On Wed, Aug 08, 2018 at 12:24:20PM -0700, Joel Fernandes wrote:
> On Wed, Aug 8, 2018 at 7:49 AM, Paul E. McKenney
> <paul...@linux.vnet.ibm.com> wrote:
> [...]
> >
> >> In that case based on what you're saying, the patch I sent to using
> >> different srcu_struct for NMI is still good I guess...
> >
> > As long as you wait for both SRCU grace periods. Hmmm... Maybe that means
> > that there is still a use for synchronize_rcu_mult():
> >
> > void call_srcu_nmi(struct rcu_head *rhp, rcu_callback_t func)
> > {
> > call_srcu(&trace_srcu_struct_nmi, rhp, func);
> > }
> >
> > void call_srcu_nonmi(struct rcu_head *rhp, rcu_callback_t func)
> > {
> > call_srcu(&trace_srcu_struct_nonmi, rhp, func);
> > }
> >
> > ...
> >
> > /* Wait concurrently on the two grace periods. */
> > synchronize_rcu_mult(call_srcu_nmi, call_srcu_nonmi);
> >
> > On the other hand, I bet that doing this is just fine in your use case:
> >
> > synchronize_srcu(&trace_srcu_struct_nmi);
> > synchronize_srcu(&trace_srcu_struct_nonmi);
> >
> > But please note that synchronize_rcu_mult() is no more in my -rcu tree,
> > so if you do want it please let me know (and please let me know why it
> > is important).
>
> I did the chaining thing (one callback calling another), that should
> work too right? I believe that is needed so that the tracepoint
> callbacks are freed at one point and only when both NMI and non-NMI
> read sections have completed.
Yes, that works also. It is possible to make that happen concurrently
via atomic_dec_and_test() or similar, but if the latency is not a problem,
why bother?
> >> >> It does start to seem like a show stopper :-(
> >> >
> >> > I suppose that an srcu_read_lock_nmi() and srcu_read_unlock_nmi() could
> >> > be added, which would do atomic ops on sp->sda->srcu_lock_count. Not
> >> > sure
> >> > whether this would be fast enough to be useful, but easy to provide:
> >> >
> >> > int __srcu_read_lock_nmi(struct srcu_struct *sp) /* UNTESTED. */
> >> > {
> >> > int idx;
> >> >
> >> > idx = READ_ONCE(sp->srcu_idx) & 0x1;
> >> > atomic_inc(&sp->sda->srcu_lock_count[idx]);
> >> > smp_mb__after_atomic(); /* B */ /* Avoid leaking critical
> >> > section. */
> >> > return idx;
> >> > }
> >> >
> >> > void __srcu_read_unlock_nmi(struct srcu_struct *sp, int idx)
> >> > {
> >> > smp_mb__before_atomic(); /* C */ /* Avoid leaking critical
> >> > section. */
> >> > atomic_inc(&sp->sda->srcu_unlock_count[idx]);
> >> > }
> >> >
> >> > With appropriate adjustments to also allow Tiny RCU to also work.
> >> >
> >> > Note that you have to use _nmi() everywhere, not just in NMI handlers.
> >> > In fact, the NMI handlers are the one place you -don't- need to use
> >> > _nmi(), strangely enough.
> >> >
> >> > Might be worth a try -- smp_mb__{before,after}_atomic() is a no-op on
> >> > some architectures, for example.
> >>
> >> Continuing Steve's question on regular interrupts, do we need to use
> >> this atomic_inc API for regular interrupts as well? So I guess
> >
> > If NMIs use one srcu_struct and non-NMI uses another, the current
> > srcu_read_lock() and srcu_read_unlock() will work just fine. If any given
> > srcu_struct needs both NMI and non-NMI readers, then we really do need
> > __srcu_read_lock_nmi() and __srcu_read_unlock_nmi() for that srcu_struct.
>
> Yes, I believe as long as in_nmi() works reliably, we can use the
> right srcu_struct (NMI vs non-NMI) and it would be fine.
>
> Going through this thread, it sounds though that this_cpu_inc may not
> be reliable on all architectures even for non-NMI interrupts and
> local_inc may be the way to go.
My understanding is that this_cpu_inc() is defined to handle interrupts,
so any architecture on which it is unreliable needs to fix its bug. ;-)
> For next merge window (not this one), lets do that then? Paul, if you
> could provide me an SRCU API that uses local_inc, then I believe that
> coupled with this patch should be all that's needed:
> https://lore.kernel.org/patchwork/patch/972657/
>
> Steve did express concern though if in_nmi() works reliably (i.e.
> tracepoint doesn't fire from "thunk" code before in_nmi() is
> available). Any thoughts on that Steve?
Agreed, not the upcoming merge window. But we do need to work out
exactly what is the right way to do this.
Thanx, Paul
