Change XFS to support sector aligned O_DIRECT rather than filesystem block

alignment.

fs/xfs/linux/xfs_aops.c

@@ -903,7 +903,7 @@ linvfs_get_block_core(
 	 */
 	if (create &&
 	    ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
-	     (offset >= inode->i_size))) {
+	     (offset >= inode->i_size) || (pbmap.pbm_flags & PBMF_NEW))) {
 		set_buffer_new(bh_result);
 	}
 
@@ -968,10 +968,18 @@ linvfs_direct_IO(
 	loff_t			offset,
 	unsigned long		nr_segs)
 {
-	struct file *file = iocb->ki_filp;
-	struct inode *inode = file->f_dentry->d_inode->i_mapping->host;
+	struct file	*file = iocb->ki_filp;
+	struct inode	*inode = file->f_dentry->d_inode->i_mapping->host;
+	vnode_t		*vp = LINVFS_GET_VP(inode);
+	page_buf_bmap_t	pbmap;
+	int		maps = 1;
+	int		error;
+
+	VOP_BMAP(vp, offset, 0, BMAP_DEVICE, &pbmap, &maps, error);
+	if (error)
+		return -error;
 
-        return blockdev_direct_IO(rw, iocb, inode, NULL,
+        return blockdev_direct_IO(rw, iocb, inode, pbmap.pbm_target->pbr_bdev,
 		iov, offset, nr_segs,
 		linvfs_get_blocks_direct,
 		linvfs_unwritten_convert_direct);

fs/xfs/linux/xfs_ioctl.c

@@ -624,15 +624,11 @@ xfs_ioctl(
 
 	case XFS_IOC_DIOINFO: {
 		struct dioattr	da;
+		pb_target_t	*target =
+			(ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
+			mp->m_rtdev_targp : mp->m_ddev_targp;
 
-
-		/*
-		 * this only really needs to be BBSIZE.
-		 * it is set to the file system block size to
-		 * avoid having to do block zeroing on short writes.
-		 */
-		da.d_miniosz = mp->m_sb.sb_blocksize;
-		da.d_mem = mp->m_sb.sb_blocksize;
+		da.d_mem = da.d_miniosz = 1 << target->pbr_sshift;
 		/* The size dio will do in one go */
 		da.d_maxiosz = 64 * PAGE_CACHE_SIZE;
 

fs/xfs/linux/xfs_iomap.c

@@ -77,6 +77,7 @@ STATIC int
 _xfs_imap_to_bmap(
 	xfs_iocore_t	*io,
 	xfs_off_t	offset,
+	int		new,
 	xfs_bmbt_irec_t *imap,
 	page_buf_bmap_t	*pbmapp,
 	int		imaps,			/* Number of imap entries */
@@ -117,6 +118,9 @@ _xfs_imap_to_bmap(
 			pbmapp->pbm_flags |= PBMF_EOF;
 		}
 
+		if (new)
+			pbmapp->pbm_flags |= PBMF_NEW;
+
 		offset += pbmapp->pbm_bsize - pbmapp->pbm_delta;
 	}
 	return pbm;	/* Return the number filled */
@@ -134,16 +138,18 @@ xfs_iomap(
 	xfs_mount_t	*mp = io->io_mount;
 	xfs_fileoff_t	offset_fsb, end_fsb;
 	int		error = 0;
+	int		new = 0;
 	int		lockmode = 0;
 	xfs_bmbt_irec_t	imap;
 	int		nimaps = 1;
 	int		bmap_flags = 0;
 
 	if (XFS_FORCED_SHUTDOWN(mp))
-		return XFS_ERROR(EIO);
+		return -XFS_ERROR(EIO);
 
 	switch (flags &
-		(BMAP_READ|BMAP_WRITE|BMAP_ALLOCATE|BMAP_UNWRITTEN)) {
+		(BMAP_READ | BMAP_WRITE | BMAP_ALLOCATE |
+		 BMAP_UNWRITTEN | BMAP_DEVICE)) {
 	case BMAP_READ:
 		lockmode = XFS_LCK_MAP_SHARED(mp, io);
 		bmap_flags = XFS_BMAPI_ENTIRE;
@@ -168,6 +174,13 @@ xfs_iomap(
 		break;
 	case BMAP_UNWRITTEN:
 		goto phase2;
+	case BMAP_DEVICE:
+		lockmode = XFS_LCK_MAP_SHARED(mp, io);
+		pbmapp->pbm_target = io->io_flags & XFS_IOCORE_RT ?
+			mp->m_rtdev_targp : mp->m_ddev_targp;
+		error = 0;
+		*npbmaps = 1;
+		goto out;
 	default:
 		BUG();
 	}
@@ -200,6 +213,7 @@ xfs_iomap(
 			error = XFS_IOMAP_WRITE_DELAY(mp, io, offset, count,
 					flags, &imap, &nimaps);
 		}
+		new = 1;
 		break;
 	case BMAP_ALLOCATE:
 		/* If we found an extent, return it */
@@ -219,8 +233,8 @@ xfs_iomap(
 	}
 
 	if (nimaps) {
-		*npbmaps = _xfs_imap_to_bmap(io, offset, &imap,
-						pbmapp, nimaps, *npbmaps);
+		*npbmaps = _xfs_imap_to_bmap(io, offset, new, &imap,
+					pbmapp, nimaps, *npbmaps);
 	} else if (npbmaps) {
 		*npbmaps = 0;
 	}

fs/xfs/linux/xfs_lrw.c

@@ -184,8 +184,11 @@ xfs_read(
 	/* END copy & waste from filemap.c */
 
 	if (direct) {
-		if ((*offset & mp->m_blockmask) ||
-		    (size & mp->m_blockmask)) {
+		pb_target_t	*target =
+			(ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
+				mp->m_rtdev_targp : mp->m_ddev_targp;
+		if ((*offset & target->pbr_smask) ||
+		    (size & target->pbr_smask)) {
 			if (*offset == ip->i_d.di_size) {
 				return (0);
 			}
@@ -566,8 +569,12 @@ xfs_write(
 	}
 
 	if (direct) {
-		if ((*offset & mp->m_blockmask) ||
-		    (size & mp->m_blockmask)) {
+		pb_target_t	*target =
+			(xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
+				mp->m_rtdev_targp : mp->m_ddev_targp;
+
+		if ((*offset & target->pbr_smask) ||
+		    (size & target->pbr_smask)) {
 			return XFS_ERROR(-EINVAL);
 		}
 		iolock = XFS_IOLOCK_SHARED;

fs/xfs/pagebuf/page_buf.h

@@ -79,8 +79,9 @@ typedef enum {				/* pbm_flags values */
 	PBMF_EOF =		0x01,	/* mapping contains EOF		*/
 	PBMF_HOLE =		0x02,	/* mapping covers a hole	*/
 	PBMF_DELAY =		0x04,	/* mapping covers delalloc region  */
-	PBMF_UNWRITTEN =	0x20	/* mapping covers allocated	*/
+	PBMF_UNWRITTEN =	0x20,	/* mapping covers allocated	*/
 					/* but uninitialized file data	*/
+	PBMF_NEW =		0x40	/* just allocated		*/
 } bmap_flags_t;
 
 typedef enum {
@@ -95,6 +96,7 @@ typedef enum {
 	BMAP_MMAP = (1 << 6),		/* allocate for mmap write */
 	BMAP_SYNC = (1 << 7),		/* sync write */
 	BMAP_TRYLOCK = (1 << 8),	/* non-blocking request */
+	BMAP_DEVICE = (1 << 9),		/* we only want to know the device */
 } bmapi_flags_t;
 
 typedef enum page_buf_flags_e {		/* pb_flags values */