The internet of things seems to be coming any day now, but the state of embedded development seems to be deplorable. Almost everything is written in C or C++. Device drivers are written over and over, once for each RTOS, or worse. When high level languages are available, they seem to be implemented directly on the hardware, rather than on top of an existing RTOS, so there’s another chance to practice writing device drivers. Want a file system, or a network stack? You’ll need to patch one for the kernel of your choice.
This approach is no longer justifiable: the target devices typically have 64K of memory, 256K of memory mapped flash , and have throughput of 100Mips. It’s time embedded developers stopped playing at writing device drivers, and thought about composition.
Trying to put together an environment for experimenting with micro controllers is frustrating. For example, I have a STM32F3 Discovery board. The micro controller itself has a huge array of peripheral interfaces, and the board adds a gyroscope and accelerometer, and LEDs to light it up like a Christmas tree. It costs £9, which is cheap enough to buy several: in case you break one. I’m a software engineer, not an electrical engineer, so that’s going to happen. There’s 48K of RAM, and 256K of flash. It’s sleep mode uses 0.6mA, so if interrupts in you application are rare, you might even use less power than an Arduino.
So, what would be a productive environment?
- Device support. I want this to run on whatever board fits my project, and support all the peripherals.
- High level. I don’t manage my own memory any more, and I like abstraction.
- Interactive. I don’t want to have to compile and install new firmware just to test a bit of wiring I’ve just done.
- Interoperable. This is for the internet of things. I’m going to need to implement network protocols.
- Composable. I want to add other peoples code to mine, and I don’t mean by forking it.
How do I get it? Well, now the pain begins.
LwIP is a small footprint TCP/IP stack written in C. Almost everything mentioned below includes some support, so you can plug it in. Using it doesn’t require anything beyond supporting C binding. Some extra work might be required if you want to provide network drivers in the language of your choice.
These are the closest thing there is to an all inclusive solution.
Espruino doesn’t have much in the way of architecture documentation. There’s no description of the interpreter, so unless you read the code, you can’t know anything about the competence of the authors, or the sophistication of the interpreter. I’d make a guess that it’s based on Tiny-JS. There’s no intermediate code form, which guarantees your code takes up a lot of RAM.
Lua has co-routines, which is a big step up on being completely event oriented. eLua can execute byte code straight out of flash – rather than having to use RAM to hold programs, which is a pretty useful optimisation in this context. Lua also has a great C-binding.
Elua runs on my device, but only to the extent that it can run a REPL on the serial port. No other peripherals are supported. eLua’s concept is to be Lua as far down as possible. From the point of view of making eLua as easy to improve as possible, this is good design decision. It’s a long game though, and I don’t see anything in it’s roadmap that suggests it’s going to tackle the issue of memory management during interrupts, or compilation, when higher performance is needed. I think that’s going to mean device drivers keep getting written in C. Given that, hitching a ride on an RTOS which has momentum in this area – E.g. Chibios, seems like a pragmatic way forward, but seems to get rejected on the mailing list.
That’s not to say that eLua isn’t the right starting point to tackle these problems: it may well be.
This RTOS offers Posix support and DLLs. This would mean, for example, that it’s reasonably easy to compile various interpreters, and lot’s of open source source software. It has limited support for the F3 Discovery board – basically no more than eLua. I could choose the F4 discovery board to solve this problem. There’s an open source project to run (full) Lua under NuttX, which I hope to try out.
This doesn’t offer any sort of standard interface. It does however support a huge array of boards, including support for all the F3 Discovery peripherals. It also seems to get new boards and drivers supported very quickly. For example, Chibios supports the ADC/Timer/DMA feature I mentioned above, and had that support a month or so after the board was released. This is also the only thing I’ve actually run on the board. It’s easy to set up a tool chain and build. The samples are readable, by C standards.
Because Chibios has good support for the boards I have, and because FreeRTOS (for example) appears to have very similar features to Chibios, I haven’t investigated much further in this category.
None of the above?
Scheme might be a good choice as an embedded interpreter. I could build a system on top of Chibios. There are at least two compilers I could choose between: Chicken and Stalin. Chicken has a REPL, so it appeals more. It lacks really good GC, but I guess that might not be such a big problem in the short term. Chickens first generation is on the stack, and I can see how that might make it possible to write interrupt handlers directly in scheme, although if the stack ran out, the interrupt handler would fail.
I must admit, I’d assumed that a TCP/IP stack written in scheme was available, but I haven’t found one. Or a file system, for that matter. Still, there’s LwIP, and writing a file system in Scheme isn’t so daunting. I’m not sure I’ll be convincing a lot of people to write electricity meter firmware in scheme, but I could always add interpreters for other languages.
I guess I hinted at the top that’s there’s no clear conclusion. Suggestions?