On Mon, Aug 24, 2015 at 06:10:01PM +0200, Tom Gundersen wrote:

On Thu, Mar 27, 2014 at 9:58 PM, Michael S. Tsirkin [off-list ref] wrote:

quoted

On Thu, Mar 27, 2014 at 09:24:10PM +0100, Tom Gundersen wrote:

quoted

On Thu, Mar 27, 2014 at 8:25 PM, Kay Sievers [off-list ref] wrote:

quoted

On Thu, Mar 27, 2014 at 8:13 PM, Michael S. Tsirkin [off-list ref] wrote:

quoted

On Thu, Mar 27, 2014 at 08:06:15PM +0100, Kay Sievers wrote:

quoted

On Thu, Mar 27, 2014 at 7:59 PM, Michael S. Tsirkin [off-list ref] wrote:

quoted

On Thu, Mar 27, 2014 at 07:43:34PM +0100, Kay Sievers wrote:

quoted

On Thu, Mar 27, 2014 at 6:30 PM, Michael S. Tsirkin [off-list ref] wrote:

quoted

If the virtio device is a PCI device, it is really best to
treat it like you treat any other PCI function (I guess you mean
function and not device, right? We support multifunction
devices and some people do pack multiple NICs in a single device)..

At the moment many devices in a single pci function can not happen on a
PCI system (no multiport) but if we add multiport, we'll follow some
existing standard to expose this information to the guest.

This means, that there can currently never multiple devices below one
and the same virtio parent device?

quoted

There's a single virtio device per pci function (you keep saying
device but I hope the distinction is clear and this is
just slip of the tongue).

Right, we talks about sysfs directories and they are called "device",
we don't really care about the actual bus that is implemented,
userspace does not really know much about them. :)

quoted

For net devices under a pci function that is also currently the case,
but I can't yet tell you for sure ahead of the time how we'll present
multiport devices if we ever implement them.

I'm guessing there will be multiple net devices under
a single pci device and we'll present a sysfs attribute with the port
number in this case.

Hmm maybe we should go ahead and add a place-holder
attribute so that it's future-proof?

I'll write a patch like that and we'll see how it's accepted.

Netdevs with multiple ports are represented with the standard "dev_id"
attribute identifying the instance of the driver per parent "device";
should all work already from the userspace side, if the virtio side
would use that too.

Kay

Aha. In that case it's easy - pls assume that if and when we implement
multiple we'll just follow standards and use dev_id.
For virtio pci devices specifically virtio<->pci 1:1 mapping is set
in stone in the spec.

Non pci ones need to be examined separately.

Nice, thanks. That means we can just "jump over" the "virtio*" device
(device as in sysfs view), and let the pci parent let identify the
device.

Background: The logic for device naming generally refuses to touch
devices with unknown parent devices in the chain (the directories you
see for: ls -l /sys/class/net/), because the parents *could* offer
their own bus logic that exposes multiple devices below and we would
create clashing names for them. That's why the virt vs. non-virt case
needs to be handled explicitly (seems in this case by just skipping
it) and it is not necessarily by default the same behaviour.

So Kay and I discussed this a bit more, and found that we likely
cannot handle virtio nic's nicely. As we would name them based on the
pci geo, we rely on this being stable between reboots and when
adding/removing hardware. Is there some way to make this work with
virtio, or will the 'fake' pci busses simply be enumerated as they are
created?

Cheers,

Tom

I think it's standard PCI thing.
Some pci bridges have a slot id register - I'm assuming you
are using these if present?

If not you'll either need to rely on firmware enumerating buses
consistently, or use the mac of the NIC.

Hi Michael,

Sorry to resurrect such an old thread, but this suddenly became
relevant again and I see that we never really reached a conclusion.
Rereading the discussion it is still not really clear to me what the
API promise is, so let me try to rephrase what we need and hopefully
you can help me understand where we are at.

My understanding is that the enumeration of virtioX buses is global
rather than local to the direct parent device of a given bus. That
means that we cannot base deterministic naming of child devices
(functions) on virtio enumeration, as the names would then depend on
the presence or absence of unrelated virtio buses on the system.

We have two options that would allow us to get around this:

1) if there is a guarantee now and in the future (even if only
restricted to netdevs) that no virtio bus will have a direct sibling
bus (i.e., with the same parent device); or

I think this is the case. The virtio bus is an artifact.
There's always a single one behind each pci device.
Is this sufficient?

2) if the enumeration of virtio buses was extended with a "port"
numebr which was guaranteed to be made local to their direct parent
device and moreover be made deterministic. I.e., if there are no
virtio sibling buses (but there may be other unrelated virtio buses)
the name would always be virtioX.0, and if there are two buses with
the same parent their names would always be virtioX.0 and virtioX.1
and always in the same order.

Is either of those things something that is already guaranteed or
something that could be made working?

Cheers,

Tom