DMABOUNCE in pci-rcar

[arnd@arndb.de (Arnd Bergmann)]

On Thursday 20 March 2014 17:12:43 Ben Dooks wrote:
> On 20/03/14 17:09, Arnd Bergmann wrote:
> > On Thursday 20 March 2014 16:04:36 Ben Dooks wrote:

> >> As a note, if we now boot a Lager with DT on 3.14-rc3 series the USB
> >> controllers no longer work with full 2GiB RAM enabled in the device
> >> tree.
> >
> > Did it work before these patches got applied initially?
> 
> It did, but I cannot remember if we where limiting DRAM to 1GiB
> or not.

I would assume you did, or you happened to never actually use
memory higher than that for DMA during tests.

> >> Could we work around this by having 1GiB of memory defined in the
> >> 32bit memory and then add the rest of the 3GiB from the >32bit
> >> area via LPAE? Will the kernel ever try to allocate DMA memory from
> >> anything >32bit?
> >
> > You can solve the case for dma_alloc_coherent() this way, or by
> > setting the mask correctly. It won't help you for dma_map_* though,
> > which still requires someone to add support for swiotlb or using
> > an IOMMU if present.
> 
> We do not have an IOMMU present at the moment. Not sure how
> to go about setting a mask on a pci-probed device.

Ah, right. The mask is a problem because PCI devices assume that
they can do DMA to any 32-bit masked address without calling
dma_set_mask. Your trick to describe the system memory at a different
physical alias would solve this part. Another option might be to
add a quirk in the OHCI/EHCI drivers, if you are able to detect
this special case from the PCI IDs. Whether we can allow the PCI host
controller to just set a mask is an open question. If we decide to go
that route, you should be able to do it using the add_bus() callback
in the host controller driver. However, it would be a departure from
the normal way of doing PCI DMA, and I can't foresee what the implications
would be.

	Arnd