Re: [Qemu-devel] [PATCH 04/14] virtio-net: avoid sg copy

[Anthony Liguori]

"Michael S. Tsirkin" <address@hidden> writes:

> Avoid tweaking iovec during receive. This removes
> the need to copy the vector.
> Note: we already have an evil cast in work_around_broken_dhclient
> and unfortunately this adds another one.
> const on buf is ignored by this function anyway so arguably
> this is not making things much worse.
>
> Signed-off-by: Michael S. Tsirkin <address@hidden>
> ---
>  hw/virtio-net.c | 45 +++++++++++++++++++++------------------------
>  iov.h           |  8 ++++----
>  2 files changed, 25 insertions(+), 28 deletions(-)
>
> diff --git a/hw/virtio-net.c b/hw/virtio-net.c
> index 0c5081e..6e53858 100644
> --- a/hw/virtio-net.c
> +++ b/hw/virtio-net.c
> @@ -503,7 +503,7 @@ static int virtio_net_has_buffers(VirtIONet *n, int 
> bufsize)
>   * we should provide a mechanism to disable it to avoid polluting the host
>   * cache.
>   */
> -static void work_around_broken_dhclient(struct virtio_net_hdr *hdr,
> +static void work_around_broken_dhclient(const struct virtio_net_hdr *hdr,
>                                          const uint8_t *buf, size_t size)
>  {
>      if ((hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) && /* missing csum */
> @@ -513,31 +513,27 @@ static void work_around_broken_dhclient(struct 
> virtio_net_hdr *hdr,
>          (buf[34] == 0 && buf[35] == 67)) { /* udp.srcport == bootps */
>          /* FIXME this cast is evil */
>          net_checksum_calculate((uint8_t *)buf, size);
> -        hdr->flags &= ~VIRTIO_NET_HDR_F_NEEDS_CSUM;
> +        ((struct virtio_net_hdr *)hdr)->flags &=
> ~VIRTIO_NET_HDR_F_NEEDS_CSUM;
>      }
>  }
>  
> -static int receive_header(VirtIONet *n, struct iovec *iov, int iovcnt,
> -                          const void *buf, size_t size, size_t hdr_len)
> +static int receive_header(VirtIONet *n, const struct iovec *iov, int iov_cnt,
> +                          const void *buf, size_t size)
>  {
> -    struct virtio_net_hdr *hdr = (struct virtio_net_hdr *)iov[0].iov_base;
>      int offset = 0;
>  
> -    hdr->flags = 0;
> -    hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
> -
>      if (n->has_vnet_hdr) {
> -        memcpy(hdr, buf, sizeof(*hdr));
> -        offset = sizeof(*hdr);
> -        work_around_broken_dhclient(hdr, buf + offset, size - offset);
> +        work_around_broken_dhclient(buf, buf + offset, size -
> offset);

I think it would be less ugly to cast here.

> +        offset = sizeof(struct virtio_net_hdr);
> +        iov_from_buf(iov, iov_cnt, 0, buf, offset);
> +    } else {
> +        struct virtio_net_hdr hdr = {
> +            .flags = 0,
> +            .gso_type = VIRTIO_NET_HDR_GSO_NONE
> +        };
> +        iov_from_buf(iov, iov_cnt, 0, &hdr, sizeof hdr);
>      }
>  
> -    /* We only ever receive a struct virtio_net_hdr from the tapfd,
> -     * but we may be passing along a larger header to the guest.
> -     */
> -    iov[0].iov_base += hdr_len;
> -    iov[0].iov_len  -= hdr_len;
> -
>      return offset;
>  }
>  
> @@ -598,7 +594,8 @@ static ssize_t virtio_net_receive(NetClientState *nc, 
> const uint8_t *buf, size_t
>  {
>      VirtIONet *n = DO_UPCAST(NICState, nc, nc)->opaque;
>      struct virtio_net_hdr_mrg_rxbuf *mhdr = NULL;
> -    size_t offset, i;
> +    const struct iovec *sg = elem.in_sg;
> +    size_t offset, i, guest_offset;
>  
>      if (!virtio_net_can_receive(&n->nic->nc))
>          return -1;
> @@ -615,7 +612,7 @@ static ssize_t virtio_net_receive(NetClientState *nc, 
> const uint8_t *buf, size_t
>      while (offset < size) {
>          VirtQueueElement elem;
>          int len, total;
> -        struct iovec sg[VIRTQUEUE_MAX_SIZE];
> +        const struct iovec *sg = elem.in_sg;
>  
>          total = 0;
>  
> @@ -640,20 +637,20 @@ static ssize_t virtio_net_receive(NetClientState *nc, 
> const uint8_t *buf, size_t
>              exit(1);
>          }
>  
> -        memcpy(&sg, &elem.in_sg[0], sizeof(sg[0]) * elem.in_num);
> -
>          if (i == 0) {
>              if (n->mergeable_rx_bufs)
>                  mhdr = (struct virtio_net_hdr_mrg_rxbuf *)sg[0].iov_base;
>  
>              offset += receive_header(n, sg, elem.in_num,
> -                                     buf + offset, size - offset,
> -                                     n->guest_hdr_len);
> +                                     buf + offset, size - offset);
>              total += n->guest_hdr_len;
> +            guest_offset = n->guest_hdr_len;
> +        } else {
> +            guest_offset = 0;
>          }
>  
>          /* copy in packet.  ugh */
> -        len = iov_from_buf(sg, elem.in_num, 0,
> +        len = iov_from_buf(sg, elem.in_num, guest_offset,
>                             buf + offset, size - offset);
>          total += len;
>          offset += len;
> diff --git a/iov.h b/iov.h
> index f0daefa..6c9fdf1 100644
> --- a/iov.h
> +++ b/iov.h
> @@ -90,8 +90,8 @@ void iov_hexdump(const struct iovec *iov, const unsigned 
> int iov_cnt,
>  /*
>   * Partial copy of vector from iov to dst_iov (data is not copied).
>   * dst_iov overlaps iov at a specified offset.
> - * size of dst_iov is at most bytes. Actual size is returned.
> + * size of dst_iov is at most bytes. dst vector count is returned.
>   */
> -size_t iov_cpy(struct iovec *dst_iov, unsigned int dst_iov_cnt,
> -               const struct iovec *iov, unsigned int iov_cnt,
> -               size_t offset, size_t bytes);
> +unsigned iov_cpy(struct iovec *dst_iov, unsigned int dst_iov_cnt,
> +                 const struct iovec *iov, unsigned int iov_cnt,
> +                 size_t offset, size_t bytes);

I think this belongs in an earlier patch.

Regards,

Anthony Liguori

> -- 
> MST