Re: [PATCH] mm: introduce MADV_CLR_HUGEPAGE

From: Mike Rapoport
Date: Wed May 24 2017 - 10:27:59 EST


On Wed, May 24, 2017 at 01:18:00PM +0200, Michal Hocko wrote:
> On Wed 24-05-17 13:39:48, Mike Rapoport wrote:
> > On Wed, May 24, 2017 at 09:58:06AM +0200, Vlastimil Babka wrote:
> > > On 05/24/2017 09:50 AM, Mike Rapoport wrote:
> > > > On Mon, May 22, 2017 at 05:52:47PM +0200, Vlastimil Babka wrote:
> > > >> On 05/22/2017 04:29 PM, Mike Rapoport wrote:
> > > >>>
> > > >>> Probably I didn't explained it too well.
> > > >>>
> > > >>> The range is intentionally not populated. When we combine pre- and
> > > >>> post-copy for process migration, we create memory pre-dump without stopping
> > > >>> the process, then we freeze the process without dumping the pages it has
> > > >>> dirtied between pre-dump and freeze, and then, during restore, we populate
> > > >>> the dirtied pages using userfaultfd.
> > > >>>
> > > >>> When CRIU restores a process in such scenario, it does something like:
> > > >>>
> > > >>> * mmap() memory region
> > > >>> * fill in the pages that were collected during the pre-dump
> > > >>> * do some other stuff
> > > >>> * register memory region with userfaultfd
> > > >>> * populate the missing memory on demand
> > > >>>
> > > >>> khugepaged collapses the pages in the partially populated regions before we
> > > >>> have a chance to register these regions with userfaultfd, which would
> > > >>> prevent the collapse.
> > > >>>
> > > >>> We could have used MADV_NOHUGEPAGE right after the mmap() call, and then
> > > >>> there would be no race because there would be nothing for khugepaged to
> > > >>> collapse at that point. But the problem is that we have no way to reset
> > > >>> *HUGEPAGE flags after the memory restore is complete.
> > > >>
> > > >> Hmm, I wouldn't be that sure if this is indeed race-free. Check that
> > > >> this scenario is indeed impossible?
> > > >>
> > > >> - you do the mmap
> > > >> - khugepaged will choose the process' mm to scan
> > > >> - khugepaged will get to the vma in question, it doesn't have
> > > >> MADV_NOHUGEPAGE yet
> > > >> - you set MADV_NOHUGEPAGE on the vma
> > > >> - you start populating the vma
> > > >> - khugepaged sees the vma is non-empty, collapses
> > > >>
> > > >> unless I'm wrong, the racers will have mmap_sem for reading only when
> > > >> setting/checking the MADV_NOHUGEPAGE? Might be actually considered a bug.
> > > >>
> > > >> However, can't you use prctl(PR_SET_THP_DISABLE) instead? "If arg2 has a
> > > >> nonzero value, the flag is set, otherwise it is cleared." says the
> > > >> manpage. Do it before the mmap and you avoid the race as well?
> > > >
> > > > Unfortunately, prctl(PR_SET_THP_DISABLE) didn't help :(
> > > > When I've tried to use it, I've ended up with VM_NOHUGEPAGE set on all VMAs
> > > > created after prctl(). This returns me to the state when checkpoint-restore
> > > > alters the application vma->vm_flags although it shouldn't and I do not see
> > > > a way to fix it using existing interfaces.
> > >
> > > [CC linux-api, should have been done in the initial posting already]
> >
> > Sorry, missed that.
> >
> > > Hm so the prctl does:
> > >
> > > if (arg2)
> > > me->mm->def_flags |= VM_NOHUGEPAGE;
> > > else
> > > me->mm->def_flags &= ~VM_NOHUGEPAGE;
> > >
> > > That's rather lazy implementation IMHO. Could we change it so the flag
> > > is stored elsewhere in the mm, and the code that decides to (not) use
> > > THP will check both the per-vma flag and the per-mm flag?
> >
> > I afraid I don't understand how that can help.
> > What we need is an ability to temporarily disable collapse of the pages in
> > VMAs that do not have VM_*HUGEPAGE flags set and that after we re-enable
> > THP, the vma->vm_flags for those VMAs will remain intact.
>
> Why cannot khugepaged simply skip over all VMAs which have userfault
> regions registered? This would sound like a less error prone approach to
> me.

khugepaged does skip over VMAs which have userfault. We could register the
regions with userfault before populating them to avoid collapses in the
transition period. But then we'll have to populate these regions with
UFFDIO_COPY which adds quite an overhead.

> --
> Michal Hocko
> SUSE Labs
>