[PATCH v8 1/2] arm: introduce psci_smp_ops

[Stefano Stabellini]

On Thu, 25 Apr 2013, Nicolas Pitre wrote:
> On Thu, 25 Apr 2013, Will Deacon wrote:
> 
> > On Thu, Apr 25, 2013 at 11:12:54AM +0100, Stefano Stabellini wrote:
> > > On Thu, 25 Apr 2013, Will Deacon wrote:
> > > > > +/*
> > > > > + * cpu_suspend   Suspend the execution on a CPU
> > > > > + * @state        we don't currently describe affinity levels, so just pass 0.
> > > > > + * @entry_point  the first instruction to be executed on return
> > > > > + * returns 0  success, < 0 on failure
> > > > > + *
> > > > > + * cpu_off       Power down a CPU
> > > > > + * @state        we don't currently describe affinity levels, so just pass 0.
> > > > > + * no return on successful call
> > > > > + *
> > > > > + * cpu_on        Power up a CPU
> > > > > + * @cpuid        cpuid of target CPU, as from MPIDR
> > > > > + * @entry_point  the first instruction to be executed on return
> > > > > + * returns 0  success, < 0 on failure
> > > > > + *
> > > > > + * migrate       Migrate the context to a different CPU
> > > > > + * @cpuid        cpuid of target CPU, as from MPIDR
> > > > > + * returns 0  success, < 0 on failure
> > > > > + *
> > > > > + */
> > > > 
> > > > Can you move these comments into psci-smp.c please? They're really specific
> > > > to the implementation there, and if we put them in a header we're lying to
> > > > ourselves about the parameters actually described by the PSCI specification.
> > > 
> > > You have a good point about the PSCI spec.
> > > 
> > > However from the Linux POV these comments should regard the functions
> > > exported by psci_operations, not the firmware interface, this is why I
> > > think it makes sense to keep them in psci.h.
> > > What we are saying is for example that psci_operations.cpu_on returns 0
> > > on success and < 0 on failure, and it takes a cpuid and an entry point
> > > as parameters. We are not saying anything about the firmware interface.
> > 
> > I disagree. You're explicitly stating that we pass the `cpuid of target CPU,
> > as from MPIDR'. That's simply not true -- the firmware could choose any
> > numbering scheme to identify the CPUs. For KVM and Xen, it *is* the mpidr,
> > which is why psci-smp.c works at all, but that's where the comment belongs,
> > not in this header file.
> 
> At some point, the _kernel_ API for interfacing with the firmware's PSCI 
> will have to ensure uniformity somehow.  The PSCI interface code could 
> translate the passed MPIDR into whatever the firmware decided to use for 
> identifying CPUs if needed, keeping this issue localized.

That is what I had in mind when I said to keep the comment in psci.h
before.
We have to draw the line somewhere to expose a uniform internal kernel
API. However it is a bit difficult to do now given that we have only one
user of the API.

I don't feel to strongly about this, please let me know what is the
final decision and I'll update the code accordingly. I remind you that
the merge window is approaching :-)