[PATCH 26/29] xfs: return -EAGAIN when nowait meets sync in transaction commit

[Dave Chinner]

On Fri, Aug 25, 2023 at 09:54:28PM +0800, Hao Xu wrote:
> From: Hao Xu <howeyxu at tencent.com>
> 
> if the log transaction is a sync one, let's fail the nowait try and
> return -EAGAIN directly since sync transaction means blocked by IO.
> 
> Signed-off-by: Hao Xu <howeyxu at tencent.com>
> ---
>  fs/xfs/xfs_trans.c | 14 +++++++++++++-
>  1 file changed, 13 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/xfs/xfs_trans.c b/fs/xfs/xfs_trans.c
> index 7988b4c7f36e..f1f84a3dd456 100644
> --- a/fs/xfs/xfs_trans.c
> +++ b/fs/xfs/xfs_trans.c
> @@ -968,12 +968,24 @@ __xfs_trans_commit(
>  	xfs_csn_t		commit_seq = 0;
>  	int			error = 0;
>  	int			sync = tp->t_flags & XFS_TRANS_SYNC;
> +	bool			nowait = tp->t_flags & XFS_TRANS_NOWAIT;
> +	bool			perm_log = tp->t_flags & XFS_TRANS_PERM_LOG_RES;
>  
>  	trace_xfs_trans_commit(tp, _RET_IP_);
>  
> +	if (nowait && sync) {
> +		/*
> +		 * Currently nowait is only from xfs_vn_update_time()
> +		 * so perm_log is always false here, but let's make
> +		 * code general.
> +		 */
> +		if (perm_log)
> +			xfs_defer_cancel(tp);
> +		goto out_unreserve;
> +	}

This is fundamentally broken.  We cannot about a transaction commit
with dirty items at this point with shutting down the filesystem.

This points to XFS_TRANS_NOWAIT being completely broken, too,
because we don't call xfs_trans_set_sync() until just before we
commit the transaction. At this point, it is -too late- for
nowait+sync to be handled gracefully, and it will *always* go bad.

IOWs, the whole transaction "nowait" semantics as designed and
implemented is not a workable solution....

-Dave.
-- 
Dave Chinner
david at fromorbit.com