-----------------------------------------------------------------------------
From: sequent!fubar@uunet.uu.net
Subject: how to do some Sequent stuff
Date: Fri, 27 Jul 90 16:14:17 PDT
How to do lots of neat things on a Sequent
-----------------How to boot it:
To boot a kernel from the Dynix disk (which was sd(0,0) when I
last checked), you type:
b sd(0,0)kernelname
The Sprite kernel I left there was called "vmsprite."
You may specify arguments as usual after the kernel name, e.g.,
b sd(0,0)kernelname -f
does a fastboot.
The auto reboot functionality is controlled by the setting of
the AUTO switch on the front panel. If AUTO is off, the machine will
not automatically boot after a crash or reset. If AUTO is on, then
boot string 0 will be executed to boot the machine.
You cannot boot diskless, unless you want to write an 8K disk
bootstrap that does the diskless stuff for you.
There is (or was when I was there) a SCEDMON (The boot
monitor, with a "*" prompt) quick reference card with most of the
commands on it.
-----------------Everything you never wanted to know about the SCED:
From SCEDMON, type:
ra
Which prints everything stored in the battery backed up RAM on
the SCED. The machine will respond with something like this:
Date 90/07/27 22:43:30 UTC
boot flags: 0x0
boot 0: 0x0: zd(0,0)dynix
boot 1: 0x64000: zd(0,0)stand/dump zd(0,7) 8000 /dev/rzd0h 81920
scc: 0 pf=8020 tf=8020 baud=9600 *1 pf=80A0 tf=80A0 baud=9600
ileave=1 erase=^H kill=^X int=^P
system mode: wide, tsize 16, bsize 16, copy-back cache
The "boot flags" is the default numeric argument to pass to
the booted program. Dynix uses this to set various boot conditions:
0x01: ask for file name (interpreted by Dynix /boot program)
0x02: single user
0x04: don't sync
0x08: don't reboot (if auto is set), just halt (after program is done)
0x10: name given for /etc/init
0x20: firmware: don't start controller
0x40: firmware: don't init system
0x80: boot aux ("boot 1") boot name
0x100: firmware: only build config tables
0x200: boot with cache off (don't try this at home)
Some of the possible values are recognised by the SCED firmware,
others by the Dynix kernel. The Symmetry Sprite kernel ignores this
number entirely, so not all of these will be effective under Sprite.
The "boot 0" and "boot 1" fields are the stored boot strings.
Normally, boot 0 will boot the OS, and boot 1 will boot the memory
dumper. You can change the boot strings with:
wnD=string
where "D" is boot name you want to fiddle with, and "string"
is what you want it to be. You can also boot Dynix and use the
/etc/bootflags program. You can't alter these things from Sprite
because I haven't done the SCED console memory driver necessary for
it.
The "scc" line breaks down like this:
scc: 0 pf=8020 tf=8020 baud=9600 *1 pf=80A0 tf=80A0 baud=9600
Everything from the first "0" to the "*" is the data for the
local console port, everything from the "1" to the end of the line is
for the remote port. The "*" is located either before the "0" or the
"1," depending upon the setting of the front panel LOCAL and REMOTE
switches. "pf" and "tf" are the terminal flags word for the
respective ports, 8020 is the default, 80A0 differs in that sending a
BREAK down the line will do the equivalent of pressing the RESET front
panel switch. With 8020 set, a BREAK will do baud rate rotation,
which isn't nearly as exciting. The definition of each bit in the
terminal flags word is given in the SCEDMON quick reference. The
"baud=9600" part of the line should be fairly self-explanatory.
The "ileave" specifies whether or not memory interleave is on
or off. I've never messed with this, so I don't really know what
happens if you change it. The "erase," "kill," and "int" fields are
the console erase, kill and interrupt characters active under SCEDMON.
You can change them with the "we=C" "wk=C" and "wx=C" commands,
respectively. Note that hitting the interrupt character will stop the
pre-boot self-tests. Interrupting the tests this way may leave the
machine in a weird state.
I have not ever needed anything on the "system mode" line, I
doubt you will either. I'm not entirely sure what the data
represents.
BTW, one useful command not in the SCEDMON quick reference
(well, not in mine anyway) is "zap." This will completely reset the
SCED, and is useful to me from time to time.
-----------------How to format a disk:
There are 3 wren3 disks, sd0-sd2 ("sd" for Scsi Disk, I
guess). sd0 contains Dynix, you can leave that there, or not. The
Symmetry disk label goes in sector 15, which was probably a bad place
to put it. I believe sector 16 is empty, you might want to change all
of the proper defines before you mess with the disks a lot. Anyway, I
brought a labeldisk program with me that should label the disk. The
partition tables for a wren3 are:
part start length
a 0 50
b 50 106
c 0 965
d 156 50
e 206 757
f 0 0
g 156 808
h 0 0
which looks vaguely like:
0 ---------------------- C len=965 -------------------------------- 964
| 50 156 206 !
| -- A len=50 --! -- B len=106 -- ! -- D len=50 -- ! -- E len=757 -- !
| ! ! ! !
| ! ! --------- G len=808 -------------!
If you change the partition tables on the disk label, the
standalone disk boot and possibly the kernel as well will break.
Once you label the disk properly, fsmake should do its thing
without trouble.
-----------------How to set up to boot from Sprite disks
There's a program in /sprite/src/boot/sdboot, that generates the
8K disk bootstrap. Before I get into it, here's how the boot procedure
goes:
you type: "b sd(0,0)whatever" at SCEDMON.
The SCED reads the first 16 sectors off of sd0 (or whichever
disk you've specified), loads them in at 0, and begins executing at 0.
The SMAGIC standalone magic number is really an i386 instruction to
make this work properly. Alert readers will note that the label lies
in the last sector of this data, originally I was going to have the
disk bootstrap using the label partition tables, but I ran out of code
space, and it would've been difficult to relabel the disks on our
machine.
The code goes from 0, and jumps to the a_bootstrap field of the
exec structure, which contains various i386 initialization things.
From there, it loads in whichever kernel file you've
specified, magically moves it to 0 and begins executing it. The
kernel file text must start at 0x4000, and the gap from the end of the
exec struct to 0x4000 must be padded with zeroes. This empty area is
filled in with system configuration data by the SCED.
Anyway, back to the disk boot. The source directory
(/sprite/src/boot/sdboot/sym.md) should contain a binary. If you dd
the first 15*512 bytes from this file onto the first 15 sectors of the
raw device (which I'll tell you how to make in just a second), you
should be able to boot Sprite kernels from any Sprite filesystem on
that disk.
-----------------How to make devices for the sd disks
The major numbers for the sd disks are the same as all of the
other sd disks (DEV_SCSI_DISK, which I believe is 4). The minor numbers
go like this:
part sd0 sd1 sd2
a 0 8 16
b 1 9 17
c 2 10 18
d 3 11 19
e 4 12 20
f 5 13 21
g 6 14 22
h 7 15 23
..and so on.
---------------------------------------------------------------------------
Everything we messed with for sequent support in the machine
independant kernel sources should be inside "#ifdef sequent." As I
recall, there really wasn't that much.
In the regular sources, pretty much everything either just
needed to be compiled, and was ok, or had stuff that went in separate
directories (like cc1.sym). There's some stuff in the C library that
didn't go behind ifdefs (disk labels and the like, mostly), there
should be RCS files for all the changes. The RCS files I have are:
Disk.man,v diskFragIO.c,v diskIO.c,v
disk.h,v diskHeader.c,v diskPrint.c,v
pdev.c,v
Sync_GetLock.s,v Sync_Unlock.s,v
fsStubs.s,v profStubs.s,v sysStubs.s,v vmStubs.s,v
netStubs.s,v sigStubs.s,v testStubs.s,v
procStubs.s,v syncStubs.s,v userSysCallInt.h,v
---------------------------------------------------------------------------
From: jhh@sprite.Berkeley.EDU (John H. Hartman)
Date: Thu, 1 Nov 1990 10:45:57 PST
Subject: news from fubar
I just got off the phone with fubar. Here's what we talked about.
It looks possible to poll for an ethernet packet. The routine
se_intr() in net/symm.md/netScedEther.c has a loop that processes all
pending packets. We just have to write a polling routine that does the
same thing.
Sequent found a bug in gas 1.37 that caused it to drop a byte in the
middle of a file. This only happens if you specify an input file.
It doesn't happen if you read from a pipe. Symptoms usually are a
syntax error, but it is possible you'd lose a byte out of a constant
or something.
He's going to send us whatever hardware documentation he can get his
hands on.
---------------------------------------------------------------------------
Subject: Re: debugging a kernel
Date: Fri, 19 Oct 90 12:49:09 PDT
From: fubar@sequent.com
>What is the proper procedure for debugging a kernel?
Well, about all you can really do is either load it up with
printfs (yuk), or dump vmcores and run /usr/etc/crash or adb on them.
Crash comes with Dynix, it hasn't got any documentation. It knows all
about Dynix proc structures and stacks and whatnot; about all you can
do with it under Sprite is hex dump things. I think it sucks. Adb
you're probably familiar with, for Sprite the only real advantage it
has is that you can start a stacktrace at any arbitrary location
(crash knows where the stack is "supposed to be," and won't go beyond
those limits). Adb doesn't come with Dynix, but I can give it to you
if you want it (presumably its ok for you guys to have BSD sources).
When Sprite panics under Dynix, it'll fill in a special structure
that you can examine after the crash. It's called panic_data, and
it's found in mach/sym.md/machArchdep.c:
struct panic_data {
int pd_processor; /* Panicing engine */
char *pd_sp; /* sp to saved regs in panic frame */
char *pd_dblsp; /*sp to saved regs in dblpanic frame */
Proc_ControlBlock *pd_proc; /* Panicing process, 0 if idle */
};
pd_processor is the processor number that initially panicked.
pd_sp is the pointer to the saved regs for the initial panic (more on
that in a bit), and pd_dblsp is the same thing, but for a double panic
(two panics either sequentially on the same processor, or concurrently
on two processors). There's a global kernel variable called
"dblpanic" that is set to 1 if a double panic occurs. Anyway,
finally, pd_proc is the Proc_ControlBlock for the executing process; I
have some of the byte offsets of various important elements in this
(processID at 0x44 and machStatePtr at 0xb0 are the two I used a lot).
The pd_sp is set (as is all of this stuff) in MachPanic, in
machArchdep.c. It's the %esp register after a pushal instruction, so
the registers are all tucked away just above it. They're pushed in
this order: %eax, %ecx, %edx, %ebx, %esp (before pushal began), %ebp,
%esi, %edi. In case you're not familiar with the 386 layout, %esp is
the stack pointer, %ebp is the frame pointer, and %eip is the
instruction pointer.
The better way is to follow the pd_proc through the control
block structure. Look at Mach_State structure (pointer found at
pd_proc + 0xb0), and follow it to the machTrapEntry structure. This
is the register set when the panic trap occured, and is probably the
easiest way to find what you're after.
One thing I did pretty often was a sort of checkpointing
structure, e.g., once you know the approximate source of the trouble,
set up something like:
struct XX {
int pc;
int somestate;
} XX[512];
int XXptr = -1;
#define CKPOINT { XXptr = (XXptr == 511) ? 0 : XXptr + 1;
XX[XXptr]. pc = GetPC();
}
and spread those liberally about, then use crash/adb to go through
that stuff in the vmcore.
The other thing we did to make life easier was to run a serial
line from a tty port on another machine to the "remote" console port
on the Sprite machine. If you then put the S27 on "Remote" (a front
panel button), you can tip to that tty port, and it's the console.
Oh, I don't know if I mentioned it elsewhere or not... to get
a vmcore in the first place, you do 'b 80' at SCEDMON (the boot
monitor). This should work automagically, and put a crash dump in
/usr/crash. It'll choke on a Sprite kernel, but that's ok.
If you have any troubles with this at all or whatever, feel
free to call or email. I can telnet in and poke around now, too,
should the need arise.
-J
---------------------------------------------------------------------------