On Fri, Oct 25, 2019 at 05:03:32PM +0200, Christoph Hellwig wrote:
> Even if we are asked for a write layout there is no point in logging
> the inode unless we actually modified it in some way.
>
> Signed-off-by: Christoph Hellwig <hch@xxxxxx>
> ---
> fs/xfs/xfs_pnfs.c | 43 +++++++++++++++++++------------------------
> 1 file changed, 19 insertions(+), 24 deletions(-)
>
> diff --git a/fs/xfs/xfs_pnfs.c b/fs/xfs/xfs_pnfs.c
> index 9c96493be9e0..fa90c6334c7c 100644
> --- a/fs/xfs/xfs_pnfs.c
> +++ b/fs/xfs/xfs_pnfs.c
> @@ -147,32 +147,27 @@ xfs_fs_map_blocks(
> if (error)
> goto out_unlock;
>
> - if (write) {
> - enum xfs_prealloc_flags flags = 0;
> -
> + if (write &&
> + (!nimaps || imap.br_startblock == HOLESTARTBLOCK)) {
> ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
The change in code flow makes this assert rather useless, I think, since
we only end up in this branch if we have a write and a hole. If the
condition that it checks is important (and it seems to be?) then it
ought to be hoisted up a level and turned into:
ASSERT(!write || !nimaps || imap.br_startblock != DELAYSTARTBLOCK);
Right?
Otherwise looks ok.
--D
>
> - if (!nimaps || imap.br_startblock == HOLESTARTBLOCK) {
> - /*
> - * xfs_iomap_write_direct() expects to take ownership of
> - * the shared ilock.
> - */
> - xfs_ilock(ip, XFS_ILOCK_SHARED);
> - error = xfs_iomap_write_direct(ip, offset, length,
> - &imap, nimaps);
> - if (error)
> - goto out_unlock;
> -
> - /*
> - * Ensure the next transaction is committed
> - * synchronously so that the blocks allocated and
> - * handed out to the client are guaranteed to be
> - * present even after a server crash.
> - */
> - flags |= XFS_PREALLOC_SET | XFS_PREALLOC_SYNC;
> - }
> -
> - error = xfs_update_prealloc_flags(ip, flags);
> + /*
> + * xfs_iomap_write_direct() expects to take ownership of the
> + * shared ilock.
> + */
> + xfs_ilock(ip, XFS_ILOCK_SHARED);
> + error = xfs_iomap_write_direct(ip, offset, length, &imap,
> + nimaps);
> + if (error)
> + goto out_unlock;
> +
> + /*
> + * Ensure the next transaction is committed synchronously so
> + * that the blocks allocated and handed out to the client are
> + * guaranteed to be present even after a server crash.
> + */
> + error = xfs_update_prealloc_flags(ip,
> + XFS_PREALLOC_SET | XFS_PREALLOC_SYNC);
> if (error)
> goto out_unlock;
> }
> --
> 2.20.1
>
