Re: [PATCH 4/6] xfs: call xfs_flush_unmap_range from xfs_free_file_space

["Darrick J. Wong" <djwong@kernel.org>]

On Tue, Aug 27, 2024 at 08:50:48AM +0200, Christoph Hellwig wrote:
> Call xfs_flush_unmap_range from xfs_free_file_space so that
> xfs_file_fallocate doesn't have to predict which mode will call it.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>

Hmm.  I /think/ it's ok to shift the xfs_flush_unmap_range after the
file_modified and some of the other EINVAL bailouts that can happen
before xfs_free_file_space gets called.  Effectively that means that we
can fail faster now? :)

Later on this means that the cow-around code that the rtreflink patchset
introduces will also get flushed to disk before we start
collapsing/zeroing/punching.  AFAICT that should be fine.

Reviewed-by: Darrick J. Wong <djwong@kernel.org>

--D

> ---
>  fs/xfs/xfs_bmap_util.c |  8 ++++++++
>  fs/xfs/xfs_file.c      | 21 ---------------------
>  2 files changed, 8 insertions(+), 21 deletions(-)
> 
> diff --git a/fs/xfs/xfs_bmap_util.c b/fs/xfs/xfs_bmap_util.c
> index fe2e2c93097550..187a0dbda24fc4 100644
> --- a/fs/xfs/xfs_bmap_util.c
> +++ b/fs/xfs/xfs_bmap_util.c
> @@ -848,6 +848,14 @@ xfs_free_file_space(
>  	if (len <= 0)	/* if nothing being freed */
>  		return 0;
>  
> +	/*
> +	 * Now AIO and DIO has drained we flush and (if necessary) invalidate
> +	 * the cached range over the first operation we are about to run.
> +	 */
> +	error = xfs_flush_unmap_range(ip, offset, len);
> +	if (error)
> +		return error;
> +
>  	startoffset_fsb = XFS_B_TO_FSB(mp, offset);
>  	endoffset_fsb = XFS_B_TO_FSBT(mp, offset + len);
>  
> diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> index 4cdc54dc96862e..5b9e49da06013c 100644
> --- a/fs/xfs/xfs_file.c
> +++ b/fs/xfs/xfs_file.c
> @@ -890,27 +890,6 @@ xfs_file_fallocate(
>  	 */
>  	inode_dio_wait(inode);
>  
> -	/*
> -	 * Now AIO and DIO has drained we flush and (if necessary) invalidate
> -	 * the cached range over the first operation we are about to run.
> -	 *
> -	 * We care about zero and collapse here because they both run a hole
> -	 * punch over the range first. Because that can zero data, and the range
> -	 * of invalidation for the shift operations is much larger, we still do
> -	 * the required flush for collapse in xfs_prepare_shift().
> -	 *
> -	 * Insert has the same range requirements as collapse, and we extend the
> -	 * file first which can zero data. Hence insert has the same
> -	 * flush/invalidate requirements as collapse and so they are both
> -	 * handled at the right time by xfs_prepare_shift().
> -	 */
> -	if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE |
> -		    FALLOC_FL_COLLAPSE_RANGE)) {
> -		error = xfs_flush_unmap_range(ip, offset, len);
> -		if (error)
> -			goto out_unlock;
> -	}
> -
>  	error = file_modified(file);
>  	if (error)
>  		goto out_unlock;
> -- 
> 2.43.0
> 
>