mm: vmscan: remove deadlock due to throttling failing to make progress

A soft lockup bug in kcompactd was reported in a private bugzilla with
the following visible in dmesg;

  watchdog: BUG: soft lockup - CPU#33 stuck for 26s! [kcompactd0:479]
  watchdog: BUG: soft lockup - CPU#33 stuck for 52s! [kcompactd0:479]
  watchdog: BUG: soft lockup - CPU#33 stuck for 78s! [kcompactd0:479]
  watchdog: BUG: soft lockup - CPU#33 stuck for 104s! [kcompactd0:479]

The machine had 256G of RAM with no swap and an earlier failed
allocation indicated that node 0 where kcompactd was run was potentially
unreclaimable;

  Node 0 active_anon:29355112kB inactive_anon:2913528kB active_file:0kB
    inactive_file:0kB unevictable:64kB isolated(anon):0kB isolated(file):0kB
    mapped:8kB dirty:0kB writeback:0kB shmem:26780kB shmem_thp:
    0kB shmem_pmdmapped: 0kB anon_thp: 23480320kB writeback_tmp:0kB
    kernel_stack:2272kB pagetables:24500kB all_unreclaimable? yes

Vlastimil Babka investigated a crash dump and found that a task
migrating pages was trying to drain PCP lists;

  PID: 52922  TASK: ffff969f820e5000  CPU: 19  COMMAND: "kworker/u128:3"
  Call Trace:
     __schedule
     schedule
     schedule_timeout
     wait_for_completion
     __flush_work
     __drain_all_pages
     __alloc_pages_slowpath.constprop.114
     __alloc_pages
     alloc_migration_target
     migrate_pages
     migrate_to_node
     do_migrate_pages
     cpuset_migrate_mm_workfn
     process_one_work
     worker_thread
     kthread
     ret_from_fork

This failure is specific to CONFIG_PREEMPT=n builds.  The root of the
problem is that kcompact0 is not rescheduling on a CPU while a task that
has isolated a large number of the pages from the LRU is waiting on
kcompact0 to reschedule so the pages can be released.  While
shrink_inactive_list() only loops once around too_many_isolated, reclaim
can continue without rescheduling if sc->skipped_deactivate == 1 which
could happen if there was no file LRU and the inactive anon list was not
low.

Link: https://lkml.kernel.org/r/20220203100326.GD3301@suse.de
Fixes: d818fca1 ("mm/vmscan: throttle reclaim and compaction when too may pages are isolated")
Signed-off-by: Mel Gorman <mgorman@suse.de>
Debugged-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/vmscan.c

@@ -1066,8 +1066,10 @@ void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
 	 * forward progress (e.g. journalling workqueues or kthreads).
 	 */
 	if (!current_is_kswapd() &&
-	    current->flags & (PF_IO_WORKER|PF_KTHREAD))
+	    current->flags & (PF_IO_WORKER|PF_KTHREAD)) {
+		cond_resched();
 		return;
+	}
 
 	/*
 	 * These figures are pulled out of thin air.