Hello, On Friday, December 23, 2011 10:22 AM Ming Lei wrote: > On Thu, Dec 22, 2011 at 5:28 PM, Marek Szyprowski > <m.szyprowski@xxxxxxxxxxx> wrote: > >> DMA contig memory resource is very limited and precious, also > >> accessing to it from CPU is very slow on some platform. > >> > >> For some cases(such as the comming face detection driver), DMA Streaming > >> buffer is enough, so introduce VIDEOBUF2_PAGE to allocate continuous > >> physical memory but letting video device driver to handle DMA buffer mapping > >> and unmapping things. > >> > >> Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxxxxx> > > > > Could you elaborate a bit why do you think that DMA contig memory resource > > is so limited? If dma_alloc_coherent fails because of the memory fragmentation, > > the alloc_pages() call with order > 0 will also fail. > > For example, on ARM, there is very limited kernel virtual address space reserved > for DMA coherent buffer mapping, the default size is about 2M if I > don't remember mistakenly. It can be easily increased for particular boards, there is no problem with this. > > I understand that there might be some speed issues with coherent (uncached) > > userspace mappings, but I would solve it in completely different way. The interface > > Also there is poor performance inside kernel space, see [1] Your driver doesn't access video data inside kernel space, so this is also not an issue. > > for both coherent/uncached and non-coherent/cached contig allocator should be the > > same, so exchanging them is easy and will not require changes in the driver. > > I'm planning to introduce some design changes in memory allocator api and introduce > > prepare and finish callbacks in allocator ops. I hope to post the rfc after > > Christmas. For your face detection driver using standard dma-contig allocator > > shouldn't be a big issue. > > > > Your current implementation also abuses the design and api of videobuf2 memory > > allocators. If the allocator needs to return a custom structure to the driver > > I think returning vaddr is enough. > > > you should use cookie method. vaddr is intended to provide only a pointer to > > kernel virtual mapping, but you pass a struct page * there. > > No, __get_free_pages returns virtual address instead of 'struct page *'. Then you MUST use cookie for it. vaddr method should return kernel virtual address to the buffer video data. Some parts of videobuf2 relies on this - it is used by file io emulator (read(), write() calls) and mmap equivalent for non-mmu systems. Manual casting in the driver is also a bad idea, that's why there are helper functions defined for both dma_contig and dma_sg allocators: vb2_dma_contig_plane_dma_addr() and vb2_dma_sg_plane_desc(). Best regards -- Marek Szyprowski Samsung Poland R&D Center -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
