Skip to content

Latest commit

 

History

History
204 lines (147 loc) · 11 KB

bpf-usdt.md

File metadata and controls

204 lines (147 loc) · 11 KB

Using BPF Tools: Node and JVM USDT Probes

In this lab, you will experiment with discovering USDT probes in a running process or an on-disk executable, enabling these probes at runtime, and collecting them using BPF's USDT support in the argdist and trace tools.


Task 1: Discovering USDT Probes

The tplist tool from the BCC collection can discover USDT probes in an on-disk executable, library, or running process. It can also tell you a lot of details about the probe locations and arguments. Unfortunately, it doesn't tell you the meaning of these arguments -- you often have to discover them yourself by using the documentation. In this section, you will walk through this process for the Node.js main executable, node.

If you're using an instructor-provided appliance, it should already have Node cloned and configured with USDT support. If you're using your own machine, clone the Node.js repository and build it with the --with-dtrace flag, which enables USDT support, using the following commands:

$ git clone https://github.com/nodejs/node.git
$ cd node
$ git checkout v6.2.1   # or whatever version is currently stable
$ ./configure --prefix=/opt/node --with-dtrace
$ make -j 3
$ sudo make install

Now, you can use tplist to discover USDT probes in the Node executable. Run the following command (provide the full paths to tplist and node as necessary):

$ tplist -l node 

This should print out a number of USDT probes, including node:http__server__request and node:gc__start. There are more probes embedded in a typical Node process, however -- you can see exactly what's available (possibly through additional libraries) by running node in the background and then using the following command:

$ tplist -p $(pidof node)

You should now see USDT probes from libc and libstdcxx as well. You can also ask tplist to print out more details about each probe. For example, here are the arguments for the node:http__server_request probe as reported by tplist:

$ tplist -l node -vv '*server__request'
/home/vagrant/node/out/Release/node node:http__server__request [sema 0x1606c34]
  location 0xef4854 raw args: 8@%r14 8@%rax 8@-4328(%rbp) -4@-4332(%rbp) 8@-4288(%rbp) 8@-4296(%rbp) -4@-4336(%rbp)
    8 unsigned bytes @ register %r14
    8 unsigned bytes @ register %rax
    8 unsigned bytes @ -4328(%rbp)
    4   signed bytes @ -4332(%rbp)
    8 unsigned bytes @ -4288(%rbp)
    8 unsigned bytes @ -4296(%rbp)
    4   signed bytes @ -4336(%rbp)

To figure out what these probes and their arguments mean, we need to inspect the documentation or some authoritative source. In many cases, software instrumented with USDT probes will also ship with a set of files intended for use with SystemTap or DTrace, which can consume these probes. Sure enough, there's a node.stp file that describes these events. Find the description for the node:http__server_request probe:

probe node_http_server_request = process("node").mark("http__server__request")
{
    remote = user_string($arg3);
    port = $arg4;
    method = user_string($arg5);
    url = user_string($arg6);
    fd = $arg7;

    probestr = sprintf("%s(remote=%s, port=%d, method=%s, url=%s, fd=%d)",
        $$name, remote, port, method, url, fd);
}

It is quite clear what the arguments mean now, and we can use them with the other tracing tools that can retrieve them.


Task 2: Enabling and Tracing Node USDT Probes with trace

Let's now make sure that we can trace some HTTP requests with the USDT support embedded in Node. For this experiment, you can build your own Node HTTP server, or just use the server.js file, reproduced here in its entirety:

var http = require('http');
var server = http.createServer(function (req, res) {
    res.end('Hello, world!');
});
server.listen(8080);

You can run this server using the following command:

$ node server.js

In a root shell, navigate to the tools directory under the BCC source, and use the following command to attach to the node:http__server__request USDT event and trace out a message whenever a request arrives. Note that the process id is required because the Node USDT events are not enabled by default. The tracing program must poke a global variable that the probing code reads in order to enable the probe, and that happens on a per-process basis.

# trace -p $(pidof node) 'u:/opt/node/node:http__server__request "%s %s", arg5, arg6'

Here, arg5 is going to be the HTTP method and arg6 is going to be the request URL -- as we discovered by inspecting the .stp file. Finally, make some HTTP requests to your server and make sure you see them in the trace:

$ curl localhost:8080
$ curl localhost:8080/index.html
$ curl 'localhost:8080/login?user=dave&pwd=123'

Task 3: Enabling and Tracing JVM USDT Probes with trace

OpenJDK is also instrumented with a number of USDT probes. Most of them are enabled out of the box, and some must be enabled with a special -XX:+ExtendedDTraceProbes flag, because they incur a performance penalty. To explore some of these probes, navigate to your $JAVA_HOME and take a look in the tapset directory (or online: tapset): :

$ cd /etc/alternatives/java_sdk
$ ls tapset/
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp
hotspot_gc-1.8.0.77-1.b03.fc22.x86_64.stp
hotspot_jni-1.8.0.77-1.b03.fc22.x86_64.stp
jstack-1.8.0.77-1.b03.fc22.x86_64.stp

These .stp files contain descriptions and declarations for a bunch of probes, including their arguments. As an example, try to find the gc_collect_tenured_begin and gc_collect_tenured_end probe descriptions.

Now, let's move to a more practical example. You are now going to trace a running Java application and see which classes are being loaded, by using the class_loaded probe. First, find its "documentation" by running the following command (or by looking online: hotspot-1.8.0.stp.in):

$ grep -A 10 'probe.*class_loaded' tapset/*.stp
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp:probe hotspot.class_loaded = 
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp-  process("/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.77-1.b03.fc22.x86_64/jre/lib/amd64/server/libjvm.so").mark("class__loaded")                                      
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp-{                           
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp-  name = "class_loaded";    
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp-  class = user_string_n($arg1, $arg2);                                                         
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp-  classloader_id = $arg3;   
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp-  is_shared = $arg4;        
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp-  probestr = sprintf("%s(class='%s',classloader_id=0x%x,is_shared=%d)",                        
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp-                     name, class, classloader_id, is_shared);                                  
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp-}                           
hotspot-1.8.0.77-1.b03.fc22.x86_64.stp-

It looks like there are four arguments, and the class name being loaded is given as as the first argument. At this point, we can run a Java application (this is Slowy.java):

$ java slowy/App

And now discover the available tracepoints using tplist:

# jps
31156 App
31398 Jps
# tplist -p 31156 '*class*loaded'
/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.77-1.b03.fc22.x86_64/jre/lib/amd64/server/libjvm.so hotspot:class__unloaded
/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.77-1.b03.fc22.x86_64/jre/lib/amd64/server/libjvm.so hotspot:class__loaded

And finally we can trace the interesting tracepoint with trace (replace the path to libjvm.so with the appropriate value from your system):

# trace 'u:/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.77-1.b03.fc22.x86_64/jre/lib/amd64/server/libjvm.so:class__loaded "%s", arg1'

At this point, you should get a trace message whenever a Java app loads a class. For example, let the Slowy app terminate -- are there any classes being loaded on the shutdown path?


Task 4: Using JVM Extended USDT Probes with argdist

Finally, let's experiment with one of the extended probes, which isn't always enabled because of its overhead. A couple of these probes are called on method entry and exit, and can be used for simple profiling (albeit with a pretty considerable cost). First, let's find the probes in the .stp file. They are called method_entry and method_return (or see online: :method_entry and method_return):

$ grep -A 10 'probe.*method_entry' *.stp
$ grep -A 10 'probe.*method_return' *.stp

As you can see from the results, the class name, method, and signature are available in $arg2, $arg4, and $arg6, respectively. Both probes use the same convention. Now, let's use argdist to figure out which methods are being called most often:

# argdist -C 'u:/usr/lib/.../libjvm.so:method__entry():char*:arg4' -T 5

In another shell, run the Slowy app with the extended probes configured (and disable inlining so that no methods are optimized away):

$ java -XX:-Inline -XX:+ExtendedDTraceProbes slowy/App

Every five seconds, you should now get a printout of the methods that were called most frequently. It's just the method names, but that's not too shabby. Now, suppose you actually wanted the execution time of each method. That's something argdist currently can't handle (it can trace function entry and return, but not probe pairs like we have in this case). However, we can get some basic results with trace and post-process them later -- or build a new BCC tool! For now, let's use trace to get all method entry and return events and print them out nicely:

# trace -o 'u:/usr/lib/.../libjvm.so:method__entry "%s.%s", arg2, arg4' 'u:/usr/lib/.../libjvm.so:method__return "%s.%s", arg2, arg4'

If you run the Slowy app again with the extended probes configured, you should get a bunch of printouts for the methods being entered and exited.

Unfortunately, you might also see printouts indicating that events are being lost -- we are not printing them quickly enough and the cyclic buffer used for kernel-mode and user-mode communication fills up, discarding some events. This is one of the reasons we're advocating for building custom BPF tools -- they will almost always be more efficient than trying to grab everything from the kernel and perform aggregations on a large number of events in user space.


Bonus: Discovering Probes in Oracle JVM

So, you think OpenJDK USDT probes are cool? For this bonus task, try to figure out if the Oracle HotSpot VM is also instrumented with USDT probes. See if you can trace some of them with argdist, trace, or both.