On Wed, Jun 29, 2022 at 5:48 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
>
> On Wed, Jun 29, 2022 at 10:21 AM Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
> > Regarding the amiga_chip_alloc(), I don't know what this means
> > for caching. If chip memory is cache-coherent (either uncached
> > or by snooping), then there should not be any
> > dma_map()/dma_unmap() for that case, but instead the
> > amiga_chip_alloc() function should return both the pointer
> > and the dma_addr_t token.
>
> Chip RAM is mapped uncached.
> Amifb remaps it using ioremap_wt() to get a write-through frame buffer.
Ok, so in this case the driver should skip the dma_map_single() cache
management and instead keep converting the chipram address to
a bus address directly. While the driver does an extra cache
flush on the uncached address today, it's clearly not needed and there
is probably a performance benefit to not doing it.
For simplicity, the normal bounce buffer could similarly use
dma_alloc_coherent(), which will also result in an uncached
buffer. If I read this correctly, this will also ensure that the buffer
is within the DMA mask, so if ZONE_DMA is larger than the mask,
it just returns NULL and you have to fall back to the chipram
page, rather than checking the address and freeing the buffer.
Arnd
