Re: [PATCH] aio: partial write should not return error code.

[Zach Brown <zach.brown@oracle.com>]

Rusty Russell wrote:
> When an AIO write gets an error after writing some data (eg. ENOSPC),
> it should return the amount written already, not the error.  Just like
> write() is supposed to.

Andrew, please don't queue this fix.  I think the bug is valid but the
patch is subtly dangerous.

> diff -r 18802689361a fs/aio.c
> --- a/fs/aio.c	Thu Jan 03 15:22:24 2008 +1100
> +++ b/fs/aio.c	Thu Jan 03 18:05:25 2008 +1100
> @@ -1346,6 +1350,13 @@ static ssize_t aio_rw_vect_retry(struct 
>  	/* This means we must have transferred all that we could */
>  	/* No need to retry anymore */
>  	if ((ret == 0) || (iocb->ki_left == 0))
> +		ret = iocb->ki_nbytes - iocb->ki_left;
> +
> +	/* If we managed to write some out we return that, rather than
> +	 * the eventual error. */
> +	if (opcode == IOCB_CMD_PWRITEV
> +	    && ret < 0
> +	    && iocb->ki_nbytes - iocb->ki_left)
>  		ret = iocb->ki_nbytes - iocb->ki_left;

This doesn't account for the (sigh) -EIOCB* error codes.  They must be
returned to the caller so that it can properly handle the iocb reference
counting.  Failure to do so can lead to oopses.

To be fair, I think you'll have a really hard time finding an
->aio_write() implementation which would return partial progress and
*then* one of the magical errnos.  But the infrastructure does allow it.

So maybe we could get a helper in aio.h that abstracts out the

	(ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY)

condition.  Then I think this patch would be fine.

I assigned a bug to remind myself to revisit this if you aren't excited
by continuing with the patch:

  http://bugzilla.kernel.org/show_bug.cgi?id=9681

- z