Re: [PATCH] mm: check the return value of lookup_page_ext for all call sites

[Shi, Yang]

On 6/1/2016 10:00 PM, Minchan Kim wrote:
> On Wed, Jun 01, 2016 at 01:40:48PM -0700, Shi, Yang wrote:
> > On 5/29/2016 11:11 PM, Minchan Kim wrote:
> > > On Fri, May 27, 2016 at 11:16:41AM -0700, Shi, Yang wrote:
> > >
> > > <snip>
> > >
> > > > > If we goes this way, how to guarantee this race?
> > > >
> > > > Thanks for pointing out this. It sounds reasonable. However, this
> > > > should be only possible to happen on 32 bit since just 32 bit
> > > > version page_is_idle() calls lookup_page_ext(), it doesn't do it on
> > > > 64 bit.
> > > >
> > > > And, such race condition should exist regardless of whether DEBUG_VM
> > > > is enabled or not, right?
> > > >
> > > > rcu might be good enough to protect it.
> > > >
> > > > A quick fix may look like:
> > > >
> > > > diff --git a/include/linux/page_idle.h b/include/linux/page_idle.h
> > > > index 8f5d4ad..bf0cd6a 100644
> > > > --- a/include/linux/page_idle.h
> > > > +++ b/include/linux/page_idle.h
> > > > @@ -77,8 +77,12 @@ static inline bool
> > > > test_and_clear_page_young(struct page *page)
> > > > static inline bool page_is_idle(struct page *page)
> > > > {
> > > >        struct page_ext *page_ext;
> > > > +
> > > > +       rcu_read_lock();
> > > >        page_ext = lookup_page_ext(page);
> > > > +       rcu_read_unlock();
> > > > +
> > > > 	if (unlikely(!page_ext))
> > > >                return false;
> > > >
> > > > diff --git a/mm/page_ext.c b/mm/page_ext.c
> > > > index 56b160f..94927c9 100644
> > > > --- a/mm/page_ext.c
> > > > +++ b/mm/page_ext.c
> > > > @@ -183,7 +183,6 @@ struct page_ext *lookup_page_ext(struct page *page)
> > > > {
> > > >        unsigned long pfn = page_to_pfn(page);
> > > >        struct mem_section *section = __pfn_to_section(pfn);
> > > > -#if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING)
> > > >        /*
> > > >         * The sanity checks the page allocator does upon freeing a
> > > >         * page can reach here before the page_ext arrays are
> > > > @@ -195,7 +194,7 @@ struct page_ext *lookup_page_ext(struct page *page)
> > > >         */
> > > >        if (!section->page_ext)
> > > >                return NULL;
> > > > -#endif
> > > > +
> > > >        return section->page_ext + pfn;
> > > > }
> > > >
> > > > @@ -279,7 +278,8 @@ static void __free_page_ext(unsigned long pfn)
> > > >                return;
> > > >        base = ms->page_ext + pfn;
> > > >        free_page_ext(base);
> > > > -       ms->page_ext = NULL;
> > > > +       rcu_assign_pointer(ms->page_ext, NULL);
> > > > +       synchronize_rcu();
> > >
> > > How does it fix the problem?
> > > I cannot understand your point.
> >
> > Assigning NULL pointer to page_Ext will be blocked until
> > rcu_read_lock critical section is done, so the lookup and writing
> > operations will be serialized. And, rcu_read_lock disables preempt
> > too.
>
> I meant your rcu_read_lock in page_idle should cover test_bit op.

Yes, definitely. Thanks for catching it.

> One more thing, you should use rcu_dereference.

I will check which one is the best since I saw some use rcu_assign_pointer.

> As well, please cover memory onlining case I mentioned in another
> thread as well as memory offlining.

I will look into it too.

Thanks,
Yang

> Anyway, to me, every caller of page_ext should prepare lookup_page_ext
> can return NULL anytime and they should use rcu_read_[un]lock, which
> is not good. :(