DMA_NONCOHERENT and dma_map_single

DMA_NONCOHERENT and dma_map_single

[Dimitri Torfs]

Hi,

  dma_map_single() is supposed to be called on a buffer that exactly
  starts and ends on a cacheline boundary, otherwise "bad things"
  (e.g. overwrite of data that was written by device, ...) (especially
  on dma non-coherent systems) may happen. 

  So what should be done when dma_map_single is not called
  with a sane (ptr, size) argument ?

    - is the driver (caller) considered buggy and should we return a 0
      return-value ?
    - is the driver (caller) considered buggy but we do the mapping
      anyway, hoping that the driver has not/will not touched/touch
      the boundary cachelines ?
    - should we take appropriate actions to make sure the
      cache-effects do not come into play (e.g. by using some kind of
      bounce buffer) ?


  Dimitri

-- 
Dimitri Torfs       |  NSCE 
dimitri@sonycom.com |  The Corporate Village
tel: +32 2 7008541  |  Da Vincilaan 7 - D1 
fax: +32 2 7008622  |  B-1935 Zaventem - Belgium

Re: DMA_NONCOHERENT and dma_map_single

[Ralf Baechle]

On Sun, Jan 18, 2004 at 08:50:06PM +0100, Dimitri Torfs wrote:

>   dma_map_single() is supposed to be called on a buffer that exactly
>   starts and ends on a cacheline boundary, otherwise "bad things"
>   (e.g. overwrite of data that was written by device, ...) (especially
>   on dma non-coherent systems) may happen. 
> 
>   So what should be done when dma_map_single is not called
>   with a sane (ptr, size) argument ?
> 
>     - is the driver (caller) considered buggy and should we return a 0
>       return-value ?
>     - is the driver (caller) considered buggy but we do the mapping
>       anyway, hoping that the driver has not/will not touched/touch
>       the boundary cachelines ?

The driver is considered buggy;  dma_map_single's behaviour is undefined so
it's perfectly ok if it paints neighbour's cat pink ;-)

>     - should we take appropriate actions to make sure the
>       cache-effects do not come into play (e.g. by using some kind of
>       bounce buffer) ?

Technically bounce buffers can be handled inside dma_map_single & co but
it's not a good idea.  Better set the appropriate flags so higher levels
can allocate memory with the appropriate GFP_* flags and thereby hopefully
avoid overly frequent buffer bouncing.

  Ralf