virtio — archive

[PATCH RFC v7 0/8] Introduce device group and device management

From: Cornelia Huck <[email protected]>
To: [email protected]
Date: Tue, 30 Aug 2022 14:31:00 -0000

[finally got around to looking at this thread] On Tue, Aug 30 2022, Halil Pasic <[email protected]> wrote: > On Mon, 29 Aug 2022 14:28:05 -0400 > "Michael S. Tsirkin" <[email protected]> wrote: > >> On Thu, Aug 18, 2022 at 03:39:58PM +0200, Halil Pasic wrote: > [..] >> > Fair point! >> > >> > I would prefer a driver normative which goes like this: >> > >> > """ >> > A driver SHOULD NOT accept features (i.e. have code that would do so if >> > the feature is offered) if the feature is not supported by the driver >> > (e.g. because unsupported by the transport), even if the specification >> > implies that the device can not offer these features in the first place >> > (e.g. because the feature is not yet supported by the transport. >> > """ >> >> ok. why not MUST NOT? > > I'm fine with MUST NOT. Since this is a general statement (i.e. not > scoped to ADMIN_VQ) I felt like SHOULD NOT is a bit safer because > provided somebody is doing this wrong for some feature already, it > wouldn't render that implementation outright non-compliant. But I > believe this is just a theoretical possibility. I'm fine with MUST NOT. I'd assume that any driver that does this (accept a feature even if not supported by the transport) today is already broken, so yes, MUST NOT is the way to go here IMHO. > >> >> > And a similar device normative as well, which just that it may not offer >> > such features. >> > >> > """ >> > Note: The rationale behind the [reference to the normative] is that >> > while some features can not be implemented within the boundaries of the >> > current virtio specification, future incarnations of the specificaton may >> > make such implementations possible. A most prominent example is optional >> > features dependent on optional virtio facilities whose transport specific >> > implementation is not yet specified for some transports. Should one end >> > gain the ability to support these features, the old implementation which >> > made the assumption that the other end will make sure these features are >> > not negotiated would end up negotiating something it can't actually >> > support. >> > """ >> > >> > >> > >> > > >> > > So, Maybe just add text >> > > >> > > Note: future versions of this specification will allow setting ADMIN_VQ >> > > for driver and device. Device MUST NOT assume driver does not >> > > acknowledge ADMIN_VQ if offered. >> > >> > I would not lean out of the window and promise something with regards to >> > future versions of this spec. >> >> s/will/might/ > > With this change it works like a charm! Works for me as well. (Although I'd use definite articles with device and driver.) > >> >> > > >> > > And similarly for drivers: >> > > >> > > Note: future versions of this specification will allow setting ADMIN_VQ >> > > for driver and device. Drivers MUST NOT assume ADMIN_VQ if not offered. Same here. >> > > >> > >> > I think we can then make a note which references the generic normative >> > for each feature affected where it suits us. >> > >> > > > >> > > > If we want, we can also state what needs to be done in general when >> > > > features are unsupported by the transport. And yes, that normative >> > > > material in my opinion. >> > > > >> > > > Regards, >> > > > Halil >> > > >> > > >> > > Are there other examples? I want to call out the list explicitly because >> > > it is so easy to enable an extra feature by mistake. >> > > >> > >> > I don't think CCW supports the shared memory yet... But I may be wrong. You are right about this one. I think we never figured out an architecture that would work with a mix of virtio-ccw and virtio-pci devices... I think ring reset and notification data should also be on that list of non-supported features. Things like SR_IOV obviously don't make sense for ccw, so they will never be implemented.