Re: [BUG Report] KASAN: slab-use-after-free in page_pool_recycle_in_ring

[Jakub Kicinski <kuba@kernel.org>]

On Thu, 22 May 2025 23:17:32 +0800 dongchenchen (A) wrote:
> Hi, Jakub
> Maybe we can fix the problem as follow:

Yes! a couple of minor nit picks below..

> diff --git a/net/core/page_pool.c b/net/core/page_pool.c
> index 7745ad924ae2..de3fa33d6775 100644
> --- a/net/core/page_pool.c
> +++ b/net/core/page_pool.c
> @@ -707,19 +707,18 @@ void page_pool_return_page(struct page_pool *pool, netmem_ref netmem)
>   
>   static bool page_pool_recycle_in_ring(struct page_pool *pool, netmem_ref netmem)
>   {
> +	bool in_softirq;
>   	int ret;
> -	/* BH protection not needed if current is softirq */
> -	if (in_softirq())
> -		ret = ptr_ring_produce(&pool->ring, (__force void *)netmem);
> -	else
> -		ret = ptr_ring_produce_bh(&pool->ring, (__force void *)netmem);
>   
> -	if (!ret) {
> +	/* BH protection not needed if current is softirq */
> +	in_softirq = page_pool_producer_lock(pool);
> +	ret = __ptr_ring_produce(&pool->ring, (__force void *)netmem);

Maybe we can flip the return value here we won't have to negate it below
and at return? Like this:

	ret = !__ptr_ring_produce(&pool->ring, (__force void *)netmem);

and adjust subsequent code

> +	if (!ret)
>   		recycle_stat_inc(pool, ring);
> -		return true;
> -	}
>   
> -	return false;
> +	page_pool_producer_unlock(pool, in_softirq);
> +
> +	return ret ? false : true;
>   }
>   
>   /* Only allow direct recycling in special circumstances, into the
> 
> @@ -1091,10 +1090,16 @@ static void page_pool_scrub(struct page_pool *pool)
>   
>   static int page_pool_release(struct page_pool *pool)
>   {
> +	bool in_softirq;
>   	int inflight;
>   
> +	/* Acquire producer lock to make sure we don't race with another thread
> +	 * returning a netmem to the ptr_ring.
> +	 */
> +	in_softirq = page_pool_producer_lock(pool);
>   	page_pool_scrub(pool);
>   	inflight = page_pool_inflight(pool, true);
> +	page_pool_producer_unlock(pool, in_softirq);

As I suggested earlier we don't have to lock the consumer, taking both
locks has lock ordering implications. My preference would be:

  	page_pool_scrub(pool);
  	inflight = page_pool_inflight(pool, true);
+	/* Acquire producer lock to make sure producers have exited. */
+	in_softirq = page_pool_producer_lock(pool);
+	page_pool_producer_unlock(pool, in_softirq);