reduce sub-block DIO serialisation v4

reduce sub-block DIO serialisation v4

[Christoph Hellwig]

This takes the approach from Dave, but adds a new flag instead of abusing
the nowait one, and keeps a simpler calling convention for iomap_dio_rw.

Changes since v3:
 - further comment improvements
 - micro-optimize an alignment check

Changes since v2:
 - add another sanity check in __iomap_dio_rw
 - rename IOMAP_DIO_UNALIGNED to IOMAP_DIO_OVERWRITE_ONLY
 - use more consistent parameter naming
 - improve a few comments
 - cleanup an if statement

Changes since v1:
 - rename the new flags
 - add an EOF check for subblock I/O
 - minor cleanups

[PATCH 01/11] xfs: factor out a xfs_ilock_iocb helper

[Christoph Hellwig]

Add a helper to factor out the nowait locking logical for the read/write
helpers.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
---
 fs/xfs/xfs_file.c | 55 +++++++++++++++++++++++++----------------------
 1 file changed, 29 insertions(+), 26 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index 5b0f93f738372d..c441cddfa4acbc 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -197,6 +197,23 @@ xfs_file_fsync(
 	return error;
 }
 
+static int
+xfs_ilock_iocb(
+	struct kiocb		*iocb,
+	unsigned int		lock_mode)
+{
+	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
+
+	if (iocb->ki_flags & IOCB_NOWAIT) {
+		if (!xfs_ilock_nowait(ip, lock_mode))
+			return -EAGAIN;
+	} else {
+		xfs_ilock(ip, lock_mode);
+	}
+
+	return 0;
+}
+
 STATIC ssize_t
 xfs_file_dio_aio_read(
 	struct kiocb		*iocb,
@@ -213,12 +230,9 @@ xfs_file_dio_aio_read(
 
 	file_accessed(iocb->ki_filp);
 
-	if (iocb->ki_flags & IOCB_NOWAIT) {
-		if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
-			return -EAGAIN;
-	} else {
-		xfs_ilock(ip, XFS_IOLOCK_SHARED);
-	}
+	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
+	if (ret)
+		return ret;
 	ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL,
 			is_sync_kiocb(iocb));
 	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
@@ -240,13 +254,9 @@ xfs_file_dax_read(
 	if (!count)
 		return 0; /* skip atime */
 
-	if (iocb->ki_flags & IOCB_NOWAIT) {
-		if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
-			return -EAGAIN;
-	} else {
-		xfs_ilock(ip, XFS_IOLOCK_SHARED);
-	}
-
+	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
+	if (ret)
+		return ret;
 	ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
 	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 
@@ -264,12 +274,9 @@ xfs_file_buffered_aio_read(
 
 	trace_xfs_file_buffered_read(ip, iov_iter_count(to), iocb->ki_pos);
 
-	if (iocb->ki_flags & IOCB_NOWAIT) {
-		if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
-			return -EAGAIN;
-	} else {
-		xfs_ilock(ip, XFS_IOLOCK_SHARED);
-	}
+	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
+	if (ret)
+		return ret;
 	ret = generic_file_read_iter(iocb, to);
 	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 
@@ -608,13 +615,9 @@ xfs_file_dax_write(
 	size_t			count;
 	loff_t			pos;
 
-	if (iocb->ki_flags & IOCB_NOWAIT) {
-		if (!xfs_ilock_nowait(ip, iolock))
-			return -EAGAIN;
-	} else {
-		xfs_ilock(ip, iolock);
-	}
-
+	ret = xfs_ilock_iocb(iocb, iolock);
+	if (ret)
+		return ret;
 	ret = xfs_file_aio_write_checks(iocb, from, &iolock);
 	if (ret)
 		goto out;
-- 
2.29.2

[PATCH 02/11] xfs: make xfs_file_aio_write_checks IOCB_NOWAIT-aware

[Christoph Hellwig]

Ensure we don't block on the iolock, or waiting for I/O in
xfs_file_aio_write_checks if the caller asked to avoid that.

Fixes: 29a5d29ec181 ("xfs: nowait aio support")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
---
 fs/xfs/xfs_file.c | 25 +++++++++++++++++++++----
 1 file changed, 21 insertions(+), 4 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index c441cddfa4acbc..fb4e6f2852bb8b 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -335,7 +335,14 @@ xfs_file_aio_write_checks(
 	if (error <= 0)
 		return error;
 
-	error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
+	if (iocb->ki_flags & IOCB_NOWAIT) {
+		error = break_layout(inode, false);
+		if (error == -EWOULDBLOCK)
+			error = -EAGAIN;
+	} else {
+		error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
+	}
+
 	if (error)
 		return error;
 
@@ -346,7 +353,11 @@ xfs_file_aio_write_checks(
 	if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
 		xfs_iunlock(ip, *iolock);
 		*iolock = XFS_IOLOCK_EXCL;
-		xfs_ilock(ip, *iolock);
+		error = xfs_ilock_iocb(iocb, *iolock);
+		if (error) {
+			*iolock = 0;
+			return error;
+		}
 		goto restart;
 	}
 	/*
@@ -368,6 +379,10 @@ xfs_file_aio_write_checks(
 	isize = i_size_read(inode);
 	if (iocb->ki_pos > isize) {
 		spin_unlock(&ip->i_flags_lock);
+
+		if (iocb->ki_flags & IOCB_NOWAIT)
+			return -EAGAIN;
+
 		if (!drained_dio) {
 			if (*iolock == XFS_IOLOCK_SHARED) {
 				xfs_iunlock(ip, *iolock);
@@ -593,7 +608,8 @@ xfs_file_dio_aio_write(
 			   &xfs_dio_write_ops,
 			   is_sync_kiocb(iocb) || unaligned_io);
 out:
-	xfs_iunlock(ip, iolock);
+	if (iolock)
+		xfs_iunlock(ip, iolock);
 
 	/*
 	 * No fallback to buffered IO after short writes for XFS, direct I/O
@@ -632,7 +648,8 @@ xfs_file_dax_write(
 		error = xfs_setfilesize(ip, pos, ret);
 	}
 out:
-	xfs_iunlock(ip, iolock);
+	if (iolock)
+		xfs_iunlock(ip, iolock);
 	if (error)
 		return error;
 
-- 
2.29.2

[PATCH 03/11] xfs: cleanup the read/write helper naming

[Christoph Hellwig]

Drop a few pointless aio_ prefixes.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
---
 fs/xfs/xfs_file.c | 30 +++++++++++++++---------------
 1 file changed, 15 insertions(+), 15 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index fb4e6f2852bb8b..ae7313ccaa11ed 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -215,7 +215,7 @@ xfs_ilock_iocb(
 }
 
 STATIC ssize_t
-xfs_file_dio_aio_read(
+xfs_file_dio_read(
 	struct kiocb		*iocb,
 	struct iov_iter		*to)
 {
@@ -265,7 +265,7 @@ xfs_file_dax_read(
 }
 
 STATIC ssize_t
-xfs_file_buffered_aio_read(
+xfs_file_buffered_read(
 	struct kiocb		*iocb,
 	struct iov_iter		*to)
 {
@@ -300,9 +300,9 @@ xfs_file_read_iter(
 	if (IS_DAX(inode))
 		ret = xfs_file_dax_read(iocb, to);
 	else if (iocb->ki_flags & IOCB_DIRECT)
-		ret = xfs_file_dio_aio_read(iocb, to);
+		ret = xfs_file_dio_read(iocb, to);
 	else
-		ret = xfs_file_buffered_aio_read(iocb, to);
+		ret = xfs_file_buffered_read(iocb, to);
 
 	if (ret > 0)
 		XFS_STATS_ADD(mp, xs_read_bytes, ret);
@@ -317,7 +317,7 @@ xfs_file_read_iter(
  * if called for a direct write beyond i_size.
  */
 STATIC ssize_t
-xfs_file_aio_write_checks(
+xfs_file_write_checks(
 	struct kiocb		*iocb,
 	struct iov_iter		*from,
 	int			*iolock)
@@ -502,7 +502,7 @@ static const struct iomap_dio_ops xfs_dio_write_ops = {
 };
 
 /*
- * xfs_file_dio_aio_write - handle direct IO writes
+ * xfs_file_dio_write - handle direct IO writes
  *
  * Lock the inode appropriately to prepare for and issue a direct IO write.
  * By separating it from the buffered write path we remove all the tricky to
@@ -527,7 +527,7 @@ static const struct iomap_dio_ops xfs_dio_write_ops = {
  * negative return values.
  */
 STATIC ssize_t
-xfs_file_dio_aio_write(
+xfs_file_dio_write(
 	struct kiocb		*iocb,
 	struct iov_iter		*from)
 {
@@ -549,7 +549,7 @@ xfs_file_dio_aio_write(
 	/*
 	 * Don't take the exclusive iolock here unless the I/O is unaligned to
 	 * the file system block size.  We don't need to consider the EOF
-	 * extension case here because xfs_file_aio_write_checks() will relock
+	 * extension case here because xfs_file_write_checks() will relock
 	 * the inode as necessary for EOF zeroing cases and fill out the new
 	 * inode size as appropriate.
 	 */
@@ -580,7 +580,7 @@ xfs_file_dio_aio_write(
 		xfs_ilock(ip, iolock);
 	}
 
-	ret = xfs_file_aio_write_checks(iocb, from, &iolock);
+	ret = xfs_file_write_checks(iocb, from, &iolock);
 	if (ret)
 		goto out;
 	count = iov_iter_count(from);
@@ -590,7 +590,7 @@ xfs_file_dio_aio_write(
 	 * in-flight at the same time or we risk data corruption. Wait for all
 	 * other IO to drain before we submit. If the IO is aligned, demote the
 	 * iolock if we had to take the exclusive lock in
-	 * xfs_file_aio_write_checks() for other reasons.
+	 * xfs_file_write_checks() for other reasons.
 	 */
 	if (unaligned_io) {
 		inode_dio_wait(inode);
@@ -634,7 +634,7 @@ xfs_file_dax_write(
 	ret = xfs_ilock_iocb(iocb, iolock);
 	if (ret)
 		return ret;
-	ret = xfs_file_aio_write_checks(iocb, from, &iolock);
+	ret = xfs_file_write_checks(iocb, from, &iolock);
 	if (ret)
 		goto out;
 
@@ -663,7 +663,7 @@ xfs_file_dax_write(
 }
 
 STATIC ssize_t
-xfs_file_buffered_aio_write(
+xfs_file_buffered_write(
 	struct kiocb		*iocb,
 	struct iov_iter		*from)
 {
@@ -682,7 +682,7 @@ xfs_file_buffered_aio_write(
 	iolock = XFS_IOLOCK_EXCL;
 	xfs_ilock(ip, iolock);
 
-	ret = xfs_file_aio_write_checks(iocb, from, &iolock);
+	ret = xfs_file_write_checks(iocb, from, &iolock);
 	if (ret)
 		goto out;
 
@@ -769,12 +769,12 @@ xfs_file_write_iter(
 		 * CoW.  In all other directio scenarios we do not
 		 * allow an operation to fall back to buffered mode.
 		 */
-		ret = xfs_file_dio_aio_write(iocb, from);
+		ret = xfs_file_dio_write(iocb, from);
 		if (ret != -ENOTBLK)
 			return ret;
 	}
 
-	return xfs_file_buffered_aio_write(iocb, from);
+	return xfs_file_buffered_write(iocb, from);
 }
 
 static void
-- 
2.29.2

[PATCH 04/11] xfs: remove the buffered I/O fallback assert

[Christoph Hellwig]

The iomap code has been designed from the start not to do magic fallback,
so remove the assert in preparation for further code cleanups.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
---
 fs/xfs/xfs_file.c | 6 ------
 1 file changed, 6 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index ae7313ccaa11ed..97836ec53397d4 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -610,12 +610,6 @@ xfs_file_dio_write(
 out:
 	if (iolock)
 		xfs_iunlock(ip, iolock);
-
-	/*
-	 * No fallback to buffered IO after short writes for XFS, direct I/O
-	 * will either complete fully or return an error.
-	 */
-	ASSERT(ret < 0 || ret == count);
 	return ret;
 }
 
-- 
2.29.2

[PATCH 05/11] xfs: simplify the read/write tracepoints

[Christoph Hellwig]

Pass the iocb and iov_iter to the tracepoints and leave decoding of
actual arguments to the code only run when tracing is enabled.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
---
 fs/xfs/xfs_file.c  | 20 ++++++++------------
 fs/xfs/xfs_trace.h | 18 +++++++++---------
 2 files changed, 17 insertions(+), 21 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index 97836ec53397d4..aa64e78fc3c467 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -220,12 +220,11 @@ xfs_file_dio_read(
 	struct iov_iter		*to)
 {
 	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
-	size_t			count = iov_iter_count(to);
 	ssize_t			ret;
 
-	trace_xfs_file_direct_read(ip, count, iocb->ki_pos);
+	trace_xfs_file_direct_read(iocb, to);
 
-	if (!count)
+	if (!iov_iter_count(to))
 		return 0; /* skip atime */
 
 	file_accessed(iocb->ki_filp);
@@ -246,12 +245,11 @@ xfs_file_dax_read(
 	struct iov_iter		*to)
 {
 	struct xfs_inode	*ip = XFS_I(iocb->ki_filp->f_mapping->host);
-	size_t			count = iov_iter_count(to);
 	ssize_t			ret = 0;
 
-	trace_xfs_file_dax_read(ip, count, iocb->ki_pos);
+	trace_xfs_file_dax_read(iocb, to);
 
-	if (!count)
+	if (!iov_iter_count(to))
 		return 0; /* skip atime */
 
 	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
@@ -272,7 +270,7 @@ xfs_file_buffered_read(
 	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
 	ssize_t			ret;
 
-	trace_xfs_file_buffered_read(ip, iov_iter_count(to), iocb->ki_pos);
+	trace_xfs_file_buffered_read(iocb, to);
 
 	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
 	if (ret)
@@ -599,7 +597,7 @@ xfs_file_dio_write(
 		iolock = XFS_IOLOCK_SHARED;
 	}
 
-	trace_xfs_file_direct_write(ip, count, iocb->ki_pos);
+	trace_xfs_file_direct_write(iocb, from);
 	/*
 	 * If unaligned, this is the only IO in-flight. Wait on it before we
 	 * release the iolock to prevent subsequent overlapping IO.
@@ -622,7 +620,6 @@ xfs_file_dax_write(
 	struct xfs_inode	*ip = XFS_I(inode);
 	int			iolock = XFS_IOLOCK_EXCL;
 	ssize_t			ret, error = 0;
-	size_t			count;
 	loff_t			pos;
 
 	ret = xfs_ilock_iocb(iocb, iolock);
@@ -633,9 +630,8 @@ xfs_file_dax_write(
 		goto out;
 
 	pos = iocb->ki_pos;
-	count = iov_iter_count(from);
 
-	trace_xfs_file_dax_write(ip, count, pos);
+	trace_xfs_file_dax_write(iocb, from);
 	ret = dax_iomap_rw(iocb, from, &xfs_direct_write_iomap_ops);
 	if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
 		i_size_write(inode, iocb->ki_pos);
@@ -683,7 +679,7 @@ xfs_file_buffered_write(
 	/* We can write back this queue in page reclaim */
 	current->backing_dev_info = inode_to_bdi(inode);
 
-	trace_xfs_file_buffered_write(ip, iov_iter_count(from), iocb->ki_pos);
+	trace_xfs_file_buffered_write(iocb, from);
 	ret = iomap_file_buffered_write(iocb, from,
 			&xfs_buffered_write_iomap_ops);
 	if (likely(ret >= 0))
diff --git a/fs/xfs/xfs_trace.h b/fs/xfs/xfs_trace.h
index 5a263ae3d4f008..a6d04d860a565e 100644
--- a/fs/xfs/xfs_trace.h
+++ b/fs/xfs/xfs_trace.h
@@ -1287,8 +1287,8 @@ TRACE_EVENT(xfs_log_assign_tail_lsn,
 )
 
 DECLARE_EVENT_CLASS(xfs_file_class,
-	TP_PROTO(struct xfs_inode *ip, size_t count, loff_t offset),
-	TP_ARGS(ip, count, offset),
+	TP_PROTO(struct kiocb *iocb, struct iov_iter *iter),
+	TP_ARGS(iocb, iter),
 	TP_STRUCT__entry(
 		__field(dev_t, dev)
 		__field(xfs_ino_t, ino)
@@ -1297,11 +1297,11 @@ DECLARE_EVENT_CLASS(xfs_file_class,
 		__field(size_t, count)
 	),
 	TP_fast_assign(
-		__entry->dev = VFS_I(ip)->i_sb->s_dev;
-		__entry->ino = ip->i_ino;
-		__entry->size = ip->i_d.di_size;
-		__entry->offset = offset;
-		__entry->count = count;
+		__entry->dev = file_inode(iocb->ki_filp)->i_sb->s_dev;
+		__entry->ino = XFS_I(file_inode(iocb->ki_filp))->i_ino;
+		__entry->size = XFS_I(file_inode(iocb->ki_filp))->i_d.di_size;
+		__entry->offset = iocb->ki_pos;
+		__entry->count = iov_iter_count(iter);
 	),
 	TP_printk("dev %d:%d ino 0x%llx size 0x%llx offset 0x%llx count 0x%zx",
 		  MAJOR(__entry->dev), MINOR(__entry->dev),
@@ -1313,8 +1313,8 @@ DECLARE_EVENT_CLASS(xfs_file_class,
 
 #define DEFINE_RW_EVENT(name)		\
 DEFINE_EVENT(xfs_file_class, name,	\
-	TP_PROTO(struct xfs_inode *ip, size_t count, loff_t offset),	\
-	TP_ARGS(ip, count, offset))
+	TP_PROTO(struct kiocb *iocb, struct iov_iter *iter),		\
+	TP_ARGS(iocb, iter))
 DEFINE_RW_EVENT(xfs_file_buffered_read);
 DEFINE_RW_EVENT(xfs_file_direct_read);
 DEFINE_RW_EVENT(xfs_file_dax_read);
-- 
2.29.2

[PATCH 06/11] xfs: improve the reflink_bounce_dio_write tracepoint

[Christoph Hellwig]

Use a more suitable event class.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
---
 fs/xfs/xfs_file.c  | 2 +-
 fs/xfs/xfs_trace.h | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index aa64e78fc3c467..a696bd34f71d21 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -560,7 +560,7 @@ xfs_file_dio_write(
 		 * files yet, as we can't unshare a partial block.
 		 */
 		if (xfs_is_cow_inode(ip)) {
-			trace_xfs_reflink_bounce_dio_write(ip, iocb->ki_pos, count);
+			trace_xfs_reflink_bounce_dio_write(iocb, from);
 			return -ENOTBLK;
 		}
 		iolock = XFS_IOLOCK_EXCL;
diff --git a/fs/xfs/xfs_trace.h b/fs/xfs/xfs_trace.h
index a6d04d860a565e..0cfd65cd67c190 100644
--- a/fs/xfs/xfs_trace.h
+++ b/fs/xfs/xfs_trace.h
@@ -1321,6 +1321,8 @@ DEFINE_RW_EVENT(xfs_file_dax_read);
 DEFINE_RW_EVENT(xfs_file_buffered_write);
 DEFINE_RW_EVENT(xfs_file_direct_write);
 DEFINE_RW_EVENT(xfs_file_dax_write);
+DEFINE_RW_EVENT(xfs_reflink_bounce_dio_write);
+
 
 DECLARE_EVENT_CLASS(xfs_imap_class,
 	TP_PROTO(struct xfs_inode *ip, xfs_off_t offset, ssize_t count,
@@ -3294,8 +3296,6 @@ DEFINE_INODE_IREC_EVENT(xfs_reflink_cow_found);
 DEFINE_INODE_IREC_EVENT(xfs_reflink_cow_enospc);
 DEFINE_INODE_IREC_EVENT(xfs_reflink_convert_cow);
 
-DEFINE_SIMPLE_IO_EVENT(xfs_reflink_bounce_dio_write);
-
 DEFINE_SIMPLE_IO_EVENT(xfs_reflink_cancel_cow_range);
 DEFINE_SIMPLE_IO_EVENT(xfs_reflink_end_cow);
 DEFINE_INODE_IREC_EVENT(xfs_reflink_cow_remap);
-- 
2.29.2

[PATCH 07/11] xfs: split the unaligned DIO write code out

[Christoph Hellwig]

From: Dave Chinner <dchinner@redhat.com>

The unaligned DIO write path is more convolted than the normal path,
and we are about to make it more complex. Keep the block aligned
fast path dio write code trim and simple by splitting out the
unaligned DIO code from it.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
[hch: rebased, fixed a few minor nits]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
---
 fs/xfs/xfs_file.c | 170 +++++++++++++++++++++++-----------------------
 1 file changed, 85 insertions(+), 85 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index a696bd34f71d21..db0d96ca4d92cb 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -500,117 +500,117 @@ static const struct iomap_dio_ops xfs_dio_write_ops = {
 };
 
 /*
- * xfs_file_dio_write - handle direct IO writes
- *
- * Lock the inode appropriately to prepare for and issue a direct IO write.
- * By separating it from the buffered write path we remove all the tricky to
- * follow locking changes and looping.
- *
- * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
- * until we're sure the bytes at the new EOF have been zeroed and/or the cached
- * pages are flushed out.
- *
- * In most cases the direct IO writes will be done holding IOLOCK_SHARED
- * allowing them to be done in parallel with reads and other direct IO writes.
- * However, if the IO is not aligned to filesystem blocks, the direct IO layer
- * needs to do sub-block zeroing and that requires serialisation against other
- * direct IOs to the same block. In this case we need to serialise the
- * submission of the unaligned IOs so that we don't get racing block zeroing in
- * the dio layer.  To avoid the problem with aio, we also need to wait for
- * outstanding IOs to complete so that unwritten extent conversion is completed
- * before we try to map the overlapping block. This is currently implemented by
- * hitting it with a big hammer (i.e. inode_dio_wait()).
- *
- * Returns with locks held indicated by @iolock and errors indicated by
- * negative return values.
+ * Handle block aligned direct I/O writes
  */
-STATIC ssize_t
-xfs_file_dio_write(
+static noinline ssize_t
+xfs_file_dio_write_aligned(
+	struct xfs_inode	*ip,
 	struct kiocb		*iocb,
 	struct iov_iter		*from)
 {
-	struct file		*file = iocb->ki_filp;
-	struct address_space	*mapping = file->f_mapping;
-	struct inode		*inode = mapping->host;
-	struct xfs_inode	*ip = XFS_I(inode);
-	struct xfs_mount	*mp = ip->i_mount;
-	ssize_t			ret = 0;
-	int			unaligned_io = 0;
-	int			iolock;
-	size_t			count = iov_iter_count(from);
-	struct xfs_buftarg      *target = xfs_inode_buftarg(ip);
+	int			iolock = XFS_IOLOCK_SHARED;
+	ssize_t			ret;
 
-	/* DIO must be aligned to device logical sector size */
-	if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
-		return -EINVAL;
+	ret = xfs_ilock_iocb(iocb, iolock);
+	if (ret)
+		return ret;
+	ret = xfs_file_write_checks(iocb, from, &iolock);
+	if (ret)
+		goto out_unlock;
 
 	/*
-	 * Don't take the exclusive iolock here unless the I/O is unaligned to
-	 * the file system block size.  We don't need to consider the EOF
-	 * extension case here because xfs_file_write_checks() will relock
-	 * the inode as necessary for EOF zeroing cases and fill out the new
-	 * inode size as appropriate.
+	 * We don't need to hold the IOLOCK exclusively across the IO, so demote
+	 * the iolock back to shared if we had to take the exclusive lock in
+	 * xfs_file_write_checks() for other reasons.
 	 */
-	if ((iocb->ki_pos & mp->m_blockmask) ||
-	    ((iocb->ki_pos + count) & mp->m_blockmask)) {
-		unaligned_io = 1;
-
-		/*
-		 * We can't properly handle unaligned direct I/O to reflink
-		 * files yet, as we can't unshare a partial block.
-		 */
-		if (xfs_is_cow_inode(ip)) {
-			trace_xfs_reflink_bounce_dio_write(iocb, from);
-			return -ENOTBLK;
-		}
-		iolock = XFS_IOLOCK_EXCL;
-	} else {
+	if (iolock == XFS_IOLOCK_EXCL) {
+		xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
 		iolock = XFS_IOLOCK_SHARED;
 	}
+	trace_xfs_file_direct_write(iocb, from);
+	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
+			   &xfs_dio_write_ops, is_sync_kiocb(iocb));
+out_unlock:
+	if (iolock)
+		xfs_iunlock(ip, iolock);
+	return ret;
+}
 
-	if (iocb->ki_flags & IOCB_NOWAIT) {
-		/* unaligned dio always waits, bail */
-		if (unaligned_io)
-			return -EAGAIN;
-		if (!xfs_ilock_nowait(ip, iolock))
-			return -EAGAIN;
-	} else {
-		xfs_ilock(ip, iolock);
+/*
+ * Handle block unaligned direct I/O writes
+ *
+ * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
+ * them to be done in parallel with reads and other direct I/O writes.  However,
+ * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
+ * to do sub-block zeroing and that requires serialisation against other direct
+ * I/O to the same block.  In this case we need to serialise the submission of
+ * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
+ *
+ * This means that unaligned dio writes always block. There is no "nowait" fast
+ * path in this code - if IOCB_NOWAIT is set we simply return -EAGAIN up front
+ * and we don't have to worry about that anymore.
+ */
+static noinline ssize_t
+xfs_file_dio_write_unaligned(
+	struct xfs_inode	*ip,
+	struct kiocb		*iocb,
+	struct iov_iter		*from)
+{
+	int			iolock = XFS_IOLOCK_EXCL;
+	ssize_t			ret;
+
+	/* unaligned dio always waits, bail */
+	if (iocb->ki_flags & IOCB_NOWAIT)
+		return -EAGAIN;
+	xfs_ilock(ip, iolock);
+
+	/*
+	 * We can't properly handle unaligned direct I/O to reflink files yet,
+	 * as we can't unshare a partial block.
+	 */
+	if (xfs_is_cow_inode(ip)) {
+		trace_xfs_reflink_bounce_dio_write(iocb, from);
+		ret = -ENOTBLK;
+		goto out_unlock;
 	}
 
 	ret = xfs_file_write_checks(iocb, from, &iolock);
 	if (ret)
-		goto out;
-	count = iov_iter_count(from);
+		goto out_unlock;
 
 	/*
-	 * If we are doing unaligned IO, we can't allow any other overlapping IO
-	 * in-flight at the same time or we risk data corruption. Wait for all
-	 * other IO to drain before we submit. If the IO is aligned, demote the
-	 * iolock if we had to take the exclusive lock in
-	 * xfs_file_write_checks() for other reasons.
+	 * If we are doing unaligned I/O, this must be the only I/O in-flight.
+	 * Otherwise we risk data corruption due to unwritten extent conversions
+	 * from the AIO end_io handler.  Wait for all other I/O to drain first.
 	 */
-	if (unaligned_io) {
-		inode_dio_wait(inode);
-	} else if (iolock == XFS_IOLOCK_EXCL) {
-		xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
-		iolock = XFS_IOLOCK_SHARED;
-	}
+	inode_dio_wait(VFS_I(ip));
 
 	trace_xfs_file_direct_write(iocb, from);
-	/*
-	 * If unaligned, this is the only IO in-flight. Wait on it before we
-	 * release the iolock to prevent subsequent overlapping IO.
-	 */
 	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
-			   &xfs_dio_write_ops,
-			   is_sync_kiocb(iocb) || unaligned_io);
-out:
+			   &xfs_dio_write_ops, true);
+out_unlock:
 	if (iolock)
 		xfs_iunlock(ip, iolock);
 	return ret;
 }
 
+static ssize_t
+xfs_file_dio_write(
+	struct kiocb		*iocb,
+	struct iov_iter		*from)
+{
+	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
+	struct xfs_buftarg      *target = xfs_inode_buftarg(ip);
+	size_t			count = iov_iter_count(from);
+
+	/* direct I/O must be aligned to device logical sector size */
+	if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
+		return -EINVAL;
+	if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
+		return xfs_file_dio_write_unaligned(ip, iocb, from);
+	return xfs_file_dio_write_aligned(ip, iocb, from);
+}
+
 static noinline ssize_t
 xfs_file_dax_write(
 	struct kiocb		*iocb,
-- 
2.29.2

[PATCH 08/11] iomap: rename the flags variable in __iomap_dio_rw

[Christoph Hellwig]

Rename flags to iomap_flags to make the usage a little more clear.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
---
 fs/iomap/direct-io.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index 933f234d5becd0..604103ab76f9c5 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -427,7 +427,7 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 	size_t count = iov_iter_count(iter);
 	loff_t pos = iocb->ki_pos;
 	loff_t end = iocb->ki_pos + count - 1, ret = 0;
-	unsigned int flags = IOMAP_DIRECT;
+	unsigned int iomap_flags = IOMAP_DIRECT;
 	struct blk_plug plug;
 	struct iomap_dio *dio;
 
@@ -461,7 +461,7 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		if (iter_is_iovec(iter))
 			dio->flags |= IOMAP_DIO_DIRTY;
 	} else {
-		flags |= IOMAP_WRITE;
+		iomap_flags |= IOMAP_WRITE;
 		dio->flags |= IOMAP_DIO_WRITE;
 
 		/* for data sync or sync, we need sync completion processing */
@@ -483,7 +483,7 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 			ret = -EAGAIN;
 			goto out_free_dio;
 		}
-		flags |= IOMAP_NOWAIT;
+		iomap_flags |= IOMAP_NOWAIT;
 	}
 
 	ret = filemap_write_and_wait_range(mapping, pos, end);
@@ -514,7 +514,7 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 
 	blk_start_plug(&plug);
 	do {
-		ret = iomap_apply(inode, pos, count, flags, ops, dio,
+		ret = iomap_apply(inode, pos, count, iomap_flags, ops, dio,
 				iomap_dio_actor);
 		if (ret <= 0) {
 			/* magic error code to fall back to buffered I/O */
-- 
2.29.2

[PATCH 09/11] iomap: pass a flags argument to iomap_dio_rw

[Christoph Hellwig]

Pass a set of flags to iomap_dio_rw instead of the boolean
wait_for_completion argument.  The IOMAP_DIO_FORCE_WAIT flag
replaces the wait_for_completion, but only needs to be passed
when the iocb isn't synchronous to start with to simplify the
callers.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
---
 fs/btrfs/file.c       |  7 +++----
 fs/ext4/file.c        |  5 ++---
 fs/gfs2/file.c        |  7 ++-----
 fs/iomap/direct-io.c  | 11 +++++------
 fs/xfs/xfs_file.c     |  7 +++----
 fs/zonefs/super.c     |  4 ++--
 include/linux/iomap.h | 10 ++++++++--
 7 files changed, 25 insertions(+), 26 deletions(-)

diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c
index 0e41459b8de667..ddfd2e2adedf58 100644
--- a/fs/btrfs/file.c
+++ b/fs/btrfs/file.c
@@ -1949,8 +1949,8 @@ static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
 		goto buffered;
 	}
 
-	dio = __iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops,
-			     &btrfs_dio_ops, is_sync_kiocb(iocb));
+	dio = __iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
+			     0);
 
 	btrfs_inode_unlock(inode, ilock_flags);
 
@@ -3622,8 +3622,7 @@ static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
 		return 0;
 
 	btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
-	ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
-			   is_sync_kiocb(iocb));
+	ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops, &btrfs_dio_ops, 0);
 	btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
 	return ret;
 }
diff --git a/fs/ext4/file.c b/fs/ext4/file.c
index 349b27f0dda0cb..194f5d00fa3267 100644
--- a/fs/ext4/file.c
+++ b/fs/ext4/file.c
@@ -74,8 +74,7 @@ static ssize_t ext4_dio_read_iter(struct kiocb *iocb, struct iov_iter *to)
 		return generic_file_read_iter(iocb, to);
 	}
 
-	ret = iomap_dio_rw(iocb, to, &ext4_iomap_ops, NULL,
-			   is_sync_kiocb(iocb));
+	ret = iomap_dio_rw(iocb, to, &ext4_iomap_ops, NULL, 0);
 	inode_unlock_shared(inode);
 
 	file_accessed(iocb->ki_filp);
@@ -550,7 +549,7 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 	if (ilock_shared)
 		iomap_ops = &ext4_iomap_overwrite_ops;
 	ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
-			   is_sync_kiocb(iocb) || unaligned_io || extend);
+			   (unaligned_io || extend) ? IOMAP_DIO_FORCE_WAIT : 0);
 	if (ret == -ENOTBLK)
 		ret = 0;
 
diff --git a/fs/gfs2/file.c b/fs/gfs2/file.c
index b39b339feddc93..89609c2997177a 100644
--- a/fs/gfs2/file.c
+++ b/fs/gfs2/file.c
@@ -797,9 +797,7 @@ static ssize_t gfs2_file_direct_read(struct kiocb *iocb, struct iov_iter *to,
 	if (ret)
 		goto out_uninit;
 
-	ret = iomap_dio_rw(iocb, to, &gfs2_iomap_ops, NULL,
-			   is_sync_kiocb(iocb));
-
+	ret = iomap_dio_rw(iocb, to, &gfs2_iomap_ops, NULL, 0);
 	gfs2_glock_dq(gh);
 out_uninit:
 	gfs2_holder_uninit(gh);
@@ -833,8 +831,7 @@ static ssize_t gfs2_file_direct_write(struct kiocb *iocb, struct iov_iter *from,
 	if (offset + len > i_size_read(&ip->i_inode))
 		goto out;
 
-	ret = iomap_dio_rw(iocb, from, &gfs2_iomap_ops, NULL,
-			   is_sync_kiocb(iocb));
+	ret = iomap_dio_rw(iocb, from, &gfs2_iomap_ops, NULL, 0);
 	if (ret == -ENOTBLK)
 		ret = 0;
 out:
diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index 604103ab76f9c5..947343730e2c93 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -420,13 +420,15 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 struct iomap_dio *
 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
-		bool wait_for_completion)
+		unsigned int dio_flags)
 {
 	struct address_space *mapping = iocb->ki_filp->f_mapping;
 	struct inode *inode = file_inode(iocb->ki_filp);
 	size_t count = iov_iter_count(iter);
 	loff_t pos = iocb->ki_pos;
 	loff_t end = iocb->ki_pos + count - 1, ret = 0;
+	bool wait_for_completion =
+		is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT);
 	unsigned int iomap_flags = IOMAP_DIRECT;
 	struct blk_plug plug;
 	struct iomap_dio *dio;
@@ -434,9 +436,6 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 	if (!count)
 		return NULL;
 
-	if (WARN_ON(is_sync_kiocb(iocb) && !wait_for_completion))
-		return ERR_PTR(-EIO);
-
 	dio = kmalloc(sizeof(*dio), GFP_KERNEL);
 	if (!dio)
 		return ERR_PTR(-ENOMEM);
@@ -598,11 +597,11 @@ EXPORT_SYMBOL_GPL(__iomap_dio_rw);
 ssize_t
 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
-		bool wait_for_completion)
+		unsigned int dio_flags)
 {
 	struct iomap_dio *dio;
 
-	dio = __iomap_dio_rw(iocb, iter, ops, dops, wait_for_completion);
+	dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags);
 	if (IS_ERR_OR_NULL(dio))
 		return PTR_ERR_OR_ZERO(dio);
 	return iomap_dio_complete(dio);
diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index db0d96ca4d92cb..c60ff7b5dd829e 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -232,8 +232,7 @@ xfs_file_dio_read(
 	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
 	if (ret)
 		return ret;
-	ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL,
-			is_sync_kiocb(iocb));
+	ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0);
 	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 
 	return ret;
@@ -529,7 +528,7 @@ xfs_file_dio_write_aligned(
 	}
 	trace_xfs_file_direct_write(iocb, from);
 	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
-			   &xfs_dio_write_ops, is_sync_kiocb(iocb));
+			   &xfs_dio_write_ops, 0);
 out_unlock:
 	if (iolock)
 		xfs_iunlock(ip, iolock);
@@ -587,7 +586,7 @@ xfs_file_dio_write_unaligned(
 
 	trace_xfs_file_direct_write(iocb, from);
 	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
-			   &xfs_dio_write_ops, true);
+			   &xfs_dio_write_ops, IOMAP_DIO_FORCE_WAIT);
 out_unlock:
 	if (iolock)
 		xfs_iunlock(ip, iolock);
diff --git a/fs/zonefs/super.c b/fs/zonefs/super.c
index bec47f2d074beb..0e7ab0bc00ae8e 100644
--- a/fs/zonefs/super.c
+++ b/fs/zonefs/super.c
@@ -780,7 +780,7 @@ static ssize_t zonefs_file_dio_write(struct kiocb *iocb, struct iov_iter *from)
 		ret = zonefs_file_dio_append(iocb, from);
 	else
 		ret = iomap_dio_rw(iocb, from, &zonefs_iomap_ops,
-				   &zonefs_write_dio_ops, sync);
+				   &zonefs_write_dio_ops, 0);
 	if (zi->i_ztype == ZONEFS_ZTYPE_SEQ &&
 	    (ret > 0 || ret == -EIOCBQUEUED)) {
 		if (ret > 0)
@@ -917,7 +917,7 @@ static ssize_t zonefs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 		}
 		file_accessed(iocb->ki_filp);
 		ret = iomap_dio_rw(iocb, to, &zonefs_iomap_ops,
-				   &zonefs_read_dio_ops, is_sync_kiocb(iocb));
+				   &zonefs_read_dio_ops, 0);
 	} else {
 		ret = generic_file_read_iter(iocb, to);
 		if (ret == -EIO)
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 5bd3cac4df9cb4..be4e1e1e01e801 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -256,12 +256,18 @@ struct iomap_dio_ops {
 			struct bio *bio, loff_t file_offset);
 };
 
+/*
+ * Wait for the I/O to complete in iomap_dio_rw even if the kiocb is not
+ * synchronous.
+ */
+#define IOMAP_DIO_FORCE_WAIT	(1 << 0)
+
 ssize_t iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
-		bool wait_for_completion);
+		unsigned int dio_flags);
 struct iomap_dio *__iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
-		bool wait_for_completion);
+		unsigned int dio_flags);
 ssize_t iomap_dio_complete(struct iomap_dio *dio);
 int iomap_dio_iopoll(struct kiocb *kiocb, bool spin);
 
-- 
2.29.2

[PATCH 10/11] iomap: add a IOMAP_DIO_OVERWRITE_ONLY flag

[Christoph Hellwig]

Add a flag to signal that only pure overwrites are allowed.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
---
 fs/iomap/direct-io.c  | 7 +++++++
 include/linux/iomap.h | 8 ++++++++
 2 files changed, 15 insertions(+)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index 947343730e2c93..65d32364345d22 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -485,6 +485,13 @@ __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		iomap_flags |= IOMAP_NOWAIT;
 	}
 
+	if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) {
+		ret = -EAGAIN;
+		if (pos >= dio->i_size || pos + count > dio->i_size)
+			goto out_free_dio;
+		iomap_flags |= IOMAP_OVERWRITE_ONLY;
+	}
+
 	ret = filemap_write_and_wait_range(mapping, pos, end);
 	if (ret)
 		goto out_free_dio;
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index be4e1e1e01e801..1a86f520de560a 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -122,6 +122,7 @@ struct iomap_page_ops {
 #define IOMAP_FAULT		(1 << 3) /* mapping for page fault */
 #define IOMAP_DIRECT		(1 << 4) /* direct I/O */
 #define IOMAP_NOWAIT		(1 << 5) /* do not block */
+#define IOMAP_OVERWRITE_ONLY	(1 << 6) /* only pure overwrites allowed */
 
 struct iomap_ops {
 	/*
@@ -262,6 +263,13 @@ struct iomap_dio_ops {
  */
 #define IOMAP_DIO_FORCE_WAIT	(1 << 0)
 
+/*
+ * Do not allocate blocks or zero partial blocks, but instead fall back to
+ * the caller by returning -EAGAIN.  Used to optimize direct I/O writes that
+ * are not aligned to the file system block size.
+  */
+#define IOMAP_DIO_OVERWRITE_ONLY	(1 << 1)
+
 ssize_t iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
 		unsigned int dio_flags);
-- 
2.29.2

[PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Christoph Hellwig]

From: Dave Chinner <dchinner@redhat.com>

Attempt shared locking for unaligned DIO, but only if the the
underlying extent is already allocated and in written state. On
failure, retry with the existing exclusive locking.

Test case is fio randrw of 512 byte IOs using AIO and an iodepth of
32 IOs.

Vanilla:

  READ: bw=4560KiB/s (4670kB/s), 4560KiB/s-4560KiB/s (4670kB/s-4670kB/s), io=134MiB (140MB), run=30001-30001msec
  WRITE: bw=4567KiB/s (4676kB/s), 4567KiB/s-4567KiB/s (4676kB/s-4676kB/s), io=134MiB (140MB), run=30001-30001msec

Patched:
   READ: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1127MiB (1182MB), run=30002-30002msec
  WRITE: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1128MiB (1183MB), run=30002-30002msec

That's an improvement from ~18k IOPS to a ~150k IOPS, which is
about the IOPS limit of the VM block device setup I'm testing on.

4kB block IO comparison:

   READ: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8868MiB (9299MB), run=30002-30002msec
  WRITE: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8878MiB (9309MB), run=30002-30002msec

Which is ~150k IOPS, same as what the test gets for sub-block
AIO+DIO writes with this patch.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
[hch: rebased, split unaligned from nowait]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
---
 fs/xfs/xfs_file.c  | 58 +++++++++++++++++++++++++++++++++++-----------
 fs/xfs/xfs_iomap.c | 29 ++++++++++++++++-------
 2 files changed, 66 insertions(+), 21 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index c60ff7b5dd829e..39695b59dfcc92 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -544,10 +544,13 @@ xfs_file_dio_write_aligned(
  * to do sub-block zeroing and that requires serialisation against other direct
  * I/O to the same block.  In this case we need to serialise the submission of
  * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
+ * In the case where sub-block zeroing is not required, we can do concurrent
+ * sub-block dios to the same block successfully.
  *
- * This means that unaligned dio writes always block. There is no "nowait" fast
- * path in this code - if IOCB_NOWAIT is set we simply return -EAGAIN up front
- * and we don't have to worry about that anymore.
+ * Optimistically submit the I/O using the shared lock first, but use the
+ * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
+ * if block allocation or partial block zeroing would be required.  In that case
+ * we try again with the exclusive lock.
  */
 static noinline ssize_t
 xfs_file_dio_write_unaligned(
@@ -555,13 +558,28 @@ xfs_file_dio_write_unaligned(
 	struct kiocb		*iocb,
 	struct iov_iter		*from)
 {
-	int			iolock = XFS_IOLOCK_EXCL;
+	size_t			isize = i_size_read(VFS_I(ip));
+	size_t			count = iov_iter_count(from);
+	int			iolock = XFS_IOLOCK_SHARED;
+	unsigned int		flags = IOMAP_DIO_OVERWRITE_ONLY;
 	ssize_t			ret;
 
-	/* unaligned dio always waits, bail */
-	if (iocb->ki_flags & IOCB_NOWAIT)
-		return -EAGAIN;
-	xfs_ilock(ip, iolock);
+	/*
+	 * Extending writes need exclusivity because of the sub-block zeroing
+	 * that the DIO code always does for partial tail blocks beyond EOF, so
+	 * don't even bother trying the fast path in this case.
+	 */
+	if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
+retry_exclusive:
+		if (iocb->ki_flags & IOCB_NOWAIT)
+			return -EAGAIN;
+		iolock = XFS_IOLOCK_EXCL;
+		flags = IOMAP_DIO_FORCE_WAIT;
+	}
+
+	ret = xfs_ilock_iocb(iocb, iolock);
+	if (ret)
+		return ret;
 
 	/*
 	 * We can't properly handle unaligned direct I/O to reflink files yet,
@@ -578,15 +596,29 @@ xfs_file_dio_write_unaligned(
 		goto out_unlock;
 
 	/*
-	 * If we are doing unaligned I/O, this must be the only I/O in-flight.
-	 * Otherwise we risk data corruption due to unwritten extent conversions
-	 * from the AIO end_io handler.  Wait for all other I/O to drain first.
+	 * If we are doing exclusive unaligned I/O, this must be the only I/O
+	 * in-flight.  Otherwise we risk data corruption due to unwritten extent
+	 * conversions from the AIO end_io handler.  Wait for all other I/O to
+	 * drain first.
 	 */
-	inode_dio_wait(VFS_I(ip));
+	if (flags & IOMAP_DIO_FORCE_WAIT)
+		inode_dio_wait(VFS_I(ip));
 
 	trace_xfs_file_direct_write(iocb, from);
 	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
-			   &xfs_dio_write_ops, IOMAP_DIO_FORCE_WAIT);
+			   &xfs_dio_write_ops, flags);
+
+	/*
+	 * Retry unaligned I/O with exclusive blocking semantics if the DIO
+	 * layer rejected it for mapping or locking reasons. If we are doing
+	 * nonblocking user I/O, propagate the error.
+	 */
+	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
+		ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
+		xfs_iunlock(ip, iolock);
+		goto retry_exclusive;
+	}
+
 out_unlock:
 	if (iolock)
 		xfs_iunlock(ip, iolock);
diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
index 7b9ff824e82d48..ef76f775fabf11 100644
--- a/fs/xfs/xfs_iomap.c
+++ b/fs/xfs/xfs_iomap.c
@@ -784,15 +784,28 @@ xfs_direct_write_iomap_begin(
 		goto allocate_blocks;
 
 	/*
-	 * NOWAIT IO needs to span the entire requested IO with a single map so
-	 * that we avoid partial IO failures due to the rest of the IO range not
-	 * covered by this map triggering an EAGAIN condition when it is
-	 * subsequently mapped and aborting the IO.
+	 * NOWAIT and OVERWRITE I/O needs to span the entire requested I/O with
+	 * a single map so that we avoid partial IO failures due to the rest of
+	 * the I/O range not covered by this map triggering an EAGAIN condition
+	 * when it is subsequently mapped and aborting the I/O.
 	 */
-	if ((flags & IOMAP_NOWAIT) &&
-	    !imap_spans_range(&imap, offset_fsb, end_fsb)) {
+	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY)) {
 		error = -EAGAIN;
-		goto out_unlock;
+		if (!imap_spans_range(&imap, offset_fsb, end_fsb))
+			goto out_unlock;
+	}
+
+	/*
+	 * For overwrite only I/O, we cannot convert unwritten extents without
+	 * requiring sub-block zeroing.  This can only be done under an
+	 * exclusive IOLOCK, hence return -EAGAIN if this is not a written
+	 * extent to tell the caller to try again.
+	 */
+	if (flags & IOMAP_OVERWRITE_ONLY) {
+		error = -EAGAIN;
+		if (imap.br_state != XFS_EXT_NORM &&
+	            ((offset | length) & mp->m_blockmask))
+			goto out_unlock;
 	}
 
 	xfs_iunlock(ip, lockmode);
@@ -801,7 +814,7 @@ xfs_direct_write_iomap_begin(
 
 allocate_blocks:
 	error = -EAGAIN;
-	if (flags & IOMAP_NOWAIT)
+	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY))
 		goto out_unlock;
 
 	/*
-- 
2.29.2

Re: [PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Darrick J. Wong]

On Fri, Jan 22, 2021 at 05:20:43PM +0100, Christoph Hellwig wrote:
> From: Dave Chinner <dchinner@redhat.com>
> 
> Attempt shared locking for unaligned DIO, but only if the the
> underlying extent is already allocated and in written state. On
> failure, retry with the existing exclusive locking.
> 
> Test case is fio randrw of 512 byte IOs using AIO and an iodepth of
> 32 IOs.
> 
> Vanilla:
> 
>   READ: bw=4560KiB/s (4670kB/s), 4560KiB/s-4560KiB/s (4670kB/s-4670kB/s), io=134MiB (140MB), run=30001-30001msec
>   WRITE: bw=4567KiB/s (4676kB/s), 4567KiB/s-4567KiB/s (4676kB/s-4676kB/s), io=134MiB (140MB), run=30001-30001msec
> 
> Patched:
>    READ: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1127MiB (1182MB), run=30002-30002msec
>   WRITE: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1128MiB (1183MB), run=30002-30002msec
> 
> That's an improvement from ~18k IOPS to a ~150k IOPS, which is
> about the IOPS limit of the VM block device setup I'm testing on.
> 
> 4kB block IO comparison:
> 
>    READ: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8868MiB (9299MB), run=30002-30002msec
>   WRITE: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8878MiB (9309MB), run=30002-30002msec
> 
> Which is ~150k IOPS, same as what the test gets for sub-block
> AIO+DIO writes with this patch.
> 
> Signed-off-by: Dave Chinner <dchinner@redhat.com>
> [hch: rebased, split unaligned from nowait]
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> Reviewed-by: Brian Foster <bfoster@redhat.com>

Looks good to me now,
Reviewed-by: Darrick J. Wong <djwong@kernel.org>

--D

> ---
>  fs/xfs/xfs_file.c  | 58 +++++++++++++++++++++++++++++++++++-----------
>  fs/xfs/xfs_iomap.c | 29 ++++++++++++++++-------
>  2 files changed, 66 insertions(+), 21 deletions(-)
> 
> diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> index c60ff7b5dd829e..39695b59dfcc92 100644
> --- a/fs/xfs/xfs_file.c
> +++ b/fs/xfs/xfs_file.c
> @@ -544,10 +544,13 @@ xfs_file_dio_write_aligned(
>   * to do sub-block zeroing and that requires serialisation against other direct
>   * I/O to the same block.  In this case we need to serialise the submission of
>   * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
> + * In the case where sub-block zeroing is not required, we can do concurrent
> + * sub-block dios to the same block successfully.
>   *
> - * This means that unaligned dio writes always block. There is no "nowait" fast
> - * path in this code - if IOCB_NOWAIT is set we simply return -EAGAIN up front
> - * and we don't have to worry about that anymore.
> + * Optimistically submit the I/O using the shared lock first, but use the
> + * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
> + * if block allocation or partial block zeroing would be required.  In that case
> + * we try again with the exclusive lock.
>   */
>  static noinline ssize_t
>  xfs_file_dio_write_unaligned(
> @@ -555,13 +558,28 @@ xfs_file_dio_write_unaligned(
>  	struct kiocb		*iocb,
>  	struct iov_iter		*from)
>  {
> -	int			iolock = XFS_IOLOCK_EXCL;
> +	size_t			isize = i_size_read(VFS_I(ip));
> +	size_t			count = iov_iter_count(from);
> +	int			iolock = XFS_IOLOCK_SHARED;
> +	unsigned int		flags = IOMAP_DIO_OVERWRITE_ONLY;
>  	ssize_t			ret;
>  
> -	/* unaligned dio always waits, bail */
> -	if (iocb->ki_flags & IOCB_NOWAIT)
> -		return -EAGAIN;
> -	xfs_ilock(ip, iolock);
> +	/*
> +	 * Extending writes need exclusivity because of the sub-block zeroing
> +	 * that the DIO code always does for partial tail blocks beyond EOF, so
> +	 * don't even bother trying the fast path in this case.
> +	 */
> +	if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
> +retry_exclusive:
> +		if (iocb->ki_flags & IOCB_NOWAIT)
> +			return -EAGAIN;
> +		iolock = XFS_IOLOCK_EXCL;
> +		flags = IOMAP_DIO_FORCE_WAIT;
> +	}
> +
> +	ret = xfs_ilock_iocb(iocb, iolock);
> +	if (ret)
> +		return ret;
>  
>  	/*
>  	 * We can't properly handle unaligned direct I/O to reflink files yet,
> @@ -578,15 +596,29 @@ xfs_file_dio_write_unaligned(
>  		goto out_unlock;
>  
>  	/*
> -	 * If we are doing unaligned I/O, this must be the only I/O in-flight.
> -	 * Otherwise we risk data corruption due to unwritten extent conversions
> -	 * from the AIO end_io handler.  Wait for all other I/O to drain first.
> +	 * If we are doing exclusive unaligned I/O, this must be the only I/O
> +	 * in-flight.  Otherwise we risk data corruption due to unwritten extent
> +	 * conversions from the AIO end_io handler.  Wait for all other I/O to
> +	 * drain first.
>  	 */
> -	inode_dio_wait(VFS_I(ip));
> +	if (flags & IOMAP_DIO_FORCE_WAIT)
> +		inode_dio_wait(VFS_I(ip));
>  
>  	trace_xfs_file_direct_write(iocb, from);
>  	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
> -			   &xfs_dio_write_ops, IOMAP_DIO_FORCE_WAIT);
> +			   &xfs_dio_write_ops, flags);
> +
> +	/*
> +	 * Retry unaligned I/O with exclusive blocking semantics if the DIO
> +	 * layer rejected it for mapping or locking reasons. If we are doing
> +	 * nonblocking user I/O, propagate the error.
> +	 */
> +	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
> +		ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
> +		xfs_iunlock(ip, iolock);
> +		goto retry_exclusive;
> +	}
> +
>  out_unlock:
>  	if (iolock)
>  		xfs_iunlock(ip, iolock);
> diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
> index 7b9ff824e82d48..ef76f775fabf11 100644
> --- a/fs/xfs/xfs_iomap.c
> +++ b/fs/xfs/xfs_iomap.c
> @@ -784,15 +784,28 @@ xfs_direct_write_iomap_begin(
>  		goto allocate_blocks;
>  
>  	/*
> -	 * NOWAIT IO needs to span the entire requested IO with a single map so
> -	 * that we avoid partial IO failures due to the rest of the IO range not
> -	 * covered by this map triggering an EAGAIN condition when it is
> -	 * subsequently mapped and aborting the IO.
> +	 * NOWAIT and OVERWRITE I/O needs to span the entire requested I/O with
> +	 * a single map so that we avoid partial IO failures due to the rest of
> +	 * the I/O range not covered by this map triggering an EAGAIN condition
> +	 * when it is subsequently mapped and aborting the I/O.
>  	 */
> -	if ((flags & IOMAP_NOWAIT) &&
> -	    !imap_spans_range(&imap, offset_fsb, end_fsb)) {
> +	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY)) {
>  		error = -EAGAIN;
> -		goto out_unlock;
> +		if (!imap_spans_range(&imap, offset_fsb, end_fsb))
> +			goto out_unlock;
> +	}
> +
> +	/*
> +	 * For overwrite only I/O, we cannot convert unwritten extents without
> +	 * requiring sub-block zeroing.  This can only be done under an
> +	 * exclusive IOLOCK, hence return -EAGAIN if this is not a written
> +	 * extent to tell the caller to try again.
> +	 */
> +	if (flags & IOMAP_OVERWRITE_ONLY) {
> +		error = -EAGAIN;
> +		if (imap.br_state != XFS_EXT_NORM &&
> +	            ((offset | length) & mp->m_blockmask))
> +			goto out_unlock;
>  	}
>  
>  	xfs_iunlock(ip, lockmode);
> @@ -801,7 +814,7 @@ xfs_direct_write_iomap_begin(
>  
>  allocate_blocks:
>  	error = -EAGAIN;
> -	if (flags & IOMAP_NOWAIT)
> +	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY))
>  		goto out_unlock;
>  
>  	/*
> -- 
> 2.29.2
> 

Re: reduce sub-block DIO serialisation v4

[Darrick J. Wong]

On Fri, Jan 22, 2021 at 05:20:32PM +0100, Christoph Hellwig wrote:
> This takes the approach from Dave, but adds a new flag instead of abusing
> the nowait one, and keeps a simpler calling convention for iomap_dio_rw.

Hm.  I realized while putting together for-next branches that I really
would have preferred the three iomap patches at the start so that I
could push those parts through the iomap tree.  The changes required to
resequence the series is minor and the iomap changes (AFAICT) are inert
if the calling fs doesn't set IOMAP_DIO_OVERWRITE_ONLY, so I think it's
low risk to push the iomap changes into iomap-for-next as a 5.12 thing.

The rest of the xfs patches in this series would form the basis of a
second week pull request (or not) since I think I ought to evaluate the
effects on performance for a little longer.

That said, if you /don't/ want me to pull any of this for 5.12, holler
and I'll pull back. :)

--D

> Changes since v3:
>  - further comment improvements
>  - micro-optimize an alignment check
> 
> Changes since v2:
>  - add another sanity check in __iomap_dio_rw
>  - rename IOMAP_DIO_UNALIGNED to IOMAP_DIO_OVERWRITE_ONLY
>  - use more consistent parameter naming
>  - improve a few comments
>  - cleanup an if statement
> 
> Changes since v1:
>  - rename the new flags
>  - add an EOF check for subblock I/O
>  - minor cleanups

Re: reduce sub-block DIO serialisation v4

[Christoph Hellwig]

On Sat, Jan 23, 2021 at 10:57:06AM -0800, Darrick J. Wong wrote:
> On Fri, Jan 22, 2021 at 05:20:32PM +0100, Christoph Hellwig wrote:
> > This takes the approach from Dave, but adds a new flag instead of abusing
> > the nowait one, and keeps a simpler calling convention for iomap_dio_rw.
> 
> Hm.  I realized while putting together for-next branches that I really
> would have preferred the three iomap patches at the start so that I
> could push those parts through the iomap tree.  The changes required to
> resequence the series is minor and the iomap changes (AFAICT) are inert
> if the calling fs doesn't set IOMAP_DIO_OVERWRITE_ONLY, so I think it's
> low risk to push the iomap changes into iomap-for-next as a 5.12 thing.
> 
> The rest of the xfs patches in this series would form the basis of a
> second week pull request (or not) since I think I ought to evaluate the
> effects on performance for a little longer.

So that is the reason why they aren't in for-next yet?  Or do you want
the remaining patches resent on top of the iomap branch?

Re: reduce sub-block DIO serialisation v4

[Darrick J. Wong]

On Mon, Feb 01, 2021 at 05:58:25PM +0100, Christoph Hellwig wrote:
> On Sat, Jan 23, 2021 at 10:57:06AM -0800, Darrick J. Wong wrote:
> > On Fri, Jan 22, 2021 at 05:20:32PM +0100, Christoph Hellwig wrote:
> > > This takes the approach from Dave, but adds a new flag instead of abusing
> > > the nowait one, and keeps a simpler calling convention for iomap_dio_rw.
> > 
> > Hm.  I realized while putting together for-next branches that I really
> > would have preferred the three iomap patches at the start so that I
> > could push those parts through the iomap tree.  The changes required to
> > resequence the series is minor and the iomap changes (AFAICT) are inert
> > if the calling fs doesn't set IOMAP_DIO_OVERWRITE_ONLY, so I think it's
> > low risk to push the iomap changes into iomap-for-next as a 5.12 thing.
> > 
> > The rest of the xfs patches in this series would form the basis of a
> > second week pull request (or not) since I think I ought to evaluate the
> > effects on performance for a little longer.
> 
> So that is the reason why they aren't in for-next yet?  Or do you want
> the remaining patches resent on top of the iomap branch?

Assuming they haven't changed, I'll just push the (slightly reordered)
series out to for-next tomorrow.  Sorry, I got totally sidetracked last
week with the quota retry series nearly tripling in size with all the
requested changes... :(

--D

reduce sub-block DIO serialisation v3

[Christoph Hellwig]

This takes the approach from Dave, but adds a new flag instead of abusing
the nowait one, and keeps a simpler calling convention for iomap_dio_rw.

Changes since v2:
 - add another sanity check in __iomap_dio_rw
 - rename IOMAP_DIO_UNALIGNED to IOMAP_DIO_OVERWRITE_ONLY
 - use more consistent parameter naming
 - improve a few comments
 - cleanup an if statement

Changes since v1:
 - rename the new flags
 - add an EOF check for subblock I/O
 - minor cleanups

[PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Christoph Hellwig]

From: Dave Chinner <dchinner@redhat.com>

Attempt shared locking for unaligned DIO, but only if the the
underlying extent is already allocated and in written state. On
failure, retry with the existing exclusive locking.

Test case is fio randrw of 512 byte IOs using AIO and an iodepth of
32 IOs.

Vanilla:

  READ: bw=4560KiB/s (4670kB/s), 4560KiB/s-4560KiB/s (4670kB/s-4670kB/s), io=134MiB (140MB), run=30001-30001msec
  WRITE: bw=4567KiB/s (4676kB/s), 4567KiB/s-4567KiB/s (4676kB/s-4676kB/s), io=134MiB (140MB), run=30001-30001msec

Patched:
   READ: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1127MiB (1182MB), run=30002-30002msec
  WRITE: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1128MiB (1183MB), run=30002-30002msec

That's an improvement from ~18k IOPS to a ~150k IOPS, which is
about the IOPS limit of the VM block device setup I'm testing on.

4kB block IO comparison:

   READ: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8868MiB (9299MB), run=30002-30002msec
  WRITE: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8878MiB (9309MB), run=30002-30002msec

Which is ~150k IOPS, same as what the test gets for sub-block
AIO+DIO writes with this patch.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
[hch: rebased, split unaligned from nowait]
Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 fs/xfs/xfs_file.c  | 87 ++++++++++++++++++++++++++++++++--------------
 fs/xfs/xfs_iomap.c | 30 +++++++++++-----
 2 files changed, 83 insertions(+), 34 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index b181db42f2f32f..33899a5cca53f9 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -544,22 +544,35 @@ xfs_file_dio_write_aligned(
 /*
  * Handle block unaligned direct IO writes
  *
- * In most cases direct IO writes will be done holding IOLOCK_SHARED, allowing
- * them to be done in parallel with reads and other direct IO writes.  However,
- * if the I/O is not aligned to filesystem blocks, the direct I/O layer may
- * need to do sub-block zeroing and that requires serialisation against other
- * direct I/Os to the same block. In this case we need to serialise the
- * submission of the unaligned I/Os so that we don't get racing block zeroing in
- * the dio layer.
+ * In most cases direct IO writes will be done holding IOLOCK_SHARED
+ * allowing them to be done in parallel with reads and other direct IO writes.
+ * However, if the IO is not aligned to filesystem blocks, the direct IO layer
+ * may need to do sub-block zeroing and that requires serialisation against other
+ * direct IOs to the same block. In the case where sub-block zeroing is not
+ * required, we can do concurrent sub-block dios to the same block successfully.
  *
- * To provide the same serialisation for AIO, we also need to wait for
+ * Hence we have two cases here - the shared, optimisitic fast path for written
+ * extents, and everything else that needs exclusive IO path access across the
+ * entire IO.
+ *
+ * For the first case, we do all the checks we need at the mapping layer in the
+ * DIO code as part of the existing NOWAIT infrastructure. Hence all we need to
+ * do to support concurrent subblock dio is first try a non-blocking submission.
+ * If that returns -EAGAIN, then we simply repeat the IO submission with full
+ * IO exclusivity guaranteed so that we avoid racing sub-block zeroing.
+ *
+ * The only wrinkle in this case is that the iomap DIO code always does
+ * partial tail sub-block zeroing for post-EOF writes. Hence for any IO that
+ * _ends_ past the current EOF we need to run with full exclusivity. Note that
+ * we also check for the start of IO being beyond EOF because then zeroing
+ * between the old EOF and the start of the IO is required and that also
+ * requires exclusivity. Hence we avoid lock cycles and blocking under
+ * IOCB_NOWAIT for this situation, too.
+ *
+ * To provide the exclusivity required when using AIO, we also need to wait for
  * outstanding IOs to complete so that unwritten extent conversion is completed
  * before we try to map the overlapping block. This is currently implemented by
  * hitting it with a big hammer (i.e. inode_dio_wait()).
- *
- * This means that unaligned dio writes always block. There is no "nowait" fast
- * path in this code - if IOCB_NOWAIT is set we simply return -EAGAIN up front
- * and we don't have to worry about that anymore.
  */
 static noinline ssize_t
 xfs_file_dio_write_unaligned(
@@ -567,13 +580,27 @@ xfs_file_dio_write_unaligned(
 	struct kiocb		*iocb,
 	struct iov_iter		*from)
 {
-	int			iolock = XFS_IOLOCK_EXCL;
+	size_t			isize = i_size_read(VFS_I(ip));
+	size_t			count = iov_iter_count(from);
+	int			iolock = XFS_IOLOCK_SHARED;
+	unsigned int		flags = IOMAP_DIO_OVERWRITE_ONLY;
 	ssize_t			ret;
 
-	/* unaligned dio always waits, bail */
-	if (iocb->ki_flags & IOCB_NOWAIT)
-		return -EAGAIN;
-	xfs_ilock(ip, iolock);
+	/*
+	 * Extending writes need exclusivity because of the sub-block zeroing
+	 * that the DIO code always does for partial tail blocks beyond EOF.
+	 */
+	if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
+retry_exclusive:
+		if (iocb->ki_flags & IOCB_NOWAIT)
+			return -EAGAIN;
+		iolock = XFS_IOLOCK_EXCL;
+		flags = IOMAP_DIO_FORCE_WAIT;
+	}
+
+	ret = xfs_ilock_iocb(iocb, iolock);
+	if (ret)
+		return ret;
 
 	/*
 	 * We can't properly handle unaligned direct I/O to reflink files yet,
@@ -590,19 +617,27 @@ xfs_file_dio_write_unaligned(
 		goto out_unlock;
 
 	/*
-	 * If we are doing unaligned I/O, we can't allow any other overlapping
-	 * I/O in-flight at the same time or we risk data corruption. Wait for
-	 * all other I/O to drain before we submit.
+	 * If we are doing exclusive unaligned IO, we can't allow any other
+	 * overlapping IO in-flight at the same time or we risk data corruption.
+	 * Wait for all other IO to drain before we submit.
 	 */
-	inode_dio_wait(VFS_I(ip));
+	if (flags & IOMAP_DIO_FORCE_WAIT)
+		inode_dio_wait(VFS_I(ip));
 
-	/*
-	 * This must be the only I/O in-flight. Wait on it before we release the
-	 * iolock to prevent subsequent overlapping I/O.
-	 */
 	trace_xfs_file_direct_write(iocb, from);
 	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
-			   &xfs_dio_write_ops, IOMAP_DIO_FORCE_WAIT);
+			   &xfs_dio_write_ops, flags);
+	/*
+	 * Retry unaligned IO with exclusive blocking semantics if the DIO
+	 * layer rejected it for mapping or locking reasons. If we are doing
+	 * nonblocking user IO, propagate the error.
+	 */
+	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
+		ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
+		xfs_iunlock(ip, iolock);
+		goto retry_exclusive;
+	}
+
 out_unlock:
 	if (iolock)
 		xfs_iunlock(ip, iolock);
diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
index 7b9ff824e82d48..596af78f910596 100644
--- a/fs/xfs/xfs_iomap.c
+++ b/fs/xfs/xfs_iomap.c
@@ -784,15 +784,29 @@ xfs_direct_write_iomap_begin(
 		goto allocate_blocks;
 
 	/*
-	 * NOWAIT IO needs to span the entire requested IO with a single map so
-	 * that we avoid partial IO failures due to the rest of the IO range not
-	 * covered by this map triggering an EAGAIN condition when it is
-	 * subsequently mapped and aborting the IO.
+	 * NOWAIT and OVERWRITE needs to span the entire requested IO with a
+	 * single map so that we avoid partial IO failures due to the rest of
+	 * the IO range not covered by this map triggering an EAGAIN condition
+	 * when it is subsequently mapped and aborting the IO.
 	 */
-	if ((flags & IOMAP_NOWAIT) &&
-	    !imap_spans_range(&imap, offset_fsb, end_fsb)) {
+	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY)) {
 		error = -EAGAIN;
-		goto out_unlock;
+		if (!imap_spans_range(&imap, offset_fsb, end_fsb))
+			goto out_unlock;
+	}
+
+	/*
+	 * For overwrite only I/O, we cannot convert unwritten extents without
+	 * requiring sub-block zeroing.  This can only be done under an
+	 * exclusive IOLOCK, hence return -EAGAIN if this is not a written
+	 * extent to tell the caller to try again.
+	 */
+	if (flags & IOMAP_OVERWRITE_ONLY) {
+		error = -EAGAIN;
+		if (imap.br_state != XFS_EXT_NORM &&
+		    ((offset & mp->m_blockmask) ||
+		     ((offset + length) & mp->m_blockmask)))
+			goto out_unlock;
 	}
 
 	xfs_iunlock(ip, lockmode);
@@ -801,7 +815,7 @@ xfs_direct_write_iomap_begin(
 
 allocate_blocks:
 	error = -EAGAIN;
-	if (flags & IOMAP_NOWAIT)
+	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY))
 		goto out_unlock;
 
 	/*
-- 
2.29.2

Re: [PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Dave Chinner]

On Thu, Jan 21, 2021 at 09:59:06AM +0100, Christoph Hellwig wrote:
> From: Dave Chinner <dchinner@redhat.com>
> 
> Attempt shared locking for unaligned DIO, but only if the the
> underlying extent is already allocated and in written state. On
> failure, retry with the existing exclusive locking.
> 
> Test case is fio randrw of 512 byte IOs using AIO and an iodepth of
> 32 IOs.
> 
> Vanilla:
> 
>   READ: bw=4560KiB/s (4670kB/s), 4560KiB/s-4560KiB/s (4670kB/s-4670kB/s), io=134MiB (140MB), run=30001-30001msec
>   WRITE: bw=4567KiB/s (4676kB/s), 4567KiB/s-4567KiB/s (4676kB/s-4676kB/s), io=134MiB (140MB), run=30001-30001msec
> 
> Patched:
>    READ: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1127MiB (1182MB), run=30002-30002msec
>   WRITE: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1128MiB (1183MB), run=30002-30002msec
> 
> That's an improvement from ~18k IOPS to a ~150k IOPS, which is
> about the IOPS limit of the VM block device setup I'm testing on.
> 
> 4kB block IO comparison:
> 
>    READ: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8868MiB (9299MB), run=30002-30002msec
>   WRITE: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8878MiB (9309MB), run=30002-30002msec
> 
> Which is ~150k IOPS, same as what the test gets for sub-block
> AIO+DIO writes with this patch.
> 
> Signed-off-by: Dave Chinner <dchinner@redhat.com>
> [hch: rebased, split unaligned from nowait]
> Signed-off-by: Christoph Hellwig <hch@lst.de>

Looks good, minor nit:

> +	/*
> +	 * For overwrite only I/O, we cannot convert unwritten extents without
> +	 * requiring sub-block zeroing.  This can only be done under an
> +	 * exclusive IOLOCK, hence return -EAGAIN if this is not a written
> +	 * extent to tell the caller to try again.
> +	 */
> +	if (flags & IOMAP_OVERWRITE_ONLY) {
> +		error = -EAGAIN;
> +		if (imap.br_state != XFS_EXT_NORM &&
> +		    ((offset & mp->m_blockmask) ||
> +		     ((offset + length) & mp->m_blockmask)))
> +			goto out_unlock;

Why not use the ((offset | length) & mp->blockmask) form of
alignment checking here?

Other than that,

Reviewed-by: Dave Chinner <dchinner@redhat.com>

-- 
Dave Chinner
david@fromorbit.com

Re: [PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Christoph Hellwig]

On Thu, Jan 21, 2021 at 08:35:49PM +1100, Dave Chinner wrote:
> Why not use the ((offset | length) & mp->blockmask) form of
> alignment checking here?

Sure. I'ĺl switch to that.

> Reviewed-by: Dave Chinner <dchinner@redhat.com>

Given that the original patch and thus credit is yours this doesn't
make sense to add, though.

Re: [PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Brian Foster]

On Thu, Jan 21, 2021 at 09:59:06AM +0100, Christoph Hellwig wrote:
> From: Dave Chinner <dchinner@redhat.com>
> 
> Attempt shared locking for unaligned DIO, but only if the the
> underlying extent is already allocated and in written state. On
> failure, retry with the existing exclusive locking.
> 
> Test case is fio randrw of 512 byte IOs using AIO and an iodepth of
> 32 IOs.
> 
> Vanilla:
> 
>   READ: bw=4560KiB/s (4670kB/s), 4560KiB/s-4560KiB/s (4670kB/s-4670kB/s), io=134MiB (140MB), run=30001-30001msec
>   WRITE: bw=4567KiB/s (4676kB/s), 4567KiB/s-4567KiB/s (4676kB/s-4676kB/s), io=134MiB (140MB), run=30001-30001msec
> 
> Patched:
>    READ: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1127MiB (1182MB), run=30002-30002msec
>   WRITE: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1128MiB (1183MB), run=30002-30002msec
> 
> That's an improvement from ~18k IOPS to a ~150k IOPS, which is
> about the IOPS limit of the VM block device setup I'm testing on.
> 
> 4kB block IO comparison:
> 
>    READ: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8868MiB (9299MB), run=30002-30002msec
>   WRITE: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8878MiB (9309MB), run=30002-30002msec
> 
> Which is ~150k IOPS, same as what the test gets for sub-block
> AIO+DIO writes with this patch.
> 
> Signed-off-by: Dave Chinner <dchinner@redhat.com>
> [hch: rebased, split unaligned from nowait]
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> ---
>  fs/xfs/xfs_file.c  | 87 ++++++++++++++++++++++++++++++++--------------
>  fs/xfs/xfs_iomap.c | 30 +++++++++++-----
>  2 files changed, 83 insertions(+), 34 deletions(-)
> 
> diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> index b181db42f2f32f..33899a5cca53f9 100644
> --- a/fs/xfs/xfs_file.c
> +++ b/fs/xfs/xfs_file.c
> @@ -544,22 +544,35 @@ xfs_file_dio_write_aligned(
>  /*
>   * Handle block unaligned direct IO writes
>   *
> - * In most cases direct IO writes will be done holding IOLOCK_SHARED, allowing
> - * them to be done in parallel with reads and other direct IO writes.  However,
> - * if the I/O is not aligned to filesystem blocks, the direct I/O layer may
> - * need to do sub-block zeroing and that requires serialisation against other
> - * direct I/Os to the same block. In this case we need to serialise the
> - * submission of the unaligned I/Os so that we don't get racing block zeroing in
> - * the dio layer.
> + * In most cases direct IO writes will be done holding IOLOCK_SHARED
> + * allowing them to be done in parallel with reads and other direct IO writes.
> + * However, if the IO is not aligned to filesystem blocks, the direct IO layer
> + * may need to do sub-block zeroing and that requires serialisation against other
> + * direct IOs to the same block. In the case where sub-block zeroing is not
> + * required, we can do concurrent sub-block dios to the same block successfully.
>   *
> - * To provide the same serialisation for AIO, we also need to wait for
> + * Hence we have two cases here - the shared, optimisitic fast path for written
> + * extents, and everything else that needs exclusive IO path access across the
> + * entire IO.
> + *
> + * For the first case, we do all the checks we need at the mapping layer in the
> + * DIO code as part of the existing NOWAIT infrastructure. Hence all we need to
> + * do to support concurrent subblock dio is first try a non-blocking submission.
> + * If that returns -EAGAIN, then we simply repeat the IO submission with full
> + * IO exclusivity guaranteed so that we avoid racing sub-block zeroing.
> + *

The above paragraph still implicitly refers to the original NOWAIT based
implementation. I'd suggest to tweak it to something like:

"The mapping layer of the dio code performs all the checks required to
distinguish between the shared (overwrite) and exclusive cases. Hence to
support concurrent unaligned dio, first submit the request in overwrite
only mode. If that returns -EAGAIN, sub-block zeroing is required.
Repeat the submission with full IO exclusivity to avoid races."

That aside, I still find the single mapping requirement a bit
unfortunate, but otherwise the code LGTM:

Reviewed-by: Brian Foster <bfoster@redhat.com>

> + * The only wrinkle in this case is that the iomap DIO code always does
> + * partial tail sub-block zeroing for post-EOF writes. Hence for any IO that
> + * _ends_ past the current EOF we need to run with full exclusivity. Note that
> + * we also check for the start of IO being beyond EOF because then zeroing
> + * between the old EOF and the start of the IO is required and that also
> + * requires exclusivity. Hence we avoid lock cycles and blocking under
> + * IOCB_NOWAIT for this situation, too.
> + *
> + * To provide the exclusivity required when using AIO, we also need to wait for
>   * outstanding IOs to complete so that unwritten extent conversion is completed
>   * before we try to map the overlapping block. This is currently implemented by
>   * hitting it with a big hammer (i.e. inode_dio_wait()).
> - *
> - * This means that unaligned dio writes always block. There is no "nowait" fast
> - * path in this code - if IOCB_NOWAIT is set we simply return -EAGAIN up front
> - * and we don't have to worry about that anymore.
>   */
>  static noinline ssize_t
>  xfs_file_dio_write_unaligned(
> @@ -567,13 +580,27 @@ xfs_file_dio_write_unaligned(
>  	struct kiocb		*iocb,
>  	struct iov_iter		*from)
>  {
> -	int			iolock = XFS_IOLOCK_EXCL;
> +	size_t			isize = i_size_read(VFS_I(ip));
> +	size_t			count = iov_iter_count(from);
> +	int			iolock = XFS_IOLOCK_SHARED;
> +	unsigned int		flags = IOMAP_DIO_OVERWRITE_ONLY;
>  	ssize_t			ret;
>  
> -	/* unaligned dio always waits, bail */
> -	if (iocb->ki_flags & IOCB_NOWAIT)
> -		return -EAGAIN;
> -	xfs_ilock(ip, iolock);
> +	/*
> +	 * Extending writes need exclusivity because of the sub-block zeroing
> +	 * that the DIO code always does for partial tail blocks beyond EOF.
> +	 */
> +	if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
> +retry_exclusive:
> +		if (iocb->ki_flags & IOCB_NOWAIT)
> +			return -EAGAIN;
> +		iolock = XFS_IOLOCK_EXCL;
> +		flags = IOMAP_DIO_FORCE_WAIT;
> +	}
> +
> +	ret = xfs_ilock_iocb(iocb, iolock);
> +	if (ret)
> +		return ret;
>  
>  	/*
>  	 * We can't properly handle unaligned direct I/O to reflink files yet,
> @@ -590,19 +617,27 @@ xfs_file_dio_write_unaligned(
>  		goto out_unlock;
>  
>  	/*
> -	 * If we are doing unaligned I/O, we can't allow any other overlapping
> -	 * I/O in-flight at the same time or we risk data corruption. Wait for
> -	 * all other I/O to drain before we submit.
> +	 * If we are doing exclusive unaligned IO, we can't allow any other
> +	 * overlapping IO in-flight at the same time or we risk data corruption.
> +	 * Wait for all other IO to drain before we submit.
>  	 */
> -	inode_dio_wait(VFS_I(ip));
> +	if (flags & IOMAP_DIO_FORCE_WAIT)
> +		inode_dio_wait(VFS_I(ip));
>  
> -	/*
> -	 * This must be the only I/O in-flight. Wait on it before we release the
> -	 * iolock to prevent subsequent overlapping I/O.
> -	 */
>  	trace_xfs_file_direct_write(iocb, from);
>  	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
> -			   &xfs_dio_write_ops, IOMAP_DIO_FORCE_WAIT);
> +			   &xfs_dio_write_ops, flags);
> +	/*
> +	 * Retry unaligned IO with exclusive blocking semantics if the DIO
> +	 * layer rejected it for mapping or locking reasons. If we are doing
> +	 * nonblocking user IO, propagate the error.
> +	 */
> +	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
> +		ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
> +		xfs_iunlock(ip, iolock);
> +		goto retry_exclusive;
> +	}
> +
>  out_unlock:
>  	if (iolock)
>  		xfs_iunlock(ip, iolock);
> diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
> index 7b9ff824e82d48..596af78f910596 100644
> --- a/fs/xfs/xfs_iomap.c
> +++ b/fs/xfs/xfs_iomap.c
> @@ -784,15 +784,29 @@ xfs_direct_write_iomap_begin(
>  		goto allocate_blocks;
>  
>  	/*
> -	 * NOWAIT IO needs to span the entire requested IO with a single map so
> -	 * that we avoid partial IO failures due to the rest of the IO range not
> -	 * covered by this map triggering an EAGAIN condition when it is
> -	 * subsequently mapped and aborting the IO.
> +	 * NOWAIT and OVERWRITE needs to span the entire requested IO with a
> +	 * single map so that we avoid partial IO failures due to the rest of
> +	 * the IO range not covered by this map triggering an EAGAIN condition
> +	 * when it is subsequently mapped and aborting the IO.
>  	 */
> -	if ((flags & IOMAP_NOWAIT) &&
> -	    !imap_spans_range(&imap, offset_fsb, end_fsb)) {
> +	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY)) {
>  		error = -EAGAIN;
> -		goto out_unlock;
> +		if (!imap_spans_range(&imap, offset_fsb, end_fsb))
> +			goto out_unlock;
> +	}
> +
> +	/*
> +	 * For overwrite only I/O, we cannot convert unwritten extents without
> +	 * requiring sub-block zeroing.  This can only be done under an
> +	 * exclusive IOLOCK, hence return -EAGAIN if this is not a written
> +	 * extent to tell the caller to try again.
> +	 */
> +	if (flags & IOMAP_OVERWRITE_ONLY) {
> +		error = -EAGAIN;
> +		if (imap.br_state != XFS_EXT_NORM &&
> +		    ((offset & mp->m_blockmask) ||
> +		     ((offset + length) & mp->m_blockmask)))
> +			goto out_unlock;
>  	}
>  
>  	xfs_iunlock(ip, lockmode);
> @@ -801,7 +815,7 @@ xfs_direct_write_iomap_begin(
>  
>  allocate_blocks:
>  	error = -EAGAIN;
> -	if (flags & IOMAP_NOWAIT)
> +	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY))
>  		goto out_unlock;
>  
>  	/*
> -- 
> 2.29.2
> 

Re: [PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Darrick J. Wong]

On Thu, Jan 21, 2021 at 08:33:34AM -0500, Brian Foster wrote:
> On Thu, Jan 21, 2021 at 09:59:06AM +0100, Christoph Hellwig wrote:
> > From: Dave Chinner <dchinner@redhat.com>
> > 
> > Attempt shared locking for unaligned DIO, but only if the the
> > underlying extent is already allocated and in written state. On
> > failure, retry with the existing exclusive locking.
> > 
> > Test case is fio randrw of 512 byte IOs using AIO and an iodepth of
> > 32 IOs.
> > 
> > Vanilla:
> > 
> >   READ: bw=4560KiB/s (4670kB/s), 4560KiB/s-4560KiB/s (4670kB/s-4670kB/s), io=134MiB (140MB), run=30001-30001msec
> >   WRITE: bw=4567KiB/s (4676kB/s), 4567KiB/s-4567KiB/s (4676kB/s-4676kB/s), io=134MiB (140MB), run=30001-30001msec
> > 
> > Patched:
> >    READ: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1127MiB (1182MB), run=30002-30002msec
> >   WRITE: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1128MiB (1183MB), run=30002-30002msec
> > 
> > That's an improvement from ~18k IOPS to a ~150k IOPS, which is
> > about the IOPS limit of the VM block device setup I'm testing on.
> > 
> > 4kB block IO comparison:
> > 
> >    READ: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8868MiB (9299MB), run=30002-30002msec
> >   WRITE: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8878MiB (9309MB), run=30002-30002msec
> > 
> > Which is ~150k IOPS, same as what the test gets for sub-block
> > AIO+DIO writes with this patch.
> > 
> > Signed-off-by: Dave Chinner <dchinner@redhat.com>
> > [hch: rebased, split unaligned from nowait]
> > Signed-off-by: Christoph Hellwig <hch@lst.de>
> > ---
> >  fs/xfs/xfs_file.c  | 87 ++++++++++++++++++++++++++++++++--------------
> >  fs/xfs/xfs_iomap.c | 30 +++++++++++-----
> >  2 files changed, 83 insertions(+), 34 deletions(-)
> > 
> > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> > index b181db42f2f32f..33899a5cca53f9 100644
> > --- a/fs/xfs/xfs_file.c
> > +++ b/fs/xfs/xfs_file.c
> > @@ -544,22 +544,35 @@ xfs_file_dio_write_aligned(
> >  /*
> >   * Handle block unaligned direct IO writes
> >   *
> > - * In most cases direct IO writes will be done holding IOLOCK_SHARED, allowing
> > - * them to be done in parallel with reads and other direct IO writes.  However,
> > - * if the I/O is not aligned to filesystem blocks, the direct I/O layer may
> > - * need to do sub-block zeroing and that requires serialisation against other
> > - * direct I/Os to the same block. In this case we need to serialise the
> > - * submission of the unaligned I/Os so that we don't get racing block zeroing in
> > - * the dio layer.
> > + * In most cases direct IO writes will be done holding IOLOCK_SHARED
> > + * allowing them to be done in parallel with reads and other direct IO writes.
> > + * However, if the IO is not aligned to filesystem blocks, the direct IO layer
> > + * may need to do sub-block zeroing and that requires serialisation against other
> > + * direct IOs to the same block. In the case where sub-block zeroing is not
> > + * required, we can do concurrent sub-block dios to the same block successfully.
> >   *
> > - * To provide the same serialisation for AIO, we also need to wait for
> > + * Hence we have two cases here - the shared, optimisitic fast path for written

"optimistic"

> > + * extents, and everything else that needs exclusive IO path access across the
> > + * entire IO.
> > + *
> > + * For the first case, we do all the checks we need at the mapping layer in the
> > + * DIO code as part of the existing NOWAIT infrastructure. Hence all we need to
> > + * do to support concurrent subblock dio is first try a non-blocking submission.
> > + * If that returns -EAGAIN, then we simply repeat the IO submission with full
> > + * IO exclusivity guaranteed so that we avoid racing sub-block zeroing.
> > + *
> 
> The above paragraph still implicitly refers to the original NOWAIT based
> implementation. I'd suggest to tweak it to something like:
> 
> "The mapping layer of the dio code performs all the checks required to
> distinguish between the shared (overwrite) and exclusive cases. Hence to
> support concurrent unaligned dio, first submit the request in overwrite
> only mode. If that returns -EAGAIN, sub-block zeroing is required.
> Repeat the submission with full IO exclusivity to avoid races."

FWIW I like this version better, because the fact that we reuse the
existing 'nowait' switching to handle unaligned direct writes is a minor
implementation detail here.

I think I'm about ready to RVB this, but there are enough small loose
ends that I'll wait for the next version. :)

--D

> That aside, I still find the single mapping requirement a bit
> unfortunate, but otherwise the code LGTM:
> 
> Reviewed-by: Brian Foster <bfoster@redhat.com>
> 
> > + * The only wrinkle in this case is that the iomap DIO code always does
> > + * partial tail sub-block zeroing for post-EOF writes. Hence for any IO that
> > + * _ends_ past the current EOF we need to run with full exclusivity. Note that
> > + * we also check for the start of IO being beyond EOF because then zeroing
> > + * between the old EOF and the start of the IO is required and that also
> > + * requires exclusivity. Hence we avoid lock cycles and blocking under
> > + * IOCB_NOWAIT for this situation, too.
> > + *
> > + * To provide the exclusivity required when using AIO, we also need to wait for
> >   * outstanding IOs to complete so that unwritten extent conversion is completed
> >   * before we try to map the overlapping block. This is currently implemented by
> >   * hitting it with a big hammer (i.e. inode_dio_wait()).
> > - *
> > - * This means that unaligned dio writes always block. There is no "nowait" fast
> > - * path in this code - if IOCB_NOWAIT is set we simply return -EAGAIN up front
> > - * and we don't have to worry about that anymore.
> >   */
> >  static noinline ssize_t
> >  xfs_file_dio_write_unaligned(
> > @@ -567,13 +580,27 @@ xfs_file_dio_write_unaligned(
> >  	struct kiocb		*iocb,
> >  	struct iov_iter		*from)
> >  {
> > -	int			iolock = XFS_IOLOCK_EXCL;
> > +	size_t			isize = i_size_read(VFS_I(ip));
> > +	size_t			count = iov_iter_count(from);
> > +	int			iolock = XFS_IOLOCK_SHARED;
> > +	unsigned int		flags = IOMAP_DIO_OVERWRITE_ONLY;
> >  	ssize_t			ret;
> >  
> > -	/* unaligned dio always waits, bail */
> > -	if (iocb->ki_flags & IOCB_NOWAIT)
> > -		return -EAGAIN;
> > -	xfs_ilock(ip, iolock);
> > +	/*
> > +	 * Extending writes need exclusivity because of the sub-block zeroing
> > +	 * that the DIO code always does for partial tail blocks beyond EOF.
> > +	 */
> > +	if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
> > +retry_exclusive:
> > +		if (iocb->ki_flags & IOCB_NOWAIT)
> > +			return -EAGAIN;
> > +		iolock = XFS_IOLOCK_EXCL;
> > +		flags = IOMAP_DIO_FORCE_WAIT;
> > +	}
> > +
> > +	ret = xfs_ilock_iocb(iocb, iolock);
> > +	if (ret)
> > +		return ret;
> >  
> >  	/*
> >  	 * We can't properly handle unaligned direct I/O to reflink files yet,
> > @@ -590,19 +617,27 @@ xfs_file_dio_write_unaligned(
> >  		goto out_unlock;
> >  
> >  	/*
> > -	 * If we are doing unaligned I/O, we can't allow any other overlapping
> > -	 * I/O in-flight at the same time or we risk data corruption. Wait for
> > -	 * all other I/O to drain before we submit.
> > +	 * If we are doing exclusive unaligned IO, we can't allow any other
> > +	 * overlapping IO in-flight at the same time or we risk data corruption.
> > +	 * Wait for all other IO to drain before we submit.
> >  	 */
> > -	inode_dio_wait(VFS_I(ip));
> > +	if (flags & IOMAP_DIO_FORCE_WAIT)
> > +		inode_dio_wait(VFS_I(ip));
> >  
> > -	/*
> > -	 * This must be the only I/O in-flight. Wait on it before we release the
> > -	 * iolock to prevent subsequent overlapping I/O.
> > -	 */
> >  	trace_xfs_file_direct_write(iocb, from);
> >  	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
> > -			   &xfs_dio_write_ops, IOMAP_DIO_FORCE_WAIT);
> > +			   &xfs_dio_write_ops, flags);
> > +	/*
> > +	 * Retry unaligned IO with exclusive blocking semantics if the DIO
> > +	 * layer rejected it for mapping or locking reasons. If we are doing
> > +	 * nonblocking user IO, propagate the error.
> > +	 */
> > +	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
> > +		ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
> > +		xfs_iunlock(ip, iolock);
> > +		goto retry_exclusive;
> > +	}
> > +
> >  out_unlock:
> >  	if (iolock)
> >  		xfs_iunlock(ip, iolock);
> > diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
> > index 7b9ff824e82d48..596af78f910596 100644
> > --- a/fs/xfs/xfs_iomap.c
> > +++ b/fs/xfs/xfs_iomap.c
> > @@ -784,15 +784,29 @@ xfs_direct_write_iomap_begin(
> >  		goto allocate_blocks;
> >  
> >  	/*
> > -	 * NOWAIT IO needs to span the entire requested IO with a single map so
> > -	 * that we avoid partial IO failures due to the rest of the IO range not
> > -	 * covered by this map triggering an EAGAIN condition when it is
> > -	 * subsequently mapped and aborting the IO.
> > +	 * NOWAIT and OVERWRITE needs to span the entire requested IO with a
> > +	 * single map so that we avoid partial IO failures due to the rest of
> > +	 * the IO range not covered by this map triggering an EAGAIN condition
> > +	 * when it is subsequently mapped and aborting the IO.
> >  	 */
> > -	if ((flags & IOMAP_NOWAIT) &&
> > -	    !imap_spans_range(&imap, offset_fsb, end_fsb)) {
> > +	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY)) {
> >  		error = -EAGAIN;
> > -		goto out_unlock;
> > +		if (!imap_spans_range(&imap, offset_fsb, end_fsb))
> > +			goto out_unlock;
> > +	}
> > +
> > +	/*
> > +	 * For overwrite only I/O, we cannot convert unwritten extents without
> > +	 * requiring sub-block zeroing.  This can only be done under an
> > +	 * exclusive IOLOCK, hence return -EAGAIN if this is not a written
> > +	 * extent to tell the caller to try again.
> > +	 */
> > +	if (flags & IOMAP_OVERWRITE_ONLY) {
> > +		error = -EAGAIN;
> > +		if (imap.br_state != XFS_EXT_NORM &&
> > +		    ((offset & mp->m_blockmask) ||
> > +		     ((offset + length) & mp->m_blockmask)))
> > +			goto out_unlock;
> >  	}
> >  
> >  	xfs_iunlock(ip, lockmode);
> > @@ -801,7 +815,7 @@ xfs_direct_write_iomap_begin(
> >  
> >  allocate_blocks:
> >  	error = -EAGAIN;
> > -	if (flags & IOMAP_NOWAIT)
> > +	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY))
> >  		goto out_unlock;
> >  
> >  	/*
> > -- 
> > 2.29.2
> > 
> 

reduce sub-block DIO serialisation v2

[Christoph Hellwig]

This takes the approach from Dave, but adds a new flag instead of abusing
the nowait one, and keeps a simpler calling convention for iomap_dio_rw.

Changes since v2:
 - rename the new flags
 - add an EOF check for subblock I/O
 - minor cleanups

[PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Christoph Hellwig]

From: Dave Chinner <dchinner@redhat.com>

Attempt shared locking for unaligned DIO, but only if the the
underlying extent is already allocated and in written state. On
failure, retry with the existing exclusive locking.

Test case is fio randrw of 512 byte IOs using AIO and an iodepth of
32 IOs.

Vanilla:

  READ: bw=4560KiB/s (4670kB/s), 4560KiB/s-4560KiB/s (4670kB/s-4670kB/s), io=134MiB (140MB), run=30001-30001msec
  WRITE: bw=4567KiB/s (4676kB/s), 4567KiB/s-4567KiB/s (4676kB/s-4676kB/s), io=134MiB (140MB), run=30001-30001msec

Patched:
   READ: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1127MiB (1182MB), run=30002-30002msec
  WRITE: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1128MiB (1183MB), run=30002-30002msec

That's an improvement from ~18k IOPS to a ~150k IOPS, which is
about the IOPS limit of the VM block device setup I'm testing on.

4kB block IO comparison:

   READ: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8868MiB (9299MB), run=30002-30002msec
  WRITE: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8878MiB (9309MB), run=30002-30002msec

Which is ~150k IOPS, same as what the test gets for sub-block
AIO+DIO writes with this patch.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
[hch: rebased, split unaligned from nowait]
Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 fs/xfs/xfs_file.c  | 87 ++++++++++++++++++++++++++++++++--------------
 fs/xfs/xfs_iomap.c | 31 ++++++++++++-----
 2 files changed, 84 insertions(+), 34 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index b181db42f2f32f..4e475e750148db 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -544,22 +544,35 @@ xfs_file_dio_write_aligned(
 /*
  * Handle block unaligned direct IO writes
  *
- * In most cases direct IO writes will be done holding IOLOCK_SHARED, allowing
- * them to be done in parallel with reads and other direct IO writes.  However,
- * if the I/O is not aligned to filesystem blocks, the direct I/O layer may
- * need to do sub-block zeroing and that requires serialisation against other
- * direct I/Os to the same block. In this case we need to serialise the
- * submission of the unaligned I/Os so that we don't get racing block zeroing in
- * the dio layer.
+ * In most cases direct IO writes will be done holding IOLOCK_SHARED
+ * allowing them to be done in parallel with reads and other direct IO writes.
+ * However, if the IO is not aligned to filesystem blocks, the direct IO layer
+ * may need to do sub-block zeroing and that requires serialisation against other
+ * direct IOs to the same block. In the case where sub-block zeroing is not
+ * required, we can do concurrent sub-block dios to the same block successfully.
  *
- * To provide the same serialisation for AIO, we also need to wait for
+ * Hence we have two cases here - the shared, optimisitic fast path for written
+ * extents, and everything else that needs exclusive IO path access across the
+ * entire IO.
+ *
+ * For the first case, we do all the checks we need at the mapping layer in the
+ * DIO code as part of the existing NOWAIT infrastructure. Hence all we need to
+ * do to support concurrent subblock dio is first try a non-blocking submission.
+ * If that returns -EAGAIN, then we simply repeat the IO submission with full
+ * IO exclusivity guaranteed so that we avoid racing sub-block zeroing.
+ *
+ * The only wrinkle in this case is that the iomap DIO code always does
+ * partial tail sub-block zeroing for post-EOF writes. Hence for any IO that
+ * _ends_ past the current EOF we need to run with full exclusivity. Note that
+ * we also check for the start of IO being beyond EOF because then zeroing
+ * between the old EOF and the start of the IO is required and that also
+ * requires exclusivity. Hence we avoid lock cycles and blocking under
+ * IOCB_NOWAIT for this situation, too.
+ *
+ * To provide the exclusivity required when using AIO, we also need to wait for
  * outstanding IOs to complete so that unwritten extent conversion is completed
  * before we try to map the overlapping block. This is currently implemented by
  * hitting it with a big hammer (i.e. inode_dio_wait()).
- *
- * This means that unaligned dio writes always block. There is no "nowait" fast
- * path in this code - if IOCB_NOWAIT is set we simply return -EAGAIN up front
- * and we don't have to worry about that anymore.
  */
 static noinline ssize_t
 xfs_file_dio_write_unaligned(
@@ -567,13 +580,27 @@ xfs_file_dio_write_unaligned(
 	struct kiocb		*iocb,
 	struct iov_iter		*from)
 {
-	int			iolock = XFS_IOLOCK_EXCL;
+	size_t			isize = i_size_read(VFS_I(ip));
+	size_t			count = iov_iter_count(from);
+	int			iolock = XFS_IOLOCK_SHARED;
+	unsigned int		flags = IOMAP_DIO_UNALIGNED;
 	ssize_t			ret;
 
-	/* unaligned dio always waits, bail */
-	if (iocb->ki_flags & IOCB_NOWAIT)
-		return -EAGAIN;
-	xfs_ilock(ip, iolock);
+	/*
+	 * Extending writes need exclusivity because of the sub-block zeroing
+	 * that the DIO code always does for partial tail blocks beyond EOF.
+	 */
+	if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
+retry_exclusive:
+		if (iocb->ki_flags & IOCB_NOWAIT)
+			return -EAGAIN;
+		iolock = XFS_IOLOCK_EXCL;
+		flags = IOMAP_DIO_FORCE_WAIT;
+	}
+
+	ret = xfs_ilock_iocb(iocb, iolock);
+	if (ret)
+		return ret;
 
 	/*
 	 * We can't properly handle unaligned direct I/O to reflink files yet,
@@ -590,19 +617,27 @@ xfs_file_dio_write_unaligned(
 		goto out_unlock;
 
 	/*
-	 * If we are doing unaligned I/O, we can't allow any other overlapping
-	 * I/O in-flight at the same time or we risk data corruption. Wait for
-	 * all other I/O to drain before we submit.
+	 * If we are doing exclusive unaligned IO, we can't allow any other
+	 * overlapping IO in-flight at the same time or we risk data corruption.
+	 * Wait for all other IO to drain before we submit.
 	 */
-	inode_dio_wait(VFS_I(ip));
+	if (!(flags & IOMAP_DIO_UNALIGNED))
+		inode_dio_wait(VFS_I(ip));
 
-	/*
-	 * This must be the only I/O in-flight. Wait on it before we release the
-	 * iolock to prevent subsequent overlapping I/O.
-	 */
 	trace_xfs_file_direct_write(iocb, from);
 	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
-			   &xfs_dio_write_ops, IOMAP_DIO_FORCE_WAIT);
+			   &xfs_dio_write_ops, flags);
+	/*
+	 * Retry unaligned IO with exclusive blocking semantics if the DIO
+	 * layer rejected it for mapping or locking reasons. If we are doing
+	 * nonblocking user IO, propagate the error.
+	 */
+	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
+		ASSERT(flags & IOMAP_DIO_UNALIGNED);
+		xfs_iunlock(ip, iolock);
+		goto retry_exclusive;
+	}
+
 out_unlock:
 	if (iolock)
 		xfs_iunlock(ip, iolock);
diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
index 7b9ff824e82d48..dc8c86e98b99bf 100644
--- a/fs/xfs/xfs_iomap.c
+++ b/fs/xfs/xfs_iomap.c
@@ -784,15 +784,30 @@ xfs_direct_write_iomap_begin(
 		goto allocate_blocks;
 
 	/*
-	 * NOWAIT IO needs to span the entire requested IO with a single map so
-	 * that we avoid partial IO failures due to the rest of the IO range not
-	 * covered by this map triggering an EAGAIN condition when it is
-	 * subsequently mapped and aborting the IO.
+	 * NOWAIT and unaligned IO needs to span the entire requested IO with a
+	 * single map so that we avoid partial IO failures due to the rest of
+	 * the IO range not covered by this map triggering an EAGAIN condition
+	 * when it is subsequently mapped and aborting the IO.
 	 */
-	if ((flags & IOMAP_NOWAIT) &&
-	    !imap_spans_range(&imap, offset_fsb, end_fsb)) {
+	if (flags & (IOMAP_NOWAIT | IOMAP_UNALIGNED)) {
 		error = -EAGAIN;
-		goto out_unlock;
+		if (!imap_spans_range(&imap, offset_fsb, end_fsb))
+			goto out_unlock;
+	}
+
+	/*
+	 * For unsigned I/O we can't convert an unwritten extents if the I/O is
+	 * not block size aligned, as such a conversion would have to do
+	 * sub-block zeroing, and that can only be done under an exclusive
+	 * IOLOCK. Hence if this is not a written extent, return EAGAIN to tell
+	 * the caller to try again.
+	 */
+	if (flags & IOMAP_UNALIGNED) {
+		error = -EAGAIN;
+		if (imap.br_state != XFS_EXT_NORM &&
+		    ((offset & mp->m_blockmask) ||
+		     ((offset + length) & mp->m_blockmask)))
+			goto out_unlock;
 	}
 
 	xfs_iunlock(ip, lockmode);
@@ -801,7 +816,7 @@ xfs_direct_write_iomap_begin(
 
 allocate_blocks:
 	error = -EAGAIN;
-	if (flags & IOMAP_NOWAIT)
+	if (flags & (IOMAP_NOWAIT | IOMAP_UNALIGNED))
 		goto out_unlock;
 
 	/*
-- 
2.29.2

Re: [PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Dave Chinner]

On Mon, Jan 18, 2021 at 08:35:16PM +0100, Christoph Hellwig wrote:
> From: Dave Chinner <dchinner@redhat.com>
> 
> Attempt shared locking for unaligned DIO, but only if the the
> underlying extent is already allocated and in written state. On
> failure, retry with the existing exclusive locking.
....
> @@ -590,19 +617,27 @@ xfs_file_dio_write_unaligned(
>  		goto out_unlock;
>  
>  	/*
> -	 * If we are doing unaligned I/O, we can't allow any other overlapping
> -	 * I/O in-flight at the same time or we risk data corruption. Wait for
> -	 * all other I/O to drain before we submit.
> +	 * If we are doing exclusive unaligned IO, we can't allow any other
> +	 * overlapping IO in-flight at the same time or we risk data corruption.
> +	 * Wait for all other IO to drain before we submit.
>  	 */
> -	inode_dio_wait(VFS_I(ip));
> +	if (!(flags & IOMAP_DIO_UNALIGNED))
> +		inode_dio_wait(VFS_I(ip));
>  
> -	/*
> -	 * This must be the only I/O in-flight. Wait on it before we release the
> -	 * iolock to prevent subsequent overlapping I/O.
> -	 */
>  	trace_xfs_file_direct_write(iocb, from);
>  	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
> -			   &xfs_dio_write_ops, IOMAP_DIO_FORCE_WAIT);
> +			   &xfs_dio_write_ops, flags);
> +	/*
> +	 * Retry unaligned IO with exclusive blocking semantics if the DIO
> +	 * layer rejected it for mapping or locking reasons. If we are doing
> +	 * nonblocking user IO, propagate the error.
> +	 */
> +	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
> +		ASSERT(flags & IOMAP_DIO_UNALIGNED);
> +		xfs_iunlock(ip, iolock);
> +		goto retry_exclusive;
> +	}
> +
>  out_unlock:
>  	if (iolock)
>  		xfs_iunlock(ip, iolock);

Do we ever get here without holding the iolock anymore?

> diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
> index 7b9ff824e82d48..dc8c86e98b99bf 100644
> --- a/fs/xfs/xfs_iomap.c
> +++ b/fs/xfs/xfs_iomap.c
> @@ -784,15 +784,30 @@ xfs_direct_write_iomap_begin(
>  		goto allocate_blocks;
>  
>  	/*
> -	 * NOWAIT IO needs to span the entire requested IO with a single map so
> -	 * that we avoid partial IO failures due to the rest of the IO range not
> -	 * covered by this map triggering an EAGAIN condition when it is
> -	 * subsequently mapped and aborting the IO.
> +	 * NOWAIT and unaligned IO needs to span the entire requested IO with a
> +	 * single map so that we avoid partial IO failures due to the rest of
> +	 * the IO range not covered by this map triggering an EAGAIN condition
> +	 * when it is subsequently mapped and aborting the IO.
>  	 */
> -	if ((flags & IOMAP_NOWAIT) &&
> -	    !imap_spans_range(&imap, offset_fsb, end_fsb)) {
> +	if (flags & (IOMAP_NOWAIT | IOMAP_UNALIGNED)) {
>  		error = -EAGAIN;
> -		goto out_unlock;
> +		if (!imap_spans_range(&imap, offset_fsb, end_fsb))
> +			goto out_unlock;
> +	}
> +
> +	/*
> +	 * For unsigned I/O we can't convert an unwritten extents if the I/O is
> +	 * not block size aligned, as such a conversion would have to do
> +	 * sub-block zeroing, and that can only be done under an exclusive
> +	 * IOLOCK. Hence if this is not a written extent, return EAGAIN to tell
> +	 * the caller to try again.
> +	 */

A few typos in that comment :)

	/*
	 * For unaligned IO, we cannot convert unwritten extents without
	 * requiring sub-block zeroing. This can only be done under an exclusive
	 * IOLOCK, hence return -EAGAIN if this is not a written extent to tell
	 * the caller to try again.
	 */

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: [PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Christoph Hellwig]

On Tue, Jan 19, 2021 at 07:55:21AM +1100, Dave Chinner wrote:
> > +			   &xfs_dio_write_ops, flags);
> > +	/*
> > +	 * Retry unaligned IO with exclusive blocking semantics if the DIO
> > +	 * layer rejected it for mapping or locking reasons. If we are doing
> > +	 * nonblocking user IO, propagate the error.
> > +	 */
> > +	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
> > +		ASSERT(flags & IOMAP_DIO_UNALIGNED);
> > +		xfs_iunlock(ip, iolock);
> > +		goto retry_exclusive;
> > +	}
> > +
> >  out_unlock:
> >  	if (iolock)
> >  		xfs_iunlock(ip, iolock);
> 
> Do we ever get here without holding the iolock anymore?

Yes, if xfs_ilock_iocb as called from xfs_file_write_checks fails.

Re: [PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Darrick J. Wong]

On Mon, Jan 18, 2021 at 08:35:16PM +0100, Christoph Hellwig wrote:
> From: Dave Chinner <dchinner@redhat.com>
> 
> Attempt shared locking for unaligned DIO, but only if the the
> underlying extent is already allocated and in written state. On
> failure, retry with the existing exclusive locking.
> 
> Test case is fio randrw of 512 byte IOs using AIO and an iodepth of
> 32 IOs.
> 
> Vanilla:
> 
>   READ: bw=4560KiB/s (4670kB/s), 4560KiB/s-4560KiB/s (4670kB/s-4670kB/s), io=134MiB (140MB), run=30001-30001msec
>   WRITE: bw=4567KiB/s (4676kB/s), 4567KiB/s-4567KiB/s (4676kB/s-4676kB/s), io=134MiB (140MB), run=30001-30001msec
> 
> Patched:
>    READ: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1127MiB (1182MB), run=30002-30002msec
>   WRITE: bw=37.6MiB/s (39.4MB/s), 37.6MiB/s-37.6MiB/s (39.4MB/s-39.4MB/s), io=1128MiB (1183MB), run=30002-30002msec
> 
> That's an improvement from ~18k IOPS to a ~150k IOPS, which is
> about the IOPS limit of the VM block device setup I'm testing on.
> 
> 4kB block IO comparison:
> 
>    READ: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8868MiB (9299MB), run=30002-30002msec
>   WRITE: bw=296MiB/s (310MB/s), 296MiB/s-296MiB/s (310MB/s-310MB/s), io=8878MiB (9309MB), run=30002-30002msec
> 
> Which is ~150k IOPS, same as what the test gets for sub-block
> AIO+DIO writes with this patch.
> 
> Signed-off-by: Dave Chinner <dchinner@redhat.com>
> [hch: rebased, split unaligned from nowait]
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> ---
>  fs/xfs/xfs_file.c  | 87 ++++++++++++++++++++++++++++++++--------------
>  fs/xfs/xfs_iomap.c | 31 ++++++++++++-----
>  2 files changed, 84 insertions(+), 34 deletions(-)
> 
> diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> index b181db42f2f32f..4e475e750148db 100644
> --- a/fs/xfs/xfs_file.c
> +++ b/fs/xfs/xfs_file.c
> @@ -544,22 +544,35 @@ xfs_file_dio_write_aligned(
>  /*
>   * Handle block unaligned direct IO writes
>   *
> - * In most cases direct IO writes will be done holding IOLOCK_SHARED, allowing
> - * them to be done in parallel with reads and other direct IO writes.  However,
> - * if the I/O is not aligned to filesystem blocks, the direct I/O layer may
> - * need to do sub-block zeroing and that requires serialisation against other
> - * direct I/Os to the same block. In this case we need to serialise the
> - * submission of the unaligned I/Os so that we don't get racing block zeroing in
> - * the dio layer.
> + * In most cases direct IO writes will be done holding IOLOCK_SHARED
> + * allowing them to be done in parallel with reads and other direct IO writes.
> + * However, if the IO is not aligned to filesystem blocks, the direct IO layer
> + * may need to do sub-block zeroing and that requires serialisation against other
> + * direct IOs to the same block. In the case where sub-block zeroing is not
> + * required, we can do concurrent sub-block dios to the same block successfully.
>   *
> - * To provide the same serialisation for AIO, we also need to wait for
> + * Hence we have two cases here - the shared, optimisitic fast path for written
> + * extents, and everything else that needs exclusive IO path access across the
> + * entire IO.
> + *
> + * For the first case, we do all the checks we need at the mapping layer in the
> + * DIO code as part of the existing NOWAIT infrastructure. Hence all we need to
> + * do to support concurrent subblock dio is first try a non-blocking submission.
> + * If that returns -EAGAIN, then we simply repeat the IO submission with full
> + * IO exclusivity guaranteed so that we avoid racing sub-block zeroing.
> + *
> + * The only wrinkle in this case is that the iomap DIO code always does
> + * partial tail sub-block zeroing for post-EOF writes. Hence for any IO that
> + * _ends_ past the current EOF we need to run with full exclusivity. Note that
> + * we also check for the start of IO being beyond EOF because then zeroing
> + * between the old EOF and the start of the IO is required and that also
> + * requires exclusivity. Hence we avoid lock cycles and blocking under
> + * IOCB_NOWAIT for this situation, too.
> + *
> + * To provide the exclusivity required when using AIO, we also need to wait for
>   * outstanding IOs to complete so that unwritten extent conversion is completed
>   * before we try to map the overlapping block. This is currently implemented by
>   * hitting it with a big hammer (i.e. inode_dio_wait()).
> - *
> - * This means that unaligned dio writes always block. There is no "nowait" fast
> - * path in this code - if IOCB_NOWAIT is set we simply return -EAGAIN up front
> - * and we don't have to worry about that anymore.
>   */
>  static noinline ssize_t
>  xfs_file_dio_write_unaligned(
> @@ -567,13 +580,27 @@ xfs_file_dio_write_unaligned(
>  	struct kiocb		*iocb,
>  	struct iov_iter		*from)
>  {
> -	int			iolock = XFS_IOLOCK_EXCL;
> +	size_t			isize = i_size_read(VFS_I(ip));
> +	size_t			count = iov_iter_count(from);
> +	int			iolock = XFS_IOLOCK_SHARED;
> +	unsigned int		flags = IOMAP_DIO_UNALIGNED;
>  	ssize_t			ret;
>  
> -	/* unaligned dio always waits, bail */
> -	if (iocb->ki_flags & IOCB_NOWAIT)
> -		return -EAGAIN;
> -	xfs_ilock(ip, iolock);
> +	/*
> +	 * Extending writes need exclusivity because of the sub-block zeroing
> +	 * that the DIO code always does for partial tail blocks beyond EOF.
> +	 */
> +	if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
> +retry_exclusive:
> +		if (iocb->ki_flags & IOCB_NOWAIT)
> +			return -EAGAIN;
> +		iolock = XFS_IOLOCK_EXCL;
> +		flags = IOMAP_DIO_FORCE_WAIT;
> +	}
> +
> +	ret = xfs_ilock_iocb(iocb, iolock);
> +	if (ret)
> +		return ret;
>  
>  	/*
>  	 * We can't properly handle unaligned direct I/O to reflink files yet,
> @@ -590,19 +617,27 @@ xfs_file_dio_write_unaligned(
>  		goto out_unlock;
>  
>  	/*
> -	 * If we are doing unaligned I/O, we can't allow any other overlapping
> -	 * I/O in-flight at the same time or we risk data corruption. Wait for
> -	 * all other I/O to drain before we submit.
> +	 * If we are doing exclusive unaligned IO, we can't allow any other
> +	 * overlapping IO in-flight at the same time or we risk data corruption.
> +	 * Wait for all other IO to drain before we submit.
>  	 */
> -	inode_dio_wait(VFS_I(ip));
> +	if (!(flags & IOMAP_DIO_UNALIGNED))
> +		inode_dio_wait(VFS_I(ip));

Er... this really confused me when I read it -- my first thought was
"How can we be in the unaligned direct write function but DIO_UNALIGNED
isn't set?  Wouldn't we be in some other function if we're doing an
aligned direct write?"

Then I looked upthread to where Christph said he'd renamed it
IOMAP_DIO_SUBBLOCK, but I didn't think that was sufficiently better:

	if (!(flags & IOMAP_DIO_SUBBLOCK))
		iomap_dio_wait(...);

This flag doesn't have a 1:1 relationship with the iocb asking for an
(fsblock-)unaligned write or the iocb saying that the write involves
sub-block io -- this flag really means "I require a stable written
mapping, no post-processing (of the disk block) allowed".

Admittedly the comment above the definition of IOMAP_DIO_UNALIGNED
actually says this, but as we all know I sometimes like to review
patchsets backwards. :P

How about...

IOMAP_DIO_REQUIRE_OVERWRITE ?

IOMAP_DIO_REQUIRE_STABLE ?

--D

>  
> -	/*
> -	 * This must be the only I/O in-flight. Wait on it before we release the
> -	 * iolock to prevent subsequent overlapping I/O.
> -	 */
>  	trace_xfs_file_direct_write(iocb, from);
>  	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
> -			   &xfs_dio_write_ops, IOMAP_DIO_FORCE_WAIT);
> +			   &xfs_dio_write_ops, flags);
> +	/*
> +	 * Retry unaligned IO with exclusive blocking semantics if the DIO
> +	 * layer rejected it for mapping or locking reasons. If we are doing
> +	 * nonblocking user IO, propagate the error.
> +	 */
> +	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
> +		ASSERT(flags & IOMAP_DIO_UNALIGNED);
> +		xfs_iunlock(ip, iolock);
> +		goto retry_exclusive;
> +	}
> +
>  out_unlock:
>  	if (iolock)
>  		xfs_iunlock(ip, iolock);
> diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
> index 7b9ff824e82d48..dc8c86e98b99bf 100644
> --- a/fs/xfs/xfs_iomap.c
> +++ b/fs/xfs/xfs_iomap.c
> @@ -784,15 +784,30 @@ xfs_direct_write_iomap_begin(
>  		goto allocate_blocks;
>  
>  	/*
> -	 * NOWAIT IO needs to span the entire requested IO with a single map so
> -	 * that we avoid partial IO failures due to the rest of the IO range not
> -	 * covered by this map triggering an EAGAIN condition when it is
> -	 * subsequently mapped and aborting the IO.
> +	 * NOWAIT and unaligned IO needs to span the entire requested IO with a
> +	 * single map so that we avoid partial IO failures due to the rest of
> +	 * the IO range not covered by this map triggering an EAGAIN condition
> +	 * when it is subsequently mapped and aborting the IO.
>  	 */
> -	if ((flags & IOMAP_NOWAIT) &&
> -	    !imap_spans_range(&imap, offset_fsb, end_fsb)) {
> +	if (flags & (IOMAP_NOWAIT | IOMAP_UNALIGNED)) {
>  		error = -EAGAIN;
> -		goto out_unlock;
> +		if (!imap_spans_range(&imap, offset_fsb, end_fsb))
> +			goto out_unlock;
> +	}
> +
> +	/*
> +	 * For unsigned I/O we can't convert an unwritten extents if the I/O is
> +	 * not block size aligned, as such a conversion would have to do
> +	 * sub-block zeroing, and that can only be done under an exclusive
> +	 * IOLOCK. Hence if this is not a written extent, return EAGAIN to tell
> +	 * the caller to try again.
> +	 */
> +	if (flags & IOMAP_UNALIGNED) {
> +		error = -EAGAIN;
> +		if (imap.br_state != XFS_EXT_NORM &&
> +		    ((offset & mp->m_blockmask) ||
> +		     ((offset + length) & mp->m_blockmask)))
> +			goto out_unlock;
>  	}
>  
>  	xfs_iunlock(ip, lockmode);
> @@ -801,7 +816,7 @@ xfs_direct_write_iomap_begin(
>  
>  allocate_blocks:
>  	error = -EAGAIN;
> -	if (flags & IOMAP_NOWAIT)
> +	if (flags & (IOMAP_NOWAIT | IOMAP_UNALIGNED))
>  		goto out_unlock;
>  
>  	/*
> -- 
> 2.29.2
> 

Re: [PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Christoph Hellwig]

[another full quote removed, guys please send properly formatted email]

On Wed, Jan 20, 2021 at 10:40:56AM -0800, Darrick J. Wong wrote:
> > +	if (!(flags & IOMAP_DIO_UNALIGNED))
> > +		inode_dio_wait(VFS_I(ip));
> 
> Er... this really confused me when I read it -- my first thought was
> "How can we be in the unaligned direct write function but DIO_UNALIGNED
> isn't set?  Wouldn't we be in some other function if we're doing an
> aligned direct write?"
> 
> Then I looked upthread to where Christph said he'd renamed it
> IOMAP_DIO_SUBBLOCK, but I didn't think that was sufficiently better:
> 
> 	if (!(flags & IOMAP_DIO_SUBBLOCK))
> 		iomap_dio_wait(...);
> 
> This flag doesn't have a 1:1 relationship with the iocb asking for an
> (fsblock-)unaligned write or the iocb saying that the write involves
> sub-block io -- this flag really means "I require a stable written
> mapping, no post-processing (of the disk block) allowed".

Would:

	if (flags & IOMAP_DIO_FORCE_WAIT)
		inode_dio_wait(VFS_I(ip));

look any better to you?  Behavior would be the same.

Re: [PATCH 11/11] xfs: reduce exclusive locking on unaligned dio

[Darrick J. Wong]

On Wed, Jan 20, 2021 at 07:44:00PM +0100, Christoph Hellwig wrote:
> [another full quote removed, guys please send properly formatted email]
> 
> On Wed, Jan 20, 2021 at 10:40:56AM -0800, Darrick J. Wong wrote:
> > > +	if (!(flags & IOMAP_DIO_UNALIGNED))
> > > +		inode_dio_wait(VFS_I(ip));
> > 
> > Er... this really confused me when I read it -- my first thought was
> > "How can we be in the unaligned direct write function but DIO_UNALIGNED
> > isn't set?  Wouldn't we be in some other function if we're doing an
> > aligned direct write?"
> > 
> > Then I looked upthread to where Christph said he'd renamed it
> > IOMAP_DIO_SUBBLOCK, but I didn't think that was sufficiently better:
> > 
> > 	if (!(flags & IOMAP_DIO_SUBBLOCK))
> > 		iomap_dio_wait(...);
> > 
> > This flag doesn't have a 1:1 relationship with the iocb asking for an
> > (fsblock-)unaligned write or the iocb saying that the write involves
> > sub-block io -- this flag really means "I require a stable written
> > mapping, no post-processing (of the disk block) allowed".
> 
> Would:
> 
> 	if (flags & IOMAP_DIO_FORCE_WAIT)
> 		inode_dio_wait(VFS_I(ip));
> 
> look any better to you?  Behavior would be the same.

Looks fine to me.

--D