lkml.org 
[lkml]   [2006]   [Dec]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: 2.6.19 file content corruption on ext3
From
Date
On Mon, 2006-12-18 at 10:03 -0800, Linus Torvalds wrote:
> Andrei,
> could you try Peter's patch (on top of Andrew's patch - it depends on
> it, and wouldn't work on an unmodified -git kernel, but add the WARN_ON()
> I mention in this email? You seem to be able to reproduce this easily..
> Thanks)

I finally beat yum into submission and I hope to have rtorrent compiled
shortly.

> On Mon, 18 Dec 2006, Peter Zijlstra wrote:
> >
> > This should be safe; page_mkclean walks the rmap and flips the pte's
> > under the pte lock and records the dirty state while iterating.
> > Concurrent faults will either do set_page_dirty() before we get around
> > to doing it or vice versa, but dirty state is not lost.
>
> Ok, I really liked this patch, but the more I thought about it, the more I
> started to doubt the reasons for liking it.
>
> I think we have some core fundamental problem here that this patch is
> needed at all.

I agree, but I suspect this is like the buffered write deadlock Nick is
working on, in that it will require some proper filesystem surgery to
get right. Having the kernel working in the meantime has my
preference ;-)

> So let's think about this: we apparently have two cases of
> "clear_page_dirty()":
>
> - the one that really wants to clear the bit unconditionally (Andrew
> calls this the "must_clean_ptes" case, which I personally find to be a
> really confusing name, but whatever)

I'm probably worse with names so I'm not even going to try and fix that.

> - the other case. The case that doesn't want to really clear the pte
> dirty bits.
>
> and I thought your patch made sense, because it saved away the pte state
> in the page dirty state, and that matches my mental model, but the more I
> think about it, the less sense that whole "the other case" situation makes
> AT ALL.
>
> Why does "the other case" exist at all? If you want to clear the dirty
> page flag, what is _ever_ the reason for not wanting to drop PTE dirty
> information? In other words, what possible reason can there ever be for
> saying "I want this page to be clean", while at the same time saying "but
> if it was dirty in the page tables, don't forget about that state".

I have tried to get my head around this, and have so far failed. Andrews
mail with the patch (great-grandparent to this mail) was the one that
made most sense explaining it afaics.

> So I absolutely detested Andrew's original patch, because that one made
> zero sense at all even from a code standpoint. With your patch on top, it
> all suddenly makes sense: at least you don't just leave dirty pages in the
> PTE's with a "struct page" that is marked clean, and the end result is
> undeniably at least _consistent_.
>
> So Andrew's patch I can't stand, because the whole point of it seems to be
> to leave the system in an inconsistent state (dirty in the pte's but
> marked "clean"), and if we want to have that state, then we should just
> revert _everything_ to the 2.6.18 situation, and not play these games at
> all.
>
> Andrew's patch with your patch on top makes me happy, because now we're
> at least honoring all the basic rules (we don't get into an inconsistent
> state), so on a local level it all makes sense. HOWEVER, I then don't
> actually understand how it could ever actually make sense to ask for
> "please clean the page, but don't actually clean it".

Somehow it looses track of actual page content dirtyness when it does
the page buffer game.

Is this because page buffers are used to do sub-page sized writes
without RMW cycles?

Cannot this case be avoided when the page is mapped, because at that
point the whole page will be resident anyway.

> So _I_ think that we should add a honking huge WARN_ON() for this case.
> Ie, do your patch, but instead of re-dirtying the page:
>
> + if (!must_clean_ptes && cleaned)
> + set_page_dirty(page);
>
> we would do
>
> + if (!must_clean_ptes && cleaned) {
> + WARN_ON(1);
> + set_page_dirty(page);
> + }
>
> and ask the people who see this problem to see if they get the WARN_ON()
> (assuming it _fixes_ their data corruption).
>
> Because whoever calls "clean_dirty_page()" without actually wanting to
> clean the PTE's is really a bug: those dirty PTE's had better not exist.
>
> Or maybe the WARN_ON() just points out _why_ somebody would want to do
> something this insane. Right now I just can't see why it's a valid thing
> to do.

Maybe, but I think Nick's mail here:
http://lkml.org/lkml/2006/12/18/59

shows a trace like that. I'm guessing that if we do the WARN_ON() some
folks might get a lot of output, perhaps WARN_ON_ONCE() ?

-
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/

\
 
 \ /
  Last update: 2006-12-18 19:31    [from the cache]
©2003-2011 Jasper Spaans