Re: [PATCH v2 13/13] vhost-user,udp: Use 2 iovec entries per element

[Laurent Vivier <lvivier@redhat.com>]

On 3/12/26 10:47, David Gibson wrote:
> On Thu, Mar 12, 2026 at 09:08:33AM +0100, Laurent Vivier wrote:
>> On 3/12/26 05:39, David Gibson wrote:
>>> On Mon, Mar 09, 2026 at 10:47:44AM +0100, Laurent Vivier wrote:
>>>> iPXE places the vnet header in one virtqueue descriptor and the payload
>>>> in another. When passt maps these descriptors, it needs two iovecs per
>>>> virtqueue element to handle this layout.
>>>>
>>>> Without this, passt crashes with:
>>>>
>>>>     ASSERTION FAILED in virtqueue_map_desc (virtio.c:403): num_sg < max_num_sg
>>>>
>>>> Signed-off-by: Laurent Vivier <lvivier@redhat.com>
>>>> ---
>>>>    udp_vu.c    |  8 ++++----
>>>>    vu_common.c | 15 +++++++++++----
>>>>    2 files changed, 15 insertions(+), 8 deletions(-)
>>>>
>>>> diff --git a/udp_vu.c b/udp_vu.c
>>>> index cb5274aa1d81..47659b0402fd 100644
>>>> --- a/udp_vu.c
>>>> +++ b/udp_vu.c
>>>> @@ -34,7 +34,7 @@
>>>>    #include "vu_common.h"
>>>>    static struct iovec     iov_vu		[VIRTQUEUE_MAX_SIZE];
>>>> -static struct vu_virtq_element	elem		[VIRTQUEUE_MAX_SIZE];
>>>> +static struct vu_virtq_element	elem	[VIRTQUEUE_MAX_SIZE / IOV_PER_ELEM];
>>>
>>> The two level structure of the queues (array of elems, each pointing
>>> to an array of iovs) confuses me a bit.  I would have thought that
>>> VIRTQUEUE_MAX_SIZE represented the maximum number of elements, even if
>>> those elements had multiple iovs.  Can you enlighten me?
>>>
>> In all the code until now I supposed the number of iovec is equal to the
>> number of element because we had only one iovec per element.
>>
>> At virtio level, VIRTQUEUE_MAX_SIZE is the number of vring_desc in a
>> virtqueue, and a vring_desc is translated to an iovec in
>> virtqueue_map_desc() that are collected to an element in
>> vu_queue_map_desc(). So in passt, the maximum number of collected iovec will
>> be
>> VIRTQUEUE_MAX_SIZE and the maximum of element is VIRTQUEUE_MAX_SIZE if we
>> have only one iovec per element.
> 
> Ok... sorry, I'm still unclear.  Is the "element" a concept that's
> part of the "hardware" interface?  Or just part of the structure on
> the software side?

It's a structure on the software side to reflect that a virtqueue entry can be split in 
several descriptors (addr, size) on the hardware side.

A hardware side entry is mapped to a software side element, a hardware side descriptor is 
mapped to a software side iovec.

Thanks,
Laurent