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

[Tao Ma <tm@tao.ma>]

On 04/28/2012 11:39 AM, Zheng Liu wrote:
> From: Zheng Liu <wenqing.lz@taobao.com>
> 
> Aligned and overwrite direct IO can be parallelized.  In ext4_file_dio_write,
> we first check whether these conditions are satisfied or not.  If so, we unlock
> the i_mutex and acquire i_data_sem directly.  Meanwhile iocb->private is set to
> indicate that this is a overwrite dio, and it will be processed in
> ext4_ext_direct_IO.
> 
> Signed-off-by: Zheng Liu <wenqing.lz@taobao.com>
> ---
>  fs/ext4/file.c |  140 ++++++++++++++++++++++++++++++++++++++++++++++++++++++-
>  1 files changed, 137 insertions(+), 3 deletions(-)
> 
> diff --git a/fs/ext4/file.c b/fs/ext4/file.c
> index e5d6be3..8a5f713 100644
> --- a/fs/ext4/file.c
> +++ b/fs/ext4/file.c
> @@ -100,9 +100,21 @@ 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;
> -	int unaligned_aio = 0;
> +	struct file *file = iocb->ki_filp;
> +	struct address_space * mapping = file->f_mapping;
> +	struct inode *inode = file->f_path.dentry->d_inode;
> +	struct blk_plug plug;
>  	ssize_t ret;
> +	ssize_t written, written_buffered;
> +	size_t length = iov_length(iov, nr_segs);
> +	size_t ocount;		/* original count */
> +	size_t count;		/* after file limit checks */
> +	int unaligned_aio = 0;
> +	int overwrite = 0;
> +	loff_t *ppos = &iocb->ki_pos;
> +	loff_t endbyte;
> +
> +	BUG_ON(iocb->ki_pos != pos);
>  
>  	if (!is_sync_kiocb(iocb))
>  		unaligned_aio = ext4_unaligned_aio(inode, iov, nr_segs, pos);
> @@ -121,7 +133,129 @@ 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);
> +	mutex_lock(&inode->i_mutex);
> +	blk_start_plug(&plug);
> +
> +	ocount = 0;
> +	ret = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
> +	if (ret)
> +		goto unlock_out;
> +
> +	count = ocount;
> +	pos = *ppos;
> +
> +	vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
> +
> +	/* We can write back this queue in page reclaim */
> +	current->backing_dev_info = mapping->backing_dev_info;
> +	written = 0;
> +
> +	ret = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
> +	if (ret)
> +		goto out;
> +
> +	if (count == 0)
> +		goto out;
> +
> +	ret = file_remove_suid(file);
> +	if (ret)
> +		goto out;
> +
> +	file_update_time(file);
> +
> +	iocb->private = NULL;
> +
> +	if (!unaligned_aio && !file->f_mapping->nrpages &&
> +	    pos + length < i_size_read(inode) &&
should be pos + length <= ?
And inode->i_size should be ok since now we have i_mutex held.
> +	    ext4_should_dioread_nolock(inode)) {
> +		struct ext4_map_blocks map;
> +		unsigned int blkbits = inode->i_blkbits;
> +		int err;
> +		int 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);
> +		if (err == len && (!map.m_flags ||
> +		    map.m_flags & EXT4_MAP_MAPPED)) {
could you please add some comments about how and why map.m_flags are
checked this way?
> +			overwrite = 1;
> +			iocb->private = &overwrite;
> +			mutex_unlock(&inode->i_mutex);
> +			down_read(&EXT4_I(inode)->i_data_sem);
Is there any possibility that the metadata is changed after we dropped
the i_mutex before the down_read?
> +		}
> +	}
> +
> +	if (file->f_mapping->nrpages && overwrite) {
> +		overwrite = 0;
> +		up_read(&EXT4_I(inode)->i_data_sem);
> +		mutex_lock(&inode->i_mutex);
I am not sure whether it could happen. But if it does happen, should we
also change the value in iocb->private?
> +	}
> +
> +	written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
> +						ppos, count, ocount);
> +	if (written < 0 || written == count)
> +		goto out;
> +	/*
> +	 * direct-io write to a hole: fall through to buffered I/O
> +	 * for completing the rest of the request.
> +	 */
> +	pos += written;
> +	count -= written;
> +	written_buffered = generic_file_buffered_write(iocb, iov,
> +					nr_segs, pos, ppos, count,
> +					written);
If we fall back here, should we re-lock the i_mutex since the buffer
write isn't guaranteed?

Thanks
Tao
> +	/*
> +	 * If generic_file_buffered_write() retuned a synchronous error
> +	 * then we want to return the number of bytes which were
> +	 * direct-written, or the error code if that was zero.  Note
> +	 * that this differs from normal direct-io semantics, which
> +	 * will return -EFOO even if some bytes were written.
> +	 */
> +	if (written_buffered < 0) {
> +		ret = written_buffered;
> +		goto out;
> +	}
> +
> +	/*
> +	 * We need to ensure that the page cache pages are written to
> +	 * disk and invalidated to preserve the expected O_DIRECT
> +	 * semantics.
> +	 */
> +	endbyte = pos + written_buffered - written - 1;
> +	ret = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
> +	if (ret == 0) {
> +		written = written_buffered;
> +		invalidate_mapping_pages(mapping,
> +					 pos >> PAGE_CACHE_SHIFT,
> +					 endbyte >> PAGE_CACHE_SHIFT);
> +	} else {
> +		/*
> +		 * We don't know how much we wrote, so just return
> +		 * the number of bytes which were direct-written
> +		 */
> +	}
> +
> +out:
> +	current->backing_dev_info = NULL;
> +	ret = written ? written : ret;
> +
> +unlock_out:
> +	if (overwrite)
> +		up_read(&EXT4_I(inode)->i_data_sem);
> +	else
> +		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));