readprofile - a tool to read kernel profiling information

readprofile [ options ]

This manpage documents version 2.0 of the program.

The readprofile command uses the /proc/profile information to print ascii data on standard output. The output is organized in three columns: the first is the number of clock ticks, the second is the name of the C function in the kernel where those many ticks occurred, and the third is the normalized `load' of the procedure, calculated as a ratio between the number of ticks and the length of the procedure. The output is filled with blanks to ease readability.

Available command line options are the following:

-m mapfile

Specify a mapfile, which by default is /usr/src/linux/System.map. You should specify the map file on cmdline if your current kernel isn't the last one you compiled, or if you keep System.map elsewhere. If the name of the map file ends with `.gz' it is decompressed on the fly.

-p pro-file

Specify a different profiling buffer, which by default is /proc/profile. Using a different pro-file is useful if you want to `freeze' the kernel profiling at some time and read it later. The /proc/profile file can be copied using `cat' or `cp'. There is no more support for compressed profile buffers, like in readprofile-1.1, because the program needs to know the size of the buffer in advance.

-i

Info. This makes readprofile only print the profiling step used by the kernel. The profiling step is the resolution of the profiling buffer, and is chosen during kernel configuration (through `make config'), or in the kernel's command line. If the -t (terse) switch is used together with -i only the decimal number is printed.

-a

Print all symbols in the mapfile. By default the procedures with 0 reported ticks are not printed.

-b

Print individual histogram-bin counts.

-r

Reset the profiling buffer. This can only be invoked by root, because /proc/profile is readable by everybody but writable only by the superuser. However, you can make readprofile setuid 0, in order to reset the buffer without gaining privileges.

-M multiplier

On some architectures it is possible to alter the frequency at which the kernel delivers profiling interrupts to each CPU. This option allows you to set the frequency, as a multiplier of the system clock frequency, HZ. This is supported on i386-SMP (2.2 and 2.4 kernel) and also on sparc-SMP and sparc64-SMP (2.4 kernel). This option also resets the profiling buffer, and requires superuser privileges.

-v

Verbose. The output is organized in four columns and filled with blanks. The first column is the RAM address of a kernel function, the second is the name of the function, the third is the number of clock ticks and the last is the normalized load.

-V

Version. This makes readprofile print its version number and exit.

Browse the profiling buffer ordering by clock ticks:

   readprofile | sort -nr | less

   readprofile | sort -nr +2 | head -20

   readprofile | grep _ext2

   readprofile -av | less

   readprofile -p ~/profile.freeze -m /zImage.map.gz

   sudo readprofile -M 20

readprofile only works with an 1.3.x or newer kernel, because /proc/profile changed in the step from 1.2 to 1.3

This program only works with ELF kernels. The change for a.out kernels is trivial, and left as an exercise to the a.out user.

To enable profiling, the kernel must be rebooted, because no profiling module is available, and it wouldn't be easy to build. To enable profiling, you can specify "profile=2" (or another number) on the kernel commandline. The number you specify is the two-exponent used as profiling step.

Profiling is disabled when interrupts are inhibited. This means that many profiling ticks happen when interrupts are re-enabled. Watch out for misleading information.

/proc/profile              A binary snapshot of the profiling buffer.
/usr/src/linux/System.map  The symbol table for the kernel.
/usr/src/linux/*           The program being profiled :-)