Re: [syzbot] INFO: task hung in __unmap_and_move (4)

[Hillf Danton <hdanton@sina.com>]

On 29 Dec 2022 22:01:42 +0000 Matthew Wilcox <willy@infradead.org>
> On Thu, Dec 29, 2022 at 01:48:42AM -0800, syzbot wrote:
> > INFO: task kcompactd1:32 blocked for more than 143 seconds.
> >       Not tainted 6.1.0-syzkaller-14594-g72a85e2b0a1e #0
> > "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> > task:kcompactd1      state:D stack:26360 pid:32    ppid:2      flags:0x00004000
> > Call Trace:
> >  <TASK>
> >  context_switch kernel/sched/core.c:5244 [inline]
> >  __schedule+0x9d1/0xe40 kernel/sched/core.c:6555
> >  schedule+0xcb/0x190 kernel/sched/core.c:6631
> >  io_schedule+0x83/0x100 kernel/sched/core.c:8811
> >  folio_wait_bit_common+0x8ca/0x1390 mm/filemap.c:1297
> >  folio_lock include/linux/pagemap.h:938 [inline]
> >  __unmap_and_move+0x835/0x12a0 mm/migrate.c:1040
> >  unmap_and_move+0x28f/0xd80 mm/migrate.c:1194
> >  migrate_pages+0x50f/0x14d0 mm/migrate.c:1477
> >  compact_zone+0x2893/0x37a0 mm/compaction.c:2413
> >  proactive_compact_node mm/compaction.c:2665 [inline]
> >  kcompactd+0x1b46/0x2750 mm/compaction.c:2975
> 
> OK, so kcompactd is trying to compact a zone, has called folio_lock()
> and whoever has the folio locked has had it locked for 143 seconds.
> That seems like quite a long time.  Probably it is locked waiting
> for I/O.
> 
> > NMI backtrace for cpu 1
> [...]
> >  lock_release+0x81/0x870 kernel/locking/lockdep.c:5679
> >  rcu_read_unlock include/linux/rcupdate.h:797 [inline]
> >  folio_evictable+0x1df/0x2d0 mm/internal.h:140
> >  move_folios_to_lru+0x324/0x25c0 mm/vmscan.c:2413
> >  shrink_inactive_list+0x60b/0xca0 mm/vmscan.c:2529
> >  shrink_list mm/vmscan.c:2767 [inline]
> >  shrink_lruvec+0x449/0xc50 mm/vmscan.c:5951
> >  shrink_node_memcgs+0x35c/0x780 mm/vmscan.c:6138
> >  shrink_node+0x299/0x1050 mm/vmscan.c:6169
> >  shrink_zones+0x4fb/0xc40 mm/vmscan.c:6407
> >  do_try_to_free_pages+0x215/0xcd0 mm/vmscan.c:6469
> >  try_to_free_pages+0x3e8/0xc60 mm/vmscan.c:6704
> >  __perform_reclaim mm/page_alloc.c:4750 [inline]
> >  __alloc_pages_direct_reclaim mm/page_alloc.c:4772 [inline]
> >  __alloc_pages_slowpath+0xd5c/0x2120 mm/page_alloc.c:5178
> >  __alloc_pages+0x3d4/0x560 mm/page_alloc.c:5562
> >  folio_alloc+0x1a/0x50 mm/mempolicy.c:2296
> >  filemap_alloc_folio+0xca/0x2c0 mm/filemap.c:972
> >  page_cache_ra_unbounded+0x212/0x820 mm/readahead.c:248
> >  do_sync_mmap_readahead+0x786/0x950 mm/filemap.c:3062
> >  filemap_fault+0x38d/0x1060 mm/filemap.c:3154
> 
> So dhcpd has taken a page fault, missed in the page cache, called
> readahead, is presumably partway through the readahead (ie has folios
> locked in the page cache, not uptodate and I/O hasn't been submitted
> on them).  It's trying to allocate pages, but has fallen into reclaim.
> It's trying to shrink the inactive list at this point, but is not
> having much luck.  For one thing, it's a GFP_NOFS allocation.  So
> it was probably the one who woke kcompactd.

In wakeup_kswapd(), wake up kcompactd instead of kswapd because of highly
fragmented but balanced node.
		/*
		 * There may be plenty of free memory available, but it's too
		 * fragmented for high-order allocations.  Wake up kcompactd
		 * and rely on compaction_suitable() to determine if it's
		 * needed.  If it fails, it will defer subsequent attempts to
		 * ratelimit its work.
		 */
		if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
			wakeup_kcompactd(pgdat, order, highest_zoneidx);
> 
> Should readahead be trying less hard to allocate memory?  It's already
> using __GFP_NORETRY.

Given blocked kcompactd, enough pages have been reclaimed by kswapd before
waking up kcompactd, and order0 allocations should be satisfied.