Re: [PATCH v2 3/3] tun: Set num_buffers for virtio 1.0
From: Akihiko Odaki
Date: Wed Jan 15 2025 - 00:08:05 EST
On 2025/01/13 12:04, Jason Wang wrote:
On Fri, Jan 10, 2025 at 7:12 PM Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:
On 2025/01/10 19:23, Michael S. Tsirkin wrote:
On Fri, Jan 10, 2025 at 11:27:13AM +0800, Jason Wang wrote:
On Thu, Jan 9, 2025 at 2:59 PM Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:
The specification says the device MUST set num_buffers to 1 if
VIRTIO_NET_F_MRG_RXBUF has not been negotiated.
Have we agreed on how to fix the spec or not?
As I replied in the spec patch, if we just remove this "MUST", it
looks like we are all fine?
Thanks
We should replace MUST with SHOULD but it is not all fine,
ignoring SHOULD is a quality of implementation issue.
So is this something that the driver should notice?
Should we really replace it? It would mean that a driver conformant with
the current specification may not be compatible with a device conformant
with the future specification.
I don't get this. We are talking about devices and we want to relax so
it should compatibile.
The problem is:
1) On the device side, the num_buffers can be left uninitialized due to bugs
2) On the driver side, the specification allows assuming the num_buffers
is set to one.
Relaxing the device requirement will replace "due to bugs" with
"according to the specification" in 1). It still contradicts with 2) so
does not fix compatibility.
Instead, we should make the driver requirement stricter to change 2).
That is what "[PATCH v3] virtio-net: Ignore num_buffers when unused" does:
https://lore.kernel.org/r/20250110-reserved-v3-1-2ade0a5d2090@xxxxxxxxxx
We are going to fix all implementations known to buggy (QEMU and Linux)
anyway so I think it's just fine to leave that part of specification as is.
I don't think we can fix it all.
It essentially only requires storing 16 bits. There are details we need
to work out, but it should be possible to fix.
Regards,
Akihiko Odaki