[PATCH 03/04] Freescale Ethernet combined driver

[Pantelis Antoniou]

Matt Porter wrote:
> On Tue, May 10, 2005 at 04:14:51PM +0300, Pantelis Antoniou wrote:
> 
>>Matt Porter wrote:
>>
>>>On Tue, May 10, 2005 at 08:13:48AM -0400, Dan Malek wrote:
>>>
>>>
>>>>On May 10, 2005, at 7:17 AM, Pantelis Antoniou wrote:
>>>>
>>>>
>>>>
>>>>>This patch replace iopa use with virt_to_phys.
>>>>
>>>>Not gonna work .....
>>>>
>>>>When you map uncached on 8xx you get a new vmalloc()
>>>>space.  The virt_to_xxx macros don't work on those addresses.
>>>>You need to use the dma_consistent() function, stash the
>>>>real physical address it returns and then use it where
>>>>appropriate.
>>>
>>>
>>>That and the use of virt_to_* and friends is deprecated by
>>>the DMA API. You'll never get that upstream even if it were
>>>a case where it did work.  That's a good thing to know for
>>>anybody doing other drivers...
>>>
>>>-Matt
>>>
>>>
>>
>>OK then.
>>
>>What's the recommended function to call to go from a
>>virtual -> physical address, but without doing a cache
>>flush/invalidate?
> 
> 
> There is no generic function to do that in a driver since
> no mainstream drivers in the kernel need to do it. Generally
> you can rework the driver such that you cache the DMA address
> as Dan suggested already.  I don't know your exact usage, however,
> you can allocate memory with dma_alloc_noncoherent() that is
> cached on ppc32 NOT_CACHE_COHERENT prcoessors and stash the
> dma_addr_t/void * for later use. The other way is to kmalloc and
> dma_map_single() (stashing the same way) which is basically the
> same thing.
> 
> Do you have a case where this doesn't work?
> 

No that will do just fine.

I just want to know the recommended way to do it.

> -Matt
> 
> 

Regards

Pantelis