On Thu, Feb 02, 2012 at 08:53:25PM +0100, Michal Nazarewicz wrote:

quoted

On Tue, Jan 31, 2012 at 05:23:59PM +0100, Marek Szyprowski wrote:

quoted

Pages, which have incorrect migrate type on free finally
causes pageblock migration type change from MIGRATE_CMA to MIGRATE_MOVABLE.

On Thu, 02 Feb 2012 13:47:29 +0100, Mel Gorman [off-list ref] wrote:

quoted

I'm not quite seeing this. In free_hot_cold_page(), the pageblock
type is checked so the page private should be set to MIGRATE_CMA or
MIGRATE_ISOLATE for the CMA area. It's not clear how this can change a
pageblock to MIGRATE_MOVABLE in error.

Here's what I think may happen:

When drain_all_pages() is called, __free_one_page() is called for each page on
pcp list with migrate type deducted from page_private() which is MIGRATE_CMA.
This result in the page being put on MIGRATE_CMA freelist even though its
pageblock's migrate type is MIGRATE_ISOLATE.

Ok, although it will only be allocated for MIGRATE_CMA-compatible
requests so it is not a disaster.

When allocation happens and pcp list is empty, rmqueue_bulk() will get executed
with migratetype argument set to MIGRATE_MOVABLE.  It calls __rmqueue() to grab
some pages and because the page described above is on MIGRATE_CMA freelist it
may be returned back to rmqueue_bulk().

This will allocate the page from a pageblock we are trying to isolate
pages from, but only for a movable page that can still be migrated. It
does mean that CMA is doing more work than it should of course and
the problem also impacts memory hot-remove. It's worse for memory
hot-remove because potentially an UNMOVABLE page was allocated from
a MIGRATE_ISOLATE pageblock.

But, pageblock's migrate type is not MIGRATE_CMA but MIGRATE_ISOLATE, so the
following code:

#ifdef CONFIG_CMA
		if (is_pageblock_cma(page))
			set_page_private(page, MIGRATE_CMA);
		else
#endif
			set_page_private(page, migratetype);

will set it's private to MIGRATE_MOVABLE and in the end the page lands back
on MIGRATE_MOVABLE pcp list but this time with page_private == MIGRATE_MOVABLE
and not MIGRATE_CMA.

One more drain_all_pages() (which may happen since alloc_contig_range() calls
set_migratetype_isolate() for each block) and next __rmqueue_fallback() may
convert the whole pageblock to MIGRATE_MOVABLE.

I know, this sounds crazy and improbable, but I couldn't find an easier path
to destruction.  As you pointed, once the page is allocated, free_hot_cold_page()
will do the right thing by reading pageblock's migrate type.

Ok, it's crazy but the problem is there.

Marek is currently experimenting with various patches including the following
change:

#ifdef CONFIG_CMA
                int mt = get_pageblock_migratetype(page);
                if (is_migrate_cma(mt) || mt == MIGRATE_ISOLATE)
                        set_page_private(page, mt);
                else
#endif
                        set_page_private(page, migratetype);

As a matter of fact, if __rmqueue() was changed to return migrate type of the
freelist it took page from, we could avoid this get_pageblock_migratetype() all
together.  For now, however, I'd rather not go that way just yet -- I'll be happy
to dig into it once CMA gets merged.

Ok, thanks for persisting with this.

-- 
Mel Gorman
SUSE Labs