> > - V4L2_MEMORY_USERPTR allows applications to pass in pointers to > > random userspace memory (userptr+length). > > If you do this, please keep in mind that some scatter-gather DMA engines > (especially the philips saa7146, don't know about the saa7134) can only > deal with page aligned boundaries. The saa7134 doesn't need a page-aligned buffer start, and I doubt the saa7146 does. The saa7134 has a iommu which wants a page list in physical memory. That forms a virtual, 4MB sized I/O space. The saa7146 does the same as far I know. For DMA transfer you have some control registers (start offset, line padding, ...) for each field. The start offset is needed to fill a buffer with an interlaced image and to place the overlay at a random location on the screen, thus I expect the saa7146 has this too. The start offset can also be used to let the capture buffer start at non-page-aligned addresses. > Alternatively, the user application may provide an already page aligned > memory, but I don't think that this is a good idea (and if it's possible > for a user application to request this). If page-alignment is really needed I'd prefeare this way. If you allocate memory by mapping anonymous memory it will be page aligned. shared memory is also page-aligned. And applications can figure the page size with the "getpagesize()" call, so they have a chance to do the page-alignment thing themself even for memory allocated via malloc(). Gerd -- You can't please everybody. And usually if you _try_ to please everybody, the end result is one big mess. -- Linus Torvalds, 2002-04-20
