search

kosslab-kr / uftrace

Function (graph) tracer for user-space
https://uftrace.github.io/slide/
GNU General Public License v2.0
16 stars 0 forks source link

readme

Nightly test Discord

uftrace

uftrace is a function call graph tracer for C, C++, Rust and Python programs.

It hooks into the entry and exit of each function, recording timestamps as well as the function's arguments and return values. uftrace is capable of tracing both user and kernel functions, as well as library functions and system events providing an integrated execution flow in a single timeline.

Initially, uftrace only supported function tracing with compiler support. However, it now allows users to trace function calls without recompilation by analyzing instructions in each function prologue and dynamically and selectively patching those instructions.

Users can also write and run scripts for each function entry and exit using python/luajit APIs to create custom tools for their specific purposes.

uftrace offers various filters to reduce the amount of trace data and provides visualization using Chrome trace viewer and flame graph or call-graph diagrams for graphviz and mermaid, allowing for a big picture view of the execution flow.

It was heavily inspired by the ftrace framework of the Linux kernel and the name uftrace stems from the combination of user and ftrace.

It can record data from:

With the recorded data, uftrace can:

It supports many commands and filters such as filtering by function call duration for analysis of program execution and performance.

uftrace-live-demo

Features

uftrace traces each function in the executable and shows time durations.

Usually, for this to be possible, the program needs to be compiled with -pg or -fpatchable-function-entry=5 (=2 is enough on aarch64). With full dynamic tracing (-P.|--patch=.), uftrace works on all executables (as long they are not stripped, or symbol information is available from a separate file).

uftrace hooks into the PLT in the given executable file to trace library calls and with (-l|--nest-libcall), it also hooks into the procedure linkage tables (PLTs) of shared libraries. The depth can be limited using -D<num>, where 1 is flat call tracing.

Using (-a|--auto-args), uftrace automatically records arguments and return values of known functions. Without extra debug information, this includes the API functions of standard (C language or system) libraries. This can be combined with -P. or -l: For example, -la traces nested library calls, even in stripped executables.

In addition, -a implies --srcline so it records the source line location info and it can be shown by uftrace replay --srcline and in uftrace tui. Users can directly open the editor at the source location as shown in https://uftrace.github.io/slide/#120.

If debug information for the program (gcc -g) is available, --auto-args works even on functions inside of the compiled the user programs.

In case argument information is not available, argument specifiations like (-A udev_new@arg1/s) can be passed on the command line or an options file.

Example:

$ uftrace record -la -A udev_new@arg1/s lsusb >/dev/null
$ uftrace replay -f+module
or simply:
$ uftrace -la -A udev_new@arg1/s -f+module lsusb  # -f+module adds the module name
# DURATION     TID        MODULE NAME   FUNCTION
 306.339 us [ 23561]            lsusb | setlocale(LC_TYPE, "") = "en_US.UTF-8";
   1.163 us [ 23561]            lsusb | getopt_long(1, 0x7fff7175f6a8, "D:vtP:p:s:d:Vh") = -1;
            [ 23561]            lsusb | udev_new("POSIXLY_CORRECT") {
   0.406 us [ 23561] libudev.so.1.7.2 |   malloc(16) = 0x55e07277a7b0;
   2.620 us [ 23561]            lsusb | } /* udev_new */
            [ 23561]            lsusb | udev_hwdb_new() {
   0.427 us [ 23561] libudev.so.1.7.2 |   calloc(1, 200) = 0x55e07277a7d0;
   5.829 us [ 23561] libudev.so.1.7.2 |   fopen64("/etc/systemd/hwdb/hwdb.bin", "re") = 0;

Furthermore, it can show detailed execution flow at function level, and report which functions had the longest execution time. It also shows information about the execution environment.

You can setup filters to exclude or include specific functions when tracing. In addition, function arguments and return values can be saved and shown later.

It supports multi-process and/or multi-threaded applications. With root privileges and if the kernel was built with CONFIG_FUNCTION_GRAPH_TRACER=y, kernel functions can be traced as well.

How to build and install uftrace

On Linux distros, misc/install-deps.sh can be used to install required software(s) for building uftrace. Those are for optional and advanced features, but are highly recommended.

$ sudo misc/install-deps.sh

Once you installed required software(s) on your system, it can be built and installed like following:

$ ./configure
$ make
$ sudo make install

For details about installation and dependencies, please refer to INSTALL.md

How to use uftrace

These are the commands supported by uftrace:

You can use -h or --help option to see available commands and options.

$ uftrace
uftrace -- function (graph) tracer for userspace

 usage: uftrace [COMMAND] [OPTION...] [<program>]

 COMMAND:
   record          Run a program and saves the trace data
   replay          Show program execution in the trace data
   report          Show performance statistics in the trace data
   live            Do record and replay in a row (default)
   info            Show system and program info in the trace data
   dump            Show low-level trace data
   recv            Save the trace data from network
   graph           Show function call graph in the trace data
   script          Run a script for recorded trace data
   tui             Show text user interface for graph and report

Try `uftrace --help' or `man uftrace [COMMAND]' for more information.

If omitted, it defaults to the live command which is almost same as running record and replay subcommand in a row (but does not record the trace info to files).

For recording, the executable needs to be compiled with the -pg (or -finstrument-functions) option which generates profiling code (calling mcount or __cyg_profile_func_enter/exit) for each function.

Note that, there's an experimental support for dynamic tracing on x86_64 and AArch64(ARM64) which doesn't require such (re-)compilations. Also recent compilers have some options to help uftrace to reduce tracing overhead with similar way (although it still needs recompilation of your program). Please see dynamic tracing section for more details.

$ uftrace tests/t-abc
# DURATION    TID     FUNCTION
  16.134 us [ 1892] | __monstartup();
 223.736 us [ 1892] | __cxa_atexit();
            [ 1892] | main() {
            [ 1892] |   a() {
            [ 1892] |     b() {
            [ 1892] |       c() {
   2.579 us [ 1892] |         getpid();
   3.739 us [ 1892] |       } /* c */
   4.376 us [ 1892] |     } /* b */
   4.962 us [ 1892] |   } /* a */
   5.769 us [ 1892] | } /* main */

For more analysis, you'd be better recording it first so that it can run analysis commands like replay, report, graph, dump and/or info multiple times.

$ uftrace record tests/t-abc

It'll create uftrace.data directory that contains trace data files. Other analysis commands expect the directory exists in the current directory, but one can use another using -d option.

The replay command shows execution information like above. As you can see, the t-abc is a very simple program merely calls a, b and c functions. In the c function it called getpid() which is a library function implemented in the C library (glibc) on normal systems - the same goes to __cxa_atexit().

Users can use various filter options to limit functions it records/prints. The depth filter (-D option) is to omit functions under the given call depth. The time filter (-t option) is to omit functions running less than the given time. And the function filters (-F and -N options) are to show/hide functions under the given function.

The -k option enables to trace kernel functions as well (needs root access). With the classic hello world program, the output would look like below (Note, I changed it to use fprintf() with stderr rather than the plain printf() to make it invoke system call directly):

$ sudo uftrace -k tests/t-hello
Hello world
# DURATION    TID     FUNCTION
   1.365 us [21901] | __monstartup();
   0.951 us [21901] | __cxa_atexit();
            [21901] | main() {
            [21901] |   fprintf() {
   3.569 us [21901] |     __do_page_fault();
  10.127 us [21901] |     sys_write();
  20.103 us [21901] |   } /* fprintf */
  21.286 us [21901] | } /* main */

You can see the page fault handler and the write syscall handler were called inside the fprintf() call.

Also it can record and show function arguments and return value with -A and -R options respectively. The following example records first argument and return value of 'fib' (fibonacci number) function.

$ uftrace record -A fib@arg1 -R fib@retval tests/t-fibonacci 5

$ uftrace replay
# DURATION    TID     FUNCTION
   2.853 us [22080] | __monstartup();
   2.194 us [22080] | __cxa_atexit();
            [22080] | main() {
   2.706 us [22080] |   atoi();
            [22080] |   fib(5) {
            [22080] |     fib(4) {
            [22080] |       fib(3) {
   7.473 us [22080] |         fib(2) = 1;
   0.419 us [22080] |         fib(1) = 1;
  11.452 us [22080] |       } = 2; /* fib */
   0.460 us [22080] |       fib(2) = 1;
  13.823 us [22080] |     } = 3; /* fib */
            [22080] |     fib(3) {
   0.424 us [22080] |       fib(2) = 1;
   0.437 us [22080] |       fib(1) = 1;
   2.860 us [22080] |     } = 2; /* fib */
  19.600 us [22080] |   } = 5; /* fib */
  25.024 us [22080] | } /* main */

The report command lets you know which function spends the longest time including its children (total time).

$ uftrace report
  Total time   Self time       Calls  Function
  ==========  ==========  ==========  ====================================
   25.024 us    2.718 us           1  main
   19.600 us   19.600 us           9  fib
    2.853 us    2.853 us           1  __monstartup
    2.706 us    2.706 us           1  atoi
    2.194 us    2.194 us           1  __cxa_atexit

The graph command shows function call graph of given function. In the above example, function graph of function 'main' looks like below:

$ uftrace graph  main
# Function Call Graph for 'main' (session: 073f1e84aa8b09d3)
=============== BACKTRACE ===============
 backtrace #0: hit 1, time  25.024 us
   [0] main (0x40066b)

========== FUNCTION CALL GRAPH ==========
  25.024 us : (1) main
   2.706 us :  +-(1) atoi
            :  |
  19.600 us :  +-(1) fib
  16.683 us :    (2) fib
  12.773 us :    (4) fib
   7.892 us :    (2) fib

The dump command shows raw output of each trace record. You can see the result in the chrome browser, once the data is processed with uftrace dump --chrome. Below is a trace of clang (LLVM) compiling a small C++ template metaprogram.

uftrace-chrome-dump

It also supports flame-graph output as well. The data can be processed with uftrace dump --flame-graph and passed to flamegraph.pl. Below is a flame graph result of gcc compiling a simple C program.

uftrace-flame-graph-dump

The info command shows system and program information when recorded.

$ uftrace info
# system information
# ==================
# program version     : uftrace v0.8.1
# recorded on         : Tue May 24 11:21:59 2016
# cmdline             : uftrace record tests/t-abc
# cpu info            : Intel(R) Core(TM) i7-3930K CPU @ 3.20GHz
# number of cpus      : 12 / 12 (online / possible)
# memory info         : 20.1 / 23.5 GB (free / total)
# system load         : 0.00 / 0.06 / 0.06 (1 / 5 / 15 min)
# kernel version      : Linux 4.5.4-1-ARCH
# hostname            : sejong
# distro              : "Arch Linux"
#
# process information
# ===================
# number of tasks     : 1
# task list           : 5098
# exe image           : /home/namhyung/project/uftrace/tests/t-abc
# build id            : a3c50d25f7dd98dab68e94ef0f215edb06e98434
# exit status         : exited with code: 0
# elapsed time        : 0.003219479 sec
# cpu time            : 0.000 / 0.003 sec (sys / user)
# context switch      : 1 / 1 (voluntary / involuntary)
# max rss             : 3072 KB
# page fault          : 0 / 172 (major / minor)
# disk iops           : 0 / 24 (read / write)

The script command allows user to run a custom script on a data recorded. The supported script types are Python 3, Python 2.7 and Lua 5.1 as of now.

The tui command is for interactive text-based user interface using ncurses. It provides basic functionality of graph, report and info commands as of now.

Limitations

License

The uftrace program is released under GPL v2. See COPYING file for details.