[RESEND]

On 5/5/15 15:36, Rafael J. Wysocki wrote:

On Tuesday, May 05, 2015 10:12:05 AM Suravee Suthikulpanit wrote:

quoted

diff --git a/drivers/acpi/Kconfig b/drivers/acpi/Kconfig
index ab2cbb5..dd386e9 100644
--- a/drivers/acpi/Kconfig
+++ b/drivers/acpi/Kconfig

@@ -54,6 +54,12 @@ config ACPI_GENERIC_GSI
  config ACPI_SYSTEM_POWER_STATES_SUPPORT
  	bool

+config ACPI_MUST_HAVE_CCA

ACPI_CCA_REQUIRED maybe?

Sure.

quoted

+	bool
+
+config ACPI_SUPPORT_CCA_ZERO

I guess this means "we support devices that can DMA, but are not coherent".
right?

Yes, basically when _CCA=0.

quoted

+	bool
+
  config ACPI_SLEEP
  	bool
  	depends on SUSPEND || HIBERNATION

diff --git a/drivers/acpi/acpi_platform.c b/drivers/acpi/acpi_platform.c
index 4bf7559..a6feca4 100644
--- a/drivers/acpi/acpi_platform.c
+++ b/drivers/acpi/acpi_platform.c

@@ -108,9 +108,11 @@ struct platform_device *acpi_create_platform_device(struct acpi_device *adev)
  	if (IS_ERR(pdev))
  		dev_err(&adev->dev, "platform device creation failed: %ld\n",
  			PTR_ERR(pdev));
-	else
+	else {

Please add braces to both branches when making such changes (as per CodingStyle).

OK.

quoted

+		acpi_setup_device_dma(adev, &pdev->dev);

Why do we need to do that here (for the second time)?

Because we are calling:
   acpi_create_platform_device()
     |--> platform_device_register_device_full()
       |-->platform_device_alloc()

This creates platform_device, which allocate a new platform_device->dev. 
This is not the same as the original acpi_device->dev that was created 
during acpi_add_single_object(). So, we have to set up the device 
coherency again.

quoted

diff --git a/drivers/acpi/scan.c b/drivers/acpi/scan.c
index 849b699..ac33b29 100644
--- a/drivers/acpi/scan.c
+++ b/drivers/acpi/scan.c

@@ -11,6 +11,7 @@
  #include <linux/kthread.h>
  #include <linux/dmi.h>
  #include <linux/nls.h>
+#include <linux/dma-mapping.h>

  #include <asm/pgtable.h>

@@ -2137,6 +2138,66 @@ void acpi_free_pnp_ids(struct acpi_device_pnp *pnp)
  	kfree(pnp->unique_id);
  }

+void acpi_setup_device_dma(struct acpi_device *adev, struct device *dev)
+{
+	int coherent = acpi_dma_is_coherent(adev);
+
+	/**
+	 * Currently, we only support DMA for devices that _CCA=1
+	 * since this seems to be the case on most ACPI platforms.
+	 *
+	 * For the case when _CCA=0 (i.e. is_coherent=0 && cca_seen=1),
+	 * we would rely on arch-specific cache maintenance for
+	 * non-coherence DMA operations if architecture enables
+	 * CONFIG_ACPI_SUPPORT_CCA_ZERO.
+	 *
+	 * For the case when _CCA is missing but platform requires it
+	 * (i.e. is_coherent=0 && cca_seen=0), we do not call
+	 * arch_setup_dma_ops() and fallback to arch-specific default
+	 * handling.
+	 */
+	if (adev->flags.cca_seen) {
+		if (!coherent && !IS_ENABLED(CONFIG_ACPI_SUPPORT_CCA_ZERO))
+			return;
+		arch_setup_dma_ops(dev, 0, 0, NULL, coherent);

Oh dear.

I made a mistake here. This logic should also call arch_setup_dma_ops() 
when cca_seen=0 and coherent=1 (e.g. when _CCA is not required and 
default to coherent when it is missing). The current logic doesn't do that.

What about

	if (adev->flags.cca_seen && (coherent || IS_ENABLED(CONFIG_ACPI_SUPPORT_CCA_ZERO)))
		arch_setup_dma_ops(dev, 0, 0, NULL, coherent);

What about:
	if (coherent ||
	    (adev->flags.cca_seen &&
			IS_ENABLED(CONFIG_ACPI_SUPPORT_CCA_ZERO))
		arch_setup_dma_ops(dev, 0, 0, NULL, coherent);
	

I wonder how this is going to affect x86/ia64 too?

This should not affect x86 since arch_setup_dma_ops() is currently not 
implement for x86, and default to NOP (see include/linux/dma-mapping.h). 
  Also, on x86, _CCA is not required and default to 1 if missing.

Thanks,

Suravee