When I was testing a bus-mastering device I found that the calls to mmap to produce buffer descriptors was becoming a 
major overhead at high transaction rates.

Using buffers that are pre-mapped removes this overhead (it's performed once during driver initialisation) but this 
isn't a solution for cases where it's not practical to copy data in and out of these buffers.

The thought I've had is to build up a cache of mmap entries.  This combines the efficiency of pre-computing the mmap 
translation while still allowing the use of any memory within the process, with the (hopefully) smaller cost of a 
looking up the translation in the cache.

The one concern I have with this is the consequence of the mmap translation changing.  This is an embedded system with 
no paging to disk so I think this would only occur if unused virtual memory is released to the system and then later 
reallocated.  I don't seem why QNX would move process memory around in RAM otherwise.

The consequence of the mapping changing with a bus mastering device would be catastrophic so there would need to be a 
guarantee that memory mapping would be constant.

Or is there a better approach to an efficient bus-mastering driver that avoids the overhead of mmap on each call?

Regards.