Re: [PATCH] xfs: update i_size after unwritten conversion in dio completion

[Christoph Hellwig <hch@infradead.org>]

On Thu, Sep 21, 2017 at 06:38:28PM +0800, Eryu Guan wrote:
> Since commit d531d91d6990 ("xfs: always use unwritten extents for
> direct I/O writes"), we start allocating unwritten extents for all
> direct writes to allow appending aio in XFS.
> 
> But for dio writes that could extend file size we update the in-core
> inode size first, then convert the unwritten extents to real
> allocations at dio completion time in xfs_dio_write_end_io(). Thus a
> racing direct read could see the new i_size and find the unwritten
> extents first and read zeros instead of actual data, if the direct
> writer also takes a shared iolock.
> 
> Fix it by updating the in-core inode size after the unwritten extent
> conversion. To do this, introduce a new boolean argument to
> xfs_iomap_write_unwritten() to tell if we want to update in-core
> i_size or not.
> 
> Suggested-by: Brian Foster <bfoster@redhat.com>
> Signed-off-by: Eryu Guan <eguan@redhat.com>
> ---
> 
> Patch passed the test posted by Eric[1] and a locally modified aio
> version of the test.
> 
> I also ran fstests with config xfs_4k_crc, xfs_2k_reflink, xfs_1k_rmap
> and xfs_512, and aio-dio tests from ltp, I don't see any new failures
> introduced.
> 
> [1] http://www.spinics.net/lists/fstests/msg06978.html
> 
>  fs/xfs/xfs_aops.c  |  9 ++++++++-
>  fs/xfs/xfs_file.c  | 34 ++++++++++++++++++++--------------
>  fs/xfs/xfs_iomap.c |  7 ++++++-
>  fs/xfs/xfs_iomap.h |  2 +-
>  fs/xfs/xfs_pnfs.c  |  2 +-
>  5 files changed, 36 insertions(+), 18 deletions(-)
> 
> diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
> index 29172609f2a3..f937968e9515 100644
> --- a/fs/xfs/xfs_aops.c
> +++ b/fs/xfs/xfs_aops.c
> @@ -343,7 +343,14 @@ xfs_end_io(
>  		error = xfs_reflink_end_cow(ip, offset, size);
>  		break;
>  	case XFS_IO_UNWRITTEN:
> -		error = xfs_iomap_write_unwritten(ip, offset, size);
> +		/*
> +		 * The correct in-core inode size should have been updated by
> +		 * generic_write_end, and the 'size' here is buffer head
> +		 * granularity size of the ioend, which could be larger than
> +		 * the actual bytes written. So skip in-core i_size update in
> +		 * xfs_iomap_write_unwritten()
> +		 */
> +		error = xfs_iomap_write_unwritten(ip, offset, size, false);
>  		break;
>  	default:
>  		ASSERT(!xfs_ioend_is_append(ioend) || ioend->io_append_trans);
> diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
> index 350b6d43ba23..d4796c5a88fe 100644
> --- a/fs/xfs/xfs_file.c
> +++ b/fs/xfs/xfs_file.c
> @@ -434,7 +434,6 @@ xfs_dio_write_end_io(
>  	struct inode		*inode = file_inode(iocb->ki_filp);
>  	struct xfs_inode	*ip = XFS_I(inode);
>  	loff_t			offset = iocb->ki_pos;
> -	bool			update_size = false;
>  	int			error = 0;
>  
>  	trace_xfs_end_io_direct_write(ip, offset, size);
> @@ -445,6 +444,22 @@ xfs_dio_write_end_io(
>  	if (size <= 0)
>  		return size;
>  
> +	if (flags & IOMAP_DIO_COW) {
> +		error = xfs_reflink_end_cow(ip, offset, size);
> +		if (error)
> +			return error;
> +	}

So this should be ok, but without digging in the details I wonder
if we have proper tests for COWing the last block.

e.g. something like

write data from bytes 0 to 4000 to file a
clone file a to file b
write bytes 4000 to 4008 on file b

> @@ -900,6 +902,9 @@ xfs_iomap_write_unwritten(
>  		if (i_size > offset + count)
>  			i_size = offset + count;
>  
> +		if (update_isize && (i_size > i_size_read(inode)))

no need for the inner braces here.