Re: Revert "aio: fix aio request leak when events are reaped by user space"

[Benjamin LaHaise <bcrl@kvack.org>]

On Tue, Aug 19, 2014 at 08:46:51PM -0400, Benjamin LaHaise wrote:
> You can trigger the behaviour with fio by using userspace event reaping.  
> Adding a test case for that behaviour to libaio would be a good idea.

> I thought about how to fix this, and it isn't actually that hard.  Move 
> the put_reqs_available() call back into event consumption, and then add 
> code in the submit path to call put_reqs_available() if the system runs 
> out of events by noticing that there is free space in the event ring.  
> Something along the lines below should do it (please note, this is 
> completely untested!).  I'll test and polish this off tomorrow, as it's 
> getting a bit late here.

Dan, does this patch work for you?  It seems to pass your test program 
when I run it in a vm...

		-ben
> -- 
> "Thought is the essence of where you are now."
> 
> diff --git a/fs/aio.c b/fs/aio.c
> index ae63587..749ac22 100644
> --- a/fs/aio.c
> +++ b/fs/aio.c
> @@ -142,6 +142,7 @@ struct kioctx {
>  	struct {
>  		unsigned	tail;
>  		spinlock_t	completion_lock;
> +		unsigned	completed_events;
>  	} ____cacheline_aligned_in_smp;
>  
>  	struct page		*internal_pages[AIO_RING_PAGES];
> @@ -857,6 +858,31 @@ out:
>  	return ret;
>  }
>  
> +static void refill_reqs_available(struct kioctx *ctx)
> +{
> +	spin_lock_irq(&ctx->completion_lock);
> +	if (ctx->completed_events) {
> +		unsigned head, tail, avail, completed;
> +		struct aio_ring *ring;
> +
> +		ring = kmap_atomic(ctx->ring_pages[0]);
> +		head = ACCESS_ONCE(ring->head);
> +		tail = ACCESS_ONCE(ring->tail);
> +		kunmap_atomic(ring);
> +
> +		avail = (head <= tail ?  tail : ctx->nr_events) - head;
> +		completed = ctx->completed_events;
> +		if (avail < completed)
> +			completed -= avail;
> +		else
> +			completed = 0;
> +		put_reqs_available(ctx, completed);
> +	}
> +
> +	spin_unlock_irq(&ctx->completion_lock);
> +}
> +
> +
>  /* aio_get_req
>   *	Allocate a slot for an aio request.
>   * Returns NULL if no requests are free.
> @@ -865,8 +891,11 @@ static inline struct kiocb *aio_get_req(struct kioctx *ctx)
>  {
>  	struct kiocb *req;
>  
> -	if (!get_reqs_available(ctx))
> -		return NULL;
> +	if (!get_reqs_available(ctx)) {
> +		refill_reqs_available(ctx);
> +		if (!get_reqs_available(ctx))
> +			return NULL;
> +	}
>  
>  	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
>  	if (unlikely(!req))
> @@ -1005,7 +1034,6 @@ void aio_complete(struct kiocb *iocb, long res, long res2)
>  
>  	/* everything turned out well, dispose of the aiocb. */
>  	kiocb_free(iocb);
> -	put_reqs_available(ctx, 1);
>  
>  	/*
>  	 * We have to order our ring_info tail store above and test

-- 
"Thought is the essence of where you are now."