Patchwork virtio-net: put virtio net header inline with data

login
register
mail settings
Submitter Michael S. Tsirkin
Date June 6, 2013, 9:55 a.m.
Message ID <20130606095456.GA7865@redhat.com>
Download mbox | patch
Permalink /patch/249340/
State New
Headers show

Comments

Michael S. Tsirkin - June 6, 2013, 9:55 a.m.
For small packets we can simplify xmit processing by linearizing buffers
with the header: most packets seem to have enough head room we can use
for this purpose.

Since some older hypervisors (e.g. qemu before version 1.5)
required that header is the first s/g element,
we need a feature bit for this.

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

This is a repost of my old patch, rebased to latest kernel.

Before:
Thu Jun  6 05:24:59 EDT 2013
TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
11.0.0.4 (11.0.0.4) port 0 AF_INET : demo
Local /Remote
Socket Size   Request  Resp.   Elapsed  Trans.
Send   Recv   Size     Size    Time     Rate         
bytes  Bytes  bytes    bytes   secs.    per sec   

16384  87380  1        1       10.00    12931.13   


After:
TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
11.0.0.4 (11.0.0.4) port 0 AF_INET : demo
Local /Remote
Socket Size   Request  Resp.   Elapsed  Trans.
Send   Recv   Size     Size    Time     Rate         
bytes  Bytes  bytes    bytes   secs.    per sec   

16384  87380  1        1       10.00    14151.12


 drivers/net/virtio_net.c        | 42 +++++++++++++++++++++++++++++++++--------
 include/uapi/linux/virtio_net.h |  6 +++++-
 2 files changed, 39 insertions(+), 9 deletions(-)
Jesse Larrew - June 6, 2013, 7:59 p.m.
Hi Michael!

On 06/06/2013 04:55 AM, Michael S. Tsirkin wrote:
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index c9e0038..d35a097 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -106,6 +106,9 @@ struct virtnet_info {
>  	/* Has control virtqueue */
>  	bool has_cvq;
> 
> +	/* Host can handle any s/g split between our header and packet data */
> +	bool any_header_sg;
> +
>  	/* enable config space updates */
>  	bool config_enable;
> 
> @@ -668,12 +671,28 @@ static void free_old_xmit_skbs(struct send_queue *sq)
> 
>  static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
>  {
> -	struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
> +	struct skb_vnet_hdr *hdr;
>  	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
>  	struct virtnet_info *vi = sq->vq->vdev->priv;
>  	unsigned num_sg;
> +	unsigned hdr_len;
> +	bool can_push;
> 
>  	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
> +	if (vi->mergeable_rx_bufs)
> +		hdr_len = sizeof hdr->mhdr;
> +	else
> +		hdr_len = sizeof hdr->hdr;

All conditionals need braces.

> +
> +	can_push = vi->any_header_sg &&
> +		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
> +		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
> +	/* Even if we can, don't push here yet as this would skew
> +	 * csum_start offset below. */
> +	if (can_push)
> +		hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
> +	else
> +		hdr = skb_vnet_hdr(skb);

Ditto.

> 
>  	if (skb->ip_summed == CHECKSUM_PARTIAL) {
>  		hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
> @@ -702,15 +721,18 @@ static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
>  		hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
>  	}
> 
> -	hdr->mhdr.num_buffers = 0;
> -
> -	/* Encode metadata header at front. */
>  	if (vi->mergeable_rx_bufs)
> -		sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
> -	else
> -		sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
> +		hdr->mhdr.num_buffers = 0;
> 

Here too, please.

> -	num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
> +	if (can_push) {
> +		__skb_push(skb, hdr_len);
> +		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
> +		/* Pull header back to avoid skew in tx bytes calculations. */
> +		__skb_pull(skb, hdr_len);
> +	} else {
> +		sg_set_buf(sq->sg, hdr, hdr_len);
> +		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
> +	}
>  	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
>  }
> 
> @@ -1554,6 +1576,9 @@ static int virtnet_probe(struct virtio_device *vdev)
>  	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
>  		vi->mergeable_rx_bufs = true;
> 

This is just context, but we may as well fix this as well.

> +	if (virtio_has_feature(vdev, VIRTIO_NET_F_ANY_HEADER_SG))
> +		vi->any_header_sg = true;
> +

Braces here, please.

>  	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
>  		vi->has_cvq = true;
> 

Might as well fix this too.

> @@ -1729,6 +1754,7 @@ static unsigned int features[] = {
>  	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_ANY_HEADER_SG,
>  };
> 
>  static struct virtio_driver virtio_net_driver = {
> diff --git a/include/uapi/linux/virtio_net.h b/include/uapi/linux/virtio_net.h
> index c520203..9c98b7d 100644
> --- a/include/uapi/linux/virtio_net.h
> +++ b/include/uapi/linux/virtio_net.h
> @@ -55,6 +55,8 @@
>  					 * Steering */
>  #define VIRTIO_NET_F_CTRL_MAC_ADDR 23	/* Set MAC address */
> 
> +#define VIRTIO_NET_F_ANY_HEADER_SG 25	/* Host can handle any header s/g */
> +
>  #define VIRTIO_NET_S_LINK_UP	1	/* Link is up */
>  #define VIRTIO_NET_S_ANNOUNCE	2	/* Announcement is needed */
> 
> @@ -70,7 +72,9 @@ struct virtio_net_config {
>  	__u16 max_virtqueue_pairs;
>  } __attribute__((packed));
> 
> -/* This is the first element of the scatter-gather list.  If you don't
> +/* This header comes first in the scatter-gather list.
> + * If VIRTIO_NET_F_ANY_HEADER_SG is not negotiated, it must
> + * be the first element of the scatter-gather list.  If you don't
>   * specify GSO or CSUM features, you can simply ignore the header. */
>  struct virtio_net_hdr {
>  #define VIRTIO_NET_HDR_F_NEEDS_CSUM	1	// Use csum_start, csum_offset
> 

Sincerely,

Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
jlarrew@linux.vnet.ibm.com
Dave Jones - June 6, 2013, 8:09 p.m.
On Thu, Jun 06, 2013 at 02:59:44PM -0500, Jesse Larrew wrote:
 
 > >  	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 > > +	if (vi->mergeable_rx_bufs)
 > > +		hdr_len = sizeof hdr->mhdr;
 > > +	else
 > > +		hdr_len = sizeof hdr->hdr;
 > 
 > All conditionals need braces.
 
Documentation/CodingStyle disagrees:

 "Do not unnecessarily use braces where a single statement will do."
 
	Dave
Jesse Larrew - June 6, 2013, 8:18 p.m.
On 06/06/2013 03:09 PM, Dave Jones wrote:
> On Thu, Jun 06, 2013 at 02:59:44PM -0500, Jesse Larrew wrote:
> 
>  > >  	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
>  > > +	if (vi->mergeable_rx_bufs)
>  > > +		hdr_len = sizeof hdr->mhdr;
>  > > +	else
>  > > +		hdr_len = sizeof hdr->hdr;
>  > 
>  > All conditionals need braces.
> 
> Documentation/CodingStyle disagrees:
> 
>  "Do not unnecessarily use braces where a single statement will do."
> 
> 	Dave 
> 

Ah, yes. This is kernel code. My mistake. :)

Jesse Larrew
Software Engineer, KVM Team
IBM Linux Technology Center
Phone: (512) 973-2052 (T/L: 363-2052)
jlarrew@linux.vnet.ibm.com
Rusty Russell - June 7, 2013, 2:12 a.m.
"Michael S. Tsirkin" <mst@redhat.com> writes:
> For small packets we can simplify xmit processing by linearizing buffers
> with the header: most packets seem to have enough head room we can use
> for this purpose.
>
> Since some older hypervisors (e.g. qemu before version 1.5)
> required that header is the first s/g element,
> we need a feature bit for this.

OK, we know this is horrible.  But I will sleep better knowing that we
this feature need never make it into a final 1.0 spec, since it can be
assumed at that point...

>  	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
> +	if (vi->mergeable_rx_bufs)
> +		hdr_len = sizeof hdr->mhdr;
> +	else
> +		hdr_len = sizeof hdr->hdr;
> +
> +	can_push = vi->any_header_sg &&
> +		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
> +		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;

Idle thought: how often does this fail?  Would it suck if we copied
headers which didn't let us prepend data?  Or could we bump
dev->hard_header_len appropriately?

Thanks,
Rusty.
Jason Wang - June 7, 2013, 2:52 a.m.
On 06/06/2013 05:55 PM, Michael S. Tsirkin wrote:
> For small packets we can simplify xmit processing by linearizing buffers
> with the header: most packets seem to have enough head room we can use
> for this purpose.
>
> Since some older hypervisors (e.g. qemu before version 1.5)
> required that header is the first s/g element,
> we need a feature bit for this.
>
> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> ---

Hi Michael:

The idea looks good, but there are some questions:

- What if we just use direct descriptors with sgs <=2 and double the
ring size?
- I believe we may add more things into vnet header in the future, so
this trick may not work because of limited header room.

Thanks
> This is a repost of my old patch, rebased to latest kernel.
>
> Before:
> Thu Jun  6 05:24:59 EDT 2013
> TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
> 11.0.0.4 (11.0.0.4) port 0 AF_INET : demo
> Local /Remote
> Socket Size   Request  Resp.   Elapsed  Trans.
> Send   Recv   Size     Size    Time     Rate         
> bytes  Bytes  bytes    bytes   secs.    per sec   
>
> 16384  87380  1        1       10.00    12931.13   
>
>
> After:
> TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
> 11.0.0.4 (11.0.0.4) port 0 AF_INET : demo
> Local /Remote
> Socket Size   Request  Resp.   Elapsed  Trans.
> Send   Recv   Size     Size    Time     Rate         
> bytes  Bytes  bytes    bytes   secs.    per sec   
>
> 16384  87380  1        1       10.00    14151.12
>
>
>  drivers/net/virtio_net.c        | 42 +++++++++++++++++++++++++++++++++--------
>  include/uapi/linux/virtio_net.h |  6 +++++-
>  2 files changed, 39 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index c9e0038..d35a097 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -106,6 +106,9 @@ struct virtnet_info {
>  	/* Has control virtqueue */
>  	bool has_cvq;
>  
> +	/* Host can handle any s/g split between our header and packet data */
> +	bool any_header_sg;
> +
>  	/* enable config space updates */
>  	bool config_enable;
>  
> @@ -668,12 +671,28 @@ static void free_old_xmit_skbs(struct send_queue *sq)
>  
>  static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
>  {
> -	struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
> +	struct skb_vnet_hdr *hdr;
>  	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
>  	struct virtnet_info *vi = sq->vq->vdev->priv;
>  	unsigned num_sg;
> +	unsigned hdr_len;
> +	bool can_push;
>  
>  	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
> +	if (vi->mergeable_rx_bufs)
> +		hdr_len = sizeof hdr->mhdr;
> +	else
> +		hdr_len = sizeof hdr->hdr;
> +
> +	can_push = vi->any_header_sg &&
> +		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
> +		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
> +	/* Even if we can, don't push here yet as this would skew
> +	 * csum_start offset below. */
> +	if (can_push)
> +		hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
> +	else
> +		hdr = skb_vnet_hdr(skb);
>  
>  	if (skb->ip_summed == CHECKSUM_PARTIAL) {
>  		hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
> @@ -702,15 +721,18 @@ static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
>  		hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
>  	}
>  
> -	hdr->mhdr.num_buffers = 0;
> -
> -	/* Encode metadata header at front. */
>  	if (vi->mergeable_rx_bufs)
> -		sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
> -	else
> -		sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
> +		hdr->mhdr.num_buffers = 0;
>  
> -	num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
> +	if (can_push) {
> +		__skb_push(skb, hdr_len);
> +		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
> +		/* Pull header back to avoid skew in tx bytes calculations. */
> +		__skb_pull(skb, hdr_len);
> +	} else {
> +		sg_set_buf(sq->sg, hdr, hdr_len);
> +		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
> +	}
>  	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
>  }
>  
> @@ -1554,6 +1576,9 @@ static int virtnet_probe(struct virtio_device *vdev)
>  	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
>  		vi->mergeable_rx_bufs = true;
>  
> +	if (virtio_has_feature(vdev, VIRTIO_NET_F_ANY_HEADER_SG))
> +		vi->any_header_sg = true;
> +
>  	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
>  		vi->has_cvq = true;
>  
> @@ -1729,6 +1754,7 @@ static unsigned int features[] = {
>  	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_ANY_HEADER_SG,
>  };
>  
>  static struct virtio_driver virtio_net_driver = {
> diff --git a/include/uapi/linux/virtio_net.h b/include/uapi/linux/virtio_net.h
> index c520203..9c98b7d 100644
> --- a/include/uapi/linux/virtio_net.h
> +++ b/include/uapi/linux/virtio_net.h
> @@ -55,6 +55,8 @@
>  					 * Steering */
>  #define VIRTIO_NET_F_CTRL_MAC_ADDR 23	/* Set MAC address */
>  
> +#define VIRTIO_NET_F_ANY_HEADER_SG 25	/* Host can handle any header s/g */
> +
>  #define VIRTIO_NET_S_LINK_UP	1	/* Link is up */
>  #define VIRTIO_NET_S_ANNOUNCE	2	/* Announcement is needed */
>  
> @@ -70,7 +72,9 @@ struct virtio_net_config {
>  	__u16 max_virtqueue_pairs;
>  } __attribute__((packed));
>  
> -/* This is the first element of the scatter-gather list.  If you don't
> +/* This header comes first in the scatter-gather list.
> + * If VIRTIO_NET_F_ANY_HEADER_SG is not negotiated, it must
> + * be the first element of the scatter-gather list.  If you don't
>   * specify GSO or CSUM features, you can simply ignore the header. */
>  struct virtio_net_hdr {
>  #define VIRTIO_NET_HDR_F_NEEDS_CSUM	1	// Use csum_start, csum_offset
Michael S. Tsirkin - June 9, 2013, 6:52 a.m.
On Fri, Jun 07, 2013 at 10:52:01AM +0800, Jason Wang wrote:
> On 06/06/2013 05:55 PM, Michael S. Tsirkin wrote:
> > For small packets we can simplify xmit processing by linearizing buffers
> > with the header: most packets seem to have enough head room we can use
> > for this purpose.
> >
> > Since some older hypervisors (e.g. qemu before version 1.5)
> > required that header is the first s/g element,
> > we need a feature bit for this.
> >
> > Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> > ---
> 
> Hi Michael:
> 
> The idea looks good, but there are some questions:
> 
> - What if we just use direct descriptors with sgs <=2 and double the
> ring size?

It's not easy to increase the ring size, we have trouble allocating
large rings on some systems as it is.
So far, any attempts to use direct for sg > 1 seem to hurt some
workloads.

> - I believe we may add more things into vnet header in the future, so
> this trick may not work because of limited header room.
> 
> Thanks

Didn't happen in a long while.
We can come up with more tricks when we run against this limit.

> > This is a repost of my old patch, rebased to latest kernel.
> >
> > Before:
> > Thu Jun  6 05:24:59 EDT 2013
> > TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
> > 11.0.0.4 (11.0.0.4) port 0 AF_INET : demo
> > Local /Remote
> > Socket Size   Request  Resp.   Elapsed  Trans.
> > Send   Recv   Size     Size    Time     Rate         
> > bytes  Bytes  bytes    bytes   secs.    per sec   
> >
> > 16384  87380  1        1       10.00    12931.13   
> >
> >
> > After:
> > TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
> > 11.0.0.4 (11.0.0.4) port 0 AF_INET : demo
> > Local /Remote
> > Socket Size   Request  Resp.   Elapsed  Trans.
> > Send   Recv   Size     Size    Time     Rate         
> > bytes  Bytes  bytes    bytes   secs.    per sec   
> >
> > 16384  87380  1        1       10.00    14151.12
> >
> >
> >  drivers/net/virtio_net.c        | 42 +++++++++++++++++++++++++++++++++--------
> >  include/uapi/linux/virtio_net.h |  6 +++++-
> >  2 files changed, 39 insertions(+), 9 deletions(-)
> >
> > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > index c9e0038..d35a097 100644
> > --- a/drivers/net/virtio_net.c
> > +++ b/drivers/net/virtio_net.c
> > @@ -106,6 +106,9 @@ struct virtnet_info {
> >  	/* Has control virtqueue */
> >  	bool has_cvq;
> >  
> > +	/* Host can handle any s/g split between our header and packet data */
> > +	bool any_header_sg;
> > +
> >  	/* enable config space updates */
> >  	bool config_enable;
> >  
> > @@ -668,12 +671,28 @@ static void free_old_xmit_skbs(struct send_queue *sq)
> >  
> >  static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
> >  {
> > -	struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
> > +	struct skb_vnet_hdr *hdr;
> >  	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
> >  	struct virtnet_info *vi = sq->vq->vdev->priv;
> >  	unsigned num_sg;
> > +	unsigned hdr_len;
> > +	bool can_push;
> >  
> >  	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
> > +	if (vi->mergeable_rx_bufs)
> > +		hdr_len = sizeof hdr->mhdr;
> > +	else
> > +		hdr_len = sizeof hdr->hdr;
> > +
> > +	can_push = vi->any_header_sg &&
> > +		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
> > +		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
> > +	/* Even if we can, don't push here yet as this would skew
> > +	 * csum_start offset below. */
> > +	if (can_push)
> > +		hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
> > +	else
> > +		hdr = skb_vnet_hdr(skb);
> >  
> >  	if (skb->ip_summed == CHECKSUM_PARTIAL) {
> >  		hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
> > @@ -702,15 +721,18 @@ static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
> >  		hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
> >  	}
> >  
> > -	hdr->mhdr.num_buffers = 0;
> > -
> > -	/* Encode metadata header at front. */
> >  	if (vi->mergeable_rx_bufs)
> > -		sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
> > -	else
> > -		sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
> > +		hdr->mhdr.num_buffers = 0;
> >  
> > -	num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
> > +	if (can_push) {
> > +		__skb_push(skb, hdr_len);
> > +		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
> > +		/* Pull header back to avoid skew in tx bytes calculations. */
> > +		__skb_pull(skb, hdr_len);
> > +	} else {
> > +		sg_set_buf(sq->sg, hdr, hdr_len);
> > +		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
> > +	}
> >  	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
> >  }
> >  
> > @@ -1554,6 +1576,9 @@ static int virtnet_probe(struct virtio_device *vdev)
> >  	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
> >  		vi->mergeable_rx_bufs = true;
> >  
> > +	if (virtio_has_feature(vdev, VIRTIO_NET_F_ANY_HEADER_SG))
> > +		vi->any_header_sg = true;
> > +
> >  	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
> >  		vi->has_cvq = true;
> >  
> > @@ -1729,6 +1754,7 @@ static unsigned int features[] = {
> >  	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_ANY_HEADER_SG,
> >  };
> >  
> >  static struct virtio_driver virtio_net_driver = {
> > diff --git a/include/uapi/linux/virtio_net.h b/include/uapi/linux/virtio_net.h
> > index c520203..9c98b7d 100644
> > --- a/include/uapi/linux/virtio_net.h
> > +++ b/include/uapi/linux/virtio_net.h
> > @@ -55,6 +55,8 @@
> >  					 * Steering */
> >  #define VIRTIO_NET_F_CTRL_MAC_ADDR 23	/* Set MAC address */
> >  
> > +#define VIRTIO_NET_F_ANY_HEADER_SG 25	/* Host can handle any header s/g */
> > +
> >  #define VIRTIO_NET_S_LINK_UP	1	/* Link is up */
> >  #define VIRTIO_NET_S_ANNOUNCE	2	/* Announcement is needed */
> >  
> > @@ -70,7 +72,9 @@ struct virtio_net_config {
> >  	__u16 max_virtqueue_pairs;
> >  } __attribute__((packed));
> >  
> > -/* This is the first element of the scatter-gather list.  If you don't
> > +/* This header comes first in the scatter-gather list.
> > + * If VIRTIO_NET_F_ANY_HEADER_SG is not negotiated, it must
> > + * be the first element of the scatter-gather list.  If you don't
> >   * specify GSO or CSUM features, you can simply ignore the header. */
> >  struct virtio_net_hdr {
> >  #define VIRTIO_NET_HDR_F_NEEDS_CSUM	1	// Use csum_start, csum_offset
Michael S. Tsirkin - June 9, 2013, 7:11 a.m.
On Fri, Jun 07, 2013 at 11:42:43AM +0930, Rusty Russell wrote:
> "Michael S. Tsirkin" <mst@redhat.com> writes:
> > For small packets we can simplify xmit processing by linearizing buffers
> > with the header: most packets seem to have enough head room we can use
> > for this purpose.
> >
> > Since some older hypervisors (e.g. qemu before version 1.5)
> > required that header is the first s/g element,
> > we need a feature bit for this.
> 
> OK, we know this is horrible.  But I will sleep better knowing that we
> this feature need never make it into a final 1.0 spec, since it can be
> assumed at that point...

Nod. Though if we want to require this for all devices,
virtio-blk scsi command passthrough will need to change -
I sent a spec patch a while ago
	virtio-spec: add field for scsi command size
any comments on it?

> >  	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
> > +	if (vi->mergeable_rx_bufs)
> > +		hdr_len = sizeof hdr->mhdr;
> > +	else
> > +		hdr_len = sizeof hdr->hdr;
> > +
> > +	can_push = vi->any_header_sg &&
> > +		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
> > +		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
> 
> Idle thought: how often does this fail?

I think it's mostly doesn't fail in my testing.

It's probably a good idea to add a counter here, then
if it starts triggering we can optimize.

I think things like skb_header_cloned depend on guest config
really, e.g. tcpdump running on the interface in guest can cause this.

>  Would it suck if we copied
> headers which didn't let us prepend data?

I think it will - copies are generally best avoided,
and header is easily 1K of data.

>  Or could we bump
> dev->hard_header_len appropriately?

Needs some thought, though from experience it's a pain.


> Thanks,
> Rusty.

Patch

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index c9e0038..d35a097 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -106,6 +106,9 @@  struct virtnet_info {
 	/* Has control virtqueue */
 	bool has_cvq;
 
+	/* Host can handle any s/g split between our header and packet data */
+	bool any_header_sg;
+
 	/* enable config space updates */
 	bool config_enable;
 
@@ -668,12 +671,28 @@  static void free_old_xmit_skbs(struct send_queue *sq)
 
 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 {
-	struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
+	struct skb_vnet_hdr *hdr;
 	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 	struct virtnet_info *vi = sq->vq->vdev->priv;
 	unsigned num_sg;
+	unsigned hdr_len;
+	bool can_push;
 
 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
+	if (vi->mergeable_rx_bufs)
+		hdr_len = sizeof hdr->mhdr;
+	else
+		hdr_len = sizeof hdr->hdr;
+
+	can_push = vi->any_header_sg &&
+		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
+		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
+	/* Even if we can, don't push here yet as this would skew
+	 * csum_start offset below. */
+	if (can_push)
+		hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
+	else
+		hdr = skb_vnet_hdr(skb);
 
 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
 		hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
@@ -702,15 +721,18 @@  static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 		hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 	}
 
-	hdr->mhdr.num_buffers = 0;
-
-	/* Encode metadata header at front. */
 	if (vi->mergeable_rx_bufs)
-		sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
-	else
-		sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
+		hdr->mhdr.num_buffers = 0;
 
-	num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
+	if (can_push) {
+		__skb_push(skb, hdr_len);
+		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
+		/* Pull header back to avoid skew in tx bytes calculations. */
+		__skb_pull(skb, hdr_len);
+	} else {
+		sg_set_buf(sq->sg, hdr, hdr_len);
+		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
+	}
 	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 }
 
@@ -1554,6 +1576,9 @@  static int virtnet_probe(struct virtio_device *vdev)
 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
 		vi->mergeable_rx_bufs = true;
 
+	if (virtio_has_feature(vdev, VIRTIO_NET_F_ANY_HEADER_SG))
+		vi->any_header_sg = true;
+
 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
 		vi->has_cvq = true;
 
@@ -1729,6 +1754,7 @@  static unsigned int features[] = {
 	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_ANY_HEADER_SG,
 };
 
 static struct virtio_driver virtio_net_driver = {
diff --git a/include/uapi/linux/virtio_net.h b/include/uapi/linux/virtio_net.h
index c520203..9c98b7d 100644
--- a/include/uapi/linux/virtio_net.h
+++ b/include/uapi/linux/virtio_net.h
@@ -55,6 +55,8 @@ 
 					 * Steering */
 #define VIRTIO_NET_F_CTRL_MAC_ADDR 23	/* Set MAC address */
 
+#define VIRTIO_NET_F_ANY_HEADER_SG 25	/* Host can handle any header s/g */
+
 #define VIRTIO_NET_S_LINK_UP	1	/* Link is up */
 #define VIRTIO_NET_S_ANNOUNCE	2	/* Announcement is needed */
 
@@ -70,7 +72,9 @@  struct virtio_net_config {
 	__u16 max_virtqueue_pairs;
 } __attribute__((packed));
 
-/* This is the first element of the scatter-gather list.  If you don't
+/* This header comes first in the scatter-gather list.
+ * If VIRTIO_NET_F_ANY_HEADER_SG is not negotiated, it must
+ * be the first element of the scatter-gather list.  If you don't
  * specify GSO or CSUM features, you can simply ignore the header. */
 struct virtio_net_hdr {
 #define VIRTIO_NET_HDR_F_NEEDS_CSUM	1	// Use csum_start, csum_offset