Re: 2.6.19 file content corruption on ext3

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

On Mon, 2006-12-18 at 12:14 -0800, Linus Torvalds wrote:
> 
> On Mon, 18 Dec 2006, Andrei Popa wrote:
> > 
> > I dropped that patch and added WARN_ON(1), the unified patch is
> > attached.
> > 
> > I got corruption: "Hash check on download completion found bad chunks,
> > consider using "safe_sync"."
> 
> Ok. That is actually _very_ interesting.
> 
> It's interesting because (a) the corruption obviously goes away with the 
> one-liner that effectively disables "page_mkclean_one()".
> 
> So that tells us that yes, it's a PTE dirty bit that matters.
> 
> But at the same time, it's interesting that it still happens when we try 
> to re-add the dirty bit. That would tell me that it's one of two cases:
> 
>  - there is another caller of page cleaning that should have done the same 
>    thing (we could check that by just doing this all _inside_ the 
>    page_mkclean() thing)
> 
> OR:
> 
>  - page_mkclean_one() is simply buggy.
> 
> And I'm starting to wonder about the second case. But it all LOOKS really 
> fine - I can't see anything wrong there (it uses the extremely 
> conservative "ptep_get_and_clear()", and seems to flush everything right 
> too, through "ptep_establish()").

How about this:

we get confused on what PG_dirty tells us, we fall back to pte_dirty,
transfer pte_dirty to PG_dirty and clear pte_dirty. Now it happens
again, however we don't have pte_dirty to fall back to anymore.

This would explain why disabling pte_mkclean() does make it go away and
non of the other tried approaches works.

We really need a way to sort out PG_dirty, independent of pte_dirty.