Re: [PATCH v1] virtio_pmem: populate numa information

[Mina Almasry]

On Wed, Oct 26, 2022 at 2:50 PM Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
>
> Pankaj Gupta wrote:
> > > > > Compute the numa information for a virtio_pmem device from the memory
> > > > > range of the device. Previously, the target_node was always 0 since
> > > > > the ndr_desc.target_node field was never explicitly set. The code for
> > > > > computing the numa node is taken from cxl_pmem_region_probe in
> > > > > drivers/cxl/pmem.c.
> > > > >
> > > > > Signed-off-by: Michael Sammler <sammler@xxxxxxxxxx>

Tested-by: Mina Almasry <almasrymina@xxxxxxxxxx>

I don't have much expertise on this driver, but with the help of this
patch I was able to get memory tiering [1] emulation going on qemu. As
far as I know there is no alternative to this emulation, and so I
would love to see this or equivalent merged, if possible.

This is what I have going to get memory tiering emulation:

In qemu, added these configs:
      -object memory-backend-file,id=m4,share=on,mem-path="$path_to_virtio_pmem_file",size=2G
\
      -smp 2,sockets=2,maxcpus=2  \
      -numa node,nodeid=0,memdev=m0 \
      -numa node,nodeid=1,memdev=m1 \
      -numa node,nodeid=2,memdev=m2,initiator=0 \
      -numa node,nodeid=3,initiator=0 \
      -device virtio-pmem-pci,memdev=m4,id=nvdimm1 \

On boot, ran these commands:
    ndctl_static create-namespace -e namespace0.0 -m devdax -f 1&> /dev/null
    echo dax0.0 > /sys/bus/dax/drivers/device_dax/unbind
    echo dax0.0 > /sys/bus/dax/drivers/kmem/new_id
    for i in `ls /sys/devices/system/memory/`; do
      state=$(cat "/sys/devices/system/memory/$i/state" 2&>/dev/null)
      if [ "$state" == "offline" ]; then
        echo online_movable > "/sys/devices/system/memory/$i/state"
      fi
    done

Without this CL, I see the memory onlined in node 0 always, and is not
a separate memory tier. With this CL and qemu configs, the memory is
onlined in node 3 and is set as a separate memory tier, which enables
qemu-based development:

==> /sys/devices/virtual/memory_tiering/memory_tier22/nodelist <==
3
==> /sys/devices/virtual/memory_tiering/memory_tier4/nodelist <==
0-2

AFAIK there is no alternative to enabling memory tiering emulation in
qemu, and would love to see this or equivalent merged, if possible.


[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/ABI/testing/sysfs-kernel-mm-memory-tiers

> > > > > ---
> > > > >  drivers/nvdimm/virtio_pmem.c | 11 +++++++++--
> > > > >  1 file changed, 9 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/drivers/nvdimm/virtio_pmem.c b/drivers/nvdimm/virtio_pmem.c
> > > > > index 20da455d2ef6..a92eb172f0e7 100644
> > > > > --- a/drivers/nvdimm/virtio_pmem.c
> > > > > +++ b/drivers/nvdimm/virtio_pmem.c
> > > > > @@ -32,7 +32,6 @@ static int init_vq(struct virtio_pmem *vpmem)
> > > > >  static int virtio_pmem_probe(struct virtio_device *vdev)
> > > > >  {
> > > > >         struct nd_region_desc ndr_desc = {};
> > > > > -       int nid = dev_to_node(&vdev->dev);
> > > > >         struct nd_region *nd_region;
> > > > >         struct virtio_pmem *vpmem;
> > > > >         struct resource res;
> > > > > @@ -79,7 +78,15 @@ static int virtio_pmem_probe(struct virtio_device *vdev)
> > > > >         dev_set_drvdata(&vdev->dev, vpmem->nvdimm_bus);
> > > > >
> > > > >         ndr_desc.res = &res;
> > > > > -       ndr_desc.numa_node = nid;
> > > > > +
> > > > > +       ndr_desc.numa_node = memory_add_physaddr_to_nid(res.start);
> > > > > +       ndr_desc.target_node = phys_to_target_node(res.start);
> > > > > +       if (ndr_desc.target_node == NUMA_NO_NODE) {
> > > > > +               ndr_desc.target_node = ndr_desc.numa_node;
> > > > > +               dev_dbg(&vdev->dev, "changing target node from %d to %d",
> > > > > +                       NUMA_NO_NODE, ndr_desc.target_node);
> > > > > +       }
> > > >
> > > > As this memory later gets hotplugged using "devm_memremap_pages". I don't
> > > > see if 'target_node' is used for fsdax case?
> > > >
> > > > It seems to me "target_node" is used mainly for volatile range above
> > > > persistent memory ( e.g kmem driver?).
> > > >
> > > I am not sure if 'target_node' is used in the fsdax case, but it is
> > > indeed used by the devdax/kmem driver when hotplugging the memory (see
> > > 'dev_dax_kmem_probe' and '__dax_pmem_probe').
> >
> > Yes, but not currently for FS_DAX iiuc.
>
> The target_node is only used by the dax_kmem driver. In the FSDAX case
> the memory (persistent or otherwise) is mapped behind a block-device.
> That block-device has affinity to a CPU initiator, but that memory does
> not itself have any NUMA affinity or identity as a target.
>
> So:
>
> block-device NUMA node == closest CPU initiator node to the device
>
> dax-device target node == memory only NUMA node target, after onlining