Debugging The Linux Kernel Using Gdb
- 1 Debugging the linux kernel using gdb
Debugging the linux kernel using gdb
The majority of day to day kernel debugging is done by adding print statements to code by using the famous printk function. This technique is well described in Kernel Debugging Tips. Using printk is a relatively simple, effective and cheap way to find problems. There are many other Linux grown techniques that take the debugging and profiling approach to a higher level. On this page we will discuss using the GNU debugger (GDB) to do kernel debugging. The GDB page describes some basic gdb command and also gives good links to documentation. Overall starting using gdb to do kernel debugging is relatively easy.
Most of the examples here will work in two (open source) situations. When using JTAG and when using QEMU system emulation. As the second option does not require any hardware you could go on and try it right away!
The open source JTAG debugging world is not that big. One project stands out in terms of debugging capabilities is OpenOCD and this is the tool used in this documentation. OpenOCD is pretty usable on the targets we tested ARM11 and ARM9.
You need to get yourself a GDB that is capable of understanding you target architecture. Often this come with you cross-compiler but if you have do compile it yourself you need to understand the difference between --target and --host configure options. GDB will be running on host(read x86) and will be able to understand target (read armv6). with that you might also want to have the gdbserver that can serve as stub for you user land debugging.
A JTAG Dongle:
To start debugging are kernel you will need to configure the kernel to have debug symbols. Once this is done you can do your normal kernel development. When needed you can "hook-up" your debugger. Start debugging a running kernel.
- start openocd
vmlinuz v.s zImage
When you want to debug the kernel you need a little understanding of how the kernel is composed. Most important is the difference between your vmlinux and the zImage. What you need to understand at this point is that the zImage is a container. This container gets loaded by a boot loader and that execution is handed over to the zImage. This zImage unpacks the kernel to the same memory location and starts executing the kernel. (explain that vmlinux does not have to be the real kernel as it is possible to debug a "stripped" kernel using a non stripped vmlinux). overall if we look at a compiled kernel we will see that vmlinux is located at the root of the kernel tree whiles the zImage is located under arch/arm/boot
vmlinux arch/arm/boot `-- zImage
vmlinux is what we will be using during debugging of the Linux kernel.
Debugging the kernel
The JTAG based debugging method described here is not intrusive. This means that besides debugging symbols you don't need to modify the kernel in any way. This is because we operate on the hardware, CPU core level. Overall this means that you can follow your normal development method. You can let your bootstrap and boot loader do their work and for example start debugging a running kernel. If your gdb-aware debugger is running it can be as simple as loading the vmlinuz and connecting to the remote target:
load vmlinuz target remote :3333
Loading a kernel in memory
Once you are used to using gdb to debug kernels you will want to use gdb to directly load kernels onto your target. The most practical way of doing this is to set a hardware breakpoint at the start of the kernel and reset your board using the JTAG reset signal. Your boot loader will initialize your board and the execution will stop at the start of the kernel. After that you can load a kernel into memory and run it.
Execute the following:
(gdb) file vmlinux (gdb) target remote :3333 (gdb) break __init_begin (gdb) cont (gdb) mon reset #perhaps this needs to be done from the openocd telnet session.. Breakpoint 1, 0xc0008000 in stext () (gdb) load vmlinux Loading section .text.head, size 0x240 lma 0xc0008000 Loading section .init, size 0xe4dc0 lma 0xc0008240 Loading section .text, size 0x219558 lma 0xc00ed000 Loading section .text.init, size 0x7c lma 0xc0306558 Loading section __ksymtab, size 0x4138 lma 0xc0307000 Loading section __ksymtab_gpl, size 0x1150 lma 0xc030b138 Loading section __kcrctab, size 0x209c lma 0xc030c288 Loading section __kcrctab_gpl, size 0x8a8 lma 0xc030e324 Loading section __ksymtab_strings, size 0xc040 lma 0xc030ebcc Loading section __param, size 0x2e4 lma 0xc031ac0c Loading section .data, size 0x1e76c lma 0xc031c000 Start address 0xc0008000, load size 3345456 Transfer rate: 64 KB/sec, 15632 bytes/write. (gdb) cont
This will boot your kernel that was loaded into memory via JTAG.
Getting the kernel log buffer
Sometimes the kernel will panic before the serial is up and running. In such situations is it very handy to be able to dump the kernel log buffer. This can be done by looking at the content of the __log_buf in the kernel. In gdb this can be done by issuing
p (char*) &__log_buf[log_start]
There must be a simple way of printing the memory area between log_start and log_end.
The problem is that gdb stops after the first line. Currently we use this routine that copied from wchar.gdb until something "normal" came out. We defined dmesg it like this:
define dmesg set $__log_buf = $arg0 set $log_start = $arg1 set $log_end = $arg2 set $x = $log_start echo " while ($x < $log_end) set $c = (char)(($__log_buf)[$x++]) printf "%c" , $c end echo "\n end document dmesg dmesg __log_buf log_start log_end Print the content of the kernel message buffer end
and call it like this:
dmesg __log_buf log_start log_end
Debugging a kernel module (.o and .ko )
Debugging a kernel module is harder.
Determining the module load address
gdb itself does not have knowledge about kernel modules and when debugging a kernel module. We will need to help gdb a little. One problem with modules is that it is not possible to determine where in the memory a module will be loaded before it is actually loaded so only once it is loaded we need to determine the address in memory it is loaded and tell gdb about it. There are many ways of determining this information. Here are 3 ways:
#gdb implementation of the linux lsmod define lsmod # The Linux kernel contains a generic double linked list implementation. # The "modules struct" is such a linked list and it contains information about modules. set $current = modules.next #The list implementation is done by adding a list_head struct to a container #containing the data that is to be "listed". If one know the offset in bytes #between the start of the struct and the "list_head" a simple formular can #be defined to determine the "content" of the list item. This value #is dependent on alignment and storage sized of the data in the struct #and is determined here for the struct module.. set $container_offset = ((int)&((struct module *)0).list) #Iterate over the list printing modules information while($current != modules.prev) printf "%s 0x%08x\n", \ ((struct module *) (((char*) ($current)) - $container_offset ) )->name ,\ ((struct module *) (((char*) ($current)) - $container_offset ) )->module_core set $current = $current.next end end
Loading Symbols of a loaded module
add-symbol-file drivers/mydrivers/mydriver.o 0xbf098000
Note that we use the .o file and not the .ko one. The address at the end is the address where the kernel decided to load the module
In the Linux Documentation directory under the kdump you will find file called gdbmacros.txt and it looks very promising as among other things it contains the a dmesg implementation
head linux-220.127.116.11/Documentation/kdump/gdbmacros.txt # # This file contains a few gdb macros (user defined commands) to extract # useful information from kernel crashdump (kdump) like stack traces of # all the processes or a particular process and trapinfo. #
A "find . -name "*gdb*" in the linux kernel directory also shows up a few interesting .gdbinit files that apparently can perform low level initialization.