On Thursday 18 December 2014 17:07:04 Murali Karicheri wrote:

Add of_pci_dma_configure() to allow updating the dma configuration
of the pci device using the configuration from the parent of
the root bridge device.

Signed-off-by: Murali Karicheri <redacted>

Much better!

There is one detail that we should get right from the start here,
which is currently wrong in the platform device code, but I have so
far not dared change it:

+	/*
+	 * Set default dma-mask to 32 bit. Drivers are expected to setup
+	 * the correct supported dma_mask.
+	 */
+	dev->coherent_dma_mask = DMA_BIT_MASK(32);

The 32-bit mask is indeed correct as the default for almost all cases,
and we must not set it larger than DMA_BIT_MASK(32). There is one
corner case though, which happens on some shmobile machines that have
a 31-bit mask on the PCI host, and I think we should use that as the
default.

+	/*
+	 * Set it to coherent_dma_mask by default if the architecture
+	 * code has not set it.
+	 */
+	if (!dev->dma_mask)
+		dev->dma_mask = &dev->coherent_dma_mask;
+
+	ret = of_dma_get_range(parent_np, &dma_addr, &paddr, &size);
+	if (ret < 0) {
+		dma_addr = offset = 0;
+		size = dev->coherent_dma_mask;

Can you check one thing here? I believe the size argument as returned
from of_dma_get_range is inclusive (e.g. 0x100000000), while the coherent
mask by definition is exlusive (e.g. 0xffffffff), so the size needs to
be adapted here. I haven't checked all the code here though, so I may
be wrong.

+	} else {
+		offset = PFN_DOWN(paddr - dma_addr);
+		dev_dbg(dev, "dma_pfn_offset(%#08lx)\n", dev->dma_pfn_offset);
+	}
+	dev->dma_pfn_offset = offset;
+
+	coherent = of_dma_is_coherent(parent_np);
+	dev_dbg(dev, "device is%sdma coherent\n",
+		coherent ? " " : " not ");
+
+	arch_setup_dma_ops(dev, dma_addr, size, NULL, coherent);

Basically, I would use limit the size argument we pass into
arch_setup_dma_ops to the minimum of 'size' and 'dma_mask' here,
after converting into the same format. We should make sure we do the
same thing for platform_device as well, so it might be better to do
it inside of arch_setup_dma_ops instead.

	Arnd