Re: [net-next PATCH v5 3/6] virtio_net: Add XDP support

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Dec 07, 2016 at 12:11:57PM -0800, John Fastabend wrote:
> From: John Fastabend <john.fastabend@gmail.com>
> 
> This adds XDP support to virtio_net. Some requirements must be
> met for XDP to be enabled depending on the mode. First it will
> only be supported with LRO disabled so that data is not pushed
> across multiple buffers. Second the MTU must be less than a page
> size to avoid having to handle XDP across multiple pages.
> 
> If mergeable receive is enabled this patch only supports the case
> where header and data are in the same buf which we can check when
> a packet is received by looking at num_buf. If the num_buf is
> greater than 1 and a XDP program is loaded the packet is dropped
> and a warning is thrown. When any_header_sg is set this does not
> happen and both header and data is put in a single buffer as expected
> so we check this when XDP programs are loaded.  Subsequent patches
> will process the packet in a degraded mode to ensure connectivity
> and correctness is not lost even if backend pushes packets into
> multiple buffers.
> 
> If big packets mode is enabled and MTU/LRO conditions above are
> met then XDP is allowed.
> 
> This patch was tested with qemu with vhost=on and vhost=off where
> mergeable and big_packet modes were forced via hard coding feature
> negotiation. Multiple buffers per packet was forced via a small
> test patch to vhost.c in the vhost=on qemu mode.
> 
> Suggested-by: Shrijeet Mukherjee <shrijeet@gmail.com>
> Signed-off-by: John Fastabend <john.r.fastabend@intel.com>

I'd like to note that I don't think disabling LRO is a good
plan long-term. It's really important for virtio performance,
so IMHO we need a fix for that.
I'm guessing that a subset of XDP programs would be quite
happy with just looking at headers, and that is there in the 1st buffer.
So how about teaching XDP that there could be a truncated packet?

Then we won't have to disable LRO.


> ---
>  drivers/net/virtio_net.c |  175 +++++++++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 170 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index a5c47b1..a009299 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -22,6 +22,7 @@
>  #include <linux/module.h>
>  #include <linux/virtio.h>
>  #include <linux/virtio_net.h>
> +#include <linux/bpf.h>
>  #include <linux/scatterlist.h>
>  #include <linux/if_vlan.h>
>  #include <linux/slab.h>
> @@ -81,6 +82,8 @@ struct receive_queue {
>  
>  	struct napi_struct napi;
>  
> +	struct bpf_prog __rcu *xdp_prog;
> +
>  	/* Chain pages by the private ptr. */
>  	struct page *pages;
>  
> @@ -324,6 +327,38 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
>  	return skb;
>  }
>  
> +static u32 do_xdp_prog(struct virtnet_info *vi,
> +		       struct bpf_prog *xdp_prog,
> +		       struct page *page, int offset, int len)
> +{
> +	int hdr_padded_len;
> +	struct xdp_buff xdp;
> +	u32 act;
> +	u8 *buf;
> +
> +	buf = page_address(page) + offset;
> +
> +	if (vi->mergeable_rx_bufs)
> +		hdr_padded_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
> +	else
> +		hdr_padded_len = sizeof(struct padded_vnet_hdr);
> +
> +	xdp.data = buf + hdr_padded_len;
> +	xdp.data_end = xdp.data + (len - vi->hdr_len);
> +
> +	act = bpf_prog_run_xdp(xdp_prog, &xdp);
> +	switch (act) {
> +	case XDP_PASS:
> +		return XDP_PASS;
> +	default:
> +		bpf_warn_invalid_xdp_action(act);
> +	case XDP_TX:
> +	case XDP_ABORTED:
> +	case XDP_DROP:
> +		return XDP_DROP;
> +	}
> +}
> +
>  static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
>  {
>  	struct sk_buff * skb = buf;
> @@ -340,14 +375,32 @@ static struct sk_buff *receive_big(struct net_device *dev,
>  				   void *buf,
>  				   unsigned int len)
>  {
> +	struct bpf_prog *xdp_prog;
>  	struct page *page = buf;
> -	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
> +	struct sk_buff *skb;
>  
> +	rcu_read_lock();
> +	xdp_prog = rcu_dereference(rq->xdp_prog);
> +	if (xdp_prog) {
> +		struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
> +		u32 act;
> +
> +		if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
> +			goto err_xdp;
> +		act = do_xdp_prog(vi, xdp_prog, page, 0, len);
> +		if (act == XDP_DROP)
> +			goto err_xdp;
> +	}
> +	rcu_read_unlock();
> +
> +	skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
>  	if (unlikely(!skb))
>  		goto err;
>  
>  	return skb;
>  
> +err_xdp:
> +	rcu_read_unlock();
>  err:
>  	dev->stats.rx_dropped++;
>  	give_pages(rq, page);
> @@ -365,11 +418,42 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
>  	u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
>  	struct page *page = virt_to_head_page(buf);
>  	int offset = buf - page_address(page);
> -	unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
> +	struct sk_buff *head_skb, *curr_skb;
> +	struct bpf_prog *xdp_prog;
> +	unsigned int truesize;
> +
> +	rcu_read_lock();
> +	xdp_prog = rcu_dereference(rq->xdp_prog);
> +	if (xdp_prog) {
> +		u32 act;
> +
> +		/* No known backend devices should send packets with
> +		 * more than a single buffer when XDP conditions are
> +		 * met. However it is not strictly illegal so the case
> +		 * is handled as an exception and a warning is thrown.
> +		 */
> +		if (unlikely(num_buf > 1)) {
> +			bpf_warn_invalid_xdp_buffer();
> +			goto err_xdp;
> +		}
>  
> -	struct sk_buff *head_skb = page_to_skb(vi, rq, page, offset, len,
> -					       truesize);
> -	struct sk_buff *curr_skb = head_skb;
> +		/* Transient failure which in theory could occur if
> +		 * in-flight packets from before XDP was enabled reach
> +		 * the receive path after XDP is loaded. In practice I
> +		 * was not able to create this condition.
> +		 */
> +		if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
> +			goto err_xdp;
> +
> +		act = do_xdp_prog(vi, xdp_prog, page, offset, len);
> +		if (act == XDP_DROP)
> +			goto err_xdp;
> +	}
> +	rcu_read_unlock();
> +
> +	truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
> +	head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
> +	curr_skb = head_skb;
>  
>  	if (unlikely(!curr_skb))
>  		goto err_skb;
> @@ -423,6 +507,8 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
>  	ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
>  	return head_skb;
>  
> +err_xdp:
> +	rcu_read_unlock();
>  err_skb:
>  	put_page(page);
>  	while (--num_buf) {
> @@ -1328,6 +1414,13 @@ static int virtnet_set_channels(struct net_device *dev,
>  	if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
>  		return -EINVAL;
>  
> +	/* For now we don't support modifying channels while XDP is loaded
> +	 * also when XDP is loaded all RX queues have XDP programs so we only
> +	 * need to check a single RX queue.
> +	 */
> +	if (vi->rq[0].xdp_prog)
> +		return -EINVAL;
> +
>  	get_online_cpus();
>  	err = virtnet_set_queues(vi, queue_pairs);
>  	if (!err) {
> @@ -1449,6 +1542,69 @@ static int virtnet_set_features(struct net_device *netdev,
>  	return 0;
>  }
>  
> +static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog)
> +{
> +	unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
> +	struct virtnet_info *vi = netdev_priv(dev);
> +	struct bpf_prog *old_prog;
> +	int i;
> +
> +	if ((dev->features & NETIF_F_LRO) && prog) {
> +		netdev_warn(dev, "can't set XDP while LRO is on, disable LRO first\n");
> +		return -EINVAL;
> +	}
> +
> +	if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
> +		netdev_warn(dev, "XDP expects header/data in single page\n");
> +		return -EINVAL;
> +	}
> +
> +	if (dev->mtu > max_sz) {
> +		netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
> +		return -EINVAL;
> +	}
> +
> +	if (prog) {
> +		prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
> +		if (IS_ERR(prog))
> +			return PTR_ERR(prog);
> +	}
> +
> +	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);
> +	}
> +
> +	return 0;
> +}
> +
> +static bool virtnet_xdp_query(struct net_device *dev)
> +{
> +	struct virtnet_info *vi = netdev_priv(dev);
> +	int i;
> +
> +	for (i = 0; i < vi->max_queue_pairs; i++) {
> +		if (vi->rq[i].xdp_prog)
> +			return true;
> +	}
> +	return false;
> +}
> +
> +static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
> +{
> +	switch (xdp->command) {
> +	case XDP_SETUP_PROG:
> +		return virtnet_xdp_set(dev, xdp->prog);
> +	case XDP_QUERY_PROG:
> +		xdp->prog_attached = virtnet_xdp_query(dev);
> +		return 0;
> +	default:
> +		return -EINVAL;
> +	}
> +}
> +
>  static const struct net_device_ops virtnet_netdev = {
>  	.ndo_open            = virtnet_open,
>  	.ndo_stop   	     = virtnet_close,
> @@ -1466,6 +1622,7 @@ static int virtnet_set_features(struct net_device *netdev,
>  	.ndo_busy_poll		= virtnet_busy_poll,
>  #endif
>  	.ndo_set_features	= virtnet_set_features,
> +	.ndo_xdp		= virtnet_xdp,
>  };
>  
>  static void virtnet_config_changed_work(struct work_struct *work)
> @@ -1527,12 +1684,20 @@ static void virtnet_free_queues(struct virtnet_info *vi)
>  
>  static void free_receive_bufs(struct virtnet_info *vi)
>  {
> +	struct bpf_prog *old_prog;
>  	int i;
>  
> +	rtnl_lock();
>  	for (i = 0; i < vi->max_queue_pairs; i++) {
>  		while (vi->rq[i].pages)
>  			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
> +
> +		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
> +		RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
> +		if (old_prog)
> +			bpf_prog_put(old_prog);
>  	}
> +	rtnl_unlock();
>  }
>  
>  static void free_receive_page_frags(struct virtnet_info *vi)