On 29/04/2022 14:02, Laurent Pinchart wrote: > On Fri, Apr 29, 2022 at 01:17:27PM +0200, Hans Verkuil wrote: >> On 29/04/2022 12:20, Laurent Pinchart wrote: >>> On Fri, Apr 29, 2022 at 12:13:46PM +0200, Hans Verkuil wrote: >>>> On 29/04/2022 12:07, Laurent Pinchart wrote: >>>>> On Fri, Apr 29, 2022 at 11:58:48AM +0200, Jacopo Mondi wrote: >>>>>> On Fri, Apr 29, 2022 at 10:43:09AM +0200, Hans Verkuil wrote: >>>>>>> On 29/04/2022 10:28, Eugen.Hristev@xxxxxxxxxxxxx wrote: >>>>>>>> On 4/29/22 11:17 AM, Hans Verkuil wrote: >>>>>>>>> On 10/03/2022 10:51, Eugen Hristev wrote: >>>>>>>>>> As a top MC video driver, the atmel-isc should not propagate the format to the >>>>>>>>>> subdevice, it should rather check at start_streaming() time if the subdev is properly >>>>>>>>>> configured with a compatible format. >>>>>>>>>> Removed the whole format finding logic, and reworked the format verification >>>>>>>>>> at start_streaming time, such that the ISC will return an error if the subdevice >>>>>>>>>> is not properly configured. To achieve this, media_pipeline_start >>>>>>>>>> is called and a link_validate callback is created to check the formats. >>>>>>>>>> With this being done, the module parameter 'sensor_preferred' makes no sense >>>>>>>>>> anymore. The ISC should not decide which format the sensor is using. The >>>>>>>>>> ISC should only cope with the situation and inform userspace if the streaming >>>>>>>>>> is possible in the current configuration. >>>>>>>>>> The redesign of the format propagation has also risen the question of the >>>>>>>>>> enumfmt callback. If enumfmt is called with an mbus_code, the enumfmt handler >>>>>>>>>> should only return the formats that are supported for this mbus_code. >>>>>>>>>> Otherwise, the enumfmt will report all the formats that the ISC could output. >>>>>>>>>> With this rework, the dynamic list of user formats is removed. It makes no >>>>>>>>>> more sense to identify at complete time which formats the sensor could emit, >>>>>>>>>> and add those into a separate dynamic list. >>>>>>>>>> The ISC will start with a simple preconfigured default format, and at >>>>>>>>>> link validate time, decide whether it can use the format that is configured >>>>>>>>>> on the sink or not. >>>>>>>>>> >>>>>>>>>> Signed-off-by: Eugen Hristev <eugen.hristev@xxxxxxxxxxxxx> >>>>>>>>>> Reviewed-by: Jacopo Mondi <jacopo@xxxxxxxxxx> >>>>>>>>>> --- >>>>>>>>>> Changes in v9: >>>>>>>>>> - isc_link_validate now static >>>>>>>>>> >>>>>>>>>> Changes in v7: >>>>>>>>>> - minor typos as suggested by Jacopo >>>>>>>>>> - small changes, reduce some indentation, modified an index, as suggested by >>>>>>>>>> Jacopo >>>>>>>>>> >>>>>>>>>> Changes in v6: >>>>>>>>>> - reworked a bit enum_fmt as suggested by Jacopo >>>>>>>>>> >>>>>>>>>> Changes in v5: >>>>>>>>>> - removed user_formats dynamic list as it is now pointless >>>>>>>>>> - greatly simplified the enum_fmt function >>>>>>>>>> - removed some init code that was useless now >>>>>>>>>> >>>>>>>>>> Changes in v4: >>>>>>>>>> - moved validation code into link_validate and used media_pipeline_start >>>>>>>>>> - merged this patch with the enum_fmt patch which was previously in v3 of >>>>>>>>>> the series >>>>>>>>>> >>>>>>>>>> Changes in v3: >>>>>>>>>> - clamp to maximum resolution once the frame size from the subdev is found >>>>>>>>>> drivers/media/platform/atmel/atmel-isc-base.c | 412 ++++++++---------- >>>>>>>>>> .../media/platform/atmel/atmel-isc-scaler.c | 5 + >>>>>>>>>> drivers/media/platform/atmel/atmel-isc.h | 13 +- >>>>>>>>>> .../media/platform/atmel/atmel-sama5d2-isc.c | 20 + >>>>>>>>>> .../media/platform/atmel/atmel-sama7g5-isc.c | 20 + >>>>>>>>>> 5 files changed, 236 insertions(+), 234 deletions(-) >>>>>>>>>> >>>>>>>>>> diff --git a/drivers/media/platform/atmel/atmel-isc-base.c b/drivers/media/platform/atmel/atmel-isc-base.c >>>>>>>>>> index ee1dda6707a0..fe2c0af58060 100644 >>>>>>>>>> --- a/drivers/media/platform/atmel/atmel-isc-base.c >>>>>>>>>> +++ b/drivers/media/platform/atmel/atmel-isc-base.c >>>>>>>>>> @@ -36,11 +36,6 @@ static unsigned int debug; >>>>>>>>>> module_param(debug, int, 0644); >>>>>>>>>> MODULE_PARM_DESC(debug, "debug level (0-2)"); >>>>>>>>>> >>>>>>>>>> -static unsigned int sensor_preferred = 1; >>>>>>>>>> -module_param(sensor_preferred, uint, 0644); >>>>>>>>>> -MODULE_PARM_DESC(sensor_preferred, >>>>>>>>>> - "Sensor is preferred to output the specified format (1-on 0-off), default 1"); >>>>>>>>>> - >>>>>>>>>> #define ISC_IS_FORMAT_RAW(mbus_code) \ >>>>>>>>>> (((mbus_code) & 0xf000) == 0x3000) >>>>>>>>>> >>>>>>>>>> @@ -337,6 +332,10 @@ static int isc_start_streaming(struct vb2_queue *vq, unsigned int count) >>>>>>>>>> unsigned long flags; >>>>>>>>>> int ret; >>>>>>>>>> >>>>>>>>>> + ret = media_pipeline_start(&isc->video_dev.entity, &isc->mpipe); >>>>>>>>> >>>>>>>>> The pipeline validation is done in start_streaming, but I don't think that >>>>>>>>> is the best place: if STREAMON is called before buffers are queued, then >>>>>>>>> an invalid pipeline isn't discovered until enough buffers are queued to >>>>>>>>> kick off start_streaming. >>>>>>>>> >>>>>>>>> Drivers like vsp1, omap3isp and the samsung drivers all do this in streamon(). >>>>>>>>> >>>>>>>>> I think that is the correct time to do this. >>>>>>>> >>>>>>>> Hello Hans, >>>>>>>> >>>>>>>> Initially (v2, v3) I had this in streamon(). The problem that I faced at >>>>>>>> that time was that streamoff was never called, so I could not call >>>>>>>> media_pipeline_stop(). Then Jacopo told me to move it to start_streaming >>>>>>>> (see change log for v4) , and I did not face any more problems. >>>>>> >>>>>> Yes indeed, seems I suggested to use media_pipeline_handler in a >>>>>> comment on your v3 >>>>>> >>>>>> "at s_stream time your top driver calls media_pipeline_start()" >>>>>> >>>>>> sorry about that, I should have looked around a bit more carefully and >>>>>> notice most drivers do so at vb2 streamon >>>>>> >>>>>> However I don't see media_pipeline_start being called at all in v3 of >>>>>> the patch >>>>>> >>>>>>> It's a mess. Looking at some drivers I see that omap3isp calls media_pipeline_stop >>>>>>> in streamoff (so will have the same problem as you described if VIDIOC_STREAMOFF >>>>>>> isn't called), exynos4-is does the same, but it also checks the streaming state in >>>>>>> the release() fop callback, so that would fix this problem. And vimc does this >>>>>>> in stop_streaming. >>>>>>> >>>>>>> I'm in favor of fixing this in vb2, that framework knows exactly when this needs >>>>>>> to be called. >>>>>> >>>>>> Are you suggesting to have vb2 to call media_pipeline_start() or is it >>>>>> more complex than this ? >>>>> >>>>> I think Hans meant adding a .validate() operation to vb2. >>>>> >>>>> vb2 is already quite complex, I don't think adding more features is a >>>>> good idea. I'd rather have vb2 focus on buffer management only >>>>> (.start_streaming() and .stop_streaming() shouldn't have been in there >>>>> in my opinion), and handle validation in the .streamon() handler. I'd >>>>> expect most drivers that deal with media pipelines to do more work in >>>>> .streamon() anyway. >>>> >>>> I disagree with that :-) >>>> >>>> It's vb2 that keeps track of the streaming state and when what actions >>>> need to be taken. Drivers really shouldn't need to care about the ioctls >>>> themselves, and just implement the relevant vb2 callbacks. Relying on >>>> drivers to handle any of the streaming ioctls is asking for problems, >>>> as this shows: most drivers implement this wrong today. >>>> >>>> The vb2 framework knows when e.g. the pipeline needs to be started or >>>> stopped, and can do this at the best time, without drivers needing to >>>> keep track of when streamon/off/release is called. Keep that logic in >>>> vb2. >>> >>> Pipeline management and buffer management are two different issues. >>> Don't forget about devices that have multiple video nodes, part of the >>> same pipeline (possibly a combination of output and capture nodes, or >>> all of the same type). Forcing drivers to go through vb2 operations to >>> handle the pipeline will be messy, will result in more bloat in vb2, and >>> make the result more bug-prone and harder to maintain. >>> >>> If pipeline management is too complex, let's simplify it, new helpers >>> can make sense, but not through vb2. >>> >> >> But it is vb2 that knows when streaming starts and stops. > > That's right, but pipeline start (which includes validation and resource > reservation) needs to be performed synchronously with VIDIOC_STREAMON. > The streaming state managed by vb2 is not relevant, > media_pipeline_start() must not be delayed the same way > .start_streaming() is. It will be the first thing that vb2_streamon calls. This has nothing to do with start_streaming: that's called when sufficient number of buffers are queued up to be able to start the DMA. This proposed prepare_streaming op will be called when VIDIOC_STREAMON is called. > >> The driver just >> needs to be informed (e.g. prepare_streaming and unprepare_streaming ops). >> >> vb2 deals with buffer management and it keeps track of the streaming state >> and makes the streaming state transitions. That *is* an integral part of >> vb2. What is missing at the moment are callbacks done at streamon time and >> when the streaming stops (streamoff, or close() when is_streaming is true). >> >> If you want to implement stream validation in a driver, then there are a >> lot of things you need to do: >> >> - override streamon, make sure you call vb2_queue_is_busy(), validate the >> pipeline, then call vb2_streamon, if that fails, remember to stop the >> pipeline. >> >> - override streamoff, make sure you call vb2_queue_is_busy(), stop the >> pipeline and call vb2_streamoff. >> >> - in the release() function when the fh is closed, you have to check >> vb2_is_streaming(), check that you are the owner of the queue, and if true, >> stop the pipeline. > > I'm not opposed to helper functions to implement that, they can bundle > vb2 calls and pipeline management. > >> By moving this to vb2 ops all you need to implement are the prepare and >> unprepare ops. >> >> Esp. the release() implementation is tricky. I'm pretty sure that >> drivers/media/platform/samsung/exynos4-is/fimc-lite.c is wrong, since it >> should only call media_pipeline_stop() for the owner of the queue. Instead >> it calls it for the last user of the queue. >> >> I see that fimc_lite_streamoff() is wrong too: you can safely call >> VIDIOC_STREAMOFF twice: the second streamoff just returns 0 without >> doing anything. Instead media_pipeline_stop is called without testing if >> the queue is streaming. >> >> And yes, this is in part because V4L2 has quite some history and certainly >> API choice were made in the past that we wouldn't make today. But vb2 >> shields you from that, and behaves much more like a proper state machine. >> >> I know you prefer to give a lot more control to driver developers, but >> in my experience very few developers can do things like this right. And >> it is really hard as a reviewer to check if all the corner cases are handled >> correctly in a driver. If vb2 is used, then I know things are called at the >> right time, and that makes my life as reviewer so much easier. > > It's not just about giving more control to drivers, it's about > organizing the software layers in a way that keeps them maintainable, > with layered abstractions and not midlayers. > > We are extensively reworking the media pipeline management as part of > the stream series, and there will be more work on top of that that will > make even more fundamental changes. I would like to at least postpone > any work on vb2 until then, to be able to evaluate the impact. I'll make an RFC patch for vb2 so you have a better idea of what it does. Regards, Hans > >> There may still be a few drivers that really need to do this manually, and >> that's OK, but a driver like the atmel-isc doesn't need that at all. >