Re: problem with rtnetlink 'reference' count

[Florian Westphal <fw@strlen.de>]

Peter Zijlstra <peterz@infradead.org> wrote:
> On Mon, Oct 23, 2017 at 06:37:44PM +0200, Florian Westphal wrote:
> 
> > Is refcount_t only supposed to be used with dec_and_test patterns?
> 
> Yes, for reference counting objects.

Hmm, I still feel its appropriate, but anyway:

> > > This rtnetlink_rcv_msg() is called from softirq-context, right? Also,
> > > all that stuff happens with rcu_read_lock() held.
> > 
> > No, its called from process context.
> 
> OK, so then why not do something like so?
> @@ -260,10 +259,18 @@ void rtnl_unregister_all(int protocol)
>  	RCU_INIT_POINTER(rtnl_msg_handlers[protocol], NULL);
>  	rtnl_unlock();
>  
> +	/*
> +	 * XXX explain what this is for...
> +	 */
>  	synchronize_net();
>  
> -	while (refcount_read(&rtnl_msg_handlers_ref[protocol]) > 1)
> -		schedule();
> +	/*
> +	 * This serializes against the rcu_read_lock() section in
> +	 * rtnetlink_rcv_msg() such that after this, all prior instances have
> +	 * completed and future instances must observe the NULL written above.
> +	 */
> +	synchronize_rcu();

Yes, but that won't help with running dumpers, see below.

> @@ -4218,7 +4223,6 @@ static int rtnetlink_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh,
>  			};
>  			err = netlink_dump_start(rtnl, skb, nlh, &c);

This will copy .dumper function address to nlh->cb for later invocation
when dump gets resumed (its called from netlink_recvmsg()),
so this can return to userspace and dump can be resumed on next recv().

Because the dumper function was stored in the socket, NULLing the
rtnl_msg_handlers[] only prevents new dumps from starting but not
already set-up dumps from resuming.