On Mon, Jul 05, 2021 at 07:45:26PM +0100, Matthew Wilcox wrote:
> On Mon, Jul 05, 2021 at 11:04:21AM -0700, Jaegeuk Kim wrote:
> > On 07/05, Matthew Wilcox wrote:
> > > I think freshly allocated pages have a page->private of 0. ie this
> > > code in mm/page_alloc.c:
> > >
> > > page = rmqueue(ac->preferred_zoneref->zone, zone, order,
> > > gfp_mask, alloc_flags, ac->migratetype);
> > > if (page) {
> > > prep_new_page(page, order, gfp_mask, alloc_flags);
> > >
> > > where prep_new_page() calls post_alloc_hook() which contains:
> > > set_page_private(page, 0);
> >
> > Hmm, I can see it in 4.14 and 5.10 kernel.
> >
> > The trace is on:
> >
> > 30875 [ 1065.118750] c3 87 f2fs_migrate_page+0x354/0x45c
> > 30876 [ 1065.123872] c3 87 move_to_new_page+0x70/0x30c
> > 30877 [ 1065.128813] c3 87 migrate_pages+0x3a0/0x964
> > 30878 [ 1065.133583] c3 87 compact_zone+0x608/0xb04
> > 30879 [ 1065.138257] c3 87 kcompactd+0x378/0x4ec
> > 30880 [ 1065.142664] c3 87 kthread+0x11c/0x12c
> > 30881 [ 1065.146897] c3 87 ret_from_fork+0x10/0x18
> >
> > It seems compaction_alloc() gets a free page which doesn't reset the fields?
>
> I'm not really familiar with the compaction code. Mel, I see a call
> to post_alloc_hook() in split_map_pages(). Are there other ways of
> getting the compaction code to allocate a page which don't go through
> split_map_pages()?
I don't *think* so but I didn't look too hard as I had limited time
available before a meeting. compaction_alloc calls isolate_freepages
and that calls split_map_pages whether fast or slow isolating pages. The
problem *may* be in split_page because only the head page gets order set
to 0 but it's a bad fit because tail pages should be cleared of private
state by del_page_from_free_list. It might be worth adding a debugging
patch to split_pages that prints a warning once if a tail page has private
state and dump the contents of private to see if it looks like an order.
--
Mel Gorman
SUSE Labs
