Re: [Qemu-ppc] [PATCH v3 22/22] pc: support for virtio-pmem

[David Hildenbrand]

On 21/09/2018 18:55, Dr. David Alan Gilbert wrote:
> * David Hildenbrand (address@hidden) wrote:
>> On 21/09/2018 18:00, Dr. David Alan Gilbert wrote:
>>> * David Hildenbrand (address@hidden) wrote:
>>>> Compile virtio-pmem for x86 and properly hotplug virtio-pmem devices
>>>> (memory-device part) from our pc machine hotplug handler.
>>>>
>>>> Signed-off-by: David Hildenbrand <address@hidden>
>>>
>>> I see a few other places in the pc subdir that have checks for
>>> TYPE_NVDIMM; can you just explain why they are untouched:
>>>   i386/pc.c: pc_memory_unplug_request
>>
>> I think I had an answer why that isn't needed, but for unplug we might
>> still be missing something.
>>
>> In general: unplug requests are always directed at the proxy device (via
>> qdev_unplug()) (e.g. virtio-pmem-pci, see below). virtio-pmem is a child
>> device of virtio-pmem-pci. Once virtio-pmem-pci is finally unrealized,
>> the virtio-mem device will be unrealized.
>>
>> Now, there are no hotplug handlers getting called as far as I can see
>> for that child device. there would have to be an proper call to a
>> hotplug handler somewhere when setting the device to unrealized. (and
>> that's the place where a unplug requests does not make really sense, but
>> only a straight unplug - because we are half way through removing the
>> hierarchy of devices - one of the reason why also failures of unrealize
>> are usually ignored).
>>
>> I'll look into the details. Thanks for pointing that out, this stuff
>> always messes with my head.
>>
>>>   acpi/ich9.c: ich9_pm_device_plug_cb
>>>   acpi/piix4.c:piix4_device_plug_cb
>>
>> virtio-pmem is not an acpi device, that's why these don't apply.
>>
>> E.g. if virtio-pmem-pci is plugged, it will most probably trigger these
>> hotplug handlers for the proxy virtio pci device. The proxy device
>> creates and realized the virtio-mem (memory-device) - without any
>> hotplug handlers involved. That#s the point where we can jump in and
>> realize the device via the machine instead.
> 
> Does not being an acpi-device make life easier or harder?

My opinion is: easier.

Being an acpi device, we would have devices that are both, ACPI and
virtio. Or you have to create separate device and somehow link them
together (both in QEMU and in Linux)

I don't like gluing virtio devices to different, unrelated technologies.
Especially, you could never support architectures that don't support
ACPI. (with virtio-pmem, we could eventually even support architectures
that don't support something like nvdimm or acpi).

> Again, not
> knowing anything about the pmem acpi; if it looked like a normal nvdimm
> through ACPI would OSs be able to boot off it?

As far as I remember, booting from NVDIMM is not implemented anywhere. I
guess if it would be, we could also realize booting from virtio-pmem
(e.g. via grub extension).

> 
> Dave


-- 

Thanks,

David / dhildenb