Re: [PATCH v5 07/12] khugepaged: add mTHP support

[David Hildenbrand]

On 02.05.25 00:29, Nico Pache wrote:
> On Wed, Apr 30, 2025 at 2:53 PM Jann Horn <jannh@xxxxxxxxxx> wrote:
> > On Mon, Apr 28, 2025 at 8:12 PM Nico Pache <npache@xxxxxxxxxx> wrote:
> > > Introduce the ability for khugepaged to collapse to different mTHP sizes.
> > > While scanning PMD ranges for potential collapse candidates, keep track
> > > of pages in KHUGEPAGED_MIN_MTHP_ORDER chunks via a bitmap. Each bit
> > > represents a utilized region of order KHUGEPAGED_MIN_MTHP_ORDER ptes. If
> > > mTHPs are enabled we remove the restriction of max_ptes_none during the
> > > scan phase so we dont bailout early and miss potential mTHP candidates.
> > >
> > > After the scan is complete we will perform binary recursion on the
> > > bitmap to determine which mTHP size would be most efficient to collapse
> > > to. max_ptes_none will be scaled by the attempted collapse order to
> > > determine how full a THP must be to be eligible.
> > >
> > > If a mTHP collapse is attempted, but contains swapped out, or shared
> > > pages, we dont perform the collapse.
> > [...]
> > > @@ -1208,11 +1211,12 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
> > >          vma_start_write(vma);
> > >          anon_vma_lock_write(vma->anon_vma);
> > >
> > > -       mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address,
> > > -                               address + HPAGE_PMD_SIZE);
> > > +       mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, _address,
> > > +                               _address + (PAGE_SIZE << order));
> > >          mmu_notifier_invalidate_range_start(&range);
> > >
> > >          pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
> > > +
> > >          /*
> > >           * This removes any huge TLB entry from the CPU so we won't allow
> > >           * huge and small TLB entries for the same virtual address to
> >
> > It's not visible in this diff, but we're about to do a
> > pmdp_collapse_flush() here. pmdp_collapse_flush() tears down the
> > entire page table, meaning it tears down 2MiB of address space; and it
> > assumes that the entire page table exclusively corresponds to the
> > current VMA.
> >
> > I think you'll need to ensure that the pmdp_collapse_flush() only
> > happens for full-size THP, and that mTHP only tears down individual
> > PTEs in the relevant range. (That code might get a bit messy, since
> > the existing THP code tears down PTEs in a detached page table, while
> > mTHP would have to do it in a still-attached page table.)
> Hi Jann!
>
> I was under the impression that this is needed to prevent GUP-fast
> races (and potentially others).
> As you state here, conceptually the PMD case is, detach the PMD, do
> the collapse, then reinstall the PMD (similarly to how the system
> recovers from a failed PMD collapse). I tried to keep the current
> locking behavior as it seemed the easiest way to get it right (and not
> break anything). So I keep the PMD detaching and reinstalling for the
> mTHP case too. As Hugh points out I am releasing the anon lock too
> early. I will comment further on his response.
>
> As I familiarize myself with the code more, I do see potential code
> improvements/cleanups and locking improvements, but I was going to
> leave those to a later series.

Right, the simplest approach on top of the current PMD collapse is to do 
exactly what we do in the PMD case, including the locking: which 
apparently is no completely the same yet :).

Instead of installing a PMD THP, we modify the page table and remap that.

Moving from the PMD lock to the PTE lock will not make a big change in 
practice for most cases: we already must disable essentially all page 
table walkers (vma lock, mmap lock in write, rmap lock in write).

The PMDP clear+flush is primarily to disable the last possible set of 
page table walkers: (1) HW modifications and (2) GUP-fast.

So after the PMDP clear+flush we know that (A) HW can not modify the 
pages concurrently and (B) GUP-fast cannot succeed anymore.

The issue with PTEP clear+flush is that we will have to remember all PTE 
values, to reset them if anything goes wrong. Using a single PMD value 
is arguably simpler. And then, the benefit vs. complexity is unclear.

Certainly something to look into later, but not a requirement for the 
first support,

The real challenge/benefit will be looking into avoiding taking all the 
heavy weight locks. Dev has already been thinking about that. For mTHP 
it might be easier than for THPs. Probably it will involve setting PTE 
migration entries whenever we drop the PTL, and dealing with the 
possibility of concurrent zapping of these migration entries.

--
Cheers,

David / dhildenb