On 09/07/2015 01:40 PM, Kirill A. Shutemov wrote:
On Sun, Sep 06, 2015 at 03:21:05PM -0400, Sasha Levin wrote:
==================================================================
ThreadSanitizer: data-race in munlock_vma_pages_range
Write of size 8 by thread T378 (K2633, CPU3):
[<ffffffff81212579>] munlock_vma_pages_range+0x59/0x3e0 mm/mlock.c:425
[<ffffffff81212ac9>] mlock_fixup+0x1c9/0x280 mm/mlock.c:549
[<ffffffff81212ccc>] do_mlock+0x14c/0x180 mm/mlock.c:589
[< inlined >] SyS_munlock+0x74/0xb0 SYSC_munlock mm/mlock.c:651
[<ffffffff812130b4>] SyS_munlock+0x74/0xb0 mm/mlock.c:643
[<ffffffff81eb352e>] entry_SYSCALL_64_fastpath+0x12/0x71
arch/x86/entry/entry_64.S:186
...
Previous read of size 8 by thread T398 (K2623, CPU2):
[<ffffffff8121d198>] try_to_unmap_one+0x78/0x4f0 mm/rmap.c:1208
[< inlined >] rmap_walk+0x147/0x450 rmap_walk_file mm/rmap.c:1540
[<ffffffff8121e7b7>] rmap_walk+0x147/0x450 mm/rmap.c:1559
[<ffffffff8121ef72>] try_to_munlock+0xa2/0xc0 mm/rmap.c:1423
[<ffffffff81211bb0>] __munlock_isolated_page+0x30/0x60 mm/mlock.c:129
[<ffffffff81212066>] __munlock_pagevec+0x236/0x3f0 mm/mlock.c:331
[<ffffffff812128a0>] munlock_vma_pages_range+0x380/0x3e0 mm/mlock.c:476
[<ffffffff81212ac9>] mlock_fixup+0x1c9/0x280 mm/mlock.c:549
[<ffffffff81212ccc>] do_mlock+0x14c/0x180 mm/mlock.c:589
[< inlined >] SyS_munlock+0x74/0xb0 SYSC_munlock mm/mlock.c:651
[<ffffffff812130b4>] SyS_munlock+0x74/0xb0 mm/mlock.c:643
[<ffffffff81eb352e>] entry_SYSCALL_64_fastpath+0x12/0x71
arch/x86/entry/entry_64.S:186
Okay, the detected race is mlock/munlock vs. rmap.
On rmap side we check vma->vm_flags in few places without taking
vma->vm_mm->mmap_sem. The vma cannot be freed since we hold i_mmap_rwsem
or anon_vma_lock, but nothing prevent vma->vm_flags from changing under
us.
In this particular case, speculative check in beginning of
try_to_unmap_one() is fine, since we re-check it under mmap_sem later in
the function.
False-negative is fine too here, since we will mlock the page in
__mm_populate() on mlock side after mlock_fixup().
BUT.
We *must* have all speculative vm_flags accesses wrapped READ_ONCE() to
avoid all compiler trickery, like duplication vm_flags access with
inconsistent results.
Doesn't taking a semaphore, as in try_to_unmap_one(), already imply a
compiler barrier forcing vm_flags to be re-read?
I looked only on VM_LOCKED checks, but there are few other flags checked
in rmap. All of them must be handled carefully. At least READ_ONCE() is
required.
Other solution would be to introduce per-vma spinlock to protect
vma->vm_flags and probably other vma fields and offload this duty
from mmap_sem.
But that's much bigger project.
Sounds like an overkill, unless we find something more serious than this.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>