Re: 2.6.19 file content corruption on ext3

[Nick Piggin <nickpiggin@yahoo.com.au>]

Andrew Morton wrote:
> On Mon, 18 Dec 2006 15:51:52 +1100
> Nick Piggin <nickpiggin@yahoo.com.au> wrote:
>
>
> > I think the problem Andrew identified is real.
>
>
> I don't.  In fact I don't think I described any problem (well, I tried to,
> but then I contradicted myself).

By saying that there shouldn't be any dirty ptes if there are no
dirty buffers? But in that case the _page_ shouldn't be dirty either,
so that clear_page_dirty would be redundant. But presumably it isn't.

> Six hours here of fsx-linux plus high memory pressure on SMP on 1k
> blocksize ext3, mainline.  Zero failures.  It's unlikely that this testing
> would pass, yet people running normal workloads are able to easily trigger
> failures.  I suspect we're looking in the wrong place.

Yes I could believe it the corruption is caused by something else
completely.

> > The issue is the disconnect between the pte dirtiness and a filesystem
> > bringing buffers clean.
>
>
> Really?  The dirtying direction goes pte_dirty->PG_dirty->BH_Dirty and the
> cleaning direction goes !BH_Dirty->!PG_dirty->!pte_dirty.  That's pretty
> simple, setting aside races.
>
> In the try_to_free_buffers case there's a large time inverval between
> !BH_Dirty and !PG_dirty, but that shouldn't affect anything.

After try_to_free_buffers detaches the buffers from the page, a
pagefault can come in, and mark the pte writeable, then set_page_dirty
(which finds no buffers, so only sets PG_dirty).

The page can now get dirtied through this mapping.

try_to_free_buffers then goes on to clean the page and ptes.

I really thought you were the one who identified this race, and I didn't
see where you showed it is safe.

It may be very unlikely with small SMPs, but less so with preempt. All
we have to do is preempt at spin_unlock in try_to_free_buffers AFAIKS.
Were you testing with preempt?

--
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com 
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/