On Mon, Dec 03, 2018 at 05:24:58PM -0500, Christoph Hellwig wrote:
> While using delalloc for extsize hints is generally a good idea, the
> current code that does so only for COW doesn't help us much and creates
> a lot of special cases. Switch it to use real allocations like we
> do for direct I/O.
>
> Signed-off-by: Christoph Hellwig <hch@xxxxxx>
> ---
> fs/xfs/xfs_iomap.c | 28 +++++++++++++++++-----------
> fs/xfs/xfs_reflink.c | 5 ++++-
> fs/xfs/xfs_reflink.h | 5 ++---
> 3 files changed, 23 insertions(+), 15 deletions(-)
>
> diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
> index 9f1fd224bb06..d851abac16a9 100644
> --- a/fs/xfs/xfs_iomap.c
> +++ b/fs/xfs/xfs_iomap.c
> @@ -1039,22 +1039,28 @@ xfs_file_iomap_begin(
> * been done up front, so we don't need to do them here.
> */
> if (xfs_is_reflink_inode(ip)) {
> + struct xfs_bmbt_irec orig = imap;
> +
> /* if zeroing doesn't need COW allocation, then we are done. */
> if ((flags & IOMAP_ZERO) &&
> !needs_cow_for_zeroing(&imap, nimaps))
> goto out_found;
>
> - if (flags & IOMAP_DIRECT) {
> - /* may drop and re-acquire the ilock */
> - error = xfs_reflink_allocate_cow(ip, &imap, &shared,
> - &lockmode);
> - if (error)
> - goto out_unlock;
> - } else {
> - error = xfs_reflink_reserve_cow(ip, &imap);
> - if (error)
> - goto out_unlock;
> - }
> + error = xfs_reflink_allocate_cow(ip, &imap, &shared, &lockmode,
> + flags);
> + if (error)
> + goto out_unlock;
> +
> + /*
> + * For buffered writes we need to report the address of the
> + * previous block (if there was any) so that the higher level
> + * write code can perform read-modify-write operations. For
> + * direct I/O code, which must be block aligned we need to
> + * report the newly allocated address.
> + */
> + if (!(flags & IOMAP_DIRECT) &&
> + orig.br_startblock != HOLESTARTBLOCK)
> + imap = orig;
>
> end_fsb = imap.br_startoff + imap.br_blockcount;
> length = XFS_FSB_TO_B(mp, end_fsb) - offset;
> diff --git a/fs/xfs/xfs_reflink.c b/fs/xfs/xfs_reflink.c
> index a8c32632090c..bdbaff1b3fb7 100644
> --- a/fs/xfs/xfs_reflink.c
> +++ b/fs/xfs/xfs_reflink.c
> @@ -397,7 +397,8 @@ xfs_reflink_allocate_cow(
> struct xfs_inode *ip,
> struct xfs_bmbt_irec *imap,
> bool *shared,
> - uint *lockmode)
> + uint *lockmode,
> + unsigned flags)
I'm not thrilled with passing iomap flags into the reflink code here...
> {
> struct xfs_mount *mp = ip->i_mount;
> xfs_fileoff_t offset_fsb = imap->br_startoff;
> @@ -471,6 +472,8 @@ xfs_reflink_allocate_cow(
> if (nimaps == 0)
> return -ENOSPC;
> convert:
> + if (!(flags & IOMAP_DIRECT))
...because I feel that it's easy to miss the subtlety here that for
buffered writes we don't care if the cow extent is unwritten or written,
but for directio we very /much/ care that the cow extent is written,
because we're writing to it immediately. Can this grow a comment to
reinforce why we skip the conversion?
Also, can we call this 'iomap_flags' to make it clearer which flags
we're talking about?
/*
* COW fork extents are supposed to remain unwritten until we're ready
* to initiate a disk write. For directio we /are/ going to write the
* data and need the conversion, but for buffered writes we're done.
*/
if (!(iomap_flags & IOMAP_DIRECT))
return 0;
return xfs_reflink_convert_cow_extent(...);
--D
> + return 0;
> return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb);
>
> out_unreserve:
> diff --git a/fs/xfs/xfs_reflink.h b/fs/xfs/xfs_reflink.h
> index 6d73daef1f13..d76fc520cac8 100644
> --- a/fs/xfs/xfs_reflink.h
> +++ b/fs/xfs/xfs_reflink.h
> @@ -12,10 +12,9 @@ extern int xfs_reflink_find_shared(struct xfs_mount *mp, struct xfs_trans *tp,
> extern int xfs_reflink_trim_around_shared(struct xfs_inode *ip,
> struct xfs_bmbt_irec *irec, bool *shared);
>
> -extern int xfs_reflink_reserve_cow(struct xfs_inode *ip,
> - struct xfs_bmbt_irec *imap);
> extern int xfs_reflink_allocate_cow(struct xfs_inode *ip,
> - struct xfs_bmbt_irec *imap, bool *shared, uint *lockmode);
> + struct xfs_bmbt_irec *imap, bool *shared, uint *lockmode,
> + unsigned flags);
> extern int xfs_reflink_convert_cow(struct xfs_inode *ip, xfs_off_t offset,
> xfs_off_t count);
>
> --
> 2.19.1
>
