On 2014?09?22? 23:54, Arnd Bergmann wrote:
> On Monday 22 September 2014 17:15:06 Boris BREZILLON wrote:
>>>> +
>>>> + /* TODO(djkurtz): fetch the mapping start/size from somewhere */
>>>> + mapping = arm_iommu_create_mapping(&platform_bus_type, 0x10000000,
>>>> + SZ_1G);
>>>> + if (IS_ERR(mapping)) {
>>>> + ret = PTR_ERR(mapping);
>>>> + goto err_config_cleanup;
>>>> + }
>>>> +
>>>> + dma_set_coherent_mask(dev, DMA_BIT_MASK(32));
>>> This is the default coherent mask. If you call this function, you
>>> should normally check the return value, or call dma_set_mask first,
>>> which you apparently don't do here, and in another place in this
>>> patch.
>> By "This is the default mask" do you mean it shouldn't be called at
>> all ? Cause I ran into some trouble when not calling this in my
>> atmel-hlcdc driver.
> We used to get this wrong in the of_platform code, but it should
> work now.
>
>> Actually, in my case the platform device is created by the MFD core
>> which seems to let the coherent_dma_mask uninitialized.
> That may well be different, but it seems like a good idea to allow
> the MFD core to set this up as well.
>
> In general, we expect that devices that are capable of doing DMA
> start with a 32-bit mask for both dma_mask and dma_coherent_mask,
> and a driver that requires a smaller masks or wants a larger mask
> has to call the appropriate interface to set these, and check the
> return value.
so I think we can use dma_set_mask_and_coherent(...) to set dma_mask and
dma_coherent_mark
at the same time, and check the return value.
> Arnd
>
>
>
