Prof interprets files produced automatically by programs loaded
using the –p option of 8l(1) or other loader. The symbol table
in the named program file (8.out etc., according to $objtype,
by default) is read and correlated with the profile file (prof.out
by default). For each symbol, the percentage of time (in
seconds) spent executing between that symbol and the next is printed
(in decreasing order), together with the time spent there and
the number of times that routine was called.
Under option –d, prof prints the dynamic call graph of the target
program, annotating the calls with the time spent in each routine
and those it calls, recursively. The output is indented two spaces
for each call, and is formatted as|
where symbol is the entry point of the call, time is in milliseconds,
and ncall is the number of times that entry point was called at
that point in the call graph. If ncall is one, the /ncall is elided.
Normally recursive calls are compressed to keep the output brief;
option –r prints the full call graph.
The size of the buffer in program used to hold the profiling data,
by default 10000 entries, may be controlled by setting the environment
variable profsize before running program. If the buffer fills,
subsequent function calls may not be recorded.
The profiling code provided by the linker initializes itself to
profile the current pid, producing a file called prof.pid. If
a process forks, only the parent will continue to be profiled.
Forked children can cause themselves to be profile by calling
which causes the function fn(arg) to be profiled. When fn returns
prof.pid is produced for the current process pid.
The environment variable proftype can be set to one of user, kernel,
elapsed, or sample, to profile time measured spent in user mode,
time spent in user+kernel mode, or elapsed time, using the cycle
counter, or the time in user mode using the kernel's HZ clock.
The cycle counter is currently only available
on modern PCs and on the PowerPC. Default profiling measures user
time, using the cycle counter if it is available.
Tprof is similar to prof, but is intended for profiling multiprocess
programs. It uses the /proc/pid/profile file to collect instruction
frequency counts for the text image associated with the process,
for all processes that share that text. It must be run while the
program is still active, since the data is stored with
the running program. To enable tprof profiling for a given process,
prof(fn, arg, entries, what)|
and then, after the program has run for a while, execute
echo profile > /proc/pid/ctl|
Since the data collected for tprof is based on interrupt–time sampling
of the program counter, tprof has no –d or –r options.
Kprof is similar to prof, but presents the data accumulated by
the kernel profiling device, kprof(3). The symbol table file,
that of the operating system kernel, and the data file, typically
/dev/kpdata, must be provided. Kprof has no options and cannot
present dynamic data.