Re: [Qemu-devel] [PATCH v2 6/8] pc: memhp: do not export nvdimm's memory via _CRS

[Igor Mammedov]

On Sat, 8 Oct 2016 15:42:41 +0800
Xiao Guangrong <address@hidden> wrote:

> On 10/03/2016 09:21 PM, Igor Mammedov wrote:
> > On Fri, 12 Aug 2016 14:54:08 +0800
> > Xiao Guangrong <address@hidden> wrote:
> >  
> >> nvdimm's memory info can not exported via _CRS, instead, it is reported
> >> by NFIT/FIT
> >>
> >> This patch let _CRS return zero for both memory address and memory size
> >> if it is a nvdimm device inserted to the slot  
> > I'm not sure if it's right thing to do.
> > As it's relatively new, spec isn't clear about NVDIMM hotplug
> > process and how it's related to PNP0C80 memory devices.
> >
> > The thing is that notify to PNP0C80 will trigger regular
> > memory hotplug which would expect a valid _CRS and
> > I won't even try to predict reaction of different guests
> > on such behavior.
> >
> > So far exposing NFIT was sufficient for guest to work with
> > NVDIMMs at startup. So I'd assume NVDIMM_ROOT._FIT() would
> > provide sufficient info to hot-plug NVDIMM and allow
> > NVDIMM driver to handle it.
> >
> > The only case of using PNP0C80 with NVDIMM, I'd imagine, is
> > when one would like to expose NVDIMM as conventional RAM and
> > make guest OS use it as such, which is probably not what you'd
> > intended do here.
> >
> > Question is should/could NVDIMM hotplug work without using
> > PNP0C80?  
> 
> Some confusion i got when i was reading ACPI spec. I thing you
> are right that notifying memory device (PNP0C80) is needed only
> if the nvdimm is trying to export conventional RAM.
> 
> >
> > One could reuse/share memhotplug GPE._E03 to notify NVDIMM_ROOT
> > but even that is not necessary as NVDIMM has its own QEMU<->guest
> > interface and could just take the next free _E04 handler.
> > So I'd suggest to redo this and 7/8 patches to implement
> > independent (of memhotplug) NVDIMM hotplug as a starting point.  
> 
> Based on ACPI Spec, the memory device (PNP0C80) and nvdimm device
> is notified by the same interruption:
> 
> Scope (\_GPE)
> {
> Method (_L00) {
> Notify (\_SB.NVDR, 0x80) // Notify to NVDIMM root device
> Notify (\_SB.MEM0, 1) // Device Check to Memory Module
> }
> }
> 
> So, it is better to reuse _E03 and only notify NVDIMM device if the
> hotplug event is triggered on a slot with nvdimm device attached?
It's only an example in spec but there isn't requirement to do it that way.
So I'd use a separate handler for now (we can always chain calls in future
is need arises or merge them together)