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

[Robin Murphy]

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@xxxxxxxx>
> > > 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.