Virtualization - Jul 2017 - [PATCH net-next 2/5] virtio-net: pack headroom into ctx for mergeable buffer

Hi:

This series brings two optimizations for virtio-net XDP:

- avoid reset during XDP set
- turn off offloads on demand

Please review.

Thanks

Jason Wang (5):
  virtio_ring: allow to store zero as the ctx
  virtio-net: pack headroom into ctx for mergeable buffer
  virtio-net: switch to use new ctx API for small buffer
  virtio-net: do not reset during XDP set
  virtio-net: switch off offloads on demand if possible on XDP set

 drivers/net/virtio_net.c     | 325 +++++++++++++++++++++++++------------------
 drivers/virtio/virtio_ring.c |   2 +-
 2 files changed, 194 insertions(+), 133 deletions(-)

-- 
2.7.4

Allow zero to be store as a ctx, with this we could store e.g zero
value which could be meaningful for the case of storing headroom
through ctx.

Signed-off-by: Jason Wang <jasowang at redhat.com>
---
 drivers/virtio/virtio_ring.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 5e1b548..9aaa177 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -391,7 +391,7 @@ static inline int virtqueue_add(struct virtqueue *_vq,
 	vq->desc_state[head].data = data;
 	if (indirect)
 		vq->desc_state[head].indir_desc = desc;
-	if (ctx)
+	else
 		vq->desc_state[head].indir_desc = ctx;
 
 	/* Put entry in available array (but don't update avail->idx until they
-- 
2.7.4

Pack headroom into ctx, then during XDP set, we could know the size of
headroom and copy if needed. This is required for avoiding reset on
XDP.

Signed-off-by: Jason Wang <jasowang at redhat.com>
---
 drivers/net/virtio_net.c | 29 ++++++++++++++++++++++++-----
 1 file changed, 24 insertions(+), 5 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 1f8c15c..8fae9a8 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -270,6 +270,23 @@ static void skb_xmit_done(struct virtqueue *vq)
 		netif_wake_subqueue(vi->dev, vq2txq(vq));
 }
 
+#define MRG_CTX_HEADER_SHIFT 22
+static void *mergeable_len_to_ctx(unsigned int truesize,
+				  unsigned int headroom)
+{
+	return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) |
truesize);
+}
+
+static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
+{
+	return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
+}
+
+static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
+{
+	return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
+}
+
 /* Called from bottom half context */
 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 				   struct receive_queue *rq,
@@ -639,13 +656,14 @@ static struct sk_buff *receive_mergeable(struct net_device
*dev,
 	}
 	rcu_read_unlock();
 
-	if (unlikely(len > (unsigned long)ctx)) {
+	truesize = mergeable_ctx_to_truesize(ctx);
+	if (unlikely(len > truesize)) {
 		pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 			 dev->name, len, (unsigned long)ctx);
 		dev->stats.rx_length_errors++;
 		goto err_skb;
 	}
-	truesize = (unsigned long)ctx;
+
 	head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
 	curr_skb = head_skb;
 
@@ -665,13 +683,14 @@ static struct sk_buff *receive_mergeable(struct net_device
*dev,
 		}
 
 		page = virt_to_head_page(buf);
-		if (unlikely(len > (unsigned long)ctx)) {
+
+		truesize = mergeable_ctx_to_truesize(ctx);
+		if (unlikely(len > truesize)) {
 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 				 dev->name, len, (unsigned long)ctx);
 			dev->stats.rx_length_errors++;
 			goto err_skb;
 		}
-		truesize = (unsigned long)ctx;
 
 		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 		if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
@@ -889,7 +908,7 @@ static int add_recvbuf_mergeable(struct virtnet_info *vi,
 
 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 	buf += headroom; /* advance address leaving hole at front of pkt */
-	ctx = (void *)(unsigned long)len;
+	ctx = mergeable_len_to_ctx(len, headroom);
 	get_page(alloc_frag->page);
 	alloc_frag->offset += len + headroom;
 	hole = alloc_frag->size - alloc_frag->offset;
-- 
2.7.4

Switch to use ctx API for small buffer, this is need for avoiding
reset on XDP.

Signed-off-by: Jason Wang <jasowang at redhat.com>
---
 drivers/net/virtio_net.c | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 8fae9a8..e31b5b2 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -410,7 +410,8 @@ static unsigned int virtnet_get_headroom(struct virtnet_info
*vi)
 static struct sk_buff *receive_small(struct net_device *dev,
 				     struct virtnet_info *vi,
 				     struct receive_queue *rq,
-				     void *buf, unsigned int len)
+				     void *buf, void *ctx,
+				     unsigned int len)
 {
 	struct sk_buff *skb;
 	struct bpf_prog *xdp_prog;
@@ -773,7 +774,7 @@ static int receive_buf(struct virtnet_info *vi, struct
receive_queue *rq,
 	else if (vi->big_packets)
 		skb = receive_big(dev, vi, rq, buf, len);
 	else
-		skb = receive_small(dev, vi, rq, buf, len);
+		skb = receive_small(dev, vi, rq, buf, ctx, len);
 
 	if (unlikely(!skb))
 		return 0;
@@ -812,6 +813,7 @@ static int add_recvbuf_small(struct virtnet_info *vi, struct
receive_queue *rq,
 	struct page_frag *alloc_frag = &rq->alloc_frag;
 	char *buf;
 	unsigned int xdp_headroom = virtnet_get_headroom(vi);
+	void *ctx = (void *)(unsigned long)xdp_headroom;
 	int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
 	int err;
 
@@ -825,7 +827,7 @@ static int add_recvbuf_small(struct virtnet_info *vi, struct
receive_queue *rq,
 	alloc_frag->offset += len;
 	sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
 		    vi->hdr_len + GOOD_PACKET_LEN);
-	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
+	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
 	if (err < 0)
 		put_page(virt_to_head_page(buf));
 
@@ -1034,7 +1036,7 @@ static int virtnet_receive(struct receive_queue *rq, int
budget)
 	void *buf;
 	struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 
-	if (vi->mergeable_rx_bufs) {
+	if (!vi->big_packets || vi->mergeable_rx_bufs) {
 		void *ctx;
 
 		while (received < budget &&
@@ -2198,7 +2200,7 @@ static int virtnet_find_vqs(struct virtnet_info *vi)
 	names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
 	if (!names)
 		goto err_names;
-	if (vi->mergeable_rx_bufs) {
+	if (!vi->big_packets || vi->mergeable_rx_bufs) {
 		ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
 		if (!ctx)
 			goto err_ctx;
-- 
2.7.4

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

Signed-off-by: Jason Wang <jasowang at redhat.com>
---
 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.
+ * 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))) {
+			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.
 	 */
-	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);
 
 	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);
 	}
 
 	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

Current XDP implementation want guest offloads feature to be disabled
on qemu cli. This is inconvenient and means guest can't benefit from
offloads if XDP is not used. This patch tries to address this
limitation by disable the offloads on demand through control guest
offloads. Guest offloads will be disabled and enabled on demand on XDP
set.

Signed-off-by: Jason Wang <jasowang at redhat.com>
---
 drivers/net/virtio_net.c | 70 ++++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 65 insertions(+), 5 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index e732bd6..d970c2d 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -57,6 +57,11 @@ DECLARE_EWMA(pkt_len, 0, 64)
 
 #define VIRTNET_DRIVER_VERSION "1.0.0"
 
+const unsigned long guest_offloads[] = { VIRTIO_NET_F_GUEST_TSO4,
+					 VIRTIO_NET_F_GUEST_TSO6,
+					 VIRTIO_NET_F_GUEST_ECN,
+					 VIRTIO_NET_F_GUEST_UFO };
+
 struct virtnet_stats {
 	struct u64_stats_sync tx_syncp;
 	struct u64_stats_sync rx_syncp;
@@ -164,10 +169,13 @@ struct virtnet_info {
 	u8 ctrl_promisc;
 	u8 ctrl_allmulti;
 	u16 ctrl_vid;
+	u64 ctrl_offloads;
 
 	/* Ethtool settings */
 	u8 duplex;
 	u32 speed;
+
+	unsigned long guest_offloads;
 };
 
 struct padded_vnet_hdr {
@@ -1889,6 +1897,47 @@ static int virtnet_restore_up(struct virtio_device *vdev)
 	return err;
 }
 
+static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
+{
+	struct scatterlist sg;
+	vi->ctrl_offloads = cpu_to_virtio64(vi->vdev, offloads);
+
+	sg_init_one(&sg, &vi->ctrl_offloads, sizeof(vi->ctrl_offloads));
+
+	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
+				  VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
+		dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
+static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
+{
+	u64 offloads = 0;
+
+	if (!vi->guest_offloads)
+		return 0;
+
+	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
+		offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
+
+	return virtnet_set_guest_offloads(vi, offloads);
+}
+
+static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
+{
+	u64 offloads = vi->guest_offloads;
+
+	if (!vi->guest_offloads)
+		return 0;
+	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
+		offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
+
+	return virtnet_set_guest_offloads(vi, offloads);
+}
+
 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
 			   struct netlink_ext_ack *extack)
 {
@@ -1898,10 +1947,11 @@ static int virtnet_xdp_set(struct net_device *dev,
struct bpf_prog *prog,
 	u16 xdp_qp = 0, curr_qp;
 	int i, err;
 
-	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
-	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
-	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
-	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO)) {
+	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
+	    && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
+	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
+	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
+		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
 		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing
LRO, disable LRO first");
 		return -EOPNOTSUPP;
 	}
@@ -1950,6 +2000,12 @@ static int virtnet_xdp_set(struct net_device *dev, struct
bpf_prog *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 (i == 0) {
+			if (!old_prog)
+				virtnet_clear_guest_offloads(vi);
+			if (!prog)
+				virtnet_restore_guest_offloads(vi);
+		}
 		if (old_prog)
 			bpf_prog_put(old_prog);
 		napi_enable(&vi->rq[i].napi);
@@ -2583,6 +2639,10 @@ static int virtnet_probe(struct virtio_device *vdev)
 		netif_carrier_on(dev);
 	}
 
+	for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
+		if (virtio_has_feature(vi->vdev, guest_offloads[i]))
+			set_bit(guest_offloads[i], &vi->guest_offloads);
+
 	pr_debug("virtnet: registered device %s with %d RX and TX
vq's\n",
 		 dev->name, max_queue_pairs);
 
@@ -2679,7 +2739,7 @@ static struct virtio_device_id id_table[] = {
 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
 	VIRTIO_NET_F_CTRL_MAC_ADDR, \
-	VIRTIO_NET_F_MTU
+	VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
 
 static unsigned int features[] = {
 	VIRTNET_FEATURES,
-- 
2.7.4

From: Jason Wang <jasowang at redhat.com>
Date: Mon, 17 Jul 2017 20:43:56 +0800
> This series brings two optimizations for virtio-net XDP:
> 
> - avoid reset during XDP set
> - turn off offloads on demand
> 
> Please review.
Michael, please review Jason's changes.

Thanks.

On Tue, Jul 18, 2017 at 11:24:42AM -0700, David Miller
wrote:
> From: Jason Wang <jasowang at redhat.com>
> Date: Mon, 17 Jul 2017 20:43:56 +0800
> 
> > This series brings two optimizations for virtio-net XDP:
> > 
> > - avoid reset during XDP set
> > - turn off offloads on demand
> > 
> > Please review.
> 
> Michael, please review Jason's changes.
> 
> Thanks.
Doing that, thanks for the reminder.

-- 
MST

On Mon, Jul 17, 2017 at 08:43:58PM +0800, Jason Wang
wrote:
> Pack headroom into ctx, then during XDP set, we could know the size of
> headroom and copy if needed. This is required for avoiding reset on
> XDP.
Not really when XDP is set - it's when buffers are used.

virtio-net: pack headroom into ctx for mergeable buffers

Pack headroom into ctx - this way when we get a buffer we can figure out
the actual headroom that was allocated for the buffer. Will be helpful
to optimize switching between XDP and non-XDP modes which have different
headroom requirements.

> 
> Signed-off-by: Jason Wang <jasowang at redhat.com>
> ---
>  drivers/net/virtio_net.c | 29 ++++++++++++++++++++++++-----
>  1 file changed, 24 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 1f8c15c..8fae9a8 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -270,6 +270,23 @@ static void skb_xmit_done(struct virtqueue *vq)
>  		netif_wake_subqueue(vi->dev, vq2txq(vq));
>  }
>  
> +#define MRG_CTX_HEADER_SHIFT 22
> +static void *mergeable_len_to_ctx(unsigned int truesize,
> +				  unsigned int headroom)
> +{
> +	return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) |
truesize);
> +}
> +
> +static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
> +{
> +	return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
> +}
> +
> +static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
> +{
> +	return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) -
1);
> +}
> +
>  /* Called from bottom half context */
>  static struct sk_buff *page_to_skb(struct virtnet_info *vi,
>  				   struct receive_queue *rq,
> @@ -639,13 +656,14 @@ static struct sk_buff *receive_mergeable(struct
net_device *dev,
>  	}
>  	rcu_read_unlock();
>  
> -	if (unlikely(len > (unsigned long)ctx)) {
> +	truesize = mergeable_ctx_to_truesize(ctx);
> +	if (unlikely(len > truesize)) {
>  		pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
>  			 dev->name, len, (unsigned long)ctx);
>  		dev->stats.rx_length_errors++;
>  		goto err_skb;
>  	}
> -	truesize = (unsigned long)ctx;
> +
>  	head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
>  	curr_skb = head_skb;
>  
> @@ -665,13 +683,14 @@ static struct sk_buff *receive_mergeable(struct
net_device *dev,
>  		}
>  
>  		page = virt_to_head_page(buf);
> -		if (unlikely(len > (unsigned long)ctx)) {
> +
> +		truesize = mergeable_ctx_to_truesize(ctx);
> +		if (unlikely(len > truesize)) {
>  			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
>  				 dev->name, len, (unsigned long)ctx);
>  			dev->stats.rx_length_errors++;
>  			goto err_skb;
>  		}
> -		truesize = (unsigned long)ctx;
>  
>  		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
>  		if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
> @@ -889,7 +908,7 @@ static int add_recvbuf_mergeable(struct virtnet_info
*vi,
>  
>  	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
>  	buf += headroom; /* advance address leaving hole at front of pkt */
> -	ctx = (void *)(unsigned long)len;
> +	ctx = mergeable_len_to_ctx(len, headroom);
>  	get_page(alloc_frag->page);
>  	alloc_frag->offset += len + headroom;
>  	hole = alloc_frag->size - alloc_frag->offset;
> -- 
> 2.7.4

what's needed is ability to store the headroom there.

virtio-net: switch to use ctx API for small buffers

Use ctx API to store headroom for small buffers.
Following patches will retrieve this info and use it for XDP.

On Mon, Jul 17, 2017 at 08:43:59PM +0800, Jason Wang
wrote:
> Switch to use ctx API for small buffer, this is need for avoiding
> reset on XDP.
> 
> Signed-off-by: Jason Wang <jasowang at redhat.com>
> ---
>  drivers/net/virtio_net.c | 12 +++++++-----
>  1 file changed, 7 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 8fae9a8..e31b5b2 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -410,7 +410,8 @@ static unsigned int virtnet_get_headroom(struct
virtnet_info *vi)
>  static struct sk_buff *receive_small(struct net_device *dev,
>  				     struct virtnet_info *vi,
>  				     struct receive_queue *rq,
> -				     void *buf, unsigned int len)
> +				     void *buf, void *ctx,
> +				     unsigned int len)
>  {
>  	struct sk_buff *skb;
>  	struct bpf_prog *xdp_prog;
> @@ -773,7 +774,7 @@ static int receive_buf(struct virtnet_info *vi, struct
receive_queue *rq,
>  	else if (vi->big_packets)
>  		skb = receive_big(dev, vi, rq, buf, len);
>  	else
> -		skb = receive_small(dev, vi, rq, buf, len);
> +		skb = receive_small(dev, vi, rq, buf, ctx, len);
>  
>  	if (unlikely(!skb))
>  		return 0;
> @@ -812,6 +813,7 @@ static int add_recvbuf_small(struct virtnet_info *vi,
struct receive_queue *rq,
Let's document that ctx API is used a bit differently here:

/* Unlike mergeable buffers, all buffers are allocated to the same size,
 * except for the headroom. For this reason we do not need to use
 * mergeable_len_to_ctx here - it is enough to store the headroom as the
 * context ignoring the truesize.
 */

as an alternative, reuse the same format as mergeable buffers.
>  	struct page_frag *alloc_frag = &rq->alloc_frag;
>  	char *buf;
>  	unsigned int xdp_headroom = virtnet_get_headroom(vi);
> +	void *ctx = (void *)(unsigned long)xdp_headroom;
>  	int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN +
xdp_headroom;
>  	int err;
>  
> @@ -825,7 +827,7 @@ static int add_recvbuf_small(struct virtnet_info *vi,
struct receive_queue *rq,
>  	alloc_frag->offset += len;
>  	sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
>  		    vi->hdr_len + GOOD_PACKET_LEN);
> -	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
> +	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
>  	if (err < 0)
>  		put_page(virt_to_head_page(buf));
>  
> @@ -1034,7 +1036,7 @@ static int virtnet_receive(struct receive_queue *rq,
int budget)
>  	void *buf;
>  	struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>  
> -	if (vi->mergeable_rx_bufs) {
> +	if (!vi->big_packets || vi->mergeable_rx_bufs) {
>  		void *ctx;
>  
>  		while (received < budget &&
> @@ -2198,7 +2200,7 @@ static int virtnet_find_vqs(struct virtnet_info *vi)
>  	names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
>  	if (!names)
>  		goto err_names;
> -	if (vi->mergeable_rx_bufs) {
> +	if (!vi->big_packets || vi->mergeable_rx_bufs) {
>  		ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
>  		if (!ctx)
>  			goto err_ctx;
> -- 
> 2.7.4

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 at redhat.com>
> ---
>  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

On Mon, Jul 17, 2017 at 08:44:01PM +0800, Jason Wang
wrote:
> Current XDP implementation want guest offloads feature to be disabled
s/want/wants/
> on qemu cli.
on the device.
> This is inconvenient and means guest can't benefit from
> offloads if XDP is not used. This patch tries to address this
> limitation by disable
disabling
> the offloads on demand through control guest
> offloads. Guest offloads will be disabled and enabled on demand on XDP
> set.
> 
> Signed-off-by: Jason Wang <jasowang at redhat.com>
In fact, since we no longer reset when XDP is set,
here device might have offloads enabled, buffers are
used but not consumed, then XDP is set.

This can result in
- packet scattered across multiple buffers
  (handled correctly but need to update the comment)
- packet may have VIRTIO_NET_HDR_F_NEEDS_CSUM, in that case
  the spec says "The checksum on the packet is incomplete".
  (probably needs to be handled by calculating the checksum).


Ideas for follow-up patches:

- skip looking at packet data completely
  won't work if you play with checksums dynamically
  but can be done if disabled on device
- allow ethtools to tweak offloads from userspace as well

> ---
>  drivers/net/virtio_net.c | 70
++++++++++++++++++++++++++++++++++++++++++++----
>  1 file changed, 65 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index e732bd6..d970c2d 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -57,6 +57,11 @@ DECLARE_EWMA(pkt_len, 0, 64)
>  
>  #define VIRTNET_DRIVER_VERSION "1.0.0"
>  
> +const unsigned long guest_offloads[] = { VIRTIO_NET_F_GUEST_TSO4,
> +					 VIRTIO_NET_F_GUEST_TSO6,
> +					 VIRTIO_NET_F_GUEST_ECN,
> +					 VIRTIO_NET_F_GUEST_UFO };
> +
>  struct virtnet_stats {
>  	struct u64_stats_sync tx_syncp;
>  	struct u64_stats_sync rx_syncp;
> @@ -164,10 +169,13 @@ struct virtnet_info {
>  	u8 ctrl_promisc;
>  	u8 ctrl_allmulti;
>  	u16 ctrl_vid;
> +	u64 ctrl_offloads;
>  
>  	/* Ethtool settings */
>  	u8 duplex;
>  	u32 speed;
> +
> +	unsigned long guest_offloads;
>  };
>  
>  struct padded_vnet_hdr {
> @@ -1889,6 +1897,47 @@ static int virtnet_restore_up(struct virtio_device
*vdev)
>  	return err;
>  }
>  
> +static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64
offloads)
> +{
> +	struct scatterlist sg;
> +	vi->ctrl_offloads = cpu_to_virtio64(vi->vdev, offloads);
> +
> +	sg_init_one(&sg, &vi->ctrl_offloads,
sizeof(vi->ctrl_offloads));
> +
> +	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
> +				  VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
> +		dev_warn(&vi->dev->dev, "Fail to set guest offload.
\n");
> +		return -EINVAL;
> +	}
> +
> +	return 0;
> +}
> +
> +static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
> +{
> +	u64 offloads = 0;
> +
> +	if (!vi->guest_offloads)
> +		return 0;
> +
> +	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
> +		offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
> +
> +	return virtnet_set_guest_offloads(vi, offloads);
> +}
> +
> +static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
> +{
> +	u64 offloads = vi->guest_offloads;
> +
> +	if (!vi->guest_offloads)
> +		return 0;
> +	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
> +		offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
> +
> +	return virtnet_set_guest_offloads(vi, offloads);
> +}
> +
>  static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
>  			   struct netlink_ext_ack *extack)
>  {
> @@ -1898,10 +1947,11 @@ static int virtnet_xdp_set(struct net_device *dev,
struct bpf_prog *prog,
>  	u16 xdp_qp = 0, curr_qp;
>  	int i, err;
>  
> -	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
> -	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
> -	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
> -	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO)) {
> +	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
> +	    && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4)
||
> +	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
> +	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
> +		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
>  		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is
implementing LRO, disable LRO first");
>  		return -EOPNOTSUPP;
>  	}
> @@ -1950,6 +2000,12 @@ static int virtnet_xdp_set(struct net_device *dev,
struct bpf_prog *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 (i == 0) {
> +			if (!old_prog)
> +				virtnet_clear_guest_offloads(vi);
> +			if (!prog)
> +				virtnet_restore_guest_offloads(vi);
> +		}
>  		if (old_prog)
>  			bpf_prog_put(old_prog);
>  		napi_enable(&vi->rq[i].napi);
> @@ -2583,6 +2639,10 @@ static int virtnet_probe(struct virtio_device *vdev)
>  		netif_carrier_on(dev);
>  	}
>  
> +	for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
> +		if (virtio_has_feature(vi->vdev, guest_offloads[i]))
> +			set_bit(guest_offloads[i], &vi->guest_offloads);
> +
>  	pr_debug("virtnet: registered device %s with %d RX and TX
vq's\n",
>  		 dev->name, max_queue_pairs);
>  
> @@ -2679,7 +2739,7 @@ static struct virtio_device_id id_table[] = {
>  	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
>  	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
>  	VIRTIO_NET_F_CTRL_MAC_ADDR, \
> -	VIRTIO_NET_F_MTU
> +	VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
>  
>  static unsigned int features[] = {
>  	VIRTNET_FEATURES,
> -- 
> 2.7.4

On Mon, Jul 17, 2017 at 08:43:56PM +0800, Jason Wang
wrote:
> Hi:
> 
> This series brings two optimizations for virtio-net XDP:
> 
> - avoid reset during XDP set
> - turn off offloads on demand
I'm glad to see this take shape - this can be
extended to optimize virtnet_get_headroom so we don't
waste room if adjust_head is enabled.

I see a couple of issues, responded to individual patches.

> Please review.
> 
> Thanks
> 
> Jason Wang (5):
>   virtio_ring: allow to store zero as the ctx
>   virtio-net: pack headroom into ctx for mergeable buffer
>   virtio-net: switch to use new ctx API for small buffer
>   virtio-net: do not reset during XDP set
>   virtio-net: switch off offloads on demand if possible on XDP set
> 
>  drivers/net/virtio_net.c     | 325
+++++++++++++++++++++++++------------------
>  drivers/virtio/virtio_ring.c |   2 +-
>  2 files changed, 194 insertions(+), 133 deletions(-)
> 
> -- 
> 2.7.4

On 2017?07?19? 04:13, Michael S. Tsirkin wrote:
> On Mon, Jul 17, 2017 at 08:43:56PM +0800, Jason Wang wrote:
>> Hi:
>>
>> This series brings two optimizations for virtio-net XDP:
>>
>> - avoid reset during XDP set
>> - turn off offloads on demand
> I'm glad to see this take shape - this can be
> extended to optimize virtnet_get_headroom so we don't
> waste room if adjust_head is enabled.
Right, we can do it on top.
> I see a couple of issues, responded to individual patches.
>
>
Thanks for the reviewing.