Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1345590 > unrolled thread

[PATCH 1/2] iommu/mediatek: select ARM_DMA_USE_IOMMU

Back to article view | Back to linux.kernel

Contents

  [PATCH 1/2] iommu/mediatek: select ARM_DMA_USE_IOMMU Arnd Bergmann <arnd@arndb.de> - 2016-02-29 10:30 +0100
    [PATCH 2/2] iommu/mediatek: mark PM functions as __maybe_unused Arnd Bergmann <arnd@arndb.de> - 2016-02-29 10:30 +0100
      Re: [PATCH 2/2] iommu/mediatek: mark PM functions as __maybe_unused Yong Wu <yong.wu@mediatek.com> - 2016-03-01 06:10 +0100
    Re: [PATCH 1/2] iommu/mediatek: select ARM_DMA_USE_IOMMU Robin Murphy <robin.murphy@arm.com> - 2016-02-29 11:40 +0100
      Re: [PATCH 1/2] iommu/mediatek: select ARM_DMA_USE_IOMMU Arnd Bergmann <arnd@arndb.de> - 2016-02-29 12:00 +0100
        Re: [PATCH 1/2] iommu/mediatek: select ARM_DMA_USE_IOMMU Robin Murphy <robin.murphy@arm.com> - 2016-02-29 12:30 +0100
          Re: [PATCH 1/2] iommu/mediatek: select ARM_DMA_USE_IOMMU Arnd Bergmann <arnd@arndb.de> - 2016-02-29 12:40 +0100
            Re: [PATCH 1/2] iommu/mediatek: select ARM_DMA_USE_IOMMU Robin Murphy <robin.murphy@arm.com> - 2016-02-29 12:50 +0100
    Re: [PATCH 1/2] iommu/mediatek: select ARM_DMA_USE_IOMMU Joerg Roedel <joro@8bytes.org> - 2016-02-29 16:50 +0100

#1345590 — [PATCH 1/2] iommu/mediatek: select ARM_DMA_USE_IOMMU

The newly added Mediatek IOMMU driver uses the IOMMU_DMA infrastructure,
but unlike other such drivers, it does not select 'ARM_DMA_USE_IOMMU',
which is a prerequisite, leading to a link error:

warning: (MTK_IOMMU) selects IOMMU_DMA which has unmet direct dependencies (IOMMU_SUPPORT && NEED_SG_DMA_LENGTH)
drivers/iommu/built-in.o: In function `iommu_put_dma_cookie':
mtk_iommu.c:(.text+0x11fe): undefined reference to `put_iova_domain'
drivers/iommu/built-in.o: In function `iommu_dma_init_domain':
mtk_iommu.c:(.text+0x1316): undefined reference to `init_iova_domain'
drivers/iommu/built-in.o: In function `__iommu_dma_unmap':
mtk_iommu.c:(.text+0x1380): undefined reference to `find_iova'

This adds the same select that the other drivers have. On a related
note, I wonder if we should just always select ARM_DMA_USE_IOMMU
whenever any IOMMU driver is enabled. Are there any cases where
we would enable an IOMMU but not use it?

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Fixes: 0df4fabe208d ("iommu/mediatek: Add mt8173 IOMMU driver")
---
 drivers/iommu/Kconfig | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig
index b325954cf8f8..ea0998921702 100644
--- a/drivers/iommu/Kconfig
+++ b/drivers/iommu/Kconfig
@@ -341,6 +341,7 @@ config MTK_IOMMU
 	bool "MTK IOMMU Support"
 	depends on ARM || ARM64
 	depends on ARCH_MEDIATEK || COMPILE_TEST
+	select ARM_DMA_USE_IOMMU
 	select IOMMU_API
 	select IOMMU_DMA
 	select IOMMU_IO_PGTABLE_ARMV7S
-- 
2.7.0

[toc] | [next] | [standalone]

#1345594 — [PATCH 2/2] iommu/mediatek: mark PM functions as __maybe_unused

When CONFIG_PM is unset, we get a harmless warning for this driver:

drivers/iommu/mtk_iommu.c:665:12: error: 'mtk_iommu_suspend' defined but not used [-Werror=unused-function]
drivers/iommu/mtk_iommu.c:680:12: error: 'mtk_iommu_resume' defined but not used [-Werror=unused-function]

Marking the functions as __maybe_unused gits rid of the two functions
and lets the compiler silently drop the object code, while still
doing syntax checking on them for build-time verification.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Fixes: 0df4fabe208d ("iommu/mediatek: Add mt8173 IOMMU driver")
---
 drivers/iommu/mtk_iommu.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/iommu/mtk_iommu.c b/drivers/iommu/mtk_iommu.c
index 721ffdb296d6..f3c160e4c25d 100644
--- a/drivers/iommu/mtk_iommu.c
+++ b/drivers/iommu/mtk_iommu.c
@@ -662,7 +662,7 @@ static int mtk_iommu_remove(struct platform_device *pdev)
 	return 0;
 }
 
-static int mtk_iommu_suspend(struct device *dev)
+static int __maybe_unused mtk_iommu_suspend(struct device *dev)
 {
 	struct mtk_iommu_data *data = dev_get_drvdata(dev);
 	struct mtk_iommu_suspend_reg *reg = &data->reg;
@@ -677,7 +677,7 @@ static int mtk_iommu_suspend(struct device *dev)
 	return 0;
 }
 
-static int mtk_iommu_resume(struct device *dev)
+static int __maybe_unused mtk_iommu_resume(struct device *dev)
 {
 	struct mtk_iommu_data *data = dev_get_drvdata(dev);
 	struct mtk_iommu_suspend_reg *reg = &data->reg;
-- 
2.7.0

[toc] | [prev] | [next] | [standalone]

#1346367 — Re: [PATCH 2/2] iommu/mediatek: mark PM functions as __maybe_unused

On Mon, 2016-02-29 at 10:19 +0100, Arnd Bergmann wrote:
> When CONFIG_PM is unset, we get a harmless warning for this driver:
> 
> drivers/iommu/mtk_iommu.c:665:12: error: 'mtk_iommu_suspend' defined but not used [-Werror=unused-function]
> drivers/iommu/mtk_iommu.c:680:12: error: 'mtk_iommu_resume' defined but not used [-Werror=unused-function]
> 
> Marking the functions as __maybe_unused gits rid of the two functions
> and lets the compiler silently drop the object code, while still
> doing syntax checking on them for build-time verification.
> 
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> Fixes: 0df4fabe208d ("iommu/mediatek: Add mt8173 IOMMU driver")

Hi Arnd,
  Thanks very much for both fixes.

[toc] | [prev] | [next] | [standalone]

#1345635

Hi Arnd,

On 29/02/16 09:19, Arnd Bergmann wrote:
> The newly added Mediatek IOMMU driver uses the IOMMU_DMA infrastructure,
> but unlike other such drivers, it does not select 'ARM_DMA_USE_IOMMU',
> which is a prerequisite, leading to a link error:
>
> warning: (MTK_IOMMU) selects IOMMU_DMA which has unmet direct dependencies (IOMMU_SUPPORT && NEED_SG_DMA_LENGTH)

Going off on a tangent, is it actually right for that to depend on a 
NEED_* symbol, or should that really be a select instead?

> drivers/iommu/built-in.o: In function `iommu_put_dma_cookie':
> mtk_iommu.c:(.text+0x11fe): undefined reference to `put_iova_domain'
> drivers/iommu/built-in.o: In function `iommu_dma_init_domain':
> mtk_iommu.c:(.text+0x1316): undefined reference to `init_iova_domain'
> drivers/iommu/built-in.o: In function `__iommu_dma_unmap':
> mtk_iommu.c:(.text+0x1380): undefined reference to `find_iova'
>
> This adds the same select that the other drivers have. On a related
> note, I wonder if we should just always select ARM_DMA_USE_IOMMU
> whenever any IOMMU driver is enabled. Are there any cases where
> we would enable an IOMMU but not use it?

You could use one solely for VFIO without caring about DMA ops - I think 
that's mostly how ARM SMMUs are being used in practice at the moment - 
but DMA-focused 'media' IOMMUs vastly outnumber 'virtualisation' IOMMUs 
on ARM, so it would probably make sense. We already have the equivalent 
"select IOMMU_DMA if IOMMU_SUPPORT" on arm64.

> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> Fixes: 0df4fabe208d ("iommu/mediatek: Add mt8173 IOMMU driver")
> ---
>   drivers/iommu/Kconfig | 1 +
>   1 file changed, 1 insertion(+)
>
> diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig
> index b325954cf8f8..ea0998921702 100644
> --- a/drivers/iommu/Kconfig
> +++ b/drivers/iommu/Kconfig
> @@ -341,6 +341,7 @@ config MTK_IOMMU
>   	bool "MTK IOMMU Support"
>   	depends on ARM || ARM64
>   	depends on ARCH_MEDIATEK || COMPILE_TEST
> +	select ARM_DMA_USE_IOMMU

If going down this route, I'd be inclined to add an "if ARM" there, just 
for clarity.

>   	select IOMMU_API
>   	select IOMMU_DMA

As above, this is already selected on arm64, and isn't used on 32-bit*, 
so could probably just be removed, especially if it leads to build issues.

Robin.

*yet, of course. I need to have a proper look over Marek's RFC ;)

>   	select IOMMU_IO_PGTABLE_ARMV7S
>

[toc] | [prev] | [next] | [standalone]

#1345645

On Monday 29 February 2016 10:37:40 Robin Murphy wrote:
> Hi Arnd,
> 
> On 29/02/16 09:19, Arnd Bergmann wrote:
> > The newly added Mediatek IOMMU driver uses the IOMMU_DMA infrastructure,
> > but unlike other such drivers, it does not select 'ARM_DMA_USE_IOMMU',
> > which is a prerequisite, leading to a link error:
> >
> > warning: (MTK_IOMMU) selects IOMMU_DMA which has unmet direct dependencies (IOMMU_SUPPORT && NEED_SG_DMA_LENGTH)
> 
> Going off on a tangent, is it actually right for that to depend on a 
> NEED_* symbol, or should that really be a select instead?

ARM_DMA_USE_IOMMU gets this right, it selects NEED_SG_DMA_LENGTH as it
actually needs it.

The IOMMU_DMA symbol is a bit strange, and the dependency can probably
get dropped altogether, but at least here it told us what went wrong.

> > drivers/iommu/built-in.o: In function `iommu_put_dma_cookie':
> > mtk_iommu.c:(.text+0x11fe): undefined reference to `put_iova_domain'
> > drivers/iommu/built-in.o: In function `iommu_dma_init_domain':
> > mtk_iommu.c:(.text+0x1316): undefined reference to `init_iova_domain'
> > drivers/iommu/built-in.o: In function `__iommu_dma_unmap':
> > mtk_iommu.c:(.text+0x1380): undefined reference to `find_iova'
> >
> > This adds the same select that the other drivers have. On a related
> > note, I wonder if we should just always select ARM_DMA_USE_IOMMU
> > whenever any IOMMU driver is enabled. Are there any cases where
> > we would enable an IOMMU but not use it?
> 
> You could use one solely for VFIO without caring about DMA ops - I think 
> that's mostly how ARM SMMUs are being used in practice at the moment - 
> but DMA-focused 'media' IOMMUs vastly outnumber 'virtualisation' IOMMUs 
> on ARM, so it would probably make sense. We already have the equivalent 
> "select IOMMU_DMA if IOMMU_SUPPORT" on arm64.

Ok.

> > Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> > Fixes: 0df4fabe208d ("iommu/mediatek: Add mt8173 IOMMU driver")
> > ---
> >   drivers/iommu/Kconfig | 1 +
> >   1 file changed, 1 insertion(+)
> >
> > diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig
> > index b325954cf8f8..ea0998921702 100644
> > --- a/drivers/iommu/Kconfig
> > +++ b/drivers/iommu/Kconfig
> > @@ -341,6 +341,7 @@ config MTK_IOMMU
> >   	bool "MTK IOMMU Support"
> >   	depends on ARM || ARM64
> >   	depends on ARCH_MEDIATEK || COMPILE_TEST
> > +	select ARM_DMA_USE_IOMMU
> 
> If going down this route, I'd be inclined to add an "if ARM" there, just 
> for clarity.

That would run into the NEED_SG_DMA_LENGTH problem on other architectures
that don't already set it, right?

	Arnd

[toc] | [prev] | [next] | [standalone]

#1345707

On 29/02/16 10:53, Arnd Bergmann wrote:
> On Monday 29 February 2016 10:37:40 Robin Murphy wrote:
>> Hi Arnd,
>>
>> On 29/02/16 09:19, Arnd Bergmann wrote:
>>> The newly added Mediatek IOMMU driver uses the IOMMU_DMA infrastructure,
>>> but unlike other such drivers, it does not select 'ARM_DMA_USE_IOMMU',
>>> which is a prerequisite, leading to a link error:
>>>
>>> warning: (MTK_IOMMU) selects IOMMU_DMA which has unmet direct dependencies (IOMMU_SUPPORT && NEED_SG_DMA_LENGTH)
>>
>> Going off on a tangent, is it actually right for that to depend on a
>> NEED_* symbol, or should that really be a select instead?
>
> ARM_DMA_USE_IOMMU gets this right, it selects NEED_SG_DMA_LENGTH as it
> actually needs it.
>
> The IOMMU_DMA symbol is a bit strange, and the dependency can probably
> get dropped altogether, but at least here it told us what went wrong.

IOMMU_DMA uses sg_dma_len() unconditionally all over the place, hence 
the "dependency". Thanks for the clarification; fix sent.

>>> drivers/iommu/built-in.o: In function `iommu_put_dma_cookie':
>>> mtk_iommu.c:(.text+0x11fe): undefined reference to `put_iova_domain'
>>> drivers/iommu/built-in.o: In function `iommu_dma_init_domain':
>>> mtk_iommu.c:(.text+0x1316): undefined reference to `init_iova_domain'
>>> drivers/iommu/built-in.o: In function `__iommu_dma_unmap':
>>> mtk_iommu.c:(.text+0x1380): undefined reference to `find_iova'
>>>
>>> This adds the same select that the other drivers have. On a related
>>> note, I wonder if we should just always select ARM_DMA_USE_IOMMU
>>> whenever any IOMMU driver is enabled. Are there any cases where
>>> we would enable an IOMMU but not use it?
>>
>> You could use one solely for VFIO without caring about DMA ops - I think
>> that's mostly how ARM SMMUs are being used in practice at the moment -
>> but DMA-focused 'media' IOMMUs vastly outnumber 'virtualisation' IOMMUs
>> on ARM, so it would probably make sense. We already have the equivalent
>> "select IOMMU_DMA if IOMMU_SUPPORT" on arm64.
>
> Ok.
>
>>> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
>>> Fixes: 0df4fabe208d ("iommu/mediatek: Add mt8173 IOMMU driver")
>>> ---
>>>    drivers/iommu/Kconfig | 1 +
>>>    1 file changed, 1 insertion(+)
>>>
>>> diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig
>>> index b325954cf8f8..ea0998921702 100644
>>> --- a/drivers/iommu/Kconfig
>>> +++ b/drivers/iommu/Kconfig
>>> @@ -341,6 +341,7 @@ config MTK_IOMMU
>>>    	bool "MTK IOMMU Support"
>>>    	depends on ARM || ARM64
>>>    	depends on ARCH_MEDIATEK || COMPILE_TEST
>>> +	select ARM_DMA_USE_IOMMU
>>
>> If going down this route, I'd be inclined to add an "if ARM" there, just
>> for clarity.
>
> That would run into the NEED_SG_DMA_LENGTH problem on other architectures
> that don't already set it, right?

Sorry, I'm lost - wouldn't "depends on ARM || ARM64" make other 
architectures moot? arm64 already has NEED_SG_DMA_LENGTH=y by default.

Robin.

[toc] | [prev] | [next] | [standalone]

#1345726

On Monday 29 February 2016 11:22:24 Robin Murphy wrote:
> >>> diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig
> >>> index b325954cf8f8..ea0998921702 100644
> >>> --- a/drivers/iommu/Kconfig
> >>> +++ b/drivers/iommu/Kconfig
> >>> @@ -341,6 +341,7 @@ config MTK_IOMMU
> >>>     bool "MTK IOMMU Support"
> >>>     depends on ARM || ARM64
> >>>     depends on ARCH_MEDIATEK || COMPILE_TEST
> >>> +   select ARM_DMA_USE_IOMMU
> >>
> >> If going down this route, I'd be inclined to add an "if ARM" there, just
> >> for clarity.
> >
> > That would run into the NEED_SG_DMA_LENGTH problem on other architectures
> > that don't already set it, right?
> 
> Sorry, I'm lost - wouldn't "depends on ARM || ARM64" make other 
> architectures moot? arm64 already has NEED_SG_DMA_LENGTH=y by default.
> 

Nevermind, I didn't notice the dependency on the architecture.

What is keeping us from having 'depends on ARM || ARM64 || COMPILE_TEST'?

I assume it doesn't work yet, but it would be nice to get that done
at some point so we can take advantage of automated build testing like
coverity.

	Arnd

[toc] | [prev] | [next] | [standalone]

#1345732

On 29/02/16 11:29, Arnd Bergmann wrote:
> On Monday 29 February 2016 11:22:24 Robin Murphy wrote:
>>>>> diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig
>>>>> index b325954cf8f8..ea0998921702 100644
>>>>> --- a/drivers/iommu/Kconfig
>>>>> +++ b/drivers/iommu/Kconfig
>>>>> @@ -341,6 +341,7 @@ config MTK_IOMMU
>>>>>      bool "MTK IOMMU Support"
>>>>>      depends on ARM || ARM64
>>>>>      depends on ARCH_MEDIATEK || COMPILE_TEST
>>>>> +   select ARM_DMA_USE_IOMMU
>>>>
>>>> If going down this route, I'd be inclined to add an "if ARM" there, just
>>>> for clarity.
>>>
>>> That would run into the NEED_SG_DMA_LENGTH problem on other architectures
>>> that don't already set it, right?
>>
>> Sorry, I'm lost - wouldn't "depends on ARM || ARM64" make other
>> architectures moot? arm64 already has NEED_SG_DMA_LENGTH=y by default.
>>
>
> Nevermind, I didn't notice the dependency on the architecture.
>
> What is keeping us from having 'depends on ARM || ARM64 || COMPILE_TEST'?

IIRC it got changed because the kbuild test robot was consistently 
blowing up due to missing definitions on things like parisc.

Robin.

> I assume it doesn't work yet, but it would be nice to get that done
> at some point so we can take advantage of automated build testing like
> coverity.
>
> 	Arnd
>

[toc] | [prev] | [next] | [standalone]

#1345928

On Mon, Feb 29, 2016 at 10:19:06AM +0100, Arnd Bergmann wrote:
> The newly added Mediatek IOMMU driver uses the IOMMU_DMA infrastructure,
> but unlike other such drivers, it does not select 'ARM_DMA_USE_IOMMU',
> which is a prerequisite, leading to a link error:
> 
> warning: (MTK_IOMMU) selects IOMMU_DMA which has unmet direct dependencies (IOMMU_SUPPORT && NEED_SG_DMA_LENGTH)
> drivers/iommu/built-in.o: In function `iommu_put_dma_cookie':
> mtk_iommu.c:(.text+0x11fe): undefined reference to `put_iova_domain'
> drivers/iommu/built-in.o: In function `iommu_dma_init_domain':
> mtk_iommu.c:(.text+0x1316): undefined reference to `init_iova_domain'
> drivers/iommu/built-in.o: In function `__iommu_dma_unmap':
> mtk_iommu.c:(.text+0x1380): undefined reference to `find_iova'
> 
> This adds the same select that the other drivers have. On a related
> note, I wonder if we should just always select ARM_DMA_USE_IOMMU
> whenever any IOMMU driver is enabled. Are there any cases where
> we would enable an IOMMU but not use it?
> 
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> Fixes: 0df4fabe208d ("iommu/mediatek: Add mt8173 IOMMU driver")
> ---
>  drivers/iommu/Kconfig | 1 +
>  1 file changed, 1 insertion(+)

Applied both, thanks.

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web