Re: a saner API for allocating DMA addressable pages

Hi Christoph,

On 19.08.2020 08:55, Christoph Hellwig wrote:

this series replaced the DMA_ATTR_NON_CONSISTENT flag to dma_alloc_attrs
with a separate new dma_alloc_pages API, which is available on all
platforms.  In addition to cleaning up the convoluted code path, this
ensures that other drivers that have asked for better support for
non-coherent DMA to pages with incurring bounce buffering over can finally
be properly supported.

I'm still a little unsure about the API naming, as alloc_pages sort of
implies a struct page return value, but we return a kernel virtual
address.  The other alternative would be to name the API
dma_alloc_noncoherent, but the whole non-coherent naming seems to put
people off.  As a follow up I plan to move the implementation of the
DMA_ATTR_NO_KERNEL_MAPPING flag over to this framework as well, given
that is also is a fundamentally non coherent allocation.  The replacement
for that flag would then return a struct page, as it is allowed to
actually return pages without a kernel mapping as the name suggested
(although most of the time they will actually have a kernel mapping..)

In addition to the conversions of the existing non-coherent DMA users
the last three patches also convert the DMA coherent allocations in
the NVMe driver to use this new framework through a dmapool addition.
This was both to give me a good testing vehicle, but also because it
should speed up the NVMe driver on platforms with non-coherent DMA
nicely, without a downside on platforms with cache coherent DMA.

I really wonder what is the difference between this new API and 
alloc_pages(GFP_DMA, n). Is this API really needed? I thought that this 
is legacy thing to be removed one day...

Maybe it would make more sense to convert the few remaining drivers to 
regular dma_map_page()/dma_sync_*()/dma_unmap_page() or have I missed 
something?

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland