On Tue, 2014-03-11 at 13:30 -0700, Andrew Morton wrote:
> On Tue, 11 Mar 2014 13:07:33 -0700 Davidlohr Bueso <davidlohr@xxxxxx> wrote:
>
> > On Tue, 2014-03-11 at 15:39 -0400, Sasha Levin wrote:
> > > Hi all,
> > >
> > > I've ended up deleting the log file by mistake, but this bug does seem to be important
> > > so I'd rather not wait before the same issue is triggered again.
> > >
> > > The call chain is:
> > >
> > > mlock (mm/mlock.c:745)
> > > __mm_populate (mm/mlock.c:700)
> > > __mlock_vma_pages_range (mm/mlock.c:229)
> > > VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
> >
> > So __mm_populate() is only called by mlock(2) and this VM_BUG_ON seems
> > wrong as we call it without the lock held:
> >
> > up_write(¤t->mm->mmap_sem);
> > if (!error)
> > error = __mm_populate(start, len, 0);
> > return error;
> > }
>
> __mm_populate() pretty clearly calls __mlock_vma_pages_range() under
> down_read(mm->mmap_sem).
>
> I worry about what happens if __get_user_pages decides to do
>
> if (ret & VM_FAULT_RETRY) {
> if (nonblocking)
> *nonblocking = 0;
> return i;
> }
>
> uh-oh, that just cleared __mm_populate()'s `locked' variable and we'll
> forget to undo mmap_sem. That won't explain this result, but it's a
> potential problem.
>
>
> All I can think is that find_vma() went and returned a vma from a
> different mm, which would be odd. How about I toss this in there?
... and we know that there is a bug (https://lkml.org/lkml/2014/3/9/201)
with stale caches going on. We seem to be missing an invalidation and/or
flush somewhere.
--
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>