Re: [PATCH] mm/pgtable: return null if no ptl in __pte_offset_map_lock

[Hugh Dickins]

On Wed, 15 Nov 2023, Matthew Wilcox wrote:
> On Wed, Nov 15, 2023 at 06:05:30PM +0200, José Pekkarinen wrote:
> 
> > > I don't think we should be changing ptlock_ptr().
> > 
> >     This is where the null ptr dereference originates, so the only
> > alternative I can think of is to protect the life cycle of the ptdesc
> > to prevent it to die between the pte check and the spin_unlock of
> > __pte_offset_map_lock. Would that work for you?

Thanks for pursuing this, José, but I agree with Matthew: I don't
think your patch is right at all.  The change in ptlock_ptr() did not
make sense to me, and the change in __pte_offset_map_lock() leaves us
still wondering what has gone wrong (and misses an rcu_read_unlock()).

You mentioned "I tested the syzbot reproducer in x86 and it doesn't
produce this kasan report anymore": are you saying that you were able
to reproduce the issue on x86 (without your patch)?  That would be very
interesting (and I think would disprove my hypothesis below).  I ought
to try on x86 if you managed to reproduce on it, but it's not worth
the effort if you did not.  If you have an x86 stack and registers,
please show (though I'm uncertain how much KASAN spoils the output).

> 
> Ah!  I think I found the problem.
> 
> If ALLOC_SPLIT_PTLOCKS is not set, there is no problem as ->ptl
> is embedded in the struct page.  But if ALLOC_SPLIT_PTLOCKS is set
> (eg you have LOCKDEP enabled), we can _return_ a NULL pointer from
> ptlock_ptr.  The NULL pointer dereference isn't in ptlock_ptr(), it's
> in __pte_offset_map_lock().
> 
> So, how to solve this?  We can't just check the ptl against NULL; the
> memory that ptl points to may have been freed.  We could grab a reference
> to the pmd_page, possibly conditionally on ALLOC_SPLIT_LOCK being set,
> but that will slow down everything.  We could make page_ptl_cachep
> SLAB_TYPESAFE_BY_RCU, preventing the memory from being freed (even if
> the lock might not be associated with this page any more).

But I don't think that's quite right either: or, I hope it's not right,
because it would say that all my business of rcu_read_lock() and
pte_free_defer() etc was a failing waste of everyone's time: if the
page table (and/or its lock) can be freed while someone is in that RCU-
protected area, then I've simply got it wrong.

What I thought, when yesterday's flurry of syzbot messages came in,
was perhaps the problem is at the other end - when the new page table
is allocated, rather than when it's being freed: a barrier missing there?

But pmd_install() does have an smp_wmb(), with a (suspiciously) long
comment about why no smp_rmb()s are needed, except on alpha.

I'm not much good on barriers, but the thought now in my mind is that
that comment is not quite accurate: that at the bottom level an smp_rmb()
is (very seldom!) needed - not just to avoid a NULL (or previous garbage)
ptl pointer in the ALLOC_SPLIT_LOCK case, but to make sure that the ptl
is visibly correctly initialized (or would spin_lock on it achieve that?);
and that what pte_offset_map() points to is visibly empty.  (I'm imagining
stale cache lines left behind on the oopsing CPU, which need to be
refetched after pmdval has been read.)

And that this issue has, strictly speaking, always been there, but in
practice never a problem, because of mmap_lock held for write while (or
prior to) freeing page table, and racers holding mmap_lock for read
(vma lock not changing the calculus); but collapse_pte_mapped_thp()
can now be done with mmap_lock for read, letting those racers in
(and maybe your filemap_map_pages() has helped speed these races up -
any blame I can shift on to you, the better ;)

But I may well be spouting nonsense.  And even if I'm right, is adding
an smp_rmb() in there too high a price to pay on some architectures?

I did make an attempt to reproduce on an arm64, but there's too much
more I'd need to muck around with to get it working right.  Let's ask for
a syz test on top of v6.7-rc1 - hmm, how do I tell syzbot that it's arm64
that it needs to try on?  I hope it gets that from the Cc'ed syz number.

#syz test: git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git b85ea95d086471afb4ad062012a4d73cd328fa86

[PATCH] mm/pgtable: smp_rmb() to match smp_wmb() in pmd_install()

Not-Yet-Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
---
 mm/memory.c          | 2 ++
 mm/pgtable-generic.c | 5 +++++
 2 files changed, 7 insertions(+)

diff --git a/mm/memory.c b/mm/memory.c
index 1f18ed4a5497..8939357f1509 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -425,6 +425,8 @@ void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
 		 * being the notable exception) will already guarantee loads are
 		 * seen in-order. See the alpha page table accessors for the
 		 * smp_rmb() barriers in page table walking code.
+		 *
+		 * See __pte_offset_map() for the smp_rmb() at the pte level.
 		 */
 		smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 		pmd_populate(mm, pmd, *pte);
diff --git a/mm/pgtable-generic.c b/mm/pgtable-generic.c
index 4fcd959dcc4d..3330b666e9c3 100644
--- a/mm/pgtable-generic.c
+++ b/mm/pgtable-generic.c
@@ -297,6 +297,11 @@ pte_t *__pte_offset_map(pmd_t *pmd, unsigned long addr, pmd_t *pmdvalp)
 		pmd_clear_bad(pmd);
 		goto nomap;
 	}
+	/*
+	 * Pair with the smp_wmb() in pmd_install(): make sure that the
+	 * page table lock and page table contents are visibly initialized.
+	 */
+	smp_rmb();
 	return __pte_map(&pmdval, addr);
 nomap:
 	rcu_read_unlock();
-- 
2.35.3