Re: [PATCH 11/15] accel/qda: Add PRIME DMA-BUF import support

["Christian König" <christian.koenig@amd.com>]

On 6/3/26 08:11, Ekansh Gupta wrote:
> On 19-05-2026 12:25, Christian König wrote:
>> On 5/19/26 08:16, Ekansh Gupta via B4 Relay wrote:
>>> From: Ekansh Gupta <ekansh.gupta@oss.qualcomm.com>
...
>>> +static int qda_memory_manager_map_imported(struct qda_memory_manager *mem_mgr,
>>> +                                          struct qda_gem_obj *gem_obj,
>>> +                                          struct qda_iommu_device *iommu_dev)
>>> +{
>>> +       struct scatterlist *sg;
>>> +       dma_addr_t dma_addr;
>>> +
>>> +       if (!gem_obj->is_imported || !gem_obj->sgt || !iommu_dev) {
>>> +               drm_err(gem_obj->base.dev, "Invalid parameters for imported buffer mapping\n");
>>> +               return -EINVAL;
>>> +       }
>>> +
>>> +       sg = gem_obj->sgt->sgl;
>>> +       if (!sg) {
>>> +               drm_err(gem_obj->base.dev, "Invalid scatter-gather list for imported buffer\n");
>>> +               return -EINVAL;
>>> +       }
>>> +
>>> +       gem_obj->iommu_dev = iommu_dev;
>>> +
>>> +       /*
>>> +        * After dma_buf_map_attachment_unlocked(), sg_dma_address() returns the
>>> +        * IOMMU virtual address, not the physical address. The IOMMU maps the
>>> +        * entire buffer as a contiguous range in the IOMMU address space even if
>>> +        * the underlying physical memory is non-contiguous. Therefore the first
>>> +        * sg entry's DMA address is the start of the complete contiguous
>>> +        * IOMMU-mapped range and is sufficient to describe the buffer to the DSP.
>>> +        */
>>> +       dma_addr = sg_dma_address(sg);
>>> +       dma_addr += ((u64)iommu_dev->sid << 32);
>>> +       gem_obj->dma_addr = dma_addr;
>>
>> That handling here is completely broken since it assumes that the exporter maps the buffer as contigious range.
>>
>> But that's in no way guaranteed.
> I'll collect more details and will try to implement this in the right
> way, maybe by iterating the full sg_table.>

You could also document explicitly that you can only import contiguous buffers (e.g. DMA-buf heap CMA etc....) and then cleanly reject non contiguous buffers here.

We have quite a number of drivers/HW with that limitation, so only accepting contiguous buffers is perfectly ok.

You just can't silently assume that IOMMU would always map the entire buffer as one contiguous range, cause that is certainly not true.

Regards,
Christian.


>> Regards,
>> Christian.