Re: [PATCH] dax: Fix deadlock in dax_lock_mapping_entry()

[Dan Williams <dan.j.williams@intel.com>]

On Fri, Oct 5, 2018 at 2:56 AM Jan Kara <jack@suse.cz> wrote:
>
> On Thu 04-10-18 21:28:14, Dan Williams wrote:
> > On Thu, Oct 4, 2018 at 9:01 PM Dan Williams <dan.j.williams@intel.com> wrote:
> > >
> > > On Thu, Oct 4, 2018 at 7:52 PM Matthew Wilcox <willy@infradead.org> wrote:
> > > >
> > > > On Thu, Oct 04, 2018 at 06:57:52PM -0700, Dan Williams wrote:
> > > > > On Thu, Oct 4, 2018 at 9:27 AM Jan Kara <jack@suse.cz> wrote:
> > > > > >
> > > > > > On Thu 27-09-18 11:22:22, Dan Williams wrote:
> > > > > > > On Thu, Sep 27, 2018 at 6:41 AM Jan Kara <jack@suse.cz> wrote:
> > > > > > > >
> > > > > > > > On Thu 27-09-18 06:28:43, Matthew Wilcox wrote:
> > > > > > > > > On Thu, Sep 27, 2018 at 01:23:32PM +0200, Jan Kara wrote:
> > > > > > > > > > When dax_lock_mapping_entry() has to sleep to obtain entry lock, it will
> > > > > > > > > > fail to unlock mapping->i_pages spinlock and thus immediately deadlock
> > > > > > > > > > against itself when retrying to grab the entry lock again. Fix the
> > > > > > > > > > problem by unlocking mapping->i_pages before retrying.
> > > > > > > > >
> > > > > > > > > It seems weird that xfstests doesn't provoke this ...
> > > > > > > >
> > > > > > > > The function currently gets called only from mm/memory-failure.c. And yes,
> > > > > > > > we are lacking DAX hwpoison error tests in fstests...
> > > > > > >
> > > > > > > I have an item on my backlog to port the ndctl unit test that does
> > > > > > > memory_failure() injection vs ext4 over to fstests. That said I've
> > > > > > > been investigating a deadlock on ext4 caused by this test. When I saw
> > > > > > > this patch I hoped it was root cause, but the test is still failing
> > > > > > > for me. Vishal is able to pass the test on his system, so the failure
> > > > > > > mode is timing dependent. I'm running this patch on top of -rc5 and
> > > > > > > still seeing the following deadlock.
> > > > > >
> > > > > > I went through the code but I don't see where the problem could be. How can
> > > > > > I run that test? Is KVM enough or do I need hardware with AEP dimms?
> > > > >
> > > > > KVM is enough... however, I have found a hack that makes the test pass:
> > > > >
> > > > > diff --git a/mm/filemap.c b/mm/filemap.c
> > > > > index 52517f28e6f4..d7f035b1846e 100644
> > > > > --- a/mm/filemap.c
> > > > > +++ b/mm/filemap.c
> > > > > @@ -1668,6 +1668,9 @@ unsigned find_get_entries(struct address_space *mapping,
> > > > >                         goto repeat;
> > > > >                 }
> > > > >  export:
> > > > > +               if (iter.index < start)
> > > > > +                       continue;
> > > > > +
> > > > >                 indices[ret] = iter.index;
> > > > >                 entries[ret] = page;
> > > > >                 if (++ret == nr_entries)
> > > > >
> > > > > Is this a radix bug? I would never expect:
> > > > >
> > > > >     radix_tree_for_each_slot(slot, &mapping->i_pages, &iter, start)
> > > > >
> > > > > ...to return entries with an index < start. Without that change above
> > > > > we loop forever because dax_layout_busy_page() can't make forward
> > > > > progress. I'll dig into the radix code tomorrow, but maybe someone
> > > > > else will be me to it overnight.
> > > >
> > > > If 'start' is within a 2MB entry, iter.index can absolutely be less
> > > > than start.  I forget exactly what the radix tree code does, but I think
> > > > it returns index set to the canonical/base index of the entry.
> > >
> > > Ok, that makes sense. Then the bug is in dax_layout_busy_page() which
> > > needs to increment 'index' by the entry size. This might also explain
> > > why not every run sees it because you may get lucky and have a 4K
> > > entry.
> >
> > Hmm, no 2MB entry here.
> >
> > We go through the first find_get_entries and export:
> >
> >     export start: 0x0 index: 0x0 page: 0x822000a
> >     export start: 0x0 index: 0x200 page: 0xcc3801a
> >
> > Then dax_layout_busy_page sets 'start' to 0x201, and find_get_entries returns:
> >
> >     export start: 0x201 index: 0x200 page: 0xcc3801a
> >
> > ...forevermore.
>
> Are you sure there's not 2MB entry starting at index 0x200? Because if
> there was, we'd get infinite loop exactly as you describe in
> dax_layout_busy_page()

My debug code was buggy, these are 2MB entries, I'll send out a fix.

> AFAICT. And it seems to me lot of other places
> iterating over entries are borked in a similar way as they all assume that
> doing +1 to the current index is guaranteeing them forward progress. Now
> actual breakage resulting from this is limited as only DAX uses multiorder
> entries and thus not many of these iterators actually ever get called for
> radix tree with multiorder entries (e.g. tmpfs still inserts every 4k subpage
> of THP into the radix tree and iteration functions usually handles
> tail subpages in a special way). But this would really deserve larger
> cleanup.

Yeah, it's a subtle detail waiting to trip up new multi-order-radix
users. The shift reported in the iterator is 6 in this case.