[PATCH v3 3/3] arm: kernel: implement cpuidle_ops with psci backend

[lorenzo.pieralisi@arm.com (Lorenzo Pieralisi)]

On Wed, Jul 15, 2015 at 03:45:07PM +0100, Russell King - ARM Linux wrote:
> On Wed, Jul 15, 2015 at 02:46:03PM +0100, Lorenzo Pieralisi wrote:
> > On Tue, Jul 14, 2015 at 09:41:38PM +0100, Russell King - ARM Linux wrote:
> > > Sorry, NAK, and end of discussion.  There is nothing more to be said
> > > here.
> > 
> > I beg to differ. To solve the issue that you brought up with this series,
> > we can create an arch function that allows to set CPUidle operations, which
> > would remove the need for the OF construct you did not like, this mirrors
> > what's done for PSCI smp operations (something similar to smp_set_ops),
> > does that sound a reasonable approach to you ?
> 
> What's the point of that?
> 
> This is _all_ that arch/arm/kernel/psci_cpuidle.c will contain after this
> series:
> 
> +#include <linux/cpuidle.h>
> +#include <linux/psci.h>
> +
> +#include <asm/cpuidle.h>
> +
> +static struct cpuidle_ops psci_cpuidle_ops __initdata = {
> +       .suspend = cpu_psci_cpu_suspend,
> +       .init = cpu_psci_cpu_init_idle,
> +};
> +CPUIDLE_METHOD_OF_DECLARE(psci_cpuidle, "psci", &psci_cpuidle_ops);
> 
> There is _nothing_ ARM specific there.  All it's doing is providing an
> entry in the DT linker-built table for that data structure, and that
> data structure contains a pair of function pointers to code which _will_
> be located in drivers/firmware/psci.c.
> 
> I'm saying _that_ is totally pointless, that data structure and
> CPUIDLE_METHOD_OF_DECLARE() _should_ live with the code in
> drivers/firmware/psci.c.
> 
> There's nothing in the above quoted code which points anything to any
> 32bit ARM specific stuff at all (so there's no need for a smp_set_ops()
> type thing), it's all about giving DT a way to specify the CPU idle
> operations which should be used, in this case, allowing DT to specify
> that the _generic_ PSCI CPU idle operations are to be used.

No it is not. struct cpuidle_ops (as struct smp_operations) is ARM
specific. You can't add that to generic code (unless guarded
by CONFIG_ARM).

You can't move that struct declaration above to drivers/firmware
for the same reason you can't move CPU_METHOD_OF_DECLARE code (that
is used as a mechanism to initialize SMP ops on ARM) to
drivers/firmware.

On ARM64 this is done differently. On ARM64 SMP operations and
CPUidle operations are merged into struct cpu_operations, that
is initialized through DT at boot (it does not use linker script
mechanism because all enable-method are statically defined in
the kernel in an array to avoid SMP methods du jour
proliferation - see arch/arm64/kernel/cpu_ops.c and the
supported_cpu_ops array).

Now, we have to initialize cpuidle_ops on ARM (and SMP ops too) if
those operations are implemented with PSCI.

For SMP operations, this is done in arch/arm/kernel/setup.c through
smp_set_ops. What I was saying is that we can do the same, without
resorting to the linker script based approch above for CPUidle operations.

We can't move:

static struct cpuidle_ops psci_cpuidle_ops __initdata = {
	.suspend = cpu_psci_cpu_suspend,
	.init = cpu_psci_cpu_init_idle,
};
CPUIDLE_METHOD_OF_DECLARE(psci_cpuidle, "psci", &psci_cpuidle_ops);

to drivers/firmware unless it is guarded by CONFIG_ARM, that's what
I am saying, is that clear ?

How do you want us to do it ?

> > As for M.Rutland's series[1], and the patch that moves common PSCI code to
> > drivers/firmware I reiterate my point, please review it[1] and provide us
> > with feedback if any otherwise acknowledge the code move, which is the
> > basis on top of which everything else can be developed, I do not understand
> > why you are stopping [1] from getting into the kernel since it moves
> > code from arch/arm to drivers/firmware to have it in a common and shared
> > place between ARM and ARM64, what's the issue you have with that or put
> > it differently why are you NAKing it ?
> 
> I'm NAKing everything related to moving the PSCI code around until
> someone in the PSCI maintainer pool (that's not me - that's ARM Ltd)
> clearly demonstrates that they know what they're talking about, and
> has a plan behind this reorganisation.

M.Rutland's patch series represents the code reorganisation and as far
as I am concerned that's a complete and well defined plan, I still do
not understand why you are NAKing that piece of code, it is completely
orthogonal to what we are debating above, please have a look at it
otherwise we are going around in circles here.

> Right now, I'm getting the impression that there is _no_ plan, or if
> there is, it's an absurd plan which results in data structures which
> should not be in arch/arm being left behind there.

Mark's series does not leave any data structure behind, it is moving
code to a common place in the kernel so that the _current_ PSCI
implementation can be shared between ARM and ARM64, again please
have a look at it, we can tackle how to define cpuidle_ops in a way
that you deem reasonable later, it is a separate issue, please
understand.

Thank you,
Lorenzo