Re: [libvirt] [PATCH 2/2] qemu: Enable NUMA node tag in pci-root for PPC64

On 11/08/2016 05:59 PM, Daniel P. Berrange wrote: > That doesn't really solve the bug above properly - your solution here > only works if all the PCI devices (whether emulated or host-assigned) > are to be associated with the same guest NUMA node. This is a pretty > restrictive setup and so of limited use IMHO. Yes. It is restrictive. There can be no 1:1 pinning of host to guest nodes. But, one can have multiple host nodes pinned to 1 guest node which the host bridge is pinned to. This helps only to the extent that, the guest drivers wont find -1 in the sysfs/numa_node. Some drivers are known to crash when they find -1. The patch addresses that part of the problem. > It seems like to solve this requires that we create multiple PCI Host > Bridges ? Yes. That will be the actual fix to have 1:1 pinning. The current approach keeps that in mind and we can have the numa_node attribute set at the indivudual spapr-pci-host-bridge level when we have it. Then, the individual <pci-root id=X> can have different target node as I have in this patch. Thanks, Shivaprasad > Daniel