Ken's last note led me to look at magtape error handling, and it was
not right. This has resulted in a lot of code changes in sigma_mt.c, but
the changes are mostly mechanical. Instead of trying to be clever and
do channel handling in mt_map_err or at the bottom of the mt_svc
routine, every magtape command now does its own error handling.
This makes the code repetitive, but I hope it is much clearer as a result.
The problems all stem from a single root cause: when an interrupt is
requested - via any one of dozens of mechanisms (uend, normal end,
special request, error, etc), the interrupt routine must be presented
with the full device address, including the unit, which is stored in the
channel data structures. This is so that AIO (acknowledge interrupt)
can retrieve the unit number from the channel data structures and
give it back to the invoking program.
Unfortunately, service routines, error routines, etc were simply using
the base device address, which is good enough for all channel operations
EXCEPT AIO. So the four multi-unit device (RAD, DK, DP, MT) have had
changes to reconstruct the full device address, with unit number, before
any calls into the channel are made.
DP, funnily enough, required the fewest changes. Ken was exactly right
about the need to OR the unit number into dva at the start of the unit
service routine. That had to be done in a few more places where my code
was being lazy.
RAD, DK, MT required many more changes, but they were mostly mechanical.
All references to xx_dib.dva had to be examined and replaced with a full
device address. The full device address had to be reconstructed from the
UNIT pointer and the base device address, through the magic of pointer
arithmetic.
Another sore spot was HIO. HIO stops all the active unit in its tracks
and does a channel UEND on it. The calculation of the unit number in the
UEND was not correct in a few places, notably if DP did a controller reset.
Fixes for the four modules; an updated bug history; and an expanded
design document; are attached.
- TIO status routine always returns DVS_AUTO instead of proper status.
- DP_SEEK definition is off by 1.
- reset does not properly initialize controller, seek threads.
The PDP11 and VAX have length switches - B, W, and (VAX) L - to
override the default parsing of numberic display or input as words
(PDP11) or longwords (VAX CPU) or bytes (VAX byte devices).
This worked fine if every value is filled in. On the PDP11:
sim> ideposit -b 0-3
0: 1
1: 2
2: 3
3: 4
sim>
But if the user skips an input, IDEPOSIT advances by 2, not 1:
sim> ideposit -b 0-3
0: 1
1: (cr)
3: 4
sim>
This change now produces the correct behavior.
PDP-11 simulator V3.12-2
sim> id -b 0-3
0: 1
1:
2: 3
3: 4
sim>
When performing floating point multiplication, the prior code overwrote
an additional word of the floating point fraction with zeros. This is
harmless with standard FP, as the floating variables always have space
for EP-length vars. When doing an EP multiply, this causes a word on
the stack to be zeroed. For the latest Raspbian release, this causes a
segfault as there's no padding past that var on the stack.
This fix, which has been tested against the original crashing code plus
validated using the FPP-8 diagnostics, avoids the overwrite.
Problem: Currently when viewing the Nova CPU history the carry flag is always
displayed as 0 regardless of it's actual value at the time.
Cause: The carry bit is stored in bit 17 and is lost when stored into the
int16 carry member of struct Hist_entry
Solution: Shift carry into bit 0 before storing it in the CPU history.
Improve code for displaying carry bit.
