Re: [PATCH 07/19] mm/fork: Accept huge pfnmap entries

[Peter Xu <peterx@redhat.com>]

On Fri, Aug 09, 2024 at 07:59:58PM +0200, David Hildenbrand wrote:
> On 09.08.24 19:15, Peter Xu wrote:
> > On Fri, Aug 09, 2024 at 06:32:44PM +0200, David Hildenbrand wrote:
> > > On 09.08.24 18:08, Peter Xu wrote:
> > > > Teach the fork code to properly copy pfnmaps for pmd/pud levels.  Pud is
> > > > much easier, the write bit needs to be persisted though for writable and
> > > > shared pud mappings like PFNMAP ones, otherwise a follow up write in either
> > > > parent or child process will trigger a write fault.
> > > > 
> > > > Do the same for pmd level.
> > > > 
> > > > Signed-off-by: Peter Xu <peterx@redhat.com>
> > > > ---
> > > >    mm/huge_memory.c | 27 ++++++++++++++++++++++++---
> > > >    1 file changed, 24 insertions(+), 3 deletions(-)
> > > > 
> > > > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > > > index 6568586b21ab..015c9468eed5 100644
> > > > --- a/mm/huge_memory.c
> > > > +++ b/mm/huge_memory.c
> > > > @@ -1375,6 +1375,22 @@ int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
> > > >    	pgtable_t pgtable = NULL;
> > > >    	int ret = -ENOMEM;
> > > > +	pmd = pmdp_get_lockless(src_pmd);
> > > > +	if (unlikely(pmd_special(pmd))) {
> > > > +		dst_ptl = pmd_lock(dst_mm, dst_pmd);
> > > > +		src_ptl = pmd_lockptr(src_mm, src_pmd);
> > > > +		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
> > > > +		/*
> > > > +		 * No need to recheck the pmd, it can't change with write
> > > > +		 * mmap lock held here.
> > > > +		 */
> > > > +		if (is_cow_mapping(src_vma->vm_flags) && pmd_write(pmd)) {
> > > > +			pmdp_set_wrprotect(src_mm, addr, src_pmd);
> > > > +			pmd = pmd_wrprotect(pmd);
> > > > +		}
> > > > +		goto set_pmd;
> > > > +	}
> > > > +
> > > 
> > > I strongly assume we should be using using vm_normal_page_pmd() instead of
> > > pmd_page() further below. pmd_special() should be mostly limited to GUP-fast
> > > and vm_normal_page_pmd().
> > 
> > One thing to mention that it has this:
> > 
> > 	if (!vma_is_anonymous(dst_vma))
> > 		return 0;
> 
> Another obscure thing in this function. It's not the job of copy_huge_pmd()
> to make the decision whether to copy, it's the job of vma_needs_copy() in
> copy_page_range().
> 
> And now I have to suspect that uffd-wp is broken with this function, because
> as vma_needs_copy() clearly states, we must copy, and we don't do that for
> PMDs. Ugh.
> 
> What a mess, we should just do what we do for PTEs and we will be fine ;)

IIUC it's not a problem: file uffd-wp is different from anonymous, in that
it pushes everything down to ptes.

It means if we skipped one huge pmd here for file, then it's destined to
have nothing to do with uffd-wp, otherwise it should have already been
split at the first attempt to wr-protect.

> 
> Also, we call copy_huge_pmd() only if "is_swap_pmd(*src_pmd) ||
> pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)"
> 
> Would that even be the case with PFNMAP? I suspect that pmd_trans_huge()
> would return "true" for special pfnmap, which is rather "surprising", but
> fortunate for us.

It's definitely not surprising to me as that's the plan.. and I thought it
shoulidn't be surprising to you - if you remember before I sent this one, I
tried to decouple that here with the "thp agnostic" series:

  https://lore.kernel.org/r/20240717220219.3743374-1-peterx@redhat.com

in which you reviewed it (which I appreciated).

So yes, pfnmap on pmd so far will report pmd_trans_huge==true.

> 
> Likely we should be calling copy_huge_pmd() if pmd_leaf() ... cleanup for
> another day.

Yes, ultimately it should really be a pmd_leaf(), but since I didn't get
much feedback there, and that can further postpone this series from being
posted I'm afraid, then I decided to just move on with "taking pfnmap as
THPs".  The corresponding change on this path is here in that series:

https://lore.kernel.org/all/20240717220219.3743374-7-peterx@redhat.com/

@@ -1235,8 +1235,7 @@ copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
 	src_pmd = pmd_offset(src_pud, addr);
 	do {
 		next = pmd_addr_end(addr, end);
-		if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
-			|| pmd_devmap(*src_pmd)) {
+		if (is_swap_pmd(*src_pmd) || pmd_is_leaf(*src_pmd)) {
 			int err;
 			VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
 			err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,

> 
> > 
> > So it's only about anonymous below that.  In that case I feel like the
> > pmd_page() is benign, and actually good.
> 
> Yes, it would likely currently work.
> 
> > 
> > Though what you're saying here made me notice my above check doesn't seem
> > to be necessary, I mean, "(is_cow_mapping(src_vma->vm_flags) &&
> > pmd_write(pmd))" can't be true when special bit is set, aka, pfnmaps.. and
> > if it's writable for CoW it means it's already an anon.
> > 
> > I think I can probably drop that line there, perhaps with a
> > VM_WARN_ON_ONCE() making sure it won't happen.
> > 
> > > 
> > > Again, we should be doing this similar to how we handle PTEs.
> > > 
> > > I'm a bit confused about the "unlikely(!pmd_trans_huge(pmd)" check, below:
> > > what else should we have here if it's not a migration entry but a present
> > > entry?
> > 
> > I had a feeling that it was just a safety belt since the 1st day of thp
> > when Andrea worked that out, so that it'll work with e.g. file truncation
> > races.
> > 
> > But with current code it looks like it's only anonymous indeed, so looks
> > not possible at least from that pov.
> 
> Yes, as stated above, likely broken with UFFD-WP ...
> 
> I really think we should make this code just behave like it would with PTEs,
> instead of throwing in more "different" handling.

So it could simply because file / anon uffd-wp work very differently.

Let me know if you still spot something that is suspicious, but in all
cases I guess we can move on with this series, but maybe if you can find
something I can tackle them together when I decide to go back to the
mremap() issues in the other thread.

Thanks,

-- 
Peter Xu