From:	MERC::"uunet!CRVAX.SRI.COM!RELAY-INFO-VAX" 16-MAR-1993 00:01:26.52
To:	Info-VAX@KL.SRI.COM
CC:	
Subj:	Re: kernel routines in C (solved)

In article <1993Mar3.152100.76@tachyon.lonestar.org>, wayne@tachyon.lonestar.org (Wayne Sewell) writes:
> Is anyone writing kernel mode code in C?  If so, how do you prevent the
> change-mode dispatcher from destroying registers?
> 
> I have been writing kernel code for a long time, and have no problem with the
> mechanics of change-mode dipatchers or the cmkrnl system service.  Due to the 
> coming of Alpha, I have decided not to write any more macro code, and have
> started implementing privileged code in C instead.
> 
> As long as you are careful what you do, there is no reason you can't have
> kernel code in C.  I have been using C for FDT-stealing routines, user-written
> system services, and other things.
> 
> The problem is that I have to fight the system a little to do so.
> 
> When you call a user written system service, the change-mode dispatcher puts
> the address of the PCB into register 4.  Same thing with the $CMKRNL system
> service.  In both cases the routine being called must specify register 4 in its
> save mask.  Otherwise the calling program will suddenly find the PCB in
> register 4.
> 
> When calling a system service, user or otherwise, the registers are saved
> according to the save mask of the transfer vector, which is itself derived from
> the save mask of the actual procedure.  Thus only the registers specified by
> the called procedure are saved.
> 
> This is no problem in macro, but in C you have no control over the register
> save mask of a function, as far as I know.  C keeps track of its own register 
> usage and saves only the registers it has personally destroyed within the 
> procedure.  It doesn't know that the dispatcher clobbered register 4 before
> calling the function.
> 
> I tried to force C to save register 4 by calling the builtin function
> _WRITE_GPR (write general purpose register).  I thought if I deliberately
> destroyed register 4 myself within the procedure, that would force C to save
> it.  No effect, much to my surprise (that makes me wonder about the viability
> of those builtin functions).
> 
> I finally solved the problem by dorking around with the transfer vectors.  In
> my transfer vectors for kernel C routines, the .transfer statement points to
> the actual routine the way it's supposed to, but the .mask statement points to
> a dummy mask that saves all registers.  Kludgy, I know, but it works.  Now my
> system services written in C operate normally and register 4 isn't destroyed
> any more.
> 
> I do something similar for cmkrnl by calling a macro jacket routine.
> 
> These methods work okay.  I was just wondering if there was a better way.
> 
> Wayne
> 
> -- 
> Wayne Sewell                            |INET: wayne@tachyon.lonestar.org

The problem has been solved over in vmsnet.internals.  

> You can do:
>         .MASK   foo,^m<r4>
> 
> (I just digested a V4 MACRO manual last night :-)
> Kenn Humborg
> elehumborg@orbsen.ucg.ie

This is the solution to the problem.  It turns out you *can* OR an arbitrary 
value into the mask.  I have been programming in macro for over ten years, but 
I didn't know that .mask could do that.  I wonder if it was a recent addition, 
or the capability has been there all along (obviously it was in V4, according
to the above).

This technique can be used for *all* user written system services, not just
those written in C.  If you do this, you don't have to remember to add R4 to
the save mask of a macro routine.

Wayne
-- 
==============================================================================
Wayne Sewell                            |INET: wayne@tachyon.lonestar.org
Tachyon Software Consulting             |UUCP: ...!{letni,mic}!tachyon!wayne
P. O. Box 550937, Dallas TX  75355-0937 |Voice: (214)-553-9760, Fax: -553-0077
==============================================================================
Curly: "I keep tryin' to think, but nuthin' happens!"