Re: [PATCH 2/2] tcp_vu, udp_vu: Account for virtio net header in minimum frame size

[Laurent Vivier <laurent@vivier.eu>]

On 2/15/26 11:56, Stefano Brivio wrote:
> On Thu, 12 Feb 2026 12:39:32 +0100
> Laurent Vivier <lvivier@redhat.com> wrote:
> 
>> 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.
>>
>> Fixes: 0cb8f9003654 ("tcp, udp: Pad batched frames for vhost-user modes to 60 bytes (802.3 minimum)")
>> Signed-off-by: Laurent Vivier <lvivier@redhat.com>
>> ---
>>   tcp_vu.c | 8 ++++----
>>   udp_vu.c | 2 +-
>>   2 files changed, 5 insertions(+), 5 deletions(-)
>>
>> diff --git a/tcp_vu.c b/tcp_vu.c
>> index f7bda4943e43..2d593d534d68 100644
>> --- a/tcp_vu.c
>> +++ b/tcp_vu.c
>> @@ -90,12 +90,12 @@ int tcp_vu_send_flag(const struct ctx *c, struct tcp_tap_conn *conn, int flags)
>>   	vu_set_element(&flags_elem[0], NULL, &flags_iov[0]);
>>   
>>   	elem_cnt = vu_collect(vdev, vq, &flags_elem[0], 1,
>> -			      MAX(hdrlen + sizeof(*opts), ETH_ZLEN), NULL);
>> +			      MAX(hdrlen + sizeof(*opts), ETH_ZLEN + VNET_HLEN), NULL);
>>   	if (elem_cnt != 1)
>>   		return -1;
>>   
>>   	ASSERT(flags_elem[0].in_sg[0].iov_len >=
>> -	       MAX(hdrlen + sizeof(*opts), ETH_ZLEN));
>> +	       MAX(hdrlen + sizeof(*opts), ETH_ZLEN + VNET_HLEN));
>>   
>>   	vu_set_vnethdr(vdev, flags_elem[0].in_sg[0].iov_base, 1);
>>   
>> @@ -208,7 +208,7 @@ static ssize_t tcp_vu_sock_recv(const struct ctx *c, struct vu_virtq *vq,
>>   
>>   		cnt = vu_collect(vdev, vq, &elem[elem_cnt],
>>   				 VIRTQUEUE_MAX_SIZE - elem_cnt,
>> -				 MAX(MIN(mss, fillsize) + hdrlen, ETH_ZLEN),
>> +				 MAX(MIN(mss, fillsize) + hdrlen, ETH_ZLEN + VNET_HLEN),
>>   				 &frame_size);
>>   		if (cnt == 0)
>>   			break;
>> @@ -302,7 +302,7 @@ static void tcp_vu_prepare(const struct ctx *c, struct tcp_tap_conn *conn,
>>   	/* we guess the first iovec provided by the guest can embed
>>   	 * all the headers needed by L2 frame, including any padding
>>   	 */
>> -	ASSERT(iov[0].iov_len >= MAX(hdrlen, ETH_ZLEN));
>> +	ASSERT(iov[0].iov_len >= MAX(hdrlen, ETH_ZLEN + VNET_HLEN));
> 
> This triggers in the passt_vu_in_ns/tcp, "TCP/IPv4: host to guest: big
> transfer" test case, that is, the first time we connect to the guest
> (probably on the initial SYN segment).
> 
> I didn't check why.

After tcp_vu_sock_recv(), iov[0].iov_len is set to the size of hdrlen + received data, 
it's why in tcp_vu_prepare() iov_len can be shorter than expected. As the actual buffer 
size is already guaranteed by vu_collect(), I think the ASSERT() should be reverted to 
ASSERT(iov[0].iov_len >= hdrlen).

Moreover there is a problem with vu_pad() in tcp_vu.c that includes the VNET_HLEN (it 
should not), see udp_vu.c

I'm fixing that and send a new version. I was not able to reproduce the problem so I rely 
on you to test it.

Thanks,
Laurent