RFC, GIC based smp_cross_call cleanup suggestion

Colin Cross ccross at google.com
Sun Apr 3 02:36:41 EDT 2011

On Sat, Apr 2, 2011 at 11:23 PM, Santosh Shilimkar
<santosh.shilimkar at ti.com> wrote:
> + Thomas G,
> On 4/2/2011 2:40 PM, Colin Cross wrote:
>> On Sat, Apr 2, 2011 at 1:51 AM, Russell King - ARM Linux
>> <linux at arm.linux.org.uk>  wrote:
>>> On Fri, Apr 01, 2011 at 04:55:02PM -0600, Grant Likely wrote:
>>>> On Fri, Apr 1, 2011 at 4:26 PM, John Linn<John.Linn at xilinx.com>  wrote:
>>>>> I’m getting ready to submit a patch to add SMP to Xilinx code. I notice
>>>>> that
>>>>> smp_cross_call for all GIC based platforms is duplicated across each
>>>>> platform in smp.h.
>>>>> I thought I’d try to jump in to help with some cleanup, although I
>>>>> realize
>>>>> it’s minimal, I have to start somewhere.
>>>>> What about moving the smp_cross_call for GIC based designs into gic.h?
>>>> Go for it.  It's an obvious cleanup.
>>> That assumes that all SMP implementations will always have a GIC.  It
>>> looks to me like this is conditional on shmobile, and so I don't think
>>> its that trivial - maybe Paul or Magnus can first indicate why this is.
>> OMAP4 may also require a custom smp_cross_call implementation if CPU
>> idle is going to be supported in SMP - in CPU off idle modes, a GIC
>> SGI will not wake the CPU, and a write directly to the CPU's power
>> management controller or an external interrupt source would be
>> required.
> This can be done without making smp_cross_call() platform
> specific.
> While working on broad-cast notifiers for ARM with Thomas G, this
> point was discussed.
> Where the TWD can't wakeup its own local CPU from C3 mode, how do we
> provide a platform specific method to perform this wakeup ?
> Thomas Quoted....
> "It would not complicate the OMAP code that much. All it needs is
> extending the clock event device callbacks by an broadcast_affinity()
> function which would be called from the broadcast code when the
> broadcast device is armed. The argument would be a cpumask which would
> tell you which core(s) to wake up when the broadcast timer fires next.
> So OMAP would fill in that hook and implement the wakeup redirector
> setup, which I guess would be a couple of lines."
> From above it's should be trivial once the broad-cast notifiers
> are extended to have "broadcast_affinity()" supported

That fixes the localtimer, but not calls to generic_exec_single, which
also calls smp_cross_call.

More information about the linux-arm-kernel mailing list