[PATCH] arm64/pci: fix dma coherency inheritance for PCI devices

[msalter@redhat.com (Mark Salter)]

On Fri, 2014-09-19 at 17:28 +0100, Will Deacon wrote:
> On Fri, Sep 19, 2014 at 04:51:26PM +0100, Mark Salter wrote:
> > The default dma_ops for devices on arm64 systems are noncoherent in
> > nature and rely upon special operations and bounce buffers to
> > perform a device DMA operation to/from memory. Some drivers rely
> > upon coherent operations involving suitably capable hardware. In
> > this case, a "dma-coherent" property will exist on the corresponding
> > Device Tree node for the bridge device, or one of its ancestors.
> > This patch adds support for applying a DMA coherent dma_ops for
> > PCI devices in the case of such a property.
> > 
> > Signed-off-by: Jon Masters <jcm@redhat.com>
> > [added search for device with of_node]
> > Signed-off-by: Mark Salter <msalter@redhat.com>
> > ---
> >  arch/arm64/mm/dma-mapping.c | 21 +++++++++++++++++++--
> >  1 file changed, 19 insertions(+), 2 deletions(-)
> > 
> > diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
> > index 4164c5a..0e26bd7 100644
> > --- a/arch/arm64/mm/dma-mapping.c
> > +++ b/arch/arm64/mm/dma-mapping.c
> > @@ -23,10 +23,12 @@
> >  #include <linux/dma-mapping.h>
> >  #include <linux/dma-contiguous.h>
> >  #include <linux/of.h>
> > +#include <linux/of_address.h>
> >  #include <linux/platform_device.h>
> >  #include <linux/vmalloc.h>
> >  #include <linux/swiotlb.h>
> >  #include <linux/amba/bus.h>
> > +#include <linux/pci.h>
> >  
> >  #include <asm/cacheflush.h>
> >  
> > @@ -316,8 +318,18 @@ static int dma_bus_notifier(struct notifier_block *nb,
> >  	if (event != BUS_NOTIFY_ADD_DEVICE)
> >  		return NOTIFY_DONE;
> >  
> > -	if (of_property_read_bool(dev->of_node, "dma-coherent"))
> > -		set_dma_ops(dev, &coherent_swiotlb_dma_ops);
> > +	/*
> > +	 * Some devices won't have an of_node but a bus controller/bridge will.
> > +	 * Search up the device chain until we find an of_node to check.
> > +	 */
> > +	while (dev) {
> > +		if (dev->of_node) {
> > +			if (of_dma_is_coherent(dev->of_node))
> > +				set_dma_ops(_dev, &coherent_swiotlb_dma_ops);
> > +			break;
> > +		}
> > +		dev = dev->parent;
> > +	}
> 
> Hmm, how does this interact with of_dma_configure? I don't think the
> dma-coherent property is inherited there, so I'm uneasy about doing this for
> buses other than PCI (i.e. the platform bus).

of_dma_configure() only gets called for platform bus and it
uses of_dma_is_coherent() which searches device ancestors for
dma-coherent property. But it uses set_arch_dma_coherent_ops()
which is a no-op on arm64 now. Catalin mentions this in another
thread: https://lkml.org/lkml/2014/9/15/363

So maybe we should fix set_arch_dma_coherent_ops() and drop
platform bus from the notifier handler.

> 
> In other words, we should move to using of_dma_configure for the platform
> and amba buses, and have a separate notifier for the PCI bus. Is it likely
> that people will have a mixture of coherent and non-coherent devices behind
> the same host controller?
> 
> Will
> 
> >  	return NOTIFY_OK;
> >  }
> > @@ -330,6 +342,10 @@ static struct notifier_block amba_bus_nb = {
> >  	.notifier_call = dma_bus_notifier,
> >  };
> >  
> > +static struct notifier_block pci_bus_nb = {
> > +	.notifier_call = dma_bus_notifier,
> > +};
> > +
> >  extern int swiotlb_late_init_with_default_size(size_t default_size);
> >  
> >  static int __init swiotlb_late_init(void)
> > @@ -341,6 +357,7 @@ static int __init swiotlb_late_init(void)
> >  	 */
> >  	bus_register_notifier(&platform_bus_type, &platform_bus_nb);
> >  	bus_register_notifier(&amba_bustype, &amba_bus_nb);
> > +	bus_register_notifier(&pci_bus_type, &pci_bus_nb);
> >  
> >  	dma_ops = &noncoherent_swiotlb_dma_ops;
> >  
> > -- 
> > 1.9.3
> > 
> >