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

[Hugh Dickins <hugh@veritas.com>]

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

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

>  - 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.

>  - 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.

> 
> 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.

(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.)

And trivial point: please undo those distracting "pte" to "ptep" mods:
if you want to call pte pointers ptep, throughout rmap.c and throughout
mm, that's another patch entirely (which I won't welcome, but others may).

Hugh