25.06.2020 12:27, Mikko Perttunen пишет:
> On 6/25/20 1:33 AM, Dmitry Osipenko wrote:
>> 23.06.2020 15:09, Mikko Perttunen пишет:
>>> struct drm_tegra_submit_relocation {
>>> /* [in] Index of GATHER or GATHER_UPTR command in commands. */
>>> __u32 gather_command_index;
>>>
>>> /*
>>> * [in] Mapping handle (obtained through CHANNEL_MAP) of the
>>> memory
>>> * the address of which will be patched in.
>>> */
>>> __u32 mapping_id;
>>>
>>> /*
>>> * [in] Offset in the gather that will be patched.
>>> */
>>> __u64 gather_offset;
>>>
>>> /*
>>> * [in] Offset in target buffer whose paddr/IOVA will be
>>> written
>>> * to the gather.
>>> */
>>> __u64 target_offset;
>>>
>>> /*
>>> * [in] Number of bits the resulting address will be logically
>>> * shifted right before writing to gather.
>>> */
>>> __u32 shift;
>>>
>>> __u32 reserved[1];
>>> };
>>
>> We will also need read/write direction flag here for the
>> DMA-reservations set up, it will be used for the implicit BO fencing by
>> the job's scheduler.
>>
>
> Ideally on newer chips with context isolation, we no longer know what
> DMA-BUFs are being used by the job at submit time - they would just be
> pointers after being MAPped.
The DMA-BUFs themselves shouldn't be needed, but GEMs should.
> Do you know how other GPUs deal with DMA reservations - I expect
> separate MAP and SUBMIT steps would be pretty common?
I can't instantly recall what other drivers do, could be worthwhile to
take a closer look.
> Do you have to
> pass the DMA-BUF to both steps (i.e. do IOMMU mapping during MAP, and
> manage reservations at SUBMIT)?
Yes, this sounds good to me if DMA-BUF part is replaced with a GEM.
Please see my other reply regarding the MAP IOCTL where I'm suggesting
to back mapping IDs with a GEM.
