Re: [PATCH 5/5] mm/thp: Split huge pmds/puds if they're pinned when fork()

From: Peter Xu
Date: Wed Sep 23 2020 - 11:24:17 EST


On Tue, Sep 22, 2020 at 09:05:05AM -0300, Jason Gunthorpe wrote:
> On Mon, Sep 21, 2020 at 05:20:31PM -0400, Peter Xu wrote:
> > Pinned pages shouldn't be write-protected when fork() happens, because follow
> > up copy-on-write on these pages could cause the pinned pages to be replaced by
> > random newly allocated pages.
> >
> > For huge PMDs, we split the huge pmd if pinning is detected. So that future
> > handling will be done by the PTE level (with our latest changes, each of the
> > small pages will be copied). We can achieve this by let copy_huge_pmd() return
> > -EAGAIN for pinned pages, so that we'll fallthrough in copy_pmd_range() and
> > finally land the next copy_pte_range() call.
> >
> > Huge PUDs will be even more special - so far it does not support anonymous
> > pages. But it can actually be done the same as the huge PMDs even if the split
> > huge PUDs means to erase the PUD entries. It'll guarantee the follow up fault
> > ins will remap the same pages in either parent/child later.
> >
> > This might not be the most efficient way, but it should be easy and clean
> > enough. It should be fine, since we're tackling with a very rare case just to
> > make sure userspaces that pinned some thps will still work even without
> > MADV_DONTFORK and after they fork()ed.
> >
> > Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
> > mm/huge_memory.c | 26 ++++++++++++++++++++++++++
> > 1 file changed, 26 insertions(+)
> >
> > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > index 7ff29cc3d55c..c40aac0ad87e 100644
> > +++ b/mm/huge_memory.c
> > @@ -1074,6 +1074,23 @@ int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
> >
> > src_page = pmd_page(pmd);
> > VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
> > +
> > + /*
> > + * If this page is a potentially pinned page, split and retry the fault
> > + * with smaller page size. Normally this should not happen because the
> > + * userspace should use MADV_DONTFORK upon pinned regions. This is a
> > + * best effort that the pinned pages won't be replaced by another
> > + * random page during the coming copy-on-write.
> > + */
> > + if (unlikely(READ_ONCE(src_mm->has_pinned) &&
> > + page_maybe_dma_pinned(src_page))) {
> > + pte_free(dst_mm, pgtable);
> > + spin_unlock(src_ptl);
> > + spin_unlock(dst_ptl);
> > + __split_huge_pmd(vma, src_pmd, addr, false, NULL);
> > + return -EAGAIN;
> > + }
>
> Not sure why, but the PMD stuff here is not calling is_cow_mapping()
> before doing the write protect. Seems like it might be an existing
> bug?

IMHO it's not a bug, because splitting a huge pmd should always be safe.

One thing I can think of that might be special here is when the pmd is
anonymously mapped but also shared (shared, tmpfs thp, I think?), then here
we'll also mark it as wrprotected even if we don't need to (or maybe we need it
for some reason..). But again I think it's safe anyways - when page fault
happens, wp_huge_pmd() should split it into smaller pages unconditionally. I
just don't know whether it's the ideal way for the shared case. Andrea should
definitely know it better (because it is there since the 1st day of thp).

>
> In any event, the has_pinned logic shouldn't be used without also
> checking is_cow_mapping(), so it should be added to that test. Same
> remarks for PUD

I think the case mentioned above is also the special case here when we didn't
check is_cow_mapping(). The major difference is whether we'll split the page
right now, or postpone it until the next write to each mm. But I think, yes,
maybe I should better still keep the is_cow_mapping() to be explicit.

Thanks,

--
Peter Xu