DMA buffer synchronisation

[Phil Nitschke]

Hi,

I'm transferring data using DMA from a custom PCI device memory into
RAM.  Can someone advise me on the correct buffer synchronisation
strategy for the following 3 scenarios:

Assuming I have three buffers, buf_A, and buf_B, both created using
dma_alloc_noncoherent(NULL, 32 * PAGE_SIZE, ...), and buf_C (much
larger), remapped into the processor's virtual address space using
ioremap(), and 'n' is the number of bytes to copy using DMA channel
'dmanr', ...

1/. DMA copy from PCI device memory into buf_B:

   /* ...prepare DMA engine, then: */
   WARN_ON(buf_B != L1_CACHE_ALIGN(buf_B));
   dma_sync_single_range_for_device(NULL, buf_B, 0, n, DMA_FROM_DEVICE);

   mv64x6x_enable_dma(dmanr);
   wait_for_completion_interruptible_timeout(...)

   dma_sync_single_range_for_cpu(NULL, buf_B, 0, n, DMA_FROM_DEVICE);

2/. DMA copy from buf_A to buf_B (same as above, except now sync the src
buffer prior to DMA):

   /* ...prepare DMA engine, check buffer alignment (x2), then: */
   dma_sync_single_range_for_device(NULL, buf_A,  0, n, DMA_TO_DEVICE);
   dma_sync_single_range_for_device(NULL, buf_B, 0, n, DMA_FROM_DEVICE);

   mv64x6x_enable_dma(dmanr);
   wait_for_completion_interruptible_timeout(...)

   dma_sync_single_range_for_cpu(NULL, buf_B, 0, n, DMA_FROM_DEVICE);

3/. DMA from PCI device memory into ioremap()'ed buffer, buf_C:

   /* ...prepare DMA engine, check buffer alignment, then: */
   /* No buffer sync-ing is required? */

   mv64x6x_enable_dma(dmanr);
   wait_for_completion_interruptible_timeout(...)

   /* No buffer sync'ing is required? */

Will this approach work (particularly case 3/.)?  I tried dma_sync-ing
the region, but got an 'oops' (kernel access of bad area).

I'm using the 2.6.16 kernel (compiled with CONFIG_NOT_COHERENT_CACHE
defined) on an Artesyn PmPPC7448 processor.  The processor utilises its
Marvell MV64460 as the DMA controller for the transaction.  

Thanks,

-- 
Phil