Re: [PATCH] mm: fix page_mkclean_one (was: 2.6.19 file content corruption on ext3)

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Wed, 2006-12-20 at 13:00 +0000, Hugh Dickins wrote:
> On Wed, 20 Dec 2006, Peter Zijlstra wrote:
> > 
> > fix page_mkclean_one()
> 
> Congratulations on getting to the bottom of it, Peter (if you have:
> I haven't digested enough of the thread to tell).

Well, I thought I understood, you just shattered that.

>   I'm mostly offline at
> present, no time for dialogue, I'll throw out a few remarks and run...

I wondered where you were ;-) Enjoy your time away from the computer.

> > 
> > it had several issues:
> >  - it failed to flush the cache
> 
> It's unclear to me why it should need to flush the cache, but I don't
> know much about that, and mprotect does flush the cache in advance -
> I think others will tell you that if it does need to be flushed,

I was still thinking about why exactly, but indeed since mprotect does I
thought it prudent to also do it.

> it must
> be flushed while there's still a valid pte (on some arches at least).

Ah, good point, makes sense I guess.

> >  - it failed to flush the tlb
> 
> Eh?  It flushed the TLB inside ptep_establish, didn't it?
> I guess you mean you've found a race before it flushed the TLB.

Hmm, quite right indeed. I missed that. So moving the flush inside the
pte cleared section closed a race. It seems I must have a long hard look
at these architecture manuals...

> >  - it failed to do s390 (s390 guys, please verify this is now correct)
> 
> Hmm, I thought we cleared it with them back at the time.

/me queries mail folder...
can't seem to find it.

> > 
> > Also, clear in a loop to ensure SMP safeness as suggested by Arjan.
> 
> Yikes.  Well, please compare with mprotect's change_pte_range.  I think
> I took that as the relevant standard when checking your implementation,
> and back then satisfied myself that what you were doing was equivalent.
> If page_mkclean_one is now agreed to be significantly defective, then
> I suspect change_pte_range is also; perhaps others too.

Arjan argued that mprotect and msync would mostly race with themselves
in userspace. 

> (But I haven't found time to do more than skim through the thread,
> I've not thought through the issues at all: I am surprised that it's
> now found defective, we looked at it long and hard back then.)

---

page_mkclean_one() fix

it had several issues:
 - it failed to flush the cache
 - a race wrt tlb flushing
 - it failed to do s390 (s390 guys, please verify this is now correct)

Also, clear in a loop to ensure SMP safeness as suggested by Arjan.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 mm/rmap.c |   23 +++++++++++++----------
 1 file changed, 13 insertions(+), 10 deletions(-)

Index: linux-2.6/mm/rmap.c
===================================================================
--- linux-2.6.orig/mm/rmap.c
+++ linux-2.6/mm/rmap.c
@@ -432,7 +432,7 @@ static int page_mkclean_one(struct page 
 {
 	struct mm_struct *mm = vma->vm_mm;
 	unsigned long address;
-	pte_t *pte, entry;
+	pte_t *pte;
 	spinlock_t *ptl;
 	int ret = 0;
 
@@ -444,17 +444,20 @@ static int page_mkclean_one(struct page 
 	if (!pte)
 		goto out;
 
-	if (!pte_dirty(*pte) && !pte_write(*pte))
-		goto unlock;
+	while (pte_dirty(*pte) || pte_write(*pte)) {
+		pte_t entry;
 
-	entry = ptep_get_and_clear(mm, address, pte);
-	entry = pte_mkclean(entry);
-	entry = pte_wrprotect(entry);
-	ptep_establish(vma, address, pte, entry);
-	lazy_mmu_prot_update(entry);
-	ret = 1;
+		flush_cache_page(vma, address, pte_pfn(*pte));
+		entry = ptep_get_and_clear(mm, address, pte);
+		flush_tlb_page(vma, address);
+		(void)page_test_and_clear_dirty(page); /* do the s390 thing */
+		entry = pte_wrprotect(entry);
+		entry = pte_mkclean(entry);
+		set_pte_at(vma, address, pte, entry);
+		lazy_mmu_prot_update(entry);
+		ret = 1;
+	}
 
-unlock:
 	pte_unmap_unlock(pte, ptl);
 out:
 	return ret;