CRASH(8) CRASH(8)
NAME
crash - what to do when the system crashes
DESCRIPTION
This section gives at least a few clues about how to proceed
if the system crashes. It can't pretend to be complete.
Bringing it back up. If the reason for the crash is not evi-
dent (see below for guidance on `evident') you may want to
try to dump the system if you feel up to debugging. At the
moment a dump can be taken only on magtape. With a tape
mounted and ready, stop the machine, load address 44, and
start. This should write a copy of all of core on the tape
with an EOF mark. Caution: Any error is taken to mean the
end of core has been reached. This means that you must be
sure the ring is in, the tape is ready, and the tape is
clean and new. If the dump fails, you can try again, but
some of the registers will be lost. See below for what to
do with the tape.
In restarting after a crash, always bring up the system
single-user. This is accomplished by following the direc-
tions in boot(8) as modified for your particular installa-
tion; a single-user system is indicated by having a particu-
lar value in the switches (173030 unless you've changed
init) as the system starts executing. When it is running,
perform a dcheck and icheck(1) on all file systems which
could have been in use at the time of the crash. If any
serious file system problems are found, they should be
repaired. When you are satisfied with the health of your
disks, check and set the date if necessary, then come up
multi-user. This is most easily accomplished by changing
the single-user value in the switches to something else,
then logging out by typing an EOT.
To even boot UNIX at all, three files (and the directories
leading to them) must be intact. First, the initialization
program /etc/init must be present and executable. If it is
not, the CPU will loop in user mode at location 6. For init
to work correctly, /dev/tty8 and /bin/sh must be present.
If either does not exist, the symptom is best described as
thrashing. Init will go into a fork/exec loop trying to
create a Shell with proper standard input and output.
If you cannot get the system to boot, a runnable system must
be obtained from a backup medium. The root file system may
then be doctored as a mounted file system as described
below. If there are any problems with the root file system,
it is probably prudent to go to a backup system to avoid
working on a mounted file system.
CRASH(8) CRASH(8)
Repairing disks. The first rule to keep in mind is that an
addled disk should be treated gently; it shouldn't be
mounted unless necessary, and if it is very valuable yet in
quite bad shape, perhaps it should be dumped before trying
surgery on it. This is an area where experience and
informed courage count for much.
The problems reported by icheck typically fall into two
kinds. There can be problems with the free list: duplicates
in the free list, or free blocks also in files. These can
be cured easily with an icheck -s. If the same block appears
in more than one file or if a file contains bad blocks, the
files should be deleted, and the free list reconstructed.
The best way to delete such a file is to use clri(1), then
remove its directory entries. If any of the affected files
is really precious, you can try to copy it to another device
first.
Dcheck may report files which have more directory entries
than links. Such situations are potentially dangerous; clri
discusses a special case of the problem. All the directory
entries for the file should be removed. If on the other
hand there are more links than directory entries, there is
no danger of spreading infection, but merely some disk space
that is lost for use. It is sufficient to copy the file (if
it has any entries and is useful) then use clri on its inode
and remove any directory entries that do exist.
Finally, there may be inodes reported by dcheck that have 0
links and 0 entries. These occur on the root device when
the system is stopped with pipes open, and on other file
systems when the system stops with files that have been
deleted while still open. A clri will free the inode, and
an icheck -s will recover any missing blocks.
Why did it crash? UNIX types a message on the console type-
writer when it voluntarily crashes. Here is the current
list of such messages, with enough information to provide a
hope at least of the remedy. The message has the form
`panic: ...', possibly accompanied by other information.
Left unstated in all cases is the possibility that hardware
or software error produced the message in some unexpected
way.
blkdev
The getblk routine was called with a nonexistent major
device as argument. Definitely hardware or software
error.
devtab
Null device table entry for the major device used as
argument to getblk. Definitely hardware or software
CRASH(8) CRASH(8)
error.
iinit
An I/O error reading the super-block for the root file
system during initialization.
out of inodes
A mounted file system has no more i-nodes when creating
a file. Sorry, the device isn't available; the icheck
should tell you.
no fs
A device has disappeared from the mounted-device table.
Definitely hardware or software error.
no imt
Like `no fs', but produced elsewhere.
no inodes
The in-core inode table is full. Try increasing NINODE
in param.h. Shouldn't be a panic, just a user error.
no clock
During initialization, neither the line nor pro-
grammable clock was found to exist.
swap error
An unrecoverable I/O error during a swap. Really
shouldn't be a panic, but it is hard to fix.
unlink - iget
The directory containing a file being deleted can't be
found. Hardware or software.
out of swap space
A program needs to be swapped out, and there is no more
swap space. It has to be increased. This really
shouldn't be a panic, but there is no easy fix.
out of text
A pure procedure program is being executed, and the
table for such things is full. This shouldn't be a
panic.
trap
An unexpected trap has occurred within the system.
This is accompanied by three numbers: a `ka6', which is
the contents of the segmentation register for the area
in which the system's stack is kept; `aps', which is
the location where the hardware stored the program sta-
tus word during the trap; and a `trap type' which
encodes which trap occurred. The trap types are:
CRASH(8) CRASH(8)
0 bus error
1 illegal instruction
2 BPT/trace
3 IOT
4 power fail
5 EMT
6 recursive system call (TRAP instruction)
7 11/70 cache parity, or programmed interrupt
10 floating point trap
11 segmentation violation
In some of these cases it is possible for octal 20 to be
added into the trap type; this indicates that the processor
was in user mode when the trap occurred. If you wish to
examine the stack after such a trap, either dump the system,
or use the console switches to examine core; the required
address mapping is described below.
Interpreting dumps. All file system problems should be taken
care of before attempting to look at dumps. The dump should
be read into the file /usr/sys/core; cp(1) will do. At this
point, you should execute ps -alxk and who to print the pro-
cess table and the users who were on at the time of the
crash. You should dump ( od(1)) the first 30 bytes of
/usr/sys/core. Starting at location 4, the registers R0, R1,
R2, R3, R4, R5, SP and KDSA6 (KISA6 for 11/40s) are stored.
If the dump had to be restarted, R0 will not be correct.
Next, take the value of KA6 (location 022(8) in the dump)
multiplied by 0100(8) and dump 01000(8) bytes starting from
there. This is the per-process data associated with the
process running at the time of the crash. Relabel the
addresses 140000 to 141776. R5 is C's frame or display
pointer. Stored at (R5) is the old R5 pointing to the pre-
vious stack frame. At (R5)+2 is the saved PC of the calling
procedure. Trace this calling chain until you obtain an R5
value of 141756, which is where the user's R5 is stored. If
the chain is broken, you have to look for a plausible R5, PC
pair and continue from there. Each PC should be looked up
in the system's name list using adb(1) and its `:' command,
to get a reverse calling order. In most cases this proce-
dure will give an idea of what is wrong. A more complete
discussion of system debugging is impossible here.
SEE ALSO
clri(1), icheck(1), dcheck(1), boot(8)