Re: [PATCH] mm/khugepaged: Don't collapse uffd-minor-registered VMAs

[Kiryl Shutsemau]

On Mon, Aug 31, 2026 at 09:51:33AM -0700, James Houghton wrote:
> On Fri, Aug 28, 2026 at 10:26 PM Lance Yang <lance.yang@xxxxxxxxx> wrote:
> > >On Fri, Aug 28, 2026 at 12:50:04AM +0000, James Houghton wrote:
> > >>If userspace implements post-copy live migration using userfaultfd minor
> > >>faults, this situation is currently possible:
> > >>1. The VMA for guest memory is userfaultfd-minor-registered and nothing
> > >>   is mapped in the page tables.
> > >>2. A stale copy of a page is present in a naturally-aligned THP (from
> > >>   pre-copy live migration).
> > >>3. khugepaged collapses the mapping of the THP, installs a PMD.
> >
> > Ouch ... I missed this earlier. The problem is real, but this commit
> > message describes the wrong trigger.
> >
> > Background khugepaged calls try_collapse_pte_mapped_thp() with
> > install_pmd=false, so it cannot install the PMD or trigger this
> > sequence. MADV_COLLAPSE passes install_pmd=true and installs the PMD.
> >
> > So the problem described here can only be triggered by MADV_COLLAPSE,
> > whether it comes through madvise() or process_madvise(), no?
> 
> Oh yeah, you're right. In our downstream version, `install_pmd` is set
> to true for the khugepaged collapse path; I missed that it was false
> upstream. Sorry about that. :(
> 
> So we can definitely remove cc:stable, there is no bug here.
> 
> WRT the semantics of process_madvise(MADV_COLLAPSE) and normal
> madvise(MADV_COLLAPSE), I think it makes sense to always disallow
> collapsing of UFFD-minor VMAs, whether MADV_COLLAPSE comes from our
> own process or another. UFFD-minor faults are intended to be resolved
> only with UFFDIO_CONTINUE, and until someone asks for it, I don't
> really see a need to make MADV_COLLAPSE actually work to resolve minor
> faults. I don't think userspace can *rely* on MADV_COLLAPSE resolving
> minor faults either.

It is safer to go with stricter policy: reject always. We can always
relax it later if there is a use-case and reasoning about the interface
is solid. We don't have it yet.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov