[PATCHv2 3/4] ARM: mm: add support for HW coherent systems in PL310

Dear Catalin Marinas,

On Wed, 14 May 2014 15:34:48 +0100, Catalin Marinas wrote:

On Tue, May 13, 2014 at 11:10:38AM +0100, Thomas Petazzoni wrote:

quoted

--- a/Documentation/devicetree/bindings/arm/l2cc.txt
+++ b/Documentation/devicetree/bindings/arm/l2cc.txt

@@ -8,6 +8,8 @@ Required properties:
 
 - compatible : should be one of:
   "arm,pl310-cache"
+  "arm,pl310-coherent-cache", used for I/O coherent platforms using
+     the PL310 cache
   "arm,l220-cache"
   "arm,l210-cache"
   "bcm,bcm11351-a2-pl310-cache": DEPRECATED by "brcm,bcm11351-a2-pl310-cache"

The binding name kind of implies that we have a transparent PL310 cache
(at least to me), which is not the case. I would rather have a specific
dma-coherent property or something similar since it's not another type
of PL310 but rather a different SoC topology.

Fine with me. A separate property or a different compatible string is
the same.

But I recall you mentioned that you can't make this decision at the DT
level since you don't know before whether the SoC is I/O coherent or
not.

As of today, hardware I/O coherency can only be enabled if CONFIG_SMP
is enabled *and* the SoC is actually multi-core, due to limitations in
the ARM core code. So if you look at my PATCH 4/4 in this series, you
will see that I adjust the compatible string of the cache controller at
run time, depending on whether hardware I/O coherency is enabled or not:

+static void __init mvebu_l2x0_pl310_coherent(void)
+{
+	struct device_node *np;
+
+	if (!coherency_available())
+		return;
+
+	for_each_compatible_node(np, NULL, "arm,pl310-cache") {
+		struct property *new_compat;
+
+		new_compat = kzalloc(sizeof(*new_compat), GFP_KERNEL);
+		new_compat->name = kstrdup("compatible", GFP_KERNEL);
+		new_compat->value = kstrdup("arm,pl310-coherent-cache",
+					    GFP_KERNEL);
+		new_compat->length = strlen(new_compat->value) + 1;
+		of_update_property(np, new_compat);
+	}
+}

So, this code can just as well be changed to add a property instead of
adjusting the compatible property. It's the same thing for me.

Also, I will send (hopefully today) a series that allows to enable
hardware I/O coherency even on !SMP configurations. But I believe it
might take a while to get merged.

Best regards,

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com