Reading the linux kernel documentation on KGDB wasn’t enough for me to be able to use the newly built-in KGDB kernel debugger with version 2.6.26 or 2.6.27. The breakthrough for me was reading part of Jason Wessel’s guide.

I have two machines:

• developer – where I run gdb
• target – where the kernel is being debugged, running in VMware

Configure VMware on the developer machine

• Power down the guest (target)
• Edit the VM guest settings
• Add a serial port
  • Use named pipe /tmp/com_1 (it’s really a UNIX domain socket)
  • Configure it to “Yield CPU on poll” (under Advanced)
• Install ‘socat’, if not already installed

Configure and Compile the kernel on the developer or the target machine

• Get kernel 2.6.26 or newer
• make menuconfig # or make gconfig
• Under Kernel Hacking:
  • enable KGDB
  • enable the Magic SysRq key
  • enable “Compile the kernel with debug info”
• Build kernel: make

Configure target

• Enable Magic SysRq key on target:
  • Edit /etc/sysctl.conf and set kernel.sysrq = 1
  • or run sysctl -w kernel.sysrq=1 # this doesn’t survive a reboot
• Install developer kernel
  • On the developer machine: rsync -av --exclude .git ./ root@target.host.name:/mnt/work/linux-2.6.26
  • On the target, a RedHat based system: make install make modules_install
• Edit /boot/grub/grub.conf and set timeout=15
• Boot into the newly installed kernel

Start debugging

• On target: echo ttyS0 > /sys/module/kgdboc/parameters/kgdboc
• On developer: socat -d -d /tmp/com_1 PTY: # notice what pty is allocated — /dev/pts/1 in my case gdb vmlinux set remotebaud 115200 target remote /dev/pts/1
• On target, do one of the following:
  • echo "g" > /proc/sysrq-trigger
  • Type ALT-SysRq-G
• Ready, get set, go! Go back to developer machine and use gdb to set breakpoints, continue, etc.

I set up debugging because I wanted to understand the behavior of the kernel when loading a module. It turns out that loading of the module failed because sitting in a debugger delayed the execution, causing a timeout in module load by the time I stepped through the code. Use of printk turned out to work better.