Re: [PATCH v2 2/3] MIPS: Loongson: Add DMA support for 7A1000

[Tiezhu Yang]

On 03/25/2020 09:27 PM, Christoph Hellwig wrote:
> On Mon, Mar 23, 2020 at 10:59:14AM +0800, Tiezhu Yang wrote:
> > Implement __phys_to_dma() and __dma_to_phys() according to the
> > node id offset in the 7A1000 DMA route config register.
> Can you please try to just use the dma_pfn_offset field in struct device
> for all loongson platforms?  I'm pretty sure I asked for that last time
> around..

Hi Christoph,

In the current market, the most used bridge chip on the Loongson platform
are AMD RS780E and Loongson 7A1000, the AMD RS780E is already supported by
the mainline kernel.

If use the default implementation of __phys_to_dma() and __dma_to_phys()
in dma-direct.h when CONFIG_ARCH_HAS_PHYS_TO_DMA is not set, it works
well used with 7A1000 on the Loongson single-way and multi-way platform,
and also works well used with RS780E on the Loongson single-way platform,
but the DMA address will be wrong on the non-node0 used with RS780E on
the Loongson multi-way platform.

Just as the description in the code comment, the devices get node id from
40 bit of HyperTransport bus, so we extract 2 bit node id (bit 44~45) from
48 bit address space of Loongson CPU and embed it into HyperTransport bus
(bit 37-38), this operation can be done only at the software level used
with RS780E on the Loongson multi-way platform, because it has no hardware
function to translate address of node id, this is a hardware compatibility
problem.

Device
    |
    | DMA address
    |
Host Bridge
    |
    | HT bus address (40 bit)
    |
   CPU
    |
    | physical address (48 bit)
    |
   RAM

The Loongson 7A1000 has dma_node_id_offset field in the DMA route config
register, the hardware can use the dma_node_id_offset to translate address
of node id automatically, so we can get correct address when just use the
dma_pfn_offset field in struct device.

For the above reasons, in order to maintain downward compatibility to
support the AMD RS780E bridge chip, it is better to use the platform
dependent implementation of __phys_to_dma() and __dma_to_phys().

Thanks,

Tiezhu Yang