NAME
crash
—
system failure and diagnosis
DESCRIPTION
This section explains what happens when the system crashes and (very briefly) how to analyze crash dumps.
When the system crashes voluntarily it prints a message of the form
panic: why i gave up the ghost
on the console and enters the kernel debugger, ddb(4).
If you wish to report this panic, you should include the output of
the ps
and trace
commands.
Unless the ‘ddb.log’ sysctl has been disabled, anything output
to screen will be appended to the system message buffer, from where it may
be possible to retrieve it through the dmesg(8) command after a warm reboot. If the debugger command
boot dump
is entered, or if the debugger was not
compiled into the kernel, or the debugger was disabled with
sysctl(8), then the system
dumps the contents of physical memory onto a mass storage peripheral device.
The particular device used is determined by the ‘dumps on’
directive in the config(8)
file used to build the kernel.
After the dump has been written, the system then invokes the automatic reboot procedure as described in reboot(8). If auto-reboot is disabled (in a machine dependent way), the system will simply halt at this point.
Upon rebooting, and unless some unexpected inconsistency is encountered in the state of the file systems due to hardware or software failure, the system will copy the previously written dump into /var/crash using savecore(8), before resuming multi-user operations.
Causes of system failure
The system has a large number of internal consistency checks; if one of these fails, then it will panic with a very short message indicating which one failed. In many instances, this will be the name of the routine which detected the error, or a two-word description of the inconsistency. A full understanding of most panic messages requires perusal of the source code for the system.
The most common cause of system failures is hardware failure (e.g., bad memory) which can reflect itself in different ways. Here are the messages which are most likely, with some hints as to causes. Left unstated in all cases is the possibility that a hardware or software error produced the message in some unexpected way.
- no init
- This panic message indicates filesystem problems, and reboots are likely to be futile. Late in the bootstrap procedure, the system was unable to locate and execute the initialization process, init(8). The root filesystem is incorrect or has been corrupted, or the mode or type of /sbin/init forbids execution.
- trap type %d, code=%x, pc=%x
- A unexpected trap has occurred within the system; the trap types are
machine dependent and can be found listed in
/sys/arch/ARCH/include/trap.h.
The code is the referenced address, and the pc is the program counter at the time of the fault is printed. Hardware flakiness will sometimes generate this panic, but if the cause is a kernel bug, the kernel debugger ddb(4) can be used to locate the instruction and subroutine inside the kernel corresponding to the PC value. If that is insufficient to suggest the nature of the problem, more detailed examination of the system status at the time of the trap usually can produce an explanation.
- init died
- The system initialization process has exited. This is bad news, as no new users will then be able to log in. Rebooting is the only fix, so the system just does it right away.
- out of mbufs: map full
- The network has exhausted its private page map for network buffers. This usually indicates that buffers are being lost, and rather than allow the system to slowly degrade, it reboots immediately. The map may be made larger if necessary.
That completes the list of panic types you are likely to see.
Analyzing a dump
When the system crashes it writes (or at least attempts to write) an image of memory, including the kernel image, onto the dump device. On reboot, the kernel image and memory image are separated and preserved in the directory /var/crash.
To analyze the kernel and memory images preserved as bsd.0 and bsd.0.core, you should run gdb(1), loading in the images with the following commands:
# gdb GNU gdb 6.3 Copyright 2004 Free Software Foundation, Inc. GDB is free software, covered by the GNU General Public License, and you are welcome to change it and/or distribute copies of it under certain conditions. Type "show copying" to see the conditions. There is absolutely no warranty for GDB. Type "show warranty" for details. This GDB was configured as "i386-unknown-openbsd4.6". (gdb) file /var/crash/bsd.0 Reading symbols from /var/crash/bsd.0...(no debugging symbols found)...done. (gdb) target kvm /var/crash/bsd.0.core
[Note that the “kvm” target is currently only supported by gdb(1) on some architectures.]
After this, you can use the where
command
to show trace of procedure calls that led to the crash.
For custom-built kernels, you should use bsd.gdb instead of bsd, thus allowing gdb(1) to show symbolic names for addresses and line numbers from the source.
Analyzing saved system images is sometimes called post-mortem debugging. There are a class of analysis tools designed to work on both live systems and saved images, most of them are linked with the kvm(3) library and share option flags to specify the kernel and memory image. These tools typically take the following flags:
-M
core- Normally this core is an image produced by savecore(8) but it can be /dev/mem too, if you are looking at the live system.
-N
system- Takes a kernel system image as an argument. This is where the symbolic information is gotten from, which means the image cannot be stripped. In some cases, using a bsd.gdb version of the kernel can assist even more.
The following commands understand these options:
fstat(1),
netstat(1),
nfsstat(1),
ps(1),
w(1),
dmesg(8),
iostat(8),
kgmon(8),
pstat(8),
trpt(8),
vmstat(8) and many others.
There are exceptions, however. For instance, ipcs(1) has renamed the -M
argument
to be -C
instead.
Examples of use:
# ps -N /var/crash/bsd.0 -M /var/crash/bsd.0.core -O paddr
The -O
paddr option
prints each process' struct proc address. This is very
useful information if you are analyzing process contexts in
gdb(1).
# vmstat -N /var/crash/bsd.0 -M /var/crash/bsd.0.core -m
This analyzes memory allocations at the time of the crash. Perhaps some resource was starving the system?
Analyzing a live kernel
Like the tools mentioned above, gdb(1) can be used to analyze a live system as well. This can be accomplished by not specifying a crash dump when selecting the “kvm” target:
(gdb) target kvm
It is possible to inspect processes that entered the kernel by
specifying a process' struct proc address to the
kvm proc
command:
(gdb) kvm proc 0xd69dada0 #0 0xd0355d91 in sleep_finish (sls=0x0, do_sleep=0) at ../../../../kern/kern_synch.c:217 217 mi_switch();
After this, the where
command will show a
trace of procedure calls, right back to where the selected process entered
the kernel.
CRASH LOCATION DETERMINATION
The following example should make it easier for a novice kernel developer to find out where the kernel crashed.
First, in ddb(4) find the function that caused the crash. It is either the
function at the top of the traceback or the function under the call to
panic
() or
uvm_fault
().
The point of the crash usually looks something like this "function+0x4711".
Find the function in the sources, let's say that the function is in "foo.c".
Go to the kernel build directory, e.g., /sys/arch/ARCH/compile/GENERIC, and do the following:
# objdump -S foo.o | less
Find the function in the output. The function will look something like this:
0: 17 47 11 42 foo %x, bar, %y 4: foo bar allan %kaka 8: XXXX boink %bloyt etc.
The first number is the offset. Find the offset that you got in the ddb trace (in this case it's 4711).
When reporting data collected in this way, include ~20 lines before and ~10 lines after the offset from the objdump output in the crash report, as well as the output of ddb(4)'s "show registers" command. It's important that the output from objdump includes at least two or three lines of C code.
REPORTING
If you are sure you have found a reproducible software bug in the kernel, and need help in further diagnosis, or already have a fix, use sendbug(1) to send the developers a detailed description including the entire session from gdb(1).