Re: [RFC] locking/rwsem: Avoid issuing wakeup before setting the reader waiter to nil

From: Davidlohr Bueso
Date: Thu Nov 29 2018 - 13:43:40 EST


On Thu, 29 Nov 2018, Peter Zijlstra wrote:

On Thu, Nov 29, 2018 at 02:12:32PM +0100, Peter Zijlstra wrote:
Yes, I think this is real, and worse, I think we need to go audit all
wake_q_add() users and document this behaviour.

In the ideal case we'd delay the actual wakeup to the last wake_up_q(),
but I don't think we can easily fix that.

See commit: 1d0dcb3ad9d3 ("futex: Implement lockless wakeups"), I think
that introduces the exact same bug.

Something like the below perhaps, altough this pattern seems to want a
wake_a_add() variant that already assumes get_task_struct().

So I was looking at ways to avoid the redundant reference counting,
but given how wake_q_add() and wake_up_q() are so loose I can't
see how to avoid it -- we hold reference across the calls to maintain
valid mem.

For example, wake_q will grab reference iff the cmpxchg succeeds,
likewise it will enter the wakeup loop in wake_up_q(), and there is no
awareness of which caller had the failed cmpxchg because another wakup
was in progress.

And yes, afaict all wake_q users suffer from the same issue, so we have
to move the wake_q_add() after the condition, while explicitly doing
the task ref counting.

Thanks,
Davidlohr