Re: [PATCH v6 1/3] udp_vu: Allow virtqueue elements with multiple iovec entries

[Laurent Vivier <lvivier@redhat.com>]

On 4/3/26 13:53, Stefano Brivio wrote:
> On Wed,  1 Apr 2026 21:23:24 +0200
> Laurent Vivier <lvivier@redhat.com> wrote:
> 
>> The previous code assumed a 1:1 mapping between virtqueue elements and
>> iovec entries (enforced by an assert).  Drop that assumption to allow
>> elements that span multiple iovecs: track elem_used separately by
>> walking the element list against the iov count returned after padding.
>> This also fixes vu_queue_rewind() and vu_flush() to use the element
>> count rather than the iov count.
>>
>> Use iov_tail_clone() in udp_vu_sock_recv() to handle header offset,
>> replacing the manual base/len adjustment and restore pattern.
>>
>> Signed-off-by: Laurent Vivier <lvivier@redhat.com>
>> ---
>>   udp_vu.c | 29 ++++++++++++++---------------
>>   1 file changed, 14 insertions(+), 15 deletions(-)
>>
>> diff --git a/udp_vu.c b/udp_vu.c
>> index 30af64034516..5608a3a96ff5 100644
>> --- a/udp_vu.c
>> +++ b/udp_vu.c
>> @@ -64,30 +64,25 @@ static size_t udp_vu_hdrlen(bool v6)
>>    */
>>   static ssize_t udp_vu_sock_recv(struct iovec *iov, size_t *cnt, int s, bool v6)
>>   {
>> +	struct iovec msg_iov[*cnt];
> 
> Variable-length Arrays (VLAs) are allowed starting from C99 but we
> should really really avoid them. If 'cnt' is big enough, we risk
> writing all over the place. That's the main reason why they were more
> or less banned from the Linux kernel some years ago and eventually
> eradicated:

I can use alloca() if you prefer ;)

> 
>    https://lore.kernel.org/lkml/CA+55aFzCG-zNmZwX4A2FQpadafLfEzK6CC=qPXydAacU1RqZWA@mail.gmail.com/
> 
>    https://lore.kernel.org/lkml/20181028172401.GA41102@beast/
> 
> Can we use VIRTQUEUE_MAX_SIZE as upper bound like udp_vu_sock_to_tap()
> does?

Yes, but the idea here is we have always *cnt < VIRTQUEUE_MAX_SIZE
(because the value comes from vu_collect() and in vu_collect():
*in_sg (&iov_cnt or *cnt) < max_in_sg (ARRAY_SIZE(iov_vu) or VIRTQUEUE_MAX_SIZE or 1024)

And vu_collect(), in this case, sets generally *in_sg to a value lower than 44:
we want to create a frame of ETH_MAX_MTU by coalescing kernel buffers of size ETH_FRAME_LEN)

For me 16 kB on the stack is a lot of memory (but I started programming on a 48 kB RAM 
computer...).

Thanks,
Laurent