Re: [PATCH] locking/mutex: Mark racy reads of owner->on_cpu

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Dec 02, 2021 at 12:53:14PM +0100, Marco Elver wrote:
> On Thu, 2 Dec 2021 at 11:13, Marco Elver <elver@google.com> wrote:
> > One of the more frequent data races reported by KCSAN is the racy read
> > in mutex_spin_on_owner(), which is usually reported as "race of unknown
> > origin" without showing the writer. This is due to the racing write
> > occurring in kernel/sched. Locally enabling KCSAN in kernel/sched shows:
> >
> >  | write (marked) to 0xffff97f205079934 of 4 bytes by task 316 on cpu 6:
> >  |  finish_task                kernel/sched/core.c:4632 [inline]
> >  |  finish_task_switch         kernel/sched/core.c:4848
> >  |  context_switch             kernel/sched/core.c:4975 [inline]
> >  |  __schedule                 kernel/sched/core.c:6253
> >  |  schedule                   kernel/sched/core.c:6326
> >  |  schedule_preempt_disabled  kernel/sched/core.c:6385
> >  |  __mutex_lock_common        kernel/locking/mutex.c:680
> >  |  __mutex_lock               kernel/locking/mutex.c:740 [inline]
> >  |  __mutex_lock_slowpath      kernel/locking/mutex.c:1028
> >  |  mutex_lock                 kernel/locking/mutex.c:283
> >  |  tty_open_by_driver         drivers/tty/tty_io.c:2062 [inline]
> >  |  ...
> >  |
> >  | read to 0xffff97f205079934 of 4 bytes by task 322 on cpu 3:
> >  |  mutex_spin_on_owner        kernel/locking/mutex.c:370
> >  |  mutex_optimistic_spin      kernel/locking/mutex.c:480
> >  |  __mutex_lock_common        kernel/locking/mutex.c:610
> >  |  __mutex_lock               kernel/locking/mutex.c:740 [inline]
> >  |  __mutex_lock_slowpath      kernel/locking/mutex.c:1028
> >  |  mutex_lock                 kernel/locking/mutex.c:283
> >  |  tty_open_by_driver         drivers/tty/tty_io.c:2062 [inline]
> >  |  ...
> >  |
> >  | value changed: 0x00000001 -> 0x00000000
> >
> > This race is clearly intentional, and the potential for miscompilation
> > is slim due to surrounding barrier() and cpu_relax(), and the value
> > being used as a boolean.
> >
> > Nevertheless, marking this reader would more clearly denote intent and
> > make it obvious that concurrency is expected. Use READ_ONCE() to avoid
> > having to reason about compiler optimizations now and in future.
> >
> > Similarly, mark the read to owner->on_cpu in mutex_can_spin_on_owner(),
> > which immediately precedes the loop executing mutex_spin_on_owner().
> >
> > Signed-off-by: Marco Elver <elver@google.com>
> [...]
> 
> Kefeng kindly pointed out that there is an alternative, which would
> refactor owner_on_cpu() from rwsem that would address both mutex and
> rwsem:
> https://lore.kernel.org/all/b641f1ea-6def-0fe4-d273-03c35c4aa7d6@huawei.com/

That seems to make sense, except it should probably go under CONFIG_SMP,
since ->on_cpu doesn't otherwise exist.