Re: [PATCH RFC 12/18] accel/qda: Add PRIME dma-buf import support

[Ekansh Gupta <ekansh.gupta@oss.qualcomm.com>]

On 3/9/2026 12:29 PM, Ekansh Gupta wrote:
>
> On 2/24/2026 2:42 PM, Christian König wrote:
>> On 2/23/26 20:09, Ekansh Gupta wrote:
>>> [Sie erhalten nicht häufig E-Mails von ekansh.gupta@oss.qualcomm.com. Weitere Informationen, warum dies wichtig ist, finden Sie unter https://aka.ms/LearnAboutSenderIdentification ]
>>>
>>> Add PRIME dma-buf import support for QDA GEM buffer objects and integrate
>>> it with the existing per-process memory manager and IOMMU device model.
>>>
>>> The implementation extends qda_gem_obj to represent imported dma-bufs,
>>> including dma_buf references, attachment state, scatter-gather tables
>>> and an imported DMA address used for DSP-facing book-keeping. The
>>> qda_gem_prime_import() path handles reimports of buffers originally
>>> exported by QDA as well as imports of external dma-bufs, attaching them
>>> to the assigned IOMMU device
>> That is usually an absolutely clear NO-GO for DMA-bufs. Where exactly in the code is that?
> dma_buf_attach* to comute-cb iommu devices are critical for DSPs to access the buffer.
> This is needed if the buffer is exported by anyone other than QDA(say system heap). If this is not
> the correct way, what should be the right way here? On the current fastrpc driver also,
> the DMABUF is getting attached with iommu device[1] due to the same requirement.
>
> [1] https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/drivers/misc/fastrpc.c#n779

Hi Christian,

Do you have any suggestions for the shared requirements?

I'm reworking on the next version and currently I don't see any other way
to handle dma_buf_attach* cases.

//Ekansh

>>> and mapping them through the memory manager
>>> for DSP access. The GEM free path is updated to unmap and detach
>>> imported buffers while preserving the existing behaviour for locally
>>> allocated memory.
>>>
>>> The PRIME fd-to-handle path is implemented in qda_prime_fd_to_handle(),
>>> which records the calling drm_file in a driver-private import context
>>> before invoking the core DRM helpers. The GEM import callback retrieves
>>> this context to ensure that an IOMMU device is assigned to the process
>>> and that imported buffers follow the same per-process IOMMU selection
>>> rules as natively allocated GEM objects.
>>>
>>> This patch prepares the driver for interoperable buffer sharing between
>>> QDA and other dma-buf capable subsystems while keeping IOMMU mapping and
>>> lifetime handling consistent with the existing GEM allocation flow.
>>>
>>> Signed-off-by: Ekansh Gupta <ekansh.gupta@oss.qualcomm.com>
>> ...
>>
>>> @@ -15,23 +16,29 @@ static int validate_gem_obj_for_mmap(struct qda_gem_obj *qda_gem_obj)
>>>                 qda_err(NULL, "Invalid GEM object size\n");
>>>                 return -EINVAL;
>>>         }
>>> -       if (!qda_gem_obj->iommu_dev || !qda_gem_obj->iommu_dev->dev) {
>>> -               qda_err(NULL, "Allocated buffer missing IOMMU device\n");
>>> -               return -EINVAL;
>>> -       }
>>> -       if (!qda_gem_obj->iommu_dev->dev) {
>>> -               qda_err(NULL, "Allocated buffer missing IOMMU device\n");
>>> -               return -EINVAL;
>>> -       }
>>> -       if (!qda_gem_obj->virt) {
>>> -               qda_err(NULL, "Allocated buffer missing virtual address\n");
>>> -               return -EINVAL;
>>> -       }
>>> -       if (qda_gem_obj->dma_addr == 0) {
>>> -               qda_err(NULL, "Allocated buffer missing DMA address\n");
>>> -               return -EINVAL;
>>> +       if (qda_gem_obj->is_imported) {
>> Absolutely clear NAK to that. Imported buffers *can't* be mmaped through the importer!
>>
>> Userspace needs to mmap() them through the exporter.
>>
>> If you absolutely have to map them through the importer for uAPI backward compatibility then there is dma_buf_mmap() for that, but this is clearly not the case here.
>>
>> ...
> Okay, the requirement is slightly different here. Any buffer which is not allocated using the
> QDA GEM interface needs to be attached to the iommu device for that particular process to
> enable DSP for the access. I should not call it `mmap` instead it should be called importing the
> buffer to a particular iommu context bank. With this definition, is it fine to keep it this way? Or
> should the dma_buf_attach* calls be moved to some other place?
>>> +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;
>>> +       int ret = 0;
>>> +
>>> +       if (!gem_obj->is_imported || !gem_obj->sgt || !iommu_dev) {
>>> +               qda_err(NULL, "Invalid parameters for imported buffer mapping\n");
>>> +               return -EINVAL;
>>> +       }
>>> +
>>> +       gem_obj->iommu_dev = iommu_dev;
>>> +
>>> +       sg = gem_obj->sgt->sgl;
>>> +       if (sg) {
>>> +               dma_addr = sg_dma_address(sg);
>>> +               dma_addr += ((u64)iommu_dev->sid << 32);
>>> +
>>> +               gem_obj->imported_dma_addr = dma_addr;
>> Well that looks like you are only using the first DMA address from the imported sgt. What about the others?
> I might have a proper appach for this now, will update in the next spin.
>> Regards,
>> Christian.