Re: [V2 PATCH 1/5] ACPI / scan: Parse _CCA and setup device coherency

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Tuesday, May 05, 2015 11:15:37 PM Suravee Suthikulpanit wrote:
> [RESEND]
> 
> On 5/5/15 15:36, Rafael J. Wysocki wrote:
> > On Tuesday, May 05, 2015 10:12:05 AM Suravee Suthikulpanit wrote:
> >> 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.
> 
> >
> >> +	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.

So what about

	ARCH_SUPPORT_CACHE_INCOHERENT_DMA

or something similar?

> >> +	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.
> 
> >> +		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.

Ah, so the second arg is different now.

Well, in that case, why do we need to set it up for the adev's dev member?

> >> 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);

Yes, that works.

> > 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). 

OK

>   Also, on x86, _CCA is not required and default to 1 if missing.

Well, that's the point. :-)


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