Re: [PATCH 0/2] Follow-up patches for Cedrus v9

[Hans Verkuil]

On 09/08/2018 09:42 PM, Paul Kocialkowski wrote:
> Hi,
> 
> Le samedi 08 septembre 2018 Ã 13:24 +0200, Hans Verkuil a Ãcrit :
>> On 09/08/2018 12:22 PM, Chen-Yu Tsai wrote:
>>> On Sat, Sep 8, 2018 at 6:06 PM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
>>>>
>>>> On 09/07/2018 06:33 PM, Paul Kocialkowski wrote:
>>>>> This brings the requested modifications on top of version 9 of the
>>>>> Cedrus VPU driver, that implements stateless video decoding using the
>>>>> Request API.
>>>>>
>>>>> Paul Kocialkowski (2):
>>>>>   media: cedrus: Fix error reporting in request validation
>>>>>   media: cedrus: Add TODO file with tasks to complete before unstaging
>>>>>
>>>>>  drivers/staging/media/sunxi/cedrus/TODO     |  7 +++++++
>>>>>  drivers/staging/media/sunxi/cedrus/cedrus.c | 15 ++++++++++++---
>>>>>  2 files changed, 19 insertions(+), 3 deletions(-)
>>>>>  create mode 100644 drivers/staging/media/sunxi/cedrus/TODO
>>>>>
>>>>
>>>> So close...
>>>>
>>>> When compiling under e.g. intel I get errors since it doesn't know about
>>>> the sunxi_sram_claim/release function and the PHYS_PFN_OFFSET define.
>>>>
>>>> Is it possible to add stub functions to linux/soc/sunxi/sunxi_sram.h
>>>> if CONFIG_SUNXI_SRAM is not defined? That would be the best fix for that.
>>>>
>>>> The use of PHYS_PFN_OFFSET is weird: are you sure this is the right
>>>> way? I see that drivers/of/device.c also sets dev->dma_pfn_offset, which
>>>> makes me wonder if this information shouldn't come from the device tree.
>>>>
>>>> You are the only driver that uses this define directly, which makes me
>>>> suspicious.
>>>
>>> On Allwinner platforms, some devices do DMA directly on the memory BUS
>>> with the DRAM controller. In such cases, the DRAM has no offset. In all
>>> other cases where the DMA goes through the common system bus and the DRAM
>>> offset is either 0x40000000 or 0x20000000, depending on the SoC. Since the
>>> former case is not described in the device tree (this is being worked on
>>> by Maxime BTW), the dma_pfn_offset is not the value it should be. AFAIK
>>> only the display and media subsystems (VPU, camera, TS) are wired this
>>> way.
>>>
>>> In drivers/gpu/drm/sun4i/sun4i_backend.c (the display driver) we use
>>> PHYS_OFFSET, which is pretty much the same thing.
>>>
>>
>> OK, in that case just put #ifdef PHYS_PFN_OFFSET around that line together
>> with a comment that this will eventually come from the device tree.
> 
> That seems fine, although I'm less certain about what to do for the
> SRAM situation. Other drivers that use SUNXI_SRAM have a Kconfig select
> on it (that Cedrus lacks). Provided that the SRAM driver builds fine
> for non-sunxi configs, bringing-in that select would remove the need
> for dummy functions and also ensure that the actual implementation is
> always used on sunxi. Otherwise, there'd be a risk of having the dummy
> functions used (if the SRAM driver is not explicitly selected in the
> config), causing a hang when accessing the VPU.

You should certainly select this kernel config.

But the real problem seems to to be drivers/soc/Makefile where it says:

obj-$(CONFIG_ARCH_SUNXI)       += sunxi/

I think this should be:

obj-y       += sunxi/

Now all compiles fine on i686 with the cedrus driver selecting SUNXI_SRAM.

Regards,

	Hans