Re: concurrent direct IO write in xfs

[Christoph Hellwig <hch@xxxxxxxxxxxxx>]

On Thu, Feb 09, 2012 at 05:09:20PM +1100, Dave Chinner wrote:
>       if (create) {
> -             lockmode = XFS_ILOCK_EXCL;
> +             /*
> +              * For direct IO, we lock in shared mode so that write
> +              * operations that don't require allocation can occur
> +              * concurrently. The ilock has to be dropped over the allocation
> +              * transaction reservation, so the only thing the ilock is
> +              * providing here is modification exclusion. i.e. there is no
> +              * need to hold the lock exclusive.
> +              *
> +              * For buffered IO, if we need to do delayed allocation then
> +              * hold the ilock exclusive so that the lookup and delalloc
> +              * reservation is atomic.
> +              */
> +             if (direct)
> +                     lockmode = XFS_ILOCK_SHARED;
> +             else
> +                     lockmode = XFS_ILOCK_EXCL;
>               xfs_ilock(ip, lockmode);
>       } else {
>               lockmode = xfs_ilock_map_shared(ip);

We'll actually need to use xfs_ilock_map_shared for the the direct
create case too, to make sure we have the exclusive lock when we first
read the extent list in.

Also xfs_qm_dqattach_locked really wants the inode locked exclusively,
which your current code doesn't handle.