Re: [PATCH] migration_entry_wait fix.

[Nick Piggin <nickpiggin-/E1597aS9LT0CCvOHzKKcA@public.gmane.org>]

On Wednesday 18 June 2008 16:04, KAMEZAWA Hiroyuki wrote:
> On Wed, 18 Jun 2008 15:35:57 +1000
>
> Nick Piggin <nickpiggin-/E1597aS9LT0CCvOHzKKcA@public.gmane.org> wrote:
> > On Wednesday 18 June 2008 11:54, KAMEZAWA Hiroyuki wrote:
> > > On Wed, 18 Jun 2008 10:13:49 +0900
> > >
> > > KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org> wrote:
> > > > +	if (!page_cache_get_speculative())
> > > > +		goto out;
> > >
> > > This is obviously buggy....sorry..quilt refresh miss..
> > >
> > > ==
> > > In speculative page cache lookup protocol, page_count(page) is set to 0
> > > while radix-tree modification is going on, truncation, migration,
> > > etc...
> >
> > These tend to all happen while the page is locked, and in particular
> > while the page does not have any references other than the current
> > code path and the pagecache. So no page tables should point to it.
> >
> > So migration_entry_wait should not find pages with a refcount of zero.
> >
> > > While page migration, a page fault to page under migration should wait
> > > unlock_page() and migration_entry_wait() waits for the page from its
> > > pte entry. It does get_page() -> wait_on_page_locked() -> put_page()
> > > now.
> > >
> > > In page migration, page_freeze_refs() -> page_unfreeze_refs() is
> > > called.
> > >
> > > Here, page_unfreeze_refs() expects page_count(page) == 0 and panics
> > > if page_count(page) != 0. To avoid this, we shouldn't touch
> > > page_count() if it is zero. This patch uses
> > > page_cache_get_speculative() to avoid the panic.
> >
> > At any rate, page_cache_get_speculative() should not be used for this
> > purpose, but for when we _really_ don't have any references to a page.
>
> Then, I got NAK. what should I do ?

Well, not nack as such as just wanting to find out a bit more about
how this happens (I'm a little bit slow...)

> (This fix is not related to lock_page() problem.)
>
> If I read your advice correctly, we shouldn't use lock_page() here.
>
> Before speculative page cache, page_table_entry of a page under migration
> has a pte entry which encodes pfn as special pte entry. and wait for the
> end of page migration by lock_page().

What I don't think I understand, is how we can have a page in the
page tables (and with the ptl held) but with a zero refcount... Oh,
it's not actually a page but a migration entry! I'm not quite so
familiar with that code.

Hmm, so we might possibly see a page there that has a zero refcount
due to page_freeze_refs? In which case, I think the direction of you
fix is good. Sorry for my misunderstanding the problem, and thank
you for fixing up my code!

I would ask you to use get_page_unless_zero rather than
page_cache_get_speculative(), because it's not exactly a speculative
reference -- a speculative reference is one where we elevate _count
and then must recheck that the page we have is correct.

Also, please add a comment. It would really be nicer to hide this
transiently-frozen state away from migration_entry_wait, but I can't
see any lock that would easily solve it.

Thanks,
Nick
--
To unsubscribe from this list: send the line "unsubscribe kernel-testers" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html