On Mon, Nov 20, 2017 at 01:28:01PM -0800, Palmer Dabbelt wrote:
> > @@ -0,0 +1,20 @@
> > +RISC-V Supervisor Binary Interface (SBI)
> > +
> > +The RISC-V privileged ISA specification mandates the presence of a supervisor
> > +binary interface that performs some operations which might otherwise require
> > +particularly complicated instructions. This interface includes
> > +inter-processor interrupts, TLB flushes, i-cache and TLB shootdowns, a
> > +console, and power management.
> > +
> > +Required properties:
> > +- compatible: must contain one of the following
> > + * "riscv,sbi" for the SBI defined by the privileged specification of the
> > + system.
> > "of the system" seems to imply that different RISC-V systems (different
> RISC-V implementations) can have different privileged specifications.
Actually, that was intentional -- I wrote it this way because different
RISC-V systems do have different privileged specifications. The RISC-V
specifications aren't frozen in time, they're just guaranteed to be
compatible in the future. For example, the user ISA document has been
updated multiple times (the C spec, eliminating some unspecified behavior)
and will continue to be updated (V and other extensions, the memory model).
The privileged spec will be updated in a compatible way just like the user
spec will be -- I know there's at least hypervisor support in the works, and
I saw some things to remove undefined behavior go past as well.
In a similar fashion, the ABI and SBI will continue to evolve. For example,
we'll probably add new system calls to extend the user ABI and new hyper
calls to extend the SBI.
My problem with the wording was that the OS somehow has to know which
version and variant of the SBI it is talking to -- either through
in-band communication (an SBI call to request SBI information, etc.), or
through devicetree or similar mechanisms.