Re: Infinite looping observed in __offline_pages

[John Allen]

On Wed, Jul 25, 2018 at 10:03:36PM +0200, Michal Hocko wrote:
> On Wed 25-07-18 13:11:15, John Allen wrote:
> [...]
> > Does a failure in do_migrate_range indicate that the range is unmigratable
> > and the loop in __offline_pages should terminate and goto failed_removal? Or
> > should we allow a certain number of retrys before we
> > give up on migrating the range?
>
> Unfortunatelly not. Migration code doesn't tell a difference between
> ephemeral and permanent failures. We are relying on
> start_isolate_page_range to tell us this. So the question is, what kind
> of page is not migratable and for what reason.
>
> Are you able to add some debugging to give us more information. The
> current debugging code in the hotplug/migration sucks...

After reproducing the problem a couple times, it seems that it can occur 
for different types of pages. Running page-types on the offending page 
over two separate instances produced the following:

# tools/vm/page-types -a 307968-308224
            flags	page-count       MB  symbolic-flags			long-symbolic-flags
0x0000000000000400	         1        0  __________B________________________________	buddy
	     total	         1        0

And the following on a separate run:

# tools/vm/page-types -a 313088-313344
            flags	page-count       MB  symbolic-flags			long-symbolic-flags
0x000000000000006c	         1        0  __RU_lA____________________________________	referenced,uptodate,lru,active
            total	         1        0

The source of the failure in migrate_pages actually doesn't seem to be 
that we're hitting the case of the permanent failure, but instead the 
-EAGAIN case. I traced the EAGAIN return back to 
migrate_page_move_mapping which I've seen return EAGAIN in two places:

mm/migrate.c:453
	if (!mapping) {
		/* Anonymous page without mapping */
		if (page_count(page) != expected_count)
                       return -EAGAIN;

mm/migrate.c:476
	if (page_count(page) != expected_count ||
               radix_tree_deref_slot_protected(pslot,
                                       &mapping->i_pages.xa_lock) != page) {
               xa_unlock_irq(&mapping->i_pages);
               return -EAGAIN;
	}

So it seems in each case, the actual reference count for the page is not 
what it is expected to be.