[1/2] mm/uffd: Fix pte marker when fork() without fork event

  When fork(), dst_vma is not guaranteed to have VM_UFFD_WP even if src may
have it and has pte marker installed.  The warning is improper along with
the comment.  The right thing is to inherit the pte marker when needed, or
keep the dst pte empty.

A vague guess is this happened by an accident when there's the prior patch
to introduce src/dst vma into this helper during the uffd-wp feature got
developed and I probably messed up in the rebase, since if we replace
dst_vma with src_vma the warning & comment it all makes sense too.

Hugetlb did exactly the right here (copy_hugetlb_page_range()).  Fix the
general path.

Reproducer:

https://github.com/xupengfe/syzkaller_logs/blob/main/221208_115556_copy_page_range/repro.c

Cc: <stable@vger.kernel.org> # 5.19+
Fixes: c56d1b62cce8 ("mm/shmem: handle uffd-wp during fork()")
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=216808
Reported-by: Pengfei Xu <pengfei.xu@intel.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
---
 mm/memory.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)
  

On 14.12.22 21:04, Peter Xu wrote:
> When fork(), dst_vma is not guaranteed to have VM_UFFD_WP even if src may
> have it and has pte marker installed.  The warning is improper along with
> the comment.  The right thing is to inherit the pte marker when needed, or
> keep the dst pte empty.
> 
> A vague guess is this happened by an accident when there's the prior patch
> to introduce src/dst vma into this helper during the uffd-wp feature got
> developed and I probably messed up in the rebase, since if we replace
> dst_vma with src_vma the warning & comment it all makes sense too.
> 
> Hugetlb did exactly the right here (copy_hugetlb_page_range()).  Fix the
> general path.
> 
> Reproducer:
> 
> https://github.com/xupengfe/syzkaller_logs/blob/main/221208_115556_copy_page_range/repro.c
> 
> Cc: <stable@vger.kernel.org> # 5.19+
> Fixes: c56d1b62cce8 ("mm/shmem: handle uffd-wp during fork()")
> Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=216808
> Reported-by: Pengfei Xu <pengfei.xu@intel.com>
> Signed-off-by: Peter Xu <peterx@redhat.com>
> ---
>   mm/memory.c | 8 ++------
>   1 file changed, 2 insertions(+), 6 deletions(-)
> 
> diff --git a/mm/memory.c b/mm/memory.c
> index aad226daf41b..032ef700c3e8 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -828,12 +828,8 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
>   			return -EBUSY;
>   		return -ENOENT;
>   	} else if (is_pte_marker_entry(entry)) {
> -		/*
> -		 * We're copying the pgtable should only because dst_vma has
> -		 * uffd-wp enabled, do sanity check.
> -		 */
> -		WARN_ON_ONCE(!userfaultfd_wp(dst_vma));
> -		set_pte_at(dst_mm, addr, dst_pte, pte);
> +		if (userfaultfd_wp(dst_vma))
> +			set_pte_at(dst_mm, addr, dst_pte, pte);
>   		return 0;
>   	}
>   	if (!userfaultfd_wp(dst_vma))

Staring at the code first made me go "what about other PTE markers". I 
then looked into the discussion in patch #2. The fix as is is 
suboptimal, because it

1) Removes the warning which is good, but
2) Silently drops swapin errors now

So it silently breaks something else temporarily ...


I remember, that theoretically we could have multiple markers stored in 
a single PTE marker.

Wouldn't it be cleaner to be able to "clean" specific markers from a PTE 
marker without having to special case on each and everyone? I mean, only 
uffd-wp is really special such that it might disappear for the target.

Something like (pseudocode):

if (!userfaultfd_wp(dst_vma))
	pte_marker_clear_uff_wp(entry);
if (!pte_marker_empty(entry)) {
	pte = make_pte_marker(pte_marker_get(entry));
	set_pte_at(dst_mm, addr, dst_pte, pte);
}

Then this fix would be correct and backport-able even without #2. And it
would work for new types of markers :)


I'd prefer a fix that doesn't break something else temporarily, even if 
the stable backport might require 5 additional minutes to do. So 
squashing #2 into #1 would also work.
  

On Fri, Dec 16, 2022 at 10:04:27AM +0100, David Hildenbrand wrote:
> On 14.12.22 21:04, Peter Xu wrote:
> > When fork(), dst_vma is not guaranteed to have VM_UFFD_WP even if src may
> > have it and has pte marker installed.  The warning is improper along with
> > the comment.  The right thing is to inherit the pte marker when needed, or
> > keep the dst pte empty.
> > 
> > A vague guess is this happened by an accident when there's the prior patch
> > to introduce src/dst vma into this helper during the uffd-wp feature got
> > developed and I probably messed up in the rebase, since if we replace
> > dst_vma with src_vma the warning & comment it all makes sense too.
> > 
> > Hugetlb did exactly the right here (copy_hugetlb_page_range()).  Fix the
> > general path.
> > 
> > Reproducer:
> > 
> > https://github.com/xupengfe/syzkaller_logs/blob/main/221208_115556_copy_page_range/repro.c
> > 
> > Cc: <stable@vger.kernel.org> # 5.19+
> > Fixes: c56d1b62cce8 ("mm/shmem: handle uffd-wp during fork()")
> > Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=216808
> > Reported-by: Pengfei Xu <pengfei.xu@intel.com>
> > Signed-off-by: Peter Xu <peterx@redhat.com>
> > ---
> >   mm/memory.c | 8 ++------
> >   1 file changed, 2 insertions(+), 6 deletions(-)
> > 
> > diff --git a/mm/memory.c b/mm/memory.c
> > index aad226daf41b..032ef700c3e8 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -828,12 +828,8 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
> >   			return -EBUSY;
> >   		return -ENOENT;
> >   	} else if (is_pte_marker_entry(entry)) {
> > -		/*
> > -		 * We're copying the pgtable should only because dst_vma has
> > -		 * uffd-wp enabled, do sanity check.
> > -		 */
> > -		WARN_ON_ONCE(!userfaultfd_wp(dst_vma));
> > -		set_pte_at(dst_mm, addr, dst_pte, pte);
> > +		if (userfaultfd_wp(dst_vma))
> > +			set_pte_at(dst_mm, addr, dst_pte, pte);
> >   		return 0;
> >   	}
> >   	if (!userfaultfd_wp(dst_vma))
> 
> Staring at the code first made me go "what about other PTE markers". I then
> looked into the discussion in patch #2. The fix as is is suboptimal, because
> it
> 
> 1) Removes the warning which is good, but
> 2) Silently drops swapin errors now
> 
> So it silently breaks something else temporarily ...
> 
> 
> I remember, that theoretically we could have multiple markers stored in a
> single PTE marker.
> 
> Wouldn't it be cleaner to be able to "clean" specific markers from a PTE
> marker without having to special case on each and everyone? I mean, only
> uffd-wp is really special such that it might disappear for the target.

Quotting the commit message in patch 2:

  Currently there is a priority difference between the uffd-wp bit and the
  swapin error entry, in which the swapin error always has higher priority
  (e.g. we don't need to wr-protect a swapin error pte marker).

  If there will be a 3rd bit introduced, we'll probably need to consider a
  more involved approach so we may need to start operate on the bits.
  Let's leave that for later.

I actually started the fix with something like that, but I noticed it's not
needed to add more code if there's no 3rd bit introduced so I dropped that.
I decided to go the simpler change approach and leave that for later.

> 
> Something like (pseudocode):
> 
> if (!userfaultfd_wp(dst_vma))
> 	pte_marker_clear_uff_wp(entry);
> if (!pte_marker_empty(entry)) {
> 	pte = make_pte_marker(pte_marker_get(entry));
> 	set_pte_at(dst_mm, addr, dst_pte, pte);
> }
> 
> Then this fix would be correct and backport-able even without #2. And it
> would work for new types of markers :)

When that comes, we may need one set_pte_marker_at() taking care of empty
pte markers, otherwise there can be a lot of such check.

> 
> 
> I'd prefer a fix that doesn't break something else temporarily, even if the
> stable backport might require 5 additional minutes to do. So squashing #2
> into #1 would also work.

The thing is whether do we care about someone: (1) explicitly checkout at
the commit of patch 1, then (2) runs the kernel, hit a swapnin error, (3)
fork(), and (4) access the swapin error page in the child.

To me I don't care even starting from (1).. because it really shouldn't
happen at all in any serious environment.

The other reason is these are indeed two issues to solve.  Even if by
accident we kept the swapin error in old code we'll probably dump an
warning which is not wanted either.  It's not something someone will really
get benefit from..

So like many other places, I don't have a strong opinion, but personally I
prefer the current approach.

Andrew, let me know if you want me to repost with a squashed version.

Thanks,
  

>>
>> Wouldn't it be cleaner to be able to "clean" specific markers from a PTE
>> marker without having to special case on each and everyone? I mean, only
>> uffd-wp is really special such that it might disappear for the target.
> 
> Quotting the commit message in patch 2:
> 
>    Currently there is a priority difference between the uffd-wp bit and the
>    swapin error entry, in which the swapin error always has higher priority
>    (e.g. we don't need to wr-protect a swapin error pte marker).
> 
>    If there will be a 3rd bit introduced, we'll probably need to consider a
>    more involved approach so we may need to start operate on the bits.
>    Let's leave that for later.
> 
> I actually started the fix with something like that, but I noticed it's not
> needed to add more code if there's no 3rd bit introduced so I dropped that.
> I decided to go the simpler change approach and leave that for later.

Okay, makes sense.

> 
>>
>> Something like (pseudocode):
>>
>> if (!userfaultfd_wp(dst_vma))
>> 	pte_marker_clear_uff_wp(entry);
>> if (!pte_marker_empty(entry)) {
>> 	pte = make_pte_marker(pte_marker_get(entry));
>> 	set_pte_at(dst_mm, addr, dst_pte, pte);
>> }
>>
>> Then this fix would be correct and backport-able even without #2. And it
>> would work for new types of markers :)
> 
> When that comes, we may need one set_pte_marker_at() taking care of empty
> pte markers, otherwise there can be a lot of such check.

Right. In the future it might be cleaner.

> 
>>
>>
>> I'd prefer a fix that doesn't break something else temporarily, even if the
>> stable backport might require 5 additional minutes to do. So squashing #2
>> into #1 would also work.
> 
> The thing is whether do we care about someone: (1) explicitly checkout at
> the commit of patch 1, then (2) runs the kernel, hit a swapnin error, (3)
> fork(), and (4) access the swapin error page in the child.

I'm more concerned about backports, when one backports #1 but not #2. In 
theory, patch #2 fixes patch #1, because that introduced IMHO a real 
regression -- a possible memory corruption when discarding a hwpoison 
marker. Warnings are not nice but at least indicate that something needs 
a second look.

> 
> To me I don't care even starting from (1).. because it really shouldn't
> happen at all in any serious environment.
> 
> The other reason is these are indeed two issues to solve.  Even if by
> accident we kept the swapin error in old code we'll probably dump an
> warning which is not wanted either.  It's not something someone will really
> get benefit from..
> 
> So like many other places, I don't have a strong opinion, but personally I
> prefer the current approach.


Me neither, two patches just felt more complicated than it should be.

Anyhow, the final code change LGTM.
  

On Fri, Dec 16, 2022 at 04:57:33PM +0100, David Hildenbrand wrote:
> I'm more concerned about backports, when one backports #1 but not #2. In
> theory, patch #2 fixes patch #1, because that introduced IMHO a real
> regression -- a possible memory corruption when discarding a hwpoison
> marker. Warnings are not nice but at least indicate that something needs a
> second look.

Note that backporting patch 1 only is exactly what I wanted to do here - it
means his/her tree should not have the swapin error pte markers at all.

The swapin error pte marker change only existed for a few days in Linus's
tree, so no one should be backporting patch 2.

Thanks,
  

On 16.12.22 17:24, Peter Xu wrote:
> On Fri, Dec 16, 2022 at 04:57:33PM +0100, David Hildenbrand wrote:
>> I'm more concerned about backports, when one backports #1 but not #2. In
>> theory, patch #2 fixes patch #1, because that introduced IMHO a real
>> regression -- a possible memory corruption when discarding a hwpoison
>> marker. Warnings are not nice but at least indicate that something needs a
>> second look.
> 
> Note that backporting patch 1 only is exactly what I wanted to do here - it
> means his/her tree should not have the swapin error pte markers at all.
> 
> The swapin error pte marker change only existed for a few days in Linus's
> tree, so no one should be backporting patch 2.

Right, and these patches are supposed to land in 6.2 as well. Makes 
sense to me then.

Especially, the other parts in patch #2 are worth being in a separate patch.

Acked-by: David Hildenbrand <david@redhat.com>