Re: [PATCH v1 0/4] vfio: report NUMA nodes for device memory

[Alex Williamson]

On Tue, 26 Sep 2023 18:54:53 +0200
David Hildenbrand <david@redhat.com> wrote:

> On 26.09.23 16:52, Ankit Agrawal wrote:
> >>>>> Good idea.  Fundamentally the device should not be creating NUMA
> >>>>> nodes, the VM should be configured with NUMA nodes and the device
> >>>>> memory associated with those nodes.  
> >>>>
> >>>> +1. That would also make it fly with DIMMs and virtio-mem, where you
> >>>> would want NUMA-less nodes ass well (imagine passing CXL memory to a VM
> >>>> using virtio-mem).
> >>>>  
> >>>
> >>> We actually do not add the device memory on the host, instead
> >>> map it into the Qemu VMA using remap_pfn_range(). Please checkout the
> >>> mmap function in vfio-pci variant driver code managing the device.
> >>> https://lore.kernel.org/all/20230915025415.6762-1-ankita@nvidia.com/
> >>> And I think host memory backend would need memory that is added on the
> >>> host.
> >>>
> >>> Moreover since we want to passthrough the entire device memory, the
> >>> -object memory-backend-ram would have to be passed a size that is equal
> >>> to the device memory. I wonder if that would be too much of a trouble
> >>> for an admin (or libvirt) triggering the Qemu process.
> >>>
> >>> Both these items are avoided by exposing the device memory as BAR as in 
> >>> the
> >>> current  implementation (referenced above) since it lets Qemu to naturally
> >>> discover the device memory region and do mmap.
> >>>  
> >>
> >> Just to clarify: nNUMA nodes for DIMMs/NVDIMMs/virtio-mem are configured
> >> on the device, not on the memory backend.
> >>
> >> e.g., -device pc-dimm,node=3,memdev=mem1,...  
> >   
> 
> Alco CCing Gavin, I remember he once experimented with virtio-mem + 
> multiple memory-less nodes and it was quite working (because of 
> MEM_AFFINITY_HOTPLUGGABLE only on the last node, below).
> 
> > Agreed, but still we will have the aforementioned issues viz.
> > 1. The backing memory for the memory device would need to be allocated
> > on the host. However, we do not add the device memory on the host in this
> > case. Instead the Qemu VMA is mapped to the device memory physical
> > address using remap_pfn_range().  
> 
> I don't see why that would be necessary ...
> 
> > 2. The memory device need to be passed an allocation size such that all of
> > the device memory is mapped into the Qemu VMA. This may not be readily
> > available to the admin/libvirt.  
> 
> ... or that. But your proposal roughly looks like what I had in mind, so 
> let's focus on that.
> 
> > 
> > Based on the suggestions here, can we consider something like the
> > following?
> > 1. Introduce a new -numa subparam 'devnode', which tells Qemu to mark
> > the node with MEM_AFFINITY_HOTPLUGGABLE in the SRAT's memory affinity
> > structure to make it hotpluggable.  
> 
> Is that "devnode=on" parameter required? Can't we simply expose any node 
> that does *not* have any boot memory assigned as MEM_AFFINITY_HOTPLUGGABLE?
> 
> Right now, with "ordinary", fixed-location memory devices 
> (DIMM/NVDIMM/virtio-mem/virtio-pmem), we create an srat entry that 
> covers the device memory region for these devices with 
> MEM_AFFINITY_HOTPLUGGABLE. We use the highest NUMA node in the machine, 
> which does not quite work IIRC. All applicable nodes that don't have 
> boot memory would need MEM_AFFINITY_HOTPLUGGABLE for Linux to create them.
> 
> In your example, which memory ranges would we use for these nodes in SRAT?
> 
> > 2. Create several NUMA nodes with 'devnode' which are supposed to be
> > associated with the vfio-pci device.
> > 3. Pass the numa node start and count to associate the nodes created.
> > 
> > So, the command would look something like the following.
> > ...
> >          -numa node,nodeid=2,devnode=on \
> >          -numa node,nodeid=3,devnode=on \
> >          -numa node,nodeid=4,devnode=on \
> >          -numa node,nodeid=5,devnode=on \
> >          -numa node,nodeid=6,devnode=on \
> >          -numa node,nodeid=7,devnode=on \
> >          -numa node,nodeid=8,devnode=on \
> >          -numa node,nodeid=9,devnode=on \
> >          -device 
> > vfio-pci-nohotplug,host=0009:01:00.0,bus=pcie.0,addr=04.0,rombar=0,numa-node-start=2,numa-node-count=8
> >  \  

I don't see how these numa-node args on a vfio-pci device have any
general utility.  They're only used to create a firmware table, so why
don't we be explicit about it and define the firmware table as an
object?  For example:

        -numa node,nodeid=2 \
        -numa node,nodeid=3 \
        -numa node,nodeid=4 \
        -numa node,nodeid=5 \
        -numa node,nodeid=6 \
        -numa node,nodeid=7 \
        -numa node,nodeid=8 \
        -numa node,nodeid=9 \
        -device 
vfio-pci-nohotplug,host=0009:01:00.0,bus=pcie.0,addr=04.0,rombar=0,id=nvgrace0 \
        -object nvidia-gpu-mem-acpi,devid=nvgrace0,nodeset=2-9 \

There are some suggestions in this thread that CXL could have similar
requirements, but I haven't found any evidence that these
dev-mem-pxm-{start,count} attributes in the _DSD are standardized in
any way.  If they are, maybe this would be a dev-mem-pxm-acpi object
rather than an NVIDIA specific one.

It seems like we could almost meet the requirement for this table via
-acpitable, but I think we'd like to avoid the VM orchestration tool
from creating, compiling, and passing ACPI data blobs into the VM.
Thanks,

Alex