Hi syed, thanks for your response. On Fri, Dec 4, 2015 at 9:16 AM, sk.syed2 <sk.syed2@gmail.com> wrote:

...

I have a device with a piece of RAM. That memory is accessible to CPU. I write a driver for that device and it maps device memory directly to userspace processes. This mapping is cacheable. So far so good.

How did you map this device memory? Using ioremap? What platform is this?

No, I provide mmap for the device file and then I do remap_pfn_range using vma->vm_pgoff as an offset into the device memory block. That's custom R&D ARM-based thing.

...

Now I'd like to flush or invalidate cache on specific parts of that memory. If it was just normal system memory I could use streaming DMA API: dma_map_*/dma_sync_*/dma_unmap_*. But the device memory is not a part of system memory and so there are no struct page for its pages, and thus streaming DMA support functions don't work.

Once you map the io memory, you get VAs for that memory and you should be able to use dma_sync APIs.

I've looked at the following function calls: arm_dma_sync_single_for_cpu -> dma_cache_maint_page -> __dma_page_dev_to_cpu and I see them mapping DMA handle to struct page and using that struct page. I'm sure that will not work, because as I said the device memory is not covered by struct page's.