Re: [PATCH] events: Annotate parent_ctx with __rcu

[Joel Fernandes]

On Mon, Feb 10, 2020 at 10:17:27PM +0530, Amol Grover wrote:
> On Mon, Feb 10, 2020 at 02:34:59PM +0100, Peter Zijlstra wrote:
> > On Mon, Feb 10, 2020 at 06:29:48PM +0530, Amol Grover wrote:
> > > On Mon, Feb 10, 2020 at 10:36:24AM +0100, Peter Zijlstra wrote:
> > > > On Sat, Feb 08, 2020 at 08:16:49PM +0530, Amol Grover wrote:
> > 
> > > > > @@ -3106,26 +3106,31 @@ static void ctx_sched_out(struct perf_event_context *ctx,
> > > > >  static int context_equiv(struct perf_event_context *ctx1,
> > > > >  			 struct perf_event_context *ctx2)
> > > > >  {
> > > > > +	struct perf_event_context *parent_ctx1, *parent_ctx2;
> > > > > +
> > > > >  	lockdep_assert_held(&ctx1->lock);
> > > > >  	lockdep_assert_held(&ctx2->lock);
> > > > >  
> > > > > +	parent_ctx1 = rcu_dereference(ctx1->parent_ctx);
> > > > > +	parent_ctx2 = rcu_dereference(ctx2->parent_ctx);

You can probably remove the earlier lockdep_assert_held(s) if you're going to
use rcu_dereference_protected() here, since that would do the checking anyway.

> > > > 
> > > > Bah.
> > > > 
> > > > Why are you  fixing all this sparse crap and making the code worse?
> > > 
> > > Hi Peter,
> > > 
> > > Sparse is quite noisy and we need to eliminate false-positives, right?
> > 
> > Dunno, I've been happy just ignoring it all.

FWIW some of the sparse fixes Amol made recently did uncover so existing
"bugs" :) (Not in perf but other code).

> > > __rcu will tell the developer, this pointer could change and he needs to
> > > take the required steps to make sure the code doesn't break.
> > 
> > I know what it does; what I don't know is why you need to make the code
> > worse. In paricular, __rcu doesn't mandate rcu_dereference(), esp. not
> > when you're actually holding the write side lock.
> 
> I might've misinterpreted the code. How does replacing rcu_dereference()
> with
> parent_ctx1 = rcu_dereference_protected(ctx1->parent_ctx,
> 					lockdep_is_held(&ctx1->lock));
> sound?

FWIW, some maintainers do hate calling RCU APIs when write side lock is held.
Evidently it does make the code readability a bit worse and I can see Peter's
point of view because the existing code is correct. I leave it to you guys to
decide how you want to handle it.

thanks!

 - Joel