dma_cache_sync replacement call

dma_cache_sync replacement call

[y bhanu]

Hi guys,

I am porting a Linux wireless driver from MIPS to ARM (ixp4xx).

As dma_cache_sync() is not implemented for ARM ,I get some nasty errors:
I am not sure whats the replacement for this deprecated call under ARM.

can  anyone help figure out what use instread of dma_cache_sync() &
dma_cache_wback_inv() calls?

Missing dma_cache_sync & dma_cache_wback_inv under arch/arm/*

Erorr log
~~~~~~~
if_bus.c:32: error: implicit declaration of function 'dma_cache_sync'
if_bus.c: In function 'bus_map_single':
if_bus.c:49: error: implicit declaration of function 'dma_cache_wback_inv'
if_bus.o] Error 1
make[8]: *** [th_dev] Error 2

Thanks,
Bhan

dma_cache_sync replacement call

[Russell King - ARM Linux]

On Wed, Dec 15, 2010 at 02:54:57PM -0800, y bhanu wrote:
> I am porting a Linux wireless driver from MIPS to ARM (ixp4xx).
> 
> As dma_cache_sync() is not implemented for ARM ,I get some nasty errors:
> I am not sure whats the replacement for this deprecated call under ARM.
> 
> can  anyone help figure out what use instread of dma_cache_sync() &
> dma_cache_wback_inv() calls?

The answer is to complain to MIPS people as to why they are using their
own private DMA cache support rather than the DMA API (whose documentation
can be found in Documentation/DMA-API.txt).

The answer is there is no straight replacement - the answer is to fix the
driver to use the proper portable interfaces.

dma_cache_sync replacement call

[y bhanu]

Hi Russel,

Here is understanding after reading DMA-API.txt and looking though
different kernel headers (arch specific & generic)
Correct me if I missed anything.

Purpose of dma_cache_sync:
  to force sync DMA buffers to be able to be accessed by CPU (programs/driver)

Alternative for dma_cache_sync:
 dma_sync_single_for_cpu()  and dma_unmap_single()
 These do the same operation of (i.e.__dma_single_dev_to_cpu ) to
trasfer the buffer ownership.

But, I really dont understand what is the differnec between
dma_sync_single_for_cpu()  and dma_unmap_single()? (Is it is just to
make the name clear?)
I think  dma_sync_single_for_cpu() bests suites to situation.

Thanks
bhanu


On Wed, Dec 15, 2010 at 2:59 PM, Russell King - ARM Linux
<linux@arm.linux.org.uk> wrote:
> On Wed, Dec 15, 2010 at 02:54:57PM -0800, y bhanu wrote:
>> I am porting a Linux wireless driver from MIPS to ARM (ixp4xx).
>>
>> As dma_cache_sync() is not implemented for ARM ,I get some nasty errors:
>> I am not sure whats the replacement for this deprecated call under ARM.
>>
>> can ?anyone help figure out what use instread of dma_cache_sync() &
>> dma_cache_wback_inv() calls?
>
> The answer is to complain to MIPS people as to why they are using their
> own private DMA cache support rather than the DMA API (whose documentation
> can be found in Documentation/DMA-API.txt).
>
> The answer is there is no straight replacement - the answer is to fix the
> driver to use the proper portable interfaces.
>

dma_cache_sync replacement call

[Arnd Bergmann]

On Thursday 16 December 2010, y bhanu wrote:
> Alternative for dma_cache_sync:
>  dma_sync_single_for_cpu()  and dma_unmap_single()
>  These do the same operation of (i.e.__dma_single_dev_to_cpu ) to
> trasfer the buffer ownership.
> 
> But, I really dont understand what is the differnec between
> dma_sync_single_for_cpu()  and dma_unmap_single()? (Is it is just to
> make the name clear?)
> I think  dma_sync_single_for_cpu() bests suites to situation.

The difference is mainly on platforms with an IOMMU, where you have
to explicitly manage the bus addresses using dma_map_* / dma_unmap_*.

dma_sync_single_for_cpu is for cases where you have a long-lived
(mapping with multiple accesses from both CPU and device, e.g. doing

dma_map_single(); /* gives returns a dma address
		     and allows the device to access it */

while (running) {
	dma_sync_single_for_cpu(); /* lets the CPU access the buffer */

	dma_sync_single_for_device(); /* lets the device access the buffer */
}

dma_unmap_single(); /*  invalidates the dma address and returns it
			to the CPU */

If the device only accesses the buffer once, you don't need to do an
explicit sync at all, just unmap after the access is done.

	Arnd