DMA transfer over keystone PCIe

Hi Vivek,

Thanks for materials indicated. My intention is NOT to use outbound transfers, as in case of buggy DSP software, outbound writes to wrong areas may affect software running on host machine. Bus mastering is DISABLED in my case. I am rather interested in DMA transfer made by host DMA engine, between pcie BAR window and memory. And today I found the solution. Before I was using virt_to_page for both memory regions. The function works well for kernel memory but not virtual memory or iomapped PCIe BAR. For PCIe BAR I changed it to vmalloc_to_page. Now it works as expected. Summing up, that was my bug.

So code copying data from BAR to memory  looks as follows:

/* At the beginning, once ... */

dma_cap_mask_t mask;

dma_cap_zero(mask);

dma_cap_set(DMA_MEMCPY, mask);

dma = dma_request_channel(mask, NULL, NULL);

/* vmalloc_to_page is more expensive as it must traverse vm structures. So let's do it just once here. */

bar1_first_page = vmalloc_to_page(bar_virtual_address);

/* And later... */

cookie = dma_async_memcpy_pg_to_pg(dma, virt_to_page(memory_virtual_address), memory_virtual_address & (PAGE_SIZE - 1), bar1_first_page, bar_offset, size);

r = dma_sync_wait(dma, cookie);

The solution works with kernel memory regions (allocated with "kmalloc", not "vmalloc"). For buffers containing few kB or more, throughput is more than 100x higher than in ordinal, software memcpy from ioremapped BAR to memory.

Paul