On Wed, 6 Jun 2001, Hubert Feyrer wrote:

: It sure would be nice to have NetBSD support MMU-less systems, but maybe
: you can go into some details on how uClinux works around an MMU?

I don't know about uClinux, but the principles are pretty simple (if a bit
tedious to accomplish from scratch):

* PIC everywhere, except perhaps the kernel.  This could require a new
  MACHINE_ARCH definition on some platforms, or simply a little more ELF
  glue for program startup on others.  *Everything*, except perhaps .data
  and .rodata relocs, needs to go through the PLT and GOT.  RSS .text
  relocations are fatal, unless the .text is loaded separately per-process
  (effectively making it act like .data); a memory waste.

* Supplant the VM system with a buffer cache and malloc() manager only.
  This could be done on NetBSD with proper abstraction/encapsulation.

* Supplant malloc() with something like a pool allocator, which frees
  its memory list at exit().

* Replace mmap(), fork(), and other VM-dependent calls with ENOSYS ... or
  don't even provide the libc symbols at all (so that they don't appear to
  exist to things like autoconf).

* Provide only the "proper" version of vfork() (we call it __vfork14()).

* Provide a special "in-place" version of daemon() (which creates an
  orphaned proc structure, but doesn't change memory).

and optionally:

* Provide the spawn*() family.  These are well recognized and their presence
  encourages friendly fork/exec use by encapsulating it in one call.

* Encourage use of daemon() for backgrounding processes (rather than
  fork/exit).

* Make sure you run code that is friendly to memory.  Code that can't pass
  Purify is fatal.

...Feel free to transcribe this braindump somewhere if you'd like.  8-)

-- 
-- Todd Vierling <tv@wasabisystems.com>  *  Wasabi NetBSD:  Run with it.
-- NetBSD 1.5 now available on CD-ROM  --  http://www.wasabisystems.com/