Proposal for versioning ucontext

Why?

- Architectures sometimes add new reasonably per-thread CPU state, like
  larger and larger SIMD units and control/status registers.

- Some ports have some unused padding left in ucontext_t, but on other
  ports like VAX it's full.

What?

1. New syscalls in 11.0:

	int getcontextlen(struct __ucontext *ucp, size_t len);
	int setcontextlen(const struct __ucontext *ucp, size_t len);
	int _lwp_createlen(const struct __ucontext *ucp, size_t len, ...);

   (We could also call these __getcontext110 &c., since they are sort
   of versioned, but I'm calling the syscalls getcontextlen &c. to
   avoid confusion with the versioned libc stubs below.)

   New libc stubs to match, with a __RENAME in ucontext.h:

	/* sys/ucontext.h */
	int getcontext(ucontext_t *) __RENAME(__getcontext110);

   Defined roughly as:

	/* namespace.h */
	#define	getcontext	__getcontext110

	/* libc/sys/getcontext.c */
	int
	getcontext(ucontext_t *ucp)
	{
		SYSCALL2(getcontextlen)(ucp, sizeof(*ucp));
	}

   That way, existing programs that invoke the `getcontext' syscall or
   refer to the `getcontext' symbol in libc don't change behaviour.

   This applies to dlsym("getcontext") too.  New programs that want to
   use dlsym on libc.so for getcontext have to use
   dlsym("__getcontext110"), but that's not a new problem, and in
   principle we could add __builtin_asm_name(getcontext) =
   "__getcontext110" (i.e., expand to the associated symbol rename) to
   resolve this generally: https://wiki.NetBSD.org/symbol_versions

2. New libc function _lwp_makecontextlen(ucontext_t *ucp, size_t len,
   ...).

   (We could call it _lwp_makecontext and do a symbol rename dance but
   that's not too important because it's a NetBSD-specific internal API
   mainly for libpthread.)

3. Ports may add members to the end of mcontext_t, but may not delete
   members, insert members earlier, or change alignment of existing
   members.

   That way, existing libraries that use sigaction with SA_SIGINFO
   continue to work.

   Existing compiled code can't get at the new members of mcontext_t
   via ucontext_t since it wasn't built with new header files, of
   course, but it retains access to the existing members.

4. Every time _any_ port expands its mcontext_t, say on VAX in NetBSD
   12.0, we:

   (a) sed -e 's/mcontext_t/mcontext110_t/g'
           <vax/include/mcontext.h >vax/include/mcontext110.h
       I.e., save the old mcontext_t under a new name mcontext110_t.

   (b) sed -e 's/ucontext_t/ucontext110_t/g'
           -e 's/mcontext_t/mcontext110_t/g'
           <sys/ucontext.h >compat/ucontext.h
       I.e., save the old ucontext_t under a new name ucontext110_t.

   (c) Rename the symbol in ucontext.h for any ports which have
       expanded getcontext in 12.0:

	/* sys/ucontext.h */
	int getcontext(ucontext_t *)
	#ifdef __vax__
	    __RENAME(__getcontext120)
	#else
	    __RENAME(__getcontext110)
	#endif
	    ;

       (Or, alternatively, we could

	#define	__GETCONTEXT_RENAME	__RENAME(__getcontext120)

       in vax/include/mcontext.h, and use

	int getcontext(ucontext_t *) __GETCONTEXT_RENAME;

       in sys/ucontext.h, to keep it tidier.)

   (d) Create a stub for compat __getcontext110 so existing programs
       continue to work:

	#include <compat/ucontext110.h>
	int
	__getcontext110(ucontext110_t *ucp)
	{
		SYSCALL2(getcontextlen)(ucp, sizeof(*ucp));
	}

       Note that the definition is the essentially same as before, and
       isn't machine-dependent, so this is a low risk for auditing and
       testing.

   (e) Teach the kernel to pass the size through to MD logic to
       dispatch on it, like we do with compat ioctls.