Better way to implement I/O in a virtual machine?

I'm writing a virtual machine - not an existing architecture emulator like Virtualbox, but rather something like the JVM or BEAM - with its own instruction set, memory model, etc. Eventually I'm planning to implement a very small and simple (but turing-complete) high-level language that would compile into its bytecode, just for fun.

Of course, the machine must have some support of I/O, but I do not want to limit it only to manipulations with stdin/stdout. I imagine something like modular "virtual devices", which can be implemented as shared libraries so that the VM can load them at runtime and communicate with them through a standard interface. This way, for example, we can have "virtual devices" for standard input/output, graphics (imagine a virtual device that lets your 开发者_运维问答VM program draw stuff inside an SDL window) or maybe even network.

The question is: how should the programs written for the VM communicate with the virtual devices? I decided to mimic techniques which are employed with actual hardware and learned about port-based I/O and memory-mapped I/O. However, I'm not sure which one of them is more suitable for my goals. Can you suggest which one is better or maybe even point out a totally different technique for dealing with input/output?

Thanks in advance.

---

Both memory-mapped and port based are inappropriate for most I/O.

DMA request with block-copy is usually what you want.