On Thu, Dec 8, 2022 at 4:28 PM David Matlack <dmatlack@xxxxxxxxxx> wrote: > > On Thu, Dec 01, 2022 at 11:57:18AM -0800, Vipin Sharma wrote: > > Page table pages of a VM are currently allocated based on the current > > task's NUMA node or its mempolicy. This can cause suboptimal remote > > accesses by the vCPU if it is accessing physical pages local to its NUMA > > node but the page table pages mapping those physcal pages were created > > by some other vCPU which was on different NUMA node or had different > > policy. > > > > Allocate page table pages on the same NUMA node where underlying > > physical page exists. Page table at level 5, 4, and 3 might not end up > > on the same NUMA node as they can span multiple NUMA nodes. > > > > Signed-off-by: Vipin Sharma <vipinsh@xxxxxxxxxx> > > --- > ... > > @@ -6284,13 +6326,16 @@ static int shadow_mmu_try_split_huge_page(struct kvm *kvm, > > gfn = kvm_mmu_page_get_gfn(huge_sp, spte_index(huge_sptep)); > > level = huge_sp->role.level; > > spte = *huge_sptep; > > + nid = kvm_pfn_to_refcounted_page_nid(spte_to_pfn(spte)); > > + if (nid == NUMA_NO_NODE) > > + nid = numa_mem_id(); > > What do you think about renaming kvm_pfn_to_refcounted_page_nid() to > kvm_pfn_to_page_table_nid() and having it return numa_mem_id() instead > of NUMA_NO_NODE (with a comment)? I think that will clean up this patch > quite a bit by getting rid of all the NUMA_NO_NODE checks. Yeah, this will clean up this patch. I will make this change in the next version.
