Re: [PATCH v6 3/4] perf,x86: add Intel RAPL PMU support

[Stephane Eranian]

Peter,


I spent some time trying to understand the hotplug code in kernel/cpu.c.
I still see this cpu_hotplug_begin()/cpu_hotplug_done() pair which
serializes access. So it seems that hotplug is still serialized.
That appears to be true also for the CPU_STARTING phase of
hiotplug.

I don't know where Andi got that cpu starting was parallel now.
Andi?

If hotplug is not parallel, then we can drop that RAPL hotplug_lock
I added.



On Mon, Nov 11, 2013 at 5:16 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Mon, Nov 11, 2013 at 05:07:22PM +0100, Stephane Eranian wrote:
>> On Mon, Nov 11, 2013 at 4:59 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>> > On Fri, Nov 08, 2013 at 08:46:02PM +0100, Stephane Eranian wrote:
>> >
>> >
>> >> +static DEFINE_SPINLOCK(rapl_hotplug_lock);
>> >
>> > What is this thing protecting?
>> >
>> > Like last time it appears to be used only from hotplug notifier
>> > callbacks and those are already fully serialized.
>> >
>> I thought you or somebody else had said, that hotplug in now parallel.
>
> There's a little something about hotplug here I think:
>
>  http://marc.info/?l=linux-kernel&m=138118235612441
>
> But I'm not entirely sure that's the origin of this lock.
>
> Andi seems worried about the IPI vs hotplug/migrate. I'm not entrely
> sure the lock as proposed actually fixes that.
>
> The 'easy' way would be to put a synchronize_sched() right after the
> migrate code, that forces all CPUs to have scheduled once and thus also
> ensures all pending/running interrupts are serviced.
>
> But I don't think that's strictly required as I seem to recall hotplug
> actually already does something like this -- it pretty much has to.
--
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/