Hi Geoff,

On Wed, Jun 25, 2014 at 09:52:00PM +0100, Geoff Levand wrote:

Hi Lorenzo,

On Wed, 2014-06-25 at 15:10 +0100, Lorenzo Pieralisi wrote:

quoted

diff --git a/arch/arm64/kernel/psci.c b/arch/arm64/kernel/psci.c

quoted

+	/*
+	 * This is belt-and-braces: make sure that if the idle
+	 * specified protocol is psci, the cpu_ops have been
+	 * initialized to psci operations. Anything else is
+	 * a recipe for mayhem.
+	 */
+	for_each_cpu(cpu, drv->cpumask) {
+		cpu_ops_ptr = cpu_ops[cpu];
+		if (WARN_ON(!cpu_ops_ptr || strcmp(cpu_ops_ptr->name, "psci")))
+			return -EOPNOTSUPP;
+	}

I'm not sure how drv->cpumask is setup, but if a system has mixed enable
methods, say some cpus 'spin-table' and some 'psci', will this give a
false error?

I do not think that's a false error. If the idle states specify an
entry-method == psci, and cpu_ops for some cpus are not set to PSCI,
obviously because the enable-method specified that, that's a firmware bug.

If drv->cpumask should only include 'psci' cpus, then should this be a
BUG()?

Ok, if we got here, it is because the idle-states entry-method was set
to PSCI. Now, if any of the CPUs in the driver mask has a cpu_ops
pointer != PSCI, we have a problem and we should warn on that. I do
not think that justifies a BUG_ON, but that's one of those things, it is
debatable.

Question is whether the check should also be carried out at cpu_ops
initialization (ie to check for mixed cpu_ops), for certain if the
idle states entry-method is PSCI and cpu_ops != PSCI we should
WARN/BUG on that. Or embed this idle state parameters initialization at
cpu_ops init (see other thread you started) so that we can kill two
birds with one stone.

quoted

+
+	psci_states = kcalloc(drv->state_count, sizeof(*psci_states),
+			      GFP_KERNEL);
+
+	if (!psci_states) {
+		pr_warn("psci idle state allocation failed\n");
+		return -ENOMEM;
+	}
+
+	for_each_cpu(cpu, drv->cpumask) {
+		if (per_cpu(psci_power_state, cpu)) {
+			pr_warn("idle states already initialized on cpu %u\n",
+				cpu);

This seems like an implementation problem, if so, shouldn't this be
pr_debug()?

Maybe, I will give it some thought.

quoted

 #endif
 
+#ifdef CONFIG_ARM64_CPU_SUSPEND
+static int cpu_psci_cpu_suspend(unsigned long index)
+{
+	struct psci_power_state *state = __get_cpu_var(psci_power_state);
+
+	if (!state)
+		return -EOPNOTSUPP;
+
+	return psci_ops.cpu_suspend(state[index], virt_to_phys(cpu_resume));
+}
+#endif

Why not put a __maybe_unused attribute on cpu_psci_cpu_suspend() and
remove the preprocessor conditional.  That way this code will always be
compiled, and with therefor always get a build test.  The linker should
strip out the unused code when CONFIG_ARM64_CPU_SUSPEND=n and the code
below is not compiled. 

It can make sense, yes.

Thanks,
Lorenzo