On Sat, Feb 19, 2022 at 08:19:00AM +0100, Christoph Hellwig wrote:

On Sat, Feb 19, 2022 at 08:52:21AM +0800, Baoquan He wrote:

quoted

Use dma_alloc_noncoherent() instead of directly allocating buffer
from kmalloc with GFP_DMA. DMA API will try to allocate buffer
depending on devices addressing limitation.

I think it would be better to still allocate the buffer at allocation
time and then just transfer ownership using dma_sync_single* in the I/O
path to avoid the GFP_ATOMIC allocation.

This driver allocates the buffer at initialization step and maps the buffer
for DMA_TO_DEVICE and DMA_FROM_DEVICE when processing IO.

But after making this driver to use dma_alloc_noncoherent(), remapping
dma_alloc_noncoherent()-ed buffer is strange So I just made it to allocate
the buffer in IO path.

At this point I thought we need an API that allocates based on
address bit mask (like dma_alloc_noncoherent()), which does not maps buffer
into dma address. __get_free_pages/kmalloc(GFP_DMA) has been so confusing..

Hmm.. for this specific case, What about allocating two buffers
for DMA_TO_DEVICE and DMA_FROM_DEVICE at initialization time?

Thanks,
Hyeonggon