On Thu, Sep 28, 2023 at 05:29:42PM +0200, Peter Zijlstra wrote:
> On Thu, Sep 28, 2023 at 07:54:10AM -0700, Boqun Feng wrote:
> > On Thu, Sep 28, 2023 at 10:09:00AM +0200, Peter Zijlstra wrote:
> > > On Wed, Sep 27, 2023 at 11:06:09PM -0700, Boqun Feng wrote:
> > >
> > > > I think this is a "side-effect" of commit f0f44752f5f6 ("rcu: Annotate
> > > > SRCU's update-side lockdep dependencies"). In verify_lock_unused(), i.e.
> > > > the checking for NMI lock usages, the logic is that
> > >
> > > I think I'm having a problem with this commit -- that is, by adding
> > > lockdep you're adding tracepoint, which rely on RCU being active.
> > >
> > > The result is that SRCU is now no longer usable from !RCU regions.
> > >
> >
> > Interesting
> >
> > > Was this considered and intended?
> > >
> >
> > No, I don't think I have considered this before, I think I may still
> > miss something here, maybe you or Paul can provide an example for such
> > a case?
>
> The whole trace_.*_rcuidle() machinery. Which I thought I had fully
> eradicated, but apparently still exists (with *one* user) :-/
>
> Search for rcuidle in include/linux/tracepoint.h
>
> Also, git grep trace_.*_rcuidle
>
Thanks! But as I mentioned in the IRC, trace_.*_rcuidle() call the
special APIs srcu_read_{un,}lock_notrace(), and these won't call lockdep
annotation functions. And what the commit did was only changing the
lockdep annotation of srcu_read_{un,}lock(), so we are still fine here?
Regards,
Boqun
>
