Re: [pve-devel] [PATCH v3 qemu-server 11/13] memory: add virtio-mem support

["DERUMIER, Alexandre" <Alexandre.DERUMIER@groupe-cyllene.com>]

Le vendredi 03 février 2023 à 14:46 +0100, Fiona Ebner a écrit :
> Nit: If the number of $sockets is not a power of 2, I think this
> breaks.
> But I guess we already don't support it. Running current version
> without
> your patches (for a VM with memory hotplug):
> root@pve701 ~ # qm set 131 --sockets 3
> update VM 131: -sockets 3
> root@pve701 ~ # qm set 131 -memory 8192
> update VM 131: -memory 8192
> root@pve701 ~ # qm start 131
> kvm: total memory for NUMA nodes (0x3fffffff) should equal RAM size
> (0x40000000)
> start failed: QEMU exited with code 1
> 
> I guess we can just fix it up together with the existing rounding
> issue
> if/when somebody complains about it ;)

Yes, indeed, we never have supported it with numa, as we try to match
real hardware numa. (and real hardware are always power of 2).


with auto-numabalancing in kernel, the host kernel should try to
balance guest numa node to physical numanode.

Personnaly, I'll prefer to add a check somewhere and forbid it.

(I really don't known the behaviour of guest kernel scheduler with non-
power of 2 numa)




I'll rework the patches series for a v4 for end the next week.


(BTW, I'm going to fosdem tomorrow, so if some proxmox devs are there,
I'll happy to drink a beer with you guys ;)