They are programs, provided by user space, that are sandboxed to run safely in kernel space. Anything that isn't a RISC-esque instruction in BPF is a jump to a kernel function, and each flavor of BPF (BPF_PROG_TYPE_*) has it's own table of allowed functions. Most of the program types that are mainlined call functions in the net stack, but that's not intrinsic to the idea of BPF.

So in the I2C case, I could see device drivers that get called to handle the weird things that I2C devices do where you're not really sure until master interrupt time exactly what needs to happen next and you need to make a real time decision. So a BPF program that implements a state machine and is run directly from the master interrupt could be more performant and more power efficient than doing it in user space, and safer than doing it in kernel space.

Did that answer your question?