Re: [PATCH v4 01/17] perf: Protect perf_guest_cbs with RCU

[Peter Zijlstra]

On Thu, Nov 11, 2021 at 08:26:58AM +0100, Paolo Bonzini wrote:
> On 11/11/21 03:07, Sean Christopherson wrote:

> >   EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
> >   int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
> >   {
> > -	perf_guest_cbs = NULL;
> > +	if (WARN_ON_ONCE(rcu_access_pointer(perf_guest_cbs) != cbs))
> > +		return -EINVAL;
> > +
> > +	rcu_assign_pointer(perf_guest_cbs, NULL);
> > +	synchronize_rcu();
> This technically could be RCU_INIT_POINTER but it's not worth a respin.
> There are dozens of other occurrences, and if somebody wanted they
> could use Coccinelle to fix all of them.

I've been pushing the other way, trying to get rid of RCU_INIT_POINTER()
since rcu_assign_pointer(, NULL) actualy DTRT per __builtin_constant_p()
etc.

There's a very few sites where we use RCU_INIT_POINTER() with a !NULL
argument, and those are 'special'.