On Mon, Oct 19, 2015 at 04:20:05AM -0700, Hugh Dickins wrote: > > Note how munlock_vma_pages_range() via __munlock_pagevec() does > > TestClearPageMlocked() without (or "between") pte or page lock. But the pte > > lock is being taken after clearing VM_LOCKED, so perhaps it's safe against > > try_to_unmap_one... > > A mind-trick I found helpful for understanding the barriers here, is > to imagine that the munlocker repeats its "vma->vm_flags &= ~VM_LOCKED" > every time it takes the pte lock: it does not actually do that, it > doesn't need to of course; but that does help show that ~VM_LOCKED > must be visible to anyone getting that pte lock afterwards. How can you make sure that any other codepath that changes vm_flags would not make (vm_flags & VM_LOCKED) temporary true while dealing with other flags? Compiler can convert things like "vma->vm_flags &= ~VM_FOO;" to whatever it wants as long as end result is the same. It's very unlikely that it will generate code to set all bits to one and then clear all which should be cleared, but it's theoretically possible. I think we need to have at lease WRITE_ONCE() everywhere we update vm_flags and READ_ONCE() where we read it without mmap_sem taken. -- Kirill A. Shutemov -- 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>