Penguin

You might also be interested in the CommonProgrammingBugs page. Beware of HeisenBugs. If you aren't a programmer, you can help by WritingBugReports.


Contents

  • gcc compile-time options
  • GDB
  • strace
  • signals
  • Core files
  • Debugging running processes
  • Other tools/commands

GCC Compile Options

  • have -g in your CFLAGS (for C) or CXXFLAGS (for C++) environment variable so that debugging symbols are stored in your binary objects.
  • Compile with -Wall to get all 'standard' compiler warnings.
  • Compile with -Wshadow to get a warning when you declare a variable with the same name as one in an outer scope.
  • If you are using the GNU C library (eg linux systems), you can #include <mcheck.h> to get some extra debugging for malloc(3) - see the mcheck page

GDB

Debugging under Linux is done mostly by using gdb(1).

If you want to debug a program:

  1. Compile it with debugging support. This is done by adding the -g option to the gcc(1) command line. If you are using make(1), you can export CFLAGS=-g, you should also set -Wall too, but only because it's a good idea in general. An example: gcc -g -Wall foo.c -o foo
  2. Load your program in gdb: "gdb ./programname"
  3. type "run" at the prompt to run the program.
  4. when the program crashes you should be able to type "bt full" to get a full backtrace of what the program was doing at the time.

useful gdb(1) commands:

bt full
give a complete backtrace of a program. If a program crashes this is what the programmer will want from you.
print
This lets you print out various expressions, eg: "print node" "print *node" "print node->key" "print node->next->key" etc.
break
This lets you set breakpoints at functions or lines in the source code, eg: "break main" or "break sourcefile.cpp:55". This can be abbreviated as just "b".
run
run a program up until it encounters a break point or completes.
cont
start running from the current command until the next break point or the end.
step
step to the next command, or into a function call (ie go to the instructions within that function).
next
step to the next command, or over a function call (ie treat the call as a single command)
frame
change which frame you are working on. eg: "frame 1" will change the scope to frame 1.

For a little more info see http://wand.net.nz/iam4/208/gdb.html


Other useful debugging tricks and traps:

strace

strace(1) lets you see what a program is doing in a coarse kind of way, if you think strace(1) is too quiet, perhaps ltrace(1)? is for you. for the bsdites amongst us, I believe these are called struss(1)? and sotrace(1)?. Darwin (MacOSX) has ptrace and ktrace (and kdump to read the created file).

The command for this is:

 strace ''programname''

if the program is already running:

 strace -p ''pid''

will also work.

signals

If your program hangs, you can press Alt-\ to send it a SIGQUIT and force it to dump core. You can also force them to dump core with the command:

 kill -QUIT ''programpid''

Core files

To allow crashing programs to create CoreDumps you have to remove the ulimit(1) on them. This can be done with the command:

 ulimit -c unlimited

Note, this is for the shell (and all its children) only.

By default core files are placed in the working directory (often the same directory the executable is in). This may not be ideal for you if the executable is on a read only file system. To change this behaviour you can use the following command. </verbatim>

echo /var/cores/core.%e.%p >/proc/sys/kernel/core_pattern

</verbatim> %e is replaced by the executable name and %p is replaced by the pid of the process. For more possible replacements see fs/exec.c in your nearest kernel source.

gdb(1) can also do postmortem analysis on core files like so:

 gdb ./''program'' ./''corefile''

If you run gdb(1) on your program and it displays the names of the functions but doesn't display their types (eg: what arguments they have or line number information) you probably didn't compile them with "-g".

modifying a running process

You can use gdb to attach to a currently running process. For example, to change where its stderr is going:

 $ gdb <executable> <process_id>
 (gdb) call close(2)
 $1 = 0
 (gdb) call open("/tmp/prog-debug", 0101)
 $2 = 2
 (gdb) cont

Note that the octal 0101 stands for O_CREAT|O_WRONLY, since gdb will complain about no debugging symbols for resolving those words otherwise. Check with your /usr/include files... the c library with debian testing at least has these definitions in /usr/include/bits/fcntl.h. (0100 + 01).

Graphical Debuggers

ddd(1)? appears to be a reasonable GUI interface to gdb(1) for those that are afraid of CommandLines.

Insight is another.

assert

use assert(3) everywhere in your source code. It's much nicer at finding your bugs closer to where the bug actually hides.

Electric Fence

Make use of electricfence (libefence) for tracking memory allocation errors. In debian this can be enabled on the fly by setting an LD_PRELOAD variable, like so
LD_PRELOAD=libefence.so.0.0 your-buggy-programme

threads

Note: When using gdb(1) to debug a threaded program, gdb(1) catches two signals (SIGPWR & SIGXCPU) which are used internally by pthreads on Linux. Use

 (gdb) handle SIGPWR pass nostop noprint
 (gdb) handle SIGXCPU pass nostop noprint

to stop gdb halting on receiving these signals.

Other useful commands

  • catchsegv - prints out the backtrace of function calls for a program receiving a SIGSEGV. - don't forget to compile with -g if you want file names and line numbers to be accessible.
  • c++filt(1) - reverses the C++ symbol name-mangling that gcc(1) does, so if the program uses c++ then you can read the function names.

Other neat tools for diagnosing memory errors are:


CategoryProgrammingBugs