Re: [PATCH v8 01/11] ARM: brcmstb: add infrastructure for ARM-based Broadcom STB SoCs

From: Brian Norris
Date: Wed Aug 13 2014 - 19:47:15 EST


Hi Russell,

Picking up this thread again, as things are now set for dropping this
patch and resubmitting SMP support for 3.18.

On Sat, Aug 02, 2014 at 10:27:56AM +0100, Russell King wrote:
> On Thu, Jul 31, 2014 at 03:06:42PM -0700, Brian Norris wrote:
> > Yes, I noticed this. What I meant is that smp_ops.cpu_die() and
> > smp_ops.cpu_kill() are not synchronized.
> ...
> > We're not relying on the L1 cache, though. Don't sync_cache_{r,w}()
> > ensure all reads/writes reach at least the L2?
>
> What exactly are you trying to achieve here?

Synchronization between v7_exit_coherency_flush() (on the dying CPU) and
yanking the power (brcmstb_cpu_kill(), on the murderous CPU). The core
completion-based synchronization is not sufficient, since it allows
brcmstb_smp_ops.smp_kill and brcmstb_smp_ops.smp_die to race.

Am I somehow not achieving what I intend here?

> > How does that ensure that the CPU is down by the time the work is
> > scheduled? It seems like this would just defer the work long enough that
> > it *most likely* has quiesced, but I don't see how this gives us a
> > better guarantee. Or maybe I'm missing something. (If so, please do
> > enlighten!)
>
> Note that I said a delayed work queue. The dying CPU runs a predictable
> sequence once cpu_die() has been entered - interrupts at the GIC have
> been programmed to be routed to other CPUs, interrupts at the CPU are
> masked, so the CPU isn't going to be doing anything else except executing
> that code path. So, it's going to be a predictable number of CPU cycles.
>
> That allows you to arrange a delayed workqueue (or a timer) to fire
> after a period of time that you can guarantee that the dying CPU has
> reached that wfi().

OK, that sounds workable for the active hotplug case.

But what about for the suspend case? CPUs are hot-unplugged during
disable_nonboot_cpus(), and I don't see that this would guarantee the
workqueue will complete before we enter suspend.

> Another point which raises itself in your patch is this:
>
> + /* Settle-time from Broadcom-internal DVT reference code */
> + udelay(7);
>
> 7us looks very precise, but udelay() may not be that precise. What is
> the actual specification? Does it say "you must wait at least 7us"?
>
> udelay() _may_ return early, especially if it is using the CPU delay
> loop to perform the delay - I've explained why this happens previously,
> and why it isn't a bug.
>
> If you're using a timer-based delay for udelay() (which you should be
> using if you support cpufreq) then the delay should be more accurate,
> but it's still good practise to give a little leeway on the figure.

I'm looking into this specific delay. I'd bet it's just "wait at least
7us." I could probably factor in some leeway to be safe.

Thanks,
Brian
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/