On Wed, Jun 03, 2020 at 07:13:20PM +0200, Peter Zijlstra wrote:
> On Wed, Jun 03, 2020 at 09:46:00AM -0700, Paul E. McKenney wrote:
>
> > > --- a/kernel/rcu/tree.c
> > > +++ b/kernel/rcu/tree.c
> > > @@ -250,7 +250,7 @@ static noinstr void rcu_dynticks_eqs_ent
> > > * next idle sojourn.
> > > */
> > > rcu_dynticks_task_trace_enter(); // Before ->dynticks update!
> > > - seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
> > > + seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
> >
> > To preserve KCSAN's ability to see this, there would be something like
> > instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks)) prior
> > to the instrumentation_end() invoked before rcu_dynticks_eqs_enter()
> > in each of rcu_eqs_enter() and rcu_nmi_exit(), correct?
>
> Yes.
>
> > > // RCU is no longer watching. Better be in extended quiescent state!
> > > WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
> > > (seq & RCU_DYNTICK_CTRL_CTR));
> > > @@ -274,13 +274,13 @@ static noinstr void rcu_dynticks_eqs_exi
> > > * and we also must force ordering with the next RCU read-side
> > > * critical section.
> > > */
> > > - seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
> > > + seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
> >
> > And same here, but after the instrumentation_begin() following
> > rcu_dynticks_eqs_exit() in both rcu_eqs_exit() and rcu_nmi_enter(),
> > correct?
>
> Yep.
>
> > > // RCU is now watching. Better not be in an extended quiescent state!
> > > rcu_dynticks_task_trace_exit(); // After ->dynticks update!
> > > WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
> > > !(seq & RCU_DYNTICK_CTRL_CTR));
> > > if (seq & RCU_DYNTICK_CTRL_MASK) {
> > > - atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks);
> > > + arch_atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks);
> >
> > This one is gone in -rcu.
>
> Good, because that would make things 'complicated' with the external
> instrumentation call. And is actually the reason I didn't even attempt
> it this time around.
>
> > > smp_mb__after_atomic(); /* _exit after clearing mask. */
> > > }
> > > }
> > > @@ -313,7 +313,7 @@ static __always_inline bool rcu_dynticks
> > > {
> > > struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
> > >
> > > - return !(atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR);
> > > + return !(arch_atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR);
>
> The above is actually instrumented by KCSAN, due to arch_atomic_read()
> being a READ_ONCE() and it now understanding volatile.
>
> > Also instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks)) as
Right, this should instead be instrument_read(...).
Though if KCSAN is unconditionally instrumenting volatile, how does
this help? Or does KCSAN's instrumentation of volatile somehow avoid
causing trouble?
> > follows:
> >
> > o rcu_nmi_exit(): After each following instrumentation_begin().
>
> Yes
>
> > o In theory in rcu_irq_exit_preempt(), but as this generates code
> > only in lockdep builds, it might not be worth worrying about.
> >
> > o Ditto for rcu_irq_exit_check_preempt().
> >
> > o Ditto for __rcu_irq_enter_check_tick().
>
> Not these, afaict they're all the above arch_atomic_read(), which is
> instrumented due to volatile in these cases.
>
> > o rcu_nmi_enter(): After each following instrumentation_begin().
>
> Yes
>
> > o __rcu_is_watching() is itself noinstr:
> >
> > o idtentry_enter_cond_rcu(): After each following
> > instrumentation_begin().
> >
> > o rcu_is_watching(): Either before or after the call to
> > rcu_dynticks_curr_cpu_in_eqs().
>
> Something like that yes.
>
> > > }
> > >
> > > /*
> > > @@ -692,6 +692,7 @@ noinstr void rcu_nmi_exit(void)
> > > {
> > > struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
> > >
> > > + instrumentation_begin();
> > > /*
> > > * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
> > > * (We are exiting an NMI handler, so RCU better be paying attention
> > > @@ -705,7 +706,6 @@ noinstr void rcu_nmi_exit(void)
> > > * leave it in non-RCU-idle state.
> > > */
> > > if (rdp->dynticks_nmi_nesting != 1) {
> > > - instrumentation_begin();
> > > trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting,
> > > rdp->dynticks_nmi_nesting - 2,
> > > atomic_read(&rdp->dynticks));
> > > WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
> > > @@ -714,7 +714,6 @@ noinstr void rcu_nmi_exit(void)
> > > return;
> > > }
> > >
> > > - instrumentation_begin();
> > > /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
> > > trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0,
> > > atomic_read(&rdp->dynticks));
> > > WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
> >
> > This one looks to be having no effect on instrumentation of atomics, but
> > rather coalescing a pair of instrumentation_begin() into one.
> >
> > Do I understand correctly?
>
> Almost, it puts the WARN_ON_ONCE()s under instrumentation_begin() too,
> and that makes a differnce, iirc it was the
> rcu_dynticks_curr_cpu_in_eqs() call that stood out. But that could've
> been before I switched it to arch_atomic_read(). In any case, I find
> this form a lot clearer.
Got it, thank you.
Thanx, Paul
