Query on dma_set_mask_and_coherent() Usage

Query on dma_set_mask_and_coherent() Usage

[Kedareswararao Appana]

Hi All,

              On arm64 platform I have booted Linux only with > 32-bit
Address i.e from 0x800000000 (reg =  <0x8 0x00000000 0x0 0x80000000>)

In my driver, I am using dma_set_mask_and_coherent(&pdev->dev,
DMA_BIT_MASK(32)); which should fail if I boot Linux
with the above configuration but the API is returning success.

I have debugged it further and observed that, in the arm64
dma-mapping.c swiotlb is zero resulting dma_supported ops always
returning 1

https://elixir.bootlin.com/linux/v4.17.8/source/arch/arm64/mm/dma-mapping.c#L321

 I will be grateful if someone can send me some pointers for fixing this issue

Regards,
Kedar.

Re: Query on dma_set_mask_and_coherent() Usage

[Catalin Marinas]

On Thu, Aug 23, 2018 at 08:20:36PM +0530, Kedareswararao Appana wrote:
>               On arm64 platform I have booted Linux only with > 32-bit
> Address i.e from 0x800000000 (reg =  <0x8 0x00000000 0x0 0x80000000>)

So you have 2GB of RAM starting at 0x8_0000_0000.

> In my driver, I am using dma_set_mask_and_coherent(&pdev->dev,
> DMA_BIT_MASK(32)); which should fail if I boot Linux
> with the above configuration but the API is returning success.
> 
> I have debugged it further and observed that, in the arm64
> dma-mapping.c swiotlb is zero resulting dma_supported ops always
> returning 1
> 
> https://elixir.bootlin.com/linux/v4.17.8/source/arch/arm64/mm/dma-mapping.c#L321
> 
>  I will be grateful if someone can send me some pointers for fixing this issue

It's an assumption we've made here:

https://elixir.bootlin.com/linux/v4.17.8/source/arch/arm64/mm/init.c#L224

32-bit devices are expected to have a dma_pfn_offset (and the hardware
configured accordingly).

-- 
Catalin