[PATCH v7 3/8] arm64: introduce is_device_dma_coherent

[catalin.marinas@arm.com (Catalin Marinas)]

On Wed, Nov 05, 2014 at 06:15:38PM +0000, Stefano Stabellini wrote:
> On Wed, 5 Nov 2014, Catalin Marinas wrote:
> > On Mon, Nov 03, 2014 at 11:10:18AM +0000, Stefano Stabellini wrote:
> > > On Mon, 3 Nov 2014, Will Deacon wrote:
> > > > On Mon, Nov 03, 2014 at 10:46:03AM +0000, Stefano Stabellini wrote:
> > > > > On Mon, 27 Oct 2014, Stefano Stabellini wrote:
> > > > > > Introduce a boolean flag and an accessor function to check whether a
> > > > > > device is dma_coherent. Set the flag from set_arch_dma_coherent_ops.
> > > > > > 
> > > > > > Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
> > > > > > Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
> > > > > > CC: will.deacon at arm.com
> > > > > 
> > > > > Will, Catalin,
> > > > > are you OK with this patch?
> > > > 
> > > > It would be nicer if the dma_coherent flag didn't have to be duplicated by
> > > > each architecture in dev_archdata. Is there any reason not to put it in the
> > > > core code?
> > > 
> > > Yes, there is a reason for it: if I added a boolean dma_coherent flag in
> > > struct device as Catalin initially suggested, what would be the default
> > > for each architecture? Where would I set it for arch that don't use
> > > device tree?
> > 
> > You don't need to. An architecture that has coherent DMA always doesn't
> > need to do anything. One that has non-coherent DMA always only needs to
> > select HAVE_DMA_NONCOHERENT. One that has a mix of both needs to find a
> > way to set dev->dma_coherent. Since that's a new API you introduce, it
> > doesn't break any existing architectures.
> 
> I am not sure that this is better than the current patch but I can see
> that this approach is not too controversial, so I am happy to go with
> whatever the maintainers prefer.

Functionally it is the same, but just less code duplication.

> > Note that if !is_device_dma_coherent(), it doesn't always mean that
> > standard cache maintenance would be enough (but that's a Xen problem,
> > not sure how to solve).
> 
> It is a thorny issue indeed.
> Xen would need to know how to do non-standard cache maintenance
> operations.

Is EL2 hyp or EL1 dom0 doing such maintenance? If the latter, you could
just use the currently registered dma ops.

> Otherwise we would need to resurrect XENFEAT_grant_map_identity (that I
> am reverting in this series) and be content with having CONFIG_XEN_DOM0
> depend on CONFIG_ARM_LPAE.

So what does buy you? Is it just the hope that with LPAE you won't have
weird system caches?

> > diff --git a/arch/Kconfig b/arch/Kconfig
> > index 05d7a8a458d5..8462b2e7491b 100644
> > --- a/arch/Kconfig
> > +++ b/arch/Kconfig
> > @@ -203,6 +203,9 @@ config HAVE_DMA_ATTRS
> >  config HAVE_DMA_CONTIGUOUS
> >  	bool
> >  
> > +config HAVE_DMA_NONCOHERENT
> > +	bool
> > +
> >  config GENERIC_SMP_IDLE_THREAD
> >         bool
> >  
> > diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
> > index 89c4b5ccc68d..fd7d5522764c 100644
> > --- a/arch/arm/Kconfig
> > +++ b/arch/arm/Kconfig
> > @@ -40,6 +40,7 @@ config ARM
> >  	select HAVE_DMA_API_DEBUG
> >  	select HAVE_DMA_ATTRS
> >  	select HAVE_DMA_CONTIGUOUS if MMU
> > +	select HAVE_DMA_NONCOHERENT if OF
> >  	select HAVE_DYNAMIC_FTRACE if (!XIP_KERNEL)
> >  	select HAVE_EFFICIENT_UNALIGNED_ACCESS if (CPU_V6 || CPU_V6K || CPU_V7) && MMU
> >  	select HAVE_FTRACE_MCOUNT_RECORD if (!XIP_KERNEL)
> > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > index 9532f8d5857e..eb7a5aa64e0e 100644
> > --- a/arch/arm64/Kconfig
> > +++ b/arch/arm64/Kconfig
> > @@ -46,6 +46,7 @@ config ARM64
> >  	select HAVE_DMA_API_DEBUG
> >  	select HAVE_DMA_ATTRS
> >  	select HAVE_DMA_CONTIGUOUS
> > +	select HAVE_DMA_NONCOHERENT
> >  	select HAVE_DYNAMIC_FTRACE
> >  	select HAVE_EFFICIENT_UNALIGNED_ACCESS
> >  	select HAVE_FTRACE_MCOUNT_RECORD
> > diff --git a/drivers/of/platform.c b/drivers/of/platform.c
> > index 3b64d0bf5bba..7e827726b702 100644
> > --- a/drivers/of/platform.c
> > +++ b/drivers/of/platform.c
> > @@ -183,6 +183,7 @@ static void of_dma_configure(struct device *dev)
> >  	 * dma coherent operations.
> >  	 */
> >  	if (of_dma_is_coherent(dev->of_node)) {
> > +		dev->dma_coherent = true;
> >  		set_arch_dma_coherent_ops(dev);
> >  		dev_dbg(dev, "device is dma coherent\n");
> >  	}
> 
> I think that this would need to be #ifdef'ed as it is possible to have
> OF support but no HAVE_DMA_NONCOHERENT (PPC?).

The field is always there. But with !HAVE_DMA_NONCOHERENT,
is_device_dma_coherent() would always return 1. You could avoid
defining is_device_dma_coherent() entirely when !HAVE_DMA_NONCOHERENT,
it wouldn't be worse than your patch in terms of an undefined function.

> > diff --git a/include/linux/device.h b/include/linux/device.h
> > index ce1f21608b16..e00ca876db01 100644
> > --- a/include/linux/device.h
> > +++ b/include/linux/device.h
> > @@ -796,6 +796,7 @@ struct device {
> >  
> >  	bool			offline_disabled:1;
> >  	bool			offline:1;
> > +	bool			dma_coherent:1;
> >  };
> 
> I guess we would have to #ifdef CONFIG_HAVE_DMA_NONCOHERENT the
> dma_coherent flag, right? Otherwise architecures that do not select
> CONFIG_HAVE_DMA_NONCOHERENT (x86 for example) would end up with a flag
> in struct device that doesn't reflect the properties of the device (dma
> coherent devices with dev->dma_coherent == 0).

In my proposal you should not read this field directly but rather access
it only via is_device_dma_coherent() (you can add a function for setting
it as well).

-- 
Catalin