Hello,
On Tuesday, December 27, 2011 6:53 PM James Bottomley wrote:
On Tue, 2011-12-27 at 09:25 +0100, Marek Szyprowski wrote: [...]
Usually these drivers don't touch the buffer data at all, so the mapping in kernel virtual address space is not needed. We can introduce DMA_ATTRIB_NO_KERNEL_MAPPING attribute which lets kernel to skip/ignore creation of kernel virtual mapping. This way we can save previous vmalloc area and simply some mapping operation on a few architectures.
I really think this wants to be a separate function. dma_alloc_coherent is for allocating memory to be shared between the kernel and a driver; we already have dma_map_sg for mapping userspace I/O as an alternative interface. This feels like it's something different again rather than an option to dma_alloc_coherent.
That is just a starting point for the discussion.
I thought about this API a bit and came to conclusion that there is no much difference between a dma_alloc_coherent which creates a mapping in kernel virtual space and the one that does not. It is just a hint from the driver that it will not use that mapping at all. Of course this attribute makes sense only together with adding a dma_mmap_attrs() call, because otherwise drivers won't be able to get access to the buffer data.
This depends. On Virtually indexed systems like PA-RISC, there are two ways of making a DMA range coherent. One is to make the range uncached. This is incredibly slow and not what we do by default, but it can be used to make multiple mappings coherent. The other is to load the virtual address up as a coherence index into the IOMMU. This makes it a full peer in the coherence process, but means we can only designate a single virtual range to be coherent (not multiple mappings unless they happen to be congruent). Perhaps it doesn't matter that much, since I don't see a use for this on PA, but if any other architecture works the same, you'd have to designate a single mapping as the coherent one and essentially promise not to use the other mapping if we followed our normal coherence protocols.
Obviously, the usual range we currently make coherent is the kernel mapping (that's actually the only virtual address we have by the time we're deep in the iommu code), so designating a different virtual address would need some surgery to the guts of the iommu code.
I see, in this case not much can be achieved by dropping the kernel mapping for the allocated buffer. I'm also not sure how to mmap the buffer into userspace meet the cpu requirements? Is it possible to use non-cached mapping in userspace together with coherent mapping in kernel virtual space?
However on some other architectures this attribute allows using HIGH_MEM for the allocated coherent buffer. The other possibility is to allocate it in chunks and map them contiguously into dma address space. With NO_KERNEL_MAPPING attribute we avoid consuming vmalloc range for the newly allocated buffer for which we cannot use the linear mapping (because it is scattered).
Of course this attribute will be implemented by the architectures where it gives some benefits. All other can simply ignore it and return plain coherent buffer with ordinary kernel virtual mapping. The driver will just ignore it.
Best regards