Predicates in DTrace

Predicates are expressions enclosed in slashes / / that are evaluated at probe firing time to determine whether the associated actions should be executed. Predicates are the primary conditional construct used for building more complex control flow in a D program. You can omit the predicate section of the probe clause entirely for any probe, in which case the actions are always executed when the probe fires.

Predicate expressions can use any of the previously described D operators and may refer to any D data objects such as variables and constants. The predicate expression must evaluate to a value of integer or pointer type so that it can be considered as true or false. As with all D expressions, a zero value is interpreted as false and any non-zero value is interpreted as true.

One major difference between D and other programming languages such as C, C++, and the Java programming language is the absence of control-flow constructs such as if-statements and loops. D program clauses are written as single straight-line statement lists that trace an optional, fixed amount of data. D does provide the ability to conditionally trace data and modify control flow using logical expressions called predicates that can be used to prefix program clauses. A predicate expression is evaluated at probe firing time prior to executing any of the statements associated with the corresponding clause. If the predicate evaluates to true, represented by any non-zero value, the statement list is executed. If the predicate is false, represented by a zero value, none of the statements are executed and the probe firing is ignored.

Example 2-1 Creating a 10-Second Countdown Timer

Type the following source code and save it in a file named countdown.d:

dtrace:::BEGIN
{
        i = 10;
}

profile:::tick-1sec
/i > 0/
{
        trace(i--);
}

profile:::tick-1sec
/i == 0/
{
        trace("blastoff!");
        exit(0);
}

This D program implements a 10-second countdown timer using predicates. When executed, countdown.d counts down from 10 and then prints a message and exits:

# dtrace -s countdown.d
dtrace: script 'countdown.d' matched 3 probes
CPU     ID                    FUNCTION:NAME
        0  25499                       :tick-1sec 10
        0  25499                       :tick-1sec 9
        0  25499                       :tick-1sec 8
        0  25499                       :tick-1sec 7
        0  25499                       :tick-1sec 6
        0  25499                       :tick-1sec 5
        0  25499                       :tick-1sec 4
        0  25499                       :tick-1sec 3
        0  25499                       :tick-1sec 2
        0  25499                       :tick-1sec 1
        0  25499                       :tick-1sec blastoff!
#

This example uses the BEGIN probe to initialize an integer i to 10 to begin the countdown. The program uses the tick-1sec probe to implement a timer that fires once per second. Notice that in countdown.d, the tick-1sec probe description is used in two different clauses, each with a different predicate and action list. The predicate is a logical expression surrounded by enclosing slashes / / that appears after the probe name and before the braces { } that surround the clause statement list.

The first predicate tests whether i is greater than zero, indicating that the timer is still running:

profile:::tick-1sec
/i > 0/
{
        trace(i--);
}

The relational operator > means greater than and returns the integer value zero for false and one for true. All of the C relational operators are supported in D; the complete list is found in Types, Operators, and Expressions in DTrace. If i is not yet zero, the script traces i and then decrements it by one using the - operator.

The second predicate uses the == operator to return true when i is exactly equal to zero, indicating that the countdown is complete:

profile:::tick-1sec
/i == 0/
{
        trace("blastoff!");
        exit(0);
}

The countdown.d program uses a sequence of characters enclosed in double quotes, called a string constant, to print a final message when the countdown is complete. The exit() function is then used to exit dtrace and return to the shell prompt.

If you look back at the structure of countdown.d, you will see that by creating two clauses with the same probe description, but different predicates and actions, effectively created the logical flow:

i = 10
once per second,
        if i is greater than zero
                trace(i--);
        otherwise if i is equal to zero
                trace("blastoff!");
                exit(0);

When you want to write complex programs using predicates, try to first visualize your algorithm in this manner, and then transform each path of your conditional constructs into a separate clause and predicate.

Now combine predicates with a new provider, the syscall provider, and create the first real D tracing program. The syscall provider permits you to enable probes on entry to or return from any Oracle Solaris system call. The next example uses DTrace to observe every time your shell performs a read or write system call. For more information, see the read(2) and write(2) man page. First, open two terminal windows, one to use for DTrace and the other containing the shell process you are going to watch. In the second window, type the following command to obtain the process ID of this shell:

# echo $$
12345

Now go back to your first terminal window and type the following D program and save it in a file named rw.d. As you type in the program, replace the integer constant 12345 with the process ID of the shell that was printed in response to your echo command.

syscall::read:entry,
syscall::write:entry
/pid == 12345/
{
}

Notice that the body of rw.d's probe clause is left empty because the program is only intended to trace notification of probe firings and not to trace any additional data. Once you have typed in rw.d, use dtrace to start your experiment and then go to your second shell window and type a few commands, pressing return after each command. As you type, you should see dtrace report probe firings in your first window, similar to the following example:

# dtrace -s rw.d
dtrace: script 'rw.d' matched 2 probes
CPU     ID                    FUNCTION:NAME
        0     34                      write:entry
        0     32                      read:entry
        0     34                      write:entry
        0     32                      read:entry
        0     34                      write:entry
        0     32                      read:entry
        0     34                      write:entry
        0     32                      read:entry
...

The read and write system calls read a character from your terminal window and echo back the result. This example includes many of the concepts described so far and a few new ones as well. First, to instrument read and write in the same manner, the script uses a single probe clause with multiple probe descriptions by separating the descriptions with commas like this:

syscall::read:entry,
syscall::write:entry

For readability, each probe description appears on its own line. This arrangement is not strictly required, but it makes for a more readable script. Next the script defines a predicate that matches only those system calls that are executed by your shell process:

/pid == 12345/

The predicate uses the predefined DTrace variable pid, which always evaluates to the process ID associated with the thread that fired the corresponding probe. DTrace provides many built-in variable definitions for useful things like the process ID. The following table lists a few DTrace variables you can use to write your first D program.

Variable	Data Type	Meaning
errno	int	Current errno value for system calls
execname	string	Name of the current process's executable file
pid	pid_t	Process ID of the current process
tid	id_t	Thread ID of the current thread
probeprov	string	Current probe description's provider field
probemod	string	Current probe description's module field
probefunc	string	Current probe description's function field
probename	string	Current probe description's name field

Now that you have written a real instrumentation program, try experimenting with it on different processes running on your system by changing the process ID and the system call probes that are instrumented. Then, you can make one more simple change and turn rw.d into a very simple version of a system call tracing tool like truss. An empty probe description field acts as a wildcard, matching any probe, so change the program to the following new source code to trace any system call executed by your shell:

syscall:::entry
/pid == 12345/
{
}

Try typing a few commands in the shell such as cd, ls, and date to see what your DTrace program reports.