Re: [PATCH net-next 4/5] virtio-net: do not reset during XDP set

[Michael S. Tsirkin]

On Mon, Jul 17, 2017 at 08:44:00PM +0800, Jason Wang wrote:
> We used to reset during XDP set, the main reason is we need allocate
> extra headroom for header adjustment but there's no way to know the
> headroom of exist receive buffer. This works buy maybe complex and may
> cause the network down for a while which is bad for user
> experience. So this patch tries to avoid this by:
> 
> - packing headroom into receive buffer ctx
> - check the headroom during XDP, and if it was not sufficient, copy
>   the packet into a location which has a large enough headroom

The packing is actually done by previous patches. Here is a
corrected version:

We currently reset the device during XDP set, the main reason is
that we allocate more headroom with XDP (for header adjustment).

This works but causes network downtime for users.

Previous patches encoded the headroom in the buffer context,
this makes it possible to detect the case where a buffer
with headroom insufficient for XDP is added to the queue and
XDP is enabled afterwards.

Upon detection, we handle this case by copying the packet
(slow, but it's a temporary condition).



> Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx>
> ---
>  drivers/net/virtio_net.c | 230 ++++++++++++++++++++++-------------------------
>  1 file changed, 105 insertions(+), 125 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index e31b5b2..e732bd6 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -407,6 +407,67 @@ static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
>  	return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
>  }
>  
> +/* We copy and linearize packet in the following cases:
> + *
> + * 1) Packet across multiple buffers, this happens normally when rx
> + *    buffer size is underestimated. Rarely, since spec does not
> + *    forbid using more than one buffer even if a single buffer is
> + *    sufficient for the packet, we should also deal with this case.

Latest SVN of the spec actually forbids this. See:
    net: clarify device rules for mergeable buffers


> + * 2) The header room is smaller than what XDP required. In this case
> + *    we should copy the packet and reserve enough headroom for this.
> + *    This would be slow but we at most we can copy times of queue
> + *    size, this is acceptable. What's more important, this help to
> + *    avoid resetting.

Last part of the comment applies to both cases. So

+/* We copy the packet for XDP in the following cases:
+ *
+ * 1) Packet is scattered across multiple rx buffers.
+ * 2) Headroom space is insufficient.
+ *
+ * This is inefficient but it's a temporary condition that
+ * we hit right after XDP is enabled and until queue is refilled
+ * with large buffers with sufficient headroom - so it should affect
+ * at most queue size packets.

+ * Afterwards, the conditions to enable
+ * XDP should preclude the underlying device from sending packets
+ * across multiple buffers (num_buf > 1), and we make sure buffers
+ * have enough headroom.
+ */



> + * 2) The header room is smaller than what XDP required. In this case
> + *    we should copy the packet and reserve enough headroom for this.
> + *    This would be slow but we at most we can copy times of queue
> + *    size, this is acceptable. What's more important, this help to
> + *    avoid resetting.



> + */
> +static struct page *xdp_linearize_page(struct receive_queue *rq,
> +				       u16 *num_buf,
> +				       struct page *p,
> +				       int offset,
> +				       int page_off,
> +				       unsigned int *len)
> +{
> +	struct page *page = alloc_page(GFP_ATOMIC);
> +
> +	if (!page)
> +		return NULL;
> +
> +	memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
> +	page_off += *len;
> +
> +	while (--*num_buf) {
> +		unsigned int buflen;
> +		void *buf;
> +		int off;
> +
> +		buf = virtqueue_get_buf(rq->vq, &buflen);
> +		if (unlikely(!buf))
> +			goto err_buf;
> +
> +		p = virt_to_head_page(buf);
> +		off = buf - page_address(p);
> +
> +		/* guard against a misconfigured or uncooperative backend that
> +		 * is sending packet larger than the MTU.
> +		 */
> +		if ((page_off + buflen) > PAGE_SIZE) {
> +			put_page(p);
> +			goto err_buf;
> +		}
> +
> +		memcpy(page_address(page) + page_off,
> +		       page_address(p) + off, buflen);
> +		page_off += buflen;
> +		put_page(p);
> +	}
> +
> +	/* Headroom does not contribute to packet length */
> +	*len = page_off - VIRTIO_XDP_HEADROOM;
> +	return page;
> +err_buf:
> +	__free_pages(page, 0);
> +	return NULL;
> +}
> +
>  static struct sk_buff *receive_small(struct net_device *dev,
>  				     struct virtnet_info *vi,
>  				     struct receive_queue *rq,
> @@ -415,12 +476,14 @@ static struct sk_buff *receive_small(struct net_device *dev,
>  {
>  	struct sk_buff *skb;
>  	struct bpf_prog *xdp_prog;
> -	unsigned int xdp_headroom = virtnet_get_headroom(vi);
> +	unsigned int xdp_headroom = (unsigned long)ctx;
>  	unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
>  	unsigned int headroom = vi->hdr_len + header_offset;
>  	unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
>  			      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
> +	struct page *page = virt_to_head_page(buf);
>  	unsigned int delta = 0;
> +	struct page *xdp_page;
>  	len -= vi->hdr_len;
>  
>  	rcu_read_lock();
> @@ -434,6 +497,27 @@ static struct sk_buff *receive_small(struct net_device *dev,
>  		if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
>  			goto err_xdp;
>  
> +		if (unlikely(xdp_headroom != virtnet_get_headroom(vi))) {

Should this be xdp_headroom < virtnet_get_headroom(vi)?
Just in case we add more modes down the road.


> +			int offset = buf - page_address(page) + header_offset;
> +			unsigned int tlen = len + vi->hdr_len;
> +			u16 num_buf = 1;
> +
> +			xdp_headroom = virtnet_get_headroom(vi);
> +			header_offset = VIRTNET_RX_PAD + xdp_headroom;
> +			headroom = vi->hdr_len + header_offset;
> +			buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
> +				 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
> +			xdp_page = xdp_linearize_page(rq, &num_buf, page,
> +						      offset, header_offset,
> +						      &tlen);
> +			if (!xdp_page)
> +				goto err_xdp;
> +
> +			buf = page_address(xdp_page);
> +			put_page(page);
> +			page = xdp_page;
> +		}
> +
>  		xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
>  		xdp.data = xdp.data_hard_start + xdp_headroom;
>  		xdp.data_end = xdp.data + len;
> @@ -462,7 +546,7 @@ static struct sk_buff *receive_small(struct net_device *dev,
>  
>  	skb = build_skb(buf, buflen);
>  	if (!skb) {
> -		put_page(virt_to_head_page(buf));
> +		put_page(page);
>  		goto err;
>  	}
>  	skb_reserve(skb, headroom - delta);
> @@ -478,7 +562,7 @@ static struct sk_buff *receive_small(struct net_device *dev,
>  err_xdp:
>  	rcu_read_unlock();
>  	dev->stats.rx_dropped++;
> -	put_page(virt_to_head_page(buf));
> +	put_page(page);
>  xdp_xmit:
>  	return NULL;
>  }
> @@ -503,66 +587,6 @@ static struct sk_buff *receive_big(struct net_device *dev,
>  	return NULL;
>  }
>  
> -/* The conditions to enable XDP should preclude the underlying device from
> - * sending packets across multiple buffers (num_buf > 1). However per spec
> - * it does not appear to be illegal to do so but rather just against convention.
> - * So in order to avoid making a system unresponsive the packets are pushed
> - * into a page and the XDP program is run. This will be extremely slow and we
> - * push a warning to the user to fix this as soon as possible. Fixing this may
> - * require resolving the underlying hardware to determine why multiple buffers
> - * are being received or simply loading the XDP program in the ingress stack
> - * after the skb is built because there is no advantage to running it here
> - * anymore.
> - */
> -static struct page *xdp_linearize_page(struct receive_queue *rq,
> -				       u16 *num_buf,
> -				       struct page *p,
> -				       int offset,
> -				       unsigned int *len)
> -{
> -	struct page *page = alloc_page(GFP_ATOMIC);
> -	unsigned int page_off = VIRTIO_XDP_HEADROOM;
> -
> -	if (!page)
> -		return NULL;
> -
> -	memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
> -	page_off += *len;
> -
> -	while (--*num_buf) {
> -		unsigned int buflen;
> -		void *buf;
> -		int off;
> -
> -		buf = virtqueue_get_buf(rq->vq, &buflen);
> -		if (unlikely(!buf))
> -			goto err_buf;
> -
> -		p = virt_to_head_page(buf);
> -		off = buf - page_address(p);
> -
> -		/* guard against a misconfigured or uncooperative backend that
> -		 * is sending packet larger than the MTU.
> -		 */
> -		if ((page_off + buflen) > PAGE_SIZE) {
> -			put_page(p);
> -			goto err_buf;
> -		}
> -
> -		memcpy(page_address(page) + page_off,
> -		       page_address(p) + off, buflen);
> -		page_off += buflen;
> -		put_page(p);
> -	}
> -
> -	/* Headroom does not contribute to packet length */
> -	*len = page_off - VIRTIO_XDP_HEADROOM;
> -	return page;
> -err_buf:
> -	__free_pages(page, 0);
> -	return NULL;
> -}
> -
>  static struct sk_buff *receive_mergeable(struct net_device *dev,
>  					 struct virtnet_info *vi,
>  					 struct receive_queue *rq,
> @@ -577,6 +601,7 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
>  	struct sk_buff *head_skb, *curr_skb;
>  	struct bpf_prog *xdp_prog;
>  	unsigned int truesize;
> +	unsigned int headroom = mergeable_ctx_to_headroom(ctx);
>  
>  	head_skb = NULL;
>  
> @@ -589,10 +614,13 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
>  		u32 act;
>  
>  		/* This happens when rx buffer size is underestimated */
> -		if (unlikely(num_buf > 1)) {
> +		if (unlikely(num_buf > 1 ||
> +			     headroom < virtnet_get_headroom(vi))) {
>  			/* linearize data for XDP */
>  			xdp_page = xdp_linearize_page(rq, &num_buf,
> -						      page, offset, &len);
> +						      page, offset,
> +						      VIRTIO_XDP_HEADROOM,
> +						      &len);
>  			if (!xdp_page)
>  				goto err_xdp;
>  			offset = VIRTIO_XDP_HEADROOM;
> @@ -830,7 +858,6 @@ static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
>  	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
>  	if (err < 0)
>  		put_page(virt_to_head_page(buf));
> -
>  	return err;
>  }
>  
> @@ -1834,7 +1861,6 @@ static void virtnet_freeze_down(struct virtio_device *vdev)
>  }
>  
>  static int init_vqs(struct virtnet_info *vi);
> -static void _remove_vq_common(struct virtnet_info *vi);
>  
>  static int virtnet_restore_up(struct virtio_device *vdev)
>  {
> @@ -1863,39 +1889,6 @@ static int virtnet_restore_up(struct virtio_device *vdev)
>  	return err;
>  }
>  
> -static int virtnet_reset(struct virtnet_info *vi, int curr_qp, int xdp_qp)
> -{
> -	struct virtio_device *dev = vi->vdev;
> -	int ret;
> -
> -	virtio_config_disable(dev);
> -	dev->failed = dev->config->get_status(dev) & VIRTIO_CONFIG_S_FAILED;
> -	virtnet_freeze_down(dev);
> -	_remove_vq_common(vi);
> -
> -	virtio_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE);
> -	virtio_add_status(dev, VIRTIO_CONFIG_S_DRIVER);
> -
> -	ret = virtio_finalize_features(dev);
> -	if (ret)
> -		goto err;
> -
> -	vi->xdp_queue_pairs = xdp_qp;
> -	ret = virtnet_restore_up(dev);
> -	if (ret)
> -		goto err;
> -	ret = _virtnet_set_queues(vi, curr_qp);
> -	if (ret)
> -		goto err;
> -
> -	virtio_add_status(dev, VIRTIO_CONFIG_S_DRIVER_OK);
> -	virtio_config_enable(dev);
> -	return 0;
> -err:
> -	virtio_add_status(dev, VIRTIO_CONFIG_S_FAILED);
> -	return ret;
> -}
> -
>  static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
>  			   struct netlink_ext_ack *extack)
>  {
> @@ -1942,35 +1935,31 @@ static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
>  			return PTR_ERR(prog);
>  	}
>  
> -	/* Changing the headroom in buffers is a disruptive operation because
> -	 * existing buffers must be flushed and reallocated. This will happen
> -	 * when a xdp program is initially added or xdp is disabled by removing
> -	 * the xdp program resulting in number of XDP queues changing.
> +	/* synchronize with NAPI which may do XDP_TX based on queue
> +	 * pair numbers.

I think you mean

 	/* Make sure NAPI is not using any XDP TX queues for RX. */

is that it?


> -	if (vi->xdp_queue_pairs != xdp_qp) {
> -		err = virtnet_reset(vi, curr_qp + xdp_qp, xdp_qp);
> -		if (err) {
> -			dev_warn(&dev->dev, "XDP reset failure.\n");
> -			goto virtio_reset_err;
> -		}
> -	}
> +	for (i = 0; i < vi->max_queue_pairs; i++)
> +		napi_disable(&vi->rq[i].napi);
>  

This is pretty slow if queues are busy.  Should we avoid this for queues
which aren't effected?


>  	netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
> +	err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
> +	if (err)
> +		goto err;
> +	vi->xdp_queue_pairs = xdp_qp;
>  
>  	for (i = 0; i < vi->max_queue_pairs; i++) {
>  		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
>  		rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
>  		if (old_prog)
>  			bpf_prog_put(old_prog);
> +		napi_enable(&vi->rq[i].napi);

This seems racy. See comment around virtnet_napi_enable.

>  	}
>  
>  	return 0;
>  
> -virtio_reset_err:
> -	/* On reset error do our best to unwind XDP changes inflight and return
> -	 * error up to user space for resolution. The underlying reset hung on
> -	 * us so not much we can do here.
> -	 */
> +err:
> +	for (i = 0; i < vi->max_queue_pairs; i++)
> +		napi_enable(&vi->rq[i].napi);
>  	if (prog)
>  		bpf_prog_sub(prog, vi->max_queue_pairs - 1);
>  	return err;
> @@ -2614,15 +2603,6 @@ static int virtnet_probe(struct virtio_device *vdev)
>  	return err;
>  }
>  
> -static void _remove_vq_common(struct virtnet_info *vi)
> -{
> -	vi->vdev->config->reset(vi->vdev);
> -	free_unused_bufs(vi);
> -	_free_receive_bufs(vi);
> -	free_receive_page_frags(vi);
> -	virtnet_del_vqs(vi);
> -}
> -
>  static void remove_vq_common(struct virtnet_info *vi)
>  {
>  	vi->vdev->config->reset(vi->vdev);
> -- 
> 2.7.4