On 09.03.23 14:52, Andy Shevchenko wrote:
On Thu, Mar 09, 2023 at 02:40:10PM +0100, Oliver Neukum wrote:
On 09.03.23 08:10, Ye Xiang wrote:
+ u8 obuf[LJCA_GPIO_BUF_SIZE];
+ u8 ibuf[LJCA_GPIO_BUF_SIZE];
And here we have a violation of DMA coherency rules.
Basically you cannot embed buffers into other data structures
if they can be subject to DMA.
Huh?!
The problem here is alignment. But other than that I can't see the issue with
embedding into structures the instances of which will be allocated on the heap.
Hi,
These constraints are documented in dma-api-howto, but perhaps this
is inconvenient to read through.
Let me explain for the input case. On certain CPUs DMA does not
update CPU caches. Hence when you want to read data generated by DMA
you must read from RAM. Hence you invalidate the cache line with dma_map_*
operations.
Those cache lines must stay invalidated. If you wish to guarantee that,
you cannot access a data structure that shares a cache line with a buffer,
until you are sure that DMA is finished.
On the affected architectures kmalloc() makes sure that no allocation straddles
cache lines.
Regards
Oliver
