Re: [OPW kernel] dma_set_coherent_mask

[Don Dutile <ddutile@redhat.com>]

On 05/16/2013 01:19 PM, Sarah Sharp wrote:
> On Thu, May 16, 2013 at 10:11:00AM -0400, Alan Stern wrote:
>> On Wed, 15 May 2013, Sarah Sharp wrote:
>>
>>>> If you're using 64-bit DMA then you almost certainly do want to call
>>>> dma_set_coherent_mask().  On the plus side, it is guaranteed that if
>>>> dma_set_mask() succeeds with a particular mask value then
>>>> dma_set_coherent_mask() for the same mask value will also succeed.
>>>
>>> So we need to call both dma_set_coherent_mask() and dma_set_mask()?  Or
>>> just dma_set_coherent_mask()?
>>
>> It depends on what kind of DMA transfers you're going to do.  For
>> streaming transfers (the ones that use dma_map_single() or
>> dma_map_sg(), for example), you need to call dma_set_mask().  For
>> coherent transfers (the ones that use dma_alloc_coherent() or
>> dma_pool_create()), you need to call dma_set_coherent_mask().
>>
>> If you want to do both kinds of transfers then you need to call both
>> routines.
>
> I think we need the host to be able to do DMA to URB buffers that are
> mapped with dma_map_single() or dma_map_sg(), since that's what
> usb_hcd_submit_urb() uses.  So the driver needs to call dma_set_mask(),
> which it does.
>
> The xHCI endpoint rings are allocated from DMA pools, so we need to
> call dma_set_coherent_mask() as well.  Missing that call explains why
> I've never seen 64-bit endpoint rings, but I have seen 64-bit URB buffer
> pointers.
>
Correct.  Otherwise, that DMA is going through the (more limited 64MB) bounce buffer
which is allocated out of low memory since... a device can't do 64-bit addressing;


> Ksenia, do you want to add code to add the call to
> dma_set_coherent_mask() in those two places in xhci_gen_setup()?  As I
> mentioned, let's add this function call first, and then have a separate
> commit refactor the copy-paste code into a new function.
>
> Again, thanks for catching this and asking questions!
>
> Sarah Sharp
> --
> To unsubscribe from this list: send the line "unsubscribe linux-pci" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html