Hello,
On Monday, August 20, 2012 10:01 PM Arnd Bergmann wrote:
On Monday 20 August 2012, Marek Szyprowski wrote:
Contiguous Memory Allocator requires only paging and MMU enabled not particular CPU architectures, so there is no need for strict dependency on CPU type. This enables to use CMA on some older ARM v5 systems which also might need large contiguous blocks for the multimedia processing hw modules.
Reported-by: Prabhakar Lad prabhakar.csengg@gmail.com Signed-off-by: Marek Szyprowski m.szyprowski@samsung.com
The patch looks simple, but I want a better explanation for it. When we went through all possible cases, we decided that:
- ARMv6+ need CMA to avoid the double mapping problem.
Right. CMA can be used to avoid double mapping issues, but we also should keep in mind that right now no one observed any issues with the real hw, so this discussion is still a bit hypothetical.
- ARMv4/v5 cannot generally use CMA because it doesn't work together with DMABOUNCE. I don't remember if it was the only problem, but I definitely remember this was intentional.
DMABOUNCE is used only by a few machines. I also tested and see no reason why it might cause problems with such machines. I've used DMABOUNCE code to test atomic allocations many times and found no problems. The only issue I might suspect is a very limited DMA zone. CMA needs 4Mib alignment of the contiguous area base and size to meet requirements of the memory management core (mainly for migration and page isolation purposes), so it might not fit into some small DMA zones.
- We want a common kernel for all ARMv6+ eventually, and a separate kernel for all ARMv4/v5 ones.
I see no problem here, if required why may have even a separate dma_map_ops for CMA and use it only for selected devices.
If the reasoning has changed, please try to explain the full situation. On a related topic, what happened to the idea that ARMv6+ is broken without CMA? I noticed that it's optional now.
Right. I removed unconditional dependency on CMA on Russell's request. CMA is still experimental and there are still some known issues with it, which we are investigating. It doesn't make sense to make the whole architecture depending on the experimental stuff. The old method of allocating and managing coherent buffers was stable and worked well enough for the drivers and platforms already present in the mainline kernel, so if one want to use stable stuff I see no reason to disable it.
Best regards