On Fri, 30 Nov 2018 at 16:14, Andrew Jones <drjo...@redhat.com> wrote: > > On Fri, Nov 30, 2018 at 03:57:13PM +0000, Peter Maydell wrote: > > On Fri, 30 Nov 2018 at 15:56, Peter Maydell <peter.mayd...@linaro.org> > > wrote: > > > I suspect the reason you've done this is that you're > > > trying to get "-device pvpanic" to work on the command > > > line. I recommend not trying to get that to work. > > > MMIO devices are not pluggable devices, and -device > > > is not expected to work with them. > > > > If you do want a pluggable pvpanic device for the virt > > board then you should implement it as a PCI device, > > incidentally. > > > > We'd have to allocate it a PCI device ID, but I guess that's OK as > there are plenty of IDs left for 1b36. I'm not sure it's worth it > though. Phil asked that this device by user creatable, but maybe > that's not necessary. Maybe we just need a mach-virt compat boolean > and then to always provide this device for 4.0 and later machines?
Yes, if it's just an mmio device then we should either: * default to providing it, with the usual flag to say "don't create in older virt board versions", and also a machine property to disable it manually (like we do with the ITS) * default to not providing it at all, and have a machine property to enable it (like we do with the IOMMU) Which is all doable, but every time we do that it makes the virt board code that extra bit more complicated (we have half a dozen machine properties on it already). This kind of thing is why a PCI device is cleaner -- it just works on any machine with a PCI controller, it by default is something that the user can provide if they want to and not if they don't, and it's not a custom UI that management layers might need to special-case. The guest does need to do a bit of PCI probing and setup initially, but then it can just leave the MMIO BAR permanently mapped and write to that address when the guest panics. thanks -- PMM