hugetlbfs: fix anon huge page migration race
authorMike Kravetz <mike.kravetz@oracle.com>
Sat, 14 Nov 2020 06:52:16 +0000 (22:52 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 14 Nov 2020 19:26:04 +0000 (11:26 -0800)
commit336bf30eb76580b579dc711ded5d599d905c0217
tree03f4a092b5921e02bb64e4721650f4edff09d280
parent8b21ca0218d29cc6bb7028125c7e5a10dfb4730c
hugetlbfs: fix anon huge page migration race

Qian Cai reported the following BUG in [1]

  LTP: starting move_pages12
  BUG: unable to handle page fault for address: ffffffffffffffe0
  ...
  RIP: 0010:anon_vma_interval_tree_iter_first+0xa2/0x170 avc_start_pgoff at mm/interval_tree.c:63
  Call Trace:
    rmap_walk_anon+0x141/0xa30 rmap_walk_anon at mm/rmap.c:1864
    try_to_unmap+0x209/0x2d0 try_to_unmap at mm/rmap.c:1763
    migrate_pages+0x1005/0x1fb0
    move_pages_and_store_status.isra.47+0xd7/0x1a0
    __x64_sys_move_pages+0xa5c/0x1100
    do_syscall_64+0x5f/0x310
    entry_SYSCALL_64_after_hwframe+0x44/0xa9

Hugh Dickins diagnosed this as a migration bug caused by code introduced
to use i_mmap_rwsem for pmd sharing synchronization.  Specifically, the
routine unmap_and_move_huge_page() is always passing the TTU_RMAP_LOCKED
flag to try_to_unmap() while holding i_mmap_rwsem.  This is wrong for
anon pages as the anon_vma_lock should be held in this case.  Further
analysis suggested that i_mmap_rwsem was not required to he held at all
when calling try_to_unmap for anon pages as an anon page could never be
part of a shared pmd mapping.

Discussion also revealed that the hack in hugetlb_page_mapping_lock_write
to drop page lock and acquire i_mmap_rwsem is wrong.  There is no way to
keep mapping valid while dropping page lock.

This patch does the following:

 - Do not take i_mmap_rwsem and set TTU_RMAP_LOCKED for anon pages when
   calling try_to_unmap.

 - Remove the hacky code in hugetlb_page_mapping_lock_write. The routine
   will now simply do a 'trylock' while still holding the page lock. If
   the trylock fails, it will return NULL. This could impact the
   callers:

    - migration calling code will receive -EAGAIN and retry up to the
      hard coded limit (10).

    - memory error code will treat the page as BUSY. This will force
      killing (SIGKILL) instead of SIGBUS any mapping tasks.

   Do note that this change in behavior only happens when there is a
   race. None of the standard kernel testing suites actually hit this
   race, but it is possible.

[1] https://lore.kernel.org/lkml/20200708012044.GC992@lca.pw/
[2] https://lore.kernel.org/linux-mm/alpine.LSU.2.11.2010071833100.2214@eggly.anvils/

Fixes: c0d0381ade79 ("hugetlbfs: use i_mmap_rwsem for more pmd sharing synchronization")
Reported-by: Qian Cai <cai@lca.pw>
Suggested-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: <stable@vger.kernel.org>
Link: https://lkml.kernel.org/r/20201105195058.78401-1-mike.kravetz@oracle.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/hugetlb.c
mm/memory-failure.c
mm/migrate.c
mm/rmap.c