Description
Currently our Spidev wrapper struct implements both SpiBus
and SpiDevice
from embedded-hal. I think this is likely to lead to some confusion, especially since the SpiBus
methods will still actually be using a shared SPI bus, toggling CS, and multiplexing with any other spidev devices on the same bus.
We discussed this a bit in #87 and while the main problem there is resolved by the new embedded-hal design, I think the split suggested is still worth considering. The proposal is to have two wrapper structs, SpidevBus
and SpidevDevice
perhaps, where:
SpidevBus
implements just theSpiBus
trait- we configure
spidev
to not drive the associated CS pin - we explain in the documentation that when this struct is used, it must be the only active
spidev
device on that bus - we encourage users to give it a CS pin that is not connected to anything, since whatever pin is assigned to it in the device tree can't be re-used as GPIO (and therefore can't be used as a CS)
- this struct can then be used with any of the bus-multiplexing crates (like embedded-hal-bus) to generate lots of
SpiDevice
objects, using generic GPIO for CS control, or used when dedicated bus access is required (eg to generate smart LED waveforms) - we are basically lying to the kernel about it being a single device, because there's no way to request exclusive bus access with
spidev
- we configure
SpidevDevice
implements just theSpiDevice
trait- it's already a perfect fit for
SpiDevice
, the kernel handles sharing and CS for us, even with other programs - it will work just as users expect
- but their device tree has to be set up to create the device and use the right CS
- it's already a perfect fit for
I'm happy to implement this as a PR if it sounds good to people, but it'd be great to get some feedback first or any suggestions on how it might be better done, especially if there's any clever spidev-related tricks that might be useful.