On 5.x kernels with the USB mmap problems fixed, the distributed librtlsdr
will use a zero-copy mapping for USB buffers. Unfortunately, there is
something about the nature of the mapping on ARM (at least on Pis) that
makes most access to the data extremely slow. The uc8_nodc converter is
about 35x slower in this case compared to working on a heap-allocated buffer.
Luckily, a plain memcpy() of the buffer is still reasonably fast, so
we can use a bounce buffer and copy the data out of the slow mapping, then
pass the copy to the converter. This mitigates most of the problem,
at the expense of always needing that extra copy (which does somewhat
defeat the purpose of zero-copy!)
Unfortunately, librtlsdr provides no reliable way to control or
detect the use of zero-copy mappings, so we have to assume the problem
is always there (at least on ARM) and pay the cost of an unnecessary
copy when zerocopy is _not_ in use, too.
Update all the SDR implementation to use it.
This was getting pretty ugly with code getting copy&pasted in all the SDR
implementations. Unify it all and give it a simpler API. Linked list works out
much simpler than the circular buffer. Also, simplify copying the overlap region
around by just using a separate buffer (it's only a few hundred bytes long, so
the double copy is not a big deal).
In cases where we do get an odd-length buffer (_very_ rare!)
it seems to be more about dropped USB data and not librtlsdr giving
us a partial callback; subsequent callbacks will still be aligned
with the I byte first despite the odd count.
Oliver Jowett
Rosen Penev
Oliver Jowett
Mictronics