Kernel Debugging Tips

From eLinux.org
Revision as of 18:34, 13 August 2008 by Tim Bird (Talk | contribs)

Jump to: navigation, search

Here are some miscellaneous tips for debugging a kernel:

Using printks

To add your own debug message to the kernel, you can place a "printk()" in the kernel code.

These messages are logged to an internal buffer, and eventually displayed by the kernel, being emitted to whatever is the active console, at the time of display. Since the console is not initialized until part-way through the boot sequence, several messages build up in the log until the console is initialize, at which point all buffered messages are displayed at once.

After the messages are displayed on the kernel console, they are left in the log buffer (which is a ring buffer), and they can be (and usually are) retrieved later by various commands.

printk is a variable-argument function, just like printf. Note that there is no "printf()" function for the kernel. "printk()" is used to specifically remind kernel developers that the environment is different.

printk can use many standard C-style printf format characters, but not all of them. See the kernel source for examples of what you can do.

Log levels

Each kernel message can be pre-pended with a tag indicating the importance of the message. The message tags are shown in the file include/linux/kernel.h. The available tags are: KERN_EMERG, KERN_ALERT, KERN_CRIT, KERN_WARNING, KERN_NOTICE, KERN_INFO, KERN_DEBUG.

Note that these tags are defined as simple strings, with angle brackets and a number:

#define      KERN_INFO       "<6>"

You could place the tag onto front of the string without using the #define, but that is considered very bad form, and is discouraged.

The way the tag is placed on the string is with C string concatenation:

printk(KERN_DEBUG "This is a debug message\n");

Note that there is no comma between the tag macro and the message string.

Adding timing information

Sometimes, it is useful to add timing information to the printk values, so you can see when a particular event occurred. The kernel includes an feature for doing this called printk times.

See the help for CONFIG_PRINTK_TIMES in the file lib/Kconfig.debug for more information on this feature. This option is found on the "Kernel Hacking" menu when configuring the kernel.

The timestamps which are inserted into the printk output consist of seconds and microseconds, as absolute values from the start of machine operation (or from the start of kernel timekeeping).

There is also tool in the kernel source which will convert the timestamp values to relative values (so you can see the interval between events). This tools is called show_delta and is located in the kernel 'scripts' directory.

See Printk Times for more information.

Viewing log messages

The klogd program will extract messages from the kernel log buffer, and send them to the system log (which winds up in /var/log/messages on most systems). This command runs in the background on most desktop or server systems, and continually transfers messages from the kernel log buffer to the system log.

You can view the contents of the log buffer directly, using the dmesg command. Note that by default dmesg displays the messages from the buffer, bug does not remove them. So this command can be run multiple times to view the kernel printk messages.

Controlling console output

In order to have the kernel boot be less "noisy", or in order to boot more quickly, it is sometimes useful to control the amount of messages displayed to the console during boot. You can do this by setting the kernel log level at boot time via a kernel command line option. See the "loglevel=" argument in Documentation/kernel-parameters.txt.

You can turn off all messages using the kernel command line option "quiet". See Disable Console for information on how much time this save at boot up.

Note that even if the log level is changed, or "quiet" is used, although the printk messages are not print to console, they are still entered into the log buffer, and they can still be extracted and displayed later using the dmesg command.

=== Changing the size of the printk buffer (referred to in the kernel source ss the "log buffer") You can make the log buffer larger when you build the kernel. This is useful for when additional printk data (such as extra messages inserted for debugging) is produced by the kernel and it overflows the regular buffer. The log buffer is a ring buffer, so later messages will overwrite the earliest messages in the buffer if too much data is written before the log buffer is drained (by, e.g. klogd or dmesg)

Step 1: Change the kernel log buffer size:

On the General Setup menu, change "Kernel log buffer size" to a larger number. This corresponds to LOG_BUF_SHIFT, which is the power-of-two size of the log buffer. By default, this is 14, which yeilds a log buffer size of 16KB. Changing this to 17 will yield a log buffer size of 128KB.

Set the option, then re-build and install the kernel for booting on the target board.

Step 2: Tell dmesg to retrieve the larger buffer: Even though the buffer is larger in the kernel, dmesg will still default to only reading 16K, unless told otherwise. To read the full buffer, use the '-s' option like so: "dmesg -s 128000"

Debugging early boot problems

Accessing the printk buffer after a silent hang on boot

Sometimes, if the kernel hangs early in the boot process, you get no messages on the console before the hang. However, there may still be messages in the printk buffer, which can give you an idea of where the problem is.

The kernel starts putting messages into the printk buffer as soon as it starts. They stay buffered there until the console code has a chance to initialize the console device (often the serial port for embedded devices). Eve though these messages are no printed before the hang, it is still possible in some circumstances to dump the printk buffer and see the messages.

Quinn Jensen writes:

Something I've found handy when the console is silent is to dump the printk buffer from the boot loader. To do it you have to know how your boot loader maps memory compared to the kernel.

Redboot example on a Freescale ADS board

Quinn says: Here's what I do with Redboot on i.MX31:

fgrep printk_buf System.map

this shows the linked address of the printk_buf, e.g.:

c02338f0 b printk_buf.16194

The address "c02338f0" is in kernel virtual, which, in the case of i.MX31 ADS, redboot will have mapped to 0x802338f0. So, after resetting the target board, but without letting it try to boot again, at the redboot prompt,

dump -b 0x802338f0 -l 10000

And you see the printk buffer that never got flushed to the UART. Knowing what's there can be very useful in debugging your console.

U-boot example on an OMAP OSK board

Tim Bird tried these steps and they worked:

grep __log_buf System.map

or

grep __log_buf /proc/kallsyms

These show:

c0352d88 B __log_buf

In the case of the OSK, this address maps to 0x10352d88. So I reset the target board and use the following:

OMAP5912 OSK # md 10352d88
10352d88: 4c3e353c 78756e69 72657620 6e6f6973    <5>Linux version
10352d98: 362e3220 2e32322e 612d3631 6e5f706c     2.6.22.16-alp_n
10352da8: 7428206c 64726962 6d697440 6b736564    l (tbird@timdesk
10352db8: 2e6d612e 796e6f73 6d6f632e 67282029    .am.sony.com) (g
10352dc8: 76206363 69737265 33206e6f 342e342e    cc version 3.4.4
10352dd8: 34232029 45525020 54504d45 65755420    ) #4 PREEMPT Tue
...

Using CONFIG_DEBUG_LL and printascii() (ARM only)

If the kernel fails before the serial console is enabled, you can use CONFIG_DEBUG_LL to add extra debug output routines to the kernel. These are printascii, printch and printhex. These routines print directly to the serial port, bypassing the console code, and are available earlier in machine initialization.

Here is an e-mail exchange seen on the linux-embedded mailing list (with answer by George Davis):

> When we try to boot a 2.6.21 kernel after uncompressing the kernel the boot process dies somehow.
> We've figured out so far that the kernel dies somewhere between  the gunzip and start_kernel.

Try enabling DEBUG_LL to see if it's an machine ID error.  If you see:

Error: unrecognized/unsupported processor variant.

or:

Error: unrecognized/unsupported machine ID...

Then you either don't have proper processor support enabled for your target
or your bootloader is passing in the wrong machine number.

If you still don't see anything, try hacking printk.c to call
printascii() (enabled for the DEBUG_LL case) to print directly to the
serial port w/o a driver, etc.,.  You can find more details on these
low-level debugging hacks via a little googling...

Triggering a kernel event

overloading the sync system call

Sometimes, it is nice to trigger an event to happen in the kernel from user space. Instead of creating infrastructure to handle a /proc event, an ioctl() or making a new syscall, it can be quick and easy to just overload an existing function. One function not used very often is sync. (I have found that the sync system call is not normally called by user space programs (or during standard linux booting).

It is quite easy to put a hook to your own kernel program in the sys_sync() routine (located in fs/sync.c) and cause it to execute by issuing 'sync' from the shell command line. This is handy as a temporary mechanism to test things that you have put in the kernel.