On Tue, Dec 04, 2018 at 09:40:07AM +0000, Peter Maydell wrote: > On Tue, 4 Dec 2018 at 00:41, <peng.h...@zte.com.cn> wrote: > > > > >I would still prefer to see a more detailed examination of whether > > >we can do this with a PCI device before we commit to taking the > > >MMIO version into the virt board. > > > > I'm sorry I thought I had sent an email. yesterday when I wrote an email to > > explain the reason, I was interrupted and forgot to send it out. > > > > Now the pvpanic device is implemented as a mmio device or an ACPI device in > > the kernel, > > and only one device can be seen at the same time. If the kernel parses FDT > > first, then pvpanic > > is a mmio device. The kernel parses ACPI table first(and virtual machine is > > configured with ACPI), > > and pvpanic is an ACPI device. > > If pvpanic is implemented as a PCI device, then the PCI device must still > > be seen when the ACPI table > > is first parsed by the kernel, because ACPI device relies on the mmio space > > of the PCI device. > > Mmio devices can be thought of as just an address space rather than a > > device in the strict sense. > > I'm afraid I don't understand. If it's a PCI device then > it does not need to be listed in the device tree or the > ACPI tables at all, because it is probeable by the guest. > This also significantly simplifies the changes needed in QEMU. > > > Secondly, I don't want it to be a pluggable device. If the user > > deletes the device by mistake, it may lead to unpredictable results. > > If the user deletes the PCI device they're using for their > disk or networking this will also lead to unpredictable > results. We expect users not to randomly unplug things from > their system if they want it to continue to work. In any > case your guest driver can easily handle the unplug: the > guest would then just lose the ability to notify on panic, > falling back to as if the pvpanic device had never been > present. >
To muddy the waters a bit more, while I'm not opposed to this device being a PCI device, there is a chance that someone will still want a platform-mmio version as well. I'm not sure how everything will eventually fall into place, but I've seen some super minimal guest configs proposed for the VMs-used-like-containers use cases, even configs that choose to use virtio-mmio over virtio-pci, and then not provide a PCI bus at all to the vm. Maybe this series and the current kernel series can be allowed to continue as they are, and if later there's a demand for a pci version, it could just be yet another variant added later? Thanks, drew
