Re: hugepage related lockdep trace.

[Michal Hocko]

On Mon 29-07-13 17:24:53, Minchan Kim wrote:
> Hi Michal,
> 
> On Thu, Jul 25, 2013 at 03:30:40PM +0200, Michal Hocko wrote:
> > On Wed 24-07-13 11:44:28, Minchan Kim wrote:
> > > On Tue, Jul 23, 2013 at 04:01:20PM +0200, Michal Hocko wrote:
> > > > On Fri 19-07-13 09:13:03, Minchan Kim wrote:
> > > > > On Thu, Jul 18, 2013 at 11:12:24PM +0530, Aneesh Kumar K.V wrote:
> > > > [...]
> > > > > > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> > > > > > index 83aff0a..2cb1be3 100644
> > > > > > --- a/mm/hugetlb.c
> > > > > > +++ b/mm/hugetlb.c
> > > > > > @@ -3266,8 +3266,8 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
> > > > > >  		put_page(virt_to_page(spte));
> > > > > >  	spin_unlock(&mm->page_table_lock);
> > > > > >  out:
> > > > > > -	pte = (pte_t *)pmd_alloc(mm, pud, addr);
> > > > > >  	mutex_unlock(&mapping->i_mmap_mutex);
> > > > > > +	pte = (pte_t *)pmd_alloc(mm, pud, addr);
> > > > > >  	return pte;
> > > > > 
> > > > > I am blind on hugetlb but not sure it doesn't break eb48c071.
> > > > > Michal?
> > > > 
> > > > Well, it is some time since I debugged the race and all the details
> > > > vanished in the meantime. But this part of the changelog suggests that
> > > > this indeed breaks the fix:
> > > > "
> > > >     This patch addresses the issue by moving pmd_alloc into huge_pmd_share
> > > >     which guarantees that the shared pud is populated in the same critical
> > > >     section as pmd.  This also means that huge_pte_offset test in
> > > >     huge_pmd_share is serialized correctly now which in turn means that the
> > > >     success of the sharing will be higher as the racing tasks see the pud
> > > >     and pmd populated together.
> > > > "
> > > > 
> > > > Besides that I fail to see how moving pmd_alloc down changes anything.
> > > > Even if pmd_alloc triggered reclaim then we cannot trip over the same
> > > > i_mmap_mutex as hugetlb pages are not reclaimable because they are not
> > > > on the LRU.
> > > 
> > > I thought we could map some part of binary with normal page and other part
> > > of the one with MAP_HUGETLB so that a address space could have both normal
> > > page and HugeTLB page. Okay, it's impossible so HugeTLB pages are not on LRU.
> > > Then, above lockdep warning is totally false positive.
> > > Best solution is avoiding pmd_alloc with holding i_mmap_mutex but we need it
> > > to fix eb48c071 so how about this if we couldn't have a better idea?
> > 
> > Shouldn't we rather use a hugetlb specific lock_class_key. I am not
> > familiar with lockdep much but something like bellow should do the
> > trick?
> 
> Looks good to me.
> Could you resend it with formal patch with Ccing Peter for Dave to confirm it?
> Below just a nitpick.

I would have posted it already but I have to confess I am really not
familiar with lockdep and what is the good way to fix such a false
positive.

Peter, for you context the lockdep splat has been reported
here: https://lkml.org/lkml/2013/7/17/381

Minchan has proposed to workaround it by using SINGLE_DEPTH_NESTING
https://lkml.org/lkml/2013/7/23/812

my idea was to use a separate class key for hugetlb as it is quite
special in many ways:
https://lkml.org/lkml/2013/7/25/277

What is the preferred way of fixing such an issue?

Thanks!
-- 
Michal Hocko
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/