Re: [PATCH linux-next v3 2/6] vdpa: Introduce query of device config layout | virtualization

quoted

在 2021/6/25 下午2:45, Parav Pandit 写道:

From: Jason Wang <jasowang@redhat.com>
Sent: Friday, June 25, 2021 8:59 AM

在 2021/6/24 下午3:59, Parav Pandit 写道:
From: Jason Wang <jasowang@redhat.com>
Sent: Thursday, June 24, 2021 12:35 PM

Consider we had a mature set of virtio specific uAPI for config space.
It would be a burden if we need an unnecessary translation layer of
netlink in the middle:

[vDPA parent (virtio_net_config)] <-> [netlink
(VDPA_ATTR_DEV_NET_XX)] <-> [userspace
(VDPA_ATTR_DEV_NET_XX)]
<-> [ user (virtio_net_config)]
This translation is not there. We show relevant net config fields as
VDPA_ATTR_DEV_NET individually.
It is not a binary dump which is harder for users to parse and make
any use
of it.

The is done implicitly, user needs to understand the semantic of
virtio_net_config and map the individual fields to the vdpa tool sub-
command.
Mostly not virtio_net_config is for the producer and consumer sw entities.
Here user doesn't know about such layout and where its located.
User only sets config params that gets set in the config space.
(without understanding what is config layout and its location).

It is only one level of translation from virtio_net_config (kernel)
-> netlink
vdpa fields.
It is similar to 'struct netdevice' -> rtnl info fields.
I think not, the problem is that the netdevice is not a part of uAPI
but virtio_net_config is.
Virtio_net_config is a UAPI for sw consumption.
That way yes, netlink can also do it, however it requires side channel
communicate what is valid.
If we make netlink simply a transport, it would be much easier. And
we
had
the chance to unify the logic of build_config() and set_config() in
the
driver.
How? We need bit mask to tell that out of 21 fields which fields to
update
and which not.
And that is further mixed with offset and length.
So set_config() could be called from userspace, so did build_config().
The only difference is:

1) they're using different transport, ioctl vs netlink
2) build_config() is only expected to be called by the management
tool

If qemu works well via set_config ioctl, netlink should work as well.
mlx5 set_config is noop.
vdpa_set_config() need to return an error code. I don't vp_vdpa.c
blindly writes the config as its passthrough.
Parsing which fields to write and which not, using offset and length is a
messy code with typecast and compare old values etc.

I don't see why it needs typecast, virtio_net_config is also uABI, you can
deference the fields directly.
User wants set only the mac address of the config space. How do user space tell this?

Good question, but we need first answer:

"Do we allow userspace space to modify one specific field of all the 
config?"

Pass the whole virtio_net_config and inform via side channel?

That could be a method.

Or vendor driver is expected to compare what fields changed from old config space?

So I think we need solve them all, but netlink is probably the wrong 
layer, we need to solve them at virtio level and let netlink a transport 
for them virtio uAPI/ABI.

And we need to figure out if we want to allow the userspace to modify 
the config after the device is created. If not, simply build the 
virtio_net_config and pass it to the vDPA parent during device creation. 
If not, invent new uAPI at virtio level to passing the config fields. 
Virtio or vDPA core can provide the library to compare the difference.

My feeling is that, if we restrict to only support build the config 
during the creation, it would simply a lot of things. And I didn't 
notice a use case that we need to change the config fields in the middle 
via the management API/tool.

Btw, what happens if management tool tries to modify the mac of vDPA
when the device is already used by the driver?
At present it allows modifying, but it should be improved in future to fail if
device is in use.

This is something we need to fix I think. Or if it's really useful to
allowing the attributes to be modified after the device is created.

Why not simply allow the config to be built only at device creation?
That avoids the problem of modifying fields after bind to vhost.
But UAPI issue still remains so lets resolve that first.

And actually, it's not the binary blob since uapi clearly define the
format (e.g struct virtio_net_config), can we find a way to use that?
E.g introduce device/net specific command and passing the blob with
length and negotiated features.
Length may change in future, mostly expand. And parsing based on
length
is not such a clean way.

Length is only for legal checking. The config is self contained with:
Unlikely. When structure size increases later, the parsing will change
based
on the length.
Because older kernel would return shorter length with older iproute2
tool.
This is fine since the older kernel only support less features. The only
possible issue if the old iproute 2 runs on new kernel. With the current
proposal, it may cause some config fields can't not be showed.
Not showing is ok.
But the code is messy to typecast on size.

I think it might be useful to introduce a command to simply dump the
config space.

So user space always have to deal and have nasty parsing/typecasting
based on the length.
Such nasty parsing is not required for netlink interface.

That's how userspace (Qemu) is expected to work now. The userspace
should determine the semantic of the fields based on the features.

Differentiate config fields doesn't help much, e.g userspace still need
to differ LINK_UP and ANNOUNCE for the status field.
Yes, this parsing is from constant size u16 status.
[..]

Its not about performance. By the time 1st call is made, features got
updated and it is out of sync with config.
1) get config
2) get device id
3) get features
This requires using features from 3rd netlink output to decode output of
1st netlink output.
Which is a bit odd of netlink.
Other netlink nla_put() probably sending whole structure doesn’t need
to
do it.

Well, we can pack them all into a single skb isn't it? (probably with a
config len).
You want to pack features and config both in the single nla_put()?
If so, it isn't necessary. There are more examples in kernel that adds
individual fields instead of nla_put(blob).
I wouldn’t follow those nla_put() callers.
No, a single skb not single nla_put().

Actually git grep told me a very good example of carrying uABI via
netlink, that is the ndt_config:

1) we had ndt_config definition in the uAPI
2) netlink simply carries the structure in neightbl_fill_info():

                  if (nla_put(skb, NDTA_CONFIG, sizeof(ndc), &ndc))
Sure. But the reverse path doesn’t have this that requires side band mask.
My concern is not for existing virtio_net_config layout, but the future increase of it requires size based typecasting on both directions.

For virito_net_config, why not simply:

len = ops->get_config_len();
config = kmalloc(len, GFP_KERNEL);
ops->get_config(vdev, 0, config, len);
nla_put(skb, VIRTIO_CONFIG, config, len);
User space need to parse content based on this length as it can change in future.
Length telling how to typecast is want I want to avoid here.

So there's no real difference, using xxx_is_valid, is just a implicit 
length checking as what is done via config_len:

if (a_is_valid) {
     /* dump a */
} else if (b_is_valid) {
     /* dump b */
}

vs.

if (length < offsetof(struct virtio_net_config, next field of a)) {
     /* dump a*/
}

Actually, Qemu has solved the similar issues via the uAPI:

https://git.qemu.org/?p=qemu.git;a=blob;f=hw/net/virtio-net.c;h=bd7958b9f0eed2705e0d6a2feaeaefb5e63bd6a4;hb=HEAD#l92

If the current uAPI is not sufficient, let's tweak it.

nla_put_le64(skb, VIRTIO_FETURES, features);

For build_config, we can simply do thing reversely. Then everything
works via the existing virtio uAPI/ABI.
In reverse path how do you tell which fields of the config space to set and which to ignore?

See my above reply.

Shall we use u64 features for it?
Will type of device able to describe their config space via a feature bit?

I think not. They're a lot of fields can not be deduced from the 
features (mtu, queue paris, mac etc).

But I agree the the config fields can not work without the feature bits.

For build config, it's only one

1) build config

I prefer to follow rest of the kernel style to return self contained
invidividual fields.

But I saw a lot of kernel codes choose to use e.g nla_put() directly with
module specific structure.
It might be self-contained structure that probably has not found the
need
to expand.

I think it's just a matter of putting the config length with the config
data. Note that we've already had .get_config_size() ops for validating
inputs through VHOST_SET_CONFIG/VHOST_GET_CONFIG.
This length comes as part of the netlink interface already, no need for extra
length.
The whole point is to avoid parsing based on length.
Well, it doesn't do anything difference compared to xxx_is_valid which
just calculating the offset implicitly (via the compiler).

We cannot change the virtio_net_config UAPI in use, but netlink code
doesn’t need to be bound to size based typecasting and compare fields
during build_config().

The points are:

1) Avoid duplicating the existing uAPIs
2) Avoid unnecessary parsing in the netlink, netlink is just the
transport, it's the charge of the vDPA parent to do that
All those parsing will move to vendor drivers to validate offset/length to update only specific fields of config space.

Or the vDPA or virtio core can provide helpers to compare the difference 
if it's necessary.

Thanks

It is a transport to carry fields which is what we are using for.
I agree there that these config fields are exposed individually in both directions to keep it safe from structure layout increments.

Thanks
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

`h`	back out one level
`j`	next message in thread
`k`	previous message in thread
`l`	drill in
`Esc`	close help / fold thread tree
`?`	toggle this help