12:37 <sven> i currently have a snps,dwc3 node on a 64bit system with RAM starting at 0x8_0000_0000. i have dma-ranges in the bus, setup the iommu to bypass and i'd expect dwc3 to be able to grab dma buffers now. it doesn't though because it seems to be stuck with a 32bit dma mask.

12:37 <sven> and, well, there's no ram at the beginning of my address space. i can manually add a dma_set_mask_and_coherent(dev, DMA_BIT_MASK(64)) to dwc3_probe and it works then, but i thought that's what dma-ranges is for

12:37 <sven> I'm still new to the kernel. am I missing something here?

19:54 <headless> OK, found the new list server :-)

20:12 <Emantor> sven: reading drivers/of/device.c you might be able to find out whats happening by sprinkling in some prints and enabling debugging in the file. Looks like the bus is involved as well.

20:45 <sven> Emantor: i found https://www.kernel.org/doc/Documentation/DMA-API-HOWTO.txt now which mentions "By default, the kernel assumes that your device can address 32-bits of DMA

20:45 <sven> addressing. For a 64-bit capable device, this needs to be increased,"

20:45 <sven> so maybe the dwc3 driver is just missing that

21:34 <arnd> sven: the dma-ranges property tells the kernel what the bus supports, but the kernel doesn't know what the driver can do, so it still has to call dma_set_mask_and_coherent()

21:35 <arnd> I don't know why the generic xhci doesn't do that, but I think there have been problems in xhci implementations in the past

21:36 <arnd> xhci-plat.c does call dma_set_mask_and_coherent()

21:37 <sven> xhci actually works fine, it's the dwc3 core that tries to allocate some event buffer (which i'm not even sure it needs in host mode) that isn't

21:38 <arnd> sven: ah, so it probably hasn't bothered anyone in the past -- if it's only a couple of allocations for the coherent API rather than a streaming mapping, then on any other machine it will just end up using ZONE_DMA32.

21:38 <arnd> in theory that's a limited resource, but in practice nobody would have noticed as long as there is some memory at all on that zone

21:39 <arnd> I suppose if you use the iommu, it works, right?

21:39 <sven> yes

21:40 <sven> and thanks, makes sense!

21:40 <arnd> patching dwc3 to add this call should be fairly low risk then, as long as you set the mask in the exact same conditions that xhci uses it

21:43 <sven> ok, i'll take a look at xhci then and prepare a patch