thp: fix deadlock in split_huge_pmd()
authorKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Thu, 17 Mar 2016 21:20:13 +0000 (14:20 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 17 Mar 2016 22:09:34 +0000 (15:09 -0700)
commit5f7377147ce0b5d521c0ee2aab2ec3ead51297db
tree4525af234d84b4ff3fe3338ce6faee08630b11f2
parentfec89c109f3a7737fe3a7bf0f40d1fb7709d353b
thp: fix deadlock in split_huge_pmd()

split_huge_pmd() tries to munlock page with munlock_vma_page().  That
requires the page to locked.

If the is locked by caller, we would get a deadlock:

Unable to find swap-space signature
INFO: task trinity-c85:1907 blocked for more than 120 seconds.
      Not tainted 4.4.0-00032-gf19d0bdced41-dirty #1606
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
trinity-c85     D ffff88084d997608     0  1907    309 0x00000000
Call Trace:
  schedule+0x9f/0x1c0
  schedule_timeout+0x48e/0x600
  io_schedule_timeout+0x1c3/0x390
  bit_wait_io+0x29/0xd0
  __wait_on_bit_lock+0x94/0x140
  __lock_page+0x1d4/0x280
  __split_huge_pmd+0x5a8/0x10f0
  split_huge_pmd_address+0x1d9/0x230
  try_to_unmap_one+0x540/0xc70
  rmap_walk_anon+0x284/0x810
  rmap_walk_locked+0x11e/0x190
  try_to_unmap+0x1b1/0x4b0
  split_huge_page_to_list+0x49d/0x18a0
  follow_page_mask+0xa36/0xea0
  SyS_move_pages+0xaf3/0x1570
  entry_SYSCALL_64_fastpath+0x12/0x6b
2 locks held by trinity-c85/1907:
 #0:  (&mm->mmap_sem){++++++}, at:  SyS_move_pages+0x933/0x1570
 #1:  (&anon_vma->rwsem){++++..}, at:  split_huge_page_to_list+0x402/0x18a0

I don't think the deadlock is triggerable without split_huge_page()
simplifilcation patchset.

But munlock_vma_page() here is wrong: we want to munlock the page
unconditionally, no need in rmap lookup, that munlock_vma_page() does.

Let's use clear_page_mlock() instead.  It can be called under ptl.

Fixes: e90309c9f772 ("thp: allow mlocked THP again")
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/huge_memory.c