Re: [PATCH net-next 2/2] vhost-net: reduce one userspace copy when building XDP buff

On Fri, Jun 13, 2025 at 10:16 AM Willem de Bruijn
[off-list ref] wrote:

Jason Wang wrote:

quoted

We used to do twice copy_from_iter() to copy virtio-net and packet
separately. This introduce overheads for userspace access hardening as
well as SMAP (for x86 it's stac/clac). So this patch tries to use one
copy_from_iter() to copy them once and move the virtio-net header
afterwards to reduce overheads.

Testpmd + vhost_net shows 10% improvement from 5.45Mpps to 6.0Mpps.

Signed-off-by: Jason Wang <redacted>

Acked-by: Willem de Bruijn <willemb@google.com>

quoted

---
 drivers/vhost/net.c | 13 ++++---------
 1 file changed, 4 insertions(+), 9 deletions(-)

diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
index 777eb6193985..2845e0a473ea 100644
--- a/drivers/vhost/net.c
+++ b/drivers/vhost/net.c

@@ -690,13 +690,13 @@ static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
      if (unlikely(!buf))
              return -ENOMEM;

-     copied = copy_from_iter(buf, sock_hlen, from);
-     if (copied != sock_hlen) {
+     copied = copy_from_iter(buf + pad - sock_hlen, len, from);
+     if (copied != len) {
              ret = -EFAULT;
              goto err;
      }

-     gso = buf;
+     gso = buf + pad - sock_hlen;

      if (!sock_hlen)
              memset(buf, 0, pad);

@@ -715,12 +715,7 @@ static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
              }
      }

-     len -= sock_hlen;
-     copied = copy_from_iter(buf + pad, len, from);
-     if (copied != len) {
-             ret = -EFAULT;
-             goto err;
-     }
+     memcpy(buf, buf + pad - sock_hlen, sock_hlen);

It's not trivial to see that the dst and src do not overlap, and does
does not need memmove.

Minimal pad that I can find is 32B and and maximal sock_hlen is 12B.

So this is safe. But not obviously so. Unfortunately, these offsets
are not all known at compile time, so a BUILD_BUG_ON is not possible.

We had this:

int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
int sock_hlen = nvq->sock_hlen;

So pad - sock_len is guaranteed to be greater than zero.

If this is not obvious, I can add a comment in the next version.

Thanks

quoted

      xdp_init_buff(xdp, buflen, NULL);
      xdp_prepare_buff(xdp, buf, pad, len, true);
--
2.34.1