$NetBSD: TODO.hp300,v 1.4 2023/09/08 19:04:28 andvar Exp $

1. Create and use an interrupt stack.
   Well actually, use the master SP for kernel stacks instead of
   the interrupt SP.  Right now we use the interrupt stack for
   everything.  Allows for more accurate accounting of systime.
   In theory, could also allow for smaller kernel stacks but we
   only use one page anyway.

2. Copy/clear primitives could be tuned.
   What is best is highly CPU and cache dependent.  One thing to look
   at are the copyin/copyout primitives.  Rather than looping using
   MOVS instructions, you could map an entire page at a time and use
   bcopy, MOVE16, or whatever.  This would lose big on the VAC models
   however.

3. Sendsig/sigreturn are pretty bogus.
   Currently we can call a signal handler even if an exception
   occurs in the middle of an instruction.  This causes the handler
   to return right back to the middle of the offending instruction
   which will most likely lead to another exception/signal.
   Technically, I feel this is the correct behavior but it requires
   saving a lot of state on the user's stack, state that we don't
   really want the user messing with.  Other 68k implementations
   (e.g. Sun) will delay signals or abort execution of the current
   instruction to reduce saved state.  Even if we stick with the
   current philosophy, the code could be cleaned up.

4. Ditto for AST and software interrupt emulation.
   Both are possibly over-elaborate and inefficiently implemented.
   We could possibly handle them by using an appropriately planted
   PS trace bit.

5. Make use of transparent translation registers on 030/040 MMU.
   With a little rearranging of the KVA space we could use one to
   map the entire external IO space [ 600000 - 20000000 ).  Since
   the translation must be 1-1, this would limit the kernel to 6mb
   (some would say that is hardly a limit) or divide it into two
   pieces.  Another promising use would be to map physical memory
   within the kernel.  This allows a much simpler and more efficient
   implementation of /dev/mem, pmap_zero_page, pmap_copy_page and
   possible even kernel-user cross address space copies.  However,
   it does eat up a significant piece of kernel address space.

6. Create a 32-bit timer.
   Timers 2 and 3 on the MC6840 clock chip can be concatenated together to
   get a 32-bit countdown timer.  There are at least three uses for this:
   1. Monitoring the interval timer ("clock") to detect lost "ticks".
      (Idea from Scott Marovich)
   2. Implement the DELAY macro properly instead of approximating with
      the current "while (--count);" loop.  Because of caches, the current
      method is potentially way off.
   3. Export as a user-mappable timer for high-precision (4us) timing.
   Note that by doing this we can no longer use timer 3 as a separate
   statistics/profiling timer.  Should be able to compile-time (runtime?)
   select between the two.

7. Conditional MMU code should be restructured.
   Right now it reflects the evolutionary path of the code: 320/350 MMU
   was supported and PMMU support was glued on.  The latter can be ifdef'ed
   out when not needed, but not all of the former (e.g. ``mmutype'' tests).
   Also, PMMU is made to look like the HP MMU somewhat ham-stringing it.
   Since HP MMU models are dead, the excess baggage should be there (though
   it could be argued that they benefit more from the minor performance
   impact).  MMU code should probably not be ifdef'ed on model type, but
   rather on more relevant tags (e.g. MMU_HP, MMU_MOTO).

8. Redo cache handling.
   There are way too many routines which are specific to particular
   cache types.  We should be able to come up with a more coherent
   scheme (though HP 68k boxes have just about every caching scheme
   imaginable: internal/external, physical/virtual, writeback/writethrough)
   See, for example, Wheeler and Bershad in ASPLOS 92.

9. Sort the free page list.
   The DMA hardware on the 300 cannot do scatter/gather IO.  For example,
   if an 8k system buffer consists of two non-contiguous physical pages
   it will require two DMA transfers (and hence two interrupts) to do the
   operation.  It would take only one transfer if they were physically
   contiguous.  By keeping the free list ordered we could potentially
   allocate contiguous pages and reduce the number of interrupts.  We can
   consider doing this since pages in the free list are not reclaimed and
   thus we don't have to worry about distorting any LRU behavior.
----
Mike Hibler
University of Utah CSS group
mike@cs.utah.edu