Re: kernel BUG at include/linux/swapops.h:LINE!

[Matthew Wilcox <willy@infradead.org>]

On Wed, Jul 29, 2020 at 08:54:32PM +0100, Matthew Wilcox wrote:
> On Wed, Jul 29, 2020 at 10:21:51PM +0300, Kirill A. Shutemov wrote:
> > On Mon, Jul 27, 2020 at 01:03:10PM +0100, Matthew Wilcox wrote:
> > > > It would be interesting to know if the migration entires ever got removed
> > > > for pfn. I mean if remove_migration_pte() got called for it.
> > > > 
> > > > It can be rmap issue too. Maybe it misses PMD on remove_migration_ptes()
> > > > or something.
> > > 
> > > It's not mapped with a PMD.  I tweaked my debugging slightly:
> > > 
> > >  static inline swp_entry_t make_migration_entry(struct page *page, int write)
> > >  {
> > > -       BUG_ON(!PageLocked(compound_head(page)));
> > > +       VM_BUG_ON_PAGE(!PageLocked(page), page);
> > >  
> > > +if (PageHead(page)) dump_page(page, "make entry");
> > > +if (PageTail(page)) printk("pfn %lx order %d\n", page_to_pfn(page), thp_order(thp_head(page)));
> > > 
> > > 1523 page:0000000006f62206 refcount:490 mapcount:1 mapping:0000000000000000 index:0x562b12a00 pfn:0x1dc00
> > > 1523 head:0000000006f62206 order:9 compound_mapcount:0 compound_pincount:0
> > > 1523 anon flags: 0x400000000009003d(locked|uptodate|dirty|lru|active|head|swapbacked)
> > > 1523 raw: 400000000009003d ffffecfd41301308 ffffecfd41b08008 ffff9e9971c00059
> > > 1523 raw: 0000000562b12a00 0000000000000000 000001ea00000000 0000000000000000
> > > 1523 page dumped because: make entry
> > > 1523 pfn 1dc01 order 9
> > > 1523 pfn 1dc02 order 9
> > > 1523 pfn 1dc03 order 9
> > > ...
> > > 
> > > Notice that it's an anonymous page, so it's not related to my work.
> > 
> > I don't have much hope, but could you try if the patch below would blow
> > up?
> 
> Running it now.  Results probably in twenty minutes.

It didn't blow up.  I added a dump_stack() after the call to dump_page()
and got this ...

2922 page:0000000085a5c107 refcount:474 mapcount:1 mapping:0000000000000000 index:0x559e98a00 pfn:0x35200
2922 head:0000000085a5c107 order:9 compound_mapcount:0 compound_pincount:0
2922 anon flags: 0x400000000009003d(locked|uptodate|dirty|lru|active|head|swapbacked)
2922 raw: 400000000009003d ffffe8e5c1bbaf48 ffffe8e5c046ee88 ffffa2e7f3787ec9
2922 raw: 0000000559e98a00 0000000000000000 000001da00000000 0000000000000000
2922 page dumped because: make entry
2922 CPU: 5 PID: 23471 Comm: dd Kdump: loaded Tainted: G        W         5.8.0-rc6-00067-gd8b18bdf9870-dirty #358
2922 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1 04/01/2014
2922 Call Trace:
2922  dump_stack+0x5e/0x7a
2922  try_to_unmap_one+0x846/0x860
2922  rmap_walk_anon+0x13d/0x2a0
2922  rmap_walk_locked+0x23/0x30
2922  try_to_unmap+0x64/0xbc
2922  split_huge_page_to_list+0x188/0xdb0
2922  deferred_split_scan+0x148/0x240
2922  shrink_slab.constprop.0+0x198/0x330
2922  shrink_node+0x1a8/0x440
2922  try_to_free_pages+0x18f/0x480
2922  __alloc_pages_slowpath.constprop.0+0x297/0xca0
2922  __alloc_pages_nodemask+0x1ba/0x1e0
2922  pagecache_get_page+0xd8/0x330
2922  grab_cache_page_write_begin+0x1c/0x40
2922  iomap_write_begin+0x2d6/0x6d0
2922  iomap_write_actor+0x8b/0x1c0
2922  iomap_apply+0xe3/0x310
2922  iomap_file_buffered_write+0x5c/0x80
2922  xfs_file_buffered_aio_write+0xbd/0x310
2922  xfs_file_write_iter+0xa8/0xc0
2922  new_sync_write+0xf5/0x170
2922  vfs_write+0x191/0x1e0

I think that's interesting because it's not trying to allocate a huge
page itself (I didn't touch the write_begin path).  Rather, I presume
the additional memory pressure from allocating huge pages is causing
anonymous pages to be split to free up memory.

It survived all the way to generic/224 this run, but I don't think
that's relevant.