PAGE_SIZE pages (4k). An average small program uses ~6 4k pages (24k),
while the same program only uses ~20 512 byte pages (10k), so it will be
a small memory usage improvement. The large improvement will be the upcoming
ability to share page table pages between processes for shared libraries.
Remaining: should be able to give back ptp pages to the system.
- Reinstall the "dynamic page table length" that was removed some
years ago.
- Limit the user page table submap to max 5% of available memory.
- Free the page table space when a process is swapped out.
- If the UPT submap runs out of space, throw away pmap mappings
using the same algorithm as for swapping processes.
As a result of this, 4MB machines are useable again and it's even possible
to compile a kernel for 2MB machines (but it will be slow... :-)
Still to do:
- Multiprocessor fixes.
- More profiling.
This will allow improvements to the pmaps so that they can more easily defer expensive operations, eg tlb/cache flush, til the last possible moment.
Currently this is a no-op on most platforms, so they should see no difference.
Reviewed by Jason.
add bus_dma support for the vs4000 dma translation map (sgmap).
enhance the vsbus for the lance to work with it.
remove the sgmap init code from ka46/48/49.c
made a whole bunch of annoying bugs disappear; mostly depending on
bad use of NBPG in non-MD code. The VAX port was the only port that
used this historical "feature".
The CL* macros should probably go away totally, there is no reason
at all to keep them.
valid bit. This is faster than the "unmap all" solution that were described
in that Mach paper _and_ it eliminates the need for checking the wired bit.
As a result of this; swapping started working again on vax :-)
page {0,1} base and length registers in the pmap structure. They will
be loaded in to the PCB when the process's address space is activated
by way of pmap_activate().
- Remove pmap_pinit() macro; it's now a real function.
- Prototypes for pmap_pinit(), pmap_activate(), pmap_deactivate().
- Map in all physical memory first in system space. This reduces
pmap_copy_page() and pmap_zero_page to 3 resp. 2 instructions.
- Have fized position user page tables. Makes the pv_table smaller
and all reverse references faster (and simpler).
- Remove the wiring code. Nobody doesn't even know what a DR32 is anymore.
- Simulate page reference bit by setting page invalid, as suggested by
Rich Draves in a paper for 1991 Mach Usenix Symposium.
This reduced the time spent in the pmap module to between 70-75% of
the previous; and made process startup _much_ faster.
and will probably work with Massbus adapters as well. (Not tested,
but it's the same code as for 11/780). Ubareset's may cause crashes
on 8600 also like 11/780, but they are more uncommon. No support
for console RL02 yet, but it's likely to come.
ragge