[dm-devel] error propagation problem on xfs over dm stripe

[Christoph Hellwig]

On Tue, Jan 31, 2017 at 09:12:07PM -0600, Eric Sandeen wrote:
> diff --git a/drivers/md/dm.c b/drivers/md/dm.c
> index 3086da5..3555ba8 100644
> --- a/drivers/md/dm.c
> +++ b/drivers/md/dm.c
> @@ -808,7 +808,9 @@ static void dec_pending(struct dm_io *io, int error)
>  		} else {
>  			/* done with normal IO or empty flush */
>  			trace_block_bio_complete(md->queue, bio, io_error);
> -			bio->bi_error = io_error;
> +			/* don't overwrite or clear existing errors */
> +			if (!bio->bi_error)
> +				bio->bi_error = io_error;
>  			bio_endio(bio);
>  		}
>  	}
> 
> but Mike was a little uneasy, not knowing for sure how we got here to
> overwrite this bio's error (hopefully I'm representing his concerns
> fairly and correctly).

FYI, what we do both in the XFS buffer cache and the new direct I/O
code is to use a

	        cmpxchg(&dio->error, 0, ret);

in a similar situation, which should work here, too.