[PATCH v2 0/2] Fix minimum frame size checks in vhost-user paths
In the vhost-user code paths, buffers from the virtio queue include a virtio net header prepended to the Ethernet frame. The minimum frame size checks (ETH_ZLEN, i.e. 60 bytes per IEEE 802.3) must account for this extra header, otherwise the size requested from vu_collect() and validated in ASSERT() is too small. v2: - PATCH 1/2 is merged, rebase the series on top of master - Add a patch to remove VNET_HLEN from l2len in vu_pad() callers - Remove assert on ETH_ZLEN in tcp_vu_prepare() as the size is trimmed to the actual data size Laurent Vivier (2): tcp_vu: vu_pad() expects l2 length tcp_vu, udp_vu: Account for virtio net header in minimum frame size tcp_vu.c | 17 ++++++++++------- udp_vu.c | 2 +- 2 files changed, 11 insertions(+), 8 deletions(-) -- 2.53.0
tcp_vu_hdrlen() returns a length that includes VNET_HLEN (the virtio
net header), but vu_pad() expects the Layer-2 frame length, which
should not include the virtio header. Passing the inflated length
means short frames aren't padded to the minimum 60-byte Ethernet
frame size (ETH_ZLEN).
Subtract VNET_HLEN from hdrlen when computing the l2 length passed
to vu_pad() in both tcp_vu_send_flag() and tcp_vu_data_from_sock().
Fixes: 0cb8f9003654 ("tcp, udp: Pad batched frames for vhost-user modes to 60 bytes (802.3 minimum)")
Signed-off-by: Laurent Vivier
In the vhost-user paths, the buffers provided by the virtio queue
include the virtio net header (VNET_HLEN) prepended to the Ethernet
frame. The minimum size checks using ETH_ZLEN must therefore account
for this additional header length, otherwise we underestimate the
minimum buffer size needed.
Use ETH_ZLEN + VNET_HLEN instead of bare ETH_ZLEN in vu_collect()
calls and the corresponding ASSERT() checks.
In tcp_vu_prepare(), revert the ASSERT to just check for hdrlen,
because at that point iov[0].iov_len has been trimmed to the actual
received data size plus headers.
Fixes: 0cb8f9003654 ("tcp, udp: Pad batched frames for vhost-user modes to 60 bytes (802.3 minimum)")
Signed-off-by: Laurent Vivier
On Mon, 23 Feb 2026 15:10:26 +0100
Laurent Vivier
tcp_vu_hdrlen() returns a length that includes VNET_HLEN (the virtio net header),
No need for this patch/series but: /** * tcp_vu_hdrlen() - return the size of the header in level 2 frame (TCP) * @v6: Set for IPv6 packet * * Return: return the size of the header */ ...can you fix that at some point? :) It's really not clear what it returns. I assumed "level 2" meant Layer-2 length (802.3-like), without virtio-net stuff, hence this issue.
but vu_pad() expects the Layer-2 frame length, which should not include the virtio header. Passing the inflated length means short frames aren't padded to the minimum 60-byte Ethernet frame size (ETH_ZLEN).
Subtract VNET_HLEN from hdrlen when computing the l2 length passed to vu_pad() in both tcp_vu_send_flag() and tcp_vu_data_from_sock().
Fixes: 0cb8f9003654 ("tcp, udp: Pad batched frames for vhost-user modes to 60 bytes (802.3 minimum)") Signed-off-by: Laurent Vivier
--- tcp_vu.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/tcp_vu.c b/tcp_vu.c index 3847a4e92087..2c9ddba0c9a4 100644 --- a/tcp_vu.c +++ b/tcp_vu.c @@ -71,8 +71,8 @@ int tcp_vu_send_flag(const struct ctx *c, struct tcp_tap_conn *conn, int flags) { struct vu_dev *vdev = c->vdev; struct vu_virtq *vq = &vdev->vq[VHOST_USER_RX_QUEUE]; - size_t optlen, hdrlen; struct vu_virtq_element flags_elem[2]; + size_t optlen, hdrlen, l2len; struct ipv6hdr *ip6h = NULL; struct iphdr *ip4h = NULL; struct iovec flags_iov[2]; @@ -137,7 +137,8 @@ int tcp_vu_send_flag(const struct ctx *c, struct tcp_tap_conn *conn, int flags) tcp_fill_headers(c, conn, eh, ip4h, ip6h, th, &payload, NULL, seq, !*c->pcap);
- vu_pad(&flags_elem[0].in_sg[0], hdrlen + optlen); + l2len = optlen + hdrlen - VNET_HLEN; + vu_pad(&flags_elem[0].in_sg[0], l2len);
if (*c->pcap) pcap_iov(&flags_elem[0].in_sg[0], 1, VNET_HLEN); @@ -446,6 +447,7 @@ int tcp_vu_data_from_sock(const struct ctx *c, struct tcp_tap_conn *conn) int buf_cnt = head[i + 1] - head[i]; ssize_t dlen = iov_size(iov, buf_cnt) - hdrlen; bool push = i == head_cnt - 1; + size_t l2len;
vu_set_vnethdr(vdev, iov->iov_base, buf_cnt);
@@ -457,7 +459,8 @@ int tcp_vu_data_from_sock(const struct ctx *c, struct tcp_tap_conn *conn) tcp_vu_prepare(c, conn, iov, buf_cnt, &check, !*c->pcap, push);
/* Pad first/single buffer only, it's at least ETH_ZLEN long */ - vu_pad(iov, dlen + hdrlen); + l2len = dlen + hdrlen - VNET_HLEN; + vu_pad(iov, l2len);
if (*c->pcap) pcap_iov(iov, buf_cnt, VNET_HLEN);
-- Stefano
On Mon, 23 Feb 2026 15:10:25 +0100
Laurent Vivier
In the vhost-user code paths, buffers from the virtio queue include a virtio net header prepended to the Ethernet frame. The minimum frame size checks (ETH_ZLEN, i.e. 60 bytes per IEEE 802.3) must account for this extra header, otherwise the size requested from vu_collect() and validated in ASSERT() is too small.
Applied. Yes, it fixes the issue I reported previously in the TCP tests. -- Stefano
participants (2)
-
Laurent Vivier
-
Stefano Brivio