Re: [PATCH RESEND v4] locking/pvqspinlock: Fix double hash race

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Aug 09, 2016 at 05:37:47PM +0800, Wanpeng Li wrote:
> From: Wanpeng Li <wanpeng.li@hotmail.com>
> 
> When the lock holder vCPU is racing with the queue head vCPU:
> 
> lock holder vCPU             queue head vCPU
> =====================        ==================
> 
> node->locked = 1;
> <preemption>                 READ_ONCE(node->locked)
>    ...                       pv_wait_head_or_lock():
>                                SPIN_THRESHOLD loop;
>                                pv_hash();
>                                lock->locked = _Q_SLOW_VAL;
>                                node->state  = vcpu_hashed;
> pv_kick_node():
>   cmpxchg(node->state,
>      vcpu_halted, vcpu_hashed);
>   lock->locked = _Q_SLOW_VAL;
>   pv_hash();

So here the example is 'wrong' in that it doesn't illustrate the fail
case, namely having vcpu_halted win while we're hashed.

> +++ b/kernel/locking/qspinlock_paravirt.h
> @@ -450,7 +450,28 @@ pv_wait_head_or_lock(struct qspinlock *lock, struct mcs_spinlock *node)
>  				goto gotlock;
>  			}
>  		}
> -		WRITE_ONCE(pn->state, vcpu_halted);
> +		/*
> +		 * lock holder vCPU             queue head vCPU
> +		 * ----------------             ---------------
> +		 * node->locked = 1;
> +		 * <preemption>                 READ_ONCE(node->locked)
> +		 *    ...                       pv_wait_head_or_lock():
> +		 *                                SPIN_THRESHOLD loop;
> +		 *                                pv_hash();
> +		 *                                lock->locked = _Q_SLOW_VAL;
> +		 *                                node->state  = vcpu_hashed;
> +		 * pv_kick_node():
> +		 *   cmpxchg(node->state,
> +		 *      vcpu_halted, vcpu_hashed);
> +		 *   lock->locked = _Q_SLOW_VAL;
> +		 *   pv_hash();
> +		 *
> +		 * With preemption at the right moment, it is possible that both the
> +		 * lock holder and queue head vCPUs can be racing to set node->state.
> +		 * Making sure the state is never set to vcpu_halted will prevent this
> +		 * racing from happening.
> +		 */
> +		WRITE_ONCE(pn->state, vcpu_hashed);
>  		qstat_inc(qstat_pv_wait_head, true);
>  		qstat_inc(qstat_pv_wait_again, waitcnt);
>  		pv_wait(&l->locked, _Q_SLOW_VAL);

And I completely fail to see the point of this comment, still.

Yes, if we would have used vcpu_halted, there would be a problem, but we
don't so there isn't. What does this comment tell us about the current
code?

In any case, I have an older version (possibly v1) queued up. That fixes
the bug.