)
+			ret = err;
+	}
 
 	if (!(dio->flags & DIO_SKIP_DIO_COUNT))
 		inode_dio_end(dio->inode);
diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
index 0662b285dc8a..56c12df107ab 100644
--- a/fs/ext4/ext4.h
+++ b/fs/ext4/ext4.h
@@ -1504,15 +1504,6 @@ static inline int ext4_valid_inum(struct super_block *sb, unsigned long ino)
 		 ino <= le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count));
 }
 
-static inline void ext4_set_io_unwritten_flag(struct inode *inode,
-					      struct ext4_io_end *io_end)
-{
-	if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
-		io_end->flag |= EXT4_IO_END_UNWRITTEN;
-		atomic_inc(&EXT4_I(inode)->i_unwritten);
-	}
-}
-
 static inline ext4_io_end_t *ext4_inode_aio(struct inode *inode)
 {
 	return inode->i_private;
@@ -3293,6 +3284,27 @@ extern struct mutex ext4__aio_mutex[EXT4_WQ_HASH_SZ];
 extern int ext4_resize_begin(struct super_block *sb);
 extern void ext4_resize_end(struct super_block *sb);
 
+static inline void ext4_set_io_unwritten_flag(struct inode *inode,
+					      struct ext4_io_end *io_end)
+{
+	if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
+		io_end->flag |= EXT4_IO_END_UNWRITTEN;
+		atomic_inc(&EXT4_I(inode)->i_unwritten);
+	}
+}
+
+static inline void ext4_clear_io_unwritten_flag(ext4_io_end_t *io_end)
+{
+	struct inode *inode = io_end->inode;
+
+	if (io_end->flag & EXT4_IO_END_UNWRITTEN) {
+		io_end->flag &= ~EXT4_IO_END_UNWRITTEN;
+		/* Wake up anyone waiting on unwritten extent conversion */
+		if (atomic_dec_and_test(&EXT4_I(inode)->i_unwritten))
+			wake_up_all(ext4_ioend_wq(inode));
+	}
+}
+
 #endif	/* __KERNEL__ */
 
 #define EFSBADCRC	EBADMSG		/* Bad CRC detected */
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 83bc8bfb3bea..2b98171a9432 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -3161,14 +3161,14 @@ out:
 }
 #endif
 
-static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
+static int ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
 			    ssize_t size, void *private)
 {
         ext4_io_end_t *io_end = iocb->private;
 
 	/* if not async direct IO just return */
 	if (!io_end)
-		return;
+		return 0;
 
 	ext_debug("ext4_end_io_dio(): io_end 0x%p "
 		  "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
@@ -3176,9 +3176,19 @@ static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
 		  size);
 
 	iocb->private = NULL;
+	/*
+	 * Error during AIO DIO. We cannot convert unwritten extents as the
+	 * data was not written. Just clear the unwritten flag and drop io_end.
+	 */
+	if (size <= 0) {
+		ext4_clear_io_unwritten_flag(io_end);
+		size = 0;
+	}
 	io_end->offset = offset;
 	io_end->size = size;
 	ext4_put_io_end(io_end);
+
+	return 0;
 }
 
 /*
@@ -3301,16 +3311,6 @@ static ssize_t ext4_ext_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
 	if (io_end) {
 		ext4_inode_aio_set(inode, NULL);
 		ext4_put_io_end(io_end);
-		/*
-		 * When no IO was submitted ext4_end_io_dio() was not
-		 * called so we have to put iocb's reference.
-		 */
-		if (ret <= 0 && ret != -EIOCBQUEUED && iocb->private) {
-			WARN_ON(iocb->private != io_end);
-			WARN_ON(io_end->flag & EXT4_IO_END_UNWRITTEN);
-			ext4_put_io_end(io_end);
-			iocb->private = NULL;
-		}
 	}
 	if (ret > 0 && !overwrite && ext4_test_inode_state(inode,
 						EXT4_STATE_DIO_UNWRITTEN)) {
diff --git a/fs/ext4/page-io.c b/fs/ext4/page-io.c
index 090b3498638e..f49a87c4fb63 100644
--- a/fs/ext4/page-io.c
+++ b/fs/ext4/page-io.c
@@ -139,16 +139,6 @@ static void ext4_release_io_end(ext4_io_end_t *io_end)
 	kmem_cache_free(io_end_cachep, io_end);
 }
 
-static void ext4_clear_io_unwritten_flag(ext4_io_end_t *io_end)
-{
-	struct inode *inode = io_end->inode;
-
-	io_end->flag &= ~EXT4_IO_END_UNWRITTEN;
-	/* Wake up anyone waiting on unwritten extent conversion */
-	if (atomic_dec_and_test(&EXT4_I(inode)->i_unwritten))
-		wake_up_all(ext4_ioend_wq(inode));
-}
-
 /*
  * Check a range of space and convert unwritten extents to written. Note that
  * we are protected from truncate touching same part of extent tree by the
diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c
index 794fd1587f34..5dcc5f5a842e 100644
--- a/fs/ocfs2/aops.c
+++ b/fs/ocfs2/aops.c
@@ -620,7 +620,7 @@ bail:
  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
  * to protect io on one node from truncation on another.
  */
-static void ocfs2_dio_end_io(struct kiocb *iocb,
+static int ocfs2_dio_end_io(struct kiocb *iocb,
 			     loff_t offset,
 			     ssize_t bytes,
 			     void *private)
@@ -628,6 +628,9 @@ static void ocfs2_dio_end_io(struct kiocb *iocb,
 	struct inode *inode = file_inode(iocb->ki_filp);
 	int level;
 
+	if (bytes <= 0)
+		return 0;
+
 	/* this io's submitter should not have unlocked this before we could */
 	BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 
@@ -644,6 +647,8 @@ static void ocfs2_dio_end_io(struct kiocb *iocb,
 		level = ocfs2_iocb_rw_locked_level(iocb);
 		ocfs2_rw_unlock(inode, level);
 	}
+
+	return 0;
 }
 
 static int ocfs2_releasepage(struct page *page, gfp_t wait)
diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
index 379c089fb051..8163910bc49f 100644
--- a/fs/xfs/xfs_aops.c
+++ b/fs/xfs/xfs_aops.c
@@ -36,6 +36,10 @@
 #include <linux/pagevec.h>
 #include <linux/writeback.h>
 
+/* flags for direct write completions */
+#define XFS_DIO_FLAG_UNWRITTEN	(1 << 0)
+#define XFS_DIO_FLAG_APPEND	(1 << 1)
+
 void
 xfs_count_page_state(
 	struct page		*page,
@@ -1238,27 +1242,8 @@ xfs_vm_releasepage(
 }
 
 /*
- * When we map a DIO buffer, we may need to attach an ioend that describes the
- * type of write IO we are doing. This passes to the completion function the
- * operations it needs to perform. If the mapping is for an overwrite wholly
- * within the EOF then we don't need an ioend and so we don't allocate one.
- * This avoids the unnecessary overhead of allocating and freeing ioends for
- * workloads that don't require transactions on IO completion.
- *
- * If we get multiple mappings in a single IO, we might be mapping different
- * types. But because the direct IO can only have a single private pointer, we
- * need to ensure that:
- *
- * a) i) the ioend spans the entire region of unwritten mappings; or
- *    ii) the ioend spans all the mappings that cross or are beyond EOF; and
- * b) if it contains unwritten extents, it is *permanently* marked as such
- *
- * We could do this by chaining ioends like buffered IO does, but we only
- * actually get one IO completion callback from the direct IO, and that spans
- * the entire IO regardless of how many mappings and IOs are needed to complete
- * the DIO. There is only going to be one reference to the ioend and its life
- * cycle is constrained by the DIO completion code. hence we don't need
- * reference counting here.
+ * When we map a DIO buffer, we may need to pass flags to
+ * xfs_end_io_direct_write to tell it what kind of write IO we are doing.
  *
  * Note that for DIO, an IO to the highest supported file block offset (i.e.
  * 2^63 - 1FSB bytes) will result in the offset + count overflowing a signed 64
@@ -1266,68 +1251,26 @@ xfs_vm_releasepage(
  * extending the file size. We won't know for sure until IO completion is run
  * and the actual max write offset is communicated to the IO completion
  * routine.
- *
- * For DAX page faults, we are preparing to never see unwritten extents here,
- * nor should we ever extend the inode size. Hence we will soon have nothing to
- * do here for this case, ensuring we don't have to provide an IO completion
- * callback to free an ioend that we don't actually need for a fault into the
- * page at offset (2^63 - 1FSB) bytes.
  */
-
 static void
 xfs_map_direct(
 	struct inode		*inode,
 	struct buffer_head	*bh_result,
 	struct xfs_bmbt_irec	*imap,
-	xfs_off_t		offset,
-	bool			dax_fault)
+	xfs_off_t		offset)
 {
-	struct xfs_ioend	*ioend;
+	uintptr_t		*flags = (uintptr_t *)&bh_result->b_private;
 	xfs_off_t		size = bh_result->b_size;
-	int			type;
-
-	if (ISUNWRITTEN(imap))
-		type = XFS_IO_UNWRITTEN;
-	else
-		type = XFS_IO_OVERWRITE;
-
-	trace_xfs_gbmap_direct(XFS_I(inode), offset, size, type, imap);
 
-	if (dax_fault) {
-		ASSERT(type == XFS_IO_OVERWRITE);
-		trace_xfs_gbmap_direct_none(XFS_I(inode), offset, size, type,
-					    imap);
-		return;
-	}
-
-	if (bh_result->b_private) {
-		ioend = bh_result->b_private;
-		ASSERT(ioend->io_size > 0);
-		ASSERT(offset >= ioend->io_offset);
-		if (offset + size > ioend->io_offset + ioend->io_size)
-			ioend->io_size = offset - ioend->io_offset + size;
-
-		if (type == XFS_IO_UNWRITTEN && type != ioend->io_type)
-			ioend->io_type = XFS_IO_UNWRITTEN;
-
-		trace_xfs_gbmap_direct_update(XFS_I(inode), ioend->io_offset,
-					      ioend->io_size, ioend->io_type,
-					      imap);
-	} else if (type == XFS_IO_UNWRITTEN ||
-		   offset + size > i_size_read(inode) ||
-		   offset + size < 0) {
-		ioend = xfs_alloc_ioend(inode, type);
-		ioend->io_offset = offset;
-		ioend->io_size = size;
+	trace_xfs_get_blocks_map_direct(XFS_I(inode), offset, size,
+		ISUNWRITTEN(imap) ? XFS_IO_UNWRITTEN : XFS_IO_OVERWRITE, imap);
 
-		bh_result->b_private = ioend;
+	if (ISUNWRITTEN(imap)) {
+		*flags |= XFS_DIO_FLAG_UNWRITTEN;
+		set_buffer_defer_completion(bh_result);
+	} else if (offset + size > i_size_read(inode) || offset + size < 0) {
+		*flags |= XFS_DIO_FLAG_APPEND;
 		set_buffer_defer_completion(bh_result);
-
-		trace_xfs_gbmap_direct_new(XFS_I(inode), offset, size, type,
-					   imap);
-	} else {
-		trace_xfs_gbmap_direct_none(XFS_I(inode), offset, size, type,
-					    imap);
 	}
 }
 
@@ -1498,9 +1441,12 @@ __xfs_get_blocks(
 		if (ISUNWRITTEN(&imap))
 			set_buffer_unwritten(bh_result);
 		/* direct IO needs special help */
-		if (create && direct)
-			xfs_map_direct(inode, bh_result, &imap, offset,
-				       dax_fault);
+		if (create && direct) {
+			if (dax_fault)
+				ASSERT(!ISUNWRITTEN(&imap));
+			else
+				xfs_map_direct(inode, bh_result, &imap, offset);
+		}
 	}
 
 	/*
@@ -1570,42 +1516,50 @@ xfs_get_blocks_dax_fault(
 	return __xfs_get_blocks(inode, iblock, bh_result, create, true, true);
 }
 
-static void
-__xfs_end_io_direct_write(
-	struct inode		*inode,
-	struct xfs_ioend	*ioend,
+/*
+ * Complete a direct I/O write request.
+ *
+ * xfs_map_direct passes us some flags in the private data to tell us what to
+ * do.  If no flags are set, then the write IO is an overwrite wholly within
+ * the existing allocated file size and so there is nothing for us to do.
+ *
+ * Note that in this case the completion can be called in interrupt context,
+ * whereas if we have flags set we will always be called in task context
+ * (i.e. from a workqueue).
+ */
+STATIC int
+xfs_end_io_direct_write(
+	struct kiocb		*iocb,
 	loff_t			offset,
-	ssize_t			size)
+	ssize_t			size,
+	void			*private)
 {
-	struct xfs_mount	*mp = XFS_I(inode)->i_mount;
+	struct inode		*inode = file_inode(iocb->ki_filp);
+	struct xfs_inode	*ip = XFS_I(inode);
+	struct xfs_mount	*mp = ip->i_mount;
+	uintptr_t		flags = (uintptr_t)private;
+	int			error = 0;
 
-	if (XFS_FORCED_SHUTDOWN(mp) || ioend->io_error)
-		goto out_end_io;
+	trace_xfs_end_io_direct_write(ip, offset, size);
 
-	/*
-	 * dio completion end_io functions are only called on writes if more
-	 * than 0 bytes was written.
-	 */
-	ASSERT(size > 0);
+	if (XFS_FORCED_SHUTDOWN(mp))
+		return -EIO;
 
-	/*
-	 * The ioend only maps whole blocks, while the IO may be sector aligned.
-	 * Hence the ioend offset/size may not match the IO offset/size exactly.
-	 * Because we don't map overwrites within EOF into the ioend, the offset
-	 * may not match, but only if the endio spans EOF.  Either way, write
-	 * the IO sizes into the ioend so that completion processing does the
-	 * right thing.
-	 */
-	ASSERT(offset + size <= ioend->io_offset + ioend->io_size);
-	ioend->io_size = size;
-	ioend->io_offset = offset;
+	if (size <= 0)
+		return size;
 
 	/*
-	 * The ioend tells us whether we are doing unwritten extent conversion
+	 * The flags tell us whether we are doing unwritten extent conversions
 	 * or an append transaction that updates the on-disk file size. These
 	 * cases are the only cases where we should *potentially* be needing
 	 * to update the VFS inode size.
-	 *
+	 */
+	if (flags == 0) {
+		ASSERT(offset + size <= i_size_read(inode));
+		return 0;
+	}
+
+	/*
 	 * We need to update the in-core inode size here so that we don't end up
 	 * with the on-disk inode size being outside the in-core inode size. We
 	 * have no other method of updating EOF for AIO, so always do it here
@@ -1616,91 +1570,56 @@ __xfs_end_io_direct_write(
 	 * here can result in EOF moving backwards and Bad Things Happen when
 	 * that occurs.
 	 */
-	spin_lock(&XFS_I(inode)->i_flags_lock);
+	spin_lock(&ip->i_flags_lock);
 	if (offset + size > i_size_read(inode))
 		i_size_write(inode, offset + size);
-	spin_unlock(&XFS_I(inode)->i_flags_lock);
+	spin_unlock(&ip->i_flags_lock);
 
-	/*
-	 * If we are doing an append IO that needs to update the EOF on disk,
-	 * do the transaction reserve now so we can use common end io
-	 * processing. Stashing the error (if there is one) in the ioend will
-	 * result in the ioend processing passing on the error if it is
-	 * possible as we can't return it from here.
-	 */
-	if (ioend->io_type == XFS_IO_OVERWRITE)
-		ioend->io_error = xfs_setfilesize_trans_alloc(ioend);
+	if (flags & XFS_DIO_FLAG_UNWRITTEN) {
+		trace_xfs_end_io_direct_write_unwritten(ip, offset, size);
 
-out_end_io:
-	xfs_end_io(&ioend->io_work);
-	return;
-}
+		error = xfs_iomap_write_unwritten(ip, offset, size);
+	} else if (flags & XFS_DIO_FLAG_APPEND) {
+		struct xfs_trans *tp;
 
-/*
- * Complete a direct I/O write request.
- *
- * The ioend structure is passed from __xfs_get_blocks() to tell us what to do.
- * If no ioend exists (i.e. @private == NULL) then the write IO is an overwrite
- * wholly within the EOF and so there is nothing for us to do. Note that in this
- * case the completion can be called in interrupt context, whereas if we have an
- * ioend we will always be called in task context (i.e. from a workqueue).
- */
-STATIC void
-xfs_end_io_direct_write(
-	struct kiocb		*iocb,
-	loff_t			offset,
-	ssize_t			size,
-	void			*private)
-{
-	struct inode		*inode = file_inode(iocb->ki_filp);
-	struct xfs_ioend	*ioend = private;
+		trace_xfs_end_io_direct_write_append(ip, offset, size);
 
-	trace_xfs_gbmap_direct_endio(XFS_I(inode), offset, size,
-				     ioend ? ioend->io_type : 0, NULL);
-
-	if (!ioend) {
-		ASSERT(offset + size <= i_size_read(inode));
-		return;
+		tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS);
+		error = xfs_trans_reserve(tp, &M_RES(mp)->tr_fsyncts, 0, 0);
+		if (error) {
+			xfs_trans_cancel(tp);
+			return error;
+		}
+		error = xfs_setfilesize(ip, tp, offset, size);
 	}
 
-	__xfs_end_io_direct_write(inode, ioend, offset, size);
+	return error;
 }
 
-static inline ssize_t
-xfs_vm_do_dio(
-	struct inode		*inode,
+STATIC ssize_t
+xfs_vm_direct_IO(
 	struct kiocb		*iocb,
 	struct iov_iter		*iter,
-	loff_t			offset,
-	void			(*endio)(struct kiocb	*iocb,
-					 loff_t		offset,
-					 ssize_t	size,
-					 void		*private),
-	int			flags)
+	loff_t			offset)
 {
+	struct inode		*inode = iocb->ki_filp->f_mapping->host;
+	dio_iodone_t		*endio = NULL;
+	int			flags = 0;
 	struct block_device	*bdev;
 
-	if (IS_DAX(inode))
+	if (iov_iter_rw(iter) == WRITE) {
+		endio = xfs_end_io_direct_write;
+		flags = DIO_ASYNC_EXTEND;
+	}
+
+	if (IS_DAX(inode)) {
 		return dax_do_io(iocb, inode, iter, offset,
 				 xfs_get_blocks_direct, endio, 0);
+	}
 
 	bdev = xfs_find_bdev_for_inode(inode);
 	return  __blockdev_direct_IO(iocb, inode, bdev, iter, offset,
-				     xfs_get_blocks_direct, endio, NULL, flags);
-}
-
-STATIC ssize_t
-xfs_vm_direct_IO(
-	struct kiocb		*iocb,
-	struct iov_iter		*iter,
-	loff_t			offset)
-{
-	struct inode		*inode = iocb->ki_filp->f_mapping->host;
-
-	if (iov_iter_rw(iter) == WRITE)
-		return xfs_vm_do_dio(inode, iocb, iter, offset,
-				     xfs_end_io_direct_write, DIO_ASYNC_EXTEND);
-	return xfs_vm_do_dio(inode, iocb, iter, offset, NULL, 0);
+			xfs_get_blocks_direct, endio, NULL, flags);
 }
 
 /*
diff --git a/fs/xfs/xfs_trace.h b/fs/xfs/xfs_trace.h
index 391d797cb53f..c8d58426008e 100644
--- a/fs/xfs/xfs_trace.h
+++ b/fs/xfs/xfs_trace.h
@@ -1296,11 +1296,7 @@ DEFINE_IOMAP_EVENT(xfs_map_blocks_found);
 DEFINE_IOMAP_EVENT(xfs_map_blocks_alloc);
 DEFINE_IOMAP_EVENT(xfs_get_blocks_found);
 DEFINE_IOMAP_EVENT(xfs_get_blocks_alloc);
-DEFINE_IOMAP_EVENT(xfs_gbmap_direct);
-DEFINE_IOMAP_EVENT(xfs_gbmap_direct_new);
-DEFINE_IOMAP_EVENT(xfs_gbmap_direct_update);
-DEFINE_IOMAP_EVENT(xfs_gbmap_direct_none);
-DEFINE_IOMAP_EVENT(xfs_gbmap_direct_endio);
+DEFINE_IOMAP_EVENT(xfs_get_blocks_map_direct);
 
 DECLARE_EVENT_CLASS(xfs_simple_io_class,
 	TP_PROTO(struct xfs_inode *ip, xfs_off_t offset, ssize_t count),
@@ -1340,6 +1336,9 @@ DEFINE_SIMPLE_IO_EVENT(xfs_unwritten_convert);
 DEFINE_SIMPLE_IO_EVENT(xfs_get_blocks_notfound);
 DEFINE_SIMPLE_IO_EVENT(xfs_setfilesize);
 DEFINE_SIMPLE_IO_EVENT(xfs_zero_eof);
+DEFINE_SIMPLE_IO_EVENT(xfs_end_io_direct_write);
+DEFINE_SIMPLE_IO_EVENT(xfs_end_io_direct_write_unwritten);
+DEFINE_SIMPLE_IO_EVENT(xfs_end_io_direct_write_append);
 
 DECLARE_EVENT_CLASS(xfs_itrunc_class,
 	TP_PROTO(struct xfs_inode *ip, xfs_fsize_t new_size),
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 1a2046275cdf..d7f37bfcbdce 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -70,7 +70,7 @@ extern int sysctl_protected_hardlinks;
 struct buffer_head;
 typedef int (get_block_t)(struct inode *inode, sector_t iblock,
 			struct buffer_head *bh_result, int create);
-typedef void (dio_iodone_t)(struct kiocb *iocb, loff_t offset,
+typedef int (dio_iodone_t)(struct kiocb *iocb, loff_t offset,
 			ssize_t bytes, void *private);
 typedef void (dax_iodone_t)(struct buffer_head *bh_map, int uptodate);