[PATCH 0/3 v3] ext4: dio overwrite nolock

[PATCH 0/3 v3] ext4: dio overwrite nolock

[Zheng Liu]

Hello list,

Here is the third version of dio overwrite nolock.  In this version, I rework
some codes according to Robin's reply.

This patch set can improve the performance of ext4 when the user does a dio
overwrite because, when a dio overwrite occurs, we don't need to take i_mutex
lock in some conditons.  The condition includes the size of file doesn't be
changed, and align aio.  So dio write can be parallelized in these conditions.

In patch 1, ext4_file_dio_write is defined to split buffered I/O and direct I/O
in ext4_file_write so that some codes can be added to check whether we can do a
dio overwrite without i_mutex lock later.

In patch 2, a new flag called EXT4_GET_BLOCKS_NOLOCK and a new get_block
function that is named ext4_get_block_write_nolock are defined to do a lookup to
let me know whether the extent of the file at this offset has been initialized
because we need to know whether a dio overwrite needs to modify the metadata of
the file or not.

In patch 3, we implement dio overwrite nolock.  In ext4_file_dio_write, we check
whether we can do a dio overwrite without lock.  Then we use 'iocb->private'
to store this flag to tell ext4_ext_direct_IO to handle it because
file_update_time will start a new journal and it will cause a deadlock.  So we
need to finish to update file time with i_mutex lock, and release lock in
ext4_ext_direct_IO.

v3 <- v2:
 * change the condition check of dio overwrite nolock in ext4_file_dio_write
 * rename 'create' variable to 'flags' in ext4_get_block_write_nolock
 * remove a duplicated if statement in ext4_ext_direct_IO

v2 <- v1:
 * rebase to 3.5
 * rework ext4_file_dio_write to avoid to copy vfs's code back into ext4
 * add some comments to explain how to determine whether we can do a nolocking
   overwrite dio

Any comments are welcomed.  Thanks.

Regards,
Zheng

Zheng Liu (3):
      ext4: split ext4_file_write into buffered IO and direct IO
      ext4: add a new flag for ext4_map_blocks
      ext4: add dio overwrite nolock

 fs/ext4/ext4.h  |    2 +
 fs/ext4/file.c  |  107 +++++++++++++++++++++++++++++++++++++++++++-----------
 fs/ext4/inode.c |   76 ++++++++++++++++++++++++++++++++++-----
 3 files changed, 153 insertions(+), 32 deletions(-)

[PATCH 1/3 v3] ext4: split ext4_file_write into buffered IO and direct IO

[Zheng Liu]

From: Zheng Liu <wenqing.lz@taobao.com>

ext4_file_dio_write is defined in order to split buffered IO and
direct IO in ext4.  This patch just refactor some stuff in write path.

CC: Tao Ma <tm@tao.ma>
CC: Eric Sandeen <sandeen@redhat.com>
CC: Robin Dong <hao.bigrat@gmail.com>
Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>
---
 fs/ext4/file.c |   60 +++++++++++++++++++++++++++++++++++--------------------
 1 files changed, 38 insertions(+), 22 deletions(-)

diff --git a/fs/ext4/file.c b/fs/ext4/file.c
index 8c7642a..a10dc77 100644
--- a/fs/ext4/file.c
+++ b/fs/ext4/file.c
@@ -90,34 +90,16 @@ ext4_unaligned_aio(struct inode *inode, const struct iovec *iov,
 }
 
 static ssize_t
-ext4_file_write(struct kiocb *iocb, const struct iovec *iov,
-		unsigned long nr_segs, loff_t pos)
+ext4_file_dio_write(struct kiocb *iocb, const struct iovec *iov,
+		    unsigned long nr_segs, loff_t pos)
 {
 	struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 	int unaligned_aio = 0;
 	ssize_t ret;
 
-	/*
-	 * If we have encountered a bitmap-format file, the size limit
-	 * is smaller than s_maxbytes, which is for extent-mapped files.
-	 */
-
-	if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
-		struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
-		size_t length = iov_length(iov, nr_segs);
-
-		if ((pos > sbi->s_bitmap_maxbytes ||
-		    (pos == sbi->s_bitmap_maxbytes && length > 0)))
-			return -EFBIG;
-
-		if (pos + length > sbi->s_bitmap_maxbytes) {
-			nr_segs = iov_shorten((struct iovec *)iov, nr_segs,
-					      sbi->s_bitmap_maxbytes - pos);
-		}
-	} else if (unlikely((iocb->ki_filp->f_flags & O_DIRECT) &&
-		   !is_sync_kiocb(iocb))) {
+	if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) &&
+	    !is_sync_kiocb(iocb))
 		unaligned_aio = ext4_unaligned_aio(inode, iov, nr_segs, pos);
-	}
 
 	/* Unaligned direct AIO must be serialized; see comment above */
 	if (unaligned_aio) {
@@ -141,6 +123,40 @@ ext4_file_write(struct kiocb *iocb, const struct iovec *iov,
 	return ret;
 }
 
+static ssize_t
+ext4_file_write(struct kiocb *iocb, const struct iovec *iov,
+		unsigned long nr_segs, loff_t pos)
+{
+	struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
+	ssize_t ret;
+
+	/*
+	 * If we have encountered a bitmap-format file, the size limit
+	 * is smaller than s_maxbytes, which is for extent-mapped files.
+	 */
+
+	if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
+		struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
+		size_t length = iov_length(iov, nr_segs);
+
+		if ((pos > sbi->s_bitmap_maxbytes ||
+		    (pos == sbi->s_bitmap_maxbytes && length > 0)))
+			return -EFBIG;
+
+		if (pos + length > sbi->s_bitmap_maxbytes) {
+			nr_segs = iov_shorten((struct iovec *)iov, nr_segs,
+					      sbi->s_bitmap_maxbytes - pos);
+		}
+	}
+
+	if (unlikely(iocb->ki_filp->f_flags & O_DIRECT))
+		ret = ext4_file_dio_write(iocb, iov, nr_segs, pos);
+	else
+		ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
+
+	return ret;
+}
+
 static const struct vm_operations_struct ext4_file_vm_ops = {
 	.fault		= filemap_fault,
 	.page_mkwrite   = ext4_page_mkwrite,
-- 
1.7.4.1

[PATCH 2/3 v3] ext4: add a new flag for ext4_map_blocks

[Zheng Liu]

From: Zheng Liu <wenqing.lz@taobao.com>

EXT4_GET_BLOCKS_NO_LOCK flag is added to indicate that we don't need to acquire
i_data_sem lock in ext4_map_blocks.  Meanwhile, it lets _ext4_get_block do not
start a new journal because when we do a overwrite dio, there is no any
metadata that needs to be modified.

We define a new function called ext4_get_block_write_nolock, which is used in
dio overwrite nolock.  In this function, it doesn't try to acquire i_data_sem
lock and doesn't start a new journal as it does a lookup.

CC: Tao Ma <tm@tao.ma>
CC: Eric Sandeen <sandeen@redhat.com>
CC: Robin Dong <hao.bigrat@gmail.com>
Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>
---
 fs/ext4/ext4.h  |    2 +
 fs/ext4/inode.c |   59 +++++++++++++++++++++++++++++++++++++++++++++---------
 2 files changed, 51 insertions(+), 10 deletions(-)

diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
index cfc4e01..d1a2b1e 100644
--- a/fs/ext4/ext4.h
+++ b/fs/ext4/ext4.h
@@ -571,6 +571,8 @@ enum {
 #define EXT4_GET_BLOCKS_NO_NORMALIZE		0x0040
 	/* Request will not result in inode size update (user for fallocate) */
 #define EXT4_GET_BLOCKS_KEEP_SIZE		0x0080
+	/* Do not take i_data_sem locking in ext4_map_blocks */
+#define EXT4_GET_BLOCKS_NO_LOCK			0x0100
 
 /*
  * Flags used by ext4_free_blocks
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 02bc8cb..76cb3b1 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -544,7 +544,8 @@ int ext4_map_blocks(handle_t *handle, struct inode *inode,
 	 * Try to see if we can get the block without requesting a new
 	 * file system block.
 	 */
-	down_read((&EXT4_I(inode)->i_data_sem));
+	if (!(flags & EXT4_GET_BLOCKS_NO_LOCK))
+		down_read((&EXT4_I(inode)->i_data_sem));
 	if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
 		retval = ext4_ext_map_blocks(handle, inode, map, flags &
 					     EXT4_GET_BLOCKS_KEEP_SIZE);
@@ -552,7 +553,8 @@ int ext4_map_blocks(handle_t *handle, struct inode *inode,
 		retval = ext4_ind_map_blocks(handle, inode, map, flags &
 					     EXT4_GET_BLOCKS_KEEP_SIZE);
 	}
-	up_read((&EXT4_I(inode)->i_data_sem));
+	if (!(flags & EXT4_GET_BLOCKS_NO_LOCK))
+		up_read((&EXT4_I(inode)->i_data_sem));
 
 	if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
 		int ret = check_block_validity(inode, map);
@@ -2818,6 +2820,32 @@ static int ext4_get_block_write(struct inode *inode, sector_t iblock,
 			       EXT4_GET_BLOCKS_IO_CREATE_EXT);
 }
 
+static int ext4_get_block_write_nolock(struct inode *inode, sector_t iblock,
+		   struct buffer_head *bh_result, int flags)
+{
+	handle_t *handle = ext4_journal_current_handle();
+	struct ext4_map_blocks map;
+	int ret = 0;
+
+	ext4_debug("ext4_get_block_write_nolock: inode %lu, flag %d\n",
+		   inode->i_ino, flags);
+
+	flags = EXT4_GET_BLOCKS_NO_LOCK;
+
+	map.m_lblk = iblock;
+	map.m_len = bh_result->b_size >> inode->i_blkbits;
+
+	ret = ext4_map_blocks(handle, inode, &map, flags);
+	if (ret > 0) {
+		map_bh(bh_result, inode->i_sb, map.m_pblk);
+		bh_result->b_state = (bh_result->b_state & ~EXT4_MAP_FLAGS) |
+					map.m_flags;
+		bh_result->b_size = inode->i_sb->s_blocksize * map.m_len;
+		ret = 0;
+	}
+	return ret;
+}
+
 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
 			    ssize_t size, void *private, int ret,
 			    bool is_async)
@@ -2966,6 +2994,8 @@ static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
 
 	loff_t final_size = offset + count;
 	if (rw == WRITE && final_size <= inode->i_size) {
+		int overwrite = 0;
+
 		/*
  		 * We could direct write to holes and fallocate.
 		 *
@@ -3005,13 +3035,22 @@ static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
 			EXT4_I(inode)->cur_aio_dio = iocb->private;
 		}
 
-		ret = __blockdev_direct_IO(rw, iocb, inode,
-					 inode->i_sb->s_bdev, iov,
-					 offset, nr_segs,
-					 ext4_get_block_write,
-					 ext4_end_io_dio,
-					 NULL,
-					 DIO_LOCKING);
+		if (overwrite)
+			ret = __blockdev_direct_IO(rw, iocb, inode,
+						 inode->i_sb->s_bdev, iov,
+						 offset, nr_segs,
+						 ext4_get_block_write_nolock,
+						 ext4_end_io_dio,
+						 NULL,
+						 0);
+		else
+			ret = __blockdev_direct_IO(rw, iocb, inode,
+						 inode->i_sb->s_bdev, iov,
+						 offset, nr_segs,
+						 ext4_get_block_write,
+						 ext4_end_io_dio,
+						 NULL,
+						 DIO_LOCKING);
 		if (iocb->private)
 			EXT4_I(inode)->cur_aio_dio = NULL;
 		/*
@@ -3031,7 +3070,7 @@ static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
 		if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
 			ext4_free_io_end(iocb->private);
 			iocb->private = NULL;
-		} else if (ret > 0 && ext4_test_inode_state(inode,
+		} else if (ret > 0 && !overwrite && ext4_test_inode_state(inode,
 						EXT4_STATE_DIO_UNWRITTEN)) {
 			int err;
 			/*
-- 
1.7.4.1

[PATCH 3/3 v3] ext4: add dio overwrite nolock

[Zheng Liu]

From: Zheng Liu <wenqing.lz@taobao.com>

Aligned and overwrite direct I/O can be parallelized.  In ext4_file_dio_write,
we first check whether these conditions are satisfied or not.  If so, we
take i_data_sem and release i_mutex lock directly.  Meanwhile iocb->private is
set to indicate that this is a dio overwrite, and it will be handled in
ext4_ext_direct_IO.

CC: Tao Ma <tm@tao.ma>
CC: Eric Sandeen <sandeen@redhat.com>
CC: Robin Dong <hao.bigrat@gmail.com>
Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>
---
 fs/ext4/file.c  |   51 +++++++++++++++++++++++++++++++++++++++++++++++++--
 fs/ext4/inode.c |   17 +++++++++++++++++
 2 files changed, 66 insertions(+), 2 deletions(-)

diff --git a/fs/ext4/file.c b/fs/ext4/file.c
index a10dc77..1c81509 100644
--- a/fs/ext4/file.c
+++ b/fs/ext4/file.c
@@ -93,9 +93,13 @@ static ssize_t
 ext4_file_dio_write(struct kiocb *iocb, const struct iovec *iov,
 		    unsigned long nr_segs, loff_t pos)
 {
-	struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
+	struct file *file = iocb->ki_filp;
+	struct inode *inode = file->f_mapping->host;
+	struct blk_plug plug;
 	int unaligned_aio = 0;
 	ssize_t ret;
+	int overwrite = 0;
+	size_t length = iov_length(iov, nr_segs);
 
 	if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) &&
 	    !is_sync_kiocb(iocb))
@@ -115,7 +119,50 @@ ext4_file_dio_write(struct kiocb *iocb, const struct iovec *iov,
 		ext4_aiodio_wait(inode);
 	}
 
-	ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
+	BUG_ON(iocb->ki_pos != pos);
+
+	mutex_lock(&inode->i_mutex);
+	blk_start_plug(&plug);
+
+	iocb->private = &overwrite;
+
+	/* check whether we do a DIO overwrite or not */
+	if (ext4_should_dioread_nolock(inode) && !unaligned_aio &&
+	    !file->f_mapping->nrpages && pos + length <= i_size_read(inode)) {
+		struct ext4_map_blocks map;
+		unsigned int blkbits = inode->i_blkbits;
+		int err, len;
+
+		map.m_lblk = pos >> blkbits;
+		map.m_len = (EXT4_BLOCK_ALIGN(pos + length, blkbits) >> blkbits)
+			- map.m_lblk;
+		len = map.m_len;
+
+		err = ext4_map_blocks(NULL, inode, &map, 0);
+		/*
+		 * 'err==len' means that all of blocks has been preallocated no
+		 * matter they are initialized or not.  For excluding
+		 * uninitialized extents, we need to check m_flags.  There are
+		 * two conditions that indicate for initialized extents.
+		 * 1) If we hit extent cache, EXT4_MAP_MAPPED flag is returned;
+		 * 2) If we do a real lookup, non-flags are returned.
+		 * So we should check these two conditions.
+		 */
+		if (err == len && (map.m_flags & EXT4_MAP_MAPPED))
+			overwrite = 1;
+	}
+
+	ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
+	mutex_unlock(&inode->i_mutex);
+
+	if (ret > 0 || ret == -EIOCBQUEUED) {
+		ssize_t err;
+
+		err = generic_write_sync(file, pos, ret);
+		if (err < 0 && ret > 0)
+			ret = err;
+	}
+	blk_finish_plug(&plug);
 
 	if (unaligned_aio)
 		mutex_unlock(ext4_aio_mutex(inode));
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 76cb3b1..8c84070 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -2996,6 +2996,16 @@ static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
 	if (rw == WRITE && final_size <= inode->i_size) {
 		int overwrite = 0;
 
+		BUG_ON(iocb->private == NULL);
+
+		/* If we do a overwrite dio, i_mutex locking can be released */
+		overwrite = *((int *)iocb->private);
+
+		if (overwrite) {
+			down_read(&EXT4_I(inode)->i_data_sem);
+			mutex_unlock(&inode->i_mutex);
+		}
+
 		/*
  		 * We could direct write to holes and fallocate.
 		 *
@@ -3083,6 +3093,13 @@ static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
 				ret = err;
 			ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
 		}
+
+		/* take i_mutex locking again if we do a ovewrite dio */
+		if (overwrite) {
+			up_read(&EXT4_I(inode)->i_data_sem);
+			mutex_lock(&inode->i_mutex);
+		}
+
 		return ret;
 	}
 
-- 
1.7.4.1

Re: [PATCH 1/3 v3] ext4: split ext4_file_write into buffered IO and direct IO

[Theodore Ts'o]

On Sun, Jul 01, 2012 at 10:21:44PM +0800, Zheng Liu wrote:
> From: Zheng Liu <wenqing.lz@taobao.com>
> 
> ext4_file_dio_write is defined in order to split buffered IO and
> direct IO in ext4.  This patch just refactor some stuff in write path.
> 
> CC: Tao Ma <tm@tao.ma>
> CC: Eric Sandeen <sandeen@redhat.com>
> CC: Robin Dong <hao.bigrat@gmail.com>
> Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>

Applied, thanks.

						 -Ted

Re: [PATCH 2/3 v3] ext4: add a new flag for ext4_map_blocks

[Theodore Ts'o]

On Sun, Jul 01, 2012 at 10:21:45PM +0800, Zheng Liu wrote:
> From: Zheng Liu <wenqing.lz@taobao.com>
> 
> EXT4_GET_BLOCKS_NO_LOCK flag is added to indicate that we don't need to acquire
> i_data_sem lock in ext4_map_blocks.  Meanwhile, it lets _ext4_get_block do not
> start a new journal because when we do a overwrite dio, there is no any
> metadata that needs to be modified.
> 
> We define a new function called ext4_get_block_write_nolock, which is used in
> dio overwrite nolock.  In this function, it doesn't try to acquire i_data_sem
> lock and doesn't start a new journal as it does a lookup.
> 
> CC: Tao Ma <tm@tao.ma>
> CC: Eric Sandeen <sandeen@redhat.com>
> CC: Robin Dong <hao.bigrat@gmail.com>
> Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>

Thanks, applied.

						- Ted

Re: [PATCH 3/3 v3] ext4: add dio overwrite nolock

[Theodore Ts'o]

On Sun, Jul 01, 2012 at 10:21:46PM +0800, Zheng Liu wrote:
> From: Zheng Liu <wenqing.lz@taobao.com>
> 
> Aligned and overwrite direct I/O can be parallelized.  In ext4_file_dio_write,
> we first check whether these conditions are satisfied or not.  If so, we
> take i_data_sem and release i_mutex lock directly.  Meanwhile iocb->private is
> set to indicate that this is a dio overwrite, and it will be handled in
> ext4_ext_direct_IO.
> 
> CC: Tao Ma <tm@tao.ma>
> CC: Eric Sandeen <sandeen@redhat.com>
> CC: Robin Dong <hao.bigrat@gmail.com>
> Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>

Thanks, applied.

					- Ted