On Thu, Aug 08, 2024, Peter Xu wrote:
> On Thu, Aug 08, 2024 at 02:31:19PM -0700, Sean Christopherson wrote:
> > On Thu, Aug 08, 2024, Peter Xu wrote:
> > > Hi, Sean,
> > >
> > > On Thu, Aug 08, 2024 at 08:33:59AM -0700, Sean Christopherson wrote:
> > > > On Wed, Aug 07, 2024, Peter Xu wrote:
> > > > > mprotect() does mmu notifiers in PMD levels. It's there since 2014 of
> > > > > commit a5338093bfb4 ("mm: move mmu notifier call from change_protection to
> > > > > change_pmd_range").
> > > > >
> > > > > At that time, the issue was that NUMA balancing can be applied on a huge
> > > > > range of VM memory, even if nothing was populated. The notification can be
> > > > > avoided in this case if no valid pmd detected, which includes either THP or
> > > > > a PTE pgtable page.
> > > > >
> > > > > Now to pave way for PUD handling, this isn't enough. We need to generate
> > > > > mmu notifications even on PUD entries properly. mprotect() is currently
> > > > > broken on PUD (e.g., one can easily trigger kernel error with dax 1G
> > > > > mappings already), this is the start to fix it.
> > > > >
> > > > > To fix that, this patch proposes to push such notifications to the PUD
> > > > > layers.
> > > > >
> > > > > There is risk on regressing the problem Rik wanted to resolve before, but I
> > > > > think it shouldn't really happen, and I still chose this solution because
> > > > > of a few reasons:
> > > > >
> > > > > 1) Consider a large VM that should definitely contain more than GBs of
> > > > > memory, it's highly likely that PUDs are also none. In this case there
> > > >
> > > > I don't follow this. Did you mean to say it's highly likely that PUDs are *NOT*
> > > > none?
> > >
> > > I did mean the original wordings.
> > >
> > > Note that in the previous case Rik worked on, it's about a mostly empty VM
> > > got NUMA hint applied. So I did mean "PUDs are also none" here, with the
> > > hope that when the numa hint applies on any part of the unpopulated guest
> > > memory, it'll find nothing in PUDs. Here it's mostly not about a huge PUD
> > > mapping as long as the guest memory is not backed by DAX (since only DAX
> > > supports 1G huge pud so far, while hugetlb has its own path here in
> > > mprotect, so it must be things like anon or shmem), but a PUD entry that
> > > contains pmd pgtables. For that part, I was trying to justify "no pmd
> > > pgtable installed" with the fact that "a large VM that should definitely
> > > contain more than GBs of memory", it means the PUD range should hopefully
> > > never been accessed, so even the pmd pgtable entry should be missing.
> >
> > Ah, now I get what you were saying.
> >
> > Problem is, walking the rmaps for the shadow MMU doesn't benefit (much) from
> > empty PUDs, because KVM needs to blindly walk the rmaps for every gfn covered by
> > the PUD to see if there are any SPTEs in any shadow MMUs mapping that gfn. And
> > that walk is done without ever yielding, which I suspect is the source of the
> > soft lockups of yore.
> >
> > And there's no way around that conundrum (walking rmaps), at least not without a
> > major rewrite in KVM. In a nested TDP scenario, KVM's stage-2 page tables (for
> > L2) key off of L2 gfns, not L1 gfns, and so the only way to find mappings is
> > through the rmaps.
>
> I think the hope here is when the whole PUDs being hinted are empty without
> pgtable installed, there'll be no mmu notifier to be kicked off at all.
>
> To be explicit, I meant after this patch applied, the pud loop for numa
> hints look like this:
>
> FOR_EACH_PUD() {
> ...
> if (pud_none(pud))
> continue;
>
> if (!range.start) {
> mmu_notifier_range_init(&range,
> MMU_NOTIFY_PROTECTION_VMA, 0,
> vma->vm_mm, addr, end);
> mmu_notifier_invalidate_range_start(&range);
> }
> ...
> }
>
> So the hope is that pud_none() is always true for the hinted area (just
> like it used to be when pmd_none() can be hopefully true always), then we
> skip the mmu notifier as a whole (including KVM's)!
Gotcha, that makes sense. Too many page tables flying around :-)
