On Fri, Jul 28, 2023 at 11:31:49PM +0200, David Hildenbrand wrote: > * vfio triggers FOLL_PIN|FOLL_LONGTERM from a random QEMU thread. > Where should we migrate that page to? Would it actually be counter- > productive to migrate it to the NUMA node of the setup thread? The > longterm pin will turn the page unmovable, yes, but where to migrate > it to? For VFIO & KVM you actively don't get any kind of numa balancing or awareness. In this case qemu should probably strive to put the memory on the numa node of the majorty of CPUs early on because it doesn't get another shot at it. In other cases it depends quite alot. Eg DPDK might want its VFIO buffers to NUMA'd to the node that is close to the device, not the CPU. Or vice versa. There is alot of micro sensitivity here at high data rates. I think people today manually tune this by deliberately allocating the memory to specific numas and then GUP should just leave it alone. FWIW, I'm reading this thread and I have no idea what the special semantic is KVM needs from GUP, so I'm all for better documentation on the GUP flag :) Jason
