Re: [virtio-dev] Re: [PATCH] vdpa/mlx5: set_features should allow reset to zero

From: Jason Wang
Date: Thu Mar 04 2021 - 03:26:37 EST



On 2021/3/3 4:29 下午, Cornelia Huck wrote:
On Wed, 3 Mar 2021 12:01:01 +0800
Jason Wang <jasowang@xxxxxxxxxx> wrote:

On 2021/3/2 8:08 下午, Cornelia Huck wrote:
On Mon, 1 Mar 2021 11:51:08 +0800
Jason Wang <jasowang@xxxxxxxxxx> wrote:
On 2021/3/1 5:25 上午, Michael S. Tsirkin wrote:
On Fri, Feb 26, 2021 at 04:19:16PM +0800, Jason Wang wrote:
On 2021/2/26 2:53 上午, Michael S. Tsirkin wrote:
Confused. What is wrong with the above? It never reads the
field unless the feature has been offered by device.
So the spec said:

"

The following driver-read-only field, max_virtqueue_pairs only exists if
VIRTIO_NET_F_MQ is set.

"

If I read this correctly, there will be no max_virtqueue_pairs field if the
VIRTIO_NET_F_MQ is not offered by device? If yes the offsetof() violates
what spec said.

Thanks
I think that's a misunderstanding. This text was never intended to
imply that field offsets change beased on feature bits.
We had this pain with legacy and we never wanted to go back there.

This merely implies that without VIRTIO_NET_F_MQ the field
should not be accessed. Exists in the sense "is accessible to driver".

Let's just clarify that in the spec, job done.
Ok, agree. That will make things more eaiser.
Yes, that makes much more sense.

What about adding the following to the "Basic Facilities of a Virtio
Device/Device Configuration Space" section of the spec:

"If an optional configuration field does not exist, the corresponding
space is still present, but reserved."

This became interesting after re-reading some of the qemu codes.

E.g in virtio-net.c we had:

*static VirtIOFeature feature_sizes[] = {
    {.flags = 1ULL << VIRTIO_NET_F_MAC,
     .end = endof(struct virtio_net_config, mac)},
    {.flags = 1ULL << VIRTIO_NET_F_STATUS,
     .end = endof(struct virtio_net_config, status)},
    {.flags = 1ULL << VIRTIO_NET_F_MQ,
     .end = endof(struct virtio_net_config, max_virtqueue_pairs)},
    {.flags = 1ULL << VIRTIO_NET_F_MTU,
     .end = endof(struct virtio_net_config, mtu)},
    {.flags = 1ULL << VIRTIO_NET_F_SPEED_DUPLEX,
     .end = endof(struct virtio_net_config, duplex)},
    {.flags = (1ULL << VIRTIO_NET_F_RSS) | (1ULL <<
VIRTIO_NET_F_HASH_REPORT),
     .end = endof(struct virtio_net_config, supported_hash_types)},
    {}
};*

*It has a implict dependency chain. E.g MTU doesn't presnet if
DUPLEX/RSS is not offered ...
*
But I think it covers everything up to the relevant field, no? So MTU
is included if we have the feature bit, even if we don't have
DUPLEX/RSS.

Given that a config space may be shorter (but must not collapse
non-existing fields), maybe a better wording would be:

"If an optional configuration field does not exist, the corresponding
space will still be present if it is not at the end of the
configuration space (i.e., further configuration fields exist.)


This should work but I think we need to define the end of configuration space first?


This
implies that a given field, if it exists, is always at the same offset
from the beginning of the configuration space."


(Do we need to specify what a device needs to do if the driver tries to
access a non-existing field? We cannot really make assumptions about
config space accesses; virtio-ccw can just copy a chunk of config space
that contains non-existing fields...

Right, it looks to me ccw doesn't depends on config len which is kind of
interesting. I think the answer depends on the implementation of both
transport and device:
(virtio-ccw is a bit odd, because channel I/O does not have the concept
of a config space and I needed to come up with something)


Ok.



Let's take virtio-net-pci as an example, it fills status unconditionally
in virtio_net_set_config() even if VIRTIO_NET_F_STATUS is not
negotiated. So with the above feature_sizes:

1) if one of the MQ, MTU, DUPLEX or RSS is implemented, driver can still
read status in this case
2) if none of the above four is negotied, driver can only read a 0xFF
(virtio_config_readb())


I guess the device could ignore
writes and return zeroes on read?)

So consider the above, it might be too late to fix/clarify that in the
spec on the expected behaviour of reading/writing non-exist fields.
We could still advise behaviour via SHOULD, though I'm not sure it adds
much at this point in time.

It really feels a bit odd that a driver can still read and write fields
for features that it did not negotiate, but I fear we're stuck with it.


Yes, since the device (at least virtio-pci) implment thing like this.

Thanks



Thanks


I've opened https://github.com/oasis-tcs/virtio-spec/issues/98 for the
spec issues.

---------------------------------------------------------------------
To unsubscribe, e-mail: virtio-dev-unsubscribe@xxxxxxxxxxxxxxxxxxxx
For additional commands, e-mail: virtio-dev-help@xxxxxxxxxxxxxxxxxxxx