Re: [PATCH] hw/arm/boot: Use NUMA node ID in memory node name

[Andrew Jones]

On Wed, Jun 23, 2021 at 02:43:49PM +1000, Gavin Shan wrote:
> Hi Drew,
> 
> On 6/22/21 5:13 PM, Andrew Jones wrote:
> > On Tue, Jun 22, 2021 at 06:53:41PM +1000, Gavin Shan wrote:
> > > On 6/3/21 2:48 PM, Gavin Shan wrote:
> > > > On 6/2/21 9:36 PM, Andrew Jones wrote:
> > > > > On Wed, Jun 02, 2021 at 11:09:32AM +1000, Gavin Shan wrote:
> > > > > > On 6/1/21 5:50 PM, Andrew Jones wrote:
> > > > > > > On Tue, Jun 01, 2021 at 03:30:04PM +0800, Gavin Shan wrote:
> > > > > > > > We possibly populate empty nodes where memory isn't included 
> > > > > > > > and might
> > > > > > > > be hot added at late time. The FDT memory nodes can't be 
> > > > > > > > created due
> > > > > > > > to conflicts on their names if multiple empty nodes are 
> > > > > > > > specified.
> > > > > > > > For example, the VM fails to start with the following error 
> > > > > > > > messages.
> > > > > > > > 
> > > > > > > >      /home/gavin/sandbox/qemu.main/build/qemu-system-aarch64    
> > > > > > > >       \
> > > > > > > >      -accel kvm -machine virt,gic-version=host                  
> > > > > > > >       \
> > > > > > > >      -cpu host -smp 4,sockets=2,cores=2,threads=1 -m 
> > > > > > > > 1024M,maxmem=64G \
> > > > > > > >      -object memory-backend-ram,id=mem0,size=512M               
> > > > > > > >       \
> > > > > > > >      -object memory-backend-ram,id=mem1,size=512M               
> > > > > > > >       \
> > > > > > > >      -numa node,nodeid=0,cpus=0-1,memdev=mem0                   
> > > > > > > >       \
> > > > > > > >      -numa node,nodeid=1,cpus=2-3,memdev=mem1                   
> > > > > > > >       \
> > > > > > > >      -numa node,nodeid=2                                        
> > > > > > > >       \
> > > > > > > >      -numa node,nodeid=3                                        
> > > > > > > >       \
> > > > > > > >        :
> > > > > > > >      -device 
> > > > > > > > virtio-balloon-pci,id=balloon0,free-page-reporting=yes
> > > > > > > > 
> > > > > > > >      qemu-system-aarch64: FDT: Failed to create subnode 
> > > > > > > > /memory@80000000: \
> > > > > > > >                           FDT_ERR_EXISTS
> > > > > > > > 
> > > > > > > > This fixes the issue by using NUMA node ID or zero in the 
> > > > > > > > memory node
> > > > > > > > name to avoid the conflicting memory node names. With this 
> > > > > > > > applied, the
> > > > > > > > VM can boot successfully with above command lines.
> > > > > > > > 
> > > > > > > > Signed-off-by: Gavin Shan <gshan@redhat.com>
> > > > > > > > ---
> > > > > > > >     hw/arm/boot.c | 7 ++++++-
> > > > > > > >     1 file changed, 6 insertions(+), 1 deletion(-)
> > > > > > > > 
> > > > > > > > diff --git a/hw/arm/boot.c b/hw/arm/boot.c
> > > > > > > > index d7b059225e..3169bdf595 100644
> > > > > > > > --- a/hw/arm/boot.c
> > > > > > > > +++ b/hw/arm/boot.c
> > > > > > > > @@ -432,7 +432,12 @@ static int fdt_add_memory_node(void *fdt, 
> > > > > > > > uint32_t acells, hwaddr mem_base,
> > > > > > > >         char *nodename;
> > > > > > > >         int ret;
> > > > > > > > -    nodename = g_strdup_printf("/memory@%" PRIx64, mem_base);
> > > > > > > > +    if (numa_node_id >= 0) {
> > > > > > > > +        nodename = g_strdup_printf("/memory@%d", numa_node_id);
> > > > > > > > +    } else {
> > > > > > > > +        nodename = g_strdup("/memory@0");
> > > > > > > > +    }
> > > > > > > > +
> > > > > > > >         qemu_fdt_add_subnode(fdt, nodename);
> > > > > > > >         qemu_fdt_setprop_string(fdt, nodename, "device_type", 
> > > > > > > > "memory");
> > > > > > > >         ret = qemu_fdt_setprop_sized_cells(fdt, nodename, 
> > > > > > > > "reg", acells, mem_base,
> > > 
> > > [...]
> > > 
> > > > 
> > > > I've sent one separate mail to check with Rob Herring. Hopefully he have
> > > > ideas as he is maintaining linux FDT subsystem. You have been included 
> > > > to
> > > > that thread. I didn't find something meaningful to this thread after 
> > > > doing
> > > > some google search either.
> > > > 
> > > > Yes, I agree with you we need to follow the specification strictly. It 
> > > > seems
> > > > it's uncertain about the 'physical memory map' bus binding requirements.
> > > > 
> > > 
> > > I didn't get expected answers from device-tree experts.
> > 
> > What response did you get? Can you please provide a link to the discussion?
> > 
> 
> Sorry about the confusion. I meant "no response" by "expected answers". Here
> is the mail sent to Rob before, no reply so far:
> 
> https://lkml.org/lkml/2021/6/2/1446
> 
> > > After rethinking about it,
> > > I plan to fix this like this way, but please let me know if it sounds 
> > > sensible
> > > to you.
> > > 
> > > The idea is to assign a (not overlapped) dummy base address to each memory
> > > node in the device-tree. The dummy is (last_valid_memory_address + NUMA 
> > > ID).
> > > The 'length' of the 'reg' property in the device-tree nodes, corresponding
> > > to empty NUMA nodes, is still zero. This ensures the nodes are still 
> > > invalid
> > > until memory is added to these nodes.

Since we don't know of any other established way to do this, then
proposing a solution makes sense. However, QEMU isn't the place to put it
first. Please write a Documentation/devicetree/bindings/ document and post
the patch to the kernel. If it gets accepted, then we can implement what
you've documented there.

Hopefully you'll get more interest and activity on your patch than on your
inquiry.

Thanks,
drew

> > > 
> > > I had the temporary patch for the implementation. It works fine and VM can
> > > boot up successfully.
> > 
> > I would like to be sure that the kernel developers for NUMA, memory
> > hotplug, and devicetree specifications are all happy with the approach
> > before adding it to QEMU.
> > 
> 
> As I understood, it won't break anything from perspectives of NUMA
> and device-tree specification. First of all, NUMA cares the so-called
> distance map and 'numa-node-id' property in the individual device-tree
> nodes. The device-tree specification doesn't say 'length' in 'reg'
> property of memory node can't be zero. Also, the linux device-tree
> implementation has the check on 'length', the memory block won't be
> added if it's zero.
> 
> Documentation/devicetree/bindings/numa.txt has more details about
> the required device-tree NUMA properties.
> 
> I'm not sure about memory hotplug. I tried memory hot add and it seems
> working finely. Memory hot-add/remove are working without issues:
> 
> /home/gavin/sandbox/qemu.main/build/qemu-system-aarch64 \
> -accel kvm -machine virt,gic-version=host               \
> -cpu host -smp 4,sockets=2,cores=2,threads=1            \
> -m 4096M,slots=16,maxmem=64G                            \
> -object memory-backend-ram,id=mem0,size=2048M           \
> -object memory-backend-ram,id=mem1,size=2048M           \
> -numa node,nodeid=0,cpus=0-1,memdev=mem0                \
> -numa node,nodeid=1,cpus=2-3,memdev=mem1                \
> -numa node,nodeid=2                                     \
> -numa node,nodeid=3
>    :
> 
> Memory hot-add
> ===============
> 
> # cat /proc/meminfo  | grep MemTotal
> MemTotal:        4027472 kB
> # cat /sys/devices/system/node/node2/meminfo | grep MemTotal
> Node 2 MemTotal:       0 kB
> 
> (qemu) object_add memory-backend-ram,id=hp-mem0,size=1G
> (qemu) device_add pc-dimm,id=dimm0,memdev=hp-mem0,node=3
> 
> # cat /proc/meminfo  | grep MemTotal
> MemTotal:        5076048 kB
> # cat /sys/devices/system/node/node2/meminfo | grep MemTotal
> Node 2 MemTotal: 1048576 kB
> 
> Memory hot-remove
> =================
> 
> (qemu) device_del dimm0
> (qemu) object_del hp-mem0
> 
> # cat /proc/meminfo  | grep MemTotal
> MemTotal:        4027472 kB
> # cat /sys/devices/system/node/node2/meminfo | grep MemTotal
> cat: can't open '/sys/devices/system/node/node2/meminfo': No such file or 
> directory
> 
> After this point, the memory can be added again without issues with
> "object_add" and "device_add". So it sounds reasonable to remove
> the empty NUMA node during memory hot-remove.
> 
> Thanks,
> Gavin
>