mm/migrate: fix refcount handling when !hugepage_migration_supported()
authorPunit Agrawal <[email protected]>
Fri, 2 Jun 2017 21:46:40 +0000 (14:46 -0700)
committerLinus Torvalds <[email protected]>
Fri, 2 Jun 2017 22:07:38 +0000 (15:07 -0700)
commit30809f559a0d348c2dfd7ab05e9a451e2384962e
treebb2982cca50de6390a03b4585a6a7e6cfef134c3
parente2093926a098a8ccf0f1d10f6df8dad452cb28d3
mm/migrate: fix refcount handling when !hugepage_migration_supported()

On failing to migrate a page, soft_offline_huge_page() performs the
necessary update to the hugepage ref-count.

But when !hugepage_migration_supported() , unmap_and_move_hugepage()
also decrements the page ref-count for the hugepage.  The combined
behaviour leaves the ref-count in an inconsistent state.

This leads to soft lockups when running the overcommitted hugepage test
from mce-tests suite.

  Soft offlining pfn 0x83ed600 at process virtual address 0x400000000000
  soft offline: 0x83ed600: migration failed 1, type 1fffc00000008008 (uptodate|head)
  INFO: rcu_preempt detected stalls on CPUs/tasks:
   Tasks blocked on level-0 rcu_node (CPUs 0-7): P2715
    (detected by 7, t=5254 jiffies, g=963, c=962, q=321)
    thugetlb_overco R  running task        0  2715   2685 0x00000008
    Call trace:
      dump_backtrace+0x0/0x268
      show_stack+0x24/0x30
      sched_show_task+0x134/0x180
      rcu_print_detail_task_stall_rnp+0x54/0x7c
      rcu_check_callbacks+0xa74/0xb08
      update_process_times+0x34/0x60
      tick_sched_handle.isra.7+0x38/0x70
      tick_sched_timer+0x4c/0x98
      __hrtimer_run_queues+0xc0/0x300
      hrtimer_interrupt+0xac/0x228
      arch_timer_handler_phys+0x3c/0x50
      handle_percpu_devid_irq+0x8c/0x290
      generic_handle_irq+0x34/0x50
      __handle_domain_irq+0x68/0xc0
      gic_handle_irq+0x5c/0xb0

Address this by changing the putback_active_hugepage() in
soft_offline_huge_page() to putback_movable_pages().

This only triggers on systems that enable memory failure handling
(ARCH_SUPPORTS_MEMORY_FAILURE) but not hugepage migration
(!ARCH_ENABLE_HUGEPAGE_MIGRATION).

I imagine this wasn't triggered as there aren't many systems running
this configuration.

[[email protected]: remove dead comment, per Naoya]
Link: http://lkml.kernel.org/r/[email protected]
Reported-by: Manoj Iyer <[email protected]>
Tested-by: Manoj Iyer <[email protected]>
Suggested-by: Naoya Horiguchi <[email protected]>
Signed-off-by: Punit Agrawal <[email protected]>
Cc: Joonsoo Kim <[email protected]>
Cc: Wanpeng Li <[email protected]>
Cc: Christoph Lameter <[email protected]>
Cc: Mel Gorman <[email protected]>
Cc: <[email protected]> [3.14+]
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
mm/memory-failure.c