Date: Fri, 18 Nov 2016 13:48:48 +0000
From: Will Deacon <will.deacon@arm.com>
To: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>,
        linux-kernel@vger.kernel.org, rt@linuxtronix.de,
        Mark Rutland <mark.rutland@arm.com>,
        Russell King <linux@armlinux.org.uk>,
        linux-arm-kernel@lists.infradead.org
Subject: Re: [PATCH 15/20] ARM/hw_breakpoint: Convert to hotplug state machine
Message-ID: <20161118134847.GO13470@arm.com>
References: <20161117183541.8588-1-bigeasy@linutronix.de>
 <20161117183541.8588-16-bigeasy@linutronix.de>
 <20161118120453.GI13470@arm.com>
 <alpine.DEB.2.20.1611181401180.3615@nanos>
 <20161118132912.GM13470@arm.com>
 <alpine.DEB.2.20.1611181430070.3615@nanos>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <alpine.DEB.2.20.1611181430070.3615@nanos>
User-Agent: Mutt/1.5.23 (2014-03-12)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1378
Lines: 31

On Fri, Nov 18, 2016 at 02:42:15PM +0100, Thomas Gleixner wrote:
> On Fri, 18 Nov 2016, Will Deacon wrote:
> > On Fri, Nov 18, 2016 at 02:11:58PM +0100, Thomas Gleixner wrote:
> > > But it's guaranteed that cpuhp_setup_state() will not return before the
> > > callback has been invoked on each online cpu.
> > 
> > Ok, that's good.
> > 
> > > If cpus are not yet online when that code is invoked, then it's the same
> > > behaviour as before. It will be invoked when the cpu comes online.
> > 
> > Just to check, but what stops a CPU from coming online between the call
> > to cpuhp_setup_state and the call to cpuhp_remove_state_nocalls in the
> > case of failure (debug_err_mask isn't empty)?
> 
> Indeed! I missed that part. So we still need a get/put_online_cpus()
> protection around all of this.

Yes, that should do it.

> Just for curiosity sake. Wouldn't it be simpler and less error prone to
> make the ARM_DBG_READ/WRITE macros use the exception table and handle that
> in the undefined instruction handler to avoid this hook dance?

That would be an option, but it's only the reset sequence that could
generate this fault so it's simpler to isolate it there. We'd also have
to take into account SMP if we toggle the handler in the READ/WRITE
accessors, since the fault handler framework is system-wide as opposed
to per-cpu. The whole thing is grotty as hell.

Will