Re: [PATCH v3 2/3] spapr: Add NVDIMM device support

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 8022 bytes --]

On Thu, Dec 12, 2019 at 02:22:56PM +0530, Shivaprasad G Bhat wrote:
> 
> 
> On 12/11/2019 01:35 PM, Igor Mammedov wrote:
> > On Wed, 11 Dec 2019 09:44:11 +0530
> > Shivaprasad G Bhat <sbhat@linux.ibm.com> wrote:
> > 
> > > On 12/06/2019 07:22 AM, David Gibson wrote:
> > > > On Wed, Nov 27, 2019 at 09:50:54AM +0530, Bharata B Rao wrote:
> > > > > On Fri, Nov 22, 2019 at 10:42 AM David Gibson
> > > > > <david@gibson.dropbear.id.au> wrote:
> > > > > > Ok.  A number of queries about this.
> > > > > > 
> > > > > > 1) The PAPR spec for ibm,dynamic-memory-v2 says that the first word in
> > > > > > each entry is the number of LMBs, but for NVDIMMs you use the
> > > > > > not-necessarily-equal scm_block_size instead.  Does the NVDIMM
> > > > > > amendment for PAPR really specify to use different block sizes for
> > > > > > these cases?  (In which case that's a really stupid spec decision, but
> > > > > > that wouldn't surprise me at this point).
> > > > > SCM block sizes can be different from LMB sizes, but here we enforce
> > > > > that SCM device size (excluding metadata) to multiple of LMB size so
> > > > > that we don't end up memory range that is not aligned to LMB size.
> > > > Right, but it still doesn't make sense to use scm_block_size when you
> > > > create the dynamic-memory-v2 property.
> > > Right, I should use LMB size here as I will be creating holes here to
> > > disallow DIMMs
> > > to claim those LMBs marking them INVALID as Bharata Suggested before.
> > > 
> > > >    As far as the thing
> > > > interpreting that goes, it *must* be LMB size, not SCM block size.  If
> > > > those are required to be the same at this point, you should use an
> > > > assert().
> > > SCM block size should be a multiple for LMB size, need not be equal.
> > > I'll add an assert
> > > for that, checking if equal. There is no benefit I see as of now having
> > > higher
> > > SCM block size as the bind/unbind are already done before the bind hcall.
> > > 
> > > > > > 2) Similarly, the ibm,dynamic-memory-v2 description says that the
> > > > > > memory block described by the entry has a whole batch of contiguous
> > > > > > DRCs starting at the DRC index given and continuing for #LMBs DRCs.
> > > > > > For NVDIMMs it appears that you just have one DRC for the whole
> > > > > > NVDIMM.  Is that right?
> > > > > One NVDIMM has one DRC, In our case, we need to mark the LMBs
> > > > > corresponding to that address range in ibm,dynamic-memory-v2 as
> > > > > reserved and invalid.
> > > > Ok, that fits very weirdly with the DRC allocation for the rest of
> > > > pluggable memory, but I suppose that's PAPR for you.
> > > > 
> > > > Having these in together is very inscrutable though, and relies on a
> > > > heap of non-obvious constraints about placement of DIMMs and NVDIMMs
> > > > relative to each other.  I really wonder if it would be better to have
> > > > a completely different address range for the NVDIMMs.
> > > The backend object for both DIMM and NVDIMM are memory-backend-*
> > > and they use the address from the same space. Separating it would mean
> > > using/introducing different backend object. I dont think we have a
> > > choice here.
> > What address-space(s) are are talking about here exactly?
> >  From my point of view memory-backend-* provides RAM block at
> > some HVA, which shouldn't not have anything to do with how NVDIMM
> > partitions and maps it to GPA.
> 
> Ah, you are right! I got confused with the HVA.
> 
> Nonetheless, I don't see a need for having vNVDIMM in different
> guest physical address range as the existing code has support for marking
> memory ranges distinctly for DIMM/NVDIMM.

The problem is that the way you create the dynamic-memory-v2 property
relies on knowing whether a GPA is DIMM or NVDIMM -but you can't in
the presence of hotplug.  Using the default address allocation, DIMMs
and NVDIMMs are dynamically assigned addresses in the hotplug memory
area.

So, if you have an NVDIMM plugged at boot time, you'll mark that range
of LMBs as invalid.  If you then unplug it, and instead plug in a
regular DIMM (before the guest has even tried to online the NVDIMM),
it will probably use the same GPA range.  You therefore need regular
LMB DRCs for that range, and we have no way to communicate that to the
guest.

Similar problems if you go the other way (unplug DIMM, plug NVDIMM).

> On another note, the x86 too does it the same way. There is no separate
> range defined there.

Yes, but AIUI, the way PC describes DIMMs and NVDIMMs in ACPI are
pretty similar.  In PAPR they are gratuitously different - we don't
even have the concept of a DIMM, only the individual LMBs that go into
it.  The match between the DIMM backend and the PAPR LMB frontend is
already pretty poor, covering NVDIMMs in the same range pushes it past
breaking point.

> > > > > > 3) You're not setting *any* extra flags on the entry.  How is the
> > > > > > guest supposed to know which are NVDIMM entries and which are regular
> > > > > > DIMM entries?  AFAICT in this version the NVDIMM slots are
> > > > > > indistinguishable from the unassigned hotplug memory (which makes the
> > > > > > difference in LMB and DRC numbering even more troubling).
> > > > > For NVDIMM case, this patch should populate the LMB set in
> > > > > ibm,dynamic-memory-v2 something like below:
> > > > >               elem = spapr_get_drconf_cell(size /lmb_size, addr,
> > > > >                                            0, -1,
> > > > > SPAPR_LMB_FLAGS_RESERVED | SPAPR_LMB_FLAGS_DRC_INVALID);
> > > > > 
> > > > > This will ensure that the NVDIMM range will never be considered as
> > > > > valid memory range for memory hotplug.
> > > > Hrm.  Ok so we already have code that does that for any gaps between
> > > > DIMMs.  I don't think there's actually anything that that code will do
> > > > differently than the code you have for NVDIMMs, so you could just skip
> > > > over the NVDIMMs here and it should do the right thing.
> > > > 
> > > > The *interpretation* of those entries will become different: for space
> > > > into which a regular DIMM is later inserted, we'll assume the DRC
> > > > index given is a base and there are more DRCs following it, but for
> > > > NVDIMMs we'll assume the same DRC throughout.  This is nuts, but IIUC
> > > > that's what PAPR says and we can't do much about it.
> > > My current patch is buggy as Bharata pointed out. The NVDIMM DRCs
> > > are not to be populated here, but mark the LMB DRCs as RESERVED and INVALID
> > > so that no malicious attempts to online those LMBs at those NVDIMM address
> > > ranges are attempted.
> > > 
> > > > > > 4) AFAICT these are _present_ NVDIMMs, so why is
> > > > > > SPAPR_LMB_FLAGS_ASSIGNED not set for them?  (and why is the node
> > > > > > forced to -1, regardless of di->node).
> > > > > > >            QSIMPLEQ_INSERT_TAIL(&drconf_queue, elem, entry);
> > > > > > >            nr_entries++;
> > > > > > >            cur_addr = addr + size;
> > > > > > > @@ -1261,6 +1273,85 @@ static void spapr_dt_hypervisor(SpaprMachineState *spapr, void *fdt)
> > > > > > >        }
> > > > > > >    }
> > > > > > > 
> > > > > > > +static void spapr_create_nvdimm_dr_connectors(SpaprMachineState *spapr)
> > > > > > > +{
> > > > > > > +    MachineState *machine = MACHINE(spapr);
> > > > > > > +    int i;
> > > > > > > +
> > > > > > > +    for (i = 0; i < machine->ram_slots; i++) {
> > > > > > > +        spapr_dr_connector_new(OBJECT(spapr), TYPE_SPAPR_DRC_PMEM, i);
> > > > > > What happens if you try to plug an NVDIMM to one of these slots, but a
> > > > > > regular DIMM has already taken it?
> > > > > NVDIMM hotplug won't get that occupied slot.
> > > > Ok.
> 

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]