Hi Marc,

On 05/05/16 15:13, Marc Zyngier wrote:

[...]

Gahhh. No. Please. Last time we did that, it took 6 months to untangle
the mess people made by adding their own hacks in this structure, 
so I definitely want to keep it completely private, forever. Same goes
for the gic_{dist,cpu.pm}_init() functions.

OK.

I've had a go at this, and came up with the following patch. I've only
briefly tested it on a host and a VM, so it is likely to break some stuff
somewhere, but you'll get the idea: The gic_chip_data struct is entirely
opaque, allocated by the GIC driver itself, with a few new fields in
it so that it becomes self-contained. This applies on top of your series.

It should also make it easy to switch to a model where we allocate
the structure dynamically instead of the old static crap.

Thoughts?

Yes I have been doing some testing and with a couple tweaks we can make
something like this work. One thing that caught me out was ...

+int gic_of_setup(struct device_node *node, struct device *dev,
+		 struct gic_chip_data **gicp)
+{
+	struct gic_chip_data *gic;
 
-	*cpu_base = of_iomap(node, 1);
-	if (WARN(!*cpu_base, "unable to map gic cpu registers\n")) {
-		iounmap(*dist_base);
-		return -ENOMEM;
+	if (!node || !gicp)
+		return -EINVAL;
+
+	if (dev) {
+		*gicp = devm_kzalloc(dev, sizeof(*gic), GFP_KERNEL);
+		if (!*gicp)
+			return -ENOMEM;
 	}
 
-	if (of_property_read_u32(node, "cpu-offset", percpu_offset))
-		*percpu_offset = 0;
+	gic = *gicp;
+
+	gic->raw_dist_base = of_iomap(node, 0);
+	if (WARN(!gic->raw_dist_base, "unable to map gic dist registers\n"))
+		goto err;
+
+	gic->raw_cpu_base = of_iomap(node, 1);
+	if (WARN(!gic->raw_cpu_base, "unable to map gic cpu registers\n"))
+		goto err;
+
+	if (of_property_read_u32(node, "cpu-offset", &gic->percpu_offset))
+		gic->percpu_offset = 0;
 
+	gic->chip.parent_device = dev;

We can't initialise the device here as it gets overwritten in the
gic_init_bases. So I have had to re-organise things a bit. Good news is
that I have eliminated the call from the platform driver to
gic_init_bases so we only have a single call to initialise the GIC.

Cheers
Jon