On Tue, Jul 12, 2022 at 11:58 AM Li Chen <me@linux.beauty> wrote:
> > On Tue, Jul 12, 2022 at 2:26 AM Li Chen <me@linux.beauty> wrote:
> > > ---- On Mon, 11 Jul 2022 21:28:10 +0800 Arnd Bergmann <arnd@xxxxxxxx> wrote ---
> > > > On Mon, Jul 11, 2022 at 2:24 PM Li Chen <me@linux.beauty> wrote:
> > > > The problem here is that the DT is meant to describe the platform in an OS
> > > > independent way, so having a binding that just corresponds to a user space
> > > > interface is not a good abstraction.
> > >
> > > Gotcha, but IMO dts + rmem is the only choice for our use case. In our real
> > > case, we use reg instead of size to specify the physical address, so
> > > memremap cannot be used.
> >
> > Does your hardware require a fixed address for the buffer? If it can be
> > anywhere in memory (or at least within a certain range) but just has to
> > be physically contiguous, the normal way would be to use a CMA area
> > to allocate from, which gives you 'struct page' backed pages.
>
> The limitation is our DSP can only access 32bit memory, but total dram is > 4G, so I cannot use
> "size = <...>" in our real case (it might get memory above 4G). I'm not sure if other vendors' DSP also has
> this limitation, if so, how do they deal with it if throughput matters.
This is a common limitation that gets handled automatically by setting
the dma_mask of the device through the dma-ranges property in DT.
When the driver does dma_alloc_coherent() or similar to gets its buffer,
it will then allocate pages below this boundary.
If you need a large contiguous memory area, then using CMA allows
you to specify a region of memory that is kept reserved for DMA
allocations, so a call to dma_alloc_coherent() on your device will
get contiguous pages from that area, and move other data in those
pages elsewhere if necessary. non-movable data is allocated from
pages outside of the CMA reserved area in this case.
Arnd
