a HAVE_GCC28 check-variable that can now be used to add other gcc-2.8
flags in cases where they may be useful, or to remove gcc 2.7.2 "bug
workaround" flags.)
up on systems that have Zorro I/O space allocated outside the Z2 I/O
region, and then only if kvm usage is high enough to begin allocating
pages already mapped for hardware mappings. Found and fixed by Niklas
Hallqvist on OpenBSD.
From niklas@cvs.openbsd.org:
Yay! This fixes a bug that has been there since day one of the amiga port.
We have never protected the kvm area that maps Zorro I/O registers in the
Z2 memory space from being allocated by the kmem_* routines. Lately kvm
usage has increased and we have needed more kvm allocated than earlier thus
this area have got allocated with random results. Most often resulting in
MMU fault panics, but also in hangs. This bug has stalled the amiga port
release builds for several weeks, but now I *hope* the amiga will have a
chance to be built and tested in time for 2.3.
it from the QuickLogic chip version byte.
If found, switch it to non-autorepeat mode (which seems to avoid the race
condition which made my keyboard driver / X server lose state under heavy
interupt load).
If not found, assume an Amiga keyboard on CIA-A.
XXX We should probe for the presence of the CIAs on the DraCo.
as this breaks C++ code that happens to indirectly include this header.
Both Matthias Scheler and I noticed this, independently.
This problem notably does not affect the atari and sun3/sun3x ports,
which have already implemented a similar solution.
are still in-order, but cached reads dont wait for the last write to finish.
Xamiga on a Altais in 8bit-mode became 30% faster servicing xanim (well, 6%
if you count xanim, too).
- make sure all sigreturn error conditions are reported to the caller,
instead of the place jumped to.
This is the bugfix part of pr 4628 by ITOH Yasufumi.
The performance optimization part will be handled seperately, after evaluating
its implications.
Testing on 68040 and removing the performance change from the proposed patch
by scottr. Half of the Amiga machdep.c change had to be done manually by me,
as the patchfile didn't apply cleanly.
XXX Yes, Amiga should be changed to use the common sig_machdep.c instead.
XXX Really soon now. I promise.
was converted to use Mach VM for Net2/4.4BSD. The user segment table
pointer was originally stored in the PCB. When Mach VM came along,
however, it was also stored in the pmap, and loaded into the PCB in
pmap_activate(). pmap_activate() would then note that the PCB's USTP
was now in sync with the pmap's USTP, and the low-level context switch
code would use the value from the PCB.
However, pmap_activate() would also load the hardware MMU context if
the pmap was the current pmap (or, in the case where pmaps can be shared,
such as in NetBSD, if the proc was the current proc). The low-level
context switch code would then reload the hardware _again_ using the
USTP from the PCB.
However, the optimization of not calling pmap_activate() if "stchanged"
was false ended up causing some processes to use stale USTP values from
the PCB when the low-level context switch code reloaded the hardware!
This was noticed by using a real vfork(2) (which worked for some time
before failing, surprisingly!)
Since I'm hard pressed to find any real optimization here (since the
hardware was always reloaded once, sometimes twice!), the code now always
calls pmap_activate(), which uses the correct USTP value (the one in the
pmap). The PCB's USTP is now ignored, and should eventually be g/c'd.
Another optimization can actually be performed, and I have added a comment
describing what it is, but have not yet implemented it.
Also note that most of the loadustp() functions where actually incomplete.
This has been corrected. These functions should probably be split up into
MMU-specific operations, and called indirectly, rather than doing constant
run-time decision making based on values that will never change during the
course of a boot's lifetime.
