Re: [PATCH 1/3] perf: convert perf_event_context.refcount to refcount_t

[Mark Rutland <mark.rutland@arm.com>]

On Fri, Feb 01, 2019 at 03:44:38PM +0000, Reshetova, Elena wrote:
> > On Tue, Jan 29, 2019 at 01:55:32PM +0000, Reshetova, Elena wrote:
> > > > On Mon, Jan 28, 2019 at 02:27:26PM +0200, Elena Reshetova wrote:
> > > > > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > > > > index 3cd13a3..a1e87d2 100644
> > > > > --- a/kernel/events/core.c
> > > > > +++ b/kernel/events/core.c
> > > > > @@ -1171,7 +1171,7 @@ static void perf_event_ctx_deactivate(struct
> > > > perf_event_context *ctx)
> > > > >
> > > > >  static void get_ctx(struct perf_event_context *ctx)
> > > > >  {
> > > > > -	WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
> > > > > +	WARN_ON(!refcount_inc_not_zero(&ctx->refcount));
> > > >
> > > > This could be refcount_inc(), remember how that already produces a WARN
> > > > when we try and increment 0.
> > >
> > > But is this true for the x86 arch-specific implementation also?
> > 
> > If you use recount_inc_checked(), it will always produce the WARN(),
> > even when using the x86-specific refcount implementation.
> > 
> > (this was one place I had intended to use the *_checked() forms of the
> > refcount ops).
> 
> Yes, with refcount_inc_checked() it would work, but I don't like it
> that much when we have functions that behave regardless of refcount 
> config. It does help for code minimization & clarity like here, but I think 
> it complicates things even more: two different configs, then functions that
> do not obey configs, etc. 

Sure. The main idea of having the _checked() forms was to not lose
warnings in a conversion to refcount_t, but I appreciate that people
might not like the existing warnings at all.

> Anyhow, I can change this to refcount_inc_checked(), if this is what everyone
> thinks is the best. 

I'll defer to Peter.

Peter, would you prefer refcount_inc() or refcount_inc_checked() here?

Thanks,
Mark.