[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <20161121201235.GX3612@linux.vnet.ibm.com>
Date: Mon, 21 Nov 2016 12:12:35 -0800
From: "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To: Steven Rostedt <rostedt@...dmis.org>
Cc: Peter Zijlstra <peterz@...radead.org>,
Andi Kleen <andi@...stfloor.org>, Jiri Olsa <jolsa@...hat.com>,
linux-kernel@...r.kernel.org, Ingo Molnar <mingo@...nel.org>,
Josh Triplett <josh@...htriplett.org>,
Jan Stancek <jstancek@...hat.com>
Subject: Re: [BUG] msr-trace.h:42 suspicious rcu_dereference_check() usage!
On Mon, Nov 21, 2016 at 01:24:38PM -0500, Steven Rostedt wrote:
>
> Paul,
>
>
> On Mon, 21 Nov 2016 12:55:01 -0500
> Steven Rostedt <rostedt@...dmis.org> wrote:
>
> > On Mon, 21 Nov 2016 18:18:53 +0100
> > Peter Zijlstra <peterz@...radead.org> wrote:
> >
> > > Its not ftrace as such though, its RCU, ftrace simply uses RCU to avoid
> > > locking, as one does.
> >
> > Just to be clear, as ftrace in the kernel mostly represents function
> > tracing, which doesn't use RCU. This is a tracepoint feature.
> >
> > >
> > > Biggest objection would be that the rcu_irq_enter_irqson() thing does
> > > POPF and rcu_irq_exit_irqson() does again. So wrapping every tracepoint
> > > with that is quite a few cycles.
> >
> > Agree. Even though this ends up being a whack-a-mole(TM) fix, I'm not
> > concerned enough to put a heavy weight rcu idle code in for all
> > tracepoints.
> >
> > Although, what about a percpu flag that can be checked in the
> > tracepoint code to see if it should enable RCU or not?
> >
> > Hmm, I wonder if "rcu_is_watching()" is light enough to have in all
> > tracepoints?
>
> Is it possible to make rcu_is_watching() an inlined call to prevent the
> overhead of doing a function call? This way we could use this in the
> fast path of the tracepoint.
It would mean exposing the rcu_dynticks structure to the rest of the
kernel, but I guess that wouldn't be the end of the world. Are you
calling rcu_is_watching() or __rcu_is_watching()? The latter is
appropriate if preemption is already disabled.
Thanx, Paul
Powered by blists - more mailing lists