On Friday, May 22, 2015 07:15:17 PM Suravee Suthikulanit wrote:

On 5/22/2015 6:05 PM, Rafael J. Wysocki wrote:

quoted

On Friday, May 22, 2015 05:24:15 PM Suravee Suthikulanit wrote:

quoted

Not sure if this went out earlier. So I am resending.

On 5/22/15 16:56, Rafael J. Wysocki wrote:

quoted

quoted

diff --git a/drivers/acpi/glue.c b/drivers/acpi/glue.c

quoted

index 39c485b..b9657af 100644

--- a/drivers/acpi/glue.c
+++ b/drivers/acpi/glue.c

@@ -13,6 +13,7 @@
  #include <linux/slab.h>
  #include <linux/rwsem.h>
  #include <linux/acpi.h>
+#include <linux/dma-mapping.h>

  #include "internal.h"

@@ -167,6 +168,7 @@ int acpi_bind_one(struct device *dev, struct acpi_device *acpi_dev)
  	struct list_head *physnode_list;
  	unsigned int node_id;
  	int retval = -EINVAL;
+	bool coherent;

  	if (has_acpi_companion(dev)) {
  		if (acpi_dev) {

@@ -223,6 +225,9 @@ int acpi_bind_one(struct device *dev, struct acpi_device *acpi_dev)
  	if (!has_acpi_companion(dev))
  		ACPI_COMPANION_SET(dev, acpi_dev);

+	if (acpi_check_dma(acpi_dev, &coherent))
+		arch_setup_dma_ops(dev, 0, 0, NULL, coherent);
+

Well, so is this going to work for PCI too after all?

No, as Bjorn suggested, PCI changes for setting DMA coherent from _CCA
(patch 3/6 in V4) will be submitted separately. We are working on
cleaning up and up-streaming the PCI ACPI support for ARM64.

OK, but acpi_bind_one() is called for PCI devices too.  Won't that be a problem?

 >
In this case, we would be going through the following call path:

   --> pci_device_add()
    |--> pci_dma_configure() ** 1 **
    |--> device_add()
      |--> platform_notify()
        |--> acpi_platform_notify()
         |--> acpi_bind_one() ** 2 **

At (1), we would be calling arch_setup_dma_ops() with the PCI host 
bridge _CCA information. So, it should have already taken care of 
setting up device coherency here.

At (2), if there is no acpi_dev for endpoint devices (which I believe 
this is normally the case), it would return early and skip 
arch_setup_dma_ops().

That's not correct.  There may be ACPI companions for endpoint devices too.

At (2), if there is an acpi_dev, the coherent_dma flag should have 
already been setup by the acpi_init_device_object during ACPI scan.

That one sets the flag for the *ACPI* *companion* of the device, which
I'm still thinking is pointless, isn't it?

However, I am not certain about this case since I don't have the DSDT 
containing  PCI endpoint devices to test with.

Every x86 PC has them (as far as I can say), but in that case there's no
_CCA and they are all coherent.


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.