Re: [RFC] [PATCH 1/6] SoC Camera: add driver for OMAP1 camera interface

[Marin Mitov <mitov@issp.bas.bg>]

On Friday, August 13, 2010 11:52:41 am Guennadi Liakhovetski wrote:
> On Fri, 13 Aug 2010, Janusz Krzysztofik wrote:
> 
> > Thursday 12 August 2010 23:38:17 Guennadi Liakhovetski napisał(a):
> > > On Sun, 1 Aug 2010, Janusz Krzysztofik wrote:
> > > > Friday 30 July 2010 20:49:05 Janusz Krzysztofik napisał(a):
> > > >
> > > > I think the right way would be if implemented at the videobuf-core level.
> > > > Then, drivers should be able to initialize both paths, providing queue
> > > > callbacks for both sets of methods, contig and sg, for videobuf sole use.
> > >
> > > Ok, here're my thoughts about this:
> > >
> > > 1. We've discussed this dynamic switching a bit on IRC today. The first
> > > reaction was - you probably should concentrate on getting the contiguous
> > > version to work reliably. I.e., to reserve the memory in the board init
> > > code similar, how other contig users currently do it. 
> > 
> > I already tried before to find out how I could allocate memory at init without 
> > reinventing a new videobuf-dma-contig implementation. Since in the 
> > Documentation/video4linux/videobuf I've read that videobuf does not currently 
> > play well with drivers that play tricks by allocating DMA space at system 
> > boot time, I've implemented the alternate sg path.
> > 
> > If it's not quite true what the documentation says and you can give me a hint 
> > how this could be done, I might try again.
> 
> For an example look at 
> arch/arm/mach-mx3/mach-pcm037.c::pcm037_camera_alloc_dma().

For preallocating dma-coherent memory for device personal use during device probe() time
(when the memory is less fragmented compared to open() time)
see also dt3155_alloc_coherent/dt3155_free_coherent in drivers/staging/dt3155v4l/dt3155vfl.c
(for x86 arch, I do not know if it works for arm arch)

> 
> > > But given problems 
> > > with this aproach in the current ARM tree [1], this might be a bit
> > > difficult. Still, those problems have to be and will be fixed somehow
> > > eventually, so, you might prefer to still just go that route.
> > 
> > My board uses two drivers that allocate dma memory at boot time: 
> > drivers/video/omap/lcdc.c and sounc/soc/omap/omap-pcm.c. Both use 
> > alloc_dma_writecombine() for this and work without problems.
> 
> dma_alloc_writecombine() also allocates contiguous RAM, right? And it 
> doesn't use device "local" memory. So, it's chances to fail are the same 
> as those of dma_alloc_coherent() in the absence of device own memory. I 
> guess, the sound driver doesn't need much RAM, but if you build your LCDC 
> driver as a module and load it later after startup, it might get problems 
> allocating RAM for the framebuffer.
> 
> > > 2. If you do want to do the switching - we also discussed, how forthcoming
> > > changes to the videobuf subsystem will affest this work. I do not think it
> > > would be possible to implement this switching in the videobuf core.
> > 
> > OK, I should have probably said that it looked not possible for me to do it 
> > without any additional support implemented at videobuf-core (or soc_camera) 
> > level.
> > 
> > > Remember, with the videobuf API you first call the respective
> > > implementation init method, which doesn't fail. Then, in REQBUFS ioctl you
> > > call videobuf_reqbufs(), which might already fail but normally doesn't.
> > > The biggest problem is the mmap call with the contig videobuf
> > > implementation. This one is likely to fail. So, you would have to catch
> > > the failing mmap, call videobuf_mmap_free(), then init the SG videobuf,
> > > request buffers and mmap them... 
> > 
> > That's what I've already discovered, but failed to identify a place in my code 
> > where I could intercept this failing mmap without replacing parts of the 
> > videobuf code.
> 
> Right, ATM soc-camera just calls videobuf_mmap_mapper() directly in its 
> mmap method. I could add a callback to struct soc_camera_host_ops like
> 
> 	int (*mmap)(struct soc_camera_device *, struct vm_area_struct *)
> 
> and modify soc_camera_mmap() to check, whether the host driver has 
> implemented it. If so - call it, otherwise call videobuf_mmap_mapper() 
> directly just like now. So, other drivers would not have to be modified. 
> And you could implement that .mmap() method, call videobuf_mmap_mapper() 
> yourself, and if it fails for contig, fall back to SG.
> 
> > > With my 2 patches from today, there is 
> > > only one process (file descriptor, to be precise), that manages the
> > > videobuf queue. So, this all can only be implemented in your driver.
> > 
> > The only way I'm yet able to invent is replacing the 
> > videobuf_queue->int_ops->mmap_mapper() callback with my own wrapper that 
> > would intercept a failing videobuf-dma-contig version of mmap_mapper(). This 
> > could be done in my soc_camera_host->ops->init_videobuf() after the 
> > videobuf-dma-contig.c version of the videobuf_queue->int_ops->mmap_mapper() 
> > is installed with the videobuf_queue_dma_contig_init().
> > 
> > Is this method close to what you have on mind?
> 
> See, if the above idea would suit your needs.
> 
> Thanks
> Guennadi
> ---
> Guennadi Liakhovetski, Ph.D.
> Freelance Open-Source Software Developer
> http://www.open-technology.de/
> --
> To unsubscribe from this list: send the line "unsubscribe linux-media" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>