Re: [PATCH 03/11] media: exynos4-is: Fix nullptr when no CSIS device present

[Jonathan Bakker]

Hi Sylwester and Tomasz,

On 2020-07-08 8:11 a.m., Sylwester Nawrocki wrote:
> Hi,
> 
> On 07.07.2020 19:55, Tomasz Figa wrote:
>> On Sat, Apr 25, 2020 at 07:26:42PM -0700, Jonathan Bakker wrote:
>>> Not all devices use the CSIS device, some may use the FIMC directly in
>>> which case the CSIS device isn't registered.  This leads to a nullptr
>>> exception when starting the stream as the CSIS device is always
>>> referenced.  Instead, if getting the CSIS device fails, try getting the
>>> FIMC directly to check if we are using the subdev API
>>>
>>> Signed-off-by: Jonathan Bakker <xc-racer2@xxxxxxx>
>>> ---
>>>  drivers/media/platform/exynos4-is/media-dev.c | 17 ++++++++++++++++-
>>>  1 file changed, 16 insertions(+), 1 deletion(-)
> 
>> Thank you for the patch. Please see my comments inline.
> 

Thank you both for taking the time to review the patches.

>>> diff --git a/drivers/media/platform/exynos4-is/media-dev.c b/drivers/media/platform/exynos4-is/media-dev.c
>>> index 9aaf3b8060d5..5c32abc7251b 100644
>>> --- a/drivers/media/platform/exynos4-is/media-dev.c
>>> +++ b/drivers/media/platform/exynos4-is/media-dev.c
>>> @@ -289,11 +289,26 @@ static int __fimc_pipeline_s_stream(struct exynos_media_pipeline *ep, bool on)
>>>  		{ IDX_CSIS, IDX_FLITE, IDX_FIMC, IDX_SENSOR, IDX_IS_ISP },
>>>  	};
>>>  	struct fimc_pipeline *p = to_fimc_pipeline(ep);
>>> -	struct fimc_md *fmd = entity_to_fimc_mdev(&p->subdevs[IDX_CSIS]->entity);
>>>  	enum fimc_subdev_index sd_id;
>>>  	int i, ret = 0;
>>>  
>>>  	if (p->subdevs[IDX_SENSOR] == NULL) {
>>> +		struct fimc_md *fmd;
>>> +		struct v4l2_subdev *sd = p->subdevs[IDX_CSIS];
>>> +
>>> +		if (!sd)
>>> +			sd = p->subdevs[IDX_FIMC];
>>> +
>>> +		if (!sd) {
>>> +			/*
>>> +			 * If neither CSIS nor FIMC was set up,
>>> +			 * it's impossible to have any sensors
>>> +			 */
>>> +			return -ENODEV;
>>> +		}
>>> +
>>> +		fmd = entity_to_fimc_mdev(&sd->entity);
>>> +
>>
>> Are you sure this is the correct thing to do here? In general, the media
>> controller should be instantiated only if there are sensors in the system.
> 
> The code being changed is only about getting a pointer to the driver private 
> data structure 'struct fimc_md', for that we need to get hold of a media 
> entity that represents a subdev that is actually available in the pipeline.
> In original code it is overlooked that there might camera pipelines without
> the CSIS subdev. 
> 
>> What do you mean by using "the FIMC directly"? Do you mean using it only as
>> an m2m image processor or with a sensor, but without the CSIS, which would
>> be the case for parallel I/F sensors?
> 
> I think it is about a use case where the sensor is connected directly to the 
> FIMC capture interface (parallel), without the MIPI-CSI2 receiverin between.
> 

Yep, that's exactly what I mean.  The device that I'm working with (first gen
Galaxy S) has two devices (S5KA3DFX and CE147) connected via parallel input
to Camera Port A (yes, both of them to the same port).

Is it more correct to say "using the parallel port instead of the CSIS device"?
I'm happy to reword the commit message if that is the case.  

>> Could you point me to the place where CSIS is always dereferenced? A quick
>> look through the code only revealed that everywhere it seems to be guarded
>> by a NULL check.
> 
> It's in this patch, above, the very first line that the patch removes.
> 

Exactly :)

>> Another thought from looking at the implementation of
>> __fimc_pipeline_s_stream() is that it probably shouldn't call s_stream on
>> all the subdevices included in seq[], but only on those that are actually
>> included as a part of the pipeline. It would be quite a waste of power to
>> enable unnecessary hardware.
> 
> As we talked on IRC, only subdev in current active media pipeline will be
> powered on/off. The p->subdevs[] array is sparsely populated and there is 
> a test for NULL in __subdev_set_power(). Perhaps the test should be moved
> to the caller instead (fimc_pipeline_s_power()), so we don't ignore any
> ENXIO errors from the s_power v4l2_subdev_call. But is a material for 
> a separate patch.
> 

Ok, I'll leave it be for now.

Thanks,
Jonathan