Re: [PATCH net-next v3 05/18] xdp: Use nested-BH locking for system_page_pool

["Toke Høiland-Jørgensen" <toke@redhat.com>]

Sebastian Andrzej Siewior <bigeasy@linutronix.de> writes:

> On 2025-05-01 12:13:24 [+0200], Toke Høiland-Jørgensen wrote:
>> > --- a/net/core/dev.c
>> > +++ b/net/core/dev.c
>> > @@ -462,7 +462,9 @@ EXPORT_PER_CPU_SYMBOL(softnet_data);
>> >   * PP consumers must pay attention to run APIs in the appropriate context
>> >   * (e.g. NAPI context).
>> >   */
>> > -DEFINE_PER_CPU(struct page_pool *, system_page_pool);
>> > +DEFINE_PER_CPU(struct page_pool_bh, system_page_pool) = {
>> > +	.bh_lock = INIT_LOCAL_LOCK(bh_lock),
>> > +};
>> 
>> I'm a little fuzzy on how DEFINE_PER_CPU() works, but does this
>> initialisation automatically do the right thing with the multiple
>> per-CPU instances?
>
> It sets the "first" per-CPU data which is then copied to all
> "possible-CPUs" during early boot when the per-CPU data is made
> available. You can initialize almost everything like that. Pointer based
> structures (such as LIST_HEAD_INIT()) is something that obviously won't
> work.

Right, I see. Cool, thanks for explaining :)

>> >  #ifdef CONFIG_LOCKDEP
>> >  /*
>> > --- a/net/core/xdp.c
>> > +++ b/net/core/xdp.c
>> > @@ -737,10 +737,10 @@ static noinline bool xdp_copy_frags_from_zc(struct sk_buff *skb,
>> >   */
>> >  struct sk_buff *xdp_build_skb_from_zc(struct xdp_buff *xdp)
>> >  {
>> > -	struct page_pool *pp = this_cpu_read(system_page_pool);
>> >  	const struct xdp_rxq_info *rxq = xdp->rxq;
>> >  	u32 len = xdp->data_end - xdp->data_meta;
>> >  	u32 truesize = xdp->frame_sz;
>> > +	struct page_pool *pp;
>> >  	struct sk_buff *skb;
>> >  	int metalen;
>> >  	void *data;
>> > @@ -748,13 +748,18 @@ struct sk_buff *xdp_build_skb_from_zc(struct xdp_buff *xdp)
>> >  	if (!IS_ENABLED(CONFIG_PAGE_POOL))
>> >  		return NULL;
>> >  
>> > +	local_lock_nested_bh(&system_page_pool.bh_lock);
>> > +	pp = this_cpu_read(system_page_pool.pool);
>> >  	data = page_pool_dev_alloc_va(pp, &truesize);
>> > -	if (unlikely(!data))
>> > +	if (unlikely(!data)) {
>> > +		local_unlock_nested_bh(&system_page_pool.bh_lock);
>> >  		return NULL;
>> > +	}
>> >  
>> >  	skb = napi_build_skb(data, truesize);
>> >  	if (unlikely(!skb)) {
>> >  		page_pool_free_va(pp, data, true);
>> > +		local_unlock_nested_bh(&system_page_pool.bh_lock);
>> >  		return NULL;
>> >  	}
>> >  
>> > @@ -773,9 +778,11 @@ struct sk_buff *xdp_build_skb_from_zc(struct xdp_buff *xdp)
>> >  
>> >  	if (unlikely(xdp_buff_has_frags(xdp)) &&
>> >  	    unlikely(!xdp_copy_frags_from_zc(skb, xdp, pp))) {
>> > +		local_unlock_nested_bh(&system_page_pool.bh_lock);
>> >  		napi_consume_skb(skb, true);
>> >  		return NULL;
>> >  	}
>> > +	local_unlock_nested_bh(&system_page_pool.bh_lock);
>> 
>> Hmm, instead of having four separate unlock calls in this function, how
>> about initialising skb = NULL, and having the unlock call just above
>> 'return skb' with an out: label?
>> 
>> Then the three topmost 'return NULL' can just straight-forwardly be
>> replaced with 'goto out', while the last one becomes 'skb = NULL; goto
>> out;'. I think that would be more readable than this repetition.
>
> Something like the following maybe? We would keep the lock during
> napi_consume_skb() which should work.
>
> diff --git a/net/core/xdp.c b/net/core/xdp.c
> index b2a5c934fe7b7..1ff0bc328305d 100644
> --- a/net/core/xdp.c
> +++ b/net/core/xdp.c
> @@ -740,8 +740,8 @@ struct sk_buff *xdp_build_skb_from_zc(struct xdp_buff *xdp)
>  	const struct xdp_rxq_info *rxq = xdp->rxq;
>  	u32 len = xdp->data_end - xdp->data_meta;
>  	u32 truesize = xdp->frame_sz;
> +	struct sk_buff *skb = NULL;
>  	struct page_pool *pp;
> -	struct sk_buff *skb;
>  	int metalen;
>  	void *data;
>  
> @@ -751,16 +751,13 @@ struct sk_buff *xdp_build_skb_from_zc(struct xdp_buff *xdp)
>  	local_lock_nested_bh(&system_page_pool.bh_lock);
>  	pp = this_cpu_read(system_page_pool.pool);
>  	data = page_pool_dev_alloc_va(pp, &truesize);
> -	if (unlikely(!data)) {
> -		local_unlock_nested_bh(&system_page_pool.bh_lock);
> -		return NULL;
> -	}
> +	if (unlikely(!data))
> +		goto out;
>  
>  	skb = napi_build_skb(data, truesize);
>  	if (unlikely(!skb)) {
>  		page_pool_free_va(pp, data, true);
> -		local_unlock_nested_bh(&system_page_pool.bh_lock);
> -		return NULL;
> +		goto out;
>  	}
>  
>  	skb_mark_for_recycle(skb);
> @@ -778,15 +775,16 @@ struct sk_buff *xdp_build_skb_from_zc(struct xdp_buff *xdp)
>  
>  	if (unlikely(xdp_buff_has_frags(xdp)) &&
>  	    unlikely(!xdp_copy_frags_from_zc(skb, xdp, pp))) {
> -		local_unlock_nested_bh(&system_page_pool.bh_lock);
>  		napi_consume_skb(skb, true);
> -		return NULL;
> +		skb = NULL;
>  	}
> +
> +out:
>  	local_unlock_nested_bh(&system_page_pool.bh_lock);
> -
> -	xsk_buff_free(xdp);
> -
> -	skb->protocol = eth_type_trans(skb, rxq->dev);
> +	if (skb) {
> +		xsk_buff_free(xdp);
> +		skb->protocol = eth_type_trans(skb, rxq->dev);
> +	}

I had in mind moving the out: label (and the unlock) below the
skb->protocol assignment, which would save the if(skb) check; any reason
we can't call xsk_buff_free() while holding the lock?

-Toke