On Wed, Jun 23, 2021 at 10:07:36AM +0200, Andrew Jones wrote:
On Wed, Jun 23, 2021 at 02:43:49PM +1000, Gavin Shan wrote:
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:
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.
Also, before posting the documentation, you may want to confirm that this
configuration is valid for real hardware. If not, then the solution to
this QEMU bug is to not allow this configuration.