As many kernel developers and hackers have known for years, loadable/unloadable kernel modules (like user-space applications) are almost never bug-free. With the continuing use and development of loadable modules growing, due in fact to the obvious benefits of the mechanism (lean kernels, reduction of kernel recompiles/reboots, etc.), developers are in an increasing need for robust debugging tools capable of aiding in the identification of problem code. Traditionally, module developers have used various debugging techniques to help identify problematic code. These techniques have included:

kgdb is a kernel patch that, once applied, allows for the use of the familiar gdb interface for source-level debugging of a running Linux kernel. The process requires the use of two machines. One machine runs the kernel being debugged while the other runs the gdb session. Communication between the running kernel and gdb transpires via a serial cable connecting the two machines.

The kgdb patch supplies the kernel with a debugging stub. This stub uses the gdb remote serial protocol to communicate with gdb through a serial driver interface (also supplied by the patch). This patch is applied to the kernel on the machine that will run the gdb session (the development machine) where it is recompiled. The newly compiled boot image is then copied to the other machine (the target machine) where it is configured as the bootable kernel. When a reboot into the transferred kernel is complete, the target machine can then be configured to halt and await a remote connection from a gdb session on the development machine. When this connection is established, the target machine's kernel can then be debugged (single-stepping, issuing of breakpoints, data examination, etc.) through gdb on the development machine as if it were a user-space application.

The first step is to download the kgdb patch for your kernel version. A patch can be obtained at http://kgdb.sourceforge.net/. As of this writing, patches only exist for the following kernels:

Once you have obtained the patch, copy the patch to the kernel source directory on the development machine, and apply.

patch  -p1 < patchfile

The patched kernel must now be recompiled. It is assumed here that /usr/src/linux/.config exists and accurately reflects your current kernel configuration. Navigate to the source directory (if you aren't there already), and do a make menuconfig. From the main menu navigate to and select kernel hacking. You should now see an option for Remote (serial) debugging with gdb. Make sure this option is selected and then exit, saving your configuration. Next, do a make clean followed by a make bzImage (or whatever image you usually make).

The recompile adds a documentation file called gdb-serial.txt to your system. This file can be found in /usr/src/linux/Documentation/i386 and includes a step-by-step description of what needs to transpire next. Basically, here are the highlights.

The newly compiled kernel image (e.g., bzImage) is copied to the target machine where it is configured for boot. For example, the image may be copied to /boot/vmlinuz-target (or whatever you want to call it) followed by an added entry in lilo.conf:

image= /boot/vmlinuz-target
     label=target_kernel
     read-only
     root=/dev/hda1

Next, run LILO at the command line, and reboot into the new kernel.

gdbstart  - s 38400 - t /dev/ttyS0

Alternatively, the documentation suggests creating a script on the target machine to deliberately call gdbstart with user-defined parameters.

Listing 1. gdb Connecting to a Remote Target

You can now issue step commands, set breakpoints, etc. Issuing a continue to gdb will return the target kernel to a running state. The kernel will continue to run until it encounters a defined breakpoint, an interrupt, a signal, a segment violation, etc., at which point control is returned to gdb on the development machine.

In order for us to use kgdb to debug loadable modules, we must understand how the remote kernel communicates with gdb on the development machine. Remember, the mechanism's default communication between gdb on the development machine and the debugging stub on the target machine transpires via a serial driver interface (called gdbserial.o). In order for the debugging process to begin, two things must happen from within this driver. First, the set_debug_traps( ) is initiated. This function is defined in the debugging stub and informs the remote kernel that all breakpoints, error conditions and other exception handling is to be intercepted and handled by gdb. Secondly, the serial driver must call the function breakpoint( ). This function is also defined in the debugging stub and is used to initiate the communication by issuing a breakpoint interrupt:

asm( " int $3");

Since gdb is now configured to intercept such a condition (i.e., the set_debug_traps( ) call), the kernel on the target machine halts and transfers control to gdb on the development machine. It is from this point that the user may begin normal debugging such as single-stepping, issuing of breakpoints, stack tracing, etc. However, if we were to begin stepping from this initial point, the code we would be examining would be in gdbserial immediately following the call to the debugging stub's breakpoint( ) (since this is where program execution has halted).

For example, Listing 2 is the excerpt of gdbserial in which the two calls to the previously explained stub functions are called.

Listing 2. An Excerpt of gdbserial

Now, if we initiate another debug session and issue a step command after gdb receives the stub's breakpoint interrupt, we would step to the next line of code in gdbserial after the breakpoint( ) is made (which should be gdb_null in the example in Listing 3).

Listing 3. Sample gdb Dialogue with Breakingpoints

As was mentioned before, gdb can return the remote kernel to a running state by issuing a continue command. If this is done, gdb patiently waits until the remote kernel returns control by issuing some sort of exception (such as a user-defined breakpoint, segmentation fault, etc.).

For a better understanding of how the entire process works review the debugging stub code found in /usr/src/linux/arch/i386/kernel/gdbstub.c and the serial driver interface, which can be found in /usr/src/linux/drivers/char/gdbserial.c.

We now have almost all the information we need in order begin using the kgdb mechanism to debug loadable modules. Remember, the important information we must retain from kgdb's communication process in order to initiate module debugging is: 1) at the beginning of the debug session, the debugging stub informs gdb on the development machine that it is responsible for intercepting and handling all exceptions from the remote target kernel for example, gdbserial's initial call to set_debug"traps( ); 2) the initial debug process begins with the serial interface's call to the debugging stub's breakpoint( ) function; 3) gdb can return the remote kernel to a running state but will regain control once the remote kernel issues any type of exception.

Additionally, we must consider that because the module will be loaded on the target machine and the gdb session runs on another machine, gdb will have no idea where in the target machine's memory the module code will be loaded. We must therefore determine this location and inform gdb of its whereabouts before the debugging of the module can begin.

During the kernel-building process, the kernel produces a file that maps addresses in memory to function names for the modules/drivers that are built during the compile process. This file is usually placed in the root of the kernel source directory (i.e., /usr/src/linux/System.map) and is used by the kernel to access those compiled devices properly in the correct memory location. However, at the construction of that file the kernel is unaware of where in memory a particular module may be loaded at a later time.

Fortunately, we can determine the memory location for modularized code during its load process. This is accomplished by using insmod with the -m parameter that informs insmod to produce a load map. This map informs us of where in memory the object-code sections reside. Ultimately, we must locate this information in order to inform gdb on our development machine of where our module's object code resides on the target machine. To illustrate, let's consider the module code shown in Listing 4, which we will refer to as the simple module.

Listing 4. Simple Module

As you can see, the simple module contains just enough functionality to to identify its memory address on the target machine when loaded.

gcc -c -O2 -g simple.c

Be sure to include the -g option during compilation in order to enable debugging symbols for gdb.

insmod -m simple.o

Listing 5. Memory Map of Kernel Module Size and Location

rmmod simple

We are almost ready to begin debugging a module. But before we proceed, we have to alleviate another possible issue. In gdb we will be using the add-symbol-file to inform the debugger of the memory location for our module's object code on the target machine. However, gdb 5.0 and previous versions have had problems in correctly calculating addresses using the add-symbol-file command (the problem surrounds the issue of a module's global variables). The problem has been corrected in developmental versions of gdb. It is therefore recommended that you use a developmental version of gdb to debug modules with kgdb. The gdb version used for the remainder of this article is a developmental gdb built for Red Hat 6.2. For more information regarding this issue, visit http://kgdb.sourceforge.net/.

The last step we must account for is the configuration of our module to communicate with the remote gdb session. This is a relatively simple task. We now know that when gdb on the development machine makes the initial contact with the target machine, the gdbserial interface issues a call to the debugging stub's set_debug_traps( ) function. This function, as you recall, instructs gdb to perform all exception handling for the target kernel. The serial interface then issues a call to the stub's breakpoint( ) function, which turns control over to gdb. At this point we can inform the remote kernel to resume normal operations by issuing a continue command from gdb.

With the target kernel now configured to return control to the remote gdb session whenever an exception is triggered, we can modify our module code to guarantee that such an event will arise. Consider the following modifications made to the simple module in Listing 6.

Listing 6. Modified Sample Module to Include Further Remote Control

As you can see by the code in Listing 6, we have added a breakpoint interrupt that is to be called (in this example) when the module is both loaded and unloaded from the kernel. These interrupts will return control to the remote gdb session, thus halting execution of the module at those points. Let's try it out.

gcc -c -O2 -g simple.c

add-symbol-file /root/simple.o 0xc480004c

You may be wondering at this point how we can assign this address space when the module is not currently loaded. Or a better question may be, ``How do we know that the kernel will load the module into the exact same location in memory?'' We can make this assumption because the kernel very often will load the object code into the same memory segment as before. While this is not written in stone it does happen frequently enough to render this method quite reliable.

insmod simple.o

Listing 7. User Space gdb Dialogue Running Modified Simple Module

Note that this sequence reflects an example of stopping the module's load process. The module could be installed with the insmod -m parameter again to verify memory placement of the object code if other module functionality (other than the initialization process) was to be debugged (e.g., file operation functions: open, read, write, close, ioctl; driver resource allocation/deallocation, etc.).

rmmod simple

This of course issues the cleanup_module function, which in turn invokes another of our breakpoints, returning control to gdb on the development machine:

(gdb) c
Continuing.
Program received signal SIGTRAP, Trace/breakpoint trap.
0xc4800068 in cleanup_module () at simple.c:38
38      } // end function init_module
(gdb)

Thanks go to all those contributors at http://kgdb.sourceforge.net/ and http://oss.sgi.com/ for their continuing work on kgdb.

A special thanks goes to kgdb's original author, David Grothe, and to Amit S. Kale for his outstanding ongoing work and enhancements to the system. Thanks guys.

Resources