Re: [PATCH RFC v2 rcu 2/8] srcu: Create an srcu_read_lock_nmisafe() and srcu_read_unlock_nmisafe()

[Frederic Weisbecker]

On Sun, Oct 02, 2022 at 04:46:55PM -0700, Paul E. McKenney wrote:
> On Sun, Oct 02, 2022 at 11:47:10PM +0200, Frederic Weisbecker wrote:
> > On Sun, Oct 02, 2022 at 09:09:57AM -0700, Paul E. McKenney wrote:
> > > On Sun, Oct 02, 2022 at 05:55:16PM +0200, Frederic Weisbecker wrote:
> > > > On Thu, Sep 29, 2022 at 11:07:25AM -0700, Paul E. McKenney wrote:
> > > > > @@ -1090,7 +1121,7 @@ static unsigned long srcu_gp_start_if_needed(struct srcu_struct *ssp,
> > > > >  	int ss_state;
> > > > >  
> > > > >  	check_init_srcu_struct(ssp);
> > > > > -	idx = srcu_read_lock(ssp);
> > > > > +	idx = __srcu_read_lock_nmisafe(ssp);
> > > > 
> > > > Why do we need to force the atomic based version here (even if CONFIG_NEED_SRCU_NMI_SAFE=y)?
> > > 
> > > In kernels built with CONFIG_NEED_SRCU_NMI_SAFE=n, we of course need it.
> > > As you say, in kernels built with CONFIG_NEED_SRCU_NMI_SAFE=y, we don't.
> > > But it doesn't hurt to always use __srcu_read_lock_nmisafe() here, and
> > > this is nowhere near a fastpath, so there is little benefit to using
> > > __srcu_read_lock() when it is safe to do so.
> > > 
> > > In addition, note that it is possible that a given srcu_struct structure's
> > > first grace period is executed before its first reader.  In that
> > > case, we have no way of knowing which of __srcu_read_lock_nmisafe()
> > > or __srcu_read_lock() to choose.
> > > 
> > > So this code always does it the slow(ish) safe way.
> > 
> > But then srcu_read_lock_nmisafe() would work as well, right?
> 
> Almost.
> 
> The problem is that without the leading "__", this would convince SRCU
> that this is an NMI-safe srcu_struct.  Which it might not be.  Worse yet,
> if this srcu_struct had already done an srcu_read_lock(), it would splat.

Ah ok, now that makes sense.

> 
> > > > >  	ss_state = smp_load_acquire(&ssp->srcu_size_state);
> > > > >  	if (ss_state < SRCU_SIZE_WAIT_CALL)
> > > > >  		sdp = per_cpu_ptr(ssp->sda, 0);
> > > > > @@ -1123,7 +1154,7 @@ static unsigned long srcu_gp_start_if_needed(struct srcu_struct *ssp,
> > > > >  		srcu_funnel_gp_start(ssp, sdp, s, do_norm);
> > > > >  	else if (needexp)
> > > > >  		srcu_funnel_exp_start(ssp, sdp_mynode, s);
> > > > > -	srcu_read_unlock(ssp, idx);
> > > > > +	__srcu_read_unlock_nmisafe(ssp, idx);
> > > > >  	return s;
> > > > >  }
> > > > >  
> > > > > @@ -1427,13 +1458,13 @@ void srcu_barrier(struct srcu_struct *ssp)
> > > > >  	/* Initial count prevents reaching zero until all CBs are posted. */
> > > > >  	atomic_set(&ssp->srcu_barrier_cpu_cnt, 1);
> > > > >  
> > > > > -	idx = srcu_read_lock(ssp);
> > > > > +	idx = __srcu_read_lock_nmisafe(ssp);
> > > > 
> > > > And same here?
> > > 
> > > Yes, same here.  ;-)
> > 
> > Now bonus question: why do SRCU grace period starting/tracking
> > need to be in an SRCU read side critical section? :o)
> 
> Because I am lazy and like to keep things simple?  ;-)
> 
> More seriously, take a look at srcu_gp_start_if_needed() and the functions
> it calls and ask yourself what bad things could happen if they were
> preempted for an arbitrarily long period of time.

I can see a risk for ssp->srcu_gp_seq to overflow. Can't say that was obvious
though, at least for me. Am I missing something else?

Thanks.