Re: [RFC] Deadlock via recursive wakeup via RCU with threadirqs

[Joel Fernandes]

On Fri, Jun 28, 2019 at 03:25:47PM -0700, Paul E. McKenney wrote:
> On Fri, Jun 28, 2019 at 05:40:18PM -0400, Joel Fernandes wrote:
> > Hi Paul,
> > 
> > On Fri, Jun 28, 2019 at 01:04:23PM -0700, Paul E. McKenney wrote:
> > [snip]
> > > > > > Commit
> > > > > > - 23634ebc1d946 ("rcu: Check for wakeup-safe conditions in
> > > > > >    rcu_read_unlock_special()") does not trigger the bug within 94
> > > > > >    attempts.
> > > > > > 
> > > > > > - 48d07c04b4cc1 ("rcu: Enable elimination of Tree-RCU softirq
> > > > > >   processing") needed 12 attempts to trigger the bug.
> > > > > 
> > > > > That matches my belief that 23634ebc1d946 ("rcu: Check for wakeup-safe
> > > > > conditions in rcu_read_unlock_special()") will at least greatly decrease
> > > > > the probability of this bug occurring.
> > > > 
> > > > I was just typing a reply that I can't reproduce it with:
> > > >   rcu: Check for wakeup-safe conditions in rcu_read_unlock_special()
> > > > 
> > > > I am trying to revert enough of this patch to see what would break things,
> > > > however I think a better exercise might be to understand more what the patch
> > > > does why it fixes things in the first place ;-) It is probably the
> > > > deferred_qs thing.
> > > 
> > > The deferred_qs flag is part of it!  Looking forward to hearing what
> > > you come up with as being the critical piece of this commit.
> > 
> > The new deferred_qs flag indeed saves the machine from the dead-lock.
> > 
> > If we don't want the deferred_qs, then the below patch also fixes the issue.
> > However, I am more sure than not that it does not handle all cases (such as
> > what if we previously had an expedited grace period IPI in a previous reader
> > section and had to to defer processing. Then it seems a similar deadlock
> > would present. But anyway, the below patch does fix it for me! It is based on
> > your -rcu tree commit 23634ebc1d946f19eb112d4455c1d84948875e31 (rcu: Check
> > for wakeup-safe conditions in rcu_read_unlock_special()).
> 
> The point here being that you rely on .b.blocked rather than
> .b.deferred_qs.  Hmmm...  There are a number of places that check all
> the bits via the .s leg of the rcu_special union.  The .s check in
> rcu_preempt_need_deferred_qs() should be OK because it is conditioned
> on t->rcu_read_lock_nesting of zero or negative.
> Do rest of those also work out OK?
> 
> It would be nice to remove the flag, but doing so clearly needs careful
> review and testing.

Agreed. I am planning to do an audit of this code within the next couple of
weeks so I will be on the look out for any optimization opportunities related
to this. Will let you know if this can work. For now I like your patch better
because it is more conservative and doesn't cause any space overhead.

If you'd like, please free to included my Tested-by on it:

Tested-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>

If you had a chance, could you also point to me any tests that show
performance improvement with the irqwork patch, on the expedited GP usecase?
I'd like to try it out as well. I guess rcuperf should have some?

thanks!

 - Joel