Copyright (c) 2004, Sun Microsystems, Inc. All Rights Reserved.
The contents of this file are subject to the terms of the Common Development and Distribution License (the "License"). You may not use this file except in compliance with the License.
You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE or http://www.opensolaris.org/os/licensing. See the License for the specific language governing permissions and limitations under the License.
When distributing Covered Code, include this CDDL HEADER in each file and include the License file at usr/src/OPENSOLARIS.LICENSE. If applicable, add the following below this CDDL HEADER, with the fields enclosed by brackets "[]" replaced with your own identifying information: Portions Copyright [yyyy] [name of copyright owner]
prex [-o trace_file_name] [-l libraries] [-s kbytes_size] cmd [cmd-args]...
prex [-o trace_file_name] [-l libraries] [-s kbytes_size] -p pid
prex -k [-s kbytes_size]
The prex command is the part of the Solaris tracing architecture that controls probes in a process or the kernel. See tracing(3TNF) for an overview of this tracing architecture, including example source code using it.
prex is the application used for external control of probes. It automatically preloads the libtnfprobe library. prex locates all the probes in a target executable or the kernel and provides an interface for the user to manipulate them. It allows a probe to be turned on for tracing, debugging, or both. Tracing generates a TNF (Trace Normal Form) trace file that can be converted to ASCII by tnfdump(1) and used for performance analysis. Debugging generates a line to standard error whenever the probe is hit at run time.
prex does not work on static executables. It only works on dynamic executables.
There are three ways to invoke prex:
1. Use prex to start the target application cmd. In this case, the target application need not be built with a dependency on libtnfprobe. See TNF_PROBE(3TNF). prex sets the environment variable LD_PRELOAD to load libtnfprobe into the target process. See ld(1). prex then uses the environment variable PATH to find the target application.
2. Attach prex to a running application. In this case, the running target application should have libtnfprobe already linked in. Alternatively, the user may manually set LD_PRELOAD to include libtnfprobe.so.1 prior to invoking the target.
3. Use prex with the -k option to set prex to kernel mode. prex can then be used to control probes in the Solaris kernel. In kernel mode, additional commands are defined, and some commands that are valid in other modes are invalid. See Kernel Mode below.
In a future release of prex, the command language may be moved to a syntax that is supported by an existing scripting language like ksh(1). In the meantime, the interface to prex is uncommitted.
Commands should be in ASCII.
Each command is terminated with the NEWLINE character.
A command can be continued onto the next line by ending the previous line with the backslash ("\e") character.
Tokens in a command must be separated by whitespace (one or more spaces or tabs).
The "#" character implies that the rest of the line is a comment.
| Command Result |
| % prex a.out |
| Attaches prex to your program and starts prex. |
| prex> enable $all Enables all the probes. |
| prex> quit resume |
| Quits prex and resumes execution of program. |
There are two different methods of communicating with prex:
By specifications in a control file. During start-up, prex searches for a file named .prexrc in the directories specified below. prex does not stop at the first one it finds. This way a user can override any defaults that are set up. The search order is:
$HOME/ ./
By typing commands at the prex prompt.
The command language for both methods is the same and is specified in USAGE. The commands that return output will not make sense in a control file. The output will go to standard output.
When using prex on a target process, the target will be in one of two states, running or stopped. This can be detected by the presence or absence of the prex> prompt. If the prompt is absent, it means that the target process is running. Typing Control-C will stop the target pr ocess and return the user to the prompt. There is no guarantee that Control-C will return to a prex prompt immediately. For example, if the target process is stopped on a job control stop (SIGSTOP), then Control-C in prex will wait until the target has been continued (SIGCONT). See Signals to Target Program below for more information on signals and the target process.
The following options are supported: -k
kernel mode: prex is used to control probes in the Solaris kernel. In kernel mode, additional commands are defined, and some commands valid in other modes are invalid. See Kernel Mode below.
The libraries mentioned are linked in to the target application using LD_PRELOAD (see ld(1)). This option cannot be used when attaching to a running process. The argument to the -l option should be a space-separated string enclosed in double quotes. Each token in the string is a library name. It follows the LD_PRELOAD rules on how libraries should be specified and where they will be found.
File to be used for the trace output. trace_file_name is assumed to be relative to the current working directory of prex (that is, the directory that the user was in when prex was started). If prex attaches to a process that is already tracing, the new trace_file_name (if provided) will not be used. If no trace_file_name is specified, the default is /$TMPDIR/trace-pid where pid is the process id of the target program. If TMPDIR is not set, /tmp is used.
Maximum size of the output trace file in Kbytes. The default size of the trace kbytes_size is 4096 (2^10) bytes or 4 Mbytes for normal usage, and 384 or 384 kbytes in kernel mode. The minimum size that can be specified is 128 Kbytes. The trace file can be thought of as a least recently used circular buffer. Once the file has been filled, newer events will overwrite the older ones.
This section describes the usage of the prex utility.
Probes are specified by a list of space-separated selectors. Selectors are of the form:
attribute=value
(See TNF_PROBE(3TNF)). The "attribute=" is optional. If it is not specified, it defaults to "keys=".
The attribute or value (generically called "spec") can be any of the following: IDENT
Any sequence of letters, digits, _\|, \e\|, ., % not beginning with a digit. IDENT implies an exact match.
Usually used to escape reserved words (any commands in the command language). QUOTED_STR implies an exact match and has to be enclosed in single quotes (' ').
An ed(1) regular expression pattern match. REGEXP has to be enclosed in slashes (/ /), A / can be included in a REGEXP by escaping it with a backslash \e\|.
The following grammar explains the syntax.
selector_list ::= | /* empty */ selector_list selector selector ::= spec=spec | /* whitespace around `=' opt */ spec spec ::= IDENT | QUOTED_STR | REGEXP
The terminals in the above grammar are:
IDENT = [a-zA-Z_\e.%]{[a-zA-Z0-9_\e.%]}+
QUOTED_STR = '[^\en']*' /* any string in single quotes */
REGEXP = /[^\en/]*/ /* regexp's have to be in / / */
This is a list of the remaining grammar that is needed to understand the syntax of the command language (defined in next subsection):
filename ::= QUOTED_STR /* QUOTED_STR defined above */ spec_list ::= /* empty */ | spec_list spec /* spec defined above */ fcn_handle ::= &IDENT /* IDENT defined above */ set_name ::= $IDENT /* IDENT defined above */
1. Set Creation and Set Listing
create $set_name selector_list
list sets # list the defined sets
create can be used to define a set which contains probes that match the
selector_list. The set $all is pre-defined as /.*/ and it matches
all the probes.
2. Function Listing
list fcns # list the available fcn_handle
The user can list the different functions that can be connected to probe
points. Currently, only the debug function called &debug is available.
3. Commands to Connect and Disconnect Probe Functions
connect &fcn_handle $set_name
connect &fcn_handle selector_list
clear $set_name
clear selector_list
The connect command is used to connect probe functions (which must be
prefixed by `&') to probes. The probes are specified either as a single
set (with a `$'), or by explicitly listing the probe selectors in the
command. The probe function has to be one that is listed by the list fcns
command. This command does not enable the probes.
The clear command is used to disconnect all connected probe functions
from the specified probes.
4. Commands to Toggle the Tracing Mode
trace $set_name
trace selector_list
untrace $set_name
untrace selector_list
The trace and untrace commands are used to toggle the tracing
action of a probe point (that is, whether a probe will emit a trace record or
not if it is hit). This command does not enable the probes specified. Probes
have tracing on by default. The most efficient way to turn off tracing is by
using the disable command. untrace is useful if you want debug
output but no tracing. If so, set the state of the probe to enabled, untraced,
and the debug function connected.
5. Commands to Enable and Disable Probes
enable $set_name
enable selector_list
disable $set_name
disable selector_list
The enable and disable commands are used to control whether the
probes perform the action that they have been set up for. To trace a probe, it
has to be both enabled and traced (using the trace command). Probes are
disabled by default. The list history command is used to list the probe
control commands issued: connect, clear, trace,
untrace, enable, and disable. These are the commands that
are executed whenever a new shared object is brought in to the target program
by dlopen(3C). See the subsection, dlopen'ed Libraries, below for
more information.
The following table shows the actions that result from specific combinations of
tracing, enabling, and connecting:
Enabled or Tracing State Debug State Results Disabled (On/Off) (Connected/Cleared) In ------------------------------------------------------------ Enabled On Connected Tracing and Debugging Enabled On Cleared Tracing only Enabled Off Connected Debugging only Enabled Off Cleared Nothing Disabled On Connected Nothing Disabled On Cleared Nothing Disabled Off Connected Nothing Disabled Off Cleared Nothing
6. List History
list history # lists probe control command history
The list history command displays a list of the probe control commands
previously issued in the tracing session, for example, connect,
clear, trace, disable. Commands in the history list are
executed wherever a new shared object is brought into the target program by
dlopen(3C).
7. Commands to List Probes, List Values, or List Trace File Name
list spec_list probes $set_name # list probes $all
list spec_list probes selector_list # list name probes file=test.c
list values spec_list # list values keys given in spec_list
list tracefile # list tracefile
The first two commands list the selected attributes and values of the specified
probes. They can be used to check the state of a probe. The third command lists
the various values associated with the selected attributes. The fourth command
lists the current tracefile.
8. Help Command
help topic
To get a list of the help topics that are available, invoke the help
command with no arguments. If a topic argument is specified, help is
printed for that topic.
9. Source a File
source filename
The source command can be used to source a file of prex commands.
source can be nested (that is, a file can source another file).
filename is a quoted string.
10. Process Control
continue # resumes the target process
quit kill # quit prex, kill target
quit resume # quit prex, continue target
quit suspend # quit prex, leave target suspended
quit # quit prex (continue or kill target)
The default quit will continue the target process if prex attached
to it. Instead, if prex had started the target program, quit will
kill the target process.
Probes in shared objects that are brought in by dlopen(3C) are automatically set up according to the command history of prex. When a shared object is removed by a dlclose(3C), prex again needs to refresh its understanding of the probes in the target program. This implies that there is more work to do for dlopen(3C) and dlclose(3C) \(emso they will take slightly longer. If a user is not interested in this feature and doesn't want to interfere with dlopen(3C) and dlclose(3C), detach prex from the target to inhibit this feature.
prex does not interfere with signals that are delivered directly to the target program. However, prex receives all signals normally generated from the terminal, for example, Control-C (SIGINT), and Control-Z (SIGSTOP), and does not forward them to the target program. To signal the target program, use the kill(1) command from a shell.
Process managing applications like dbx, truss(1), and prex cannot operate on the same target program simultaneously. prex will not be able to attach to a target which is being controlled by another application. A user can trace and debug a program serially by the following method: first attach prex to target (or start target through prex), set up the probes using the command language, and then type quit suspend. The user can then attach dbx to the suspended process and debug it. A user can also suspend the target by sending it a SIGSTOP signal, and then by typing quit resume to prex. In this case, the user should also send a SIGCONT signal after invoking dbx on the stopped process (else dbx will be hung).
There are a few failure points that are possible when writing out events to a trace file, for example, system call failures. These failures result in a failure code being set in the target process. The target process continues normally, but no trace records are written. Whenever a user enters Control-C to prex to get to a prex prompt, prex will check the failure code in the target and inform the user if there was a tracing failure.
If the target program does a fork(2), any probes that the child encounters will cause events to be logged to the same trace file. Events are annotated with a process id, so it will be possible to determine which process a particular event came from. In multi-threaded programs, there is a race condition with a thread doing a fork while the other threads are still running. For the trace file not to get corrupted, the user should either use fork1(2), or make sure that all other threads are quiescent when doing a fork(2),
If the target program itself (not any children it may fork(2)) does an exec(2), prex detaches from the target and exits. The user can reconnect prex with prex -p pid.
A vfork(2) is generally followed quickly by an exec(2) in the child, and in the interim, the child borrows the parent's process while the parent waits for the exec(2). Any events logged by the child from the parent process will appear to have been logged by the parent.
Invoking prex with the -k flag causes prex to run in kernel mode. In kernel mode, prex controls probes in the Solaris kernel. See tnf_kernel_probes(4) for a list of available probes in the Solaris kernel. A few prex commands are unavailable in kernel mode; many other commands are valid in kernel mode only.
The -l, -o, and -p command-line options are not valid in kernel mode (that is, they may not be combined with the -k flag).
The rest of this section describes the differences in the prex command language when running prex in kernel mode.
1. prex will not stop the kernel When prex attaches to a running user program, it stops the user program. Obviously, it cannot do this when attaching to the kernel. Instead, prex provides a ``tracing master switch'': no probes will have any effect unless the tracing master switch is on. This allows the user to iteratively select probes to enable, then enable them all at once by turning on the master switch. The command
ktrace [ on | off ]
is used to inspect and set the value of the master switch. Without an argument,
prex reports the current state of the master switch.
Since prex will not stop or kill the kernel, the
quit resume
and
quit kill
commands are not valid in kernel mode.
2. No functions may be attached to probes in the kernel In particular, the debug function is unavailable in kernel mode.
3. Trace output is written to an in-core buffer In kernel mode, a trace output file is not generated directly, in order to allow probes to be placed in time-critical code. Instead, trace output is written to an in-core buffer, and copied out by a separate program, tnfxtract(1). The in-core buffer is not automatically created. The following prex command controls buffer allocation and deallocation:
buffer [ alloc [ size ] | dealloc ]
Without an argument, the buffer command reports the size of the currently
allocated buffer, if any. With an argument of alloc [size],
prex allocates a buffer of the given size. size is in bytes, with
an optional suffix of 'k' or 'm' specifying a multiplier of
1024 or 1048576, respectively. If no size is specified, the
size specified on the command line with the -s option is used as a
default. If the -s command line option was not used, the ``default
default'' is 384 kilobytes.
With an argument of dealloc, prex deallocates the trace buffer in
the kernel.
prex will reject attempts to turn the tracing master switch on when no
buffer is allocated, and to deallocate the buffer when the tracing master
switch is on. prex will refuse to allocate a buffer when one is already
allocated; use buffer dealloc first.
prex will not allocate a buffer larger than one-half of a machine's
physical memory.
4. prex supports per-process probe enabling in the kernel In kernel mode, it is possible to select a set of processes for which probes are enabled. No trace output will be written when other processes traverse these probe points. This is called "process filter mode". By default, process filter mode is off, and all processes cause the generation of trace records when they hit an enabled probe. Some kernel events such as interrupts cannot be associated with a particular user process. By convention, these events are considered to be generated by process id 0. prex provides commands to turn process filter mode on and off, to get the current status of the process filter mode switch, to add and delete processes (by process id) from the process filter set, and to list the current process filter set. The process filter set is maintained even when process filter mode is off, but has no effect unless process filter mode is on. When a process in the process filter set exits, its process id is automatically deleted from the process filter set. The command:
pfilter [ on | off | add pidlist | delete pidlist ]
controls the process filter switch, and process filter set membership. With no
arguments, pfilter prints the current process filter set and the state of
the process filter mode switch:
on or off
set the state of the process filter mode switch.
delete pidlist
add or delete processes from the process filter set. pidlist is a comma-separated list of one or more process ids.
See tracing(3TNF) for complete examples showing, among other things, the use of prex to do simple probe control.
When either the process or kernel is started, all probes are disabled.
Example 1 Set creation and set listing
create $out name=/out/ # $out = probes with "out" in
# value of "name" attribute
create $foo /page/ name=biodone # $foo = union of
# probes with "page" in value of keys attribute
# probes with "biodone" as value of "name" attribute
list sets # list the defined sets
list fcns # list the defined probe fcns
Example 2 Commands to trace and connect probe functions
trace foobar='on' # exact match on foobar attribute
trace $all # trace all probes (predefined set $all)
connect &debug $foo # connect debug func to probes in $foo
Example 3 Commands to enable and disable probes
enable $all # enable all probes
enable /vm/ name=alloc # enable the specified probes
disable $foo # disable probes in set $foo
list history # list probe control commands issued
Example 4 Process control
continue # resumes the target process
^C # stop target; give control to prex
quit resume # exit prex, leave process running
# and resume execution of program
Example 5 Kernel mode
buffer alloc 2m # allocate a 2 Megabyte buffer
enable $all # enable all probes
trace $all # trace all probes
ktrace on # turn tracing on
ktrace off # turn tracing back off
pfilter on # turn process filter mode on
pfilter add 1379 # add pid 1379 to process filter
ktrace on # turn tracing on
# (only pid 1379 will be traced)
local prex initialization file
user's prex initialization file
process files
ed(1), kill(1), ksh(1), ld(1), tnfdump(1), tnfxtract(1), truss(1), exec(2), fork(2), fork1(2), vfork(2), TNF_DECLARE_RECORD(3TNF), TNF_PROBE(3TNF), dlclose(3C), dlopen(3C), gethrtime(3C), libtnfctl(3TNF), tnf_process_disable(3TNF), tracing(3TNF), tnf_kernel_probes(4), attributes(5)
Currently, the only probe function that is available is the &debug function. When this function is executed, it prints out the arguments sent in to the probe as well as the value associated with the sunw%debug attribute in the detail field (if any) to stderr.
For example, for the following probe point:
TNF_PROBE_2(input_values, "testapp main", "sunw%debug 'have read input values successfully'", tnf_long, int_input, x, tnf_string, string_input, input);
If x was 100 and input was the string "success", then the output of the debug probe function would be:
probe input_values; sunw%debug "have read input values successfully"; int_input=100; string_input="success";
Some non-SPARC hardware lacks a true high-resolution timer, causing gethrtime() to return the same value multiple times in succession. This can lead to problems in how some tools interpret the trace file. This situation can be improved by interposing a version of gethrtime(), which causes these successive values to be artificially incremented by one nanosecond:
hrtime_t
gethrtime()
{
static mutex_t lock;
static hrtime_t (*real_gethrtime)(void) = NULL;
static hrtime_t last_time = 0;
hrtime_t this_time;
if (real_gethrtime == NULL) {
real_gethrtime =
(hrtime_t (*)(void)) dlsym(RTLD_NEXT, "gethrtime");
}
this_time = real_gethrtime();
mutex_lock(&lock);
if (this_time <= last_time)
this_time = ++last_time;
else
last_time = this_time;
mutex_unlock(&lock);
return (this_time);
}
Of course, this does not increase the resolution of the timer, so timestamps for individual events are still relatively inaccurate. But this technique maintains ordering, so that if event A causes event B, B never appears to happen before or at the same time as A.
dbx is available with the Sun Workshop Products.
prex should issue a notification when a process id has been automatically deleted from the filter set.
There is a known bug in prex which can result in this message:
Tracing shut down in target program due to an internal error - Please restart prex and target
When prex runs as root, and the target process is not root, and the tracefile is placed in a directory where it cannot be removed and re-created (a directory with the sticky bit on, like /tmp),mm then the target process will not be able to open the tracefile when it needs to. This results in tracing being disabled.
Changing any of the circumstances listed above should fix the problem. Either don't run prex as root, or run the target process as root, or specify the tracefile in a directory other than /tmp.