Re: [PATCH] Re: aio_poll in 2.6?

[Andrew Morton]

John Myers <[e-mail]> wrote:
>
> It's basically waiting for someone to merge the patch.  There were
> some people making unsubstantiated claims that it didn't scale, but
> the available benchmarks showed that it scaled perfectly across the
> parameters tested.

What is the testing status of this?

Have any real-life applications been converted?

Does it require libaio changes to test?

>  
> +int async_poll(struct kiocb *iocb, int events);
> +

growl.  Please don't put function prototypes in .c files.  It defeats
typechecking.  I've moved this to aio.h.

> +int async_poll(struct kiocb *iocb, int events)
> +{
> +	unsigned int mask;
> +	struct async_poll_iocb *apiocb = kiocb_to_apiocb(iocb);
> +
> +	/* Fast path */
> +	if (iocb->ki_filp->f_op && iocb->ki_filp->f_op->poll) {
> +		mask = iocb->ki_filp->f_op->poll(iocb->ki_filp, NULL);
> +		mask &= events | POLLERR | POLLHUP;
> +		if (mask & events)
> +			return events;
> +	}
> +
> +	init_poll_funcptr(&apiocb->pt, async_poll_queue_proc);
> +	apiocb->armed = &apiocb;
> +	apiocb->outofmem = 0;
> +	apiocb->events = events;
> +	apiocb->ehead = NULL;
> +
> +	iocb->ki_users++;

There is no locking around this modification of ->ki_users.  Is this
correct?

> +	wmb();

Barriers always need comments explaining why they are there.

> +
> +	mask = DEFAULT_POLLMASK;
> +	if (iocb->ki_filp->f_op && iocb->ki_filp->f_op->poll)
> +		mask = iocb->ki_filp->f_op->poll(iocb->ki_filp, &apiocb->pt);
> +	mask &= events | POLLERR | POLLHUP;
> + 	if (mask && xchg(&apiocb->armed, NULL)) {
> +		async_poll_freewait(apiocb, NULL);
> +		aio_complete(iocb, mask, 0);
> +	}
> +	if (unlikely(apiocb->outofmem) && xchg(&apiocb->armed, NULL)) {
> +		async_poll_freewait(apiocb, NULL);
> +		aio_put_req(iocb);
> +		aio_put_req(iocb);

Is the double-put intentional?