Re: [PATCH RFC 2/4] net, pidfs: prepare for handing out pidfds for reaped sk->sk_peer_pid

[Kuniyuki Iwashima]

From: Christian Brauner <brauner@xxxxxxxxxx>
Date: Thu, 24 Apr 2025 17:19:28 +0200
> > > @@ -643,6 +644,14 @@ static void unix_sock_destructor(struct sock *sk)
> > >  		return;
> > >  	}
> > > 
> > > +	if (sock_flag(sk, SOCK_RCU_FREE)) {
> > > +		pr_info("Attempting to release RCU protected socket with sleeping 
> > > locks: %p\n", sk);
> > > +		return;
> > > +	}
> > 
> > unix-sockets do not use `SOCK_RCU_FREE`,

Right, and I think we won't flag SOCK_RCU_FREE in the future.


> but even if they did, doesn't
> > this flag imply that the destructor is delayed via `call_rcu`, and
> > thus *IS* allowed to sleep? And then, sleeping in the destructor is
> > always safe, isn't it? `SOCK_RCU_FREE` just guarantees that it is
> > delayed for at least an RCU grace period, right? Not sure, what you
> > are getting at here, but I might be missing something obvious as well.
> 
> Callbacks run from call_rcu() can be called from softirq context and in
> general are not allowed to block. That's what queue_rcu_work() is for
> which uses system_unbound_wq.
> 
> > 
> > Regardless, wouldn't you want WARN_ON_ONCE() rather than pr_info?
> 
> Sure.

I prefer DEBUG_NET_WARN_ON_ONCE() or removing it as rcu_sleep_check()
in __might_sleep() has better checks.

The netdev CI enables debug.config, which has CONFIG_DEBUG_ATOMIC_SLEEP
and enables the checks, so adding a test case in
tools/testing/selftests/net/af_unix/scm_pidfd.c will catch the future
regression.