Re: [PATCH] sched: Don't call any kfree*() API in do_set_cpus_allowed()

[Frederic Weisbecker <frederic@kernel.org>]

On Tue, Oct 31, 2023 at 09:53:08AM +0100, Peter Zijlstra wrote:
> On Mon, Oct 30, 2023 at 08:14:18PM -0400, Waiman Long wrote:
> > Commit 851a723e45d1 ("sched: Always clear user_cpus_ptr in
> > do_set_cpus_allowed()") added a kfree() call to free any user
> > provided affinity mask, if present. It was changed later to use
> > kfree_rcu() in commit 9a5418bc48ba ("sched/core: Use kfree_rcu()
> > in do_set_cpus_allowed()") to avoid a circular locking dependency
> > problem.
> > 
> > It turns out that even kfree_rcu() isn't safe for avoiding
> > circular locking problem. As reported by kernel test robot,
> > the following circular locking dependency still exists:
> > 
> >   &rdp->nocb_lock --> rcu_node_0 --> &rq->__lock
> > 
> > So no kfree*() API can be used in do_set_cpus_allowed(). To prevent
> > memory leakage, the unused user provided affinity mask is now saved in a
> > lockless list to be reused later by subsequent sched_setaffinity() calls.
> > 
> > Without kfree_rcu(), the internal cpumask_rcuhead union can be removed
> > too as a lockless list entry only holds a single pointer.
> > 
> > Fixes: 851a723e45d1 ("sched: Always clear user_cpus_ptr in do_set_cpus_allowed()")
> 
> Bah, or we fix RCU...  Paul, how insane is the below?

Makes sense. We can't remove &rdp->nocb_lock --> rcu_node_0 but we can (and
should) indeed remove rcu_node_0 --> &rq->__lock

Just a detail below:

> @@ -2284,10 +2289,13 @@ static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
>  		}
>  		for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
>  			rdp = per_cpu_ptr(&rcu_data, cpu);
> -			if (f(rdp)) {
> +			ret = f(rdp);
> +			if (ret > 0) {
>  				mask |= rdp->grpmask;
>  				rcu_disable_urgency_upon_qs(rdp);
>  			}
> +			if (ret < 0)
> +				rsmask |= 1UL << (cpu - rnp->grplo);

I guess this can be simplified with rsmask |= rdp->grpmask;

Thanks.

>  		}
>  		if (mask != 0) {
>  			/* Idle/offline CPUs, report (releases rnp->lock). */
> @@ -2296,6 +2304,9 @@ static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
>  			/* Nothing to do here, so just drop the lock. */
>  			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
>  		}
> +
> +		for_each_leaf_node_cpu_mask(rnp, cpu, rsmask)
> +			resched_cpu(cpu);
>  	}
>  }
>