Re: [BUG] dma-ranges, reserved memory regions, dma_alloc_coherent: possible bug?

[Vladimir Murzin <vladimir.murzin@arm.com>]

On 10/14/19 2:54 PM, Robin Murphy wrote:
> On 13/10/2019 15:28, Daniele Alessandrelli wrote:
>> Hi,
>>
>> It looks like dma_alloc_coherent() is setting the dma_handle output
>> parameter to the memory physical address and not the device bus
>> address when the device is using reserved memory regions for DMA
>> allocation. This is despite using 'dma_ranges' in the device tree to
>> describe the DMA memory mapping. Is this expected behavior or a bug?
> 
> That does sound like a bug :(
> 
>> Here is a reduced version of the device tree I'm using:
>> \ {
>>          reserved-memory {
>>                  #address-cells = <2>;
>>                  #size-cells = <2>;
>>                  ranges;
>>                  mydev_rsvd: rsvd_mem@494800000 {
>>                          compatible = "shared-dma-pool";
>>                          reg = <0x4 0x94800000 0x0 0x200000>;
>>                          no-map;
>>                  };
>>          };
>>          soc {
>>                  compatible = "simple-bus";
>>                  #address-cells = <2>;
>>                  #size-cells = <2>;
>>                  ranges;
>>                  dma_ranges;
>>
>>                  mybus {
>>                          ranges = <>;
>>                          dma-ranges = <>;
>>                          compatible = "simple-bus";
>>                          #address-cells = <2>;
>>                          #size-cells = <2>;
>>                          ranges =     <0x0 0x0  0x0 0x0  0x0 0x80000000>;
>>                          dma-ranges = <0x0 0x80000000  0x4 0x80000000
>> 0x0 0x80000000>;
>>
>>                          mydevice {
>>                                  compatible = "my-compatible-string";
>>                                  memory-region = <&mydev_rsvd>;
>>                          }
>>                  }
>>          }
>> };
>>
>> It looks like this issue was previously fixed by commit c41f9ea998f3
>> ("drivers: dma-coherent: Account dma_pfn_offset when used with device
>> tree") which introduced a new function ('dma_get_device_base()') to
>> return the reserved memory address as seen by the device. However,
>> such a function, even if still there, is not used anymore in latest
>> code (as of v5.4-rc2). Was that done for a specific reason? Or is it
>> just a mistake?
> 
> Hmm, it looks like 43fc509c3efb ("dma-coherent: introduce interface for default DMA pool") removed the caller of dma_get_device_base() in the alloc path shortly after it was introduced, which certainly appears as if it may have been unintentional - Vladimir?

I do not remember it was intentional. Looking at history, default DMA pool was a response
to another report. However, now I'm wondering why it was not caught by STM32 - most of that
work was required to support "dma-ranges" with NOMMU+caches (Cortex-M7).

Alex (or anybody else from ST), maybe you have some input?

Cheers
Vladimir

> 
> Robin.

_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu