Hi Rob,
On 9/24/21 1:17 AM, Rob Herring wrote:
On Thu, Sep 23, 2021 at 1:32 AM Gavin Shan <gshan@xxxxxxxxxx> wrote:
On 9/22/21 9:05 PM, Ard Biesheuvel wrote:
On Tue, 21 Sept 2021 at 21:45, Rob Herring <robh@xxxxxxxxxx> wrote:
On Sun, Sep 5, 2021 at 11:16 PM Gavin Shan <gshan@xxxxxxxxxx> wrote:
The empty memory nodes, where no memory resides in, are allowed.
For these empty memory nodes, the 'len' of 'reg' property is zero.
The NUMA node IDs are still valid and parsed, but memory may be
added to them through hotplug afterwards. Currently, QEMU fails
to boot when multiple empty memory nodes are specified. It's
caused by device-tree population failure and duplicated memory
node names.
Those memory regions are known in advance, right? So wouldn't it be
better to use something like 'status = "disabled"' here?
Yes, these memory regions are known in advance. For the empty nodes,
their 'len' property is zero and it's equal to disabled state.
I still don't like the fake addresses. I can't really give suggestions
on alternative ways to fix this with you just presenting a solution.
Agreed. Please try to explain what the problem is, and why this is the
best way to solve it. Please include other solutions that were
considered and rejected if any exist.
What is the failure you see? Can we relax the kernel's expectations?
What about UEFI boot as the memory nodes aren't used (or maybe they
are for NUMA?) How does this work with ACPI?
The EFI memory map only needs to describe the memory that was present
at boot. More memory can be represented as ACPI objects, including
coldplugged memory that is already present at boot. None of this
involves the memory nodes in DT.
I'm using the following command line to start a virtual machine (VM).
There are 4 NUMA nodes specified, but the last two are empty. In QEMU,
the device-tree nodes are populated to represent these 4 NUMA nodes.
Unfortunately, QEMU fails to start because of the conflicting names
for the empty node are found, as the following error message indicates.
/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,slots=16,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
According to device-tree specification, the memory device-tree node's
name is following the format 'memory@base-address'. For the empty
NUMA nodes, their base addresses aren't determined. The device-tree
specification doesn't indicate what 'base-address' should be assigned
for the empty nodes. So I proposed this patch because I think the
linux device-tree binding documentation is best place to get this
documented.
Why even create the node?
What does IBM pSeries do here. AIUI, those platforms create/remove
nodes for hotplug. That's the reason CONFIG_OF_DYNAMIC existed
originally. Unfortunately, that's the extent of my knowledge on that.
It has been long time that I didn't read pSeries related code. I
spent some time on that and you're correct. On pSeries, the device-tree
node is added dynamically. However, the IBM private nodes or properties,
whose names start with "ibm", are used for the memory hotplug. So
ARM64 can't follow without adaption.
I agree on your suggestion, which will be reflected in v2:
(a) the memory device-tree nodes are added and removed on memory hot
add and removal.
(b) the supported NUMA nodes, including the empty ones, are identified
through "numa-distance-map" compatible device-tree node. It's
exactly same to what you suggested below.
In this way, we won't have the issue of the conflicting memory node
names, introduced by the empty NUMA nodes.
ACPI is different story. The NUMA nodes are represented by SRAT
(System Resource Affinity Table). In the above example, there are
4 SRATs. We needn't assign names to the tables and we don't have
the conflicting names as we do in device-tree case.
By the way, QEMU currently prevents to expose SRATs for empty NUMA
nodes. I need submit QEMU patch to break the limitation in future.
With the limitation, the hot-added memory is always put into the
last NUMA node and it's not exactly customer wants.
As device-tree specification indicates, the 'unit-address' of
these empty memory nodes, part of their names, are the equivalents
to 'base-address'. Unfortunately, I finds difficulty to get where
the assignment of 'base-address' is properly documented for these
empty memory nodes. So lets add a section for empty memory nodes
to cover this in NUMA binding document. The 'unit-address',
equivalent to 'base-address' in the 'reg' property of these empty
memory nodes is specified to be the summation of highest memory
address plus the NUMA node ID.
Signed-off-by: Gavin Shan <gshan@xxxxxxxxxx>
Acked-by: Randy Dunlap <rdunlap@xxxxxxxxxxxxx>
---
Documentation/devicetree/bindings/numa.txt | 60 +++++++++++++++++++++-
1 file changed, 59 insertions(+), 1 deletion(-)
diff --git a/Documentation/devicetree/bindings/numa.txt b/Documentation/devicetree/bindings/numa.txt
index 21b35053ca5a..82f047bc8dd6 100644
--- a/Documentation/devicetree/bindings/numa.txt
+++ b/Documentation/devicetree/bindings/numa.txt
@@ -103,7 +103,65 @@ Example:
};
==============================================================================
-4 - Example dts
+4 - Empty memory nodes
+==============================================================================
+
+Empty memory nodes, which no memory resides in, are allowed. The 'length'
+field of the 'reg' property is zero. However, the 'base-address' is a
+dummy and invalid address, which is the summation of highest memory address
+plus the NUMA node ID. The NUMA node IDs and distance maps are still valid
+and memory may be added into them through hotplug afterwards.
+
+Example:
+
+ memory@0 {
+ device_type = "memory";
+ reg = <0x0 0x0 0x0 0x80000000>;
+ numa-node-id = <0>;
+ };
+
+ memory@80000000 {
+ device_type = "memory";
+ reg = <0x0 0x80000000 0x0 0x80000000>;
+ numa-node-id = <1>;
+ };
+
+ /* Empty memory node */
+ memory@100000002 {
+ device_type = "memory";
+ reg = <0x1 0x2 0x0 0x0>;
+ numa-node-id = <2>;
+ };
+
+ /* Empty memory node */
+ memory@100000003 {
+ device_type = "memory";
+ reg = <0x1 0x3 0x0 0x0>;
+ numa-node-id = <3>;
+ };
Do you really need the memory nodes here or just some way to define
numa node id's 2 and 3 as valid?
It's the way to define NUMA node IDs are valid. Besides, the 'reg'
property provides 'base-address', which is part of the device-tree
node's name, as described in this patch.
The distance-matrix already lists all possible NUMA node IDs. That
should be enough information for the kernel. If not, fix the kernel.
Thanks, Rob. An extra patch will address the issue in v2.
Thanks,
Gavin
