Re: [bpf-next V2 PATCH 02/15] xdp: introduce xdp_return_frame API and use in cpumap

[Jason Wang <jasowang@redhat.com>]

On 2018年03月08日 21:07, Jesper Dangaard Brouer wrote:
> Introduce an xdp_return_frame API, and convert over cpumap as
> the first user, given it have queued XDP frame structure to leverage.
>
> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
> ---
>   include/net/xdp.h   |   32 +++++++++++++++++++++++++++
>   kernel/bpf/cpumap.c |   60 +++++++++++++++++++++++++++++++--------------------
>   net/core/xdp.c      |   18 +++++++++++++++
>   3 files changed, 86 insertions(+), 24 deletions(-)
>
> diff --git a/include/net/xdp.h b/include/net/xdp.h
> index b2362ddfa694..3cb726a6dc5b 100644
> --- a/include/net/xdp.h
> +++ b/include/net/xdp.h
> @@ -33,16 +33,48 @@
>    * also mandatory during RX-ring setup.
>    */
>   
> +enum mem_type {
> +	MEM_TYPE_PAGE_SHARED = 0, /* Split-page refcnt based model */
> +	MEM_TYPE_PAGE_ORDER0,     /* Orig XDP full page model */
> +	// Possible new ideas for types:
> +	// MEM_TYPE_PAGE_POOL,    /* Will be added later  */
> +	// MEM_TYPE_AF_XDP,
> +	// MEM_TYPE_DEVICE_OWNED -- invoking an dev->ndo?
> +	MEM_TYPE_MAX,
> +};

So if we plan to support dev->ndo, it looks to me two types AF_XDP and 
DEVICE_OWNED are sufficient? Driver can do what it wants (e.g page pool 
or ordinary page allocator) in ndo or what ever other callbacks.

> +
> +struct xdp_mem_info {
> +	u32 type; /* enum mem_type, but known size type */
> +	u32 id; // Needed later (to lookup struct xdp_rxq_info)
> +};
> +
>   struct xdp_rxq_info {
>   	struct net_device *dev;
>   	u32 queue_index;
>   	u32 reg_state;
> +	struct xdp_mem_info mem;
>   } ____cacheline_aligned; /* perf critical, avoid false-sharing */
>   
> +
> +static inline
> +void xdp_return_frame(void *data, struct xdp_mem_info *mem)
> +{
> +	if (mem->type == MEM_TYPE_PAGE_SHARED)
> +		page_frag_free(data);
> +
> +	if (mem->type == MEM_TYPE_PAGE_ORDER0) {
> +		struct page *page = virt_to_page(data); /* Assumes order0 page*/
> +
> +		put_page(page);
> +	}
> +}
> +
>   int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
>   		     struct net_device *dev, u32 queue_index);
>   void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq);
>   void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq);
>   bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq);
> +int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
> +			       enum mem_type type, void *allocator);
>   
>   #endif /* __LINUX_NET_XDP_H__ */
> diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
> index a4bb0b34375a..3e4bbcbe3e86 100644
> --- a/kernel/bpf/cpumap.c
> +++ b/kernel/bpf/cpumap.c
> @@ -19,6 +19,7 @@
>   #include <linux/bpf.h>
>   #include <linux/filter.h>
>   #include <linux/ptr_ring.h>
> +#include <net/xdp.h>
>   
>   #include <linux/sched.h>
>   #include <linux/workqueue.h>
> @@ -137,27 +138,6 @@ static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
>   	return ERR_PTR(err);
>   }
>   
> -static void __cpu_map_queue_destructor(void *ptr)
> -{
> -	/* The tear-down procedure should have made sure that queue is
> -	 * empty.  See __cpu_map_entry_replace() and work-queue
> -	 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
> -	 * gracefully and warn once.
> -	 */
> -	if (WARN_ON_ONCE(ptr))
> -		page_frag_free(ptr);
> -}
> -
> -static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
> -{
> -	if (atomic_dec_and_test(&rcpu->refcnt)) {
> -		/* The queue should be empty at this point */
> -		ptr_ring_cleanup(rcpu->queue, __cpu_map_queue_destructor);
> -		kfree(rcpu->queue);
> -		kfree(rcpu);
> -	}
> -}
> -
>   static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
>   {
>   	atomic_inc(&rcpu->refcnt);
> @@ -188,6 +168,10 @@ struct xdp_pkt {
>   	u16 len;
>   	u16 headroom;
>   	u16 metasize;
> +	/* Lifetime of xdp_rxq_info is limited to NAPI/enqueue time,
> +	 * while mem info is valid on remote CPU.
> +	 */

Can we simply move the xdp_mem_info to xdp_buff to avoid conversion?

Thanks

> +	struct xdp_mem_info mem;
>   	struct net_device *dev_rx;
>   };
>   
> @@ -213,6 +197,9 @@ static struct xdp_pkt *convert_to_xdp_pkt(struct xdp_buff *xdp)
>   	xdp_pkt->headroom = headroom - sizeof(*xdp_pkt);
>   	xdp_pkt->metasize = metasize;
>   
> +	/* rxq only valid until napi_schedule ends, convert to xdp_mem_info */
> +	xdp_pkt->mem = xdp->rxq->mem;
> +
>   	return xdp_pkt;
>   }
>   
> @@ -265,6 +252,31 @@ static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
>   	return skb;
>   }
>   
> +static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
> +{
> +	/* The tear-down procedure should have made sure that queue is
> +	 * empty.  See __cpu_map_entry_replace() and work-queue
> +	 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
> +	 * gracefully and warn once.
> +	 */
> +	struct xdp_pkt *xdp_pkt;
> +
> +	while ((xdp_pkt = ptr_ring_consume(ring)))
> +		if (WARN_ON_ONCE(xdp_pkt))
> +			xdp_return_frame(xdp_pkt, &xdp_pkt->mem);
> +}
> +
> +static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
> +{
> +	if (atomic_dec_and_test(&rcpu->refcnt)) {
> +		/* The queue should be empty at this point */
> +		__cpu_map_ring_cleanup(rcpu->queue);
> +		ptr_ring_cleanup(rcpu->queue, NULL);
> +		kfree(rcpu->queue);
> +		kfree(rcpu);
> +	}
> +}
> +
>   static int cpu_map_kthread_run(void *data)
>   {
>   	struct bpf_cpu_map_entry *rcpu = data;
> @@ -307,7 +319,7 @@ static int cpu_map_kthread_run(void *data)
>   
>   			skb = cpu_map_build_skb(rcpu, xdp_pkt);
>   			if (!skb) {
> -				page_frag_free(xdp_pkt);
> +				xdp_return_frame(xdp_pkt, &xdp_pkt->mem);
>   				continue;
>   			}
>   
> @@ -604,13 +616,13 @@ static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
>   	spin_lock(&q->producer_lock);
>   
>   	for (i = 0; i < bq->count; i++) {
> -		void *xdp_pkt = bq->q[i];
> +		struct xdp_pkt *xdp_pkt = bq->q[i];
>   		int err;
>   
>   		err = __ptr_ring_produce(q, xdp_pkt);
>   		if (err) {
>   			drops++;
> -			page_frag_free(xdp_pkt); /* Free xdp_pkt */
> +			xdp_return_frame(xdp_pkt->data, &xdp_pkt->mem);
>   		}
>   		processed++;
>   	}
> diff --git a/net/core/xdp.c b/net/core/xdp.c
> index 097a0f74e004..9eee0c431126 100644
> --- a/net/core/xdp.c
> +++ b/net/core/xdp.c
> @@ -71,3 +71,21 @@ bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
>   	return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
>   }
>   EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
> +
> +int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
> +			       enum mem_type type, void *allocator)
> +{
> +	if (type >= MEM_TYPE_MAX)
> +		return -EINVAL;
> +
> +	xdp_rxq->mem.type = type;
> +
> +	if (allocator)
> +		return -EOPNOTSUPP;
> +
> +	/* TODO: Allocate an ID that maps to allocator pointer
> +	 * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
> +	 */
> +	return 0;
> +}
> +EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
>