On 5/20/2015 5:01 AM, Catalin Marinas wrote:

On Fri, May 15, 2015 at 04:23:09PM -0500, Suravee Suthikulpanit wrote:

quoted

+static inline bool acpi_dma_is_supported(struct acpi_device *adev)
+{
+	/**
+	 * Currently, we mainly support _CCA=1 (i.e. is_coherent=1)
+	 * This should be equivalent to specifyig dma-coherent for
+	 * a device in OF.
+	 *
+	 * For the case when _CCA=0 (i.e. is_coherent=0 && cca_seen=1),
+	 * There are two approaches:
+	 * 1. Do not support and disable DMA.
+	 * 2. Support but rely on arch-specific cache maintenance for
+	 * non-coherence DMA operations. ARM64 is one example.
+	 *
+	 * For the case when _CCA is missing (i.e. cca_seen=0) but
+	 * platform specifies ACPI_CCA_REQUIRED, we do not support DMA,
+	 * and fallback to arch-specific default handling.
+	 *
+	 * See acpi_init_coherency() for more info.
+	 */
+	return adev && (adev->flags.is_coherent ||
+			(adev->flags.cca_seen && IS_ENABLED(CONFIG_ARM64)));
+}

I don't particularly like the check for CONFIG_ARM64 here but I
understand why it was added (I had the wrong impression that x86 can
cope with _CCA = 0).

Alternatively, we could leave it out (together with cca_seen) until
someone comes forward with a real use-case for _CCA = 0 on arm64. One
platform I'm aware of is Juno but even though it boot with ACPI, I
wouldn't call it a server platform.

Ok. That seems to be what Arnd would prefer as well.  Let's just leave 
the support for _CCA=0 out until it is needed then.

Thanks,
Suravee