% UFTRACE-RECORD(1) Uftrace User Manuals
% Namhyung Kim <namhyung@gmail.com>
% May, 2016

NAME
====
uftrace-record - Run a command and record its trace data

SYNOPSIS
========
uftrace record [*options*] COMMAND [*command-options*]

DESCRIPTION
===========
This command runs COMMAND and gathers function trace data from it, and saves it into files under the uftrace data directory - without displaying anything.

This data can then be inspected later on, using `uftrace replay` or `uftrace report`.

OPTIONS
=======
-b *SIZE*, \--buffer=*SIZE*
:   Size of internal buffer in which trace data will be saved.  Default size is 128k.

-F *FUNC*, \--filter=*FUNC*
:   Set filter to trace selected functions only.  This option can be used more than once.  See *FILTERS*.

-N *FUNC*, \--notrace=*FUNC*
:   Set filter not to trace selected functions (or the functions called underneath them).  This option can be used more than once.  See *FILTERS*.

-T *TRG*, \--trigger=*TRG*
:   Set trigger on selected functions.  This option can be used more than once.  See *TRIGGERS*.

-t *TIME*, \--time-filter=*TIME*
:   Do not show functions which run under the time threshold.  If some functions explicitly have the 'trace' trigger applied, those are always traced regardless of execution time.

\--force
:   Allow running uftrace even if some problems occur.  When `uftrace record` finds no mcount symbol (which is generated by compiler) in the executable, it quits with an error message since uftrace can not trace the program.  However, it is possible that the user is only interested in functions within a dynamically-linked library, in which case this option can be used to cause uftrace to run the program regardless.  Also, the `-A`/`--argument` and `-R`/`--retval` options work only for binaries built with `-pg`, so uftrace will normally exit when it tries to run binaries built without that option.  This option ignores the warning and goes on tracing without the argument and/or return value.

-L *PATH*, \--library-path=*PATH*
:   Load necessary internal libraries from this path.  This is for testing purposes.

\--no-libcall
:   Do not record library function invocations.  Library calls are normally traced by hooking the dynamic linker's resolve function in the PLT.  One can disable it with this option.

\--no-pltbind
:   Do not bind dynamic symbol address.  This option uses the `LD_BIND_NOT` environment variable to trace library function calls which might be missing due to concurrent (first) accesses.  It is not meaningful to use this option with the `--no-libcall` option.

\--nest-libcall
:   Trace function calls between libraries.  By default, uftrace only record library call from the main executable.  Implies `--force`.

-D *DEPTH*, \--depth=*DEPTH*
:   Set global trace limit in nesting level.

\--max-stack=*DEPTH*
:   Set the max function stack depth for tracing.  Default is 1024.

\--nop
:   Do not record any functions.  This is a no-op and only meaningful for performance comparisons.

\--time
:   Print running time of children in `time`(1)-style.

-k, \--kernel
:   Trace kernel functions as well as user functions.  Only kernel entry/exit functions will be traced by default.  Use the `--kernel-depth` option to override this.

-H *HOST*, \--host=*HOST*
:   Send trace data to given host via the network, not writing to files.  The `uftrace recv` command should be run on the destination host to receive the data.

\--port=*PORT*
:   When sending data to the network (with `-H`), use the given port instead of the default (8090).

\--disable
:   Start uftrace with tracing disabled.  This is only meaningful when used with a `trace_on` trigger.

-A *SPEC*, \--argument=*SPEC*
:   Record function arguments.  This option can be used more than once.  See *ARGUMENTS*.

-R *SPEC*, \--retval=*SPEC*
:   Record function return values.  This option can be used more than once.  See *ARGUMENTS*.

-a, \--auto-args
:   Automatically record arguments and return values of known functions.
    These are usually functions in standard (C language or system) libraries
    but if debug info is available it includes functions in the user program.

\--num-thread=*NUM*
:   Use NUM threads to record trace data.  Default is 1/4 of online CPUs (but when full kernel tracing is enabled, it will use the full number of CPUs).

\--libmcount-single
:   Use single thread version of libmcount for faster recording.  This is ignored if the target program calls `pthread_create()`.

\--rt-prio=*PRIO*
:   Boost priority of recording threads to real-time (FIFO) with priority of *PRIO*.  This is particularly useful for high-volume data such as full kernel tracing.

-K *DEPTH*, \--kernel-depth=*DEPTH*
:   Set kernel max function depth separately.  Implies `--kernel`.

\--kernel-buffer=*SIZE*
:   Set kernel tracing buffer size.  The default value (in the kernel) is 1408k.

-P *FUNC*, \--patch=*FUNC*
:   Patch FUNC dynamically.  This is only applicable binaries built by
    gcc with `-pg -mfentry -mnop-mcount` or clang with `-fxray-instrument`.
    This option can be used more than once.  See *DYNAMIC TRACING*.

-E *EVENT*, \--event=*EVENT*
:   Enable event tracing.  The event should be available on the system.

\--no-event
:   Disable event recording which is used by default.  Note that explicit event
    tracing by `--event` option is not affected by this.

\--keep-pid
:   Retain same pid for traced program.  For some daemon processes, it is important to have same pid when forked.  Running under uftrace normally changes pid as it calls fork() again internally.

-S *SCRIPT_PATH*, \--script=*SCRIPT_PATH*
:   Add a script to do additional work at the entry and exit of function.  The type of script is detected by the postfix such as '.py' for python.

--match=*TYPE*
:   Use pattern match using TYPE.  Possible types are `regex` and `glob`.  Default is `regex`.

--no-randomize-addr
:   Disable ASLR (Address Space Layout Randomization).  It makes the target process fix its address space layout.


FILTERS
=======
The uftrace tool supports filtering out uninteresting functions.  Filtering is highly recommended since it helps users focus on the interesting functions and reduces the data size.  When uftrace is called it receives two types of function filter; an opt-in filter with `-F`/`--filter` and an opt-out filter with `-N`/`--notrace`.  These filters can be applied either at record time or replay time.

The first one is an opt-in filter. By default, it doesn't trace anything. But when one of the specified functions is executed, tracing is started. When the function returns, tracing is stopped again.

For example, consider a simple program which calls `a()`, `b()` and `c()` in turn.

    $ cat abc.c
    void c(void) {
        /* do nothing */
    }

    void b(void) {
        c();
    }

    void a(void) {
        b();
    }

    int main(void) {
        a();
        return 0;
    }

    $ gcc -pg -o abc abc.c

Normally uftrace will trace all the functions from `main()` to `c()`.

    $ uftrace ./abc
    # DURATION    TID     FUNCTION
     138.494 us [ 1234] | __cxa_atexit();
                [ 1234] | main() {
                [ 1234] |   a() {
                [ 1234] |     b() {
       3.880 us [ 1234] |       c();
       5.475 us [ 1234] |     } /* b */
       6.448 us [ 1234] |   } /* a */
       8.631 us [ 1234] | } /* main */

But when the `-F b` filter option is used, it will not trace `main()` or `a()` but only `b()` and `c()`.

    $ uftrace record -F b ./abc
    $ uftrace replay
    # DURATION    TID     FUNCTION
                [ 1234] | b() {
       3.880 us [ 1234] |   c();
       5.475 us [ 1234] | } /* b */

The second type of filter is opt-out. By default, everything is traced, but when one of the specified functions is executed, tracing stops.  When the excluded function returns, tracing is started again.

In the above example, you can omit the function `b()` and all calls it makes with the `-N` option.

    $ uftrace record -N b ./abc
    $ uftrace replay
    # DURATION    TID     FUNCTION
     138.494 us [ 1234] | __cxa_atexit();
                [ 1234] | main() {
       6.448 us [ 1234] |   a();
       8.631 us [ 1234] | } /* main */

In addition, you can limit the print nesting level with the `-D` option.

    $ uftrace record -D 3 ./abc
    $ uftrace replay
    # DURATION    TID     FUNCTION
     138.494 us [ 1234] | __cxa_atexit();
                [ 1234] | main() {
                [ 1234] |   a() {
       5.475 us [ 1234] |     b();
       6.448 us [ 1234] |   } /* a */
       8.631 us [ 1234] | } /* main */

In the above example, uftrace only prints functions up to a depth of 3, so leaf function `c()` was omitted.  Note that the `-D` option works with `-F`.

Sometimes it's useful to see long-running functions only.  This is good because there are usually many tiny functions that are not interesting.  The `-t`/`--time-filter` option implements the time-based filter that only records functions which run longer than the given threshold.  In the above example, the user might want to see functions running more than 5 microseconds like below:

    $ uftrace record -t 5us ./abc
    $ uftrace replay
    # DURATION    TID     FUNCTION
     138.494 us [ 1234] | __cxa_atexit();
                [ 1234] | main() {
                [ 1234] |   a() {
       5.475 us [ 1234] |     b();
       6.448 us [ 1234] |   } /* a */
       8.631 us [ 1234] | } /* main */

The `-t`/`--time-filter` option works for user-level functions only.  It does not work for recording kernel functions, but they can be hidden in replay, report, dump and graph commands with `-t`/`--time-filter` option.

You can also set triggers on filtered functions.  See *TRIGGERS* section below for details.

When kernel function tracing is enabled, you can also set the filters on kernel functions by marking the symbol with the `@kernel` modifier.  The following example will show all user functions and the (kernel) page fault handler.

    $ sudo uftrace -k -F '.*page_fault@kernel' ./abc
    # DURATION    TID     FUNCTION
               [14721] | main() {
      7.713 us [14721] |   __do_page_fault();
      6.600 us [14721] |   __do_page_fault();
      6.544 us [14721] |   __do_page_fault();
               [14721] |   a() {
               [14721] |     b() {
               [14721] |       c() {
      0.860 us [14721] |         getpid();
      2.346 us [14721] |       } /* c */
      2.956 us [14721] |     } /* b */
      3.340 us [14721] |   } /* a */
     79.086 us [14721] | } /* main */


TRIGGERS
========
The uftrace tool supports triggering actions on selected function calls with or
without filters.  Currently supported triggers are listed below.
The BNF for trigger specification is:

    <trigger>    :=  <symbol> "@" <actions>
    <actions>    :=  <action>  | <action> "," <actions>
    <action>     :=  "depth="<num> | "trace" | "trace_on" | "trace_off" |
                     "time="<time_spec> | "read="<read_spec> | "finish" |
                     "filter" | "notrace" | "recover"
    <time_spec>  :=  <num> [ <time_unit> ]
    <time_unit>  :=  "ns" | "us" | "ms" | "s"
    <read_spec>  :=  "proc/statm" | "page-fault" | "pmu-cycle" | "pmu-cache" |
                     "pmu-branch"

The `depth` trigger is to change filter depth during execution of the function.
It can be used to apply different filter depths for different functions.

The following example shows how triggers work.  The global filter maximum depth
is 5, but when function `b()` is called, it is changed to 1, so functions below
`b()` will not shown.

    $ uftrace record -D 5 -T 'b@depth=1' ./abc
    $ uftrace replay
    # DURATION    TID     FUNCTION
     138.494 us [ 1234] | __cxa_atexit();
                [ 1234] | main() {
                [ 1234] |   a() {
       5.475 us [ 1234] |     b();
       6.448 us [ 1234] |   } /* a */
       8.631 us [ 1234] | } /* main */

The `backtrace` trigger is only meaningful in the replay command.

The `traceon` and `traceoff` actions (the `_` can be omitted from `trace_on`
and `trace_off`) control whether uftrace records the specified functions or not.

The 'recover' trigger is for some corner cases in which the process accesses the
callstack directly.  During tracing of the v8 javascript engine, for example, it
kept getting segfaults in the garbage collection stage.  It was because v8
incorporates the return address into compiled code objects(?).  The `recover`
trigger restores the original return address at the function entry point and
resets to the uftrace return hook address again at function exit.  I was managed
to work around the segfault by setting the `recover` trigger on the related
function (specifically `ExitFrame::Iterate`).

The 'time' trigger is to change time filter setting during execution of the
function.  It can be used to apply differernt time filter for different functions.

The `read` trigger is to read some information at runtime.  The result will be
recorded as (builtin) events at the beginning and the end of a given function.
As of now, following events are supported:

 * "proc/statm": process memory stat from /proc filesystem
 * "page-fault": number of page faults using getrusage(2)
 * "pmu-cycle":  cpu cycles and instructions using Linux perf-event syscall
 * "pmu-cache":  (cpu) cache-references and misses using Linux perf-event syscall
 * "pmu-branch": branch instructions and misses using Linux perf-event syscall

The results are printed in comments like below.

    $ uftrace record -T a@read=proc/statm ./abc
    $ uftrace replay
    # DURATION    TID     FUNCTION
                [ 1234] | main() {
                [ 1234] |   a() {
                [ 1234] |     /* read:proc/statm (size=6808KB, rss=776KB, shared=712KB) */
                [ 1234] |     b() {
                [ 1234] |       c() {
       1.448 us [ 1234] |         getpid();
      10.270 us [ 1234] |       } /* c */
      11.250 us [ 1234] |     } /* b */
                [ 1234] |     /* read2:proc/statm (size=+4KB, rss=+0KB, shared=+0KB) */
      18.380 us [ 1234] |   } /* a */
      19.537 us [ 1234] | } /* main */

The 'finish' trigger is to end recording.  The process still can run and this
can be useful to trace unterminated processes like daemon.

The 'filter' and 'notrace' triggers have same effect as `-F`/`--filter` and
`-N`/`--notrace` options respectively.

Triggers only work for user-level functions for now.


ARGUMENTS
=========
The uftrace tool supports recording function arguments and/or return values using the `-A`/`--argument` and `-R`/`--retval` options respectively.  The syntax is very similar to that of triggers:

    <argument>    :=  <symbol> "@" <specs>
    <specs>       :=  <spec> | <spec> "," <spec>
    <spec>        :=  ( <int_spec> | <float_spec> | <ret_spec> )
    <int_spec>    :=  "arg" N [ "/" <format> [ <size> ] ] [ "%" ( <reg> | <stack> ) ]
    <float_spec>  :=  "fparg" N [ "/" ( <size> | "80" ) ] [ "%" ( <reg> | <stack> ) ]
    <ret_spec>    :=  "retval" [ "/" <format> [ <size> ] ]
    <format>      :=  "d" | "i" | "u" | "x" | "s" | "c" | "f" | "S" | "p"
    <size>        :=  "8" | "16" | "32" | "64"
    <reg>         :=  <arch-specific register name>  # "rdi", "xmm0", "r0", ...
    <stack>       :=  "stack" [ "+" ] <offset>

The `-A`/`--argument` option takes argN where N is an index of the arguments.  The index starts from 1 and corresponds to the argument passing order of the calling convention on the system.  Note that the indexes of arguments are separately counted for integer (or pointer) and floating-point type, and they can interfere depending on the calling convention.  The argN is for integer arguments and fpargN is for floating-point arguments.

Users can optionally specify a format and size for the arguments and/or return values.  The "d" format or without format field, uftrace treats them as 'long int' type for integers and 'double' for floating-point numbers.  The "i" format makes it signed integer type and "u" format is for unsigned type.  Both are printed as decimal while "x" format makes it printed as hexadecimal.  The "s" format is for null-terminated string type and "c" format is for character type.  The "f" format is for floating-point type and is meaningful only for return value (generally).  Note that fpargN doesn't take the format field since it's always floating-point.  The "S" format is for std::string, but it only supports libstdc++ library as of yet.  Finally, the "p" format is for function pointer. Once the target address is recorded, it will be displayed as function name.

Please beware when using string type arguments since it can crash the program if the (pointer) value is invalid.

It is also possible to specify a certain register name or stack offset for arguments (but not for return value).  The following register names can be used for argument:

 * x86: rdi, rsi, rdx, rcx, r8, r9 (for integer), xmm[0-7] (for floating-point)
 * arm: r[0-3] (for integer), s[0-15] or d[0-7] (for floating-point)

Examples are below:

    $ uftrace record -A main@arg1/x -R main@retval/i32 ./abc
    $ uftrace replay
    # DURATION    TID     FUNCTION
     138.494 us [ 1234] | __cxa_atexit();
                [ 1234] | main(0x1) {
                [ 1234] |   a() {
                [ 1234] |     b() {
       3.880 us [ 1234] |       c();
       5.475 us [ 1234] |     } /* b */
       6.448 us [ 1234] |   } /* a */
       8.631 us [ 1234] | } = 0; /* main */

    $ uftrace record -A puts@arg1/s -R puts@retval ./hello
    Hello world

    $ uftrace replay
    # DURATION    TID     FUNCTION
       1.457 us [21534] | __monstartup();
       0.997 us [21534] | __cxa_atexit();
                [21534] | main() {
       7.226 us [21534] |   puts("Hello world") = 12;
       8.708 us [21534] | } /* main */

Note that these arguments and return value are recorded only if the executable
was built with the `-pg` option.  Executables built with `-finstrument-functions`
will ignore it except for library calls.  Recording of arguments and return
values only works with user-level functions for now.

If the target program is built with debug info like DWARF, uftrace can identify
number of arguments and their types automatically (when built with libdw).
Also arguments and return value of some well-known library functions are
provided even if the debug info is not available.  In these cases user don't
need to specify format of the arguments and return value manually - just
function name (or pattern) is enough.  In fact, manual argspec will suppress
the automatic argspec.

For example, the above example can be written like below:

    $ uftrace record -A . -R main ./hello
    Hello world

    $ uftrace replay -F main
    # DURATION     TID     FUNCTION
                [ 18948] | main(1, 0x7ffeeb7590b8) {
       7.183 us [ 18948] |   puts("Hello world");
       9.832 us [ 18948] | } = 0; /* main */

Note that argument pattern (".") matches to any character so it recorded
all (supported) functions.  It shows two arguments for "main" and a single
string argument for "puts".  If you simply want to see all arguments and
return values of every functions (if supported), use `-a`/`--auto-args` option.


DYNAMIC TRACING
===============
The uftrace tool supports dynamic function tracing which can be enabled at runtime (load-time, to be precise) on x86_64.  Before recording functions, normally you need to build the target program with `-pg` (or `-finstrument-functions`), then it has some performance impact because all functions call `mcount()`.

With dynamic tracing, you can trace specific functions only given by the `-P`/`--patch` option.  However you need to add some more compiler (gcc) options when building the target program.  The gcc 5.1 or more recent versions provide `-mfentry` and `-mnop-mcount` options which add instrumentation code (i.e. calling `mcount()` function) at the very beginning of a function and convert the instruction to a NOP.  Then it has almost zero performance overhead when running in a normal condition.  The uftrace can convert it back to call `mcount()` if users want to (using `-P` option).

The following example shows a error message when normally running uftrace with the executable built with `-pg -mfentry -mnop-mcount`. Because the binary doesn't call any instrumentation code (i.e. 'mcount').

    $ gcc -o abc -pg -mfentry -mnop-mcount tests/s-abc.c
    $ uftrace abc
    uftrace: /home/namhyung/project/uftrace/cmd-record.c:1305:check_binary
      ERROR: Can't find 'mcount' symbol in the 'abc'.
             It seems not to be compiled with -pg or -finstrument-functions flag
             which generates traceable code.  Please check your binary file.

But when the `-P a` patch option is used, and then only it can dynamically trace `a()`.

    $ uftrace record --no-libcall -P a abc
    $ uftrace replay
    # DURATION    TID     FUNCTION
       0.923 us [19379] | a();

In addition, you can enable all functions at load time using '.' that matches to any character in a regex pattern with `P` option.

    $ uftrace record --no-libcall -P . abc
    $ uftrace replay
    # DURATION    TID     FUNCTION
                [19387] | main() {
                [19387] |   a() {
                [19387] |     b() {
       0.940 us [19387] |       c();
       2.030 us [19387] |     } /* b */
       2.451 us [19387] |   } /* a */
       3.289 us [19387] | } /* main */

Clang/LLVM 4.0 provides a dynamic instrumentation technique called [X-ray](http://llvm.org/docs/XRay.html).  It's similar to a combination of `gcc -mfentry -mnop-mcount` and `-finstrument-functions`.  The uftrace also supports dynamic tracing on the executables built with the `X-ray`.

For example, you can build the target program by clang with the below option and equally use `-P` option for dynamic tracing like below:

    $ clang -fxray-instrument -fxray-instruction-threshold=1 -o abc-xray  tests/s-abc.c
    $ uftrace record -P main abc-xray
    $ uftrace replay
    # DURATION    TID     FUNCTION
                [11093] | main() {
       1.659 us [11093] |   getpid();
       5.963 us [11093] | } /* main */

    $ uftrace record -P . abc-xray
    $ uftrace replay
    # DURATION    TID     FUNCTION
                [11098] | main() {
                [11098] |   a() {
                [11098] |     b() {
                [11098] |       c() {
       0.753 us [11098] |         getpid();
       1.430 us [11098] |       } /* c */
       1.915 us [11098] |     } /* b */
       2.405 us [11098] |   } /* a */
       3.005 us [11098] | } /* main */


SCRIPT EXECUTION
================
The uftrace tool supports script execution for each function entry and exit.  The supported script is only Python 2.7 as of now.

The user can write four functions. 'uftrace_entry' and 'uftrace_exit' are executed whenever each function is executed at the entry and exit.  However 'uftrace_begin' and 'uftrace_end' are only executed once when the target program begins and ends.

    $ cat scripts/simple.py
    def uftrace_begin(ctx):
        print("program begins...")

    def uftrace_entry(ctx):
        func = ctx["name"]
        print("entry : " + func + "()")

    def uftrace_exit(ctx):
        func = ctx["name"]
        print("exit  : " + func + "()")

    def uftrace_end():
        print("program is finished")

The above script can be executed in record time as follows:

    $ uftrace record -S scripts/simple.py -F main tests/t-abc
    program begins...
    entry : main()
    entry : a()
    entry : b()
    entry : c()
    entry : getpid()
    exit  : getpid()
    exit  : c()
    exit  : b()
    exit  : a()
    exit  : main()
    program is finished

The 'ctx' variable is a dictionary type that contains the below information.

    /* context information passed to script */
    script_context = {
        int       tid;
        int       depth;
        long      timestamp;
        long      duration;    # exit only
        long      address;
        string    name;
        list      args;        # entry only (if available)
        value     retval;      # exit  only (if available)
    };

Each field in 'script_context' can be read inside the script.  Please see `uftrace-script`(1) for details about scripting.


SEE ALSO
========
`uftrace`(1), `uftrace-replay`(1), `uftrace-report`(1), `uftrace-recv`(1), `uftrace-script`(1), `uftrace-tui`(1)