Re: [RFC PATCH RT 3/4] rcu: unlock special: Treat irq and preempt disabled the same

["Paul E. McKenney" <paulmck@linux.ibm.com>]

On Tue, Jun 18, 2019 at 08:19:07PM -0500, Scott Wood wrote:
> [Note: Just before posting this I noticed that the invoke_rcu_core stuff
>  is part of the latest RCU pull request, and it has a patch that
>  addresses this in a more complicated way that appears to deal with the
>  bare irq-disabled sequence as well.

Far easier to deal with it than to debug the lack of it.  ;-)

>  Assuming we need/want to support such sequences, is the
>  invoke_rcu_core() call actually going to result in scheduling any
>  sooner?  resched_curr() just does the same setting of need_resched
>  when it's the same cpu.
> ]

Yes, invoke_rcu_core() can in some cases invoke the scheduler sooner.
Setting the CPU-local bits might not have effect until the next interrupt.

So if -rt wants the simpler and slower approach, the change needs to
use IS_ENABLED(CONFIG_PREEMPT_RT_FULL) or similar.  Not that this is
an issue until CONFIG_PREEMPT_RT_FULL hits mainline.

							Thanx, Paul

> Since special should never be getting set inside an irqs-disabled
> critical section, this is safe as long as there are no sequences of
> rcu_read_lock()/local_irq_disable()/rcu_read_unlock()/local_irq_enable()
> (without preempt_disable() wrapped around the IRQ disabling, as spinlocks
> do).  If there are such sequences, then the grace period may be delayed
> until the next time need_resched is checked.
> 
> This is needed because otherwise, in a sequence such as:
> 1. rcu_read_lock()
> 2. *preempt*, set rcu_read_unlock_special.b.blocked
> 3. preempt_disable()
> 4. rcu_read_unlock()
> 5. preempt_enable()
> 
> ...rcu_read_unlock_special.b.blocked will not be cleared during
> step 4, because of the disabled preemption.  If an interrupt is then
> taken between steps 4 and 5, and that interrupt enters scheduler code
> that takes pi/rq locks, and an rcu read lock inside that, then when
> dropping that rcu read lock we will end up in rcu_read_unlock_special()
> again -- but this time, since irqs are disabled, it will call
> invoke_rcu_core() in the RT tree (regardless of PREEMPT_RT_FULL), which
> calls wake_up_process().  This can cause a pi/rq lock deadlock.  An
> example of interrupt code that does this is scheduler_tick().
> 
> The above sequence can be found in (at least) __lock_task_sighand() (for
> !PREEMPT_RT_FULL) and d_alloc_parallel().
> 
> It's potentially an issue on non-RT as well.  While
> raise_softirq_irqoff() doesn't call wake_up_process() when in_interrupt()
> is true, if code between steps 4 and 5 directly calls into scheduler
> code, and that code uses RCU with pi/rq lock held, wake_up_process() can
> still be called.
> 
> On RT, migrate_enable() is such a codepath, so an in_interrupt() check
> alone would not work on RT.  Instead, keep track of whether we've already
> had an rcu_read_unlock_special() with preemption disabled but haven't yet
> scheduled, and rely on the preempt_enable() yet to come instead of
> calling invoke_rcu_core().
> 
> Signed-off-by: Scott Wood <swood@redhat.com>
> ---
>  kernel/rcu/tree_plugin.h | 10 ++--------
>  1 file changed, 2 insertions(+), 8 deletions(-)
> 
> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> index 5d63914b3687..d7ddbcc7231c 100644
> --- a/kernel/rcu/tree_plugin.h
> +++ b/kernel/rcu/tree_plugin.h
> @@ -630,14 +630,8 @@ static void rcu_read_unlock_special(struct task_struct *t)
>  	if (preempt_bh_were_disabled || irqs_were_disabled) {
>  		WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, false);
>  		/* Need to defer quiescent state until everything is enabled. */
> -		if (irqs_were_disabled) {
> -			/* Enabling irqs does not reschedule, so... */
> -			invoke_rcu_core();
> -		} else {
> -			/* Enabling BH or preempt does reschedule, so... */
> -			set_tsk_need_resched(current);
> -			set_preempt_need_resched();
> -		}
> +		set_tsk_need_resched(current);
> +		set_preempt_need_resched();
>  		local_irq_restore(flags);
>  		return;
>  	}
> -- 
> 1.8.3.1
>