[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20200603171320.GE2570@hirez.programming.kicks-ass.net>
Date: Wed, 3 Jun 2020 19:13:20 +0200
From: Peter Zijlstra <peterz@...radead.org>
To: "Paul E. McKenney" <paulmck@...nel.org>
Cc: tglx@...utronix.de, x86@...nel.org, elver@...gle.com,
kasan-dev@...glegroups.com, linux-kernel@...r.kernel.org,
will@...nel.org, dvyukov@...gle.com, glider@...gle.com,
andreyknvl@...gle.com
Subject: Re: [PATCH 2/9] rcu: Fixup noinstr warnings
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
> 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.
Powered by blists - more mailing lists