[PATCH 0/2] virtio net improvements

[PATCH 0/2] virtio net improvements

[Rusty Russell]

Hi Dave,

   Nice driver optimization from Shirley, but requires a new virtio hook.
Do you want to take both?  I have nothing else overlapping it.

Cheers,
Rusty.

[PATCH 1/2] virtio: Add ability to detach unused buffers from vrings

[Rusty Russell]

From: Shirley Ma <mashirle@us.ibm.com>

There's currently no way for a virtio driver to ask for unused
buffers, so it has to keep a list itself to reclaim them at shutdown.
This is redundant, since virtio_ring stores that information.  So
add a new hook to do this.

Signed-off-by: Shirley Ma <xma@us.ibm.com>
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
---
 drivers/virtio/virtio_ring.c |   25 +++++++++++++++++++++++++
 include/linux/virtio.h       |    4 ++++
 2 files changed, 29 insertions(+)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index fbd2ecd..71929ee 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -334,6 +334,30 @@ static bool vring_enable_cb(struct virtqueue *_vq)
 	return true;
 }
 
+static void *vring_detach_unused_buf(struct virtqueue *_vq)
+{
+	struct vring_virtqueue *vq = to_vvq(_vq);
+	unsigned int i;
+	void *buf;
+
+	START_USE(vq);
+
+	for (i = 0; i < vq->vring.num; i++) {
+		if (!vq->data[i])
+			continue;
+		/* detach_buf clears data, so grab it now. */
+		buf = vq->data[i];
+		detach_buf(vq, i);
+		END_USE(vq);
+		return buf;
+	}
+	/* That should have freed everything. */
+	BUG_ON(vq->num_free != vq->vring.num);
+
+	END_USE(vq);
+	return NULL;
+}
+
 irqreturn_t vring_interrupt(int irq, void *_vq)
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
@@ -360,6 +384,7 @@ static struct virtqueue_ops vring_vq_ops = {
 	.kick = vring_kick,
 	.disable_cb = vring_disable_cb,
 	.enable_cb = vring_enable_cb,
+	.detach_unused_buf = vring_detach_unused_buf,
 };
 
 struct virtqueue *vring_new_virtqueue(unsigned int num,
diff --git a/include/linux/virtio.h b/include/linux/virtio.h
index 057a2e0..f508c65 100644
--- a/include/linux/virtio.h
+++ b/include/linux/virtio.h
@@ -51,6 +51,9 @@ struct virtqueue {
  *	This re-enables callbacks; it returns "false" if there are pending
  *	buffers in the queue, to detect a possible race between the driver
  *	checking for more work, and enabling callbacks.
+ * @detach_unused_buf: detach first unused buffer
+ * 	vq: the struct virtqueue we're talking about.
+ * 	Returns NULL or the "data" token handed to add_buf
  *
  * Locking rules are straightforward: the driver is responsible for
  * locking.  No two operations may be invoked simultaneously, with the exception
@@ -71,6 +74,7 @@ struct virtqueue_ops {
 
 	void (*disable_cb)(struct virtqueue *vq);
 	bool (*enable_cb)(struct virtqueue *vq);
+	void *(*detach_unused_buf)(struct virtqueue *vq);
 };
 
 /**

[PATCH 1/2] virtio: Add ability to detach unused buffers from vrings

[Rusty Russell]

From: Shirley Ma <mashirle@us.ibm.com>

There's currently no way for a virtio driver to ask for unused
buffers, so it has to keep a list itself to reclaim them at shutdown.
This is redundant, since virtio_ring stores that information.  So
add a new hook to do this.

Signed-off-by: Shirley Ma <xma@us.ibm.com>
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
---
 drivers/virtio/virtio_ring.c |   25 +++++++++++++++++++++++++
 include/linux/virtio.h       |    4 ++++
 2 files changed, 29 insertions(+)

diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index fbd2ecd..71929ee 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -334,6 +334,30 @@ static bool vring_enable_cb(struct virtqueue *_vq)
 	return true;
 }
 
+static void *vring_detach_unused_buf(struct virtqueue *_vq)
+{
+	struct vring_virtqueue *vq = to_vvq(_vq);
+	unsigned int i;
+	void *buf;
+
+	START_USE(vq);
+
+	for (i = 0; i < vq->vring.num; i++) {
+		if (!vq->data[i])
+			continue;
+		/* detach_buf clears data, so grab it now. */
+		buf = vq->data[i];
+		detach_buf(vq, i);
+		END_USE(vq);
+		return buf;
+	}
+	/* That should have freed everything. */
+	BUG_ON(vq->num_free != vq->vring.num);
+
+	END_USE(vq);
+	return NULL;
+}
+
 irqreturn_t vring_interrupt(int irq, void *_vq)
 {
 	struct vring_virtqueue *vq = to_vvq(_vq);
@@ -360,6 +384,7 @@ static struct virtqueue_ops vring_vq_ops = {
 	.kick = vring_kick,
 	.disable_cb = vring_disable_cb,
 	.enable_cb = vring_enable_cb,
+	.detach_unused_buf = vring_detach_unused_buf,
 };
 
 struct virtqueue *vring_new_virtqueue(unsigned int num,
diff --git a/include/linux/virtio.h b/include/linux/virtio.h
index 057a2e0..f508c65 100644
--- a/include/linux/virtio.h
+++ b/include/linux/virtio.h
@@ -51,6 +51,9 @@ struct virtqueue {
  *	This re-enables callbacks; it returns "false" if there are pending
  *	buffers in the queue, to detect a possible race between the driver
  *	checking for more work, and enabling callbacks.
+ * @detach_unused_buf: detach first unused buffer
+ * 	vq: the struct virtqueue we're talking about.
+ * 	Returns NULL or the "data" token handed to add_buf
  *
  * Locking rules are straightforward: the driver is responsible for
  * locking.  No two operations may be invoked simultaneously, with the exception
@@ -71,6 +74,7 @@ struct virtqueue_ops {
 
 	void (*disable_cb)(struct virtqueue *vq);
 	bool (*enable_cb)(struct virtqueue *vq);
+	void *(*detach_unused_buf)(struct virtqueue *vq);
 };
 
 /**

[PATCH 2/2] virtio_net: Defer skb allocation in receive path Date: Wed, 13 Jan 2010 12:53:38 -0800

[Rusty Russell]

From: Shirley Ma <mashirle@us.ibm.com>

virtio_net receives packets from its pre-allocated vring buffers, then it 
delivers these packets to upper layer protocols as skb buffs. So it's not
necessary to pre-allocate skb for each mergable buffer, then frees extra 
skbs when buffers are merged into a large packet. This patch has deferred 
skb allocation in receiving packets for both big packets and mergeable buffers
to reduce skb pre-allocations and skb frees. It frees unused buffers by calling 
detach_unused_buf in vring, so recv skb queue is not needed.

Signed-off-by: Shirley Ma <xma@us.ibm.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
---
 drivers/net/virtio_net.c |  427 +++++++++++++++++++++++++++--------------------
 1 file changed, 248 insertions(+), 179 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index c708ecc..72b3f21 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -56,8 +56,7 @@ struct virtnet_info
 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
 	bool mergeable_rx_bufs;
 
-	/* Receive & send queues. */
-	struct sk_buff_head recv;
+	/* Send queue. */
 	struct sk_buff_head send;
 
 	/* Work struct for refilling if we run low on memory. */
@@ -75,34 +74,44 @@ struct skb_vnet_hdr {
 	unsigned int num_sg;
 };
 
+struct padded_vnet_hdr {
+	struct virtio_net_hdr hdr;
+	/*
+	 * virtio_net_hdr should be in a separated sg buffer because of a
+	 * QEMU bug, and data sg buffer shares same page with this header sg.
+	 * This padding makes next sg 16 byte aligned after virtio_net_hdr.
+	 */
+	char padding[6];
+};
+
 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 {
 	return (struct skb_vnet_hdr *)skb->cb;
 }
 
-static void give_a_page(struct virtnet_info *vi, struct page *page)
-{
-	page->private = (unsigned long)vi->pages;
-	vi->pages = page;
-}
-
-static void trim_pages(struct virtnet_info *vi, struct sk_buff *skb)
+/*
+ * private is used to chain pages for big packets, put the whole
+ * most recent used list in the beginning for reuse
+ */
+static void give_pages(struct virtnet_info *vi, struct page *page)
 {
-	unsigned int i;
+	struct page *end;
 
-	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
-		give_a_page(vi, skb_shinfo(skb)->frags[i].page);
-	skb_shinfo(skb)->nr_frags = 0;
-	skb->data_len = 0;
+	/* Find end of list, sew whole thing into vi->pages. */
+	for (end = page; end->private; end = (struct page *)end->private);
+	end->private = (unsigned long)vi->pages;
+	vi->pages = page;
 }
 
 static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
 {
 	struct page *p = vi->pages;
 
-	if (p)
+	if (p) {
 		vi->pages = (struct page *)p->private;
-	else
+		/* clear private here, it is used to chain pages */
+		p->private = 0;
+	} else
 		p = alloc_page(gfp_mask);
 	return p;
 }
@@ -118,99 +127,142 @@ static void skb_xmit_done(struct virtqueue *svq)
 	netif_wake_queue(vi->dev);
 }
 
-static void receive_skb(struct net_device *dev, struct sk_buff *skb,
-			unsigned len)
+static void set_skb_frag(struct sk_buff *skb, struct page *page,
+			 unsigned int offset, unsigned int *len)
 {
-	struct virtnet_info *vi = netdev_priv(dev);
-	struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
-	int err;
-	int i;
-
-	if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
-		pr_debug("%s: short packet %i\n", dev->name, len);
-		dev->stats.rx_length_errors++;
-		goto drop;
-	}
+	int i = skb_shinfo(skb)->nr_frags;
+	skb_frag_t *f;
+
+	f = &skb_shinfo(skb)->frags[i];
+	f->size = min((unsigned)PAGE_SIZE - offset, *len);
+	f->page_offset = offset;
+	f->page = page;
+
+	skb->data_len += f->size;
+	skb->len += f->size;
+	skb_shinfo(skb)->nr_frags++;
+	*len -= f->size;
+}
 
-	if (vi->mergeable_rx_bufs) {
-		unsigned int copy;
-		char *p = page_address(skb_shinfo(skb)->frags[0].page);
+static struct sk_buff *page_to_skb(struct virtnet_info *vi,
+				   struct page *page, unsigned int len)
+{
+	struct sk_buff *skb;
+	struct skb_vnet_hdr *hdr;
+	unsigned int copy, hdr_len, offset;
+	char *p;
 
-		if (len > PAGE_SIZE)
-			len = PAGE_SIZE;
-		len -= sizeof(struct virtio_net_hdr_mrg_rxbuf);
+	p = page_address(page);
 
-		memcpy(&hdr->mhdr, p, sizeof(hdr->mhdr));
-		p += sizeof(hdr->mhdr);
+	/* copy small packet so we can reuse these pages for small data */
+	skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
+	if (unlikely(!skb))
+		return NULL;
 
-		copy = len;
-		if (copy > skb_tailroom(skb))
-			copy = skb_tailroom(skb);
+	hdr = skb_vnet_hdr(skb);
 
-		memcpy(skb_put(skb, copy), p, copy);
+	if (vi->mergeable_rx_bufs) {
+		hdr_len = sizeof hdr->mhdr;
+		offset = hdr_len;
+	} else {
+		hdr_len = sizeof hdr->hdr;
+		offset = sizeof(struct padded_vnet_hdr);
+	}
 
-		len -= copy;
+	memcpy(hdr, p, hdr_len);
 
-		if (!len) {
-			give_a_page(vi, skb_shinfo(skb)->frags[0].page);
-			skb_shinfo(skb)->nr_frags--;
-		} else {
-			skb_shinfo(skb)->frags[0].page_offset +=
-				sizeof(hdr->mhdr) + copy;
-			skb_shinfo(skb)->frags[0].size = len;
-			skb->data_len += len;
-			skb->len += len;
-		}
+	len -= hdr_len;
+	p += offset;
 
-		while (--hdr->mhdr.num_buffers) {
-			struct sk_buff *nskb;
+	copy = len;
+	if (copy > skb_tailroom(skb))
+		copy = skb_tailroom(skb);
+	memcpy(skb_put(skb, copy), p, copy);
 
-			i = skb_shinfo(skb)->nr_frags;
-			if (i >= MAX_SKB_FRAGS) {
-				pr_debug("%s: packet too long %d\n", dev->name,
-					 len);
-				dev->stats.rx_length_errors++;
-				goto drop;
-			}
+	len -= copy;
+	offset += copy;
 
-			nskb = vi->rvq->vq_ops->get_buf(vi->rvq, &len);
-			if (!nskb) {
-				pr_debug("%s: rx error: %d buffers missing\n",
-					 dev->name, hdr->mhdr.num_buffers);
-				dev->stats.rx_length_errors++;
-				goto drop;
-			}
+	while (len) {
+		set_skb_frag(skb, page, offset, &len);
+		page = (struct page *)page->private;
+		offset = 0;
+	}
 
-			__skb_unlink(nskb, &vi->recv);
-			vi->num--;
+	if (page)
+		give_pages(vi, page);
 
-			skb_shinfo(skb)->frags[i] = skb_shinfo(nskb)->frags[0];
-			skb_shinfo(nskb)->nr_frags = 0;
-			kfree_skb(nskb);
+	return skb;
+}
 
-			if (len > PAGE_SIZE)
-				len = PAGE_SIZE;
+static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
+{
+	struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
+	struct page *page;
+	int num_buf, i, len;
+
+	num_buf = hdr->mhdr.num_buffers;
+	while (--num_buf) {
+		i = skb_shinfo(skb)->nr_frags;
+		if (i >= MAX_SKB_FRAGS) {
+			pr_debug("%s: packet too long\n", skb->dev->name);
+			skb->dev->stats.rx_length_errors++;
+			return -EINVAL;
+		}
 
-			skb_shinfo(skb)->frags[i].size = len;
-			skb_shinfo(skb)->nr_frags++;
-			skb->data_len += len;
-			skb->len += len;
+		page = vi->rvq->vq_ops->get_buf(vi->rvq, &len);
+		if (!page) {
+			pr_debug("%s: rx error: %d buffers missing\n",
+				 skb->dev->name, hdr->mhdr.num_buffers);
+			skb->dev->stats.rx_length_errors++;
+			return -EINVAL;
 		}
-	} else {
-		len -= sizeof(hdr->hdr);
+		if (len > PAGE_SIZE)
+			len = PAGE_SIZE;
+
+		set_skb_frag(skb, page, 0, &len);
+
+		--vi->num;
+	}
+	return 0;
+}
+
+static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
+{
+	struct virtnet_info *vi = netdev_priv(dev);
+	struct sk_buff *skb;
+	struct page *page;
+	struct skb_vnet_hdr *hdr;
 
-		if (len <= MAX_PACKET_LEN)
-			trim_pages(vi, skb);
+	if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
+		pr_debug("%s: short packet %i\n", dev->name, len);
+		dev->stats.rx_length_errors++;
+		if (vi->mergeable_rx_bufs || vi->big_packets)
+			give_pages(vi, buf);
+		else
+			dev_kfree_skb(buf);
+		return;
+	}
 
-		err = pskb_trim(skb, len);
-		if (err) {
-			pr_debug("%s: pskb_trim failed %i %d\n", dev->name,
-				 len, err);
+	if (!vi->mergeable_rx_bufs && !vi->big_packets) {
+		skb = buf;
+		len -= sizeof(struct virtio_net_hdr);
+		skb_trim(skb, len);
+	} else {
+		page = buf;
+		skb = page_to_skb(vi, page, len);
+		if (unlikely(!skb)) {
 			dev->stats.rx_dropped++;
-			goto drop;
+			give_pages(vi, page);
+			return;
 		}
+		if (vi->mergeable_rx_bufs)
+			if (receive_mergeable(vi, skb)) {
+				dev_kfree_skb(skb);
+				return;
+			}
 	}
 
+	hdr = skb_vnet_hdr(skb);
 	skb->truesize += skb->data_len;
 	dev->stats.rx_bytes += skb->len;
 	dev->stats.rx_packets++;
@@ -267,110 +319,119 @@ static void receive_skb(struct net_device *dev, struct sk_buff *skb,
 
 frame_err:
 	dev->stats.rx_frame_errors++;
-drop:
 	dev_kfree_skb(skb);
 }
 
-static bool try_fill_recv_maxbufs(struct virtnet_info *vi, gfp_t gfp)
+static int add_recvbuf_small(struct virtnet_info *vi, gfp_t gfp)
 {
 	struct sk_buff *skb;
-	struct scatterlist sg[2+MAX_SKB_FRAGS];
-	int num, err, i;
-	bool oom = false;
-
-	sg_init_table(sg, 2+MAX_SKB_FRAGS);
-	do {
-		struct skb_vnet_hdr *hdr;
+	struct skb_vnet_hdr *hdr;
+	struct scatterlist sg[2];
+	int err;
 
-		skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
-		if (unlikely(!skb)) {
-			oom = true;
-			break;
-		}
+	skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
+	if (unlikely(!skb))
+		return -ENOMEM;
 
-		skb_put(skb, MAX_PACKET_LEN);
+	skb_put(skb, MAX_PACKET_LEN);
 
-		hdr = skb_vnet_hdr(skb);
-		sg_set_buf(sg, &hdr->hdr, sizeof(hdr->hdr));
+	hdr = skb_vnet_hdr(skb);
+	sg_set_buf(sg, &hdr->hdr, sizeof hdr->hdr);
 
-		if (vi->big_packets) {
-			for (i = 0; i < MAX_SKB_FRAGS; i++) {
-				skb_frag_t *f = &skb_shinfo(skb)->frags[i];
-				f->page = get_a_page(vi, gfp);
-				if (!f->page)
-					break;
+	skb_to_sgvec(skb, sg + 1, 0, skb->len);
 
-				f->page_offset = 0;
-				f->size = PAGE_SIZE;
+	err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, 2, skb);
+	if (err < 0)
+		dev_kfree_skb(skb);
 
-				skb->data_len += PAGE_SIZE;
-				skb->len += PAGE_SIZE;
+	return err;
+}
 
-				skb_shinfo(skb)->nr_frags++;
-			}
+static int add_recvbuf_big(struct virtnet_info *vi, gfp_t gfp)
+{
+	struct scatterlist sg[MAX_SKB_FRAGS + 2];
+	struct page *first, *list = NULL;
+	char *p;
+	int i, err, offset;
+
+	/* page in sg[MAX_SKB_FRAGS + 1] is list tail */
+	for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
+		first = get_a_page(vi, gfp);
+		if (!first) {
+			if (list)
+				give_pages(vi, list);
+			return -ENOMEM;
 		}
+		sg_set_buf(&sg[i], page_address(first), PAGE_SIZE);
 
-		num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
-		skb_queue_head(&vi->recv, skb);
+		/* chain new page in list head to match sg */
+		first->private = (unsigned long)list;
+		list = first;
+	}
 
-		err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, num, skb);
-		if (err < 0) {
-			skb_unlink(skb, &vi->recv);
-			trim_pages(vi, skb);
-			kfree_skb(skb);
-			break;
-		}
-		vi->num++;
-	} while (err >= num);
-	if (unlikely(vi->num > vi->max))
-		vi->max = vi->num;
-	vi->rvq->vq_ops->kick(vi->rvq);
-	return !oom;
+	first = get_a_page(vi, gfp);
+	if (!first) {
+		give_pages(vi, list);
+		return -ENOMEM;
+	}
+	p = page_address(first);
+
+	/* sg[0], sg[1] share the same page */
+	/* a separated sg[0] for  virtio_net_hdr only during to QEMU bug*/
+	sg_set_buf(&sg[0], p, sizeof(struct virtio_net_hdr));
+
+	/* sg[1] for data packet, from offset */
+	offset = sizeof(struct padded_vnet_hdr);
+	sg_set_buf(&sg[1], p + offset, PAGE_SIZE - offset);
+
+	/* chain first in list head */
+	first->private = (unsigned long)list;
+	err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, MAX_SKB_FRAGS + 2,
+				       first);
+	if (err < 0)
+		give_pages(vi, first);
+
+	return err;
 }
 
-/* Returns false if we couldn't fill entirely (OOM). */
-static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
+static int add_recvbuf_mergeable(struct virtnet_info *vi, gfp_t gfp)
 {
-	struct sk_buff *skb;
-	struct scatterlist sg[1];
+	struct page *page;
+	struct scatterlist sg;
 	int err;
-	bool oom = false;
-
-	if (!vi->mergeable_rx_bufs)
-		return try_fill_recv_maxbufs(vi, gfp);
 
-	do {
-		skb_frag_t *f;
+	page = get_a_page(vi, gfp);
+	if (!page)
+		return -ENOMEM;
 
-		skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
-		if (unlikely(!skb)) {
-			oom = true;
-			break;
-		}
+	sg_init_one(&sg, page_address(page), PAGE_SIZE);
 
-		f = &skb_shinfo(skb)->frags[0];
-		f->page = get_a_page(vi, gfp);
-		if (!f->page) {
-			oom = true;
-			kfree_skb(skb);
-			break;
-		}
+	err = vi->rvq->vq_ops->add_buf(vi->rvq, &sg, 0, 1, page);
+	if (err < 0)
+		give_pages(vi, page);
 
-		f->page_offset = 0;
-		f->size = PAGE_SIZE;
+	return err;
+}
 
-		skb_shinfo(skb)->nr_frags++;
+/* Returns false if we couldn't fill entirely (OOM). */
+static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
+{
+	int err;
+	bool oom = false;
 
-		sg_init_one(sg, page_address(f->page), PAGE_SIZE);
-		skb_queue_head(&vi->recv, skb);
+	do {
+		if (vi->mergeable_rx_bufs)
+			err = add_recvbuf_mergeable(vi, gfp);
+		else if (vi->big_packets)
+			err = add_recvbuf_big(vi, gfp);
+		else
+			err = add_recvbuf_small(vi, gfp);
 
-		err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, 1, skb);
 		if (err < 0) {
-			skb_unlink(skb, &vi->recv);
-			kfree_skb(skb);
+			oom = true;
 			break;
 		}
-		vi->num++;
+		++vi->num;
 	} while (err > 0);
 	if (unlikely(vi->num > vi->max))
 		vi->max = vi->num;
@@ -408,15 +469,14 @@ static void refill_work(struct work_struct *work)
 static int virtnet_poll(struct napi_struct *napi, int budget)
 {
 	struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
-	struct sk_buff *skb = NULL;
+	void *buf;
 	unsigned int len, received = 0;
 
 again:
 	while (received < budget &&
-	       (skb = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
-		__skb_unlink(skb, &vi->recv);
-		receive_skb(vi->dev, skb, len);
-		vi->num--;
+	       (buf = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
+		receive_buf(vi->dev, buf, len);
+		--vi->num;
 		received++;
 	}
 
@@ -496,9 +556,9 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 
 	/* Encode metadata header at front. */
 	if (vi->mergeable_rx_bufs)
-		sg_set_buf(sg, &hdr->mhdr, sizeof(hdr->mhdr));
+		sg_set_buf(sg, &hdr->mhdr, sizeof hdr->mhdr);
 	else
-		sg_set_buf(sg, &hdr->hdr, sizeof(hdr->hdr));
+		sg_set_buf(sg, &hdr->hdr, sizeof hdr->hdr);
 
 	hdr->num_sg = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 	return vi->svq->vq_ops->add_buf(vi->svq, sg, hdr->num_sg, 0, skb);
@@ -916,8 +976,7 @@ static int virtnet_probe(struct virtio_device *vdev)
 			dev->features |= NETIF_F_HW_VLAN_FILTER;
 	}
 
-	/* Initialize our empty receive and send queues. */
-	skb_queue_head_init(&vi->recv);
+	/* Initialize our empty send queue. */
 	skb_queue_head_init(&vi->send);
 
 	err = register_netdev(dev);
@@ -952,25 +1011,35 @@ free:
 	return err;
 }
 
+static void free_unused_bufs(struct virtnet_info *vi)
+{
+	void *buf;
+	while (1) {
+		buf = vi->rvq->vq_ops->detach_unused_buf(vi->rvq);
+		if (!buf)
+			break;
+		if (vi->mergeable_rx_bufs || vi->big_packets)
+			give_pages(vi, buf);
+		else
+			dev_kfree_skb(buf);
+		--vi->num;
+	}
+	BUG_ON(vi->num != 0);
+}
+
 static void __devexit virtnet_remove(struct virtio_device *vdev)
 {
 	struct virtnet_info *vi = vdev->priv;
-	struct sk_buff *skb;
 
 	/* Stop all the virtqueues. */
 	vdev->config->reset(vdev);
 
-	/* Free our skbs in send and recv queues, if any. */
-	while ((skb = __skb_dequeue(&vi->recv)) != NULL) {
-		kfree_skb(skb);
-		vi->num--;
-	}
+	/* Free our skbs in send queue, if any. */
 	__skb_queue_purge(&vi->send);
 
-	BUG_ON(vi->num != 0);
-
 	unregister_netdev(vi->dev);
 	cancel_delayed_work_sync(&vi->refill);
+	free_unused_bufs(vi);
 
 	vdev->config->del_vqs(vi->vdev);
 

[PATCH 1/1 net-next] virtio_net: remove send queue

[Shirley Ma]

Use detach buffers API in virtio to free unused buffers in send queue when shutting down 
virtio_net to avoid maintaining skb link list for each transmit packet.

Signed-off-by: Shirley Ma <xma@us.ibm.com>
---
 drivers/net/virtio_net.c |   26 ++++++++------------------
 1 files changed, 8 insertions(+), 18 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 9d8984a..8069c08 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -56,9 +56,6 @@ struct virtnet_info
 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
 	bool mergeable_rx_bufs;
 
-	/* Send queue. */
-	struct sk_buff_head send;
-
 	/* Work struct for refilling if we run low on memory. */
 	struct delayed_work refill;
 
@@ -505,7 +502,6 @@ static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
 
 	while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
 		pr_debug("Sent skb %p\n", skb);
-		__skb_unlink(skb, &vi->send);
 		vi->dev->stats.tx_bytes += skb->len;
 		vi->dev->stats.tx_packets++;
 		tot_sgs += skb_vnet_hdr(skb)->num_sg;
@@ -588,15 +584,6 @@ again:
 	}
 	vi->svq->vq_ops->kick(vi->svq);
 
-	/*
-	 * Put new one in send queue.  You'd expect we'd need this before
-	 * xmit_skb calls add_buf(), since the callback can be triggered
-	 * immediately after that.  But since the callback just triggers
-	 * another call back here, normal network xmit locking prevents the
-	 * race.
-	 */
-	__skb_queue_head(&vi->send, skb);
-
 	/* Don't wait up for transmitted skbs to be freed. */
 	skb_orphan(skb);
 	nf_reset(skb);
@@ -977,9 +964,6 @@ static int virtnet_probe(struct virtio_device *vdev)
 			dev->features |= NETIF_F_HW_VLAN_FILTER;
 	}
 
-	/* Initialize our empty send queue. */
-	skb_queue_head_init(&vi->send);
-
 	err = register_netdev(dev);
 	if (err) {
 		pr_debug("virtio_net: registering device failed\n");
@@ -1016,6 +1000,12 @@ static void free_unused_bufs(struct virtnet_info *vi)
 {
 	void *buf;
 	while (1) {
+		buf = vi->svq->vq_ops->detach_unused_buf(vi->svq);
+		if (!buf)
+			break;
+		dev_kfree_skb(buf);
+	}
+	while (1) {
 		buf = vi->rvq->vq_ops->detach_unused_buf(vi->rvq);
 		if (!buf)
 			break;
@@ -1035,11 +1025,11 @@ static void __devexit virtnet_remove(struct virtio_device *vdev)
 	/* Stop all the virtqueues. */
 	vdev->config->reset(vdev);
 
-	/* Free our skbs in send queue, if any. */
-	__skb_queue_purge(&vi->send);
 
 	unregister_netdev(vi->dev);
 	cancel_delayed_work_sync(&vi->refill);
+
+	/* Free unused buffers in both send and recv, if any. */
 	free_unused_bufs(vi);
 
 	vdev->config->del_vqs(vi->vdev);

[PATCH 0/1 net-next] virtio_net: remove send queue

[Shirley Ma]

This patch is built on top of net-next + defer skb patch for virtio_net.
The patch has been tested and some performance has been gained by
removing skb link & unlink for each packet in send queue.

thanks
Shirley

Re: [PATCH 1/1 net-next] virtio_net: remove send queue

[Michael S. Tsirkin]

On Tue, Feb 02, 2010 at 01:48:51PM -0800, Shirley Ma wrote:
> Use detach buffers API in virtio to free unused buffers in send queue when shutting down 
> virtio_net to avoid maintaining skb link list for each transmit packet.
> 
> Signed-off-by: Shirley Ma <xma@us.ibm.com>

Acked-by: Michael S. Tsirkin <mst@redhat.com>

> ---
>  drivers/net/virtio_net.c |   26 ++++++++------------------
>  1 files changed, 8 insertions(+), 18 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 9d8984a..8069c08 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -56,9 +56,6 @@ struct virtnet_info
>  	/* Host will merge rx buffers for big packets (shake it! shake it!) */
>  	bool mergeable_rx_bufs;
>  
> -	/* Send queue. */
> -	struct sk_buff_head send;
> -
>  	/* Work struct for refilling if we run low on memory. */
>  	struct delayed_work refill;
>  
> @@ -505,7 +502,6 @@ static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
>  
>  	while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
>  		pr_debug("Sent skb %p\n", skb);
> -		__skb_unlink(skb, &vi->send);
>  		vi->dev->stats.tx_bytes += skb->len;
>  		vi->dev->stats.tx_packets++;
>  		tot_sgs += skb_vnet_hdr(skb)->num_sg;
> @@ -588,15 +584,6 @@ again:
>  	}
>  	vi->svq->vq_ops->kick(vi->svq);
>  
> -	/*
> -	 * Put new one in send queue.  You'd expect we'd need this before
> -	 * xmit_skb calls add_buf(), since the callback can be triggered
> -	 * immediately after that.  But since the callback just triggers
> -	 * another call back here, normal network xmit locking prevents the
> -	 * race.
> -	 */
> -	__skb_queue_head(&vi->send, skb);
> -
>  	/* Don't wait up for transmitted skbs to be freed. */
>  	skb_orphan(skb);
>  	nf_reset(skb);
> @@ -977,9 +964,6 @@ static int virtnet_probe(struct virtio_device *vdev)
>  			dev->features |= NETIF_F_HW_VLAN_FILTER;
>  	}
>  
> -	/* Initialize our empty send queue. */
> -	skb_queue_head_init(&vi->send);
> -
>  	err = register_netdev(dev);
>  	if (err) {
>  		pr_debug("virtio_net: registering device failed\n");
> @@ -1016,6 +1000,12 @@ static void free_unused_bufs(struct virtnet_info *vi)
>  {
>  	void *buf;
>  	while (1) {
> +		buf = vi->svq->vq_ops->detach_unused_buf(vi->svq);
> +		if (!buf)
> +			break;
> +		dev_kfree_skb(buf);
> +	}
> +	while (1) {
>  		buf = vi->rvq->vq_ops->detach_unused_buf(vi->rvq);
>  		if (!buf)
>  			break;
> @@ -1035,11 +1025,11 @@ static void __devexit virtnet_remove(struct virtio_device *vdev)
>  	/* Stop all the virtqueues. */
>  	vdev->config->reset(vdev);
>  
> -	/* Free our skbs in send queue, if any. */
> -	__skb_queue_purge(&vi->send);
>  
>  	unregister_netdev(vi->dev);
>  	cancel_delayed_work_sync(&vi->refill);
> +
> +	/* Free unused buffers in both send and recv, if any. */
>  	free_unused_bufs(vi);
>  
>  	vdev->config->del_vqs(vi->vdev);
> 
> 
> --
> To unsubscribe from this list: send the line "unsubscribe kvm" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 1/1 net-next] virtio_net: remove send queue

[Rusty Russell]

On Wed, 3 Feb 2010 08:18:51 am Shirley Ma wrote:
> Use detach buffers API in virtio to free unused buffers in send queue when shutting down 
> virtio_net to avoid maintaining skb link list for each transmit packet.
> 
> Signed-off-by: Shirley Ma <xma@us.ibm.com>

Hi Shirley,

  Nice cleanup.  Please cc: netdev@vger.kernel.org for net patches...

Your comment could be slightly enhanced to indicate that the API is new, eg;

	Now we have a virtio detach API (in commit XXXXXXXXXX), we don't need to
	track xmit skbs in the	driver.

Acked-by: Rusty Russell <rusty@rustcorp.com.au>

Thanks!
Rusty.

Re: [PATCH 2/2] virtio_net: Defer skb allocation in receive path Date: Wed, 13 Jan 2010 12:53:38 -0800

[David Miller]

From: Rusty Russell <rusty@rustcorp.com.au>
Date: Fri, 29 Jan 2010 23:50:04 +1030

> From: Shirley Ma <mashirle@us.ibm.com>
> 
> virtio_net receives packets from its pre-allocated vring buffers, then it 
> delivers these packets to upper layer protocols as skb buffs. So it's not
> necessary to pre-allocate skb for each mergable buffer, then frees extra 
> skbs when buffers are merged into a large packet. This patch has deferred 
> skb allocation in receiving packets for both big packets and mergeable buffers
> to reduce skb pre-allocations and skb frees. It frees unused buffers by calling 
> detach_unused_buf in vring, so recv skb queue is not needed.
> 
> Signed-off-by: Shirley Ma <xma@us.ibm.com>
> Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Applied.

Re: [PATCH 2/2] virtio_net: Defer skb allocation in receive path Date: Wed, 13 Jan 2010 12:53:38 -0800

[David Miller]

From: Rusty Russell <rusty@rustcorp.com.au>
Date: Fri, 29 Jan 2010 23:50:04 +1030

> From: Shirley Ma <mashirle@us.ibm.com>
> 
> virtio_net receives packets from its pre-allocated vring buffers, then it 
> delivers these packets to upper layer protocols as skb buffs. So it's not
> necessary to pre-allocate skb for each mergable buffer, then frees extra 
> skbs when buffers are merged into a large packet. This patch has deferred 
> skb allocation in receiving packets for both big packets and mergeable buffers
> to reduce skb pre-allocations and skb frees. It frees unused buffers by calling 
> detach_unused_buf in vring, so recv skb queue is not needed.
> 
> Signed-off-by: Shirley Ma <xma@us.ibm.com>
> Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Applied.

[PATCH 2/2] virtio_net: Defer skb allocation in receive path Date: Wed, 13 Jan 2010 12:53:38 -0800

[Rusty Russell]

From: Shirley Ma <mashirle@us.ibm.com>

virtio_net receives packets from its pre-allocated vring buffers, then it 
delivers these packets to upper layer protocols as skb buffs. So it's not
necessary to pre-allocate skb for each mergable buffer, then frees extra 
skbs when buffers are merged into a large packet. This patch has deferred 
skb allocation in receiving packets for both big packets and mergeable buffers
to reduce skb pre-allocations and skb frees. It frees unused buffers by calling 
detach_unused_buf in vring, so recv skb queue is not needed.

Signed-off-by: Shirley Ma <xma@us.ibm.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
---
 drivers/net/virtio_net.c |  427 +++++++++++++++++++++++++++--------------------
 1 file changed, 248 insertions(+), 179 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index c708ecc..72b3f21 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -56,8 +56,7 @@ struct virtnet_info
 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
 	bool mergeable_rx_bufs;
 
-	/* Receive & send queues. */
-	struct sk_buff_head recv;
+	/* Send queue. */
 	struct sk_buff_head send;
 
 	/* Work struct for refilling if we run low on memory. */
@@ -75,34 +74,44 @@ struct skb_vnet_hdr {
 	unsigned int num_sg;
 };
 
+struct padded_vnet_hdr {
+	struct virtio_net_hdr hdr;
+	/*
+	 * virtio_net_hdr should be in a separated sg buffer because of a
+	 * QEMU bug, and data sg buffer shares same page with this header sg.
+	 * This padding makes next sg 16 byte aligned after virtio_net_hdr.
+	 */
+	char padding[6];
+};
+
 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 {
 	return (struct skb_vnet_hdr *)skb->cb;
 }
 
-static void give_a_page(struct virtnet_info *vi, struct page *page)
-{
-	page->private = (unsigned long)vi->pages;
-	vi->pages = page;
-}
-
-static void trim_pages(struct virtnet_info *vi, struct sk_buff *skb)
+/*
+ * private is used to chain pages for big packets, put the whole
+ * most recent used list in the beginning for reuse
+ */
+static void give_pages(struct virtnet_info *vi, struct page *page)
 {
-	unsigned int i;
+	struct page *end;
 
-	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
-		give_a_page(vi, skb_shinfo(skb)->frags[i].page);
-	skb_shinfo(skb)->nr_frags = 0;
-	skb->data_len = 0;
+	/* Find end of list, sew whole thing into vi->pages. */
+	for (end = page; end->private; end = (struct page *)end->private);
+	end->private = (unsigned long)vi->pages;
+	vi->pages = page;
 }
 
 static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
 {
 	struct page *p = vi->pages;
 
-	if (p)
+	if (p) {
 		vi->pages = (struct page *)p->private;
-	else
+		/* clear private here, it is used to chain pages */
+		p->private = 0;
+	} else
 		p = alloc_page(gfp_mask);
 	return p;
 }
@@ -118,99 +127,142 @@ static void skb_xmit_done(struct virtqueue *svq)
 	netif_wake_queue(vi->dev);
 }
 
-static void receive_skb(struct net_device *dev, struct sk_buff *skb,
-			unsigned len)
+static void set_skb_frag(struct sk_buff *skb, struct page *page,
+			 unsigned int offset, unsigned int *len)
 {
-	struct virtnet_info *vi = netdev_priv(dev);
-	struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
-	int err;
-	int i;
-
-	if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
-		pr_debug("%s: short packet %i\n", dev->name, len);
-		dev->stats.rx_length_errors++;
-		goto drop;
-	}
+	int i = skb_shinfo(skb)->nr_frags;
+	skb_frag_t *f;
+
+	f = &skb_shinfo(skb)->frags[i];
+	f->size = min((unsigned)PAGE_SIZE - offset, *len);
+	f->page_offset = offset;
+	f->page = page;
+
+	skb->data_len += f->size;
+	skb->len += f->size;
+	skb_shinfo(skb)->nr_frags++;
+	*len -= f->size;
+}
 
-	if (vi->mergeable_rx_bufs) {
-		unsigned int copy;
-		char *p = page_address(skb_shinfo(skb)->frags[0].page);
+static struct sk_buff *page_to_skb(struct virtnet_info *vi,
+				   struct page *page, unsigned int len)
+{
+	struct sk_buff *skb;
+	struct skb_vnet_hdr *hdr;
+	unsigned int copy, hdr_len, offset;
+	char *p;
 
-		if (len > PAGE_SIZE)
-			len = PAGE_SIZE;
-		len -= sizeof(struct virtio_net_hdr_mrg_rxbuf);
+	p = page_address(page);
 
-		memcpy(&hdr->mhdr, p, sizeof(hdr->mhdr));
-		p += sizeof(hdr->mhdr);
+	/* copy small packet so we can reuse these pages for small data */
+	skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
+	if (unlikely(!skb))
+		return NULL;
 
-		copy = len;
-		if (copy > skb_tailroom(skb))
-			copy = skb_tailroom(skb);
+	hdr = skb_vnet_hdr(skb);
 
-		memcpy(skb_put(skb, copy), p, copy);
+	if (vi->mergeable_rx_bufs) {
+		hdr_len = sizeof hdr->mhdr;
+		offset = hdr_len;
+	} else {
+		hdr_len = sizeof hdr->hdr;
+		offset = sizeof(struct padded_vnet_hdr);
+	}
 
-		len -= copy;
+	memcpy(hdr, p, hdr_len);
 
-		if (!len) {
-			give_a_page(vi, skb_shinfo(skb)->frags[0].page);
-			skb_shinfo(skb)->nr_frags--;
-		} else {
-			skb_shinfo(skb)->frags[0].page_offset +=
-				sizeof(hdr->mhdr) + copy;
-			skb_shinfo(skb)->frags[0].size = len;
-			skb->data_len += len;
-			skb->len += len;
-		}
+	len -= hdr_len;
+	p += offset;
 
-		while (--hdr->mhdr.num_buffers) {
-			struct sk_buff *nskb;
+	copy = len;
+	if (copy > skb_tailroom(skb))
+		copy = skb_tailroom(skb);
+	memcpy(skb_put(skb, copy), p, copy);
 
-			i = skb_shinfo(skb)->nr_frags;
-			if (i >= MAX_SKB_FRAGS) {
-				pr_debug("%s: packet too long %d\n", dev->name,
-					 len);
-				dev->stats.rx_length_errors++;
-				goto drop;
-			}
+	len -= copy;
+	offset += copy;
 
-			nskb = vi->rvq->vq_ops->get_buf(vi->rvq, &len);
-			if (!nskb) {
-				pr_debug("%s: rx error: %d buffers missing\n",
-					 dev->name, hdr->mhdr.num_buffers);
-				dev->stats.rx_length_errors++;
-				goto drop;
-			}
+	while (len) {
+		set_skb_frag(skb, page, offset, &len);
+		page = (struct page *)page->private;
+		offset = 0;
+	}
 
-			__skb_unlink(nskb, &vi->recv);
-			vi->num--;
+	if (page)
+		give_pages(vi, page);
 
-			skb_shinfo(skb)->frags[i] = skb_shinfo(nskb)->frags[0];
-			skb_shinfo(nskb)->nr_frags = 0;
-			kfree_skb(nskb);
+	return skb;
+}
 
-			if (len > PAGE_SIZE)
-				len = PAGE_SIZE;
+static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
+{
+	struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
+	struct page *page;
+	int num_buf, i, len;
+
+	num_buf = hdr->mhdr.num_buffers;
+	while (--num_buf) {
+		i = skb_shinfo(skb)->nr_frags;
+		if (i >= MAX_SKB_FRAGS) {
+			pr_debug("%s: packet too long\n", skb->dev->name);
+			skb->dev->stats.rx_length_errors++;
+			return -EINVAL;
+		}
 
-			skb_shinfo(skb)->frags[i].size = len;
-			skb_shinfo(skb)->nr_frags++;
-			skb->data_len += len;
-			skb->len += len;
+		page = vi->rvq->vq_ops->get_buf(vi->rvq, &len);
+		if (!page) {
+			pr_debug("%s: rx error: %d buffers missing\n",
+				 skb->dev->name, hdr->mhdr.num_buffers);
+			skb->dev->stats.rx_length_errors++;
+			return -EINVAL;
 		}
-	} else {
-		len -= sizeof(hdr->hdr);
+		if (len > PAGE_SIZE)
+			len = PAGE_SIZE;
+
+		set_skb_frag(skb, page, 0, &len);
+
+		--vi->num;
+	}
+	return 0;
+}
+
+static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
+{
+	struct virtnet_info *vi = netdev_priv(dev);
+	struct sk_buff *skb;
+	struct page *page;
+	struct skb_vnet_hdr *hdr;
 
-		if (len <= MAX_PACKET_LEN)
-			trim_pages(vi, skb);
+	if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
+		pr_debug("%s: short packet %i\n", dev->name, len);
+		dev->stats.rx_length_errors++;
+		if (vi->mergeable_rx_bufs || vi->big_packets)
+			give_pages(vi, buf);
+		else
+			dev_kfree_skb(buf);
+		return;
+	}
 
-		err = pskb_trim(skb, len);
-		if (err) {
-			pr_debug("%s: pskb_trim failed %i %d\n", dev->name,
-				 len, err);
+	if (!vi->mergeable_rx_bufs && !vi->big_packets) {
+		skb = buf;
+		len -= sizeof(struct virtio_net_hdr);
+		skb_trim(skb, len);
+	} else {
+		page = buf;
+		skb = page_to_skb(vi, page, len);
+		if (unlikely(!skb)) {
 			dev->stats.rx_dropped++;
-			goto drop;
+			give_pages(vi, page);
+			return;
 		}
+		if (vi->mergeable_rx_bufs)
+			if (receive_mergeable(vi, skb)) {
+				dev_kfree_skb(skb);
+				return;
+			}
 	}
 
+	hdr = skb_vnet_hdr(skb);
 	skb->truesize += skb->data_len;
 	dev->stats.rx_bytes += skb->len;
 	dev->stats.rx_packets++;
@@ -267,110 +319,119 @@ static void receive_skb(struct net_device *dev, struct sk_buff *skb,
 
 frame_err:
 	dev->stats.rx_frame_errors++;
-drop:
 	dev_kfree_skb(skb);
 }
 
-static bool try_fill_recv_maxbufs(struct virtnet_info *vi, gfp_t gfp)
+static int add_recvbuf_small(struct virtnet_info *vi, gfp_t gfp)
 {
 	struct sk_buff *skb;
-	struct scatterlist sg[2+MAX_SKB_FRAGS];
-	int num, err, i;
-	bool oom = false;
-
-	sg_init_table(sg, 2+MAX_SKB_FRAGS);
-	do {
-		struct skb_vnet_hdr *hdr;
+	struct skb_vnet_hdr *hdr;
+	struct scatterlist sg[2];
+	int err;
 
-		skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
-		if (unlikely(!skb)) {
-			oom = true;
-			break;
-		}
+	skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
+	if (unlikely(!skb))
+		return -ENOMEM;
 
-		skb_put(skb, MAX_PACKET_LEN);
+	skb_put(skb, MAX_PACKET_LEN);
 
-		hdr = skb_vnet_hdr(skb);
-		sg_set_buf(sg, &hdr->hdr, sizeof(hdr->hdr));
+	hdr = skb_vnet_hdr(skb);
+	sg_set_buf(sg, &hdr->hdr, sizeof hdr->hdr);
 
-		if (vi->big_packets) {
-			for (i = 0; i < MAX_SKB_FRAGS; i++) {
-				skb_frag_t *f = &skb_shinfo(skb)->frags[i];
-				f->page = get_a_page(vi, gfp);
-				if (!f->page)
-					break;
+	skb_to_sgvec(skb, sg + 1, 0, skb->len);
 
-				f->page_offset = 0;
-				f->size = PAGE_SIZE;
+	err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, 2, skb);
+	if (err < 0)
+		dev_kfree_skb(skb);
 
-				skb->data_len += PAGE_SIZE;
-				skb->len += PAGE_SIZE;
+	return err;
+}
 
-				skb_shinfo(skb)->nr_frags++;
-			}
+static int add_recvbuf_big(struct virtnet_info *vi, gfp_t gfp)
+{
+	struct scatterlist sg[MAX_SKB_FRAGS + 2];
+	struct page *first, *list = NULL;
+	char *p;
+	int i, err, offset;
+
+	/* page in sg[MAX_SKB_FRAGS + 1] is list tail */
+	for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
+		first = get_a_page(vi, gfp);
+		if (!first) {
+			if (list)
+				give_pages(vi, list);
+			return -ENOMEM;
 		}
+		sg_set_buf(&sg[i], page_address(first), PAGE_SIZE);
 
-		num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
-		skb_queue_head(&vi->recv, skb);
+		/* chain new page in list head to match sg */
+		first->private = (unsigned long)list;
+		list = first;
+	}
 
-		err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, num, skb);
-		if (err < 0) {
-			skb_unlink(skb, &vi->recv);
-			trim_pages(vi, skb);
-			kfree_skb(skb);
-			break;
-		}
-		vi->num++;
-	} while (err >= num);
-	if (unlikely(vi->num > vi->max))
-		vi->max = vi->num;
-	vi->rvq->vq_ops->kick(vi->rvq);
-	return !oom;
+	first = get_a_page(vi, gfp);
+	if (!first) {
+		give_pages(vi, list);
+		return -ENOMEM;
+	}
+	p = page_address(first);
+
+	/* sg[0], sg[1] share the same page */
+	/* a separated sg[0] for  virtio_net_hdr only during to QEMU bug*/
+	sg_set_buf(&sg[0], p, sizeof(struct virtio_net_hdr));
+
+	/* sg[1] for data packet, from offset */
+	offset = sizeof(struct padded_vnet_hdr);
+	sg_set_buf(&sg[1], p + offset, PAGE_SIZE - offset);
+
+	/* chain first in list head */
+	first->private = (unsigned long)list;
+	err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, MAX_SKB_FRAGS + 2,
+				       first);
+	if (err < 0)
+		give_pages(vi, first);
+
+	return err;
 }
 
-/* Returns false if we couldn't fill entirely (OOM). */
-static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
+static int add_recvbuf_mergeable(struct virtnet_info *vi, gfp_t gfp)
 {
-	struct sk_buff *skb;
-	struct scatterlist sg[1];
+	struct page *page;
+	struct scatterlist sg;
 	int err;
-	bool oom = false;
-
-	if (!vi->mergeable_rx_bufs)
-		return try_fill_recv_maxbufs(vi, gfp);
 
-	do {
-		skb_frag_t *f;
+	page = get_a_page(vi, gfp);
+	if (!page)
+		return -ENOMEM;
 
-		skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
-		if (unlikely(!skb)) {
-			oom = true;
-			break;
-		}
+	sg_init_one(&sg, page_address(page), PAGE_SIZE);
 
-		f = &skb_shinfo(skb)->frags[0];
-		f->page = get_a_page(vi, gfp);
-		if (!f->page) {
-			oom = true;
-			kfree_skb(skb);
-			break;
-		}
+	err = vi->rvq->vq_ops->add_buf(vi->rvq, &sg, 0, 1, page);
+	if (err < 0)
+		give_pages(vi, page);
 
-		f->page_offset = 0;
-		f->size = PAGE_SIZE;
+	return err;
+}
 
-		skb_shinfo(skb)->nr_frags++;
+/* Returns false if we couldn't fill entirely (OOM). */
+static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
+{
+	int err;
+	bool oom = false;
 
-		sg_init_one(sg, page_address(f->page), PAGE_SIZE);
-		skb_queue_head(&vi->recv, skb);
+	do {
+		if (vi->mergeable_rx_bufs)
+			err = add_recvbuf_mergeable(vi, gfp);
+		else if (vi->big_packets)
+			err = add_recvbuf_big(vi, gfp);
+		else
+			err = add_recvbuf_small(vi, gfp);
 
-		err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, 1, skb);
 		if (err < 0) {
-			skb_unlink(skb, &vi->recv);
-			kfree_skb(skb);
+			oom = true;
 			break;
 		}
-		vi->num++;
+		++vi->num;
 	} while (err > 0);
 	if (unlikely(vi->num > vi->max))
 		vi->max = vi->num;
@@ -408,15 +469,14 @@ static void refill_work(struct work_struct *work)
 static int virtnet_poll(struct napi_struct *napi, int budget)
 {
 	struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
-	struct sk_buff *skb = NULL;
+	void *buf;
 	unsigned int len, received = 0;
 
 again:
 	while (received < budget &&
-	       (skb = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
-		__skb_unlink(skb, &vi->recv);
-		receive_skb(vi->dev, skb, len);
-		vi->num--;
+	       (buf = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
+		receive_buf(vi->dev, buf, len);
+		--vi->num;
 		received++;
 	}
 
@@ -496,9 +556,9 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 
 	/* Encode metadata header at front. */
 	if (vi->mergeable_rx_bufs)
-		sg_set_buf(sg, &hdr->mhdr, sizeof(hdr->mhdr));
+		sg_set_buf(sg, &hdr->mhdr, sizeof hdr->mhdr);
 	else
-		sg_set_buf(sg, &hdr->hdr, sizeof(hdr->hdr));
+		sg_set_buf(sg, &hdr->hdr, sizeof hdr->hdr);
 
 	hdr->num_sg = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 	return vi->svq->vq_ops->add_buf(vi->svq, sg, hdr->num_sg, 0, skb);
@@ -916,8 +976,7 @@ static int virtnet_probe(struct virtio_device *vdev)
 			dev->features |= NETIF_F_HW_VLAN_FILTER;
 	}
 
-	/* Initialize our empty receive and send queues. */
-	skb_queue_head_init(&vi->recv);
+	/* Initialize our empty send queue. */
 	skb_queue_head_init(&vi->send);
 
 	err = register_netdev(dev);
@@ -952,25 +1011,35 @@ free:
 	return err;
 }
 
+static void free_unused_bufs(struct virtnet_info *vi)
+{
+	void *buf;
+	while (1) {
+		buf = vi->rvq->vq_ops->detach_unused_buf(vi->rvq);
+		if (!buf)
+			break;
+		if (vi->mergeable_rx_bufs || vi->big_packets)
+			give_pages(vi, buf);
+		else
+			dev_kfree_skb(buf);
+		--vi->num;
+	}
+	BUG_ON(vi->num != 0);
+}
+
 static void __devexit virtnet_remove(struct virtio_device *vdev)
 {
 	struct virtnet_info *vi = vdev->priv;
-	struct sk_buff *skb;
 
 	/* Stop all the virtqueues. */
 	vdev->config->reset(vdev);
 
-	/* Free our skbs in send and recv queues, if any. */
-	while ((skb = __skb_dequeue(&vi->recv)) != NULL) {
-		kfree_skb(skb);
-		vi->num--;
-	}
+	/* Free our skbs in send queue, if any. */
 	__skb_queue_purge(&vi->send);
 
-	BUG_ON(vi->num != 0);
-
 	unregister_netdev(vi->dev);
 	cancel_delayed_work_sync(&vi->refill);
+	free_unused_bufs(vi);
 
 	vdev->config->del_vqs(vi->vdev);

Re: [PATCH 1/2] virtio: Add ability to detach unused buffers from vrings

[David Miller]

From: Rusty Russell <rusty@rustcorp.com.au>
Date: Fri, 29 Jan 2010 23:49:05 +1030

> From: Shirley Ma <mashirle@us.ibm.com>
> 
> There's currently no way for a virtio driver to ask for unused
> buffers, so it has to keep a list itself to reclaim them at shutdown.
> This is redundant, since virtio_ring stores that information.  So
> add a new hook to do this.
> 
> Signed-off-by: Shirley Ma <xma@us.ibm.com>
> Signed-off-by: Amit Shah <amit.shah@redhat.com>
> Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Applied.

Re: [PATCH 1/2] virtio: Add ability to detach unused buffers from vrings

[David Miller]

From: Rusty Russell <rusty@rustcorp.com.au>
Date: Fri, 29 Jan 2010 23:49:05 +1030

> From: Shirley Ma <mashirle@us.ibm.com>
> 
> There's currently no way for a virtio driver to ask for unused
> buffers, so it has to keep a list itself to reclaim them at shutdown.
> This is redundant, since virtio_ring stores that information.  So
> add a new hook to do this.
> 
> Signed-off-by: Shirley Ma <xma@us.ibm.com>
> Signed-off-by: Amit Shah <amit.shah@redhat.com>
> Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Applied.

Re: [PATCH 0/2] virtio net improvements

[Rusty Russell]

On Fri, 29 Jan 2010 11:46:43 pm Rusty Russell wrote:
> Hi Dave,
> 
>    Nice driver optimization from Shirley, but requires a new virtio hook.
> Do you want to take both?  I have nothing else overlapping it.

Dave, any news on this?  Shirley, I note that your final commit message did
not contain this information from your original version:

	Tests have been done for small packets, big packets and
	mergeable buffers.

	The single netperf TCP_STREAM performance improved for host to guest. 
	It also reduces UDP packets drop rate.

	The netperf laptop results were:

	mtu=1500
	netperf -H xxx -l 120

	                w/o patch       w/i patch (two runs)    
	guest to host:  3336.84Mb/s   3730.14Mb/s ~ 3582.88Mb/s

	host to guest:  3165.10Mb/s   3370.39Mb/s ~ 3407.96Mb/s

Thanks,
Rusty.

Re: [PATCH 0/2] virtio net improvements

[Rusty Russell]

On Fri, 29 Jan 2010 11:46:43 pm Rusty Russell wrote:
> Hi Dave,
> 
>    Nice driver optimization from Shirley, but requires a new virtio hook.
> Do you want to take both?  I have nothing else overlapping it.

Dave, any news on this?  Shirley, I note that your final commit message did
not contain this information from your original version:

	Tests have been done for small packets, big packets and
	mergeable buffers.

	The single netperf TCP_STREAM performance improved for host to guest. 
	It also reduces UDP packets drop rate.

	The netperf laptop results were:

	mtu=1500
	netperf -H xxx -l 120

	                w/o patch       w/i patch (two runs)    
	guest to host:  3336.84Mb/s   3730.14Mb/s ~ 3582.88Mb/s

	host to guest:  3165.10Mb/s   3370.39Mb/s ~ 3407.96Mb/s

Thanks,
Rusty.

Re: [PATCH 0/2] virtio net improvements

[David Miller]

From: Rusty Russell <rusty@rustcorp.com.au>
Date: Wed, 3 Feb 2010 09:57:06 +1030

> On Fri, 29 Jan 2010 11:46:43 pm Rusty Russell wrote:
>> Hi Dave,
>> 
>>    Nice driver optimization from Shirley, but requires a new virtio hook.
>> Do you want to take both?  I have nothing else overlapping it.
> 
> Dave, any news on this?

Just slowly creeping up the backlog :-)

Re: [PATCH 0/2] virtio net improvements

[David Miller]

From: Rusty Russell <rusty@rustcorp.com.au>
Date: Wed, 3 Feb 2010 09:57:06 +1030

> On Fri, 29 Jan 2010 11:46:43 pm Rusty Russell wrote:
>> Hi Dave,
>> 
>>    Nice driver optimization from Shirley, but requires a new virtio hook.
>> Do you want to take both?  I have nothing else overlapping it.
> 
> Dave, any news on this?

Just slowly creeping up the backlog :-)

[PATCH 1/1 net-next] virtio_net: remove send queue

[Shirley Ma]

Now we have a virtio detach API (in commit
f9bfbebf34eab707b065116cdc9699d25ba4252a), we don't need to track xmit
skbs in the virio_net driver, which improves transmission performance.

Signed-off-by: Shirley Ma <xma@us.ibm.com>
Acked-by: Rusty Russell <rusty@rustcorp.com.au>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
---
 drivers/net/virtio_net.c |   26 ++++++++------------------
 1 files changed, 8 insertions(+), 18 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 9d8984a..8069c08 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -56,9 +56,6 @@ struct virtnet_info
 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
 	bool mergeable_rx_bufs;
 
-	/* Send queue. */
-	struct sk_buff_head send;
-
 	/* Work struct for refilling if we run low on memory. */
 	struct delayed_work refill;
 
@@ -505,7 +502,6 @@ static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
 
 	while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
 		pr_debug("Sent skb %p\n", skb);
-		__skb_unlink(skb, &vi->send);
 		vi->dev->stats.tx_bytes += skb->len;
 		vi->dev->stats.tx_packets++;
 		tot_sgs += skb_vnet_hdr(skb)->num_sg;
@@ -588,15 +584,6 @@ again:
 	}
 	vi->svq->vq_ops->kick(vi->svq);
 
-	/*
-	 * Put new one in send queue.  You'd expect we'd need this before
-	 * xmit_skb calls add_buf(), since the callback can be triggered
-	 * immediately after that.  But since the callback just triggers
-	 * another call back here, normal network xmit locking prevents the
-	 * race.
-	 */
-	__skb_queue_head(&vi->send, skb);
-
 	/* Don't wait up for transmitted skbs to be freed. */
 	skb_orphan(skb);
 	nf_reset(skb);
@@ -977,9 +964,6 @@ static int virtnet_probe(struct virtio_device *vdev)
 			dev->features |= NETIF_F_HW_VLAN_FILTER;
 	}
 
-	/* Initialize our empty send queue. */
-	skb_queue_head_init(&vi->send);
-
 	err = register_netdev(dev);
 	if (err) {
 		pr_debug("virtio_net: registering device failed\n");
@@ -1016,6 +1000,12 @@ static void free_unused_bufs(struct virtnet_info *vi)
 {
 	void *buf;
 	while (1) {
+		buf = vi->svq->vq_ops->detach_unused_buf(vi->svq);
+		if (!buf)
+			break;
+		dev_kfree_skb(buf);
+	}
+	while (1) {
 		buf = vi->rvq->vq_ops->detach_unused_buf(vi->rvq);
 		if (!buf)
 			break;
@@ -1035,11 +1025,11 @@ static void __devexit virtnet_remove(struct virtio_device *vdev)
 	/* Stop all the virtqueues. */
 	vdev->config->reset(vdev);
 
-	/* Free our skbs in send queue, if any. */
-	__skb_queue_purge(&vi->send);
 
 	unregister_netdev(vi->dev);
 	cancel_delayed_work_sync(&vi->refill);
+
+	/* Free unused buffers in both send and recv, if any. */
 	free_unused_bufs(vi);
 
 	vdev->config->del_vqs(vi->vdev);

Re: [PATCH 1/1 net-next] virtio_net: remove send queue

[Shirley Ma]

netperf results showed about 20% transmission gain with this patch for
virtio_net.

Thanks
Shirley

Re: [PATCH 1/1 net-next] virtio_net: remove send queue

[David Miller]

From: Shirley Ma <mashirle@us.ibm.com>
Date: Mon, 08 Feb 2010 16:14:42 -0800

> Now we have a virtio detach API (in commit
> f9bfbebf34eab707b065116cdc9699d25ba4252a), we don't need to track xmit
> skbs in the virio_net driver, which improves transmission performance.
> 
> Signed-off-by: Shirley Ma <xma@us.ibm.com>
> Acked-by: Rusty Russell <rusty@rustcorp.com.au>
> Acked-by: Michael S. Tsirkin <mst@redhat.com>

Applied to net-next-2.6, thanks everyone.