4 Oct
2006
4 Oct
'06
7 a.m.
[ Meant to go to list, but sent to DMR only by mistake. ]
On Wed, 4 Oct 2006 dmr(a)plan9.bell-labs.com wrote:
Yep, that's the one. I regard it as a bug because it indeed happened in
user mode...
It contains
the famous Thrust Meter, a few papers by Yours Truly, and
I think it has the short assembly program that would bring a PDP-11/70
to its knees (the infamous "SPL" firmware bug).
Was this the feature (not really a bug; it's in the manual) that SPL
suppressed interrupts for one instruction after the SPL? I suppose it
was indeed a bug that this happened even in user mode where SPL was
intended to be a no-op. I remember trying this. It depends on completely
filling memory with
SPLs, which I could not figure out how to do using an instruction
sequence. However, putting a bunch of SPLs into a file and reading it
in over the program did the job.
There was a clever assembly program that did it; it relied upon the
instruction counter wrapping around (I can't remember in which direction,
or whether it first relocated itself). Anyone, it managed to fill memory
with SPLs, so the next instruction after overwriting its last instruction
was SPL, and for the foreseeable future after that...
If I find the article I'll post it here; I don't think there are too many
11/70s still in public operation.
It was a bit hard to break out of--the halt switch
didn't work. At first
I thought that power-off was the only solution, but it turned out that
holding down both reset and halt simultaneously did the job.
I'll remember that, should I ever see an emulator :-) I still remember
Ian Johnstone cursing me...
-- Dave
4 Oct
4 Oct
8:22 a.m.
New subject: [pups] Issues of AUUGN
Dave Horsfall wrote:
:-)
Oh, that would be having the HALT switch down and pressing the START
switch, by the way. That combination will trigger a Unibus reset, and
will bring the CPU out of almost all catatonic states that I've seen,
including serious bus problems.
Johnny
[ Meant to go to list, but sent to DMR only by
mistake. ]
On Wed, 4 Oct 2006 dmr(a)plan9.bell-labs.com wrote:
Yep, that's the one. I regard it as a bug because it indeed happened in
user mode...
It would have to wrap forwards. Basically, if you just have a
MOV (PC)+,(R0)+
SPL #7
and make sure that the rest of the memory don't do anything overly
foolish, and make sure that R0 starts pointing after this piece, it will
eventually fill the whole memory. The final step will be the MOV which
overwrites itself with an SPL, and then we're done.
However, another way of achieving this, if you have some kind of control
of the MMU is to just fill one page with SPLs, and then remap all of
your memory to be that page. The last page you remap is just the page
that holds all the code doing the setup.
I'm not sure you can control that under Unix well enough, but in RSX
this wouldn't be impossible, I think.
It
contains the famous Thrust Meter, a few papers by Yours Truly, and
I think it has the short assembly program that would bring a PDP-11/70
to its knees (the infamous "SPL" firmware bug).
Was this the feature (not really a bug; it's in the manual) that SPL
suppressed interrupts for one instruction after the SPL? I suppose it
was indeed a bug that this happened even in user mode where SPL was
intended to be a no-op. I remember trying this. It depends on completely
filling memory with
SPLs, which I could not figure out how to do using an instruction
sequence. However, putting a bunch of SPLs into a file and reading it
in over the program did the job.
There was a clever assembly program that did it; it relied upon the
instruction counter wrapping around (I can't remember in which direction,
or whether it first relocated itself). Anyone, it managed to fill memory
with SPLs, so the next instruction after overwriting its last instruction
was SPL, and for the foreseeable future after that... If I find the article I'll post it here; I
don't think there are too many
11/70s still in public operation.
Well, ours is occasionally. It's off at the moment, since we're not
allowed to consume that much money anymore, but Magica.Update.UU.SE is
just a key turn away from being online.
It was a bit
hard to break out of--the halt switch didn't work. At first
I thought that power-off was the only solution, but it turned out that
holding down both reset and halt simultaneously did the job.
I'll remember that, should I ever see an emulator :-) I still remember
Ian Johnstone cursing me...
10:22 a.m.
New subject: [pups] Issues of AUUGN
On Wed, 4 Oct 2006, Johnny Billquist wrote:
Yes, that's it! Except it was SPL 0; not that it made any difference, but
it was just as devastating in user mode.
Damn; I'm still trying to visualise how it works... It took me ages, the
first time I saw it; I *think* it propagates the MOV throughout memory,
leaving a trail of SPLs behind it?
Cool :-)
It was two words: "YOU XXXX!" (an indelicate term for a part of the female
anatomy) followed by the phone being slammed down...
There was a
clever assembly program that did it; it relied upon the
instruction counter wrapping around (I can't remember in which
direction, or whether it first relocated itself). Anyone, it managed
to fill memory with SPLs, so the next instruction after overwriting
its last instruction was SPL, and for the foreseeable future after
that...
It would have to wrap forwards. Basically, if you just have a
MOV (PC)+,(R0)+
SPL #7 and make sure that the rest of the memory don't
do anything overly
foolish, [...]
Not a problem in user mode?
However, another way of achieving this, if you have
some kind of control
of the MMU is to just fill one page with SPLs, and then remap all of
your memory to be that page. The last page you remap is just the page
that holds all the code doing the setup.
But you'd need kernel mode for that; this is a DoS attack (one of the
first?) launched by a user.
If I find the
article I'll post it here; I don't think there are too
many 11/70s still in public operation.
Well, ours is occasionally. It's off at the moment, since we're not
allowed to consume that much money anymore, but Magica.Update.UU.SE is
just a key turn away from being online. I'll
remember that, should I ever see an emulator :-) I still
remember Ian Johnstone cursing me...
:-) Oh, that would be having the HALT switch down and
pressing the START
switch, by the way. That combination will trigger a Unibus reset, and
will bring the CPU out of almost all catatonic states that I've seen,
including serious bus problems.
Interesting. I knew the HALT switch didn't halt the box right away; bus
transfers still completed so we were taught to W/PROT the RK-05s *after*
hitting HALT, but I didn't know it worked in combination.
-- Dave
11:11 a.m.
New subject: [pups] Issues of AUUGN
Dave Horsfall wrote:
You wouldn't need kernel mode if you were programming in RSX, since
there are system calls to remap your memory in that OS. You cannot remap
it arbitrarily, but you can remap it enough to set this up.
Cool :-)
Heck, we actually have two. But one is enough to have standing ready.
We're a computer club, so we exist at the mercy of our university.
Our second machine used to run 2.11BSD, but is now in storage. *sigh*
Wish we had a little more money designated for our fun. :-)
Yeah, the HALT switch will stop the machine at the next instruction
fetch. That is, unless you have the S.INST/S.CYCLE switch thrown to the
S.CYCLE position, which will cause the machine to stop at the next clock
cycle.
The thing to realize is that the START switch always performs a bus
reset, and then it usually also starts the clocks running in the
machine. However, if the HALT switch is active, the clocks won't start
running (obviously), so all that remains is the bus reset.
Hmm, thinking about this, I wonder how much the CPU is reset by the
start switch as well. Maybe I should check if the CPU registers are
cleared and so on... The MMU is definitely cleared.
Another thing is debugging on the front panel. You have START and CONT,
which both will start the CPU clocks. The difference is wether a reset
of the machine is done or not. If you want to debug a program from the
front panel, you keep the HALT switch down, and pull the CONT switch to
step a single instruction or single cycle (depending on the switch).
If you raise the HALT switch, CONT will cause the processor to proceed
from wherever it is at the moment, while a START also will start the
processor, but will reset a number of things so things will not actually
be able to proceed as nothing happened.
Johnny
On Wed, 4 Oct 2006, Johnny Billquist wrote:
Yes, that's it! Except it was SPL 0; not that it made any difference, but
it was just as devastating in user mode.
Interesting thing, though. I didn't know that SPL blocked interrupts for
the next instruction. Always something more to learn...
There was
a clever assembly program that did it; it relied upon the
instruction counter wrapping around (I can't remember in which
direction, or whether it first relocated itself). Anyone, it managed
to fill memory with SPLs, so the next instruction after overwriting
its last instruction was SPL, and for the foreseeable future after
that...
It would have to wrap forwards. Basically, if you just have a
MOV (PC)+,(R0)+
SPL #7 Damn; I'm still trying to visualise how it
works... It took me ages, the
first time I saw it; I *think* it propagates the MOV throughout memory,
leaving a trail of SPLs behind it?
Nope. That snippet I wrote now just copies the SPL in the instruction
after the MOV to wherever R0 is pointing. The key is that both these
instructions takes just one word.
Actually, here is an improved version:
MOV (PC)+,R0
.WORD FOO
FOO: MOV (PC)+,-(R0)
SPL #0
BR FOO
(Sorry about the DEC syntax, I've never learned as)
This code will work fine no matter what is in memory, and no matter
where we have the program.
This code will start by overwriting the .WORD FOO with the SPL, then
overwrite the MOV, and then continue backwards, wrap around, write from
the end of memory back towards the branch.
When we overwrite the branch instruction we will no longer be in our
tight loop. However all other memory than what we have here is now
containing the SPL instruction, so running through that is no problem,
and we will wrap to the start of the memory and at a later time once
again come to the MOV (PC)+,-(R0), which this time will overwrite the
SPL instruction with the SPL instruction. Basically just a decrement of
R0. Next time around, it will overwrite the MOV (PC)+,-(R0), which will
leave just the SPL instructions left.
My first version will write forwards, without an explicit loop, so it
requires that the rest of the memory didn't modify R0 and didn't branch
backwards in a never-ending loop not including the MOV.
Perhaps an even more devious piece of code is:
MOV -(PC),-(PC)
even though it won't hang the machine. :-)
However,
another way of achieving this, if you have some kind of control
of the MMU is to just fill one page with SPLs, and then remap all of
your memory to be that page. The last page you remap is just the page
that holds all the code doing the setup.
But you'd need kernel mode for that; this is a DoS attack (one of the
first?) launched by a user. If I find the article I'll post it here; I
don't think there are too
many 11/70s still in public operation.
Well, ours is occasionally. It's off at the moment, since we're not
allowed to consume that much money anymore, but Magica.Update.UU.SE is
just a key turn away from being online. Oh, that would
be having the HALT switch down and pressing the START
switch, by the way. That combination will trigger a Unibus reset, and
will bring the CPU out of almost all catatonic states that I've seen,
including serious bus problems.
Interesting. I knew the HALT switch didn't halt the box right away; bus
transfers still completed so we were taught to W/PROT the RK-05s *after*
hitting HALT, but I didn't know it worked in combination.
6849
days inactive
6849
days old
3 comments
2 participants
participants (2)
-
Dave Horsfall -
Johnny Billquist